Datasets:

urialon

/

converted_qasper

like 0

b2b0321b0aaf58c3aa9050906ade6ef35874c5c1

b2b0321b0aaf58c3aa9050906ade6ef35874c5c1_0

Q: How large is the dataset? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

$150,000$ tweets

4e9684fd68a242cb354fa6961b0e3b5c35aae4b6

4e9684fd68a242cb354fa6961b0e3b5c35aae4b6_0

Q: What is the results of multimodal compared to unimodal models? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

Unimodal LSTM vs Best Multimodal (FCM) - F score: 0.703 vs 0.704 - AUC: 0.732 vs 0.734 - Mean Accuracy: 68.3 vs 68.4

2e632eb5ad611bbd16174824de0ae5efe4892daf

2e632eb5ad611bbd16174824de0ae5efe4892daf_0

Q: What is author's opinion on why current multimodal models cannot outperform models analyzing only text? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

Noisy data, Complexity and diversity of multimodal relations, Small set of multimodal examples

d1ff6cba8c37e25ac6b261a25ea804d8e58e09c0

d1ff6cba8c37e25ac6b261a25ea804d8e58e09c0_0

Q: What metrics are used to benchmark the results? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

F-score, Area Under the ROC Curve (AUC), mean accuracy (ACC), Precision vs Recall plot, ROC curve (which plots the True Positive Rate vs the False Positive Rate)

24c0f3d6170623385283dfda7f2b6ca2c7169238

24c0f3d6170623385283dfda7f2b6ca2c7169238_0

Q: How is data collected, manual collection or Twitter api? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

Twitter API

21a9f1cddd7cb65d5d48ec4f33fe2221b2a8f62e

21a9f1cddd7cb65d5d48ec4f33fe2221b2a8f62e_0

Q: How many tweats does MMHS150k contains, 150000? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

$150,000$ tweets

a0ef0633d8b4040bf7cdc5e254d8adf82c8eed5e

a0ef0633d8b4040bf7cdc5e254d8adf82c8eed5e_0

Q: What unimodal detection models were used? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

single layer LSTM with a 150-dimensional hidden state for hate / not hate classification

b0799e26152197aeb3aa3b11687a6cc9f6c31011

b0799e26152197aeb3aa3b11687a6cc9f6c31011_0

Q: What different models for multimodal detection were proposed? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

Feature Concatenation Model (FCM), Spatial Concatenation Model (SCM), Textual Kernels Model (TKM)

4ce4db7f277a06595014db181342f8cb5cb94626

4ce4db7f277a06595014db181342f8cb5cb94626_0

Q: What annotations are available in the dataset - tweat used hate speach or not? Text: Introduction Social Media platforms such as Facebook, Twitter or Reddit have empowered individuals' voices and facilitated freedom of expression. However they have also been a breeding ground for hate speech and other types of online harassment. Hate speech is defined in legal literature as speech (or any form of expression) that expresses (or seeks to promote, or has the capacity to increase) hatred against a person or a group of people because of a characteristic they share, or a group to which they belong BIBREF0. Twitter develops this definition in its hateful conduct policy as violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease. In this work we focus on hate speech detection. Due to the inherent complexity of this task, it is important to distinguish hate speech from other types of online harassment. In particular, although it might be offensive to many people, the sole presence of insulting terms does not itself signify or convey hate speech. And, the other way around, hate speech may denigrate or threaten an individual or a group of people without the use of any profanities. People from the african-american community, for example, often use the term nigga online, in everyday language, without malicious intentions to refer to folks within their community, and the word cunt is often used in non hate speech publications and without any sexist purpose. The goal of this work is not to discuss if racial slur, such as nigga, should be pursued. The goal is to distinguish between publications using offensive terms and publications attacking communities, which we call hate speech. Modern social media content usually include images and text. Some of these multimodal publications are only hate speech because of the combination of the text with a certain image. That is because, as we have stated, the presence of offensive terms does not itself signify hate speech, and the presence of hate speech is often determined by the context of a publication. Moreover, users authoring hate speech tend to intentionally construct publications where the text is not enough to determine they are hate speech. This happens especially in Twitter, where multimodal tweets are formed by an image and a short text, which in many cases is not enough to judge them. In those cases, the image might give extra context to make a proper judgement. Fig. FIGREF5 shows some of such examples in MMHS150K. The contributions of this work are as follows: [noitemsep,leftmargin=*] We propose the novel task of hate speech detection in multimodal publications, collect, annotate and publish a large scale dataset. We evaluate state of the art multimodal models on this specific task and compare their performance with unimodal detection. Even though images are proved to be useful for hate speech detection, the proposed multimodal models do not outperform unimodal textual models. We study the challenges of the proposed task, and open the field for future research. Related Work ::: Hate Speech Detection The literature on detecting hate speech on online textual publications is extensive. Schmidt and Wiegand BIBREF1 recently provided a good survey of it, where they review the terminology used over time, the features used, the existing datasets and the different approaches. However, the field lacks a consistent dataset and evaluation protocol to compare proposed methods. Saleem et al. BIBREF2 compare different classification methods detecting hate speech in Reddit and other forums. Wassem and Hovy BIBREF3 worked on hate speech detection on twitter, published a manually annotated dataset and studied its hate distribution. Later Wassem BIBREF4 extended the previous published dataset and compared amateur and expert annotations, concluding that amateur annotators are more likely than expert annotators to label items as hate speech. Park and Fung BIBREF5 worked on Wassem datasets and proposed a classification method using a CNN over Word2Vec BIBREF6 word embeddings, showing also classification results on racism and sexism hate sub-classes. Davidson et al. BIBREF7 also worked on hate speech detection on twitter, publishing another manually annotated dataset. They test different classifiers such as SVMs and decision trees and provide a performance comparison. Malmasi and Zampieri BIBREF8 worked on Davidson's dataset improving his results using more elaborated features. ElSherief et al. BIBREF9 studied hate speech on twitter and selected the most frequent terms in hate tweets based on Hatebase, a hate expression repository. They propose a big hate dataset but it lacks manual annotations, and all the tweets containing certain hate expressions are considered hate speech. Zhang et al. BIBREF10 recently proposed a more sophisticated approach for hate speech detection, using a CNN and a GRU BIBREF11 over Word2Vec BIBREF6 word embeddings. They show experiments in different datasets outperforming previous methods. Next, we summarize existing hate speech datasets: [noitemsep,leftmargin=*] RM BIBREF10: Formed by $2,435$ tweets discussing Refugees and Muslims, annotated as hate or non-hate. DT BIBREF7: Formed by $24,783$ tweets annotated as hate, offensive language or neither. In our work, offensive language tweets are considered as non-hate. WZ-LS BIBREF5: A combination of Wassem datasets BIBREF4, BIBREF3 labeled as racism, sexism, neither or both that make a total of $18,624$ tweets. Semi-Supervised BIBREF9: Contains $27,330$ general hate speech Twitter tweets crawled in a semi-supervised manner. Although often modern social media publications include images, not too many contributions exist that exploit visual information. Zhong et al. BIBREF12 worked on classifying Instagram images as potential cyberbullying targets, exploiting both the image content, the image caption and the comments. However, their visual information processing is limited to the use of features extracted by a pre-trained CNN, the use of which does not achieve any improvement. Hosseinmardi et al. BIBREF13 also address the problem of detecting cyberbullying incidents on Instagram exploiting both textual and image content. But, again, their visual information processing is limited to use the features of a pre-trained CNN, and the improvement when using visual features on cyberbullying classification is only of 0.01%. Related Work ::: Visual and Textual Data Fusion A typical task in multimodal visual and textual analysis is to learn an alignment between feature spaces. To do that, usually a CNN and a RNN are trained jointly to learn a joint embedding space from aligned multimodal data. This approach is applied in tasks such as image captioning BIBREF14, BIBREF15 and multimodal image retrieval BIBREF16, BIBREF17. On the other hand, instead of explicitly learning an alignment between two spaces, the goal of Visual Question Answering (VQA) is to merge both data modalities in order to decide which answer is correct. This problem requires modeling very precise correlations between the image and the question representations. The VQA task requirements are similar to our hate speech detection problem in multimodal publications, where we have a visual and a textual input and we need to combine both sources of information to understand the global context and make a decision. We thus take inspiration from the VQA literature for the tested models. Early VQA methods BIBREF18 fuse textual and visual information by feature concatenation. Later methods, such as Multimodal Compact Bilinear pooling BIBREF19, utilize bilinear pooling to learn multimodal features. An important limitation of these methods is that the multimodal features are fused in the latter model stage, so the textual and visual relationships are modeled only in the last layers. Another limitation is that the visual features are obtained by representing the output of the CNN as a one dimensional vector, which losses the spatial information of the input images. In a recent work, Gao et al. BIBREF20 propose a feature fusion scheme to overcome these limitations. They learn convolution kernels from the textual information –which they call question-guided kernels– and convolve them with the visual information in an earlier stage to get the multimodal features. Margffoy-Tuay et al. BIBREF21 use a similar approach to combine visual and textual information, but they address a different task: instance segmentation guided by natural language queries. We inspire in these latest feature fusion works to build the models for hate speech detection. The MMHS150K dataset Existing hate speech datasets contain only textual data. Moreover, a reference benchmark does not exists. Most of the published datasets are crawled from Twitter and distributed as tweet IDs but, since Twitter removes reported user accounts, an important amount of their hate tweets is no longer accessible. We create a new manually annotated multimodal hate speech dataset formed by $150,000$ tweets, each one of them containing text and an image. We call the dataset MMHS150K, and made it available online . In this section, we explain the dataset creation steps. The MMHS150K dataset ::: Tweets Gathering We used the Twitter API to gather real-time tweets from September 2018 until February 2019, selecting the ones containing any of the 51 Hatebase terms that are more common in hate speech tweets, as studied in BIBREF9. We filtered out retweets, tweets containing less than three words and tweets containing porn related terms. From that selection, we kept the ones that included images and downloaded them. Twitter applies hate speech filters and other kinds of content control based on its policy, although the supervision is based on users' reports. Therefore, as we are gathering tweets from real-time posting, the content we get has not yet passed any filter. The MMHS150K dataset ::: Textual Image Filtering We aim to create a multimodal hate speech database where all the instances contain visual and textual information that we can later process to determine if a tweet is hate speech or not. But a considerable amount of the images of the selected tweets contain only textual information, such as screenshots of other tweets. To ensure that all the dataset instances contain both visual and textual information, we remove those tweets. To do that, we use TextFCN BIBREF22, BIBREF23 , a Fully Convolutional Network that produces a pixel wise text probability map of an image. We set empirical thresholds to discard images that have a substantial total text probability, filtering out $23\%$ of the collected tweets. The MMHS150K dataset ::: Annotation We annotate the gathered tweets using the crowdsourcing platform Amazon Mechanical Turk. There, we give the workers the definition of hate speech and show some examples to make the task clearer. We then show the tweet text and image and we ask them to classify it in one of 6 categories: No attacks to any community, racist, sexist, homophobic, religion based attacks or attacks to other communities. Each one of the $150,000$ tweets is labeled by 3 different workers to palliate discrepancies among workers. We received a lot of valuable feedback from the annotators. Most of them had understood the task correctly, but they were worried because of its subjectivity. This is indeed a subjective task, highly dependent on the annotator convictions and sensitivity. However, we expect to get cleaner annotations the more strong the attack is, which are the publications we are more interested on detecting. We also detected that several users annotate tweets for hate speech just by spotting slur. As already said previously, just the use of particular words can be offensive to many people, but this is not the task we aim to solve. We have not included in our experiments those hits that were made in less than 3 seconds, understanding that it takes more time to grasp the multimodal context and make a decision. We do a majority voting between the three annotations to get the tweets category. At the end, we obtain $112,845$ not hate tweets and $36,978$ hate tweets. The latest are divided in $11,925$ racist, $3,495$ sexist, $3,870$ homophobic, 163 religion-based hate and $5,811$ other hate tweets (Fig. FIGREF17). In this work, we do not use hate sub-categories, and stick to the hate / not hate split. We separate balanced validation ($5,000$) and test ($10,000$) sets. The remaining tweets are used for training. We also experimented using hate scores for each tweet computed given the different votes by the three annotators instead of binary labels. The results did not present significant differences to those shown in the experimental part of this work, but the raw annotations will be published nonetheless for further research. As far as we know, this dataset is the biggest hate speech dataset to date, and the first multimodal hate speech dataset. One of its challenges is to distinguish between tweets using the same key offensive words that constitute or not an attack to a community (hate speech). Fig. FIGREF18 shows the percentage of hate and not hate tweets of the top keywords. Methodology ::: Unimodal Treatment ::: Images. All images are resized such that their shortest size has 500 pixels. During training, online data augmentation is applied as random cropping of $299\times 299$ patches and mirroring. We use a CNN as the image features extractor which is an Imagenet BIBREF24 pre-trained Google Inception v3 architecture BIBREF25. The fine-tuning process of the Inception v3 layers aims to modify its weights to extract the features that, combined with the textual information, are optimal for hate speech detection. Methodology ::: Unimodal Treatment ::: Tweet Text. We train a single layer LSTM with a 150-dimensional hidden state for hate / not hate classification. The input dimensionality is set to 100 and GloVe BIBREF26 embeddings are used as word input representations. Since our dataset is not big enough to train a GloVe word embedding model, we used a pre-trained model that has been trained in two billion tweets. This ensures that the model will be able to produce word embeddings for slang and other words typically used in Twitter. To process the tweets text before generating the word embeddings, we use the same pipeline as the model authors, which includes generating symbols to encode Twitter special interactions such as user mentions (@user) or hashtags (#hashtag). To encode the tweet text and input it later to multimodal models, we use the LSTM hidden state after processing the last tweet word. Since the LSTM has been trained for hate speech classification, it extracts the most useful information for this task from the text, which is encoded in the hidden state after inputting the last tweet word. Methodology ::: Unimodal Treatment ::: Image Text. The text in the image can also contain important information to decide if a publication is hate speech or not, so we extract it and also input it to our model. To do so, we use Google Vision API Text Detection module BIBREF27. We input the tweet text and the text from the image separately to the multimodal models, so it might learn different relations between them and between them and the image. For instance, the model could learn to relate the image text with the area in the image where the text appears, so it could learn to interpret the text in a different way depending on the location where it is written in the image. The image text is also encoded by the LSTM as the hidden state after processing its last word. Methodology ::: Multimodal Architectures The objective of this work is to build a hate speech detector that leverages both textual and visual data and detects hate speech publications based on the context given by both data modalities. To study how the multimodal context can boost the performance compared to an unimodal context we evaluate different models: a Feature Concatenation Model (FCM), a Spatial Concatenation Model (SCM) and a Textual Kernels Model (TKM). All of them are CNN+RNN models with three inputs: the tweet image, the tweet text and the text appearing in the image (if any). Methodology ::: Multimodal Architectures ::: Feature Concatenation Model (FCM) The image is fed to the Inception v3 architecture and the 2048 dimensional feature vector after the last average pooling layer is used as the visual representation. This vector is then concatenated with the 150 dimension vectors of the LSTM last word hidden states of the image text and the tweet text, resulting in a 2348 feature vector. This vector is then processed by three fully connected layers of decreasing dimensionality $(2348, 1024, 512)$ with following corresponding batch normalization and ReLu layers until the dimensions are reduced to two, the number of classes, in the last classification layer. The FCM architecture is illustrated in Fig. FIGREF26. Methodology ::: Multimodal Architectures ::: Spatial Concatenation Model (SCM) Instead of using the latest feature vector before classification of the Inception v3 as the visual representation, in the SCM we use the $8\times 8\times 2048$ feature map after the last Inception module. Then we concatenate the 150 dimension vectors encoding the tweet text and the tweet image text at each spatial location of that feature map. The resulting multimodal feature map is processed by two Inception-E blocks BIBREF28. After that, dropout and average pooling are applied and, as in the FCM model, three fully connected layers are used to reduce the dimensionality until the classification layer. Methodology ::: Multimodal Architectures ::: Textual Kernels Model (TKM) The TKM design, inspired by BIBREF20 and BIBREF21, aims to capture interactions between the two modalities more expressively than concatenation models. As in SCM we use the $8\times 8\times 2048$ feature map after the last Inception module as the visual representation. From the 150 dimension vector encoding the tweet text, we learn $K_t$ text dependent kernels using independent fully connected layers that are trained together with the rest of the model. The resulting $K_t$ text dependent kernels will have dimensionality of $1\times 1\times 2048$. We do the same with the feature vector encoding the image text, learning $K_{it}$ kernels. The textual kernels are convolved with the visual feature map in the channel dimension at each spatial location, resulting in a $8\times 8\times (K_i+K_{it})$ multimodal feature map, and batch normalization is applied. Then, as in the SCM, the 150 dimension vectors encoding the tweet text and the tweet image text are concatenated at each spatial dimension. The rest of the architecture is the same as in SCM: two Inception-E blocks, dropout, average pooling and three fully connected layers until the classification layer. The number of tweet textual kernels $K_t$ and tweet image textual kernels $K_it$ is set to $K_t = 10$ and $K_it = 5$. The TKM architecture is illustrated in Fig. FIGREF29. Methodology ::: Multimodal Architectures ::: Training We train the multimodal models with a Cross-Entropy loss with Softmax activations and an ADAM optimizer with an initial learning rate of $1e-4$. Our dataset suffers from a high class imbalance, so we weight the contribution to the loss of the samples to totally compensate for it. One of the goals of this work is to explore how every one of the inputs contributes to the classification and to prove that the proposed model can learn concurrences between visual and textual data useful to improve the hate speech classification results on multimodal data. To do that we train different models where all or only some inputs are available. When an input is not available, we set it to zeros, and we do the same when an image has no text. Results Table TABREF31 shows the F-score, the Area Under the ROC Curve (AUC) and the mean accuracy (ACC) of the proposed models when different inputs are available. $TT$ refers to the tweet text, $IT$ to the image text and $I$ to the image. It also shows results for the LSTM, for the Davison method proposed in BIBREF7 trained with MMHS150K, and for random scores. Fig. FIGREF32 shows the Precision vs Recall plot and the ROC curve (which plots the True Positive Rate vs the False Positive Rate) of the different models. First, notice that given the subjectivity of the task and the discrepancies between annotators, getting optimal scores in the evaluation metrics is virtually impossible. However, a system with relatively low metric scores can still be very useful for hate speech detection in a real application: it will fire on publications for which most annotators agree they are hate, which are often the stronger attacks. The proposed LSTM to detect hate speech when only text is available, gets similar results as the method presented in BIBREF7, which we trained with MMHS150K and the same splits. However, more than substantially advancing the state of the art on hate speech detection in textual publications, our key purpose in this work is to introduce and work on its detection on multimodal publications. We use LSTM because it provides a strong representation of the tweet texts. The FCM trained only with images gets decent results, considering that in many publications the images might not give any useful information for the task. Fig. FIGREF33 shows some representative examples of the top hate and not hate scored images of this model. Many hate tweets are accompanied by demeaning nudity images, being sexist or homophobic. Other racist tweets are accompanied by images caricaturing black people. Finally, MEMES are also typically used in hate speech publications. The top scored images for not hate are portraits of people belonging to minorities. This is due to the use of slur inside these communities without an offensive intention, such as the word nigga inside the afro-american community or the word dyke inside the lesbian community. These results show that images can be effectively used to discriminate between offensive and non-offensive uses of those words. Despite the model trained only with images proves that they are useful for hate speech detection, the proposed multimodal models are not able to improve the detection compared to the textual models. Besides the different architectures, we have tried different training strategies, such as initializing the CNN weights with a model already trained solely with MMHS150K images or using dropout to force the multimodal models to use the visual information. Eventually, though, these models end up using almost only the text input for the prediction and producing very similar results to those of the textual models. The proposed multimodal models, such as TKM, have shown good performance in other tasks, such as VQA. Next, we analyze why they do not perform well in this task and with this data: [noitemsep,leftmargin=*] Noisy data. A major challenge of this task is the discrepancy between annotations due to subjective judgement. Although this affects also detection using only text, its repercussion is bigger in more complex tasks, such as detection using images or multimodal detection. Complexity and diversity of multimodal relations. Hate speech multimodal publications employ a lot of background knowledge which makes the relations between visual and textual elements they use very complex and diverse, and therefore difficult to learn by a neural network. Small set of multimodal examples. Fig. FIGREF5 shows some of the challenging multimodal hate examples that we aimed to detect. But although we have collected a big dataset of $150K$ tweets, the subset of multimodal hate there is still too small to learn the complex multimodal relations needed to identify multimodal hate. Conclusions In this work we have explored the task of hate speech detection on multimodal publications. We have created MMHS150K, to our knowledge the biggest available hate speech dataset, and the first one composed of multimodal data, namely tweets formed by image and text. We have trained different textual, visual and multimodal models with that data, and found out that, despite the fact that images are useful for hate speech detection, the multimodal models do not outperform the textual models. Finally, we have analyzed the challenges of the proposed task and dataset. Given that most of the content in Social Media nowadays is multimodal, we truly believe on the importance of pushing forward this research. The code used in this work is available in .

No attacks to any community, racist, sexist, homophobic, religion based attacks, attacks to other communities

62a6382157d5f9c1dce6e6c24ac5994442053002

62a6382157d5f9c1dce6e6c24ac5994442053002_0

Q: What were the evaluation metrics used? Text: Introduction Short text clustering is of great importance due to its various applications, such as user profiling BIBREF0 and recommendation BIBREF1 , for nowaday's social media dataset emerged day by day. However, short text clustering has the data sparsity problem and most words only occur once in each short text BIBREF2 . As a result, the Term Frequency-Inverse Document Frequency (TF-IDF) measure cannot work well in short text setting. In order to address this problem, some researchers work on expanding and enriching the context of data from Wikipedia BIBREF3 or an ontology BIBREF4 . However, these methods involve solid Natural Language Processing (NLP) knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another way to overcome these issues is to explore some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Yet how to design an effective model is an open question, and most of these methods directly trained based on Bag-of-Words (BoW) are shallow structures which cannot preserve the accurate semantic similarities. Recently, with the help of word embedding, neural networks demonstrate their great performance in terms of constructing text representation, such as Recursive Neural Network (RecNN) BIBREF6 , BIBREF7 and Recurrent Neural Network (RNN) BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the hidden layer computed at the last word to represent the text, is a biased model where later words are more dominant than earlier words BIBREF9 . Whereas for the non-biased models, the learned representation of one text can be extracted from all the words in the text with non-dominant learned weights. More recently, Convolution Neural Network (CNN), as the most popular non-biased model and applying convolutional filters to capture local features, has achieved a better performance in many NLP applications, such as sentence modeling BIBREF10 , relation classification BIBREF11 , and other traditional NLP tasks BIBREF12 . Most of the previous works focus CNN on solving supervised NLP tasks, while in this paper we aim to explore the power of CNN on one unsupervised NLP task, short text clustering. We systematically introduce a simple yet surprisingly powerful Self-Taught Convolutional neural network framework for Short Text Clustering, called STC INLINEFORM0 . An overall architecture of our proposed approach is illustrated in Figure FIGREF5 . We, inspired by BIBREF13 , BIBREF14 , utilize a self-taught learning framework into our task. In particular, the original raw text features are first embedded into compact binary codes INLINEFORM1 with the help of one traditional unsupervised dimensionality reduction function. Then text matrix INLINEFORM2 projected from word embeddings are fed into CNN model to learn the deep feature representation INLINEFORM3 and the output units are used to fit the pre-trained binary codes INLINEFORM4 . After obtaining the learned features, K-means algorithm is employed on them to cluster texts into clusters INLINEFORM5 . Obviously, we call our approach “self-taught” because the CNN model is learnt from the pseudo labels generated from the previous stage, which is quite different from the term “self-taught” in BIBREF15 . Our main contributions can be summarized as follows: This work is an extension of our conference paper BIBREF16 , and they differ in the following aspects. First, we put forward a general a self-taught CNN framework in this paper which can flexibly couple various semantic features, whereas the conference version can be seen as a specific example of this work. Second, in this paper we use a new short text dataset, Biomedical, in the experiment to verify the effectiveness of our approach. Third, we put much effort on studying the influence of various different semantic features integrated in our self-taught CNN framework, which is not involved in the conference paper. For the purpose of reproducibility, we make the datasets and software used in our experiments publicly available at the website. The remainder of this paper is organized as follows: In Section SECREF2 , we first briefly survey several related works. In Section SECREF3 , we describe the proposed approach STC INLINEFORM0 and implementation details. Experimental results and analyses are presented in Section SECREF4 . Finally, conclusions are given in the last Section. Related Work In this section, we review the related work from the following two perspectives: short text clustering and deep neural networks. Short Text Clustering There have been several studies that attempted to overcome the sparseness of short text representation. One way is to expand and enrich the context of data. For example, Banerjee et al. BIBREF3 proposed a method of improving the accuracy of short text clustering by enriching their representation with additional features from Wikipedia, and Fodeh et al. BIBREF4 incorporate semantic knowledge from an ontology into text clustering. However, these works need solid NLP knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another direction is to map the original features into reduced space, such as Latent Semantic Analysis (LSA) BIBREF17 , Laplacian Eigenmaps (LE) BIBREF18 , and Locality Preserving Indexing (LPI) BIBREF19 . Even some researchers explored some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Moreover, some studies even focus the above both two streams. For example, Tang et al. BIBREF20 proposed a novel framework which enrich the text features by employing machine translation and reduce the original features simultaneously through matrix factorization techniques. Despite the above clustering methods can alleviate sparseness of short text representation to some extent, most of them ignore word order in the text and belong to shallow structures which can not fully capture accurate semantic similarities. Deep Neural Networks Recently, there is a revival of interest in DNN and many researchers have concentrated on using Deep Learning to learn features. Hinton and Salakhutdinov BIBREF21 use DAE to learn text representation. During the fine-tuning procedure, they use backpropagation to find codes that are good at reconstructing the word-count vector. More recently, researchers propose to use external corpus to learn a distributed representation for each word, called word embedding BIBREF22 , to improve DNN performance on NLP tasks. The Skip-gram and continuous bag-of-words models of Word2vec BIBREF23 propose a simple single-layer architecture based on the inner product between two word vectors, and Pennington et al. BIBREF24 introduce a new model for word representation, called GloVe, which captures the global corpus statistics. In order to learn the compact representation vectors of sentences, Le and Mikolov BIBREF25 directly extend the previous Word2vec BIBREF23 by predicting words in the sentence, which is named Paragraph Vector (Para2vec). Para2vec is still a shallow window-based method and need a larger corpus to yield better performance. More neural networks utilize word embedding to capture true meaningful syntactic and semantic regularities, such as RecNN BIBREF6 , BIBREF7 and RNN BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the layer computed at the last word to represent the text, is a biased model. Recently, Long Short-Term Memory (LSTM) BIBREF26 and Gated Recurrent Unit (GRU) BIBREF27 , as sophisticated recurrent hidden units of RNN, has presented its advantages in many sequence generation problem, such as machine translation BIBREF28 , speech recognition BIBREF29 , and text conversation BIBREF30 . While, CNN is better to learn non-biased implicit features which has been successfully exploited for many supervised NLP learning tasks as described in Section SECREF1 , and various CNN based variants are proposed in the recent works, such as Dynamic Convolutional Neural Network (DCNN) BIBREF10 , Gated Recursive Convolutional Neural Network (grConv) BIBREF31 and Self-Adaptive Hierarchical Sentence model (AdaSent) BIBREF32 . In the past few days, Visin et al. BIBREF33 have attempted to replace convolutional layer in CNN to learn non-biased features for object recognition with four RNNs, called ReNet, that sweep over lower-layer features in different directions: (1) bottom to top, (2) top to bottom, (3) left to right and (4) right to left. However, ReNet does not outperform state-of-the-art convolutional neural networks on any of the three benchmark datasets, and it is also a supervised learning model for classification. Inspired by Skip-gram of word2vec BIBREF34 , BIBREF23 , Skip-thought model BIBREF35 describe an approach for unsupervised learning of a generic, distributed sentence encoder. Similar as Skip-gram model, Skip-thought model trains an encoder-decoder model that tries to reconstruct the surrounding sentences of an encoded sentence and released an off-the-shelf encoder to extract sentence representation. Even some researchers introduce continuous Skip-gram and negative sampling to CNN for learning visual representation in an unsupervised manner BIBREF36 . This paper, from a new perspective, puts forward a general self-taught CNN framework which can flexibly couple various semantic features and achieve a good performance on one unsupervised learning task, short text clustering. Methodology Assume that we are given a dataset of INLINEFORM0 training texts denoted as: INLINEFORM1 , where INLINEFORM2 is the dimensionality of the original BoW representation. Denote its tag set as INLINEFORM3 and the pre-trained word embedding set as INLINEFORM4 , where INLINEFORM5 is the dimensionality of word vectors and INLINEFORM6 is the vocabulary size. In order to learn the INLINEFORM7 -dimensional deep feature representation INLINEFORM8 from CNN in an unsupervised manner, some unsupervised dimensionality reduction methods INLINEFORM9 are employed to guide the learning of CNN model. Our goal is to cluster these texts INLINEFORM10 into clusters INLINEFORM11 based on the learned deep feature representation while preserving the semantic consistency. As depicted in Figure FIGREF5 , the proposed framework consist of three components, deep convolutional neural network (CNN), unsupervised dimensionality reduction function and K-means module. In the rest sections, we first present the first two components respectively, and then give the trainable parameters and the objective function to learn the deep feature representation. Finally, the last section describe how to perform clustering on the learned features. Deep Convolutional Neural Networks In this section, we briefly review one popular deep convolutional neural network, Dynamic Convolutional Neural Network (DCNN) BIBREF10 as an instance of CNN in the following sections, which as the foundation of our proposed method has been successfully proposed for the completely supervised learning task, text classification. Taking a neural network with two convolutional layers in Figure FIGREF9 as an example, the network transforms raw input text to a powerful representation. Particularly, each raw text vector INLINEFORM0 is projected into a matrix representation INLINEFORM1 by looking up a word embedding INLINEFORM2 , where INLINEFORM3 is the length of one text. We also let INLINEFORM4 and INLINEFORM5 denote the weights of the neural networks. The network defines a transformation INLINEFORM6 INLINEFORM7 which transforms an input raw text INLINEFORM8 to a INLINEFORM9 -dimensional deep representation INLINEFORM10 . There are three basic operations described as follows: Wide one-dimensional convolution This operation INLINEFORM0 is applied to an individual row of the sentence matrix INLINEFORM1 , and yields a resulting matrix INLINEFORM2 , where INLINEFORM3 is the width of convolutional filter. Folding In this operation, every two rows in a feature map are simply summed component-wisely. For a map of INLINEFORM0 rows, folding returns a map of INLINEFORM1 rows, thus halving the size of the representation and yielding a matrix feature INLINEFORM2 . Note that folding operation does not introduce any additional parameters. Dynamic INLINEFORM0 -max pooling Assuming the pooling parameter as INLINEFORM1 , INLINEFORM2 -max pooling selects the sub-matrix INLINEFORM3 of the INLINEFORM4 highest values in each row of the matrix INLINEFORM5 . For dynamic INLINEFORM6 -max pooling, the pooling parameter INLINEFORM7 is dynamically selected in order to allow for a smooth extraction of higher-order and longer-range features BIBREF10 . Given a fixed pooling parameter INLINEFORM8 for the topmost convolutional layer, the parameter INLINEFORM9 of INLINEFORM10 -max pooling in the INLINEFORM11 -th convolutional layer can be computed as follows: DISPLAYFORM0 where INLINEFORM0 is the total number of convolutional layers in the network. Unsupervised Dimensionality Reduction As described in Figure FIGREF5 , the dimensionality reduction function is defined as follows: DISPLAYFORM0 where, INLINEFORM0 are the INLINEFORM1 -dimensional reduced latent space representations. Here, we take four popular dimensionality reduction methods as examples in our framework. Average Embedding (AE): This method directly averages the word embeddings which are respectively weighted with TF and TF-IDF. Huang et al. BIBREF37 used this strategy as the global context in their task, and Socher et al. BIBREF7 and Lai et al. BIBREF9 used this method for text classification. The weighted average of all word vectors in one text can be computed as follows: DISPLAYFORM0 where INLINEFORM0 can be any weighting function that captures the importance of word INLINEFORM1 in the text INLINEFORM2 . Latent Semantic Analysis (LSA): LSA BIBREF17 is the most popular global matrix factorization method, which applies a dimension reducing linear projection, Singular Value Decomposition (SVD), of the corresponding term/document matrix. Suppose the rank of INLINEFORM0 is INLINEFORM1 , LSA decompose INLINEFORM2 into the product of three other matrices: DISPLAYFORM0 where INLINEFORM0 and INLINEFORM1 are the singular values of INLINEFORM2 , INLINEFORM3 is a set of left singular vectors and INLINEFORM4 is a set of right singular vectors. LSA uses the top INLINEFORM5 vectors in INLINEFORM6 as the transformation matrix to embed the original text features into a INLINEFORM7 -dimensional subspace INLINEFORM8 BIBREF17 . Laplacian Eigenmaps (LE): The top eigenvectors of graph Laplacian, defined on the similarity matrix of texts, are used in the method, which can discover the manifold structure of the text space BIBREF18 . In order to avoid storing the dense similarity matrix, many approximation techniques are proposed to reduce the memory usage and computational complexity for LE. There are two representative approximation methods, sparse similarity matrix and Nystr INLINEFORM0 m approximation. Following previous studies BIBREF38 , BIBREF13 , we select the former technique to construct the INLINEFORM1 local similarity matrix INLINEFORM2 by using heat kernel as follows: DISPLAYFORM0 where, INLINEFORM0 is a tuning parameter (default is 1) and INLINEFORM1 represents the set of INLINEFORM2 -nearest-neighbors of INLINEFORM3 . By introducing a diagonal INLINEFORM4 matrix INLINEFORM5 whose entries are given by INLINEFORM6 , the graph Laplacian INLINEFORM7 can be computed by ( INLINEFORM8 ). The optimal INLINEFORM9 real-valued matrix INLINEFORM10 can be obtained by solving the following objective function: DISPLAYFORM0 where INLINEFORM0 is the trace function, INLINEFORM1 requires the different dimensions to be uncorrelated, and INLINEFORM2 requires each dimension to achieve equal probability as positive or negative). Locality Preserving Indexing (LPI): This method extends LE to deal with unseen texts by approximating the linear function INLINEFORM0 BIBREF13 , and the subspace vectors are obtained by finding the optimal linear approximations to the eigenfunctions of the Laplace Beltrami operator on the Riemannian manifold BIBREF19 . Similar as LE, we first construct the local similarity matrix INLINEFORM1 , then the graph Laplacian INLINEFORM2 can be computed by ( INLINEFORM3 ), where INLINEFORM4 measures the local density around INLINEFORM5 and is equal to INLINEFORM6 . Compute the eigenvectors INLINEFORM7 and eigenvalues INLINEFORM8 of the following generalized eigen-problem: DISPLAYFORM0 The mapping function INLINEFORM0 can be obtained and applied to the unseen data BIBREF38 . All of the above methods claim a better performance in capturing semantic similarity between texts in the reduced latent space representation INLINEFORM0 than in the original representation INLINEFORM1 , while the performance of short text clustering can be further enhanced with the help of our framework, self-taught CNN. Learning The last layer of CNN is an output layer as follows: DISPLAYFORM0 where, INLINEFORM0 is the deep feature representation, INLINEFORM1 is the output vector and INLINEFORM2 is weight matrix. In order to incorporate the latent semantic features INLINEFORM0 , we first binary the real-valued vectors INLINEFORM1 to the binary codes INLINEFORM2 by setting the threshold to be the media vector INLINEFORM3 . Then, the output vector INLINEFORM4 is used to fit the binary codes INLINEFORM5 via INLINEFORM6 logistic operations as follows: DISPLAYFORM0 All parameters to be trained are defined as INLINEFORM0 . DISPLAYFORM0 Given the training text collection INLINEFORM0 , and the pre-trained binary codes INLINEFORM1 , the log likelihood of the parameters can be written down as follows: DISPLAYFORM0 Following the previous work BIBREF10 , we train the network with mini-batches by back-propagation and perform the gradient-based optimization using the Adagrad update rule BIBREF39 . For regularization, we employ dropout with 50% rate to the penultimate layer BIBREF10 , BIBREF40 . K-means for Clustering With the given short texts, we first utilize the trained deep neural network to obtain the semantic representations INLINEFORM0 , and then employ traditional K-means algorithm to perform clustering. Datasets We test our proposed approach on three public short text datasets. The summary statistics and semantic topics of these datasets are described in Table TABREF24 and Table TABREF25 . SearchSnippets. This dataset was selected from the results of web search transaction using predefined phrases of 8 different domains by Phan et al. BIBREF41 . StackOverflow. We use the challenge data published in Kaggle.com. The raw dataset consists 3,370,528 samples through July 31st, 2012 to August 14, 2012. In our experiments, we randomly select 20,000 question titles from 20 different tags as in Table TABREF25 . Biomedical. We use the challenge data published in BioASQ's official website. In our experiments, we randomly select 20, 000 paper titles from 20 different MeSH major topics as in Table TABREF25 . As described in Table TABREF24 , the max length of selected paper titles is 53. For these datasets, we randomly select 10% of data as the development set. Since SearchSnippets has been pre-processed by Phan et al. BIBREF41 , we do not further process this dataset. In StackOverflow, texts contain lots of computer terminology, and symbols and capital letters are meaningful, thus we do not do any pre-processed procedures. For Biomedical, we remove the symbols and convert letters into lower case. Pre-trained Word Vectors We use the publicly available word2vec tool to train word embeddings, and the most parameters are set as same as Mikolov et al. BIBREF23 to train word vectors on Google News setting, except of vector dimensionality using 48 and minimize count using 5. For SearchSnippets, we train word vectors on Wikipedia dumps. For StackOverflow, we train word vectors on the whole corpus of the StackOverflow dataset described above which includes the question titles and post contents. For Biomedical, we train word vectors on all titles and abstracts of 2014 training articles. The coverage of these learned vectors on three datasets are listed in Table TABREF32 , and the words not present in the set of pre-trained words are initialized randomly. Comparisons In our experiment, some widely used text clustering methods are compared with our approach. Besides K-means, Skip-thought Vectors, Recursive Neural Network and Paragraph Vector based clustering methods, four baseline clustering methods are directly based on the popular unsupervised dimensionality reduction methods as described in Section SECREF11 . We further compare our approach with some other non-biased neural networks, such as bidirectional RNN. More details are listed as follows: K-means K-means BIBREF42 on original keyword features which are respectively weighted with term frequency (TF) and term frequency-inverse document frequency (TF-IDF). Skip-thought Vectors (SkipVec) This baseline BIBREF35 gives an off-the-shelf encoder to produce highly generic sentence representations. The encoder is trained using a large collection of novels and provides three encoder modes, that are unidirectional encoder (SkipVec (Uni)) with 2,400 dimensions, bidirectional encoder (SkipVec (Bi)) with 2,400 dimensions and combined encoder (SkipVec (Combine)) with SkipVec (Uni) and SkipVec (Bi) of 2,400 dimensions each. K-means is employed on the these vector representations respectively. Recursive Neural Network (RecNN) In BIBREF6 , the tree structure is firstly greedy approximated via unsupervised recursive autoencoder. Then, semi-supervised recursive autoencoders are used to capture the semantics of texts based on the predicted structure. In order to make this recursive-based method completely unsupervised, we remove the cross-entropy error in the second phrase to learn vector representation and subsequently employ K-means on the learned vectors of the top tree node and the average of all vectors in the tree. Paragraph Vector (Para2vec) K-means on the fixed size feature vectors generated by Paragraph Vector (Para2vec) BIBREF25 which is an unsupervised method to learn distributed representation of words and paragraphs. In our experiments, we use the open source software released by Mesnil et al. BIBREF43 . Average Embedding (AE) K-means on the weighted average vectors of the word embeddings which are respectively weighted with TF and TF-IDF. The dimension of average vectors is equal to and decided by the dimension of word vectors used in our experiments. Latent Semantic Analysis (LSA) K-means on the reduced subspace vectors generated by Singular Value Decomposition (SVD) method. The dimension of subspace is default set to the number of clusters, we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 10 on SearchSnippets, 20 on StackOverflow and 20 on Biomedical in our experiments. Laplacian Eigenmaps (LE) This baseline, using Laplacian Eigenmaps and subsequently employing K-means algorithm, is well known as spectral clustering BIBREF44 . The dimension of subspace is default set to the number of clusters BIBREF18 , BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 70 on StackOverflow and 30 on Biomedical in our experiments. Locality Preserving Indexing (LPI) This baseline, projecting the texts into a lower dimensional semantic space, can discover both the geometric and discriminating structures of the original feature space BIBREF38 . The dimension of subspace is default set to the number of clusters BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 80 on StackOverflow and 30 on Biomedical in our experiments. bidirectional RNN (bi-RNN) We replace the CNN model in our framework as in Figure FIGREF5 with some bi-RNN models. Particularly, LSTM and GRU units are used in the experiments. In order to generate the fixed-length document representation from the variable-length vector sequences, for both bi-LSTM and bi-GRU based clustering methods, we further utilize three pooling methods: last pooling (using the last hidden state), mean pooling and element-wise max pooling. These pooling methods are respectively used in the previous works BIBREF45 , BIBREF27 , BIBREF46 and BIBREF9 . For regularization, the training gradients of all parameters with an INLINEFORM0 2 norm larger than 40 are clipped to 40, as the previous work BIBREF47 . Evaluation Metrics The clustering performance is evaluated by comparing the clustering results of texts with the tags/labels provided by the text corpus. Two metrics, the accuracy (ACC) and the normalized mutual information metric (NMI), are used to measure the clustering performance BIBREF38 , BIBREF48 . Given a text INLINEFORM0 , let INLINEFORM1 and INLINEFORM2 be the obtained cluster label and the label provided by the corpus, respectively. Accuracy is defined as: DISPLAYFORM0 where, INLINEFORM0 is the total number of texts, INLINEFORM1 is the indicator function that equals one if INLINEFORM2 and equals zero otherwise, and INLINEFORM3 is the permutation mapping function that maps each cluster label INLINEFORM4 to the equivalent label from the text data by Hungarian algorithm BIBREF49 . Normalized mutual information BIBREF50 between tag/label set INLINEFORM0 and cluster set INLINEFORM1 is a popular metric used for evaluating clustering tasks. It is defined as follows: DISPLAYFORM0 where, INLINEFORM0 is the mutual information between INLINEFORM1 and INLINEFORM2 , INLINEFORM3 is entropy and the denominator INLINEFORM4 is used for normalizing the mutual information to be in the range of [0, 1]. Hyperparameter Settings The most of parameters are set uniformly for these datasets. Following previous study BIBREF38 , the number of nearest neighbors in Eqn. ( EQREF15 ) is fixed to 15 when constructing the graph structures for LE and LPI. For CNN model, the networks has two convolutional layers. The widths of the convolutional filters are both 3. The value of INLINEFORM0 for the top INLINEFORM1 -max pooling in Eqn. ( EQREF10 ) is 5. The number of feature maps at the first convolutional layer is 12, and 8 feature maps at the second convolutional layer. Both those two convolutional layers are followed by a folding layer. We further set the dimension of word embeddings INLINEFORM2 as 48. Finally, the dimension of the deep feature representation INLINEFORM3 is fixed to 480. Moreover, we set the learning rate INLINEFORM4 as 0.01 and the mini-batch training size as 200. The output size INLINEFORM5 in Eqn. ( EQREF19 ) is set same as the best dimensions of subspace in the baseline method, as described in Section SECREF37 . For initial centroids have significant impact on clustering results when utilizing the K-means algorithms, we repeat K-means for multiple times with random initial centroids (specifically, 100 times for statistical significance) as Huang BIBREF48 . The all subspace vectors are normalized to 1 before applying K-means and the final results reported are the average of 5 trials with all clustering methods on three text datasets. Results and Analysis In Table TABREF43 and Table TABREF44 , we report the ACC and NMI performance of our proposed approaches and four baseline methods, K-means, SkipVec, RecNN and Para2vec based clustering methods. Intuitively, we get a general observation that (1) BoW based approaches, including K-means (TF) and K-means (TF-IDF), and SkipVec based approaches perform not well; (2) RecNN based approaches, both RecNN (Ave.) and RecNN (Top+Ave.), do better; (3) Para2vec makes a comparable performance with the most baselines; and (4) the evaluation clearly demonstrate the superiority of our proposed methods STC INLINEFORM0 . It is an expected results. For SkipVec based approaches, the off-the-shelf encoders are trained on the BookCorpus datasets BIBREF51 , and then applied to our datasets to extract the sentence representations. The SkipVec encoders can produce generic sentence representations but may not perform well for specific datasets, in our experiments, StackOverflow and Biomedical datasets consist of many computer terms and medical terms, such as “ASP.NET”, “XML”, “C#”, “serum” and “glycolytic”. When we take a more careful look, we find that RecNN (Top) does poorly, even worse than K-means (TF-IDF). The reason maybe that although recursive neural models introduce tree structure to capture compositional semantics, the vector of the top node mainly captures a biased semantic while the average of all vectors in the tree nodes, such as RecNN (Ave.), can be better to represent sentence level semantic. And we also get another observation that, although our proposed STC INLINEFORM1 -LE and STC INLINEFORM2 -LPI outperform both BoW based and RecNN based approaches across all three datasets, STC INLINEFORM3 -AE and STC INLINEFORM4 -LSA do just exhibit some similar performances as RecNN (Ave.) and RecNN (Top+Ave.) do in the datasets of StackOverflow and Biomedical. We further replace the CNN model in our framework as in Figure FIGREF5 with some other non-biased models, such as bi-LSTM and bi-GRU, and report the results in Table TABREF46 and Table TABREF47 . As an instance, the binary codes generated from LPI are used to guide the learning of bi-LSTM/bi-GRU models. From the results, we can see that bi-GRU and bi-LSTM based clustering methods do equally well, no clear winner, and both achieve great enhancements compared with LPI (best). Compared with these bi-LSTM/bi-GRU based models, the evaluation results still demonstrate the superiority of our approach methods, CNN based clustering model, in the most cases. As the results reported by Visin et al. BIBREF33 , despite bi-directional or multi-directional RNN models perform a good non-biased feature extraction, they yet do not outperform state-of-the-art CNN on some tasks. In order to make clear what factors make our proposed method work, we report the bar chart results of ACC and MNI of our proposed methods and the corresponding baseline methods in Figure FIGREF49 and Figure FIGREF53 . It is clear that, although AE and LSA does well or even better than LE and LPI, especially in dataset of both StackOverflow and Biomedical, STC INLINEFORM0 -LE and STC INLINEFORM1 -LPI achieve a much larger performance enhancements than STC INLINEFORM2 -AE and STC INLINEFORM3 -LSA do. The possible reason is that the information the pseudo supervision used to guide the learning of CNN model that make difference. Especially, for AE case, the input features fed into CNN model and the pseudo supervision employed to guide the learning of CNN model are all come from word embeddings. There are no different semantic features to be used into our proposed method, thus the performance enhancements are limited in STC INLINEFORM4 -AE. For LSA case, as we known, LSA is to make matrix factorization to find the best subspace approximation of the original feature space to minimize the global reconstruction error. And as BIBREF24 , BIBREF52 recently point out that word embeddings trained with word2vec or some variances, is essentially to do an operation of matrix factorization. Therefore, the information between input and the pseudo supervision in CNN is not departed very largely from each other, and the performance enhancements of STC INLINEFORM5 -AE is also not quite satisfactory. For LE and LPI case, as we known that LE extracts the manifold structure of the original feature space, and LPI extracts both geometric and discriminating structure of the original feature space BIBREF38 . We guess that our approach STC INLINEFORM6 -LE and STC INLINEFORM7 -LPI achieve enhancements compared with both LE and LPI by a large margin, because both of LE and LPI get useful semantic features, and these features are also different from word embeddings used as input of CNN. From this view, we say that our proposed STC has potential to behave more effective when the pseudo supervision is able to get semantic meaningful features, which is different enough from the input of CNN. Furthermore, from the results of K-means and AE in Table TABREF43 - TABREF44 and Figure FIGREF49 - FIGREF53 , we note that TF-IDF weighting gives a more remarkable improvement for K-means, while TF weighting works better than TF-IDF weighting for Average Embedding. Maybe the reason is that pre-trained word embeddings encode some useful information from external corpus and are able to get even better results without TF-IDF weighting. Meanwhile, we find that LE get quite unusual good performance than LPI, LSA and AE in SearchSnippets dataset, which is not found in the other two datasets. To get clear about this, and also to make a much better demonstration about our proposed approaches and other baselines, we further report 2-dimensional text embeddings on SearchSnippets in Figure FIGREF58 , using t-SNE BIBREF53 to get distributed stochastic neighbor embedding of the feature representations used in the clustering methods. We can see that the results of from AE and LSA seem to be fairly good or even better than the ones from LE and LPI, which is not the same as the results from ACC and NMI in Figure FIGREF49 - FIGREF53 . Meanwhile, RecNN (Ave.) performs better than BoW (both TF and TF-IDF) while RecNN (Top) does not, which is the same as the results from ACC and NMI in Table TABREF43 and Table TABREF44 . Then we guess that both ”the same as” and ”not the same as” above, is just a good example to illustrate that visualization tool, such as t-SNE, get some useful information for measuring results, which is different from the ones of ACC and NMI. Moreover, from this complementary view of t-SNE, we can see that our STC INLINEFORM0 -AE, STC INLINEFORM1 -LSA, STC INLINEFORM2 -LE, and STC INLINEFORM3 -LPI show more clear-cut margins among different semantic topics (that is, tags/labels), compared with AE, LSA, LE and LPI, respectively, as well as compared with both baselines, BoW and RecNN based ones. From all these results, with three measures of ACC, NMI and t-SNE under three datasets, we can get a solid conclusion that our proposed approaches is an effective approaches to get useful semantic features for short text clustering. Conclusions With the emergence of social media, short text clustering has become an increasing important task. This paper explores a new perspective to cluster short texts based on deep feature representation learned from the proposed self-taught convolutional neural networks. Our framework can be successfully accomplished without using any external tags/labels and complicated NLP pre-processing, and and our approach is a flexible framework, in which the traditional dimension reduction approaches could be used to get performance enhancement. Our extensive experimental study on three short text datasets shows that our approach can achieve a significantly better performance. In the future, how to select and incorporate more effective semantic features into the proposed framework would call for more research. Acknowledgments We would like to thank reviewers for their comments, and acknowledge Kaggle and BioASQ for making the datasets available. This work is supported by the National Natural Science Foundation of China (No. 61602479, No. 61303172, No. 61403385) and the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB02070005).

accuracy, normalized mutual information

9e04730907ad728d62049f49ac828acb4e0a1a2a

9e04730907ad728d62049f49ac828acb4e0a1a2a_0

Q: What were their performance results? Text: Introduction Short text clustering is of great importance due to its various applications, such as user profiling BIBREF0 and recommendation BIBREF1 , for nowaday's social media dataset emerged day by day. However, short text clustering has the data sparsity problem and most words only occur once in each short text BIBREF2 . As a result, the Term Frequency-Inverse Document Frequency (TF-IDF) measure cannot work well in short text setting. In order to address this problem, some researchers work on expanding and enriching the context of data from Wikipedia BIBREF3 or an ontology BIBREF4 . However, these methods involve solid Natural Language Processing (NLP) knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another way to overcome these issues is to explore some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Yet how to design an effective model is an open question, and most of these methods directly trained based on Bag-of-Words (BoW) are shallow structures which cannot preserve the accurate semantic similarities. Recently, with the help of word embedding, neural networks demonstrate their great performance in terms of constructing text representation, such as Recursive Neural Network (RecNN) BIBREF6 , BIBREF7 and Recurrent Neural Network (RNN) BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the hidden layer computed at the last word to represent the text, is a biased model where later words are more dominant than earlier words BIBREF9 . Whereas for the non-biased models, the learned representation of one text can be extracted from all the words in the text with non-dominant learned weights. More recently, Convolution Neural Network (CNN), as the most popular non-biased model and applying convolutional filters to capture local features, has achieved a better performance in many NLP applications, such as sentence modeling BIBREF10 , relation classification BIBREF11 , and other traditional NLP tasks BIBREF12 . Most of the previous works focus CNN on solving supervised NLP tasks, while in this paper we aim to explore the power of CNN on one unsupervised NLP task, short text clustering. We systematically introduce a simple yet surprisingly powerful Self-Taught Convolutional neural network framework for Short Text Clustering, called STC INLINEFORM0 . An overall architecture of our proposed approach is illustrated in Figure FIGREF5 . We, inspired by BIBREF13 , BIBREF14 , utilize a self-taught learning framework into our task. In particular, the original raw text features are first embedded into compact binary codes INLINEFORM1 with the help of one traditional unsupervised dimensionality reduction function. Then text matrix INLINEFORM2 projected from word embeddings are fed into CNN model to learn the deep feature representation INLINEFORM3 and the output units are used to fit the pre-trained binary codes INLINEFORM4 . After obtaining the learned features, K-means algorithm is employed on them to cluster texts into clusters INLINEFORM5 . Obviously, we call our approach “self-taught” because the CNN model is learnt from the pseudo labels generated from the previous stage, which is quite different from the term “self-taught” in BIBREF15 . Our main contributions can be summarized as follows: This work is an extension of our conference paper BIBREF16 , and they differ in the following aspects. First, we put forward a general a self-taught CNN framework in this paper which can flexibly couple various semantic features, whereas the conference version can be seen as a specific example of this work. Second, in this paper we use a new short text dataset, Biomedical, in the experiment to verify the effectiveness of our approach. Third, we put much effort on studying the influence of various different semantic features integrated in our self-taught CNN framework, which is not involved in the conference paper. For the purpose of reproducibility, we make the datasets and software used in our experiments publicly available at the website. The remainder of this paper is organized as follows: In Section SECREF2 , we first briefly survey several related works. In Section SECREF3 , we describe the proposed approach STC INLINEFORM0 and implementation details. Experimental results and analyses are presented in Section SECREF4 . Finally, conclusions are given in the last Section. Related Work In this section, we review the related work from the following two perspectives: short text clustering and deep neural networks. Short Text Clustering There have been several studies that attempted to overcome the sparseness of short text representation. One way is to expand and enrich the context of data. For example, Banerjee et al. BIBREF3 proposed a method of improving the accuracy of short text clustering by enriching their representation with additional features from Wikipedia, and Fodeh et al. BIBREF4 incorporate semantic knowledge from an ontology into text clustering. However, these works need solid NLP knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another direction is to map the original features into reduced space, such as Latent Semantic Analysis (LSA) BIBREF17 , Laplacian Eigenmaps (LE) BIBREF18 , and Locality Preserving Indexing (LPI) BIBREF19 . Even some researchers explored some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Moreover, some studies even focus the above both two streams. For example, Tang et al. BIBREF20 proposed a novel framework which enrich the text features by employing machine translation and reduce the original features simultaneously through matrix factorization techniques. Despite the above clustering methods can alleviate sparseness of short text representation to some extent, most of them ignore word order in the text and belong to shallow structures which can not fully capture accurate semantic similarities. Deep Neural Networks Recently, there is a revival of interest in DNN and many researchers have concentrated on using Deep Learning to learn features. Hinton and Salakhutdinov BIBREF21 use DAE to learn text representation. During the fine-tuning procedure, they use backpropagation to find codes that are good at reconstructing the word-count vector. More recently, researchers propose to use external corpus to learn a distributed representation for each word, called word embedding BIBREF22 , to improve DNN performance on NLP tasks. The Skip-gram and continuous bag-of-words models of Word2vec BIBREF23 propose a simple single-layer architecture based on the inner product between two word vectors, and Pennington et al. BIBREF24 introduce a new model for word representation, called GloVe, which captures the global corpus statistics. In order to learn the compact representation vectors of sentences, Le and Mikolov BIBREF25 directly extend the previous Word2vec BIBREF23 by predicting words in the sentence, which is named Paragraph Vector (Para2vec). Para2vec is still a shallow window-based method and need a larger corpus to yield better performance. More neural networks utilize word embedding to capture true meaningful syntactic and semantic regularities, such as RecNN BIBREF6 , BIBREF7 and RNN BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the layer computed at the last word to represent the text, is a biased model. Recently, Long Short-Term Memory (LSTM) BIBREF26 and Gated Recurrent Unit (GRU) BIBREF27 , as sophisticated recurrent hidden units of RNN, has presented its advantages in many sequence generation problem, such as machine translation BIBREF28 , speech recognition BIBREF29 , and text conversation BIBREF30 . While, CNN is better to learn non-biased implicit features which has been successfully exploited for many supervised NLP learning tasks as described in Section SECREF1 , and various CNN based variants are proposed in the recent works, such as Dynamic Convolutional Neural Network (DCNN) BIBREF10 , Gated Recursive Convolutional Neural Network (grConv) BIBREF31 and Self-Adaptive Hierarchical Sentence model (AdaSent) BIBREF32 . In the past few days, Visin et al. BIBREF33 have attempted to replace convolutional layer in CNN to learn non-biased features for object recognition with four RNNs, called ReNet, that sweep over lower-layer features in different directions: (1) bottom to top, (2) top to bottom, (3) left to right and (4) right to left. However, ReNet does not outperform state-of-the-art convolutional neural networks on any of the three benchmark datasets, and it is also a supervised learning model for classification. Inspired by Skip-gram of word2vec BIBREF34 , BIBREF23 , Skip-thought model BIBREF35 describe an approach for unsupervised learning of a generic, distributed sentence encoder. Similar as Skip-gram model, Skip-thought model trains an encoder-decoder model that tries to reconstruct the surrounding sentences of an encoded sentence and released an off-the-shelf encoder to extract sentence representation. Even some researchers introduce continuous Skip-gram and negative sampling to CNN for learning visual representation in an unsupervised manner BIBREF36 . This paper, from a new perspective, puts forward a general self-taught CNN framework which can flexibly couple various semantic features and achieve a good performance on one unsupervised learning task, short text clustering. Methodology Assume that we are given a dataset of INLINEFORM0 training texts denoted as: INLINEFORM1 , where INLINEFORM2 is the dimensionality of the original BoW representation. Denote its tag set as INLINEFORM3 and the pre-trained word embedding set as INLINEFORM4 , where INLINEFORM5 is the dimensionality of word vectors and INLINEFORM6 is the vocabulary size. In order to learn the INLINEFORM7 -dimensional deep feature representation INLINEFORM8 from CNN in an unsupervised manner, some unsupervised dimensionality reduction methods INLINEFORM9 are employed to guide the learning of CNN model. Our goal is to cluster these texts INLINEFORM10 into clusters INLINEFORM11 based on the learned deep feature representation while preserving the semantic consistency. As depicted in Figure FIGREF5 , the proposed framework consist of three components, deep convolutional neural network (CNN), unsupervised dimensionality reduction function and K-means module. In the rest sections, we first present the first two components respectively, and then give the trainable parameters and the objective function to learn the deep feature representation. Finally, the last section describe how to perform clustering on the learned features. Deep Convolutional Neural Networks In this section, we briefly review one popular deep convolutional neural network, Dynamic Convolutional Neural Network (DCNN) BIBREF10 as an instance of CNN in the following sections, which as the foundation of our proposed method has been successfully proposed for the completely supervised learning task, text classification. Taking a neural network with two convolutional layers in Figure FIGREF9 as an example, the network transforms raw input text to a powerful representation. Particularly, each raw text vector INLINEFORM0 is projected into a matrix representation INLINEFORM1 by looking up a word embedding INLINEFORM2 , where INLINEFORM3 is the length of one text. We also let INLINEFORM4 and INLINEFORM5 denote the weights of the neural networks. The network defines a transformation INLINEFORM6 INLINEFORM7 which transforms an input raw text INLINEFORM8 to a INLINEFORM9 -dimensional deep representation INLINEFORM10 . There are three basic operations described as follows: Wide one-dimensional convolution This operation INLINEFORM0 is applied to an individual row of the sentence matrix INLINEFORM1 , and yields a resulting matrix INLINEFORM2 , where INLINEFORM3 is the width of convolutional filter. Folding In this operation, every two rows in a feature map are simply summed component-wisely. For a map of INLINEFORM0 rows, folding returns a map of INLINEFORM1 rows, thus halving the size of the representation and yielding a matrix feature INLINEFORM2 . Note that folding operation does not introduce any additional parameters. Dynamic INLINEFORM0 -max pooling Assuming the pooling parameter as INLINEFORM1 , INLINEFORM2 -max pooling selects the sub-matrix INLINEFORM3 of the INLINEFORM4 highest values in each row of the matrix INLINEFORM5 . For dynamic INLINEFORM6 -max pooling, the pooling parameter INLINEFORM7 is dynamically selected in order to allow for a smooth extraction of higher-order and longer-range features BIBREF10 . Given a fixed pooling parameter INLINEFORM8 for the topmost convolutional layer, the parameter INLINEFORM9 of INLINEFORM10 -max pooling in the INLINEFORM11 -th convolutional layer can be computed as follows: DISPLAYFORM0 where INLINEFORM0 is the total number of convolutional layers in the network. Unsupervised Dimensionality Reduction As described in Figure FIGREF5 , the dimensionality reduction function is defined as follows: DISPLAYFORM0 where, INLINEFORM0 are the INLINEFORM1 -dimensional reduced latent space representations. Here, we take four popular dimensionality reduction methods as examples in our framework. Average Embedding (AE): This method directly averages the word embeddings which are respectively weighted with TF and TF-IDF. Huang et al. BIBREF37 used this strategy as the global context in their task, and Socher et al. BIBREF7 and Lai et al. BIBREF9 used this method for text classification. The weighted average of all word vectors in one text can be computed as follows: DISPLAYFORM0 where INLINEFORM0 can be any weighting function that captures the importance of word INLINEFORM1 in the text INLINEFORM2 . Latent Semantic Analysis (LSA): LSA BIBREF17 is the most popular global matrix factorization method, which applies a dimension reducing linear projection, Singular Value Decomposition (SVD), of the corresponding term/document matrix. Suppose the rank of INLINEFORM0 is INLINEFORM1 , LSA decompose INLINEFORM2 into the product of three other matrices: DISPLAYFORM0 where INLINEFORM0 and INLINEFORM1 are the singular values of INLINEFORM2 , INLINEFORM3 is a set of left singular vectors and INLINEFORM4 is a set of right singular vectors. LSA uses the top INLINEFORM5 vectors in INLINEFORM6 as the transformation matrix to embed the original text features into a INLINEFORM7 -dimensional subspace INLINEFORM8 BIBREF17 . Laplacian Eigenmaps (LE): The top eigenvectors of graph Laplacian, defined on the similarity matrix of texts, are used in the method, which can discover the manifold structure of the text space BIBREF18 . In order to avoid storing the dense similarity matrix, many approximation techniques are proposed to reduce the memory usage and computational complexity for LE. There are two representative approximation methods, sparse similarity matrix and Nystr INLINEFORM0 m approximation. Following previous studies BIBREF38 , BIBREF13 , we select the former technique to construct the INLINEFORM1 local similarity matrix INLINEFORM2 by using heat kernel as follows: DISPLAYFORM0 where, INLINEFORM0 is a tuning parameter (default is 1) and INLINEFORM1 represents the set of INLINEFORM2 -nearest-neighbors of INLINEFORM3 . By introducing a diagonal INLINEFORM4 matrix INLINEFORM5 whose entries are given by INLINEFORM6 , the graph Laplacian INLINEFORM7 can be computed by ( INLINEFORM8 ). The optimal INLINEFORM9 real-valued matrix INLINEFORM10 can be obtained by solving the following objective function: DISPLAYFORM0 where INLINEFORM0 is the trace function, INLINEFORM1 requires the different dimensions to be uncorrelated, and INLINEFORM2 requires each dimension to achieve equal probability as positive or negative). Locality Preserving Indexing (LPI): This method extends LE to deal with unseen texts by approximating the linear function INLINEFORM0 BIBREF13 , and the subspace vectors are obtained by finding the optimal linear approximations to the eigenfunctions of the Laplace Beltrami operator on the Riemannian manifold BIBREF19 . Similar as LE, we first construct the local similarity matrix INLINEFORM1 , then the graph Laplacian INLINEFORM2 can be computed by ( INLINEFORM3 ), where INLINEFORM4 measures the local density around INLINEFORM5 and is equal to INLINEFORM6 . Compute the eigenvectors INLINEFORM7 and eigenvalues INLINEFORM8 of the following generalized eigen-problem: DISPLAYFORM0 The mapping function INLINEFORM0 can be obtained and applied to the unseen data BIBREF38 . All of the above methods claim a better performance in capturing semantic similarity between texts in the reduced latent space representation INLINEFORM0 than in the original representation INLINEFORM1 , while the performance of short text clustering can be further enhanced with the help of our framework, self-taught CNN. Learning The last layer of CNN is an output layer as follows: DISPLAYFORM0 where, INLINEFORM0 is the deep feature representation, INLINEFORM1 is the output vector and INLINEFORM2 is weight matrix. In order to incorporate the latent semantic features INLINEFORM0 , we first binary the real-valued vectors INLINEFORM1 to the binary codes INLINEFORM2 by setting the threshold to be the media vector INLINEFORM3 . Then, the output vector INLINEFORM4 is used to fit the binary codes INLINEFORM5 via INLINEFORM6 logistic operations as follows: DISPLAYFORM0 All parameters to be trained are defined as INLINEFORM0 . DISPLAYFORM0 Given the training text collection INLINEFORM0 , and the pre-trained binary codes INLINEFORM1 , the log likelihood of the parameters can be written down as follows: DISPLAYFORM0 Following the previous work BIBREF10 , we train the network with mini-batches by back-propagation and perform the gradient-based optimization using the Adagrad update rule BIBREF39 . For regularization, we employ dropout with 50% rate to the penultimate layer BIBREF10 , BIBREF40 . K-means for Clustering With the given short texts, we first utilize the trained deep neural network to obtain the semantic representations INLINEFORM0 , and then employ traditional K-means algorithm to perform clustering. Datasets We test our proposed approach on three public short text datasets. The summary statistics and semantic topics of these datasets are described in Table TABREF24 and Table TABREF25 . SearchSnippets. This dataset was selected from the results of web search transaction using predefined phrases of 8 different domains by Phan et al. BIBREF41 . StackOverflow. We use the challenge data published in Kaggle.com. The raw dataset consists 3,370,528 samples through July 31st, 2012 to August 14, 2012. In our experiments, we randomly select 20,000 question titles from 20 different tags as in Table TABREF25 . Biomedical. We use the challenge data published in BioASQ's official website. In our experiments, we randomly select 20, 000 paper titles from 20 different MeSH major topics as in Table TABREF25 . As described in Table TABREF24 , the max length of selected paper titles is 53. For these datasets, we randomly select 10% of data as the development set. Since SearchSnippets has been pre-processed by Phan et al. BIBREF41 , we do not further process this dataset. In StackOverflow, texts contain lots of computer terminology, and symbols and capital letters are meaningful, thus we do not do any pre-processed procedures. For Biomedical, we remove the symbols and convert letters into lower case. Pre-trained Word Vectors We use the publicly available word2vec tool to train word embeddings, and the most parameters are set as same as Mikolov et al. BIBREF23 to train word vectors on Google News setting, except of vector dimensionality using 48 and minimize count using 5. For SearchSnippets, we train word vectors on Wikipedia dumps. For StackOverflow, we train word vectors on the whole corpus of the StackOverflow dataset described above which includes the question titles and post contents. For Biomedical, we train word vectors on all titles and abstracts of 2014 training articles. The coverage of these learned vectors on three datasets are listed in Table TABREF32 , and the words not present in the set of pre-trained words are initialized randomly. Comparisons In our experiment, some widely used text clustering methods are compared with our approach. Besides K-means, Skip-thought Vectors, Recursive Neural Network and Paragraph Vector based clustering methods, four baseline clustering methods are directly based on the popular unsupervised dimensionality reduction methods as described in Section SECREF11 . We further compare our approach with some other non-biased neural networks, such as bidirectional RNN. More details are listed as follows: K-means K-means BIBREF42 on original keyword features which are respectively weighted with term frequency (TF) and term frequency-inverse document frequency (TF-IDF). Skip-thought Vectors (SkipVec) This baseline BIBREF35 gives an off-the-shelf encoder to produce highly generic sentence representations. The encoder is trained using a large collection of novels and provides three encoder modes, that are unidirectional encoder (SkipVec (Uni)) with 2,400 dimensions, bidirectional encoder (SkipVec (Bi)) with 2,400 dimensions and combined encoder (SkipVec (Combine)) with SkipVec (Uni) and SkipVec (Bi) of 2,400 dimensions each. K-means is employed on the these vector representations respectively. Recursive Neural Network (RecNN) In BIBREF6 , the tree structure is firstly greedy approximated via unsupervised recursive autoencoder. Then, semi-supervised recursive autoencoders are used to capture the semantics of texts based on the predicted structure. In order to make this recursive-based method completely unsupervised, we remove the cross-entropy error in the second phrase to learn vector representation and subsequently employ K-means on the learned vectors of the top tree node and the average of all vectors in the tree. Paragraph Vector (Para2vec) K-means on the fixed size feature vectors generated by Paragraph Vector (Para2vec) BIBREF25 which is an unsupervised method to learn distributed representation of words and paragraphs. In our experiments, we use the open source software released by Mesnil et al. BIBREF43 . Average Embedding (AE) K-means on the weighted average vectors of the word embeddings which are respectively weighted with TF and TF-IDF. The dimension of average vectors is equal to and decided by the dimension of word vectors used in our experiments. Latent Semantic Analysis (LSA) K-means on the reduced subspace vectors generated by Singular Value Decomposition (SVD) method. The dimension of subspace is default set to the number of clusters, we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 10 on SearchSnippets, 20 on StackOverflow and 20 on Biomedical in our experiments. Laplacian Eigenmaps (LE) This baseline, using Laplacian Eigenmaps and subsequently employing K-means algorithm, is well known as spectral clustering BIBREF44 . The dimension of subspace is default set to the number of clusters BIBREF18 , BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 70 on StackOverflow and 30 on Biomedical in our experiments. Locality Preserving Indexing (LPI) This baseline, projecting the texts into a lower dimensional semantic space, can discover both the geometric and discriminating structures of the original feature space BIBREF38 . The dimension of subspace is default set to the number of clusters BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 80 on StackOverflow and 30 on Biomedical in our experiments. bidirectional RNN (bi-RNN) We replace the CNN model in our framework as in Figure FIGREF5 with some bi-RNN models. Particularly, LSTM and GRU units are used in the experiments. In order to generate the fixed-length document representation from the variable-length vector sequences, for both bi-LSTM and bi-GRU based clustering methods, we further utilize three pooling methods: last pooling (using the last hidden state), mean pooling and element-wise max pooling. These pooling methods are respectively used in the previous works BIBREF45 , BIBREF27 , BIBREF46 and BIBREF9 . For regularization, the training gradients of all parameters with an INLINEFORM0 2 norm larger than 40 are clipped to 40, as the previous work BIBREF47 . Evaluation Metrics The clustering performance is evaluated by comparing the clustering results of texts with the tags/labels provided by the text corpus. Two metrics, the accuracy (ACC) and the normalized mutual information metric (NMI), are used to measure the clustering performance BIBREF38 , BIBREF48 . Given a text INLINEFORM0 , let INLINEFORM1 and INLINEFORM2 be the obtained cluster label and the label provided by the corpus, respectively. Accuracy is defined as: DISPLAYFORM0 where, INLINEFORM0 is the total number of texts, INLINEFORM1 is the indicator function that equals one if INLINEFORM2 and equals zero otherwise, and INLINEFORM3 is the permutation mapping function that maps each cluster label INLINEFORM4 to the equivalent label from the text data by Hungarian algorithm BIBREF49 . Normalized mutual information BIBREF50 between tag/label set INLINEFORM0 and cluster set INLINEFORM1 is a popular metric used for evaluating clustering tasks. It is defined as follows: DISPLAYFORM0 where, INLINEFORM0 is the mutual information between INLINEFORM1 and INLINEFORM2 , INLINEFORM3 is entropy and the denominator INLINEFORM4 is used for normalizing the mutual information to be in the range of [0, 1]. Hyperparameter Settings The most of parameters are set uniformly for these datasets. Following previous study BIBREF38 , the number of nearest neighbors in Eqn. ( EQREF15 ) is fixed to 15 when constructing the graph structures for LE and LPI. For CNN model, the networks has two convolutional layers. The widths of the convolutional filters are both 3. The value of INLINEFORM0 for the top INLINEFORM1 -max pooling in Eqn. ( EQREF10 ) is 5. The number of feature maps at the first convolutional layer is 12, and 8 feature maps at the second convolutional layer. Both those two convolutional layers are followed by a folding layer. We further set the dimension of word embeddings INLINEFORM2 as 48. Finally, the dimension of the deep feature representation INLINEFORM3 is fixed to 480. Moreover, we set the learning rate INLINEFORM4 as 0.01 and the mini-batch training size as 200. The output size INLINEFORM5 in Eqn. ( EQREF19 ) is set same as the best dimensions of subspace in the baseline method, as described in Section SECREF37 . For initial centroids have significant impact on clustering results when utilizing the K-means algorithms, we repeat K-means for multiple times with random initial centroids (specifically, 100 times for statistical significance) as Huang BIBREF48 . The all subspace vectors are normalized to 1 before applying K-means and the final results reported are the average of 5 trials with all clustering methods on three text datasets. Results and Analysis In Table TABREF43 and Table TABREF44 , we report the ACC and NMI performance of our proposed approaches and four baseline methods, K-means, SkipVec, RecNN and Para2vec based clustering methods. Intuitively, we get a general observation that (1) BoW based approaches, including K-means (TF) and K-means (TF-IDF), and SkipVec based approaches perform not well; (2) RecNN based approaches, both RecNN (Ave.) and RecNN (Top+Ave.), do better; (3) Para2vec makes a comparable performance with the most baselines; and (4) the evaluation clearly demonstrate the superiority of our proposed methods STC INLINEFORM0 . It is an expected results. For SkipVec based approaches, the off-the-shelf encoders are trained on the BookCorpus datasets BIBREF51 , and then applied to our datasets to extract the sentence representations. The SkipVec encoders can produce generic sentence representations but may not perform well for specific datasets, in our experiments, StackOverflow and Biomedical datasets consist of many computer terms and medical terms, such as “ASP.NET”, “XML”, “C#”, “serum” and “glycolytic”. When we take a more careful look, we find that RecNN (Top) does poorly, even worse than K-means (TF-IDF). The reason maybe that although recursive neural models introduce tree structure to capture compositional semantics, the vector of the top node mainly captures a biased semantic while the average of all vectors in the tree nodes, such as RecNN (Ave.), can be better to represent sentence level semantic. And we also get another observation that, although our proposed STC INLINEFORM1 -LE and STC INLINEFORM2 -LPI outperform both BoW based and RecNN based approaches across all three datasets, STC INLINEFORM3 -AE and STC INLINEFORM4 -LSA do just exhibit some similar performances as RecNN (Ave.) and RecNN (Top+Ave.) do in the datasets of StackOverflow and Biomedical. We further replace the CNN model in our framework as in Figure FIGREF5 with some other non-biased models, such as bi-LSTM and bi-GRU, and report the results in Table TABREF46 and Table TABREF47 . As an instance, the binary codes generated from LPI are used to guide the learning of bi-LSTM/bi-GRU models. From the results, we can see that bi-GRU and bi-LSTM based clustering methods do equally well, no clear winner, and both achieve great enhancements compared with LPI (best). Compared with these bi-LSTM/bi-GRU based models, the evaluation results still demonstrate the superiority of our approach methods, CNN based clustering model, in the most cases. As the results reported by Visin et al. BIBREF33 , despite bi-directional or multi-directional RNN models perform a good non-biased feature extraction, they yet do not outperform state-of-the-art CNN on some tasks. In order to make clear what factors make our proposed method work, we report the bar chart results of ACC and MNI of our proposed methods and the corresponding baseline methods in Figure FIGREF49 and Figure FIGREF53 . It is clear that, although AE and LSA does well or even better than LE and LPI, especially in dataset of both StackOverflow and Biomedical, STC INLINEFORM0 -LE and STC INLINEFORM1 -LPI achieve a much larger performance enhancements than STC INLINEFORM2 -AE and STC INLINEFORM3 -LSA do. The possible reason is that the information the pseudo supervision used to guide the learning of CNN model that make difference. Especially, for AE case, the input features fed into CNN model and the pseudo supervision employed to guide the learning of CNN model are all come from word embeddings. There are no different semantic features to be used into our proposed method, thus the performance enhancements are limited in STC INLINEFORM4 -AE. For LSA case, as we known, LSA is to make matrix factorization to find the best subspace approximation of the original feature space to minimize the global reconstruction error. And as BIBREF24 , BIBREF52 recently point out that word embeddings trained with word2vec or some variances, is essentially to do an operation of matrix factorization. Therefore, the information between input and the pseudo supervision in CNN is not departed very largely from each other, and the performance enhancements of STC INLINEFORM5 -AE is also not quite satisfactory. For LE and LPI case, as we known that LE extracts the manifold structure of the original feature space, and LPI extracts both geometric and discriminating structure of the original feature space BIBREF38 . We guess that our approach STC INLINEFORM6 -LE and STC INLINEFORM7 -LPI achieve enhancements compared with both LE and LPI by a large margin, because both of LE and LPI get useful semantic features, and these features are also different from word embeddings used as input of CNN. From this view, we say that our proposed STC has potential to behave more effective when the pseudo supervision is able to get semantic meaningful features, which is different enough from the input of CNN. Furthermore, from the results of K-means and AE in Table TABREF43 - TABREF44 and Figure FIGREF49 - FIGREF53 , we note that TF-IDF weighting gives a more remarkable improvement for K-means, while TF weighting works better than TF-IDF weighting for Average Embedding. Maybe the reason is that pre-trained word embeddings encode some useful information from external corpus and are able to get even better results without TF-IDF weighting. Meanwhile, we find that LE get quite unusual good performance than LPI, LSA and AE in SearchSnippets dataset, which is not found in the other two datasets. To get clear about this, and also to make a much better demonstration about our proposed approaches and other baselines, we further report 2-dimensional text embeddings on SearchSnippets in Figure FIGREF58 , using t-SNE BIBREF53 to get distributed stochastic neighbor embedding of the feature representations used in the clustering methods. We can see that the results of from AE and LSA seem to be fairly good or even better than the ones from LE and LPI, which is not the same as the results from ACC and NMI in Figure FIGREF49 - FIGREF53 . Meanwhile, RecNN (Ave.) performs better than BoW (both TF and TF-IDF) while RecNN (Top) does not, which is the same as the results from ACC and NMI in Table TABREF43 and Table TABREF44 . Then we guess that both ”the same as” and ”not the same as” above, is just a good example to illustrate that visualization tool, such as t-SNE, get some useful information for measuring results, which is different from the ones of ACC and NMI. Moreover, from this complementary view of t-SNE, we can see that our STC INLINEFORM0 -AE, STC INLINEFORM1 -LSA, STC INLINEFORM2 -LE, and STC INLINEFORM3 -LPI show more clear-cut margins among different semantic topics (that is, tags/labels), compared with AE, LSA, LE and LPI, respectively, as well as compared with both baselines, BoW and RecNN based ones. From all these results, with three measures of ACC, NMI and t-SNE under three datasets, we can get a solid conclusion that our proposed approaches is an effective approaches to get useful semantic features for short text clustering. Conclusions With the emergence of social media, short text clustering has become an increasing important task. This paper explores a new perspective to cluster short texts based on deep feature representation learned from the proposed self-taught convolutional neural networks. Our framework can be successfully accomplished without using any external tags/labels and complicated NLP pre-processing, and and our approach is a flexible framework, in which the traditional dimension reduction approaches could be used to get performance enhancement. Our extensive experimental study on three short text datasets shows that our approach can achieve a significantly better performance. In the future, how to select and incorporate more effective semantic features into the proposed framework would call for more research. Acknowledgments We would like to thank reviewers for their comments, and acknowledge Kaggle and BioASQ for making the datasets available. This work is supported by the National Natural Science Foundation of China (No. 61602479, No. 61303172, No. 61403385) and the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB02070005).

On SearchSnippets dataset ACC 77.01%, NMI 62.94%, on StackOverflow dataset ACC 51.14%, NMI 49.08%, on Biomedical dataset ACC 43.00%, NMI 38.18%

5a0841cc0628e872fe473874694f4ab9411a1d10

5a0841cc0628e872fe473874694f4ab9411a1d10_0

Q: By how much did they outperform the other methods? Text: Introduction Short text clustering is of great importance due to its various applications, such as user profiling BIBREF0 and recommendation BIBREF1 , for nowaday's social media dataset emerged day by day. However, short text clustering has the data sparsity problem and most words only occur once in each short text BIBREF2 . As a result, the Term Frequency-Inverse Document Frequency (TF-IDF) measure cannot work well in short text setting. In order to address this problem, some researchers work on expanding and enriching the context of data from Wikipedia BIBREF3 or an ontology BIBREF4 . However, these methods involve solid Natural Language Processing (NLP) knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another way to overcome these issues is to explore some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Yet how to design an effective model is an open question, and most of these methods directly trained based on Bag-of-Words (BoW) are shallow structures which cannot preserve the accurate semantic similarities. Recently, with the help of word embedding, neural networks demonstrate their great performance in terms of constructing text representation, such as Recursive Neural Network (RecNN) BIBREF6 , BIBREF7 and Recurrent Neural Network (RNN) BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the hidden layer computed at the last word to represent the text, is a biased model where later words are more dominant than earlier words BIBREF9 . Whereas for the non-biased models, the learned representation of one text can be extracted from all the words in the text with non-dominant learned weights. More recently, Convolution Neural Network (CNN), as the most popular non-biased model and applying convolutional filters to capture local features, has achieved a better performance in many NLP applications, such as sentence modeling BIBREF10 , relation classification BIBREF11 , and other traditional NLP tasks BIBREF12 . Most of the previous works focus CNN on solving supervised NLP tasks, while in this paper we aim to explore the power of CNN on one unsupervised NLP task, short text clustering. We systematically introduce a simple yet surprisingly powerful Self-Taught Convolutional neural network framework for Short Text Clustering, called STC INLINEFORM0 . An overall architecture of our proposed approach is illustrated in Figure FIGREF5 . We, inspired by BIBREF13 , BIBREF14 , utilize a self-taught learning framework into our task. In particular, the original raw text features are first embedded into compact binary codes INLINEFORM1 with the help of one traditional unsupervised dimensionality reduction function. Then text matrix INLINEFORM2 projected from word embeddings are fed into CNN model to learn the deep feature representation INLINEFORM3 and the output units are used to fit the pre-trained binary codes INLINEFORM4 . After obtaining the learned features, K-means algorithm is employed on them to cluster texts into clusters INLINEFORM5 . Obviously, we call our approach “self-taught” because the CNN model is learnt from the pseudo labels generated from the previous stage, which is quite different from the term “self-taught” in BIBREF15 . Our main contributions can be summarized as follows: This work is an extension of our conference paper BIBREF16 , and they differ in the following aspects. First, we put forward a general a self-taught CNN framework in this paper which can flexibly couple various semantic features, whereas the conference version can be seen as a specific example of this work. Second, in this paper we use a new short text dataset, Biomedical, in the experiment to verify the effectiveness of our approach. Third, we put much effort on studying the influence of various different semantic features integrated in our self-taught CNN framework, which is not involved in the conference paper. For the purpose of reproducibility, we make the datasets and software used in our experiments publicly available at the website. The remainder of this paper is organized as follows: In Section SECREF2 , we first briefly survey several related works. In Section SECREF3 , we describe the proposed approach STC INLINEFORM0 and implementation details. Experimental results and analyses are presented in Section SECREF4 . Finally, conclusions are given in the last Section. Related Work In this section, we review the related work from the following two perspectives: short text clustering and deep neural networks. Short Text Clustering There have been several studies that attempted to overcome the sparseness of short text representation. One way is to expand and enrich the context of data. For example, Banerjee et al. BIBREF3 proposed a method of improving the accuracy of short text clustering by enriching their representation with additional features from Wikipedia, and Fodeh et al. BIBREF4 incorporate semantic knowledge from an ontology into text clustering. However, these works need solid NLP knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another direction is to map the original features into reduced space, such as Latent Semantic Analysis (LSA) BIBREF17 , Laplacian Eigenmaps (LE) BIBREF18 , and Locality Preserving Indexing (LPI) BIBREF19 . Even some researchers explored some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Moreover, some studies even focus the above both two streams. For example, Tang et al. BIBREF20 proposed a novel framework which enrich the text features by employing machine translation and reduce the original features simultaneously through matrix factorization techniques. Despite the above clustering methods can alleviate sparseness of short text representation to some extent, most of them ignore word order in the text and belong to shallow structures which can not fully capture accurate semantic similarities. Deep Neural Networks Recently, there is a revival of interest in DNN and many researchers have concentrated on using Deep Learning to learn features. Hinton and Salakhutdinov BIBREF21 use DAE to learn text representation. During the fine-tuning procedure, they use backpropagation to find codes that are good at reconstructing the word-count vector. More recently, researchers propose to use external corpus to learn a distributed representation for each word, called word embedding BIBREF22 , to improve DNN performance on NLP tasks. The Skip-gram and continuous bag-of-words models of Word2vec BIBREF23 propose a simple single-layer architecture based on the inner product between two word vectors, and Pennington et al. BIBREF24 introduce a new model for word representation, called GloVe, which captures the global corpus statistics. In order to learn the compact representation vectors of sentences, Le and Mikolov BIBREF25 directly extend the previous Word2vec BIBREF23 by predicting words in the sentence, which is named Paragraph Vector (Para2vec). Para2vec is still a shallow window-based method and need a larger corpus to yield better performance. More neural networks utilize word embedding to capture true meaningful syntactic and semantic regularities, such as RecNN BIBREF6 , BIBREF7 and RNN BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the layer computed at the last word to represent the text, is a biased model. Recently, Long Short-Term Memory (LSTM) BIBREF26 and Gated Recurrent Unit (GRU) BIBREF27 , as sophisticated recurrent hidden units of RNN, has presented its advantages in many sequence generation problem, such as machine translation BIBREF28 , speech recognition BIBREF29 , and text conversation BIBREF30 . While, CNN is better to learn non-biased implicit features which has been successfully exploited for many supervised NLP learning tasks as described in Section SECREF1 , and various CNN based variants are proposed in the recent works, such as Dynamic Convolutional Neural Network (DCNN) BIBREF10 , Gated Recursive Convolutional Neural Network (grConv) BIBREF31 and Self-Adaptive Hierarchical Sentence model (AdaSent) BIBREF32 . In the past few days, Visin et al. BIBREF33 have attempted to replace convolutional layer in CNN to learn non-biased features for object recognition with four RNNs, called ReNet, that sweep over lower-layer features in different directions: (1) bottom to top, (2) top to bottom, (3) left to right and (4) right to left. However, ReNet does not outperform state-of-the-art convolutional neural networks on any of the three benchmark datasets, and it is also a supervised learning model for classification. Inspired by Skip-gram of word2vec BIBREF34 , BIBREF23 , Skip-thought model BIBREF35 describe an approach for unsupervised learning of a generic, distributed sentence encoder. Similar as Skip-gram model, Skip-thought model trains an encoder-decoder model that tries to reconstruct the surrounding sentences of an encoded sentence and released an off-the-shelf encoder to extract sentence representation. Even some researchers introduce continuous Skip-gram and negative sampling to CNN for learning visual representation in an unsupervised manner BIBREF36 . This paper, from a new perspective, puts forward a general self-taught CNN framework which can flexibly couple various semantic features and achieve a good performance on one unsupervised learning task, short text clustering. Methodology Assume that we are given a dataset of INLINEFORM0 training texts denoted as: INLINEFORM1 , where INLINEFORM2 is the dimensionality of the original BoW representation. Denote its tag set as INLINEFORM3 and the pre-trained word embedding set as INLINEFORM4 , where INLINEFORM5 is the dimensionality of word vectors and INLINEFORM6 is the vocabulary size. In order to learn the INLINEFORM7 -dimensional deep feature representation INLINEFORM8 from CNN in an unsupervised manner, some unsupervised dimensionality reduction methods INLINEFORM9 are employed to guide the learning of CNN model. Our goal is to cluster these texts INLINEFORM10 into clusters INLINEFORM11 based on the learned deep feature representation while preserving the semantic consistency. As depicted in Figure FIGREF5 , the proposed framework consist of three components, deep convolutional neural network (CNN), unsupervised dimensionality reduction function and K-means module. In the rest sections, we first present the first two components respectively, and then give the trainable parameters and the objective function to learn the deep feature representation. Finally, the last section describe how to perform clustering on the learned features. Deep Convolutional Neural Networks In this section, we briefly review one popular deep convolutional neural network, Dynamic Convolutional Neural Network (DCNN) BIBREF10 as an instance of CNN in the following sections, which as the foundation of our proposed method has been successfully proposed for the completely supervised learning task, text classification. Taking a neural network with two convolutional layers in Figure FIGREF9 as an example, the network transforms raw input text to a powerful representation. Particularly, each raw text vector INLINEFORM0 is projected into a matrix representation INLINEFORM1 by looking up a word embedding INLINEFORM2 , where INLINEFORM3 is the length of one text. We also let INLINEFORM4 and INLINEFORM5 denote the weights of the neural networks. The network defines a transformation INLINEFORM6 INLINEFORM7 which transforms an input raw text INLINEFORM8 to a INLINEFORM9 -dimensional deep representation INLINEFORM10 . There are three basic operations described as follows: Wide one-dimensional convolution This operation INLINEFORM0 is applied to an individual row of the sentence matrix INLINEFORM1 , and yields a resulting matrix INLINEFORM2 , where INLINEFORM3 is the width of convolutional filter. Folding In this operation, every two rows in a feature map are simply summed component-wisely. For a map of INLINEFORM0 rows, folding returns a map of INLINEFORM1 rows, thus halving the size of the representation and yielding a matrix feature INLINEFORM2 . Note that folding operation does not introduce any additional parameters. Dynamic INLINEFORM0 -max pooling Assuming the pooling parameter as INLINEFORM1 , INLINEFORM2 -max pooling selects the sub-matrix INLINEFORM3 of the INLINEFORM4 highest values in each row of the matrix INLINEFORM5 . For dynamic INLINEFORM6 -max pooling, the pooling parameter INLINEFORM7 is dynamically selected in order to allow for a smooth extraction of higher-order and longer-range features BIBREF10 . Given a fixed pooling parameter INLINEFORM8 for the topmost convolutional layer, the parameter INLINEFORM9 of INLINEFORM10 -max pooling in the INLINEFORM11 -th convolutional layer can be computed as follows: DISPLAYFORM0 where INLINEFORM0 is the total number of convolutional layers in the network. Unsupervised Dimensionality Reduction As described in Figure FIGREF5 , the dimensionality reduction function is defined as follows: DISPLAYFORM0 where, INLINEFORM0 are the INLINEFORM1 -dimensional reduced latent space representations. Here, we take four popular dimensionality reduction methods as examples in our framework. Average Embedding (AE): This method directly averages the word embeddings which are respectively weighted with TF and TF-IDF. Huang et al. BIBREF37 used this strategy as the global context in their task, and Socher et al. BIBREF7 and Lai et al. BIBREF9 used this method for text classification. The weighted average of all word vectors in one text can be computed as follows: DISPLAYFORM0 where INLINEFORM0 can be any weighting function that captures the importance of word INLINEFORM1 in the text INLINEFORM2 . Latent Semantic Analysis (LSA): LSA BIBREF17 is the most popular global matrix factorization method, which applies a dimension reducing linear projection, Singular Value Decomposition (SVD), of the corresponding term/document matrix. Suppose the rank of INLINEFORM0 is INLINEFORM1 , LSA decompose INLINEFORM2 into the product of three other matrices: DISPLAYFORM0 where INLINEFORM0 and INLINEFORM1 are the singular values of INLINEFORM2 , INLINEFORM3 is a set of left singular vectors and INLINEFORM4 is a set of right singular vectors. LSA uses the top INLINEFORM5 vectors in INLINEFORM6 as the transformation matrix to embed the original text features into a INLINEFORM7 -dimensional subspace INLINEFORM8 BIBREF17 . Laplacian Eigenmaps (LE): The top eigenvectors of graph Laplacian, defined on the similarity matrix of texts, are used in the method, which can discover the manifold structure of the text space BIBREF18 . In order to avoid storing the dense similarity matrix, many approximation techniques are proposed to reduce the memory usage and computational complexity for LE. There are two representative approximation methods, sparse similarity matrix and Nystr INLINEFORM0 m approximation. Following previous studies BIBREF38 , BIBREF13 , we select the former technique to construct the INLINEFORM1 local similarity matrix INLINEFORM2 by using heat kernel as follows: DISPLAYFORM0 where, INLINEFORM0 is a tuning parameter (default is 1) and INLINEFORM1 represents the set of INLINEFORM2 -nearest-neighbors of INLINEFORM3 . By introducing a diagonal INLINEFORM4 matrix INLINEFORM5 whose entries are given by INLINEFORM6 , the graph Laplacian INLINEFORM7 can be computed by ( INLINEFORM8 ). The optimal INLINEFORM9 real-valued matrix INLINEFORM10 can be obtained by solving the following objective function: DISPLAYFORM0 where INLINEFORM0 is the trace function, INLINEFORM1 requires the different dimensions to be uncorrelated, and INLINEFORM2 requires each dimension to achieve equal probability as positive or negative). Locality Preserving Indexing (LPI): This method extends LE to deal with unseen texts by approximating the linear function INLINEFORM0 BIBREF13 , and the subspace vectors are obtained by finding the optimal linear approximations to the eigenfunctions of the Laplace Beltrami operator on the Riemannian manifold BIBREF19 . Similar as LE, we first construct the local similarity matrix INLINEFORM1 , then the graph Laplacian INLINEFORM2 can be computed by ( INLINEFORM3 ), where INLINEFORM4 measures the local density around INLINEFORM5 and is equal to INLINEFORM6 . Compute the eigenvectors INLINEFORM7 and eigenvalues INLINEFORM8 of the following generalized eigen-problem: DISPLAYFORM0 The mapping function INLINEFORM0 can be obtained and applied to the unseen data BIBREF38 . All of the above methods claim a better performance in capturing semantic similarity between texts in the reduced latent space representation INLINEFORM0 than in the original representation INLINEFORM1 , while the performance of short text clustering can be further enhanced with the help of our framework, self-taught CNN. Learning The last layer of CNN is an output layer as follows: DISPLAYFORM0 where, INLINEFORM0 is the deep feature representation, INLINEFORM1 is the output vector and INLINEFORM2 is weight matrix. In order to incorporate the latent semantic features INLINEFORM0 , we first binary the real-valued vectors INLINEFORM1 to the binary codes INLINEFORM2 by setting the threshold to be the media vector INLINEFORM3 . Then, the output vector INLINEFORM4 is used to fit the binary codes INLINEFORM5 via INLINEFORM6 logistic operations as follows: DISPLAYFORM0 All parameters to be trained are defined as INLINEFORM0 . DISPLAYFORM0 Given the training text collection INLINEFORM0 , and the pre-trained binary codes INLINEFORM1 , the log likelihood of the parameters can be written down as follows: DISPLAYFORM0 Following the previous work BIBREF10 , we train the network with mini-batches by back-propagation and perform the gradient-based optimization using the Adagrad update rule BIBREF39 . For regularization, we employ dropout with 50% rate to the penultimate layer BIBREF10 , BIBREF40 . K-means for Clustering With the given short texts, we first utilize the trained deep neural network to obtain the semantic representations INLINEFORM0 , and then employ traditional K-means algorithm to perform clustering. Datasets We test our proposed approach on three public short text datasets. The summary statistics and semantic topics of these datasets are described in Table TABREF24 and Table TABREF25 . SearchSnippets. This dataset was selected from the results of web search transaction using predefined phrases of 8 different domains by Phan et al. BIBREF41 . StackOverflow. We use the challenge data published in Kaggle.com. The raw dataset consists 3,370,528 samples through July 31st, 2012 to August 14, 2012. In our experiments, we randomly select 20,000 question titles from 20 different tags as in Table TABREF25 . Biomedical. We use the challenge data published in BioASQ's official website. In our experiments, we randomly select 20, 000 paper titles from 20 different MeSH major topics as in Table TABREF25 . As described in Table TABREF24 , the max length of selected paper titles is 53. For these datasets, we randomly select 10% of data as the development set. Since SearchSnippets has been pre-processed by Phan et al. BIBREF41 , we do not further process this dataset. In StackOverflow, texts contain lots of computer terminology, and symbols and capital letters are meaningful, thus we do not do any pre-processed procedures. For Biomedical, we remove the symbols and convert letters into lower case. Pre-trained Word Vectors We use the publicly available word2vec tool to train word embeddings, and the most parameters are set as same as Mikolov et al. BIBREF23 to train word vectors on Google News setting, except of vector dimensionality using 48 and minimize count using 5. For SearchSnippets, we train word vectors on Wikipedia dumps. For StackOverflow, we train word vectors on the whole corpus of the StackOverflow dataset described above which includes the question titles and post contents. For Biomedical, we train word vectors on all titles and abstracts of 2014 training articles. The coverage of these learned vectors on three datasets are listed in Table TABREF32 , and the words not present in the set of pre-trained words are initialized randomly. Comparisons In our experiment, some widely used text clustering methods are compared with our approach. Besides K-means, Skip-thought Vectors, Recursive Neural Network and Paragraph Vector based clustering methods, four baseline clustering methods are directly based on the popular unsupervised dimensionality reduction methods as described in Section SECREF11 . We further compare our approach with some other non-biased neural networks, such as bidirectional RNN. More details are listed as follows: K-means K-means BIBREF42 on original keyword features which are respectively weighted with term frequency (TF) and term frequency-inverse document frequency (TF-IDF). Skip-thought Vectors (SkipVec) This baseline BIBREF35 gives an off-the-shelf encoder to produce highly generic sentence representations. The encoder is trained using a large collection of novels and provides three encoder modes, that are unidirectional encoder (SkipVec (Uni)) with 2,400 dimensions, bidirectional encoder (SkipVec (Bi)) with 2,400 dimensions and combined encoder (SkipVec (Combine)) with SkipVec (Uni) and SkipVec (Bi) of 2,400 dimensions each. K-means is employed on the these vector representations respectively. Recursive Neural Network (RecNN) In BIBREF6 , the tree structure is firstly greedy approximated via unsupervised recursive autoencoder. Then, semi-supervised recursive autoencoders are used to capture the semantics of texts based on the predicted structure. In order to make this recursive-based method completely unsupervised, we remove the cross-entropy error in the second phrase to learn vector representation and subsequently employ K-means on the learned vectors of the top tree node and the average of all vectors in the tree. Paragraph Vector (Para2vec) K-means on the fixed size feature vectors generated by Paragraph Vector (Para2vec) BIBREF25 which is an unsupervised method to learn distributed representation of words and paragraphs. In our experiments, we use the open source software released by Mesnil et al. BIBREF43 . Average Embedding (AE) K-means on the weighted average vectors of the word embeddings which are respectively weighted with TF and TF-IDF. The dimension of average vectors is equal to and decided by the dimension of word vectors used in our experiments. Latent Semantic Analysis (LSA) K-means on the reduced subspace vectors generated by Singular Value Decomposition (SVD) method. The dimension of subspace is default set to the number of clusters, we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 10 on SearchSnippets, 20 on StackOverflow and 20 on Biomedical in our experiments. Laplacian Eigenmaps (LE) This baseline, using Laplacian Eigenmaps and subsequently employing K-means algorithm, is well known as spectral clustering BIBREF44 . The dimension of subspace is default set to the number of clusters BIBREF18 , BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 70 on StackOverflow and 30 on Biomedical in our experiments. Locality Preserving Indexing (LPI) This baseline, projecting the texts into a lower dimensional semantic space, can discover both the geometric and discriminating structures of the original feature space BIBREF38 . The dimension of subspace is default set to the number of clusters BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 80 on StackOverflow and 30 on Biomedical in our experiments. bidirectional RNN (bi-RNN) We replace the CNN model in our framework as in Figure FIGREF5 with some bi-RNN models. Particularly, LSTM and GRU units are used in the experiments. In order to generate the fixed-length document representation from the variable-length vector sequences, for both bi-LSTM and bi-GRU based clustering methods, we further utilize three pooling methods: last pooling (using the last hidden state), mean pooling and element-wise max pooling. These pooling methods are respectively used in the previous works BIBREF45 , BIBREF27 , BIBREF46 and BIBREF9 . For regularization, the training gradients of all parameters with an INLINEFORM0 2 norm larger than 40 are clipped to 40, as the previous work BIBREF47 . Evaluation Metrics The clustering performance is evaluated by comparing the clustering results of texts with the tags/labels provided by the text corpus. Two metrics, the accuracy (ACC) and the normalized mutual information metric (NMI), are used to measure the clustering performance BIBREF38 , BIBREF48 . Given a text INLINEFORM0 , let INLINEFORM1 and INLINEFORM2 be the obtained cluster label and the label provided by the corpus, respectively. Accuracy is defined as: DISPLAYFORM0 where, INLINEFORM0 is the total number of texts, INLINEFORM1 is the indicator function that equals one if INLINEFORM2 and equals zero otherwise, and INLINEFORM3 is the permutation mapping function that maps each cluster label INLINEFORM4 to the equivalent label from the text data by Hungarian algorithm BIBREF49 . Normalized mutual information BIBREF50 between tag/label set INLINEFORM0 and cluster set INLINEFORM1 is a popular metric used for evaluating clustering tasks. It is defined as follows: DISPLAYFORM0 where, INLINEFORM0 is the mutual information between INLINEFORM1 and INLINEFORM2 , INLINEFORM3 is entropy and the denominator INLINEFORM4 is used for normalizing the mutual information to be in the range of [0, 1]. Hyperparameter Settings The most of parameters are set uniformly for these datasets. Following previous study BIBREF38 , the number of nearest neighbors in Eqn. ( EQREF15 ) is fixed to 15 when constructing the graph structures for LE and LPI. For CNN model, the networks has two convolutional layers. The widths of the convolutional filters are both 3. The value of INLINEFORM0 for the top INLINEFORM1 -max pooling in Eqn. ( EQREF10 ) is 5. The number of feature maps at the first convolutional layer is 12, and 8 feature maps at the second convolutional layer. Both those two convolutional layers are followed by a folding layer. We further set the dimension of word embeddings INLINEFORM2 as 48. Finally, the dimension of the deep feature representation INLINEFORM3 is fixed to 480. Moreover, we set the learning rate INLINEFORM4 as 0.01 and the mini-batch training size as 200. The output size INLINEFORM5 in Eqn. ( EQREF19 ) is set same as the best dimensions of subspace in the baseline method, as described in Section SECREF37 . For initial centroids have significant impact on clustering results when utilizing the K-means algorithms, we repeat K-means for multiple times with random initial centroids (specifically, 100 times for statistical significance) as Huang BIBREF48 . The all subspace vectors are normalized to 1 before applying K-means and the final results reported are the average of 5 trials with all clustering methods on three text datasets. Results and Analysis In Table TABREF43 and Table TABREF44 , we report the ACC and NMI performance of our proposed approaches and four baseline methods, K-means, SkipVec, RecNN and Para2vec based clustering methods. Intuitively, we get a general observation that (1) BoW based approaches, including K-means (TF) and K-means (TF-IDF), and SkipVec based approaches perform not well; (2) RecNN based approaches, both RecNN (Ave.) and RecNN (Top+Ave.), do better; (3) Para2vec makes a comparable performance with the most baselines; and (4) the evaluation clearly demonstrate the superiority of our proposed methods STC INLINEFORM0 . It is an expected results. For SkipVec based approaches, the off-the-shelf encoders are trained on the BookCorpus datasets BIBREF51 , and then applied to our datasets to extract the sentence representations. The SkipVec encoders can produce generic sentence representations but may not perform well for specific datasets, in our experiments, StackOverflow and Biomedical datasets consist of many computer terms and medical terms, such as “ASP.NET”, “XML”, “C#”, “serum” and “glycolytic”. When we take a more careful look, we find that RecNN (Top) does poorly, even worse than K-means (TF-IDF). The reason maybe that although recursive neural models introduce tree structure to capture compositional semantics, the vector of the top node mainly captures a biased semantic while the average of all vectors in the tree nodes, such as RecNN (Ave.), can be better to represent sentence level semantic. And we also get another observation that, although our proposed STC INLINEFORM1 -LE and STC INLINEFORM2 -LPI outperform both BoW based and RecNN based approaches across all three datasets, STC INLINEFORM3 -AE and STC INLINEFORM4 -LSA do just exhibit some similar performances as RecNN (Ave.) and RecNN (Top+Ave.) do in the datasets of StackOverflow and Biomedical. We further replace the CNN model in our framework as in Figure FIGREF5 with some other non-biased models, such as bi-LSTM and bi-GRU, and report the results in Table TABREF46 and Table TABREF47 . As an instance, the binary codes generated from LPI are used to guide the learning of bi-LSTM/bi-GRU models. From the results, we can see that bi-GRU and bi-LSTM based clustering methods do equally well, no clear winner, and both achieve great enhancements compared with LPI (best). Compared with these bi-LSTM/bi-GRU based models, the evaluation results still demonstrate the superiority of our approach methods, CNN based clustering model, in the most cases. As the results reported by Visin et al. BIBREF33 , despite bi-directional or multi-directional RNN models perform a good non-biased feature extraction, they yet do not outperform state-of-the-art CNN on some tasks. In order to make clear what factors make our proposed method work, we report the bar chart results of ACC and MNI of our proposed methods and the corresponding baseline methods in Figure FIGREF49 and Figure FIGREF53 . It is clear that, although AE and LSA does well or even better than LE and LPI, especially in dataset of both StackOverflow and Biomedical, STC INLINEFORM0 -LE and STC INLINEFORM1 -LPI achieve a much larger performance enhancements than STC INLINEFORM2 -AE and STC INLINEFORM3 -LSA do. The possible reason is that the information the pseudo supervision used to guide the learning of CNN model that make difference. Especially, for AE case, the input features fed into CNN model and the pseudo supervision employed to guide the learning of CNN model are all come from word embeddings. There are no different semantic features to be used into our proposed method, thus the performance enhancements are limited in STC INLINEFORM4 -AE. For LSA case, as we known, LSA is to make matrix factorization to find the best subspace approximation of the original feature space to minimize the global reconstruction error. And as BIBREF24 , BIBREF52 recently point out that word embeddings trained with word2vec or some variances, is essentially to do an operation of matrix factorization. Therefore, the information between input and the pseudo supervision in CNN is not departed very largely from each other, and the performance enhancements of STC INLINEFORM5 -AE is also not quite satisfactory. For LE and LPI case, as we known that LE extracts the manifold structure of the original feature space, and LPI extracts both geometric and discriminating structure of the original feature space BIBREF38 . We guess that our approach STC INLINEFORM6 -LE and STC INLINEFORM7 -LPI achieve enhancements compared with both LE and LPI by a large margin, because both of LE and LPI get useful semantic features, and these features are also different from word embeddings used as input of CNN. From this view, we say that our proposed STC has potential to behave more effective when the pseudo supervision is able to get semantic meaningful features, which is different enough from the input of CNN. Furthermore, from the results of K-means and AE in Table TABREF43 - TABREF44 and Figure FIGREF49 - FIGREF53 , we note that TF-IDF weighting gives a more remarkable improvement for K-means, while TF weighting works better than TF-IDF weighting for Average Embedding. Maybe the reason is that pre-trained word embeddings encode some useful information from external corpus and are able to get even better results without TF-IDF weighting. Meanwhile, we find that LE get quite unusual good performance than LPI, LSA and AE in SearchSnippets dataset, which is not found in the other two datasets. To get clear about this, and also to make a much better demonstration about our proposed approaches and other baselines, we further report 2-dimensional text embeddings on SearchSnippets in Figure FIGREF58 , using t-SNE BIBREF53 to get distributed stochastic neighbor embedding of the feature representations used in the clustering methods. We can see that the results of from AE and LSA seem to be fairly good or even better than the ones from LE and LPI, which is not the same as the results from ACC and NMI in Figure FIGREF49 - FIGREF53 . Meanwhile, RecNN (Ave.) performs better than BoW (both TF and TF-IDF) while RecNN (Top) does not, which is the same as the results from ACC and NMI in Table TABREF43 and Table TABREF44 . Then we guess that both ”the same as” and ”not the same as” above, is just a good example to illustrate that visualization tool, such as t-SNE, get some useful information for measuring results, which is different from the ones of ACC and NMI. Moreover, from this complementary view of t-SNE, we can see that our STC INLINEFORM0 -AE, STC INLINEFORM1 -LSA, STC INLINEFORM2 -LE, and STC INLINEFORM3 -LPI show more clear-cut margins among different semantic topics (that is, tags/labels), compared with AE, LSA, LE and LPI, respectively, as well as compared with both baselines, BoW and RecNN based ones. From all these results, with three measures of ACC, NMI and t-SNE under three datasets, we can get a solid conclusion that our proposed approaches is an effective approaches to get useful semantic features for short text clustering. Conclusions With the emergence of social media, short text clustering has become an increasing important task. This paper explores a new perspective to cluster short texts based on deep feature representation learned from the proposed self-taught convolutional neural networks. Our framework can be successfully accomplished without using any external tags/labels and complicated NLP pre-processing, and and our approach is a flexible framework, in which the traditional dimension reduction approaches could be used to get performance enhancement. Our extensive experimental study on three short text datasets shows that our approach can achieve a significantly better performance. In the future, how to select and incorporate more effective semantic features into the proposed framework would call for more research. Acknowledgments We would like to thank reviewers for their comments, and acknowledge Kaggle and BioASQ for making the datasets available. This work is supported by the National Natural Science Foundation of China (No. 61602479, No. 61303172, No. 61403385) and the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB02070005).

on SearchSnippets dataset by 6.72% in ACC, by 6.94% in NMI; on Biomedical dataset by 5.77% in ACC, 3.91% in NMI

a5dd569e6d641efa86d2c2b2e970ce5871e0963f

a5dd569e6d641efa86d2c2b2e970ce5871e0963f_0

Q: Which popular clustering methods did they experiment with? Text: Introduction Short text clustering is of great importance due to its various applications, such as user profiling BIBREF0 and recommendation BIBREF1 , for nowaday's social media dataset emerged day by day. However, short text clustering has the data sparsity problem and most words only occur once in each short text BIBREF2 . As a result, the Term Frequency-Inverse Document Frequency (TF-IDF) measure cannot work well in short text setting. In order to address this problem, some researchers work on expanding and enriching the context of data from Wikipedia BIBREF3 or an ontology BIBREF4 . However, these methods involve solid Natural Language Processing (NLP) knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another way to overcome these issues is to explore some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Yet how to design an effective model is an open question, and most of these methods directly trained based on Bag-of-Words (BoW) are shallow structures which cannot preserve the accurate semantic similarities. Recently, with the help of word embedding, neural networks demonstrate their great performance in terms of constructing text representation, such as Recursive Neural Network (RecNN) BIBREF6 , BIBREF7 and Recurrent Neural Network (RNN) BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the hidden layer computed at the last word to represent the text, is a biased model where later words are more dominant than earlier words BIBREF9 . Whereas for the non-biased models, the learned representation of one text can be extracted from all the words in the text with non-dominant learned weights. More recently, Convolution Neural Network (CNN), as the most popular non-biased model and applying convolutional filters to capture local features, has achieved a better performance in many NLP applications, such as sentence modeling BIBREF10 , relation classification BIBREF11 , and other traditional NLP tasks BIBREF12 . Most of the previous works focus CNN on solving supervised NLP tasks, while in this paper we aim to explore the power of CNN on one unsupervised NLP task, short text clustering. We systematically introduce a simple yet surprisingly powerful Self-Taught Convolutional neural network framework for Short Text Clustering, called STC INLINEFORM0 . An overall architecture of our proposed approach is illustrated in Figure FIGREF5 . We, inspired by BIBREF13 , BIBREF14 , utilize a self-taught learning framework into our task. In particular, the original raw text features are first embedded into compact binary codes INLINEFORM1 with the help of one traditional unsupervised dimensionality reduction function. Then text matrix INLINEFORM2 projected from word embeddings are fed into CNN model to learn the deep feature representation INLINEFORM3 and the output units are used to fit the pre-trained binary codes INLINEFORM4 . After obtaining the learned features, K-means algorithm is employed on them to cluster texts into clusters INLINEFORM5 . Obviously, we call our approach “self-taught” because the CNN model is learnt from the pseudo labels generated from the previous stage, which is quite different from the term “self-taught” in BIBREF15 . Our main contributions can be summarized as follows: This work is an extension of our conference paper BIBREF16 , and they differ in the following aspects. First, we put forward a general a self-taught CNN framework in this paper which can flexibly couple various semantic features, whereas the conference version can be seen as a specific example of this work. Second, in this paper we use a new short text dataset, Biomedical, in the experiment to verify the effectiveness of our approach. Third, we put much effort on studying the influence of various different semantic features integrated in our self-taught CNN framework, which is not involved in the conference paper. For the purpose of reproducibility, we make the datasets and software used in our experiments publicly available at the website. The remainder of this paper is organized as follows: In Section SECREF2 , we first briefly survey several related works. In Section SECREF3 , we describe the proposed approach STC INLINEFORM0 and implementation details. Experimental results and analyses are presented in Section SECREF4 . Finally, conclusions are given in the last Section. Related Work In this section, we review the related work from the following two perspectives: short text clustering and deep neural networks. Short Text Clustering There have been several studies that attempted to overcome the sparseness of short text representation. One way is to expand and enrich the context of data. For example, Banerjee et al. BIBREF3 proposed a method of improving the accuracy of short text clustering by enriching their representation with additional features from Wikipedia, and Fodeh et al. BIBREF4 incorporate semantic knowledge from an ontology into text clustering. However, these works need solid NLP knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another direction is to map the original features into reduced space, such as Latent Semantic Analysis (LSA) BIBREF17 , Laplacian Eigenmaps (LE) BIBREF18 , and Locality Preserving Indexing (LPI) BIBREF19 . Even some researchers explored some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Moreover, some studies even focus the above both two streams. For example, Tang et al. BIBREF20 proposed a novel framework which enrich the text features by employing machine translation and reduce the original features simultaneously through matrix factorization techniques. Despite the above clustering methods can alleviate sparseness of short text representation to some extent, most of them ignore word order in the text and belong to shallow structures which can not fully capture accurate semantic similarities. Deep Neural Networks Recently, there is a revival of interest in DNN and many researchers have concentrated on using Deep Learning to learn features. Hinton and Salakhutdinov BIBREF21 use DAE to learn text representation. During the fine-tuning procedure, they use backpropagation to find codes that are good at reconstructing the word-count vector. More recently, researchers propose to use external corpus to learn a distributed representation for each word, called word embedding BIBREF22 , to improve DNN performance on NLP tasks. The Skip-gram and continuous bag-of-words models of Word2vec BIBREF23 propose a simple single-layer architecture based on the inner product between two word vectors, and Pennington et al. BIBREF24 introduce a new model for word representation, called GloVe, which captures the global corpus statistics. In order to learn the compact representation vectors of sentences, Le and Mikolov BIBREF25 directly extend the previous Word2vec BIBREF23 by predicting words in the sentence, which is named Paragraph Vector (Para2vec). Para2vec is still a shallow window-based method and need a larger corpus to yield better performance. More neural networks utilize word embedding to capture true meaningful syntactic and semantic regularities, such as RecNN BIBREF6 , BIBREF7 and RNN BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the layer computed at the last word to represent the text, is a biased model. Recently, Long Short-Term Memory (LSTM) BIBREF26 and Gated Recurrent Unit (GRU) BIBREF27 , as sophisticated recurrent hidden units of RNN, has presented its advantages in many sequence generation problem, such as machine translation BIBREF28 , speech recognition BIBREF29 , and text conversation BIBREF30 . While, CNN is better to learn non-biased implicit features which has been successfully exploited for many supervised NLP learning tasks as described in Section SECREF1 , and various CNN based variants are proposed in the recent works, such as Dynamic Convolutional Neural Network (DCNN) BIBREF10 , Gated Recursive Convolutional Neural Network (grConv) BIBREF31 and Self-Adaptive Hierarchical Sentence model (AdaSent) BIBREF32 . In the past few days, Visin et al. BIBREF33 have attempted to replace convolutional layer in CNN to learn non-biased features for object recognition with four RNNs, called ReNet, that sweep over lower-layer features in different directions: (1) bottom to top, (2) top to bottom, (3) left to right and (4) right to left. However, ReNet does not outperform state-of-the-art convolutional neural networks on any of the three benchmark datasets, and it is also a supervised learning model for classification. Inspired by Skip-gram of word2vec BIBREF34 , BIBREF23 , Skip-thought model BIBREF35 describe an approach for unsupervised learning of a generic, distributed sentence encoder. Similar as Skip-gram model, Skip-thought model trains an encoder-decoder model that tries to reconstruct the surrounding sentences of an encoded sentence and released an off-the-shelf encoder to extract sentence representation. Even some researchers introduce continuous Skip-gram and negative sampling to CNN for learning visual representation in an unsupervised manner BIBREF36 . This paper, from a new perspective, puts forward a general self-taught CNN framework which can flexibly couple various semantic features and achieve a good performance on one unsupervised learning task, short text clustering. Methodology Assume that we are given a dataset of INLINEFORM0 training texts denoted as: INLINEFORM1 , where INLINEFORM2 is the dimensionality of the original BoW representation. Denote its tag set as INLINEFORM3 and the pre-trained word embedding set as INLINEFORM4 , where INLINEFORM5 is the dimensionality of word vectors and INLINEFORM6 is the vocabulary size. In order to learn the INLINEFORM7 -dimensional deep feature representation INLINEFORM8 from CNN in an unsupervised manner, some unsupervised dimensionality reduction methods INLINEFORM9 are employed to guide the learning of CNN model. Our goal is to cluster these texts INLINEFORM10 into clusters INLINEFORM11 based on the learned deep feature representation while preserving the semantic consistency. As depicted in Figure FIGREF5 , the proposed framework consist of three components, deep convolutional neural network (CNN), unsupervised dimensionality reduction function and K-means module. In the rest sections, we first present the first two components respectively, and then give the trainable parameters and the objective function to learn the deep feature representation. Finally, the last section describe how to perform clustering on the learned features. Deep Convolutional Neural Networks In this section, we briefly review one popular deep convolutional neural network, Dynamic Convolutional Neural Network (DCNN) BIBREF10 as an instance of CNN in the following sections, which as the foundation of our proposed method has been successfully proposed for the completely supervised learning task, text classification. Taking a neural network with two convolutional layers in Figure FIGREF9 as an example, the network transforms raw input text to a powerful representation. Particularly, each raw text vector INLINEFORM0 is projected into a matrix representation INLINEFORM1 by looking up a word embedding INLINEFORM2 , where INLINEFORM3 is the length of one text. We also let INLINEFORM4 and INLINEFORM5 denote the weights of the neural networks. The network defines a transformation INLINEFORM6 INLINEFORM7 which transforms an input raw text INLINEFORM8 to a INLINEFORM9 -dimensional deep representation INLINEFORM10 . There are three basic operations described as follows: Wide one-dimensional convolution This operation INLINEFORM0 is applied to an individual row of the sentence matrix INLINEFORM1 , and yields a resulting matrix INLINEFORM2 , where INLINEFORM3 is the width of convolutional filter. Folding In this operation, every two rows in a feature map are simply summed component-wisely. For a map of INLINEFORM0 rows, folding returns a map of INLINEFORM1 rows, thus halving the size of the representation and yielding a matrix feature INLINEFORM2 . Note that folding operation does not introduce any additional parameters. Dynamic INLINEFORM0 -max pooling Assuming the pooling parameter as INLINEFORM1 , INLINEFORM2 -max pooling selects the sub-matrix INLINEFORM3 of the INLINEFORM4 highest values in each row of the matrix INLINEFORM5 . For dynamic INLINEFORM6 -max pooling, the pooling parameter INLINEFORM7 is dynamically selected in order to allow for a smooth extraction of higher-order and longer-range features BIBREF10 . Given a fixed pooling parameter INLINEFORM8 for the topmost convolutional layer, the parameter INLINEFORM9 of INLINEFORM10 -max pooling in the INLINEFORM11 -th convolutional layer can be computed as follows: DISPLAYFORM0 where INLINEFORM0 is the total number of convolutional layers in the network. Unsupervised Dimensionality Reduction As described in Figure FIGREF5 , the dimensionality reduction function is defined as follows: DISPLAYFORM0 where, INLINEFORM0 are the INLINEFORM1 -dimensional reduced latent space representations. Here, we take four popular dimensionality reduction methods as examples in our framework. Average Embedding (AE): This method directly averages the word embeddings which are respectively weighted with TF and TF-IDF. Huang et al. BIBREF37 used this strategy as the global context in their task, and Socher et al. BIBREF7 and Lai et al. BIBREF9 used this method for text classification. The weighted average of all word vectors in one text can be computed as follows: DISPLAYFORM0 where INLINEFORM0 can be any weighting function that captures the importance of word INLINEFORM1 in the text INLINEFORM2 . Latent Semantic Analysis (LSA): LSA BIBREF17 is the most popular global matrix factorization method, which applies a dimension reducing linear projection, Singular Value Decomposition (SVD), of the corresponding term/document matrix. Suppose the rank of INLINEFORM0 is INLINEFORM1 , LSA decompose INLINEFORM2 into the product of three other matrices: DISPLAYFORM0 where INLINEFORM0 and INLINEFORM1 are the singular values of INLINEFORM2 , INLINEFORM3 is a set of left singular vectors and INLINEFORM4 is a set of right singular vectors. LSA uses the top INLINEFORM5 vectors in INLINEFORM6 as the transformation matrix to embed the original text features into a INLINEFORM7 -dimensional subspace INLINEFORM8 BIBREF17 . Laplacian Eigenmaps (LE): The top eigenvectors of graph Laplacian, defined on the similarity matrix of texts, are used in the method, which can discover the manifold structure of the text space BIBREF18 . In order to avoid storing the dense similarity matrix, many approximation techniques are proposed to reduce the memory usage and computational complexity for LE. There are two representative approximation methods, sparse similarity matrix and Nystr INLINEFORM0 m approximation. Following previous studies BIBREF38 , BIBREF13 , we select the former technique to construct the INLINEFORM1 local similarity matrix INLINEFORM2 by using heat kernel as follows: DISPLAYFORM0 where, INLINEFORM0 is a tuning parameter (default is 1) and INLINEFORM1 represents the set of INLINEFORM2 -nearest-neighbors of INLINEFORM3 . By introducing a diagonal INLINEFORM4 matrix INLINEFORM5 whose entries are given by INLINEFORM6 , the graph Laplacian INLINEFORM7 can be computed by ( INLINEFORM8 ). The optimal INLINEFORM9 real-valued matrix INLINEFORM10 can be obtained by solving the following objective function: DISPLAYFORM0 where INLINEFORM0 is the trace function, INLINEFORM1 requires the different dimensions to be uncorrelated, and INLINEFORM2 requires each dimension to achieve equal probability as positive or negative). Locality Preserving Indexing (LPI): This method extends LE to deal with unseen texts by approximating the linear function INLINEFORM0 BIBREF13 , and the subspace vectors are obtained by finding the optimal linear approximations to the eigenfunctions of the Laplace Beltrami operator on the Riemannian manifold BIBREF19 . Similar as LE, we first construct the local similarity matrix INLINEFORM1 , then the graph Laplacian INLINEFORM2 can be computed by ( INLINEFORM3 ), where INLINEFORM4 measures the local density around INLINEFORM5 and is equal to INLINEFORM6 . Compute the eigenvectors INLINEFORM7 and eigenvalues INLINEFORM8 of the following generalized eigen-problem: DISPLAYFORM0 The mapping function INLINEFORM0 can be obtained and applied to the unseen data BIBREF38 . All of the above methods claim a better performance in capturing semantic similarity between texts in the reduced latent space representation INLINEFORM0 than in the original representation INLINEFORM1 , while the performance of short text clustering can be further enhanced with the help of our framework, self-taught CNN. Learning The last layer of CNN is an output layer as follows: DISPLAYFORM0 where, INLINEFORM0 is the deep feature representation, INLINEFORM1 is the output vector and INLINEFORM2 is weight matrix. In order to incorporate the latent semantic features INLINEFORM0 , we first binary the real-valued vectors INLINEFORM1 to the binary codes INLINEFORM2 by setting the threshold to be the media vector INLINEFORM3 . Then, the output vector INLINEFORM4 is used to fit the binary codes INLINEFORM5 via INLINEFORM6 logistic operations as follows: DISPLAYFORM0 All parameters to be trained are defined as INLINEFORM0 . DISPLAYFORM0 Given the training text collection INLINEFORM0 , and the pre-trained binary codes INLINEFORM1 , the log likelihood of the parameters can be written down as follows: DISPLAYFORM0 Following the previous work BIBREF10 , we train the network with mini-batches by back-propagation and perform the gradient-based optimization using the Adagrad update rule BIBREF39 . For regularization, we employ dropout with 50% rate to the penultimate layer BIBREF10 , BIBREF40 . K-means for Clustering With the given short texts, we first utilize the trained deep neural network to obtain the semantic representations INLINEFORM0 , and then employ traditional K-means algorithm to perform clustering. Datasets We test our proposed approach on three public short text datasets. The summary statistics and semantic topics of these datasets are described in Table TABREF24 and Table TABREF25 . SearchSnippets. This dataset was selected from the results of web search transaction using predefined phrases of 8 different domains by Phan et al. BIBREF41 . StackOverflow. We use the challenge data published in Kaggle.com. The raw dataset consists 3,370,528 samples through July 31st, 2012 to August 14, 2012. In our experiments, we randomly select 20,000 question titles from 20 different tags as in Table TABREF25 . Biomedical. We use the challenge data published in BioASQ's official website. In our experiments, we randomly select 20, 000 paper titles from 20 different MeSH major topics as in Table TABREF25 . As described in Table TABREF24 , the max length of selected paper titles is 53. For these datasets, we randomly select 10% of data as the development set. Since SearchSnippets has been pre-processed by Phan et al. BIBREF41 , we do not further process this dataset. In StackOverflow, texts contain lots of computer terminology, and symbols and capital letters are meaningful, thus we do not do any pre-processed procedures. For Biomedical, we remove the symbols and convert letters into lower case. Pre-trained Word Vectors We use the publicly available word2vec tool to train word embeddings, and the most parameters are set as same as Mikolov et al. BIBREF23 to train word vectors on Google News setting, except of vector dimensionality using 48 and minimize count using 5. For SearchSnippets, we train word vectors on Wikipedia dumps. For StackOverflow, we train word vectors on the whole corpus of the StackOverflow dataset described above which includes the question titles and post contents. For Biomedical, we train word vectors on all titles and abstracts of 2014 training articles. The coverage of these learned vectors on three datasets are listed in Table TABREF32 , and the words not present in the set of pre-trained words are initialized randomly. Comparisons In our experiment, some widely used text clustering methods are compared with our approach. Besides K-means, Skip-thought Vectors, Recursive Neural Network and Paragraph Vector based clustering methods, four baseline clustering methods are directly based on the popular unsupervised dimensionality reduction methods as described in Section SECREF11 . We further compare our approach with some other non-biased neural networks, such as bidirectional RNN. More details are listed as follows: K-means K-means BIBREF42 on original keyword features which are respectively weighted with term frequency (TF) and term frequency-inverse document frequency (TF-IDF). Skip-thought Vectors (SkipVec) This baseline BIBREF35 gives an off-the-shelf encoder to produce highly generic sentence representations. The encoder is trained using a large collection of novels and provides three encoder modes, that are unidirectional encoder (SkipVec (Uni)) with 2,400 dimensions, bidirectional encoder (SkipVec (Bi)) with 2,400 dimensions and combined encoder (SkipVec (Combine)) with SkipVec (Uni) and SkipVec (Bi) of 2,400 dimensions each. K-means is employed on the these vector representations respectively. Recursive Neural Network (RecNN) In BIBREF6 , the tree structure is firstly greedy approximated via unsupervised recursive autoencoder. Then, semi-supervised recursive autoencoders are used to capture the semantics of texts based on the predicted structure. In order to make this recursive-based method completely unsupervised, we remove the cross-entropy error in the second phrase to learn vector representation and subsequently employ K-means on the learned vectors of the top tree node and the average of all vectors in the tree. Paragraph Vector (Para2vec) K-means on the fixed size feature vectors generated by Paragraph Vector (Para2vec) BIBREF25 which is an unsupervised method to learn distributed representation of words and paragraphs. In our experiments, we use the open source software released by Mesnil et al. BIBREF43 . Average Embedding (AE) K-means on the weighted average vectors of the word embeddings which are respectively weighted with TF and TF-IDF. The dimension of average vectors is equal to and decided by the dimension of word vectors used in our experiments. Latent Semantic Analysis (LSA) K-means on the reduced subspace vectors generated by Singular Value Decomposition (SVD) method. The dimension of subspace is default set to the number of clusters, we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 10 on SearchSnippets, 20 on StackOverflow and 20 on Biomedical in our experiments. Laplacian Eigenmaps (LE) This baseline, using Laplacian Eigenmaps and subsequently employing K-means algorithm, is well known as spectral clustering BIBREF44 . The dimension of subspace is default set to the number of clusters BIBREF18 , BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 70 on StackOverflow and 30 on Biomedical in our experiments. Locality Preserving Indexing (LPI) This baseline, projecting the texts into a lower dimensional semantic space, can discover both the geometric and discriminating structures of the original feature space BIBREF38 . The dimension of subspace is default set to the number of clusters BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 80 on StackOverflow and 30 on Biomedical in our experiments. bidirectional RNN (bi-RNN) We replace the CNN model in our framework as in Figure FIGREF5 with some bi-RNN models. Particularly, LSTM and GRU units are used in the experiments. In order to generate the fixed-length document representation from the variable-length vector sequences, for both bi-LSTM and bi-GRU based clustering methods, we further utilize three pooling methods: last pooling (using the last hidden state), mean pooling and element-wise max pooling. These pooling methods are respectively used in the previous works BIBREF45 , BIBREF27 , BIBREF46 and BIBREF9 . For regularization, the training gradients of all parameters with an INLINEFORM0 2 norm larger than 40 are clipped to 40, as the previous work BIBREF47 . Evaluation Metrics The clustering performance is evaluated by comparing the clustering results of texts with the tags/labels provided by the text corpus. Two metrics, the accuracy (ACC) and the normalized mutual information metric (NMI), are used to measure the clustering performance BIBREF38 , BIBREF48 . Given a text INLINEFORM0 , let INLINEFORM1 and INLINEFORM2 be the obtained cluster label and the label provided by the corpus, respectively. Accuracy is defined as: DISPLAYFORM0 where, INLINEFORM0 is the total number of texts, INLINEFORM1 is the indicator function that equals one if INLINEFORM2 and equals zero otherwise, and INLINEFORM3 is the permutation mapping function that maps each cluster label INLINEFORM4 to the equivalent label from the text data by Hungarian algorithm BIBREF49 . Normalized mutual information BIBREF50 between tag/label set INLINEFORM0 and cluster set INLINEFORM1 is a popular metric used for evaluating clustering tasks. It is defined as follows: DISPLAYFORM0 where, INLINEFORM0 is the mutual information between INLINEFORM1 and INLINEFORM2 , INLINEFORM3 is entropy and the denominator INLINEFORM4 is used for normalizing the mutual information to be in the range of [0, 1]. Hyperparameter Settings The most of parameters are set uniformly for these datasets. Following previous study BIBREF38 , the number of nearest neighbors in Eqn. ( EQREF15 ) is fixed to 15 when constructing the graph structures for LE and LPI. For CNN model, the networks has two convolutional layers. The widths of the convolutional filters are both 3. The value of INLINEFORM0 for the top INLINEFORM1 -max pooling in Eqn. ( EQREF10 ) is 5. The number of feature maps at the first convolutional layer is 12, and 8 feature maps at the second convolutional layer. Both those two convolutional layers are followed by a folding layer. We further set the dimension of word embeddings INLINEFORM2 as 48. Finally, the dimension of the deep feature representation INLINEFORM3 is fixed to 480. Moreover, we set the learning rate INLINEFORM4 as 0.01 and the mini-batch training size as 200. The output size INLINEFORM5 in Eqn. ( EQREF19 ) is set same as the best dimensions of subspace in the baseline method, as described in Section SECREF37 . For initial centroids have significant impact on clustering results when utilizing the K-means algorithms, we repeat K-means for multiple times with random initial centroids (specifically, 100 times for statistical significance) as Huang BIBREF48 . The all subspace vectors are normalized to 1 before applying K-means and the final results reported are the average of 5 trials with all clustering methods on three text datasets. Results and Analysis In Table TABREF43 and Table TABREF44 , we report the ACC and NMI performance of our proposed approaches and four baseline methods, K-means, SkipVec, RecNN and Para2vec based clustering methods. Intuitively, we get a general observation that (1) BoW based approaches, including K-means (TF) and K-means (TF-IDF), and SkipVec based approaches perform not well; (2) RecNN based approaches, both RecNN (Ave.) and RecNN (Top+Ave.), do better; (3) Para2vec makes a comparable performance with the most baselines; and (4) the evaluation clearly demonstrate the superiority of our proposed methods STC INLINEFORM0 . It is an expected results. For SkipVec based approaches, the off-the-shelf encoders are trained on the BookCorpus datasets BIBREF51 , and then applied to our datasets to extract the sentence representations. The SkipVec encoders can produce generic sentence representations but may not perform well for specific datasets, in our experiments, StackOverflow and Biomedical datasets consist of many computer terms and medical terms, such as “ASP.NET”, “XML”, “C#”, “serum” and “glycolytic”. When we take a more careful look, we find that RecNN (Top) does poorly, even worse than K-means (TF-IDF). The reason maybe that although recursive neural models introduce tree structure to capture compositional semantics, the vector of the top node mainly captures a biased semantic while the average of all vectors in the tree nodes, such as RecNN (Ave.), can be better to represent sentence level semantic. And we also get another observation that, although our proposed STC INLINEFORM1 -LE and STC INLINEFORM2 -LPI outperform both BoW based and RecNN based approaches across all three datasets, STC INLINEFORM3 -AE and STC INLINEFORM4 -LSA do just exhibit some similar performances as RecNN (Ave.) and RecNN (Top+Ave.) do in the datasets of StackOverflow and Biomedical. We further replace the CNN model in our framework as in Figure FIGREF5 with some other non-biased models, such as bi-LSTM and bi-GRU, and report the results in Table TABREF46 and Table TABREF47 . As an instance, the binary codes generated from LPI are used to guide the learning of bi-LSTM/bi-GRU models. From the results, we can see that bi-GRU and bi-LSTM based clustering methods do equally well, no clear winner, and both achieve great enhancements compared with LPI (best). Compared with these bi-LSTM/bi-GRU based models, the evaluation results still demonstrate the superiority of our approach methods, CNN based clustering model, in the most cases. As the results reported by Visin et al. BIBREF33 , despite bi-directional or multi-directional RNN models perform a good non-biased feature extraction, they yet do not outperform state-of-the-art CNN on some tasks. In order to make clear what factors make our proposed method work, we report the bar chart results of ACC and MNI of our proposed methods and the corresponding baseline methods in Figure FIGREF49 and Figure FIGREF53 . It is clear that, although AE and LSA does well or even better than LE and LPI, especially in dataset of both StackOverflow and Biomedical, STC INLINEFORM0 -LE and STC INLINEFORM1 -LPI achieve a much larger performance enhancements than STC INLINEFORM2 -AE and STC INLINEFORM3 -LSA do. The possible reason is that the information the pseudo supervision used to guide the learning of CNN model that make difference. Especially, for AE case, the input features fed into CNN model and the pseudo supervision employed to guide the learning of CNN model are all come from word embeddings. There are no different semantic features to be used into our proposed method, thus the performance enhancements are limited in STC INLINEFORM4 -AE. For LSA case, as we known, LSA is to make matrix factorization to find the best subspace approximation of the original feature space to minimize the global reconstruction error. And as BIBREF24 , BIBREF52 recently point out that word embeddings trained with word2vec or some variances, is essentially to do an operation of matrix factorization. Therefore, the information between input and the pseudo supervision in CNN is not departed very largely from each other, and the performance enhancements of STC INLINEFORM5 -AE is also not quite satisfactory. For LE and LPI case, as we known that LE extracts the manifold structure of the original feature space, and LPI extracts both geometric and discriminating structure of the original feature space BIBREF38 . We guess that our approach STC INLINEFORM6 -LE and STC INLINEFORM7 -LPI achieve enhancements compared with both LE and LPI by a large margin, because both of LE and LPI get useful semantic features, and these features are also different from word embeddings used as input of CNN. From this view, we say that our proposed STC has potential to behave more effective when the pseudo supervision is able to get semantic meaningful features, which is different enough from the input of CNN. Furthermore, from the results of K-means and AE in Table TABREF43 - TABREF44 and Figure FIGREF49 - FIGREF53 , we note that TF-IDF weighting gives a more remarkable improvement for K-means, while TF weighting works better than TF-IDF weighting for Average Embedding. Maybe the reason is that pre-trained word embeddings encode some useful information from external corpus and are able to get even better results without TF-IDF weighting. Meanwhile, we find that LE get quite unusual good performance than LPI, LSA and AE in SearchSnippets dataset, which is not found in the other two datasets. To get clear about this, and also to make a much better demonstration about our proposed approaches and other baselines, we further report 2-dimensional text embeddings on SearchSnippets in Figure FIGREF58 , using t-SNE BIBREF53 to get distributed stochastic neighbor embedding of the feature representations used in the clustering methods. We can see that the results of from AE and LSA seem to be fairly good or even better than the ones from LE and LPI, which is not the same as the results from ACC and NMI in Figure FIGREF49 - FIGREF53 . Meanwhile, RecNN (Ave.) performs better than BoW (both TF and TF-IDF) while RecNN (Top) does not, which is the same as the results from ACC and NMI in Table TABREF43 and Table TABREF44 . Then we guess that both ”the same as” and ”not the same as” above, is just a good example to illustrate that visualization tool, such as t-SNE, get some useful information for measuring results, which is different from the ones of ACC and NMI. Moreover, from this complementary view of t-SNE, we can see that our STC INLINEFORM0 -AE, STC INLINEFORM1 -LSA, STC INLINEFORM2 -LE, and STC INLINEFORM3 -LPI show more clear-cut margins among different semantic topics (that is, tags/labels), compared with AE, LSA, LE and LPI, respectively, as well as compared with both baselines, BoW and RecNN based ones. From all these results, with three measures of ACC, NMI and t-SNE under three datasets, we can get a solid conclusion that our proposed approaches is an effective approaches to get useful semantic features for short text clustering. Conclusions With the emergence of social media, short text clustering has become an increasing important task. This paper explores a new perspective to cluster short texts based on deep feature representation learned from the proposed self-taught convolutional neural networks. Our framework can be successfully accomplished without using any external tags/labels and complicated NLP pre-processing, and and our approach is a flexible framework, in which the traditional dimension reduction approaches could be used to get performance enhancement. Our extensive experimental study on three short text datasets shows that our approach can achieve a significantly better performance. In the future, how to select and incorporate more effective semantic features into the proposed framework would call for more research. Acknowledgments We would like to thank reviewers for their comments, and acknowledge Kaggle and BioASQ for making the datasets available. This work is supported by the National Natural Science Foundation of China (No. 61602479, No. 61303172, No. 61403385) and the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB02070005).

K-means, Skip-thought Vectors, Recursive Neural Network and Paragraph Vector based clustering methods

785c054f6ea04701f4ab260d064af7d124260ccc

785c054f6ea04701f4ab260d064af7d124260ccc_0

Q: What datasets did they use? Text: Introduction Short text clustering is of great importance due to its various applications, such as user profiling BIBREF0 and recommendation BIBREF1 , for nowaday's social media dataset emerged day by day. However, short text clustering has the data sparsity problem and most words only occur once in each short text BIBREF2 . As a result, the Term Frequency-Inverse Document Frequency (TF-IDF) measure cannot work well in short text setting. In order to address this problem, some researchers work on expanding and enriching the context of data from Wikipedia BIBREF3 or an ontology BIBREF4 . However, these methods involve solid Natural Language Processing (NLP) knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another way to overcome these issues is to explore some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Yet how to design an effective model is an open question, and most of these methods directly trained based on Bag-of-Words (BoW) are shallow structures which cannot preserve the accurate semantic similarities. Recently, with the help of word embedding, neural networks demonstrate their great performance in terms of constructing text representation, such as Recursive Neural Network (RecNN) BIBREF6 , BIBREF7 and Recurrent Neural Network (RNN) BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the hidden layer computed at the last word to represent the text, is a biased model where later words are more dominant than earlier words BIBREF9 . Whereas for the non-biased models, the learned representation of one text can be extracted from all the words in the text with non-dominant learned weights. More recently, Convolution Neural Network (CNN), as the most popular non-biased model and applying convolutional filters to capture local features, has achieved a better performance in many NLP applications, such as sentence modeling BIBREF10 , relation classification BIBREF11 , and other traditional NLP tasks BIBREF12 . Most of the previous works focus CNN on solving supervised NLP tasks, while in this paper we aim to explore the power of CNN on one unsupervised NLP task, short text clustering. We systematically introduce a simple yet surprisingly powerful Self-Taught Convolutional neural network framework for Short Text Clustering, called STC INLINEFORM0 . An overall architecture of our proposed approach is illustrated in Figure FIGREF5 . We, inspired by BIBREF13 , BIBREF14 , utilize a self-taught learning framework into our task. In particular, the original raw text features are first embedded into compact binary codes INLINEFORM1 with the help of one traditional unsupervised dimensionality reduction function. Then text matrix INLINEFORM2 projected from word embeddings are fed into CNN model to learn the deep feature representation INLINEFORM3 and the output units are used to fit the pre-trained binary codes INLINEFORM4 . After obtaining the learned features, K-means algorithm is employed on them to cluster texts into clusters INLINEFORM5 . Obviously, we call our approach “self-taught” because the CNN model is learnt from the pseudo labels generated from the previous stage, which is quite different from the term “self-taught” in BIBREF15 . Our main contributions can be summarized as follows: This work is an extension of our conference paper BIBREF16 , and they differ in the following aspects. First, we put forward a general a self-taught CNN framework in this paper which can flexibly couple various semantic features, whereas the conference version can be seen as a specific example of this work. Second, in this paper we use a new short text dataset, Biomedical, in the experiment to verify the effectiveness of our approach. Third, we put much effort on studying the influence of various different semantic features integrated in our self-taught CNN framework, which is not involved in the conference paper. For the purpose of reproducibility, we make the datasets and software used in our experiments publicly available at the website. The remainder of this paper is organized as follows: In Section SECREF2 , we first briefly survey several related works. In Section SECREF3 , we describe the proposed approach STC INLINEFORM0 and implementation details. Experimental results and analyses are presented in Section SECREF4 . Finally, conclusions are given in the last Section. Related Work In this section, we review the related work from the following two perspectives: short text clustering and deep neural networks. Short Text Clustering There have been several studies that attempted to overcome the sparseness of short text representation. One way is to expand and enrich the context of data. For example, Banerjee et al. BIBREF3 proposed a method of improving the accuracy of short text clustering by enriching their representation with additional features from Wikipedia, and Fodeh et al. BIBREF4 incorporate semantic knowledge from an ontology into text clustering. However, these works need solid NLP knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another direction is to map the original features into reduced space, such as Latent Semantic Analysis (LSA) BIBREF17 , Laplacian Eigenmaps (LE) BIBREF18 , and Locality Preserving Indexing (LPI) BIBREF19 . Even some researchers explored some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Moreover, some studies even focus the above both two streams. For example, Tang et al. BIBREF20 proposed a novel framework which enrich the text features by employing machine translation and reduce the original features simultaneously through matrix factorization techniques. Despite the above clustering methods can alleviate sparseness of short text representation to some extent, most of them ignore word order in the text and belong to shallow structures which can not fully capture accurate semantic similarities. Deep Neural Networks Recently, there is a revival of interest in DNN and many researchers have concentrated on using Deep Learning to learn features. Hinton and Salakhutdinov BIBREF21 use DAE to learn text representation. During the fine-tuning procedure, they use backpropagation to find codes that are good at reconstructing the word-count vector. More recently, researchers propose to use external corpus to learn a distributed representation for each word, called word embedding BIBREF22 , to improve DNN performance on NLP tasks. The Skip-gram and continuous bag-of-words models of Word2vec BIBREF23 propose a simple single-layer architecture based on the inner product between two word vectors, and Pennington et al. BIBREF24 introduce a new model for word representation, called GloVe, which captures the global corpus statistics. In order to learn the compact representation vectors of sentences, Le and Mikolov BIBREF25 directly extend the previous Word2vec BIBREF23 by predicting words in the sentence, which is named Paragraph Vector (Para2vec). Para2vec is still a shallow window-based method and need a larger corpus to yield better performance. More neural networks utilize word embedding to capture true meaningful syntactic and semantic regularities, such as RecNN BIBREF6 , BIBREF7 and RNN BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the layer computed at the last word to represent the text, is a biased model. Recently, Long Short-Term Memory (LSTM) BIBREF26 and Gated Recurrent Unit (GRU) BIBREF27 , as sophisticated recurrent hidden units of RNN, has presented its advantages in many sequence generation problem, such as machine translation BIBREF28 , speech recognition BIBREF29 , and text conversation BIBREF30 . While, CNN is better to learn non-biased implicit features which has been successfully exploited for many supervised NLP learning tasks as described in Section SECREF1 , and various CNN based variants are proposed in the recent works, such as Dynamic Convolutional Neural Network (DCNN) BIBREF10 , Gated Recursive Convolutional Neural Network (grConv) BIBREF31 and Self-Adaptive Hierarchical Sentence model (AdaSent) BIBREF32 . In the past few days, Visin et al. BIBREF33 have attempted to replace convolutional layer in CNN to learn non-biased features for object recognition with four RNNs, called ReNet, that sweep over lower-layer features in different directions: (1) bottom to top, (2) top to bottom, (3) left to right and (4) right to left. However, ReNet does not outperform state-of-the-art convolutional neural networks on any of the three benchmark datasets, and it is also a supervised learning model for classification. Inspired by Skip-gram of word2vec BIBREF34 , BIBREF23 , Skip-thought model BIBREF35 describe an approach for unsupervised learning of a generic, distributed sentence encoder. Similar as Skip-gram model, Skip-thought model trains an encoder-decoder model that tries to reconstruct the surrounding sentences of an encoded sentence and released an off-the-shelf encoder to extract sentence representation. Even some researchers introduce continuous Skip-gram and negative sampling to CNN for learning visual representation in an unsupervised manner BIBREF36 . This paper, from a new perspective, puts forward a general self-taught CNN framework which can flexibly couple various semantic features and achieve a good performance on one unsupervised learning task, short text clustering. Methodology Assume that we are given a dataset of INLINEFORM0 training texts denoted as: INLINEFORM1 , where INLINEFORM2 is the dimensionality of the original BoW representation. Denote its tag set as INLINEFORM3 and the pre-trained word embedding set as INLINEFORM4 , where INLINEFORM5 is the dimensionality of word vectors and INLINEFORM6 is the vocabulary size. In order to learn the INLINEFORM7 -dimensional deep feature representation INLINEFORM8 from CNN in an unsupervised manner, some unsupervised dimensionality reduction methods INLINEFORM9 are employed to guide the learning of CNN model. Our goal is to cluster these texts INLINEFORM10 into clusters INLINEFORM11 based on the learned deep feature representation while preserving the semantic consistency. As depicted in Figure FIGREF5 , the proposed framework consist of three components, deep convolutional neural network (CNN), unsupervised dimensionality reduction function and K-means module. In the rest sections, we first present the first two components respectively, and then give the trainable parameters and the objective function to learn the deep feature representation. Finally, the last section describe how to perform clustering on the learned features. Deep Convolutional Neural Networks In this section, we briefly review one popular deep convolutional neural network, Dynamic Convolutional Neural Network (DCNN) BIBREF10 as an instance of CNN in the following sections, which as the foundation of our proposed method has been successfully proposed for the completely supervised learning task, text classification. Taking a neural network with two convolutional layers in Figure FIGREF9 as an example, the network transforms raw input text to a powerful representation. Particularly, each raw text vector INLINEFORM0 is projected into a matrix representation INLINEFORM1 by looking up a word embedding INLINEFORM2 , where INLINEFORM3 is the length of one text. We also let INLINEFORM4 and INLINEFORM5 denote the weights of the neural networks. The network defines a transformation INLINEFORM6 INLINEFORM7 which transforms an input raw text INLINEFORM8 to a INLINEFORM9 -dimensional deep representation INLINEFORM10 . There are three basic operations described as follows: Wide one-dimensional convolution This operation INLINEFORM0 is applied to an individual row of the sentence matrix INLINEFORM1 , and yields a resulting matrix INLINEFORM2 , where INLINEFORM3 is the width of convolutional filter. Folding In this operation, every two rows in a feature map are simply summed component-wisely. For a map of INLINEFORM0 rows, folding returns a map of INLINEFORM1 rows, thus halving the size of the representation and yielding a matrix feature INLINEFORM2 . Note that folding operation does not introduce any additional parameters. Dynamic INLINEFORM0 -max pooling Assuming the pooling parameter as INLINEFORM1 , INLINEFORM2 -max pooling selects the sub-matrix INLINEFORM3 of the INLINEFORM4 highest values in each row of the matrix INLINEFORM5 . For dynamic INLINEFORM6 -max pooling, the pooling parameter INLINEFORM7 is dynamically selected in order to allow for a smooth extraction of higher-order and longer-range features BIBREF10 . Given a fixed pooling parameter INLINEFORM8 for the topmost convolutional layer, the parameter INLINEFORM9 of INLINEFORM10 -max pooling in the INLINEFORM11 -th convolutional layer can be computed as follows: DISPLAYFORM0 where INLINEFORM0 is the total number of convolutional layers in the network. Unsupervised Dimensionality Reduction As described in Figure FIGREF5 , the dimensionality reduction function is defined as follows: DISPLAYFORM0 where, INLINEFORM0 are the INLINEFORM1 -dimensional reduced latent space representations. Here, we take four popular dimensionality reduction methods as examples in our framework. Average Embedding (AE): This method directly averages the word embeddings which are respectively weighted with TF and TF-IDF. Huang et al. BIBREF37 used this strategy as the global context in their task, and Socher et al. BIBREF7 and Lai et al. BIBREF9 used this method for text classification. The weighted average of all word vectors in one text can be computed as follows: DISPLAYFORM0 where INLINEFORM0 can be any weighting function that captures the importance of word INLINEFORM1 in the text INLINEFORM2 . Latent Semantic Analysis (LSA): LSA BIBREF17 is the most popular global matrix factorization method, which applies a dimension reducing linear projection, Singular Value Decomposition (SVD), of the corresponding term/document matrix. Suppose the rank of INLINEFORM0 is INLINEFORM1 , LSA decompose INLINEFORM2 into the product of three other matrices: DISPLAYFORM0 where INLINEFORM0 and INLINEFORM1 are the singular values of INLINEFORM2 , INLINEFORM3 is a set of left singular vectors and INLINEFORM4 is a set of right singular vectors. LSA uses the top INLINEFORM5 vectors in INLINEFORM6 as the transformation matrix to embed the original text features into a INLINEFORM7 -dimensional subspace INLINEFORM8 BIBREF17 . Laplacian Eigenmaps (LE): The top eigenvectors of graph Laplacian, defined on the similarity matrix of texts, are used in the method, which can discover the manifold structure of the text space BIBREF18 . In order to avoid storing the dense similarity matrix, many approximation techniques are proposed to reduce the memory usage and computational complexity for LE. There are two representative approximation methods, sparse similarity matrix and Nystr INLINEFORM0 m approximation. Following previous studies BIBREF38 , BIBREF13 , we select the former technique to construct the INLINEFORM1 local similarity matrix INLINEFORM2 by using heat kernel as follows: DISPLAYFORM0 where, INLINEFORM0 is a tuning parameter (default is 1) and INLINEFORM1 represents the set of INLINEFORM2 -nearest-neighbors of INLINEFORM3 . By introducing a diagonal INLINEFORM4 matrix INLINEFORM5 whose entries are given by INLINEFORM6 , the graph Laplacian INLINEFORM7 can be computed by ( INLINEFORM8 ). The optimal INLINEFORM9 real-valued matrix INLINEFORM10 can be obtained by solving the following objective function: DISPLAYFORM0 where INLINEFORM0 is the trace function, INLINEFORM1 requires the different dimensions to be uncorrelated, and INLINEFORM2 requires each dimension to achieve equal probability as positive or negative). Locality Preserving Indexing (LPI): This method extends LE to deal with unseen texts by approximating the linear function INLINEFORM0 BIBREF13 , and the subspace vectors are obtained by finding the optimal linear approximations to the eigenfunctions of the Laplace Beltrami operator on the Riemannian manifold BIBREF19 . Similar as LE, we first construct the local similarity matrix INLINEFORM1 , then the graph Laplacian INLINEFORM2 can be computed by ( INLINEFORM3 ), where INLINEFORM4 measures the local density around INLINEFORM5 and is equal to INLINEFORM6 . Compute the eigenvectors INLINEFORM7 and eigenvalues INLINEFORM8 of the following generalized eigen-problem: DISPLAYFORM0 The mapping function INLINEFORM0 can be obtained and applied to the unseen data BIBREF38 . All of the above methods claim a better performance in capturing semantic similarity between texts in the reduced latent space representation INLINEFORM0 than in the original representation INLINEFORM1 , while the performance of short text clustering can be further enhanced with the help of our framework, self-taught CNN. Learning The last layer of CNN is an output layer as follows: DISPLAYFORM0 where, INLINEFORM0 is the deep feature representation, INLINEFORM1 is the output vector and INLINEFORM2 is weight matrix. In order to incorporate the latent semantic features INLINEFORM0 , we first binary the real-valued vectors INLINEFORM1 to the binary codes INLINEFORM2 by setting the threshold to be the media vector INLINEFORM3 . Then, the output vector INLINEFORM4 is used to fit the binary codes INLINEFORM5 via INLINEFORM6 logistic operations as follows: DISPLAYFORM0 All parameters to be trained are defined as INLINEFORM0 . DISPLAYFORM0 Given the training text collection INLINEFORM0 , and the pre-trained binary codes INLINEFORM1 , the log likelihood of the parameters can be written down as follows: DISPLAYFORM0 Following the previous work BIBREF10 , we train the network with mini-batches by back-propagation and perform the gradient-based optimization using the Adagrad update rule BIBREF39 . For regularization, we employ dropout with 50% rate to the penultimate layer BIBREF10 , BIBREF40 . K-means for Clustering With the given short texts, we first utilize the trained deep neural network to obtain the semantic representations INLINEFORM0 , and then employ traditional K-means algorithm to perform clustering. Datasets We test our proposed approach on three public short text datasets. The summary statistics and semantic topics of these datasets are described in Table TABREF24 and Table TABREF25 . SearchSnippets. This dataset was selected from the results of web search transaction using predefined phrases of 8 different domains by Phan et al. BIBREF41 . StackOverflow. We use the challenge data published in Kaggle.com. The raw dataset consists 3,370,528 samples through July 31st, 2012 to August 14, 2012. In our experiments, we randomly select 20,000 question titles from 20 different tags as in Table TABREF25 . Biomedical. We use the challenge data published in BioASQ's official website. In our experiments, we randomly select 20, 000 paper titles from 20 different MeSH major topics as in Table TABREF25 . As described in Table TABREF24 , the max length of selected paper titles is 53. For these datasets, we randomly select 10% of data as the development set. Since SearchSnippets has been pre-processed by Phan et al. BIBREF41 , we do not further process this dataset. In StackOverflow, texts contain lots of computer terminology, and symbols and capital letters are meaningful, thus we do not do any pre-processed procedures. For Biomedical, we remove the symbols and convert letters into lower case. Pre-trained Word Vectors We use the publicly available word2vec tool to train word embeddings, and the most parameters are set as same as Mikolov et al. BIBREF23 to train word vectors on Google News setting, except of vector dimensionality using 48 and minimize count using 5. For SearchSnippets, we train word vectors on Wikipedia dumps. For StackOverflow, we train word vectors on the whole corpus of the StackOverflow dataset described above which includes the question titles and post contents. For Biomedical, we train word vectors on all titles and abstracts of 2014 training articles. The coverage of these learned vectors on three datasets are listed in Table TABREF32 , and the words not present in the set of pre-trained words are initialized randomly. Comparisons In our experiment, some widely used text clustering methods are compared with our approach. Besides K-means, Skip-thought Vectors, Recursive Neural Network and Paragraph Vector based clustering methods, four baseline clustering methods are directly based on the popular unsupervised dimensionality reduction methods as described in Section SECREF11 . We further compare our approach with some other non-biased neural networks, such as bidirectional RNN. More details are listed as follows: K-means K-means BIBREF42 on original keyword features which are respectively weighted with term frequency (TF) and term frequency-inverse document frequency (TF-IDF). Skip-thought Vectors (SkipVec) This baseline BIBREF35 gives an off-the-shelf encoder to produce highly generic sentence representations. The encoder is trained using a large collection of novels and provides three encoder modes, that are unidirectional encoder (SkipVec (Uni)) with 2,400 dimensions, bidirectional encoder (SkipVec (Bi)) with 2,400 dimensions and combined encoder (SkipVec (Combine)) with SkipVec (Uni) and SkipVec (Bi) of 2,400 dimensions each. K-means is employed on the these vector representations respectively. Recursive Neural Network (RecNN) In BIBREF6 , the tree structure is firstly greedy approximated via unsupervised recursive autoencoder. Then, semi-supervised recursive autoencoders are used to capture the semantics of texts based on the predicted structure. In order to make this recursive-based method completely unsupervised, we remove the cross-entropy error in the second phrase to learn vector representation and subsequently employ K-means on the learned vectors of the top tree node and the average of all vectors in the tree. Paragraph Vector (Para2vec) K-means on the fixed size feature vectors generated by Paragraph Vector (Para2vec) BIBREF25 which is an unsupervised method to learn distributed representation of words and paragraphs. In our experiments, we use the open source software released by Mesnil et al. BIBREF43 . Average Embedding (AE) K-means on the weighted average vectors of the word embeddings which are respectively weighted with TF and TF-IDF. The dimension of average vectors is equal to and decided by the dimension of word vectors used in our experiments. Latent Semantic Analysis (LSA) K-means on the reduced subspace vectors generated by Singular Value Decomposition (SVD) method. The dimension of subspace is default set to the number of clusters, we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 10 on SearchSnippets, 20 on StackOverflow and 20 on Biomedical in our experiments. Laplacian Eigenmaps (LE) This baseline, using Laplacian Eigenmaps and subsequently employing K-means algorithm, is well known as spectral clustering BIBREF44 . The dimension of subspace is default set to the number of clusters BIBREF18 , BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 70 on StackOverflow and 30 on Biomedical in our experiments. Locality Preserving Indexing (LPI) This baseline, projecting the texts into a lower dimensional semantic space, can discover both the geometric and discriminating structures of the original feature space BIBREF38 . The dimension of subspace is default set to the number of clusters BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 80 on StackOverflow and 30 on Biomedical in our experiments. bidirectional RNN (bi-RNN) We replace the CNN model in our framework as in Figure FIGREF5 with some bi-RNN models. Particularly, LSTM and GRU units are used in the experiments. In order to generate the fixed-length document representation from the variable-length vector sequences, for both bi-LSTM and bi-GRU based clustering methods, we further utilize three pooling methods: last pooling (using the last hidden state), mean pooling and element-wise max pooling. These pooling methods are respectively used in the previous works BIBREF45 , BIBREF27 , BIBREF46 and BIBREF9 . For regularization, the training gradients of all parameters with an INLINEFORM0 2 norm larger than 40 are clipped to 40, as the previous work BIBREF47 . Evaluation Metrics The clustering performance is evaluated by comparing the clustering results of texts with the tags/labels provided by the text corpus. Two metrics, the accuracy (ACC) and the normalized mutual information metric (NMI), are used to measure the clustering performance BIBREF38 , BIBREF48 . Given a text INLINEFORM0 , let INLINEFORM1 and INLINEFORM2 be the obtained cluster label and the label provided by the corpus, respectively. Accuracy is defined as: DISPLAYFORM0 where, INLINEFORM0 is the total number of texts, INLINEFORM1 is the indicator function that equals one if INLINEFORM2 and equals zero otherwise, and INLINEFORM3 is the permutation mapping function that maps each cluster label INLINEFORM4 to the equivalent label from the text data by Hungarian algorithm BIBREF49 . Normalized mutual information BIBREF50 between tag/label set INLINEFORM0 and cluster set INLINEFORM1 is a popular metric used for evaluating clustering tasks. It is defined as follows: DISPLAYFORM0 where, INLINEFORM0 is the mutual information between INLINEFORM1 and INLINEFORM2 , INLINEFORM3 is entropy and the denominator INLINEFORM4 is used for normalizing the mutual information to be in the range of [0, 1]. Hyperparameter Settings The most of parameters are set uniformly for these datasets. Following previous study BIBREF38 , the number of nearest neighbors in Eqn. ( EQREF15 ) is fixed to 15 when constructing the graph structures for LE and LPI. For CNN model, the networks has two convolutional layers. The widths of the convolutional filters are both 3. The value of INLINEFORM0 for the top INLINEFORM1 -max pooling in Eqn. ( EQREF10 ) is 5. The number of feature maps at the first convolutional layer is 12, and 8 feature maps at the second convolutional layer. Both those two convolutional layers are followed by a folding layer. We further set the dimension of word embeddings INLINEFORM2 as 48. Finally, the dimension of the deep feature representation INLINEFORM3 is fixed to 480. Moreover, we set the learning rate INLINEFORM4 as 0.01 and the mini-batch training size as 200. The output size INLINEFORM5 in Eqn. ( EQREF19 ) is set same as the best dimensions of subspace in the baseline method, as described in Section SECREF37 . For initial centroids have significant impact on clustering results when utilizing the K-means algorithms, we repeat K-means for multiple times with random initial centroids (specifically, 100 times for statistical significance) as Huang BIBREF48 . The all subspace vectors are normalized to 1 before applying K-means and the final results reported are the average of 5 trials with all clustering methods on three text datasets. Results and Analysis In Table TABREF43 and Table TABREF44 , we report the ACC and NMI performance of our proposed approaches and four baseline methods, K-means, SkipVec, RecNN and Para2vec based clustering methods. Intuitively, we get a general observation that (1) BoW based approaches, including K-means (TF) and K-means (TF-IDF), and SkipVec based approaches perform not well; (2) RecNN based approaches, both RecNN (Ave.) and RecNN (Top+Ave.), do better; (3) Para2vec makes a comparable performance with the most baselines; and (4) the evaluation clearly demonstrate the superiority of our proposed methods STC INLINEFORM0 . It is an expected results. For SkipVec based approaches, the off-the-shelf encoders are trained on the BookCorpus datasets BIBREF51 , and then applied to our datasets to extract the sentence representations. The SkipVec encoders can produce generic sentence representations but may not perform well for specific datasets, in our experiments, StackOverflow and Biomedical datasets consist of many computer terms and medical terms, such as “ASP.NET”, “XML”, “C#”, “serum” and “glycolytic”. When we take a more careful look, we find that RecNN (Top) does poorly, even worse than K-means (TF-IDF). The reason maybe that although recursive neural models introduce tree structure to capture compositional semantics, the vector of the top node mainly captures a biased semantic while the average of all vectors in the tree nodes, such as RecNN (Ave.), can be better to represent sentence level semantic. And we also get another observation that, although our proposed STC INLINEFORM1 -LE and STC INLINEFORM2 -LPI outperform both BoW based and RecNN based approaches across all three datasets, STC INLINEFORM3 -AE and STC INLINEFORM4 -LSA do just exhibit some similar performances as RecNN (Ave.) and RecNN (Top+Ave.) do in the datasets of StackOverflow and Biomedical. We further replace the CNN model in our framework as in Figure FIGREF5 with some other non-biased models, such as bi-LSTM and bi-GRU, and report the results in Table TABREF46 and Table TABREF47 . As an instance, the binary codes generated from LPI are used to guide the learning of bi-LSTM/bi-GRU models. From the results, we can see that bi-GRU and bi-LSTM based clustering methods do equally well, no clear winner, and both achieve great enhancements compared with LPI (best). Compared with these bi-LSTM/bi-GRU based models, the evaluation results still demonstrate the superiority of our approach methods, CNN based clustering model, in the most cases. As the results reported by Visin et al. BIBREF33 , despite bi-directional or multi-directional RNN models perform a good non-biased feature extraction, they yet do not outperform state-of-the-art CNN on some tasks. In order to make clear what factors make our proposed method work, we report the bar chart results of ACC and MNI of our proposed methods and the corresponding baseline methods in Figure FIGREF49 and Figure FIGREF53 . It is clear that, although AE and LSA does well or even better than LE and LPI, especially in dataset of both StackOverflow and Biomedical, STC INLINEFORM0 -LE and STC INLINEFORM1 -LPI achieve a much larger performance enhancements than STC INLINEFORM2 -AE and STC INLINEFORM3 -LSA do. The possible reason is that the information the pseudo supervision used to guide the learning of CNN model that make difference. Especially, for AE case, the input features fed into CNN model and the pseudo supervision employed to guide the learning of CNN model are all come from word embeddings. There are no different semantic features to be used into our proposed method, thus the performance enhancements are limited in STC INLINEFORM4 -AE. For LSA case, as we known, LSA is to make matrix factorization to find the best subspace approximation of the original feature space to minimize the global reconstruction error. And as BIBREF24 , BIBREF52 recently point out that word embeddings trained with word2vec or some variances, is essentially to do an operation of matrix factorization. Therefore, the information between input and the pseudo supervision in CNN is not departed very largely from each other, and the performance enhancements of STC INLINEFORM5 -AE is also not quite satisfactory. For LE and LPI case, as we known that LE extracts the manifold structure of the original feature space, and LPI extracts both geometric and discriminating structure of the original feature space BIBREF38 . We guess that our approach STC INLINEFORM6 -LE and STC INLINEFORM7 -LPI achieve enhancements compared with both LE and LPI by a large margin, because both of LE and LPI get useful semantic features, and these features are also different from word embeddings used as input of CNN. From this view, we say that our proposed STC has potential to behave more effective when the pseudo supervision is able to get semantic meaningful features, which is different enough from the input of CNN. Furthermore, from the results of K-means and AE in Table TABREF43 - TABREF44 and Figure FIGREF49 - FIGREF53 , we note that TF-IDF weighting gives a more remarkable improvement for K-means, while TF weighting works better than TF-IDF weighting for Average Embedding. Maybe the reason is that pre-trained word embeddings encode some useful information from external corpus and are able to get even better results without TF-IDF weighting. Meanwhile, we find that LE get quite unusual good performance than LPI, LSA and AE in SearchSnippets dataset, which is not found in the other two datasets. To get clear about this, and also to make a much better demonstration about our proposed approaches and other baselines, we further report 2-dimensional text embeddings on SearchSnippets in Figure FIGREF58 , using t-SNE BIBREF53 to get distributed stochastic neighbor embedding of the feature representations used in the clustering methods. We can see that the results of from AE and LSA seem to be fairly good or even better than the ones from LE and LPI, which is not the same as the results from ACC and NMI in Figure FIGREF49 - FIGREF53 . Meanwhile, RecNN (Ave.) performs better than BoW (both TF and TF-IDF) while RecNN (Top) does not, which is the same as the results from ACC and NMI in Table TABREF43 and Table TABREF44 . Then we guess that both ”the same as” and ”not the same as” above, is just a good example to illustrate that visualization tool, such as t-SNE, get some useful information for measuring results, which is different from the ones of ACC and NMI. Moreover, from this complementary view of t-SNE, we can see that our STC INLINEFORM0 -AE, STC INLINEFORM1 -LSA, STC INLINEFORM2 -LE, and STC INLINEFORM3 -LPI show more clear-cut margins among different semantic topics (that is, tags/labels), compared with AE, LSA, LE and LPI, respectively, as well as compared with both baselines, BoW and RecNN based ones. From all these results, with three measures of ACC, NMI and t-SNE under three datasets, we can get a solid conclusion that our proposed approaches is an effective approaches to get useful semantic features for short text clustering. Conclusions With the emergence of social media, short text clustering has become an increasing important task. This paper explores a new perspective to cluster short texts based on deep feature representation learned from the proposed self-taught convolutional neural networks. Our framework can be successfully accomplished without using any external tags/labels and complicated NLP pre-processing, and and our approach is a flexible framework, in which the traditional dimension reduction approaches could be used to get performance enhancement. Our extensive experimental study on three short text datasets shows that our approach can achieve a significantly better performance. In the future, how to select and incorporate more effective semantic features into the proposed framework would call for more research. Acknowledgments We would like to thank reviewers for their comments, and acknowledge Kaggle and BioASQ for making the datasets available. This work is supported by the National Natural Science Foundation of China (No. 61602479, No. 61303172, No. 61403385) and the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB02070005).

SearchSnippets, StackOverflow, Biomedical

3f6610d1d68c62eddc2150c460bf1b48a064e5e6

3f6610d1d68c62eddc2150c460bf1b48a064e5e6_0

Q: Does pre-training on general text corpus improve performance? Text: Introduction Students are exposed to simple arithmetic word problems starting in elementary school, and most become proficient in solving them at a young age. Automatic solvers of such problems could potentially help educators, as well as become an integral part of general question answering services. However, it has been challenging to write programs to solve even such elementary school level problems well. Solving a math word problem (MWP) starts with one or more sentences describing a transactional situation to be understood. The sentences are processed to produce an arithmetic expression, which is evaluated to provide an answer. Recent neural approaches to solving arithmetic word problems have used various flavors of recurrent neural networks (RNN) as well as reinforcement learning. Such methods have had difficulty achieving a high level of generalization. Often, systems extract the relevant numbers successfully but misplace them in the generated expressions. More problematic, they get the arithmetic operations wrong. The use of infix notation also requires pairs of parentheses to be placed and balanced correctly, bracketing the right numbers. There have been problems with parentheses placement as well. Correctly extracting the numbers in the problem is necessary. Figure FIGREF1 gives examples of some infix representations that a machine learning solver can potentially produce from a simple word problem using the correct numbers. Of the expressions shown, only the first one is correct. After carefully observing expressions that actual problem solvers have generated, we want to explore if the use of infix notation may itself be a part of the problem because it requires the generation of additional characters, the open and close parentheses, which must be balanced and placed correctly. The actual numbers appearing in MWPs vary widely from problem to problem. Real numbers take any conceivable value, making it almost impossible for a neural network to learn representations for them. As a result, trained programs sometimes generate expressions that have seemingly random numbers. For example, in some runs, a trained program could generate a potentially inexplicable expression such as $(25.01 - 4) * 9$ for the problem given in Figure FIGREF1, with one or more numbers not in the problem sentences. We hypothesize that replacing the numbers in the problem statement with generic tags like $\rm \langle n1 \rangle $, $\rm \langle n2 \rangle $, and $\rm \langle n3 \rangle $ and saving their values as a pre-processing step, does not take away from the generality of the solution, but suppresses the problem of fertility in number generation leading to the introduction of numbers not present in the question sentences. Another idea we want to test is whether a neural network which has been pre-trained to acquire language knowledge is better able to “understand" the problem sentences. Pre-training with a large amount of arithmetic-related text is likely to help develop such knowledge, but due to the lack of large such focused corpora, we want to test whether pre-training with a sufficient general corpus is beneficial. In this paper, we use the Transformer model BIBREF0 to solve arithmetic word problems as a particular case of machine translation from text to the language of arithmetic expressions. Transformers in various configurations have become a staple of NLP in the past two years. Past neural approaches did not treat this problem as pure translation like we do, and additionally, these approaches usually augmented the neural architectures with various external modules such as parse trees or used deep reinforcement learning, which we do not do. In this paper, we demonstrate that Transformers can be used to solve MWPs successfully with the simple adjustments we describe above. We compare performance on four individual datasets. In particular, we show that our translation-based approach outperforms state-of-the-art results reported by BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5 by a large margin on three of four datasets tested. On average, our best neural architecture outperforms previous results by almost 10%, although our approach is conceptually more straightforward. We organize our paper as follows. The second section presents related work. Then, we discuss our approach. We follow by an analysis of experimental results and compare them to those of other recent approaches. We also discuss our successes and shortcomings. Finally, we share our concluding thoughts and end with our direction for future work. Related Work Past strategies have used rules and templates to match sentences to arithmetic expressions. Some such approaches seemed to solve problems impressively within a narrow domain, but performed poorly when out of domain, lacking generality BIBREF6, BIBREF7, BIBREF8, BIBREF9. Kushman et al. BIBREF3 used feature extraction and template-based categorization by representing equations as expression forests and finding a near match. Such methods required human intervention in the form of feature engineering and development of templates and rules, which is not desirable for expandability and adaptability. Hosseini et al. BIBREF2 performed statistical similarity analysis to obtain acceptable results, but did not perform well with texts that were dissimilar to training examples. Existing approaches have used various forms of auxiliary information. Hosseini et al. BIBREF2 used verb categorization to identify important mathematical cues and contexts. Mitra and Baral BIBREF10 used predefined formulas to assist in matching. Koncel-Kedziorski et al. BIBREF11 parsed the input sentences, enumerated all parses, and learned to match, requiring expensive computations. Roy and Roth BIBREF12 performed searches for semantic trees over large spaces. Some recent approaches have transitioned to using neural networks. Semantic parsing takes advantage of RNN architectures to parse MWPs directly into equations or expressions in a math-specific language BIBREF9, BIBREF13. RNNs have shown promising results, but they have had difficulties balancing parenthesis, and also, sometimes incorrectly choose numbers when generating equations. Rehman et al. BIBREF14 used POS tagging and classification of equation templates to produce systems of equations from third-grade level MWPs. Most recently, Sun et al. BIBREF13 used a Bi-Directional LSTM architecture for math word problems. Huang et al. BIBREF15 used a deep reinforcement learning model to achieve character placement in both seen and novel equation templates. Wang et al. BIBREF1 also used deep reinforcement learning. Approach We view math word problem solving as a sequence-to-sequence translation problem. RNNs have excelled in sequence-to-sequence problems such as translation and question answering. The recent introduction of attention mechanisms has improved the performance of RNN models. Vaswani et al. BIBREF0 introduced the Transformer network, which uses stacks of attention layers instead of recurrence. Applications of Transformers have achieved state-of-the-art performance in many NLP tasks. We use this architecture to produce character sequences that are arithmetic expressions. The models we experiment with are easy and efficient to train, allowing us to test several configurations for a comprehensive comparison. We use several configurations of Transformer networks to learn the prefix, postfix, and infix notations of MWP equations independently. Prefix and postfix representations of equations do not contain parentheses, which has been a source of confusion in some approaches. If the learned target sequences are simple, with fewer characters to generate, it is less likely to make mistakes during generation. Simple targets also may help the learning of the model to be more robust. Experimenting with all three representations for equivalent expressions may help us discover which one works best. We train on standard datasets, which are readily available and commonly used. Our method considers the translation of English text to simple algebraic expressions. After performing experiments by training directly on math word problem corpora, we perform a different set of experiments by pre-training on a general language corpus. The success of pre-trained models such as ELMo BIBREF16, GPT-2 BIBREF17, and BERT BIBREF18 for many natural language tasks, provides reasoning that pre-training is likely to produce better learning by our system. We use pre-training so that the system has some foundational knowledge of English before we train it on the domain-specific text of math word problems. However, the output is not natural language but algebraic expressions, which is likely to limit the effectiveness of such pre-training. Approach ::: Data We work with four individual datasets. The datasets contain addition, subtraction, multiplication, and division word problems. AI2 BIBREF2. AI2 is a collection of 395 addition and subtraction problems, containing numeric values, where some may not be relevant to the question. CC BIBREF19. The Common Core dataset contains 600 2-step questions. The Cognitive Computation Group at the University of Pennsylvania gathered these questions. IL BIBREF4. The Illinois dataset contains 562 1-step algebra word questions. The Cognitive Computation Group compiled these questions also. MAWPS BIBREF20. MAWPS is a relatively large collection, primarily from other MWP datasets. We use 2,373 of 3,915 MWPs from this set. The problems not used were more complex problems that generate systems of equations. We exclude such problems because generating systems of equations is not our focus. We take a randomly sampled 95% of examples from each dataset for training. From each dataset, MWPs not included in training make up the testing data used when generating our results. Training and testing are repeated three times, and reported results are an average of the three outcomes. Approach ::: Representation Conversion We take a simple approach to convert infix expressions found in the MWPs to the other two representations. Two stacks are filled by iterating through string characters, one with operators found in the equation and the other with the operands. From these stacks, we form a binary tree structure. Traversing an expression tree in pre-order results in a prefix conversion. Post-order traversal gives us a postfix expression. Three versions of our training and testing data are created to correspond to each type of expression. By training on different representations, we expect our test results to change. Approach ::: Pre-training We pre-train half of our networks to endow them with a foundational knowledge of English. Pre-training models on significant-sized language corpora have been a common approach recently. We explore the pre-training approach using a general English corpus because the language of MWPs is regular English, interspersed with numerical values. Ideally, the corpus for pre-training should be a very general and comprehensive corpus like an English Wikipedia dump or many gigabytes of human-generated text scraped from the internet like GPT-2 BIBREF21 used. However, in this paper, we want to perform experiments to see if pre-training with a smaller corpus can help. In particular, for this task, we use the IMDb Movie Reviews dataset BIBREF22. This set contains 314,041 unique sentences. Since movie reviewers wrote this data, it is a reference to natural language not related to arithmetic. Training on a much bigger and general corpus may make the language model stronger, but we leave this for future work. We compare pre-trained models to non-pre-trained models to observe performance differences. Our pre-trained models are trained in an unsupervised fashion to improve the encodings of our fine-tuned solvers. In the pre-training process, we use sentences from the IMDb reviews with a target output of an empty string. We leave the input unlabelled, which focuses the network on adjusting encodings while providing unbiased decoding when we later change from IMDb English text to MWP-Data. Approach ::: Method: Training and Testing The input sequence is a natural language specification of an arithmetic word problem. The MWP questions and equations have been encoded using the subword text encoder provided by the TensorFlow Datasets library. The output is an expression in prefix, infix, or postfix notation, which then can be manipulated further and solved to obtain a final answer. All examples in the datasets contain numbers, some of which are unique or rare in the corpus. Rare terms are adverse for generalization since the network is unlikely to form good representations for them. As a remedy to this issue, our networks do not consider any relevant numbers during training. Before the networks attempt any translation, we pre-process each question and expression by a number mapping algorithm. This algorithm replaces each numeric value with a corresponding identifier (e.g., $\langle n1 \rangle $, $\langle n2 \rangle $, etc.), and remembers the necessary mapping. We expect that this approach may significantly improve how networks interpret each question. When translating, the numbers in the original question are tagged and cached. From the encoded English and tags, a predicted sequence resembling an expression presents itself as output. Since each network's learned output resembles an arithmetic expression (e.g., $\langle n1 \rangle + \langle n2 \rangle * \langle n3 \rangle $), we use the cached tag mapping to replace the tags with the corresponding numbers and return a final mathematical expression. Three representation models are trained and tested separately: Prefix-Transformer, Postfix-Transformer, and Infix-Transformer. For each experiment, we use representation-specific Transformer architectures. Each model uses the Adam optimizer with $beta_1=0.95$ and $beta_2=0.99$ with a standard epsilon of $1 \times e^{-9}$. The learning rate is reduced automatically in each training session as the loss decreases. Throughout the training, each model respects a 10% dropout rate. We employ a batch size of 128 for all training. Each model is trained on MWP data for 300 iterations before testing. The networks are trained on a machine using 1 Nvidia 1080 Ti graphics processing unit (GPU). We compare medium-sized, small, and minimal networks to show if network size can be reduced to increase training and testing efficiency while retaining high accuracy. Networks over six layers have shown to be non-effective for this task. We tried many configurations of our network models, but report results with only three configurations of Transformers. Transformer Type 1: This network is a small to medium-sized network consisting of 4 Transformer layers. Each layer utilizes 8 attention heads with a depth of 512 and a feed-forward depth of 1024. Transformer Type 2: The second model is small in size, using 2 Transformer layers. The layers utilize 8 attention heads with a depth of 256 and a feed-forward depth of 1024. Transformer Type 3: The third type of model is minimal, using only 1 Transformer layer. This network utilizes 8 attention heads with a depth of 256 and a feed-forward depth of 512. Approach ::: Method: Training and Testing ::: Objective Function We calculate the loss in training according to a mean of the sparse categorical cross-entropy formula. Sparse categorical cross-entropy BIBREF23 is used for identifying classes from a feature set, which assumes a large target classification set. Evaluation between the possible translation classes (all vocabulary subword tokens) and the produced class (predicted token) is the metric of performance here. During each evaluation, target terms are masked, predicted, and then compared to the masked (known) value. We adjust the model's loss according to the mean of the translation accuracy after predicting every determined subword in a translation. where $K = |Translation \; Classes|$, $J = |Translation|$, and $I$ is the number of examples. Approach ::: Method: Training and Testing ::: Experiment 1: Representation Some of the problems encountered by prior approaches seem to be attributable to the use of infix notation. In this experiment, we compare translation BLEU-2 scores to spot the differences in representation interpretability. Traditionally, a BLEU score is a metric of translation quality BIBREF24. Our presented BLEU scores represent an average of scores a given model received over each of the target test sets. We use a standard bi-gram weight to show how accurate translations are within a window of two adjacent terms. After testing translations, we calculate an average BLEU-2 score per test set, which is related to the success over that data. An average of the scores for each dataset become the presented value. where $N$ is the number of test datasets, which is 4. Approach ::: Method: Training and Testing ::: Experiment 2: State-of-the-art This experiment compares our networks to recent previous work. We count a given test score by a simple “correct versus incorrect" method. The answer to an expression directly ties to all of the translation terms being correct, which is why we do not consider partial precision. We compare average accuracies over 3 test trials on different randomly sampled test sets from each MWP dataset. This calculation more accurately depicts the generalization of our networks. Approach ::: Method: Training and Testing ::: Effect of Pre-training We also explore the effect of language pre-training, as discussed earlier. This training occurs over 30 iterations, at the start of the two experiments, to introduce a good level of language understanding before training on the MWP data. The same Transformer architectures are also trained solely on the MWP data. We calculate the reported results as: where $R$ is the number of test repetitions, which is 3; $N$ is the number of test datasets, which is 4; $P$ is the number of MWPs, and $C$ is the number of correct equation translations. Results We now present the results of our various experiments. We compare the three representations of target equations and three architectures of the Transformer model in each test. Results of Experiment 1 are given in Table TABREF21. For clarity, the number in parentheses in front of a row is the Transformer type. By using BLEU scores, we assess the translation capability of each network. This test displays how networks transform different math representations to a character summary level. We compare by average BLEU-2 accuracy among our tests in the Average column of Table TABREF21 to communicate these translation differences. To make it easier to understand the results, Table TABREF22 provides a summary of Table TABREF21. Looking at Tables TABREF21 and TABREF22, we note that both the prefix and postfix representations of our target language perform better than the generally used infix notation. The non-pre-trained models perform slightly better than the pre-trained models, and the small or Type 2 models perform slightly better than the minimal-sized and medium-sized Transformer models. The non-pre-trained type 2 prefix Transformer arrangement produced the most consistent translations. Table TABREF23 provides detailed results of Experiment 2. The numbers are absolute accuracies, i.e., they correspond to cases where the arithmetic expression generated is 100% correct, leading to the correct numeric answer. Results by BIBREF1, BIBREF2, BIBREF4, BIBREF5 are sparse but indicate the scale of success compared to recent past approaches. Prefix, postfix, and infix representations in Table TABREF23 show that network capabilities are changed by how teachable the target data is. The values in the last column of Table TABREF23 are summarized in Table TABREF24. How the models compare with respect to accuracy closely resembles the comparison of BLEU scores, presented earlier. Thus, BLEU scores seem to correlate well with accuracy values in our case. While our networks fell short of BIBREF1 AI2 testing accuracy, we present state-of-the-art results for the remaining three datasets. The AI2 dataset is tricky because it has numeric values in the word descriptions that are extraneous or irrelevant to the actual computation, whereas the other datasets have only relevant numeric values. The type 2 postfix Transformer received the highest testing average of 87.2%. Our attempt at language pre-training fell short of our expectations in all but one tested dataset. We had hoped that more stable language understanding would improve results in general. As previously mentioned, using more general and comprehensive corpora of language could help grow semantic ability. Results ::: Analysis All of the network configurations used were very successful for our task. The prefix representation overall provides the most stable network performance. To display the capability of our most successful model (type 2 postfix Transformer), we present some outputs of the network in Figure FIGREF26. The models respect the syntax of math expressions, even when incorrect. For the majority of questions, our translators were able to determine operators based solely on the context of language. Our pre-training was unsuccessful in improving accuracy, even when applied to networks larger than those reported. We may need to use more inclusive language, or pre-train on very math specific texts to be successful. Our results support our thesis of infix limitation. Results ::: Analysis ::: Error Analysis Our system, while performing above standard, could still benefit from some improvements. One issue originates from the algorithmic pre-processing of our questions and expressions. In Figure FIGREF27 we show an example of one such issue. The excerpt comes from a type 3 non-pre-trained Transformer test. The example shows an overlooked identifier, $\langle n1 \rangle $. The issue is attributed to the identifier algorithm only considering numbers in the problem. Observe in the question that the word “eight" is the number we expect to relate to $\langle n2 \rangle $. Our identifying algorithm could be improved by considering such number words and performing conversion to a numerical value. If our algorithm performed as expected, the identifier $\langle n1 \rangle $ relates with 4 (the first occurring number in the question) and $\langle n2 \rangle $ with 8 (the converted number word appearing second in the question). The overall translation was incorrect whether or not our algorithm was successful, but it is essential to analyze problems like these that may result in future improvements. Had all questions been tagged correctly, our performance would have likely improved. Conclusions and Future Work In this paper, we have shown that the use of Transformer networks improves automatic math word problem-solving. We have also shown that the use of postfix target expressions performs better than the other two expression formats. Our improvements are well-motivated but straightforward and easy to use, demonstrating that the well-acclaimed Transformer architecture for language processing can handle MWPs well, obviating the need to build specialized neural architectures for this task. Extensive pre-training over much larger corpora of language has extended the capabilities of many neural approaches. For example, networks like BERT BIBREF18, trained extensively on data from Wikipedia, perform relatively better in many tasks. Pre-training on a much larger corpus remains an extension we would like to try. We want to work with more complex MWP datasets. Our datasets contain basic arithmetic expressions of +, -, * and /, and only up to 3 of them. For example, datasets such as Dolphin18k BIBREF25, consisting of web-answered questions from Yahoo! Answers, require a wider variety of arithmetic operators to be understood by the system. We have noticed that the presence of irrelevant numbers in the sentences for MWPs limits our performance. We can think of such numbers as a sort of adversarial threat to an MWP solver that stress-test it. It may be interesting to explore how to keep a network's performance high, even in such cases. With a hope to further advance this area of research and heighten interests, all of the code and data used is available on GitHub. Acknowledgement The National Science Foundation supports the work reported in this paper under Grant No. 1659788. Any opinions, findings any conclusions or recommendations expressed in this work are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

No

4c854d33a832f3f729ce73b206ff90677e131e48

4c854d33a832f3f729ce73b206ff90677e131e48_0

Q: What neural configurations are explored? Text: Introduction Students are exposed to simple arithmetic word problems starting in elementary school, and most become proficient in solving them at a young age. Automatic solvers of such problems could potentially help educators, as well as become an integral part of general question answering services. However, it has been challenging to write programs to solve even such elementary school level problems well. Solving a math word problem (MWP) starts with one or more sentences describing a transactional situation to be understood. The sentences are processed to produce an arithmetic expression, which is evaluated to provide an answer. Recent neural approaches to solving arithmetic word problems have used various flavors of recurrent neural networks (RNN) as well as reinforcement learning. Such methods have had difficulty achieving a high level of generalization. Often, systems extract the relevant numbers successfully but misplace them in the generated expressions. More problematic, they get the arithmetic operations wrong. The use of infix notation also requires pairs of parentheses to be placed and balanced correctly, bracketing the right numbers. There have been problems with parentheses placement as well. Correctly extracting the numbers in the problem is necessary. Figure FIGREF1 gives examples of some infix representations that a machine learning solver can potentially produce from a simple word problem using the correct numbers. Of the expressions shown, only the first one is correct. After carefully observing expressions that actual problem solvers have generated, we want to explore if the use of infix notation may itself be a part of the problem because it requires the generation of additional characters, the open and close parentheses, which must be balanced and placed correctly. The actual numbers appearing in MWPs vary widely from problem to problem. Real numbers take any conceivable value, making it almost impossible for a neural network to learn representations for them. As a result, trained programs sometimes generate expressions that have seemingly random numbers. For example, in some runs, a trained program could generate a potentially inexplicable expression such as $(25.01 - 4) * 9$ for the problem given in Figure FIGREF1, with one or more numbers not in the problem sentences. We hypothesize that replacing the numbers in the problem statement with generic tags like $\rm \langle n1 \rangle $, $\rm \langle n2 \rangle $, and $\rm \langle n3 \rangle $ and saving their values as a pre-processing step, does not take away from the generality of the solution, but suppresses the problem of fertility in number generation leading to the introduction of numbers not present in the question sentences. Another idea we want to test is whether a neural network which has been pre-trained to acquire language knowledge is better able to “understand" the problem sentences. Pre-training with a large amount of arithmetic-related text is likely to help develop such knowledge, but due to the lack of large such focused corpora, we want to test whether pre-training with a sufficient general corpus is beneficial. In this paper, we use the Transformer model BIBREF0 to solve arithmetic word problems as a particular case of machine translation from text to the language of arithmetic expressions. Transformers in various configurations have become a staple of NLP in the past two years. Past neural approaches did not treat this problem as pure translation like we do, and additionally, these approaches usually augmented the neural architectures with various external modules such as parse trees or used deep reinforcement learning, which we do not do. In this paper, we demonstrate that Transformers can be used to solve MWPs successfully with the simple adjustments we describe above. We compare performance on four individual datasets. In particular, we show that our translation-based approach outperforms state-of-the-art results reported by BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5 by a large margin on three of four datasets tested. On average, our best neural architecture outperforms previous results by almost 10%, although our approach is conceptually more straightforward. We organize our paper as follows. The second section presents related work. Then, we discuss our approach. We follow by an analysis of experimental results and compare them to those of other recent approaches. We also discuss our successes and shortcomings. Finally, we share our concluding thoughts and end with our direction for future work. Related Work Past strategies have used rules and templates to match sentences to arithmetic expressions. Some such approaches seemed to solve problems impressively within a narrow domain, but performed poorly when out of domain, lacking generality BIBREF6, BIBREF7, BIBREF8, BIBREF9. Kushman et al. BIBREF3 used feature extraction and template-based categorization by representing equations as expression forests and finding a near match. Such methods required human intervention in the form of feature engineering and development of templates and rules, which is not desirable for expandability and adaptability. Hosseini et al. BIBREF2 performed statistical similarity analysis to obtain acceptable results, but did not perform well with texts that were dissimilar to training examples. Existing approaches have used various forms of auxiliary information. Hosseini et al. BIBREF2 used verb categorization to identify important mathematical cues and contexts. Mitra and Baral BIBREF10 used predefined formulas to assist in matching. Koncel-Kedziorski et al. BIBREF11 parsed the input sentences, enumerated all parses, and learned to match, requiring expensive computations. Roy and Roth BIBREF12 performed searches for semantic trees over large spaces. Some recent approaches have transitioned to using neural networks. Semantic parsing takes advantage of RNN architectures to parse MWPs directly into equations or expressions in a math-specific language BIBREF9, BIBREF13. RNNs have shown promising results, but they have had difficulties balancing parenthesis, and also, sometimes incorrectly choose numbers when generating equations. Rehman et al. BIBREF14 used POS tagging and classification of equation templates to produce systems of equations from third-grade level MWPs. Most recently, Sun et al. BIBREF13 used a Bi-Directional LSTM architecture for math word problems. Huang et al. BIBREF15 used a deep reinforcement learning model to achieve character placement in both seen and novel equation templates. Wang et al. BIBREF1 also used deep reinforcement learning. Approach We view math word problem solving as a sequence-to-sequence translation problem. RNNs have excelled in sequence-to-sequence problems such as translation and question answering. The recent introduction of attention mechanisms has improved the performance of RNN models. Vaswani et al. BIBREF0 introduced the Transformer network, which uses stacks of attention layers instead of recurrence. Applications of Transformers have achieved state-of-the-art performance in many NLP tasks. We use this architecture to produce character sequences that are arithmetic expressions. The models we experiment with are easy and efficient to train, allowing us to test several configurations for a comprehensive comparison. We use several configurations of Transformer networks to learn the prefix, postfix, and infix notations of MWP equations independently. Prefix and postfix representations of equations do not contain parentheses, which has been a source of confusion in some approaches. If the learned target sequences are simple, with fewer characters to generate, it is less likely to make mistakes during generation. Simple targets also may help the learning of the model to be more robust. Experimenting with all three representations for equivalent expressions may help us discover which one works best. We train on standard datasets, which are readily available and commonly used. Our method considers the translation of English text to simple algebraic expressions. After performing experiments by training directly on math word problem corpora, we perform a different set of experiments by pre-training on a general language corpus. The success of pre-trained models such as ELMo BIBREF16, GPT-2 BIBREF17, and BERT BIBREF18 for many natural language tasks, provides reasoning that pre-training is likely to produce better learning by our system. We use pre-training so that the system has some foundational knowledge of English before we train it on the domain-specific text of math word problems. However, the output is not natural language but algebraic expressions, which is likely to limit the effectiveness of such pre-training. Approach ::: Data We work with four individual datasets. The datasets contain addition, subtraction, multiplication, and division word problems. AI2 BIBREF2. AI2 is a collection of 395 addition and subtraction problems, containing numeric values, where some may not be relevant to the question. CC BIBREF19. The Common Core dataset contains 600 2-step questions. The Cognitive Computation Group at the University of Pennsylvania gathered these questions. IL BIBREF4. The Illinois dataset contains 562 1-step algebra word questions. The Cognitive Computation Group compiled these questions also. MAWPS BIBREF20. MAWPS is a relatively large collection, primarily from other MWP datasets. We use 2,373 of 3,915 MWPs from this set. The problems not used were more complex problems that generate systems of equations. We exclude such problems because generating systems of equations is not our focus. We take a randomly sampled 95% of examples from each dataset for training. From each dataset, MWPs not included in training make up the testing data used when generating our results. Training and testing are repeated three times, and reported results are an average of the three outcomes. Approach ::: Representation Conversion We take a simple approach to convert infix expressions found in the MWPs to the other two representations. Two stacks are filled by iterating through string characters, one with operators found in the equation and the other with the operands. From these stacks, we form a binary tree structure. Traversing an expression tree in pre-order results in a prefix conversion. Post-order traversal gives us a postfix expression. Three versions of our training and testing data are created to correspond to each type of expression. By training on different representations, we expect our test results to change. Approach ::: Pre-training We pre-train half of our networks to endow them with a foundational knowledge of English. Pre-training models on significant-sized language corpora have been a common approach recently. We explore the pre-training approach using a general English corpus because the language of MWPs is regular English, interspersed with numerical values. Ideally, the corpus for pre-training should be a very general and comprehensive corpus like an English Wikipedia dump or many gigabytes of human-generated text scraped from the internet like GPT-2 BIBREF21 used. However, in this paper, we want to perform experiments to see if pre-training with a smaller corpus can help. In particular, for this task, we use the IMDb Movie Reviews dataset BIBREF22. This set contains 314,041 unique sentences. Since movie reviewers wrote this data, it is a reference to natural language not related to arithmetic. Training on a much bigger and general corpus may make the language model stronger, but we leave this for future work. We compare pre-trained models to non-pre-trained models to observe performance differences. Our pre-trained models are trained in an unsupervised fashion to improve the encodings of our fine-tuned solvers. In the pre-training process, we use sentences from the IMDb reviews with a target output of an empty string. We leave the input unlabelled, which focuses the network on adjusting encodings while providing unbiased decoding when we later change from IMDb English text to MWP-Data. Approach ::: Method: Training and Testing The input sequence is a natural language specification of an arithmetic word problem. The MWP questions and equations have been encoded using the subword text encoder provided by the TensorFlow Datasets library. The output is an expression in prefix, infix, or postfix notation, which then can be manipulated further and solved to obtain a final answer. All examples in the datasets contain numbers, some of which are unique or rare in the corpus. Rare terms are adverse for generalization since the network is unlikely to form good representations for them. As a remedy to this issue, our networks do not consider any relevant numbers during training. Before the networks attempt any translation, we pre-process each question and expression by a number mapping algorithm. This algorithm replaces each numeric value with a corresponding identifier (e.g., $\langle n1 \rangle $, $\langle n2 \rangle $, etc.), and remembers the necessary mapping. We expect that this approach may significantly improve how networks interpret each question. When translating, the numbers in the original question are tagged and cached. From the encoded English and tags, a predicted sequence resembling an expression presents itself as output. Since each network's learned output resembles an arithmetic expression (e.g., $\langle n1 \rangle + \langle n2 \rangle * \langle n3 \rangle $), we use the cached tag mapping to replace the tags with the corresponding numbers and return a final mathematical expression. Three representation models are trained and tested separately: Prefix-Transformer, Postfix-Transformer, and Infix-Transformer. For each experiment, we use representation-specific Transformer architectures. Each model uses the Adam optimizer with $beta_1=0.95$ and $beta_2=0.99$ with a standard epsilon of $1 \times e^{-9}$. The learning rate is reduced automatically in each training session as the loss decreases. Throughout the training, each model respects a 10% dropout rate. We employ a batch size of 128 for all training. Each model is trained on MWP data for 300 iterations before testing. The networks are trained on a machine using 1 Nvidia 1080 Ti graphics processing unit (GPU). We compare medium-sized, small, and minimal networks to show if network size can be reduced to increase training and testing efficiency while retaining high accuracy. Networks over six layers have shown to be non-effective for this task. We tried many configurations of our network models, but report results with only three configurations of Transformers. Transformer Type 1: This network is a small to medium-sized network consisting of 4 Transformer layers. Each layer utilizes 8 attention heads with a depth of 512 and a feed-forward depth of 1024. Transformer Type 2: The second model is small in size, using 2 Transformer layers. The layers utilize 8 attention heads with a depth of 256 and a feed-forward depth of 1024. Transformer Type 3: The third type of model is minimal, using only 1 Transformer layer. This network utilizes 8 attention heads with a depth of 256 and a feed-forward depth of 512. Approach ::: Method: Training and Testing ::: Objective Function We calculate the loss in training according to a mean of the sparse categorical cross-entropy formula. Sparse categorical cross-entropy BIBREF23 is used for identifying classes from a feature set, which assumes a large target classification set. Evaluation between the possible translation classes (all vocabulary subword tokens) and the produced class (predicted token) is the metric of performance here. During each evaluation, target terms are masked, predicted, and then compared to the masked (known) value. We adjust the model's loss according to the mean of the translation accuracy after predicting every determined subword in a translation. where $K = |Translation \; Classes|$, $J = |Translation|$, and $I$ is the number of examples. Approach ::: Method: Training and Testing ::: Experiment 1: Representation Some of the problems encountered by prior approaches seem to be attributable to the use of infix notation. In this experiment, we compare translation BLEU-2 scores to spot the differences in representation interpretability. Traditionally, a BLEU score is a metric of translation quality BIBREF24. Our presented BLEU scores represent an average of scores a given model received over each of the target test sets. We use a standard bi-gram weight to show how accurate translations are within a window of two adjacent terms. After testing translations, we calculate an average BLEU-2 score per test set, which is related to the success over that data. An average of the scores for each dataset become the presented value. where $N$ is the number of test datasets, which is 4. Approach ::: Method: Training and Testing ::: Experiment 2: State-of-the-art This experiment compares our networks to recent previous work. We count a given test score by a simple “correct versus incorrect" method. The answer to an expression directly ties to all of the translation terms being correct, which is why we do not consider partial precision. We compare average accuracies over 3 test trials on different randomly sampled test sets from each MWP dataset. This calculation more accurately depicts the generalization of our networks. Approach ::: Method: Training and Testing ::: Effect of Pre-training We also explore the effect of language pre-training, as discussed earlier. This training occurs over 30 iterations, at the start of the two experiments, to introduce a good level of language understanding before training on the MWP data. The same Transformer architectures are also trained solely on the MWP data. We calculate the reported results as: where $R$ is the number of test repetitions, which is 3; $N$ is the number of test datasets, which is 4; $P$ is the number of MWPs, and $C$ is the number of correct equation translations. Results We now present the results of our various experiments. We compare the three representations of target equations and three architectures of the Transformer model in each test. Results of Experiment 1 are given in Table TABREF21. For clarity, the number in parentheses in front of a row is the Transformer type. By using BLEU scores, we assess the translation capability of each network. This test displays how networks transform different math representations to a character summary level. We compare by average BLEU-2 accuracy among our tests in the Average column of Table TABREF21 to communicate these translation differences. To make it easier to understand the results, Table TABREF22 provides a summary of Table TABREF21. Looking at Tables TABREF21 and TABREF22, we note that both the prefix and postfix representations of our target language perform better than the generally used infix notation. The non-pre-trained models perform slightly better than the pre-trained models, and the small or Type 2 models perform slightly better than the minimal-sized and medium-sized Transformer models. The non-pre-trained type 2 prefix Transformer arrangement produced the most consistent translations. Table TABREF23 provides detailed results of Experiment 2. The numbers are absolute accuracies, i.e., they correspond to cases where the arithmetic expression generated is 100% correct, leading to the correct numeric answer. Results by BIBREF1, BIBREF2, BIBREF4, BIBREF5 are sparse but indicate the scale of success compared to recent past approaches. Prefix, postfix, and infix representations in Table TABREF23 show that network capabilities are changed by how teachable the target data is. The values in the last column of Table TABREF23 are summarized in Table TABREF24. How the models compare with respect to accuracy closely resembles the comparison of BLEU scores, presented earlier. Thus, BLEU scores seem to correlate well with accuracy values in our case. While our networks fell short of BIBREF1 AI2 testing accuracy, we present state-of-the-art results for the remaining three datasets. The AI2 dataset is tricky because it has numeric values in the word descriptions that are extraneous or irrelevant to the actual computation, whereas the other datasets have only relevant numeric values. The type 2 postfix Transformer received the highest testing average of 87.2%. Our attempt at language pre-training fell short of our expectations in all but one tested dataset. We had hoped that more stable language understanding would improve results in general. As previously mentioned, using more general and comprehensive corpora of language could help grow semantic ability. Results ::: Analysis All of the network configurations used were very successful for our task. The prefix representation overall provides the most stable network performance. To display the capability of our most successful model (type 2 postfix Transformer), we present some outputs of the network in Figure FIGREF26. The models respect the syntax of math expressions, even when incorrect. For the majority of questions, our translators were able to determine operators based solely on the context of language. Our pre-training was unsuccessful in improving accuracy, even when applied to networks larger than those reported. We may need to use more inclusive language, or pre-train on very math specific texts to be successful. Our results support our thesis of infix limitation. Results ::: Analysis ::: Error Analysis Our system, while performing above standard, could still benefit from some improvements. One issue originates from the algorithmic pre-processing of our questions and expressions. In Figure FIGREF27 we show an example of one such issue. The excerpt comes from a type 3 non-pre-trained Transformer test. The example shows an overlooked identifier, $\langle n1 \rangle $. The issue is attributed to the identifier algorithm only considering numbers in the problem. Observe in the question that the word “eight" is the number we expect to relate to $\langle n2 \rangle $. Our identifying algorithm could be improved by considering such number words and performing conversion to a numerical value. If our algorithm performed as expected, the identifier $\langle n1 \rangle $ relates with 4 (the first occurring number in the question) and $\langle n2 \rangle $ with 8 (the converted number word appearing second in the question). The overall translation was incorrect whether or not our algorithm was successful, but it is essential to analyze problems like these that may result in future improvements. Had all questions been tagged correctly, our performance would have likely improved. Conclusions and Future Work In this paper, we have shown that the use of Transformer networks improves automatic math word problem-solving. We have also shown that the use of postfix target expressions performs better than the other two expression formats. Our improvements are well-motivated but straightforward and easy to use, demonstrating that the well-acclaimed Transformer architecture for language processing can handle MWPs well, obviating the need to build specialized neural architectures for this task. Extensive pre-training over much larger corpora of language has extended the capabilities of many neural approaches. For example, networks like BERT BIBREF18, trained extensively on data from Wikipedia, perform relatively better in many tasks. Pre-training on a much larger corpus remains an extension we would like to try. We want to work with more complex MWP datasets. Our datasets contain basic arithmetic expressions of +, -, * and /, and only up to 3 of them. For example, datasets such as Dolphin18k BIBREF25, consisting of web-answered questions from Yahoo! Answers, require a wider variety of arithmetic operators to be understood by the system. We have noticed that the presence of irrelevant numbers in the sentences for MWPs limits our performance. We can think of such numbers as a sort of adversarial threat to an MWP solver that stress-test it. It may be interesting to explore how to keep a network's performance high, even in such cases. With a hope to further advance this area of research and heighten interests, all of the code and data used is available on GitHub. Acknowledgement The National Science Foundation supports the work reported in this paper under Grant No. 1659788. Any opinions, findings any conclusions or recommendations expressed in this work are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

tried many configurations of our network models, but report results with only three configurations, Transformer Type 1, Transformer Type 2, Transformer Type 3

163c15da1aa0ba370a00c5a09294cd2ccdb4b96d

163c15da1aa0ba370a00c5a09294cd2ccdb4b96d_0

Q: Are the Transformers masked? Text: Introduction Students are exposed to simple arithmetic word problems starting in elementary school, and most become proficient in solving them at a young age. Automatic solvers of such problems could potentially help educators, as well as become an integral part of general question answering services. However, it has been challenging to write programs to solve even such elementary school level problems well. Solving a math word problem (MWP) starts with one or more sentences describing a transactional situation to be understood. The sentences are processed to produce an arithmetic expression, which is evaluated to provide an answer. Recent neural approaches to solving arithmetic word problems have used various flavors of recurrent neural networks (RNN) as well as reinforcement learning. Such methods have had difficulty achieving a high level of generalization. Often, systems extract the relevant numbers successfully but misplace them in the generated expressions. More problematic, they get the arithmetic operations wrong. The use of infix notation also requires pairs of parentheses to be placed and balanced correctly, bracketing the right numbers. There have been problems with parentheses placement as well. Correctly extracting the numbers in the problem is necessary. Figure FIGREF1 gives examples of some infix representations that a machine learning solver can potentially produce from a simple word problem using the correct numbers. Of the expressions shown, only the first one is correct. After carefully observing expressions that actual problem solvers have generated, we want to explore if the use of infix notation may itself be a part of the problem because it requires the generation of additional characters, the open and close parentheses, which must be balanced and placed correctly. The actual numbers appearing in MWPs vary widely from problem to problem. Real numbers take any conceivable value, making it almost impossible for a neural network to learn representations for them. As a result, trained programs sometimes generate expressions that have seemingly random numbers. For example, in some runs, a trained program could generate a potentially inexplicable expression such as $(25.01 - 4) * 9$ for the problem given in Figure FIGREF1, with one or more numbers not in the problem sentences. We hypothesize that replacing the numbers in the problem statement with generic tags like $\rm \langle n1 \rangle $, $\rm \langle n2 \rangle $, and $\rm \langle n3 \rangle $ and saving their values as a pre-processing step, does not take away from the generality of the solution, but suppresses the problem of fertility in number generation leading to the introduction of numbers not present in the question sentences. Another idea we want to test is whether a neural network which has been pre-trained to acquire language knowledge is better able to “understand" the problem sentences. Pre-training with a large amount of arithmetic-related text is likely to help develop such knowledge, but due to the lack of large such focused corpora, we want to test whether pre-training with a sufficient general corpus is beneficial. In this paper, we use the Transformer model BIBREF0 to solve arithmetic word problems as a particular case of machine translation from text to the language of arithmetic expressions. Transformers in various configurations have become a staple of NLP in the past two years. Past neural approaches did not treat this problem as pure translation like we do, and additionally, these approaches usually augmented the neural architectures with various external modules such as parse trees or used deep reinforcement learning, which we do not do. In this paper, we demonstrate that Transformers can be used to solve MWPs successfully with the simple adjustments we describe above. We compare performance on four individual datasets. In particular, we show that our translation-based approach outperforms state-of-the-art results reported by BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5 by a large margin on three of four datasets tested. On average, our best neural architecture outperforms previous results by almost 10%, although our approach is conceptually more straightforward. We organize our paper as follows. The second section presents related work. Then, we discuss our approach. We follow by an analysis of experimental results and compare them to those of other recent approaches. We also discuss our successes and shortcomings. Finally, we share our concluding thoughts and end with our direction for future work. Related Work Past strategies have used rules and templates to match sentences to arithmetic expressions. Some such approaches seemed to solve problems impressively within a narrow domain, but performed poorly when out of domain, lacking generality BIBREF6, BIBREF7, BIBREF8, BIBREF9. Kushman et al. BIBREF3 used feature extraction and template-based categorization by representing equations as expression forests and finding a near match. Such methods required human intervention in the form of feature engineering and development of templates and rules, which is not desirable for expandability and adaptability. Hosseini et al. BIBREF2 performed statistical similarity analysis to obtain acceptable results, but did not perform well with texts that were dissimilar to training examples. Existing approaches have used various forms of auxiliary information. Hosseini et al. BIBREF2 used verb categorization to identify important mathematical cues and contexts. Mitra and Baral BIBREF10 used predefined formulas to assist in matching. Koncel-Kedziorski et al. BIBREF11 parsed the input sentences, enumerated all parses, and learned to match, requiring expensive computations. Roy and Roth BIBREF12 performed searches for semantic trees over large spaces. Some recent approaches have transitioned to using neural networks. Semantic parsing takes advantage of RNN architectures to parse MWPs directly into equations or expressions in a math-specific language BIBREF9, BIBREF13. RNNs have shown promising results, but they have had difficulties balancing parenthesis, and also, sometimes incorrectly choose numbers when generating equations. Rehman et al. BIBREF14 used POS tagging and classification of equation templates to produce systems of equations from third-grade level MWPs. Most recently, Sun et al. BIBREF13 used a Bi-Directional LSTM architecture for math word problems. Huang et al. BIBREF15 used a deep reinforcement learning model to achieve character placement in both seen and novel equation templates. Wang et al. BIBREF1 also used deep reinforcement learning. Approach We view math word problem solving as a sequence-to-sequence translation problem. RNNs have excelled in sequence-to-sequence problems such as translation and question answering. The recent introduction of attention mechanisms has improved the performance of RNN models. Vaswani et al. BIBREF0 introduced the Transformer network, which uses stacks of attention layers instead of recurrence. Applications of Transformers have achieved state-of-the-art performance in many NLP tasks. We use this architecture to produce character sequences that are arithmetic expressions. The models we experiment with are easy and efficient to train, allowing us to test several configurations for a comprehensive comparison. We use several configurations of Transformer networks to learn the prefix, postfix, and infix notations of MWP equations independently. Prefix and postfix representations of equations do not contain parentheses, which has been a source of confusion in some approaches. If the learned target sequences are simple, with fewer characters to generate, it is less likely to make mistakes during generation. Simple targets also may help the learning of the model to be more robust. Experimenting with all three representations for equivalent expressions may help us discover which one works best. We train on standard datasets, which are readily available and commonly used. Our method considers the translation of English text to simple algebraic expressions. After performing experiments by training directly on math word problem corpora, we perform a different set of experiments by pre-training on a general language corpus. The success of pre-trained models such as ELMo BIBREF16, GPT-2 BIBREF17, and BERT BIBREF18 for many natural language tasks, provides reasoning that pre-training is likely to produce better learning by our system. We use pre-training so that the system has some foundational knowledge of English before we train it on the domain-specific text of math word problems. However, the output is not natural language but algebraic expressions, which is likely to limit the effectiveness of such pre-training. Approach ::: Data We work with four individual datasets. The datasets contain addition, subtraction, multiplication, and division word problems. AI2 BIBREF2. AI2 is a collection of 395 addition and subtraction problems, containing numeric values, where some may not be relevant to the question. CC BIBREF19. The Common Core dataset contains 600 2-step questions. The Cognitive Computation Group at the University of Pennsylvania gathered these questions. IL BIBREF4. The Illinois dataset contains 562 1-step algebra word questions. The Cognitive Computation Group compiled these questions also. MAWPS BIBREF20. MAWPS is a relatively large collection, primarily from other MWP datasets. We use 2,373 of 3,915 MWPs from this set. The problems not used were more complex problems that generate systems of equations. We exclude such problems because generating systems of equations is not our focus. We take a randomly sampled 95% of examples from each dataset for training. From each dataset, MWPs not included in training make up the testing data used when generating our results. Training and testing are repeated three times, and reported results are an average of the three outcomes. Approach ::: Representation Conversion We take a simple approach to convert infix expressions found in the MWPs to the other two representations. Two stacks are filled by iterating through string characters, one with operators found in the equation and the other with the operands. From these stacks, we form a binary tree structure. Traversing an expression tree in pre-order results in a prefix conversion. Post-order traversal gives us a postfix expression. Three versions of our training and testing data are created to correspond to each type of expression. By training on different representations, we expect our test results to change. Approach ::: Pre-training We pre-train half of our networks to endow them with a foundational knowledge of English. Pre-training models on significant-sized language corpora have been a common approach recently. We explore the pre-training approach using a general English corpus because the language of MWPs is regular English, interspersed with numerical values. Ideally, the corpus for pre-training should be a very general and comprehensive corpus like an English Wikipedia dump or many gigabytes of human-generated text scraped from the internet like GPT-2 BIBREF21 used. However, in this paper, we want to perform experiments to see if pre-training with a smaller corpus can help. In particular, for this task, we use the IMDb Movie Reviews dataset BIBREF22. This set contains 314,041 unique sentences. Since movie reviewers wrote this data, it is a reference to natural language not related to arithmetic. Training on a much bigger and general corpus may make the language model stronger, but we leave this for future work. We compare pre-trained models to non-pre-trained models to observe performance differences. Our pre-trained models are trained in an unsupervised fashion to improve the encodings of our fine-tuned solvers. In the pre-training process, we use sentences from the IMDb reviews with a target output of an empty string. We leave the input unlabelled, which focuses the network on adjusting encodings while providing unbiased decoding when we later change from IMDb English text to MWP-Data. Approach ::: Method: Training and Testing The input sequence is a natural language specification of an arithmetic word problem. The MWP questions and equations have been encoded using the subword text encoder provided by the TensorFlow Datasets library. The output is an expression in prefix, infix, or postfix notation, which then can be manipulated further and solved to obtain a final answer. All examples in the datasets contain numbers, some of which are unique or rare in the corpus. Rare terms are adverse for generalization since the network is unlikely to form good representations for them. As a remedy to this issue, our networks do not consider any relevant numbers during training. Before the networks attempt any translation, we pre-process each question and expression by a number mapping algorithm. This algorithm replaces each numeric value with a corresponding identifier (e.g., $\langle n1 \rangle $, $\langle n2 \rangle $, etc.), and remembers the necessary mapping. We expect that this approach may significantly improve how networks interpret each question. When translating, the numbers in the original question are tagged and cached. From the encoded English and tags, a predicted sequence resembling an expression presents itself as output. Since each network's learned output resembles an arithmetic expression (e.g., $\langle n1 \rangle + \langle n2 \rangle * \langle n3 \rangle $), we use the cached tag mapping to replace the tags with the corresponding numbers and return a final mathematical expression. Three representation models are trained and tested separately: Prefix-Transformer, Postfix-Transformer, and Infix-Transformer. For each experiment, we use representation-specific Transformer architectures. Each model uses the Adam optimizer with $beta_1=0.95$ and $beta_2=0.99$ with a standard epsilon of $1 \times e^{-9}$. The learning rate is reduced automatically in each training session as the loss decreases. Throughout the training, each model respects a 10% dropout rate. We employ a batch size of 128 for all training. Each model is trained on MWP data for 300 iterations before testing. The networks are trained on a machine using 1 Nvidia 1080 Ti graphics processing unit (GPU). We compare medium-sized, small, and minimal networks to show if network size can be reduced to increase training and testing efficiency while retaining high accuracy. Networks over six layers have shown to be non-effective for this task. We tried many configurations of our network models, but report results with only three configurations of Transformers. Transformer Type 1: This network is a small to medium-sized network consisting of 4 Transformer layers. Each layer utilizes 8 attention heads with a depth of 512 and a feed-forward depth of 1024. Transformer Type 2: The second model is small in size, using 2 Transformer layers. The layers utilize 8 attention heads with a depth of 256 and a feed-forward depth of 1024. Transformer Type 3: The third type of model is minimal, using only 1 Transformer layer. This network utilizes 8 attention heads with a depth of 256 and a feed-forward depth of 512. Approach ::: Method: Training and Testing ::: Objective Function We calculate the loss in training according to a mean of the sparse categorical cross-entropy formula. Sparse categorical cross-entropy BIBREF23 is used for identifying classes from a feature set, which assumes a large target classification set. Evaluation between the possible translation classes (all vocabulary subword tokens) and the produced class (predicted token) is the metric of performance here. During each evaluation, target terms are masked, predicted, and then compared to the masked (known) value. We adjust the model's loss according to the mean of the translation accuracy after predicting every determined subword in a translation. where $K = |Translation \; Classes|$, $J = |Translation|$, and $I$ is the number of examples. Approach ::: Method: Training and Testing ::: Experiment 1: Representation Some of the problems encountered by prior approaches seem to be attributable to the use of infix notation. In this experiment, we compare translation BLEU-2 scores to spot the differences in representation interpretability. Traditionally, a BLEU score is a metric of translation quality BIBREF24. Our presented BLEU scores represent an average of scores a given model received over each of the target test sets. We use a standard bi-gram weight to show how accurate translations are within a window of two adjacent terms. After testing translations, we calculate an average BLEU-2 score per test set, which is related to the success over that data. An average of the scores for each dataset become the presented value. where $N$ is the number of test datasets, which is 4. Approach ::: Method: Training and Testing ::: Experiment 2: State-of-the-art This experiment compares our networks to recent previous work. We count a given test score by a simple “correct versus incorrect" method. The answer to an expression directly ties to all of the translation terms being correct, which is why we do not consider partial precision. We compare average accuracies over 3 test trials on different randomly sampled test sets from each MWP dataset. This calculation more accurately depicts the generalization of our networks. Approach ::: Method: Training and Testing ::: Effect of Pre-training We also explore the effect of language pre-training, as discussed earlier. This training occurs over 30 iterations, at the start of the two experiments, to introduce a good level of language understanding before training on the MWP data. The same Transformer architectures are also trained solely on the MWP data. We calculate the reported results as: where $R$ is the number of test repetitions, which is 3; $N$ is the number of test datasets, which is 4; $P$ is the number of MWPs, and $C$ is the number of correct equation translations. Results We now present the results of our various experiments. We compare the three representations of target equations and three architectures of the Transformer model in each test. Results of Experiment 1 are given in Table TABREF21. For clarity, the number in parentheses in front of a row is the Transformer type. By using BLEU scores, we assess the translation capability of each network. This test displays how networks transform different math representations to a character summary level. We compare by average BLEU-2 accuracy among our tests in the Average column of Table TABREF21 to communicate these translation differences. To make it easier to understand the results, Table TABREF22 provides a summary of Table TABREF21. Looking at Tables TABREF21 and TABREF22, we note that both the prefix and postfix representations of our target language perform better than the generally used infix notation. The non-pre-trained models perform slightly better than the pre-trained models, and the small or Type 2 models perform slightly better than the minimal-sized and medium-sized Transformer models. The non-pre-trained type 2 prefix Transformer arrangement produced the most consistent translations. Table TABREF23 provides detailed results of Experiment 2. The numbers are absolute accuracies, i.e., they correspond to cases where the arithmetic expression generated is 100% correct, leading to the correct numeric answer. Results by BIBREF1, BIBREF2, BIBREF4, BIBREF5 are sparse but indicate the scale of success compared to recent past approaches. Prefix, postfix, and infix representations in Table TABREF23 show that network capabilities are changed by how teachable the target data is. The values in the last column of Table TABREF23 are summarized in Table TABREF24. How the models compare with respect to accuracy closely resembles the comparison of BLEU scores, presented earlier. Thus, BLEU scores seem to correlate well with accuracy values in our case. While our networks fell short of BIBREF1 AI2 testing accuracy, we present state-of-the-art results for the remaining three datasets. The AI2 dataset is tricky because it has numeric values in the word descriptions that are extraneous or irrelevant to the actual computation, whereas the other datasets have only relevant numeric values. The type 2 postfix Transformer received the highest testing average of 87.2%. Our attempt at language pre-training fell short of our expectations in all but one tested dataset. We had hoped that more stable language understanding would improve results in general. As previously mentioned, using more general and comprehensive corpora of language could help grow semantic ability. Results ::: Analysis All of the network configurations used were very successful for our task. The prefix representation overall provides the most stable network performance. To display the capability of our most successful model (type 2 postfix Transformer), we present some outputs of the network in Figure FIGREF26. The models respect the syntax of math expressions, even when incorrect. For the majority of questions, our translators were able to determine operators based solely on the context of language. Our pre-training was unsuccessful in improving accuracy, even when applied to networks larger than those reported. We may need to use more inclusive language, or pre-train on very math specific texts to be successful. Our results support our thesis of infix limitation. Results ::: Analysis ::: Error Analysis Our system, while performing above standard, could still benefit from some improvements. One issue originates from the algorithmic pre-processing of our questions and expressions. In Figure FIGREF27 we show an example of one such issue. The excerpt comes from a type 3 non-pre-trained Transformer test. The example shows an overlooked identifier, $\langle n1 \rangle $. The issue is attributed to the identifier algorithm only considering numbers in the problem. Observe in the question that the word “eight" is the number we expect to relate to $\langle n2 \rangle $. Our identifying algorithm could be improved by considering such number words and performing conversion to a numerical value. If our algorithm performed as expected, the identifier $\langle n1 \rangle $ relates with 4 (the first occurring number in the question) and $\langle n2 \rangle $ with 8 (the converted number word appearing second in the question). The overall translation was incorrect whether or not our algorithm was successful, but it is essential to analyze problems like these that may result in future improvements. Had all questions been tagged correctly, our performance would have likely improved. Conclusions and Future Work In this paper, we have shown that the use of Transformer networks improves automatic math word problem-solving. We have also shown that the use of postfix target expressions performs better than the other two expression formats. Our improvements are well-motivated but straightforward and easy to use, demonstrating that the well-acclaimed Transformer architecture for language processing can handle MWPs well, obviating the need to build specialized neural architectures for this task. Extensive pre-training over much larger corpora of language has extended the capabilities of many neural approaches. For example, networks like BERT BIBREF18, trained extensively on data from Wikipedia, perform relatively better in many tasks. Pre-training on a much larger corpus remains an extension we would like to try. We want to work with more complex MWP datasets. Our datasets contain basic arithmetic expressions of +, -, * and /, and only up to 3 of them. For example, datasets such as Dolphin18k BIBREF25, consisting of web-answered questions from Yahoo! Answers, require a wider variety of arithmetic operators to be understood by the system. We have noticed that the presence of irrelevant numbers in the sentences for MWPs limits our performance. We can think of such numbers as a sort of adversarial threat to an MWP solver that stress-test it. It may be interesting to explore how to keep a network's performance high, even in such cases. With a hope to further advance this area of research and heighten interests, all of the code and data used is available on GitHub. Acknowledgement The National Science Foundation supports the work reported in this paper under Grant No. 1659788. Any opinions, findings any conclusions or recommendations expressed in this work are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Yes

90dd5c0f5084a045fd6346469bc853c33622908f

90dd5c0f5084a045fd6346469bc853c33622908f_0

Q: How is this problem evaluated? Text: Introduction Students are exposed to simple arithmetic word problems starting in elementary school, and most become proficient in solving them at a young age. Automatic solvers of such problems could potentially help educators, as well as become an integral part of general question answering services. However, it has been challenging to write programs to solve even such elementary school level problems well. Solving a math word problem (MWP) starts with one or more sentences describing a transactional situation to be understood. The sentences are processed to produce an arithmetic expression, which is evaluated to provide an answer. Recent neural approaches to solving arithmetic word problems have used various flavors of recurrent neural networks (RNN) as well as reinforcement learning. Such methods have had difficulty achieving a high level of generalization. Often, systems extract the relevant numbers successfully but misplace them in the generated expressions. More problematic, they get the arithmetic operations wrong. The use of infix notation also requires pairs of parentheses to be placed and balanced correctly, bracketing the right numbers. There have been problems with parentheses placement as well. Correctly extracting the numbers in the problem is necessary. Figure FIGREF1 gives examples of some infix representations that a machine learning solver can potentially produce from a simple word problem using the correct numbers. Of the expressions shown, only the first one is correct. After carefully observing expressions that actual problem solvers have generated, we want to explore if the use of infix notation may itself be a part of the problem because it requires the generation of additional characters, the open and close parentheses, which must be balanced and placed correctly. The actual numbers appearing in MWPs vary widely from problem to problem. Real numbers take any conceivable value, making it almost impossible for a neural network to learn representations for them. As a result, trained programs sometimes generate expressions that have seemingly random numbers. For example, in some runs, a trained program could generate a potentially inexplicable expression such as $(25.01 - 4) * 9$ for the problem given in Figure FIGREF1, with one or more numbers not in the problem sentences. We hypothesize that replacing the numbers in the problem statement with generic tags like $\rm \langle n1 \rangle $, $\rm \langle n2 \rangle $, and $\rm \langle n3 \rangle $ and saving their values as a pre-processing step, does not take away from the generality of the solution, but suppresses the problem of fertility in number generation leading to the introduction of numbers not present in the question sentences. Another idea we want to test is whether a neural network which has been pre-trained to acquire language knowledge is better able to “understand" the problem sentences. Pre-training with a large amount of arithmetic-related text is likely to help develop such knowledge, but due to the lack of large such focused corpora, we want to test whether pre-training with a sufficient general corpus is beneficial. In this paper, we use the Transformer model BIBREF0 to solve arithmetic word problems as a particular case of machine translation from text to the language of arithmetic expressions. Transformers in various configurations have become a staple of NLP in the past two years. Past neural approaches did not treat this problem as pure translation like we do, and additionally, these approaches usually augmented the neural architectures with various external modules such as parse trees or used deep reinforcement learning, which we do not do. In this paper, we demonstrate that Transformers can be used to solve MWPs successfully with the simple adjustments we describe above. We compare performance on four individual datasets. In particular, we show that our translation-based approach outperforms state-of-the-art results reported by BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5 by a large margin on three of four datasets tested. On average, our best neural architecture outperforms previous results by almost 10%, although our approach is conceptually more straightforward. We organize our paper as follows. The second section presents related work. Then, we discuss our approach. We follow by an analysis of experimental results and compare them to those of other recent approaches. We also discuss our successes and shortcomings. Finally, we share our concluding thoughts and end with our direction for future work. Related Work Past strategies have used rules and templates to match sentences to arithmetic expressions. Some such approaches seemed to solve problems impressively within a narrow domain, but performed poorly when out of domain, lacking generality BIBREF6, BIBREF7, BIBREF8, BIBREF9. Kushman et al. BIBREF3 used feature extraction and template-based categorization by representing equations as expression forests and finding a near match. Such methods required human intervention in the form of feature engineering and development of templates and rules, which is not desirable for expandability and adaptability. Hosseini et al. BIBREF2 performed statistical similarity analysis to obtain acceptable results, but did not perform well with texts that were dissimilar to training examples. Existing approaches have used various forms of auxiliary information. Hosseini et al. BIBREF2 used verb categorization to identify important mathematical cues and contexts. Mitra and Baral BIBREF10 used predefined formulas to assist in matching. Koncel-Kedziorski et al. BIBREF11 parsed the input sentences, enumerated all parses, and learned to match, requiring expensive computations. Roy and Roth BIBREF12 performed searches for semantic trees over large spaces. Some recent approaches have transitioned to using neural networks. Semantic parsing takes advantage of RNN architectures to parse MWPs directly into equations or expressions in a math-specific language BIBREF9, BIBREF13. RNNs have shown promising results, but they have had difficulties balancing parenthesis, and also, sometimes incorrectly choose numbers when generating equations. Rehman et al. BIBREF14 used POS tagging and classification of equation templates to produce systems of equations from third-grade level MWPs. Most recently, Sun et al. BIBREF13 used a Bi-Directional LSTM architecture for math word problems. Huang et al. BIBREF15 used a deep reinforcement learning model to achieve character placement in both seen and novel equation templates. Wang et al. BIBREF1 also used deep reinforcement learning. Approach We view math word problem solving as a sequence-to-sequence translation problem. RNNs have excelled in sequence-to-sequence problems such as translation and question answering. The recent introduction of attention mechanisms has improved the performance of RNN models. Vaswani et al. BIBREF0 introduced the Transformer network, which uses stacks of attention layers instead of recurrence. Applications of Transformers have achieved state-of-the-art performance in many NLP tasks. We use this architecture to produce character sequences that are arithmetic expressions. The models we experiment with are easy and efficient to train, allowing us to test several configurations for a comprehensive comparison. We use several configurations of Transformer networks to learn the prefix, postfix, and infix notations of MWP equations independently. Prefix and postfix representations of equations do not contain parentheses, which has been a source of confusion in some approaches. If the learned target sequences are simple, with fewer characters to generate, it is less likely to make mistakes during generation. Simple targets also may help the learning of the model to be more robust. Experimenting with all three representations for equivalent expressions may help us discover which one works best. We train on standard datasets, which are readily available and commonly used. Our method considers the translation of English text to simple algebraic expressions. After performing experiments by training directly on math word problem corpora, we perform a different set of experiments by pre-training on a general language corpus. The success of pre-trained models such as ELMo BIBREF16, GPT-2 BIBREF17, and BERT BIBREF18 for many natural language tasks, provides reasoning that pre-training is likely to produce better learning by our system. We use pre-training so that the system has some foundational knowledge of English before we train it on the domain-specific text of math word problems. However, the output is not natural language but algebraic expressions, which is likely to limit the effectiveness of such pre-training. Approach ::: Data We work with four individual datasets. The datasets contain addition, subtraction, multiplication, and division word problems. AI2 BIBREF2. AI2 is a collection of 395 addition and subtraction problems, containing numeric values, where some may not be relevant to the question. CC BIBREF19. The Common Core dataset contains 600 2-step questions. The Cognitive Computation Group at the University of Pennsylvania gathered these questions. IL BIBREF4. The Illinois dataset contains 562 1-step algebra word questions. The Cognitive Computation Group compiled these questions also. MAWPS BIBREF20. MAWPS is a relatively large collection, primarily from other MWP datasets. We use 2,373 of 3,915 MWPs from this set. The problems not used were more complex problems that generate systems of equations. We exclude such problems because generating systems of equations is not our focus. We take a randomly sampled 95% of examples from each dataset for training. From each dataset, MWPs not included in training make up the testing data used when generating our results. Training and testing are repeated three times, and reported results are an average of the three outcomes. Approach ::: Representation Conversion We take a simple approach to convert infix expressions found in the MWPs to the other two representations. Two stacks are filled by iterating through string characters, one with operators found in the equation and the other with the operands. From these stacks, we form a binary tree structure. Traversing an expression tree in pre-order results in a prefix conversion. Post-order traversal gives us a postfix expression. Three versions of our training and testing data are created to correspond to each type of expression. By training on different representations, we expect our test results to change. Approach ::: Pre-training We pre-train half of our networks to endow them with a foundational knowledge of English. Pre-training models on significant-sized language corpora have been a common approach recently. We explore the pre-training approach using a general English corpus because the language of MWPs is regular English, interspersed with numerical values. Ideally, the corpus for pre-training should be a very general and comprehensive corpus like an English Wikipedia dump or many gigabytes of human-generated text scraped from the internet like GPT-2 BIBREF21 used. However, in this paper, we want to perform experiments to see if pre-training with a smaller corpus can help. In particular, for this task, we use the IMDb Movie Reviews dataset BIBREF22. This set contains 314,041 unique sentences. Since movie reviewers wrote this data, it is a reference to natural language not related to arithmetic. Training on a much bigger and general corpus may make the language model stronger, but we leave this for future work. We compare pre-trained models to non-pre-trained models to observe performance differences. Our pre-trained models are trained in an unsupervised fashion to improve the encodings of our fine-tuned solvers. In the pre-training process, we use sentences from the IMDb reviews with a target output of an empty string. We leave the input unlabelled, which focuses the network on adjusting encodings while providing unbiased decoding when we later change from IMDb English text to MWP-Data. Approach ::: Method: Training and Testing The input sequence is a natural language specification of an arithmetic word problem. The MWP questions and equations have been encoded using the subword text encoder provided by the TensorFlow Datasets library. The output is an expression in prefix, infix, or postfix notation, which then can be manipulated further and solved to obtain a final answer. All examples in the datasets contain numbers, some of which are unique or rare in the corpus. Rare terms are adverse for generalization since the network is unlikely to form good representations for them. As a remedy to this issue, our networks do not consider any relevant numbers during training. Before the networks attempt any translation, we pre-process each question and expression by a number mapping algorithm. This algorithm replaces each numeric value with a corresponding identifier (e.g., $\langle n1 \rangle $, $\langle n2 \rangle $, etc.), and remembers the necessary mapping. We expect that this approach may significantly improve how networks interpret each question. When translating, the numbers in the original question are tagged and cached. From the encoded English and tags, a predicted sequence resembling an expression presents itself as output. Since each network's learned output resembles an arithmetic expression (e.g., $\langle n1 \rangle + \langle n2 \rangle * \langle n3 \rangle $), we use the cached tag mapping to replace the tags with the corresponding numbers and return a final mathematical expression. Three representation models are trained and tested separately: Prefix-Transformer, Postfix-Transformer, and Infix-Transformer. For each experiment, we use representation-specific Transformer architectures. Each model uses the Adam optimizer with $beta_1=0.95$ and $beta_2=0.99$ with a standard epsilon of $1 \times e^{-9}$. The learning rate is reduced automatically in each training session as the loss decreases. Throughout the training, each model respects a 10% dropout rate. We employ a batch size of 128 for all training. Each model is trained on MWP data for 300 iterations before testing. The networks are trained on a machine using 1 Nvidia 1080 Ti graphics processing unit (GPU). We compare medium-sized, small, and minimal networks to show if network size can be reduced to increase training and testing efficiency while retaining high accuracy. Networks over six layers have shown to be non-effective for this task. We tried many configurations of our network models, but report results with only three configurations of Transformers. Transformer Type 1: This network is a small to medium-sized network consisting of 4 Transformer layers. Each layer utilizes 8 attention heads with a depth of 512 and a feed-forward depth of 1024. Transformer Type 2: The second model is small in size, using 2 Transformer layers. The layers utilize 8 attention heads with a depth of 256 and a feed-forward depth of 1024. Transformer Type 3: The third type of model is minimal, using only 1 Transformer layer. This network utilizes 8 attention heads with a depth of 256 and a feed-forward depth of 512. Approach ::: Method: Training and Testing ::: Objective Function We calculate the loss in training according to a mean of the sparse categorical cross-entropy formula. Sparse categorical cross-entropy BIBREF23 is used for identifying classes from a feature set, which assumes a large target classification set. Evaluation between the possible translation classes (all vocabulary subword tokens) and the produced class (predicted token) is the metric of performance here. During each evaluation, target terms are masked, predicted, and then compared to the masked (known) value. We adjust the model's loss according to the mean of the translation accuracy after predicting every determined subword in a translation. where $K = |Translation \; Classes|$, $J = |Translation|$, and $I$ is the number of examples. Approach ::: Method: Training and Testing ::: Experiment 1: Representation Some of the problems encountered by prior approaches seem to be attributable to the use of infix notation. In this experiment, we compare translation BLEU-2 scores to spot the differences in representation interpretability. Traditionally, a BLEU score is a metric of translation quality BIBREF24. Our presented BLEU scores represent an average of scores a given model received over each of the target test sets. We use a standard bi-gram weight to show how accurate translations are within a window of two adjacent terms. After testing translations, we calculate an average BLEU-2 score per test set, which is related to the success over that data. An average of the scores for each dataset become the presented value. where $N$ is the number of test datasets, which is 4. Approach ::: Method: Training and Testing ::: Experiment 2: State-of-the-art This experiment compares our networks to recent previous work. We count a given test score by a simple “correct versus incorrect" method. The answer to an expression directly ties to all of the translation terms being correct, which is why we do not consider partial precision. We compare average accuracies over 3 test trials on different randomly sampled test sets from each MWP dataset. This calculation more accurately depicts the generalization of our networks. Approach ::: Method: Training and Testing ::: Effect of Pre-training We also explore the effect of language pre-training, as discussed earlier. This training occurs over 30 iterations, at the start of the two experiments, to introduce a good level of language understanding before training on the MWP data. The same Transformer architectures are also trained solely on the MWP data. We calculate the reported results as: where $R$ is the number of test repetitions, which is 3; $N$ is the number of test datasets, which is 4; $P$ is the number of MWPs, and $C$ is the number of correct equation translations. Results We now present the results of our various experiments. We compare the three representations of target equations and three architectures of the Transformer model in each test. Results of Experiment 1 are given in Table TABREF21. For clarity, the number in parentheses in front of a row is the Transformer type. By using BLEU scores, we assess the translation capability of each network. This test displays how networks transform different math representations to a character summary level. We compare by average BLEU-2 accuracy among our tests in the Average column of Table TABREF21 to communicate these translation differences. To make it easier to understand the results, Table TABREF22 provides a summary of Table TABREF21. Looking at Tables TABREF21 and TABREF22, we note that both the prefix and postfix representations of our target language perform better than the generally used infix notation. The non-pre-trained models perform slightly better than the pre-trained models, and the small or Type 2 models perform slightly better than the minimal-sized and medium-sized Transformer models. The non-pre-trained type 2 prefix Transformer arrangement produced the most consistent translations. Table TABREF23 provides detailed results of Experiment 2. The numbers are absolute accuracies, i.e., they correspond to cases where the arithmetic expression generated is 100% correct, leading to the correct numeric answer. Results by BIBREF1, BIBREF2, BIBREF4, BIBREF5 are sparse but indicate the scale of success compared to recent past approaches. Prefix, postfix, and infix representations in Table TABREF23 show that network capabilities are changed by how teachable the target data is. The values in the last column of Table TABREF23 are summarized in Table TABREF24. How the models compare with respect to accuracy closely resembles the comparison of BLEU scores, presented earlier. Thus, BLEU scores seem to correlate well with accuracy values in our case. While our networks fell short of BIBREF1 AI2 testing accuracy, we present state-of-the-art results for the remaining three datasets. The AI2 dataset is tricky because it has numeric values in the word descriptions that are extraneous or irrelevant to the actual computation, whereas the other datasets have only relevant numeric values. The type 2 postfix Transformer received the highest testing average of 87.2%. Our attempt at language pre-training fell short of our expectations in all but one tested dataset. We had hoped that more stable language understanding would improve results in general. As previously mentioned, using more general and comprehensive corpora of language could help grow semantic ability. Results ::: Analysis All of the network configurations used were very successful for our task. The prefix representation overall provides the most stable network performance. To display the capability of our most successful model (type 2 postfix Transformer), we present some outputs of the network in Figure FIGREF26. The models respect the syntax of math expressions, even when incorrect. For the majority of questions, our translators were able to determine operators based solely on the context of language. Our pre-training was unsuccessful in improving accuracy, even when applied to networks larger than those reported. We may need to use more inclusive language, or pre-train on very math specific texts to be successful. Our results support our thesis of infix limitation. Results ::: Analysis ::: Error Analysis Our system, while performing above standard, could still benefit from some improvements. One issue originates from the algorithmic pre-processing of our questions and expressions. In Figure FIGREF27 we show an example of one such issue. The excerpt comes from a type 3 non-pre-trained Transformer test. The example shows an overlooked identifier, $\langle n1 \rangle $. The issue is attributed to the identifier algorithm only considering numbers in the problem. Observe in the question that the word “eight" is the number we expect to relate to $\langle n2 \rangle $. Our identifying algorithm could be improved by considering such number words and performing conversion to a numerical value. If our algorithm performed as expected, the identifier $\langle n1 \rangle $ relates with 4 (the first occurring number in the question) and $\langle n2 \rangle $ with 8 (the converted number word appearing second in the question). The overall translation was incorrect whether or not our algorithm was successful, but it is essential to analyze problems like these that may result in future improvements. Had all questions been tagged correctly, our performance would have likely improved. Conclusions and Future Work In this paper, we have shown that the use of Transformer networks improves automatic math word problem-solving. We have also shown that the use of postfix target expressions performs better than the other two expression formats. Our improvements are well-motivated but straightforward and easy to use, demonstrating that the well-acclaimed Transformer architecture for language processing can handle MWPs well, obviating the need to build specialized neural architectures for this task. Extensive pre-training over much larger corpora of language has extended the capabilities of many neural approaches. For example, networks like BERT BIBREF18, trained extensively on data from Wikipedia, perform relatively better in many tasks. Pre-training on a much larger corpus remains an extension we would like to try. We want to work with more complex MWP datasets. Our datasets contain basic arithmetic expressions of +, -, * and /, and only up to 3 of them. For example, datasets such as Dolphin18k BIBREF25, consisting of web-answered questions from Yahoo! Answers, require a wider variety of arithmetic operators to be understood by the system. We have noticed that the presence of irrelevant numbers in the sentences for MWPs limits our performance. We can think of such numbers as a sort of adversarial threat to an MWP solver that stress-test it. It may be interesting to explore how to keep a network's performance high, even in such cases. With a hope to further advance this area of research and heighten interests, all of the code and data used is available on GitHub. Acknowledgement The National Science Foundation supports the work reported in this paper under Grant No. 1659788. Any opinions, findings any conclusions or recommendations expressed in this work are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

BLEU-2, average accuracies over 3 test trials on different randomly sampled test sets

095888f6e10080a958d9cd3f779a339498f3a109

095888f6e10080a958d9cd3f779a339498f3a109_0

Q: What datasets do they use? Text: Introduction Students are exposed to simple arithmetic word problems starting in elementary school, and most become proficient in solving them at a young age. Automatic solvers of such problems could potentially help educators, as well as become an integral part of general question answering services. However, it has been challenging to write programs to solve even such elementary school level problems well. Solving a math word problem (MWP) starts with one or more sentences describing a transactional situation to be understood. The sentences are processed to produce an arithmetic expression, which is evaluated to provide an answer. Recent neural approaches to solving arithmetic word problems have used various flavors of recurrent neural networks (RNN) as well as reinforcement learning. Such methods have had difficulty achieving a high level of generalization. Often, systems extract the relevant numbers successfully but misplace them in the generated expressions. More problematic, they get the arithmetic operations wrong. The use of infix notation also requires pairs of parentheses to be placed and balanced correctly, bracketing the right numbers. There have been problems with parentheses placement as well. Correctly extracting the numbers in the problem is necessary. Figure FIGREF1 gives examples of some infix representations that a machine learning solver can potentially produce from a simple word problem using the correct numbers. Of the expressions shown, only the first one is correct. After carefully observing expressions that actual problem solvers have generated, we want to explore if the use of infix notation may itself be a part of the problem because it requires the generation of additional characters, the open and close parentheses, which must be balanced and placed correctly. The actual numbers appearing in MWPs vary widely from problem to problem. Real numbers take any conceivable value, making it almost impossible for a neural network to learn representations for them. As a result, trained programs sometimes generate expressions that have seemingly random numbers. For example, in some runs, a trained program could generate a potentially inexplicable expression such as $(25.01 - 4) * 9$ for the problem given in Figure FIGREF1, with one or more numbers not in the problem sentences. We hypothesize that replacing the numbers in the problem statement with generic tags like $\rm \langle n1 \rangle $, $\rm \langle n2 \rangle $, and $\rm \langle n3 \rangle $ and saving their values as a pre-processing step, does not take away from the generality of the solution, but suppresses the problem of fertility in number generation leading to the introduction of numbers not present in the question sentences. Another idea we want to test is whether a neural network which has been pre-trained to acquire language knowledge is better able to “understand" the problem sentences. Pre-training with a large amount of arithmetic-related text is likely to help develop such knowledge, but due to the lack of large such focused corpora, we want to test whether pre-training with a sufficient general corpus is beneficial. In this paper, we use the Transformer model BIBREF0 to solve arithmetic word problems as a particular case of machine translation from text to the language of arithmetic expressions. Transformers in various configurations have become a staple of NLP in the past two years. Past neural approaches did not treat this problem as pure translation like we do, and additionally, these approaches usually augmented the neural architectures with various external modules such as parse trees or used deep reinforcement learning, which we do not do. In this paper, we demonstrate that Transformers can be used to solve MWPs successfully with the simple adjustments we describe above. We compare performance on four individual datasets. In particular, we show that our translation-based approach outperforms state-of-the-art results reported by BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5 by a large margin on three of four datasets tested. On average, our best neural architecture outperforms previous results by almost 10%, although our approach is conceptually more straightforward. We organize our paper as follows. The second section presents related work. Then, we discuss our approach. We follow by an analysis of experimental results and compare them to those of other recent approaches. We also discuss our successes and shortcomings. Finally, we share our concluding thoughts and end with our direction for future work. Related Work Past strategies have used rules and templates to match sentences to arithmetic expressions. Some such approaches seemed to solve problems impressively within a narrow domain, but performed poorly when out of domain, lacking generality BIBREF6, BIBREF7, BIBREF8, BIBREF9. Kushman et al. BIBREF3 used feature extraction and template-based categorization by representing equations as expression forests and finding a near match. Such methods required human intervention in the form of feature engineering and development of templates and rules, which is not desirable for expandability and adaptability. Hosseini et al. BIBREF2 performed statistical similarity analysis to obtain acceptable results, but did not perform well with texts that were dissimilar to training examples. Existing approaches have used various forms of auxiliary information. Hosseini et al. BIBREF2 used verb categorization to identify important mathematical cues and contexts. Mitra and Baral BIBREF10 used predefined formulas to assist in matching. Koncel-Kedziorski et al. BIBREF11 parsed the input sentences, enumerated all parses, and learned to match, requiring expensive computations. Roy and Roth BIBREF12 performed searches for semantic trees over large spaces. Some recent approaches have transitioned to using neural networks. Semantic parsing takes advantage of RNN architectures to parse MWPs directly into equations or expressions in a math-specific language BIBREF9, BIBREF13. RNNs have shown promising results, but they have had difficulties balancing parenthesis, and also, sometimes incorrectly choose numbers when generating equations. Rehman et al. BIBREF14 used POS tagging and classification of equation templates to produce systems of equations from third-grade level MWPs. Most recently, Sun et al. BIBREF13 used a Bi-Directional LSTM architecture for math word problems. Huang et al. BIBREF15 used a deep reinforcement learning model to achieve character placement in both seen and novel equation templates. Wang et al. BIBREF1 also used deep reinforcement learning. Approach We view math word problem solving as a sequence-to-sequence translation problem. RNNs have excelled in sequence-to-sequence problems such as translation and question answering. The recent introduction of attention mechanisms has improved the performance of RNN models. Vaswani et al. BIBREF0 introduced the Transformer network, which uses stacks of attention layers instead of recurrence. Applications of Transformers have achieved state-of-the-art performance in many NLP tasks. We use this architecture to produce character sequences that are arithmetic expressions. The models we experiment with are easy and efficient to train, allowing us to test several configurations for a comprehensive comparison. We use several configurations of Transformer networks to learn the prefix, postfix, and infix notations of MWP equations independently. Prefix and postfix representations of equations do not contain parentheses, which has been a source of confusion in some approaches. If the learned target sequences are simple, with fewer characters to generate, it is less likely to make mistakes during generation. Simple targets also may help the learning of the model to be more robust. Experimenting with all three representations for equivalent expressions may help us discover which one works best. We train on standard datasets, which are readily available and commonly used. Our method considers the translation of English text to simple algebraic expressions. After performing experiments by training directly on math word problem corpora, we perform a different set of experiments by pre-training on a general language corpus. The success of pre-trained models such as ELMo BIBREF16, GPT-2 BIBREF17, and BERT BIBREF18 for many natural language tasks, provides reasoning that pre-training is likely to produce better learning by our system. We use pre-training so that the system has some foundational knowledge of English before we train it on the domain-specific text of math word problems. However, the output is not natural language but algebraic expressions, which is likely to limit the effectiveness of such pre-training. Approach ::: Data We work with four individual datasets. The datasets contain addition, subtraction, multiplication, and division word problems. AI2 BIBREF2. AI2 is a collection of 395 addition and subtraction problems, containing numeric values, where some may not be relevant to the question. CC BIBREF19. The Common Core dataset contains 600 2-step questions. The Cognitive Computation Group at the University of Pennsylvania gathered these questions. IL BIBREF4. The Illinois dataset contains 562 1-step algebra word questions. The Cognitive Computation Group compiled these questions also. MAWPS BIBREF20. MAWPS is a relatively large collection, primarily from other MWP datasets. We use 2,373 of 3,915 MWPs from this set. The problems not used were more complex problems that generate systems of equations. We exclude such problems because generating systems of equations is not our focus. We take a randomly sampled 95% of examples from each dataset for training. From each dataset, MWPs not included in training make up the testing data used when generating our results. Training and testing are repeated three times, and reported results are an average of the three outcomes. Approach ::: Representation Conversion We take a simple approach to convert infix expressions found in the MWPs to the other two representations. Two stacks are filled by iterating through string characters, one with operators found in the equation and the other with the operands. From these stacks, we form a binary tree structure. Traversing an expression tree in pre-order results in a prefix conversion. Post-order traversal gives us a postfix expression. Three versions of our training and testing data are created to correspond to each type of expression. By training on different representations, we expect our test results to change. Approach ::: Pre-training We pre-train half of our networks to endow them with a foundational knowledge of English. Pre-training models on significant-sized language corpora have been a common approach recently. We explore the pre-training approach using a general English corpus because the language of MWPs is regular English, interspersed with numerical values. Ideally, the corpus for pre-training should be a very general and comprehensive corpus like an English Wikipedia dump or many gigabytes of human-generated text scraped from the internet like GPT-2 BIBREF21 used. However, in this paper, we want to perform experiments to see if pre-training with a smaller corpus can help. In particular, for this task, we use the IMDb Movie Reviews dataset BIBREF22. This set contains 314,041 unique sentences. Since movie reviewers wrote this data, it is a reference to natural language not related to arithmetic. Training on a much bigger and general corpus may make the language model stronger, but we leave this for future work. We compare pre-trained models to non-pre-trained models to observe performance differences. Our pre-trained models are trained in an unsupervised fashion to improve the encodings of our fine-tuned solvers. In the pre-training process, we use sentences from the IMDb reviews with a target output of an empty string. We leave the input unlabelled, which focuses the network on adjusting encodings while providing unbiased decoding when we later change from IMDb English text to MWP-Data. Approach ::: Method: Training and Testing The input sequence is a natural language specification of an arithmetic word problem. The MWP questions and equations have been encoded using the subword text encoder provided by the TensorFlow Datasets library. The output is an expression in prefix, infix, or postfix notation, which then can be manipulated further and solved to obtain a final answer. All examples in the datasets contain numbers, some of which are unique or rare in the corpus. Rare terms are adverse for generalization since the network is unlikely to form good representations for them. As a remedy to this issue, our networks do not consider any relevant numbers during training. Before the networks attempt any translation, we pre-process each question and expression by a number mapping algorithm. This algorithm replaces each numeric value with a corresponding identifier (e.g., $\langle n1 \rangle $, $\langle n2 \rangle $, etc.), and remembers the necessary mapping. We expect that this approach may significantly improve how networks interpret each question. When translating, the numbers in the original question are tagged and cached. From the encoded English and tags, a predicted sequence resembling an expression presents itself as output. Since each network's learned output resembles an arithmetic expression (e.g., $\langle n1 \rangle + \langle n2 \rangle * \langle n3 \rangle $), we use the cached tag mapping to replace the tags with the corresponding numbers and return a final mathematical expression. Three representation models are trained and tested separately: Prefix-Transformer, Postfix-Transformer, and Infix-Transformer. For each experiment, we use representation-specific Transformer architectures. Each model uses the Adam optimizer with $beta_1=0.95$ and $beta_2=0.99$ with a standard epsilon of $1 \times e^{-9}$. The learning rate is reduced automatically in each training session as the loss decreases. Throughout the training, each model respects a 10% dropout rate. We employ a batch size of 128 for all training. Each model is trained on MWP data for 300 iterations before testing. The networks are trained on a machine using 1 Nvidia 1080 Ti graphics processing unit (GPU). We compare medium-sized, small, and minimal networks to show if network size can be reduced to increase training and testing efficiency while retaining high accuracy. Networks over six layers have shown to be non-effective for this task. We tried many configurations of our network models, but report results with only three configurations of Transformers. Transformer Type 1: This network is a small to medium-sized network consisting of 4 Transformer layers. Each layer utilizes 8 attention heads with a depth of 512 and a feed-forward depth of 1024. Transformer Type 2: The second model is small in size, using 2 Transformer layers. The layers utilize 8 attention heads with a depth of 256 and a feed-forward depth of 1024. Transformer Type 3: The third type of model is minimal, using only 1 Transformer layer. This network utilizes 8 attention heads with a depth of 256 and a feed-forward depth of 512. Approach ::: Method: Training and Testing ::: Objective Function We calculate the loss in training according to a mean of the sparse categorical cross-entropy formula. Sparse categorical cross-entropy BIBREF23 is used for identifying classes from a feature set, which assumes a large target classification set. Evaluation between the possible translation classes (all vocabulary subword tokens) and the produced class (predicted token) is the metric of performance here. During each evaluation, target terms are masked, predicted, and then compared to the masked (known) value. We adjust the model's loss according to the mean of the translation accuracy after predicting every determined subword in a translation. where $K = |Translation \; Classes|$, $J = |Translation|$, and $I$ is the number of examples. Approach ::: Method: Training and Testing ::: Experiment 1: Representation Some of the problems encountered by prior approaches seem to be attributable to the use of infix notation. In this experiment, we compare translation BLEU-2 scores to spot the differences in representation interpretability. Traditionally, a BLEU score is a metric of translation quality BIBREF24. Our presented BLEU scores represent an average of scores a given model received over each of the target test sets. We use a standard bi-gram weight to show how accurate translations are within a window of two adjacent terms. After testing translations, we calculate an average BLEU-2 score per test set, which is related to the success over that data. An average of the scores for each dataset become the presented value. where $N$ is the number of test datasets, which is 4. Approach ::: Method: Training and Testing ::: Experiment 2: State-of-the-art This experiment compares our networks to recent previous work. We count a given test score by a simple “correct versus incorrect" method. The answer to an expression directly ties to all of the translation terms being correct, which is why we do not consider partial precision. We compare average accuracies over 3 test trials on different randomly sampled test sets from each MWP dataset. This calculation more accurately depicts the generalization of our networks. Approach ::: Method: Training and Testing ::: Effect of Pre-training We also explore the effect of language pre-training, as discussed earlier. This training occurs over 30 iterations, at the start of the two experiments, to introduce a good level of language understanding before training on the MWP data. The same Transformer architectures are also trained solely on the MWP data. We calculate the reported results as: where $R$ is the number of test repetitions, which is 3; $N$ is the number of test datasets, which is 4; $P$ is the number of MWPs, and $C$ is the number of correct equation translations. Results We now present the results of our various experiments. We compare the three representations of target equations and three architectures of the Transformer model in each test. Results of Experiment 1 are given in Table TABREF21. For clarity, the number in parentheses in front of a row is the Transformer type. By using BLEU scores, we assess the translation capability of each network. This test displays how networks transform different math representations to a character summary level. We compare by average BLEU-2 accuracy among our tests in the Average column of Table TABREF21 to communicate these translation differences. To make it easier to understand the results, Table TABREF22 provides a summary of Table TABREF21. Looking at Tables TABREF21 and TABREF22, we note that both the prefix and postfix representations of our target language perform better than the generally used infix notation. The non-pre-trained models perform slightly better than the pre-trained models, and the small or Type 2 models perform slightly better than the minimal-sized and medium-sized Transformer models. The non-pre-trained type 2 prefix Transformer arrangement produced the most consistent translations. Table TABREF23 provides detailed results of Experiment 2. The numbers are absolute accuracies, i.e., they correspond to cases where the arithmetic expression generated is 100% correct, leading to the correct numeric answer. Results by BIBREF1, BIBREF2, BIBREF4, BIBREF5 are sparse but indicate the scale of success compared to recent past approaches. Prefix, postfix, and infix representations in Table TABREF23 show that network capabilities are changed by how teachable the target data is. The values in the last column of Table TABREF23 are summarized in Table TABREF24. How the models compare with respect to accuracy closely resembles the comparison of BLEU scores, presented earlier. Thus, BLEU scores seem to correlate well with accuracy values in our case. While our networks fell short of BIBREF1 AI2 testing accuracy, we present state-of-the-art results for the remaining three datasets. The AI2 dataset is tricky because it has numeric values in the word descriptions that are extraneous or irrelevant to the actual computation, whereas the other datasets have only relevant numeric values. The type 2 postfix Transformer received the highest testing average of 87.2%. Our attempt at language pre-training fell short of our expectations in all but one tested dataset. We had hoped that more stable language understanding would improve results in general. As previously mentioned, using more general and comprehensive corpora of language could help grow semantic ability. Results ::: Analysis All of the network configurations used were very successful for our task. The prefix representation overall provides the most stable network performance. To display the capability of our most successful model (type 2 postfix Transformer), we present some outputs of the network in Figure FIGREF26. The models respect the syntax of math expressions, even when incorrect. For the majority of questions, our translators were able to determine operators based solely on the context of language. Our pre-training was unsuccessful in improving accuracy, even when applied to networks larger than those reported. We may need to use more inclusive language, or pre-train on very math specific texts to be successful. Our results support our thesis of infix limitation. Results ::: Analysis ::: Error Analysis Our system, while performing above standard, could still benefit from some improvements. One issue originates from the algorithmic pre-processing of our questions and expressions. In Figure FIGREF27 we show an example of one such issue. The excerpt comes from a type 3 non-pre-trained Transformer test. The example shows an overlooked identifier, $\langle n1 \rangle $. The issue is attributed to the identifier algorithm only considering numbers in the problem. Observe in the question that the word “eight" is the number we expect to relate to $\langle n2 \rangle $. Our identifying algorithm could be improved by considering such number words and performing conversion to a numerical value. If our algorithm performed as expected, the identifier $\langle n1 \rangle $ relates with 4 (the first occurring number in the question) and $\langle n2 \rangle $ with 8 (the converted number word appearing second in the question). The overall translation was incorrect whether or not our algorithm was successful, but it is essential to analyze problems like these that may result in future improvements. Had all questions been tagged correctly, our performance would have likely improved. Conclusions and Future Work In this paper, we have shown that the use of Transformer networks improves automatic math word problem-solving. We have also shown that the use of postfix target expressions performs better than the other two expression formats. Our improvements are well-motivated but straightforward and easy to use, demonstrating that the well-acclaimed Transformer architecture for language processing can handle MWPs well, obviating the need to build specialized neural architectures for this task. Extensive pre-training over much larger corpora of language has extended the capabilities of many neural approaches. For example, networks like BERT BIBREF18, trained extensively on data from Wikipedia, perform relatively better in many tasks. Pre-training on a much larger corpus remains an extension we would like to try. We want to work with more complex MWP datasets. Our datasets contain basic arithmetic expressions of +, -, * and /, and only up to 3 of them. For example, datasets such as Dolphin18k BIBREF25, consisting of web-answered questions from Yahoo! Answers, require a wider variety of arithmetic operators to be understood by the system. We have noticed that the presence of irrelevant numbers in the sentences for MWPs limits our performance. We can think of such numbers as a sort of adversarial threat to an MWP solver that stress-test it. It may be interesting to explore how to keep a network's performance high, even in such cases. With a hope to further advance this area of research and heighten interests, all of the code and data used is available on GitHub. Acknowledgement The National Science Foundation supports the work reported in this paper under Grant No. 1659788. Any opinions, findings any conclusions or recommendations expressed in this work are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

AI2 BIBREF2, CC BIBREF19, IL BIBREF4, MAWPS BIBREF20

57e783f00f594e08e43a31939aedb235c9d5a102

57e783f00f594e08e43a31939aedb235c9d5a102_0

Q: What evaluation metrics were used? Text: Introduction Voice-controlled virtual assistants (VVA) such as Siri and Alexa have experienced an exponential growth in terms of number of users and provided capabilities. They are used by millions for a variety of tasks including shopping, playing music, and even telling jokes. Arguably, their success is due in part to the emotional and personalized experience they provide. One important aspect of this emotional interaction is humor, a fundamental element of communication. Not only can it create in the user a sense of personality, but also be used as fallback technique for out-of-domain queries BIBREF0. Usually, a VVA's humorous responses are invoked by users with the phrase "Tell me a joke". In order to improve the joke experience and overall user satisfaction with a VVA, we propose to personalize the response to each request. To achieve this, a method should be able to recognize and evaluate humor, a challenging task that has been the focus of extensive work. Some authors have applied traditional NLP techniques BIBREF1, while others deep learning models BIBREF2. Moreover, BIBREF3 follows a semantic-based approach, while BIBREF4 and BIBREF5 tackle the challenge from a cognitive and linguistic perspective respectively. To this end, we have developed two methods. The first one is based on traditional NLP techniques. Although relatively simple, it is robust, scalable, and has low latency, a fundamental property for real-time VVA systems. The other approaches combine multi-task learning BIBREF6 and self-attentional networks BIBREF7 to obtain better results, at the cost of added complexity. Both BERT BIBREF8 and an adapted transformer BIBREF7 architecture are considered. This choice of architecture was motivated by the advantages it presents over traditional RNN and CNN models, including better performance BIBREF9, faster training/inference (important for real-time systems), and better sense disambiguation BIBREF10 (an important component of computational humor BIBREF3). The proposed models use binary classifiers to perform point-wise ranking, and therefore require a labelled dataset. To generate it, we explore two implicit user-feedback labelling strategies: five-minute reuse and one-day return. Online A/B testing is used to determine if these labelling strategies are suited to optimize the desired user-satisfaction metrics, and offline data to evaluated and compared the system's performance. Method ::: Labelling Strategies Generating labels for this VVA skill is challenging. Label generation through explicit user feedback is unavailable since asking users for feedback creates friction and degrade the user experience. In addition, available humor datasets such as BIBREF3, BIBREF11 only contain jokes and corresponding labels, but not the additional features we need to personalize the jokes. To overcome this difficulty, it is common to resort to implicit feedback. In particular, many VVA applications use interruptions as negative labels, the rationale being that unhappy users will stop the VVA. This strategy, however, is not suitable for our use-case since responses are short and users need to hear the entire joke to decide if it is funny. Instead, we explore two other implicit feedback labelling strategies: five-minute reuse and 1-day return. Five-minute reuse labels an instance positive if it was followed by a new joke request within five-minutes. Conversely, 1-day return marks as positive all joke requests that were followed by a new one within the following 1 to 25-hour interval. Both strategies assume that if a user returns, he is happy with the jokes. This is clearly an approximation, since a returning user might be overall satisfied with the experience, but not with all the jokes. The same is true for the implied negatives; the user might have been satisfied with some or all of the jokes. Therefore, these labels are noisy and only provide weak supervision to the models. Table TABREF2 shows an example of the labels' values for a set of joke requests from one user. Method ::: Features All models have access to the same raw features, which we conceptually separate into user, item and contextual features. Examples of features in each of these categories are shown in Table TABREF4. Some of these are used directly by the models, while others need to be pre-processed. The manner in which each model consumes them is explained next. Method ::: NLP-based: LR-Model To favor simplicity over accuracy, a logistic regression (LR) model is first proposed. Significant effort was put into finding expressive features. Categorical features are one-hot encoded and numerical ones are normalized. The raw Joke Text and Timestamp features require special treatment. The Joke Text is tokenized and the stop-words are removed. We can then compute computational humor features on the clean text such as sense combination BIBREF3 and ambiguity BIBREF12. In addition, since many jokes in our corpus are related to specific events (Christmas, etc), we check for keywords that relate the jokes to them. For example, if "Santa" is included, we infer it is a Christmas joke. Finally, pre-computed word embeddings with sub-word information are used to represent jokes by taking the average and maximum vectors over the token representations. Sub-word information is important when encoding jokes since many can contain out-of-vocabulary tokens. The joke's vector representations are also used to compute a summarized view of the user's past liked and disliked jokes. We consider that a user liked a joke when the assigned label is 1, an approximation given the noisy nature of the labels. The user's liked/disliked joke vectors are also combined with the candidate joke vector by taking the cosine similarity between them. For the raw Timestamp feature, we first extract simple time/date features such as month, day and isWeekend. We then compute binary features that mark if the timestamp occurred near one of the special events mentioned before. Some of these events occur the same day every year, while others change (for example, the Super Bowl). In addition, many events are country dependent. The timestamp's event features are combined with the joke's event features to allow the model to capture if an event-related joke occurs at the right time of the year. The LR classifier is trained on the processed features and one of the labels. The model's posterior probability is used to sort the candidates, which are chosen randomly from a pool of unheard jokes. Although useful (see Validation section), this model has several shortcomings. In particular, many of the used features require significant feature engineering and/or are country/language dependent, limiting the extensibility of the model. Method ::: Deep-Learning-based: DL-Models To overcome the LR-model's limitations, we propose the following model (see Figure FIGREF7). In the input layer, features are separated into context, item and user features. Unlike the LR-model, time and text features do not require extensive feature engineering. Instead, simple features (day, month and year) are extracted from the timestamp. After tokenization and stop-word removal, text features are passed through a pre-trained word embeding layer, and later, input into the joke encoder block. The basis of the joke encoder is a modified transformer. Firstly, only the encoder is needed. Moreover, since studies suggest that humor is subjective and conditioned on the user's context BIBREF13, we add an additional sub-layer in the transformer encoder that performs attention over the user's features. This sub-layer, inserted between the two typical transformer sub-layers at certain depths of the network, allows the encoder to adapt the representations of the jokes to different user contexts. Thus, the same joke can be encoded differently depending on the user's features. In practice, this additional sub-layer works like the normal self-attention sub-layer, except it creates its query matrix Q from the sub-layer below, and its K and V matrices from the user features. As an alternative, we also test encoding the jokes using a pre-trained BERT model. Regardless of the used encoder, we average the token representations to obtain a global encoding of the jokes. The same encoder is used to represent the item's (the joke to rank) and the user's (liked and disliked jokes) textual features through weight sharing, and the cosine similarity between both representations are computed. The processed features are then concatenated and passed through a final block of fully connected layers that contains the output layers. Since experiments determined (see Validation section) that both labeling strategies can improve the desired business metrics, instead of optimizing for only one of them, we take a multi-task learning approach. Thus, we have two softmax outputs. Finally, we use a loss function that considers label uncertainty, class imbalance and the different labeling functions. We start from the traditional cross-entropy loss for one labelling function. We then apply uniform label smoothing BIBREF14, which converts the one-hot-encoded label vectors into smoothed label vectors towards $0.5$: with $\epsilon $ a hyper-parameter. Label smoothing provides a way of considering the uncertainty on the labels by encouraging the model to be less confident. We have also experimented with other alternatives, including specialized losses such as BIBREF15. However, they did not produce a significant increase in performance in our tests. To further model the possible uncertainty in the feedback, we apply sample weights calculated using an exponential decay function on the time difference between the current and the following training instance of the same customer: where $w_i$ is the weight of sample $i$, $t_i$ is the time difference between instances $i$ and $i+1$ for the same user, and $a,b$ are hyper-parameters such that $a>0$ and $0<b<1$. The rationale behind these weights is the following. If for example, we consider labeling function 1, and a user asks for consecutive jokes, first within 10 seconds and later within 4.9 minutes, both instances are labeled as positive. However, we hypothesize that there is a lower chance that in the second case the user requested an additional joke because he liked the first one. In addition, class weights are applied to each sample to account for the natural class imbalance of the dataset. Finally, the total loss to be optimized is the weighted sum of the losses for each of the considered labeling functions: where $w_{l}$ are manually set weights for each label and $\mathcal {L}_{l}$ are the losses corresponding to each label, which include all the weights mentioned before. Validation A two-step validation was conducted for English-speaking customers. An initial A/B testing for the LR model in a production setting was performed to compare the labelling strategies. A second offline comparison of the models was conducted on historical data and a selected labelling strategy. One month of data and a subset of the customers was used (approx. eighty thousand). The sampled dataset presents a fraction of positive labels of approximately 0.5 for reuse and 0.2 for one-day return. Importantly, since this evaluation is done on a subset of users, the dataset characteristic's do not necessarily represent real production traffic. The joke corpus in this dataset contains thousands of unique jokes of different categories (sci-fi, sports, etc) and types (puns, limerick, etc). The dataset was split timewise into training/validation/test sets, and hyperparameters were optimized to maximize the AUC-ROC on the validation set. As a benchmark, we also consider two additional methods: a non-personalized popularity model and one that follows BIBREF16, replacing the transformer joke encoder with a CNN network (the specialized loss and other characteristics of the DL model are kept). Hyperparameters were optimized using grid-search for the LR-Model. Due to computational constraints, random search was instead used for the DL-Model. In both cases, hyperparameters are selected to optimize the AUC-ROC on the validation set. Table TABREF11 lists some of the considered hyperparameter values and ranges for both models. The actual optimal values are sample specific. Validation ::: Online Results: A/B Testing Two treatment groups are considered, one per label. Users in the control group are presented jokes at random, without repetition. Several user-satisfaction metrics such as user interruption rate, reuse of this and other VVA skills, and number of active dialogs are monitored during the tests. The relative improvement/decline of these metrics is compared between the treatments and control, and between the treatments themselves. The statistical significance is measured when determining differences between the groups. Results show that the LR-based model consistently outperforms the heuristic method for both labeling strategies, significantly improving retention, dialogs and interruptions. These results suggest that models trained using either label can improve the VVA's joke experience. Validation ::: Offline Results One-day return was selected for the offline evaluation because models trained on it have a better AUC-ROC, and both labeling strategies were successful in the online validation. All results are expressed as relative change with respect to the popularity model. We start by evaluating the models using AUC-ROC. As seen in Table TABREF14, the transformer-based models, and in particular our custom architecture, outperform all other approaches. Similar conclusions can be reached regarding overall accuracy. However, given the class imbalance, accuracy is not necessarily the best metric to consider. In addition, to better understand the effect to the original transformer architecture, we present the performance of the model with and without the modified loss and special attention sub-layer (see Table TABREF14). Results suggest both modifications have a positive impact on the performance. Finally, to further evaluate the ranking capabilities of the proposed methods, we use top-1 accuracy. Additional positions in the ranking are not considered because only the top ranked joke is presented to the customer. Results show that the DL based models outperform the other systems, with a relative change in top-1 accuracy of 1.4 for DL-BERT and 0.43 for DL-T, compared with 0.14 for the LR method. Results show that the proposed methods provide different compromises between accuracy, scalability and robustness. On one hand, the relatively good performance of the LR model with engineered features provides a strong baseline both in terms of accuracy and training/inference performance, at the cost of being difficult to extend to new countries and languages. On the other hand, DL based methods give a significant accuracy gain and require no feature engineering, which facilitates the expansion of the joke experience to new markets and languages. This comes at a cost of added complexity if deployed in production. In addition, given the size of the BERT model (340M parameters), real-time inference using DL-BERT becomes problematic due to latency constraints. In this regard, the DL-T model could be a good compromise since its complexity can be adapted, and it provides good overall accuracy. Conclusions and Future Work This paper describes systems to personalize a VVA's joke experience using NLP and deep-learning techniques that provide different compromises between accuracy, scalability and robustness. Implicit feedback signals are used to generate weak labels and provide supervision to the ranking models. Results on production data show that models trained on any of the considered labels present a positive real-world impact on user satisfaction, and that the deep learning approaches can potentially improve the joke skill with respect to the other considered methods. In the future, we would like to compare all methods in A/B testing, and to extend the models to other languages.

AUC-ROC

9646fa1abbe3102a0364f84e0a55d107d45c97f0

9646fa1abbe3102a0364f84e0a55d107d45c97f0_0

Q: Where did the real production data come from? Text: Introduction Voice-controlled virtual assistants (VVA) such as Siri and Alexa have experienced an exponential growth in terms of number of users and provided capabilities. They are used by millions for a variety of tasks including shopping, playing music, and even telling jokes. Arguably, their success is due in part to the emotional and personalized experience they provide. One important aspect of this emotional interaction is humor, a fundamental element of communication. Not only can it create in the user a sense of personality, but also be used as fallback technique for out-of-domain queries BIBREF0. Usually, a VVA's humorous responses are invoked by users with the phrase "Tell me a joke". In order to improve the joke experience and overall user satisfaction with a VVA, we propose to personalize the response to each request. To achieve this, a method should be able to recognize and evaluate humor, a challenging task that has been the focus of extensive work. Some authors have applied traditional NLP techniques BIBREF1, while others deep learning models BIBREF2. Moreover, BIBREF3 follows a semantic-based approach, while BIBREF4 and BIBREF5 tackle the challenge from a cognitive and linguistic perspective respectively. To this end, we have developed two methods. The first one is based on traditional NLP techniques. Although relatively simple, it is robust, scalable, and has low latency, a fundamental property for real-time VVA systems. The other approaches combine multi-task learning BIBREF6 and self-attentional networks BIBREF7 to obtain better results, at the cost of added complexity. Both BERT BIBREF8 and an adapted transformer BIBREF7 architecture are considered. This choice of architecture was motivated by the advantages it presents over traditional RNN and CNN models, including better performance BIBREF9, faster training/inference (important for real-time systems), and better sense disambiguation BIBREF10 (an important component of computational humor BIBREF3). The proposed models use binary classifiers to perform point-wise ranking, and therefore require a labelled dataset. To generate it, we explore two implicit user-feedback labelling strategies: five-minute reuse and one-day return. Online A/B testing is used to determine if these labelling strategies are suited to optimize the desired user-satisfaction metrics, and offline data to evaluated and compared the system's performance. Method ::: Labelling Strategies Generating labels for this VVA skill is challenging. Label generation through explicit user feedback is unavailable since asking users for feedback creates friction and degrade the user experience. In addition, available humor datasets such as BIBREF3, BIBREF11 only contain jokes and corresponding labels, but not the additional features we need to personalize the jokes. To overcome this difficulty, it is common to resort to implicit feedback. In particular, many VVA applications use interruptions as negative labels, the rationale being that unhappy users will stop the VVA. This strategy, however, is not suitable for our use-case since responses are short and users need to hear the entire joke to decide if it is funny. Instead, we explore two other implicit feedback labelling strategies: five-minute reuse and 1-day return. Five-minute reuse labels an instance positive if it was followed by a new joke request within five-minutes. Conversely, 1-day return marks as positive all joke requests that were followed by a new one within the following 1 to 25-hour interval. Both strategies assume that if a user returns, he is happy with the jokes. This is clearly an approximation, since a returning user might be overall satisfied with the experience, but not with all the jokes. The same is true for the implied negatives; the user might have been satisfied with some or all of the jokes. Therefore, these labels are noisy and only provide weak supervision to the models. Table TABREF2 shows an example of the labels' values for a set of joke requests from one user. Method ::: Features All models have access to the same raw features, which we conceptually separate into user, item and contextual features. Examples of features in each of these categories are shown in Table TABREF4. Some of these are used directly by the models, while others need to be pre-processed. The manner in which each model consumes them is explained next. Method ::: NLP-based: LR-Model To favor simplicity over accuracy, a logistic regression (LR) model is first proposed. Significant effort was put into finding expressive features. Categorical features are one-hot encoded and numerical ones are normalized. The raw Joke Text and Timestamp features require special treatment. The Joke Text is tokenized and the stop-words are removed. We can then compute computational humor features on the clean text such as sense combination BIBREF3 and ambiguity BIBREF12. In addition, since many jokes in our corpus are related to specific events (Christmas, etc), we check for keywords that relate the jokes to them. For example, if "Santa" is included, we infer it is a Christmas joke. Finally, pre-computed word embeddings with sub-word information are used to represent jokes by taking the average and maximum vectors over the token representations. Sub-word information is important when encoding jokes since many can contain out-of-vocabulary tokens. The joke's vector representations are also used to compute a summarized view of the user's past liked and disliked jokes. We consider that a user liked a joke when the assigned label is 1, an approximation given the noisy nature of the labels. The user's liked/disliked joke vectors are also combined with the candidate joke vector by taking the cosine similarity between them. For the raw Timestamp feature, we first extract simple time/date features such as month, day and isWeekend. We then compute binary features that mark if the timestamp occurred near one of the special events mentioned before. Some of these events occur the same day every year, while others change (for example, the Super Bowl). In addition, many events are country dependent. The timestamp's event features are combined with the joke's event features to allow the model to capture if an event-related joke occurs at the right time of the year. The LR classifier is trained on the processed features and one of the labels. The model's posterior probability is used to sort the candidates, which are chosen randomly from a pool of unheard jokes. Although useful (see Validation section), this model has several shortcomings. In particular, many of the used features require significant feature engineering and/or are country/language dependent, limiting the extensibility of the model. Method ::: Deep-Learning-based: DL-Models To overcome the LR-model's limitations, we propose the following model (see Figure FIGREF7). In the input layer, features are separated into context, item and user features. Unlike the LR-model, time and text features do not require extensive feature engineering. Instead, simple features (day, month and year) are extracted from the timestamp. After tokenization and stop-word removal, text features are passed through a pre-trained word embeding layer, and later, input into the joke encoder block. The basis of the joke encoder is a modified transformer. Firstly, only the encoder is needed. Moreover, since studies suggest that humor is subjective and conditioned on the user's context BIBREF13, we add an additional sub-layer in the transformer encoder that performs attention over the user's features. This sub-layer, inserted between the two typical transformer sub-layers at certain depths of the network, allows the encoder to adapt the representations of the jokes to different user contexts. Thus, the same joke can be encoded differently depending on the user's features. In practice, this additional sub-layer works like the normal self-attention sub-layer, except it creates its query matrix Q from the sub-layer below, and its K and V matrices from the user features. As an alternative, we also test encoding the jokes using a pre-trained BERT model. Regardless of the used encoder, we average the token representations to obtain a global encoding of the jokes. The same encoder is used to represent the item's (the joke to rank) and the user's (liked and disliked jokes) textual features through weight sharing, and the cosine similarity between both representations are computed. The processed features are then concatenated and passed through a final block of fully connected layers that contains the output layers. Since experiments determined (see Validation section) that both labeling strategies can improve the desired business metrics, instead of optimizing for only one of them, we take a multi-task learning approach. Thus, we have two softmax outputs. Finally, we use a loss function that considers label uncertainty, class imbalance and the different labeling functions. We start from the traditional cross-entropy loss for one labelling function. We then apply uniform label smoothing BIBREF14, which converts the one-hot-encoded label vectors into smoothed label vectors towards $0.5$: with $\epsilon $ a hyper-parameter. Label smoothing provides a way of considering the uncertainty on the labels by encouraging the model to be less confident. We have also experimented with other alternatives, including specialized losses such as BIBREF15. However, they did not produce a significant increase in performance in our tests. To further model the possible uncertainty in the feedback, we apply sample weights calculated using an exponential decay function on the time difference between the current and the following training instance of the same customer: where $w_i$ is the weight of sample $i$, $t_i$ is the time difference between instances $i$ and $i+1$ for the same user, and $a,b$ are hyper-parameters such that $a>0$ and $0<b<1$. The rationale behind these weights is the following. If for example, we consider labeling function 1, and a user asks for consecutive jokes, first within 10 seconds and later within 4.9 minutes, both instances are labeled as positive. However, we hypothesize that there is a lower chance that in the second case the user requested an additional joke because he liked the first one. In addition, class weights are applied to each sample to account for the natural class imbalance of the dataset. Finally, the total loss to be optimized is the weighted sum of the losses for each of the considered labeling functions: where $w_{l}$ are manually set weights for each label and $\mathcal {L}_{l}$ are the losses corresponding to each label, which include all the weights mentioned before. Validation A two-step validation was conducted for English-speaking customers. An initial A/B testing for the LR model in a production setting was performed to compare the labelling strategies. A second offline comparison of the models was conducted on historical data and a selected labelling strategy. One month of data and a subset of the customers was used (approx. eighty thousand). The sampled dataset presents a fraction of positive labels of approximately 0.5 for reuse and 0.2 for one-day return. Importantly, since this evaluation is done on a subset of users, the dataset characteristic's do not necessarily represent real production traffic. The joke corpus in this dataset contains thousands of unique jokes of different categories (sci-fi, sports, etc) and types (puns, limerick, etc). The dataset was split timewise into training/validation/test sets, and hyperparameters were optimized to maximize the AUC-ROC on the validation set. As a benchmark, we also consider two additional methods: a non-personalized popularity model and one that follows BIBREF16, replacing the transformer joke encoder with a CNN network (the specialized loss and other characteristics of the DL model are kept). Hyperparameters were optimized using grid-search for the LR-Model. Due to computational constraints, random search was instead used for the DL-Model. In both cases, hyperparameters are selected to optimize the AUC-ROC on the validation set. Table TABREF11 lists some of the considered hyperparameter values and ranges for both models. The actual optimal values are sample specific. Validation ::: Online Results: A/B Testing Two treatment groups are considered, one per label. Users in the control group are presented jokes at random, without repetition. Several user-satisfaction metrics such as user interruption rate, reuse of this and other VVA skills, and number of active dialogs are monitored during the tests. The relative improvement/decline of these metrics is compared between the treatments and control, and between the treatments themselves. The statistical significance is measured when determining differences between the groups. Results show that the LR-based model consistently outperforms the heuristic method for both labeling strategies, significantly improving retention, dialogs and interruptions. These results suggest that models trained using either label can improve the VVA's joke experience. Validation ::: Offline Results One-day return was selected for the offline evaluation because models trained on it have a better AUC-ROC, and both labeling strategies were successful in the online validation. All results are expressed as relative change with respect to the popularity model. We start by evaluating the models using AUC-ROC. As seen in Table TABREF14, the transformer-based models, and in particular our custom architecture, outperform all other approaches. Similar conclusions can be reached regarding overall accuracy. However, given the class imbalance, accuracy is not necessarily the best metric to consider. In addition, to better understand the effect to the original transformer architecture, we present the performance of the model with and without the modified loss and special attention sub-layer (see Table TABREF14). Results suggest both modifications have a positive impact on the performance. Finally, to further evaluate the ranking capabilities of the proposed methods, we use top-1 accuracy. Additional positions in the ranking are not considered because only the top ranked joke is presented to the customer. Results show that the DL based models outperform the other systems, with a relative change in top-1 accuracy of 1.4 for DL-BERT and 0.43 for DL-T, compared with 0.14 for the LR method. Results show that the proposed methods provide different compromises between accuracy, scalability and robustness. On one hand, the relatively good performance of the LR model with engineered features provides a strong baseline both in terms of accuracy and training/inference performance, at the cost of being difficult to extend to new countries and languages. On the other hand, DL based methods give a significant accuracy gain and require no feature engineering, which facilitates the expansion of the joke experience to new markets and languages. This comes at a cost of added complexity if deployed in production. In addition, given the size of the BERT model (340M parameters), real-time inference using DL-BERT becomes problematic due to latency constraints. In this regard, the DL-T model could be a good compromise since its complexity can be adapted, and it provides good overall accuracy. Conclusions and Future Work This paper describes systems to personalize a VVA's joke experience using NLP and deep-learning techniques that provide different compromises between accuracy, scalability and robustness. Implicit feedback signals are used to generate weak labels and provide supervision to the ranking models. Results on production data show that models trained on any of the considered labels present a positive real-world impact on user satisfaction, and that the deep learning approaches can potentially improve the joke skill with respect to the other considered methods. In the future, we would like to compare all methods in A/B testing, and to extend the models to other languages.

jokes of different categories (sci-fi, sports, etc) and types (puns, limerick, etc)

29983f4bc8a5513a198755e474361deee93d4ab6

29983f4bc8a5513a198755e474361deee93d4ab6_0

Q: What feedback labels are used? Text: Introduction Voice-controlled virtual assistants (VVA) such as Siri and Alexa have experienced an exponential growth in terms of number of users and provided capabilities. They are used by millions for a variety of tasks including shopping, playing music, and even telling jokes. Arguably, their success is due in part to the emotional and personalized experience they provide. One important aspect of this emotional interaction is humor, a fundamental element of communication. Not only can it create in the user a sense of personality, but also be used as fallback technique for out-of-domain queries BIBREF0. Usually, a VVA's humorous responses are invoked by users with the phrase "Tell me a joke". In order to improve the joke experience and overall user satisfaction with a VVA, we propose to personalize the response to each request. To achieve this, a method should be able to recognize and evaluate humor, a challenging task that has been the focus of extensive work. Some authors have applied traditional NLP techniques BIBREF1, while others deep learning models BIBREF2. Moreover, BIBREF3 follows a semantic-based approach, while BIBREF4 and BIBREF5 tackle the challenge from a cognitive and linguistic perspective respectively. To this end, we have developed two methods. The first one is based on traditional NLP techniques. Although relatively simple, it is robust, scalable, and has low latency, a fundamental property for real-time VVA systems. The other approaches combine multi-task learning BIBREF6 and self-attentional networks BIBREF7 to obtain better results, at the cost of added complexity. Both BERT BIBREF8 and an adapted transformer BIBREF7 architecture are considered. This choice of architecture was motivated by the advantages it presents over traditional RNN and CNN models, including better performance BIBREF9, faster training/inference (important for real-time systems), and better sense disambiguation BIBREF10 (an important component of computational humor BIBREF3). The proposed models use binary classifiers to perform point-wise ranking, and therefore require a labelled dataset. To generate it, we explore two implicit user-feedback labelling strategies: five-minute reuse and one-day return. Online A/B testing is used to determine if these labelling strategies are suited to optimize the desired user-satisfaction metrics, and offline data to evaluated and compared the system's performance. Method ::: Labelling Strategies Generating labels for this VVA skill is challenging. Label generation through explicit user feedback is unavailable since asking users for feedback creates friction and degrade the user experience. In addition, available humor datasets such as BIBREF3, BIBREF11 only contain jokes and corresponding labels, but not the additional features we need to personalize the jokes. To overcome this difficulty, it is common to resort to implicit feedback. In particular, many VVA applications use interruptions as negative labels, the rationale being that unhappy users will stop the VVA. This strategy, however, is not suitable for our use-case since responses are short and users need to hear the entire joke to decide if it is funny. Instead, we explore two other implicit feedback labelling strategies: five-minute reuse and 1-day return. Five-minute reuse labels an instance positive if it was followed by a new joke request within five-minutes. Conversely, 1-day return marks as positive all joke requests that were followed by a new one within the following 1 to 25-hour interval. Both strategies assume that if a user returns, he is happy with the jokes. This is clearly an approximation, since a returning user might be overall satisfied with the experience, but not with all the jokes. The same is true for the implied negatives; the user might have been satisfied with some or all of the jokes. Therefore, these labels are noisy and only provide weak supervision to the models. Table TABREF2 shows an example of the labels' values for a set of joke requests from one user. Method ::: Features All models have access to the same raw features, which we conceptually separate into user, item and contextual features. Examples of features in each of these categories are shown in Table TABREF4. Some of these are used directly by the models, while others need to be pre-processed. The manner in which each model consumes them is explained next. Method ::: NLP-based: LR-Model To favor simplicity over accuracy, a logistic regression (LR) model is first proposed. Significant effort was put into finding expressive features. Categorical features are one-hot encoded and numerical ones are normalized. The raw Joke Text and Timestamp features require special treatment. The Joke Text is tokenized and the stop-words are removed. We can then compute computational humor features on the clean text such as sense combination BIBREF3 and ambiguity BIBREF12. In addition, since many jokes in our corpus are related to specific events (Christmas, etc), we check for keywords that relate the jokes to them. For example, if "Santa" is included, we infer it is a Christmas joke. Finally, pre-computed word embeddings with sub-word information are used to represent jokes by taking the average and maximum vectors over the token representations. Sub-word information is important when encoding jokes since many can contain out-of-vocabulary tokens. The joke's vector representations are also used to compute a summarized view of the user's past liked and disliked jokes. We consider that a user liked a joke when the assigned label is 1, an approximation given the noisy nature of the labels. The user's liked/disliked joke vectors are also combined with the candidate joke vector by taking the cosine similarity between them. For the raw Timestamp feature, we first extract simple time/date features such as month, day and isWeekend. We then compute binary features that mark if the timestamp occurred near one of the special events mentioned before. Some of these events occur the same day every year, while others change (for example, the Super Bowl). In addition, many events are country dependent. The timestamp's event features are combined with the joke's event features to allow the model to capture if an event-related joke occurs at the right time of the year. The LR classifier is trained on the processed features and one of the labels. The model's posterior probability is used to sort the candidates, which are chosen randomly from a pool of unheard jokes. Although useful (see Validation section), this model has several shortcomings. In particular, many of the used features require significant feature engineering and/or are country/language dependent, limiting the extensibility of the model. Method ::: Deep-Learning-based: DL-Models To overcome the LR-model's limitations, we propose the following model (see Figure FIGREF7). In the input layer, features are separated into context, item and user features. Unlike the LR-model, time and text features do not require extensive feature engineering. Instead, simple features (day, month and year) are extracted from the timestamp. After tokenization and stop-word removal, text features are passed through a pre-trained word embeding layer, and later, input into the joke encoder block. The basis of the joke encoder is a modified transformer. Firstly, only the encoder is needed. Moreover, since studies suggest that humor is subjective and conditioned on the user's context BIBREF13, we add an additional sub-layer in the transformer encoder that performs attention over the user's features. This sub-layer, inserted between the two typical transformer sub-layers at certain depths of the network, allows the encoder to adapt the representations of the jokes to different user contexts. Thus, the same joke can be encoded differently depending on the user's features. In practice, this additional sub-layer works like the normal self-attention sub-layer, except it creates its query matrix Q from the sub-layer below, and its K and V matrices from the user features. As an alternative, we also test encoding the jokes using a pre-trained BERT model. Regardless of the used encoder, we average the token representations to obtain a global encoding of the jokes. The same encoder is used to represent the item's (the joke to rank) and the user's (liked and disliked jokes) textual features through weight sharing, and the cosine similarity between both representations are computed. The processed features are then concatenated and passed through a final block of fully connected layers that contains the output layers. Since experiments determined (see Validation section) that both labeling strategies can improve the desired business metrics, instead of optimizing for only one of them, we take a multi-task learning approach. Thus, we have two softmax outputs. Finally, we use a loss function that considers label uncertainty, class imbalance and the different labeling functions. We start from the traditional cross-entropy loss for one labelling function. We then apply uniform label smoothing BIBREF14, which converts the one-hot-encoded label vectors into smoothed label vectors towards $0.5$: with $\epsilon $ a hyper-parameter. Label smoothing provides a way of considering the uncertainty on the labels by encouraging the model to be less confident. We have also experimented with other alternatives, including specialized losses such as BIBREF15. However, they did not produce a significant increase in performance in our tests. To further model the possible uncertainty in the feedback, we apply sample weights calculated using an exponential decay function on the time difference between the current and the following training instance of the same customer: where $w_i$ is the weight of sample $i$, $t_i$ is the time difference between instances $i$ and $i+1$ for the same user, and $a,b$ are hyper-parameters such that $a>0$ and $0<b<1$. The rationale behind these weights is the following. If for example, we consider labeling function 1, and a user asks for consecutive jokes, first within 10 seconds and later within 4.9 minutes, both instances are labeled as positive. However, we hypothesize that there is a lower chance that in the second case the user requested an additional joke because he liked the first one. In addition, class weights are applied to each sample to account for the natural class imbalance of the dataset. Finally, the total loss to be optimized is the weighted sum of the losses for each of the considered labeling functions: where $w_{l}$ are manually set weights for each label and $\mathcal {L}_{l}$ are the losses corresponding to each label, which include all the weights mentioned before. Validation A two-step validation was conducted for English-speaking customers. An initial A/B testing for the LR model in a production setting was performed to compare the labelling strategies. A second offline comparison of the models was conducted on historical data and a selected labelling strategy. One month of data and a subset of the customers was used (approx. eighty thousand). The sampled dataset presents a fraction of positive labels of approximately 0.5 for reuse and 0.2 for one-day return. Importantly, since this evaluation is done on a subset of users, the dataset characteristic's do not necessarily represent real production traffic. The joke corpus in this dataset contains thousands of unique jokes of different categories (sci-fi, sports, etc) and types (puns, limerick, etc). The dataset was split timewise into training/validation/test sets, and hyperparameters were optimized to maximize the AUC-ROC on the validation set. As a benchmark, we also consider two additional methods: a non-personalized popularity model and one that follows BIBREF16, replacing the transformer joke encoder with a CNN network (the specialized loss and other characteristics of the DL model are kept). Hyperparameters were optimized using grid-search for the LR-Model. Due to computational constraints, random search was instead used for the DL-Model. In both cases, hyperparameters are selected to optimize the AUC-ROC on the validation set. Table TABREF11 lists some of the considered hyperparameter values and ranges for both models. The actual optimal values are sample specific. Validation ::: Online Results: A/B Testing Two treatment groups are considered, one per label. Users in the control group are presented jokes at random, without repetition. Several user-satisfaction metrics such as user interruption rate, reuse of this and other VVA skills, and number of active dialogs are monitored during the tests. The relative improvement/decline of these metrics is compared between the treatments and control, and between the treatments themselves. The statistical significance is measured when determining differences between the groups. Results show that the LR-based model consistently outperforms the heuristic method for both labeling strategies, significantly improving retention, dialogs and interruptions. These results suggest that models trained using either label can improve the VVA's joke experience. Validation ::: Offline Results One-day return was selected for the offline evaluation because models trained on it have a better AUC-ROC, and both labeling strategies were successful in the online validation. All results are expressed as relative change with respect to the popularity model. We start by evaluating the models using AUC-ROC. As seen in Table TABREF14, the transformer-based models, and in particular our custom architecture, outperform all other approaches. Similar conclusions can be reached regarding overall accuracy. However, given the class imbalance, accuracy is not necessarily the best metric to consider. In addition, to better understand the effect to the original transformer architecture, we present the performance of the model with and without the modified loss and special attention sub-layer (see Table TABREF14). Results suggest both modifications have a positive impact on the performance. Finally, to further evaluate the ranking capabilities of the proposed methods, we use top-1 accuracy. Additional positions in the ranking are not considered because only the top ranked joke is presented to the customer. Results show that the DL based models outperform the other systems, with a relative change in top-1 accuracy of 1.4 for DL-BERT and 0.43 for DL-T, compared with 0.14 for the LR method. Results show that the proposed methods provide different compromises between accuracy, scalability and robustness. On one hand, the relatively good performance of the LR model with engineered features provides a strong baseline both in terms of accuracy and training/inference performance, at the cost of being difficult to extend to new countries and languages. On the other hand, DL based methods give a significant accuracy gain and require no feature engineering, which facilitates the expansion of the joke experience to new markets and languages. This comes at a cost of added complexity if deployed in production. In addition, given the size of the BERT model (340M parameters), real-time inference using DL-BERT becomes problematic due to latency constraints. In this regard, the DL-T model could be a good compromise since its complexity can be adapted, and it provides good overall accuracy. Conclusions and Future Work This paper describes systems to personalize a VVA's joke experience using NLP and deep-learning techniques that provide different compromises between accuracy, scalability and robustness. Implicit feedback signals are used to generate weak labels and provide supervision to the ranking models. Results on production data show that models trained on any of the considered labels present a positive real-world impact on user satisfaction, and that the deep learning approaches can potentially improve the joke skill with respect to the other considered methods. In the future, we would like to compare all methods in A/B testing, and to extend the models to other languages.

five-minute reuse and one-day return

6c0f97807cd83a94a4d26040286c6f89c4a0f8e0

6c0f97807cd83a94a4d26040286c6f89c4a0f8e0_0

Q: What representations for textual documents do they use? Text: Introduction Over the past few years, the term big data has become an important key point for research into data mining and information retrieval. Through the years, the quantity of data managed across enterprises has evolved from a simple and imperceptible task to an extent to which it has become the central performance improvement problem. In other words, it evolved to be the next frontier for innovation, competition and productivity BIBREF0. Extracting knowledge from data is now a very competitive environment. Many companies process vast amounts of customer/user data in order to improve the quality of experience (QoE) of their customers. For instance, a typical use-case scenario would be a book seller that performs an automatic extraction of the content of the books a customer has bought, and subsequently extracts knowledge of what customers prefer to read. The knowledge extracted could then be used to recommend other books. Book recommending systems are typical examples where data mining techniques should be considered as the primary tool for making future decisions BIBREF1. KE from TDs is an essential field of research in data mining and it certainly requires techniques that are reliable and accurate in order to neutralize (or even eliminate) uncertainty in future decisions. Grouping TDs based on their content and mutual key information is referred to as clustering. Clustering is mostly performed with respect to a measure of similarity between TDs, which must be represented as vectors in a vector space beforehand BIBREF2. News aggregation engines can be considered as a typical representative where such techniques are extensively applied as a sub-field of natural language processing (NLP). In this paper we present a new technique for measuring similarity between TDs, represented in a vector space, based on SRCC - "a statistical measure of association between two things" BIBREF3, which in this case things refer to TDs. The mathematical properties of SRCC (such as the ability to detect nonlinear correlation) make it compelling to be researched into. Our motivation is to provide a new technique of improving the quality of KE based on the well-known association measure SRCC, as opposed to other well-known TD similarity measures. The paper is organized as follows: Section SECREF2 gives a brief overview of the vector space representation of a TD and the corresponding similarity measures, in Section SECREF3 we address conducted research of the role of SRCC in data mining and trend prediction. Section SECREF4 is a detailed description of the proposed technique, and later, in Section SECREF5 we present clustering and classification experiments conducted on several sets of TDs, while Section SECREF6 summarizes our research and contribution to the broad area of statistical text analysis. Background In this section we provide a brief background of vector space representation of TDs and existing similarity measures that have been widely used in statistical text analysis. To begin with, we consider the representation of documents. Background ::: Document Representation A document $d$ can be defined as a finite sequence of terms (independent textual entities within a document, for example, words), namely $d=(t_1,t_2,\dots ,t_n)$. A general idea is to associate weight to each term $t_i$ within $d$, such that which has proven superior in prior extensive research BIBREF4. The most common weight measure is Term Frequency - Inverse Document Frequency (TF-IDF). TF is the frequency of a term within a single document, and IDF represents the importance, or uniqueness of a term within a set of documents $D=\lbrace d_1, d_2, \dots ,d_m\rbrace $. TF-IDF is defined as follows: where such that $f$ is the number of occurrences of $t$ in $d$ and $\log $ is used to avoid very small values close to zero. Having these measures defined, it becomes obvious that each $w_i$, for $i=1,\dots ,n$ is assigned the TF-IDF value of the corresponding term. It turns out that each document is represented as a vector of TF-IDF weights within a vector space model (VSM) with its properties BIBREF5. Background ::: Measures of Similarity Different ways of computing the similarity of two vector exist. There are two main approaches in similarity computation: Deterministic - similarity measures exploiting algebraic properties of vectors and their geometrical interpretation. These include, for instance, cosine similarity (CS), Jaccard coefficients (for binary representations), etc. Stochastic - similarity measures in which uncertainty is taken into account. These include, for instance, statistics such as Pearson's Correlation Coefficient (PCC) BIBREF6. Let $\mathbf {u}$ and $\mathbf {v}$ be the vector representations of two documents $d_1$ and $d_2$. Cosine similarity simply measures $cos\theta $, where $\theta $ is the angle between $\mathbf {u}$ and $\mathbf {v}$ (cosine similarity) (PCC) where All of the above measures are widely used and have proven efficient, but an important aspect is the lack of importance of the order of terms in textual data. It is easy for one to conclude that, two documents containing a single sentence each, but in a reverse order of terms, most deterministic methods fail to express that these are actually very similar. On the other hand, PCC detects only linear correlation, which constraints the diversity present in textual data. In the following section, we study relevant research in solving this problem, and then in Sections SECREF4 and SECREF5 we present our solution and results. Related Work A significant number of similarity measures have been proposed and this topic has been thoroughly elaborated. Its main application is considered to be clustering and classification of textual data organized in TDs. In this section, we provide an overview of relevant research on this topic, to which we can later compare our proposed technique for computing vector similarity. KE (also referred to as knowledge discovery) techniques are used to extract information from unstructured data, which can be subsequently used for applying supervised or unsupervised learning techniques, such as clustering and classification of the content BIBREF7. Text clustering should address several challenges such as vast amounts of data, very high dimensionality of more than 10,000 terms (dimensions), and most importantly - an understandable description of the clusters BIBREF8, which essentially implies the demand for high quality of extracted information. Regarding high quality KE and information accuracy, much effort has been put into improving similarity measurements. An improvement based on linear algebra, known as Singular Value Decomposition (SVD), is oriented towards word similarity, but instead, its main application is document similarity BIBREF9. Alluring is the fact that this measure takes the advantage of synonym recognition and has been used to achieve human-level scores on multiple-choice synonym questions from the Test of English as a Foreign Language (TOEFL) in a technique known as Latent Semantic Analysis (LSA) BIBREF10 BIBREF5. Other semantic term similarity measures have been also proposed, based on information exclusively derived from large corpora of words, such as Pointwise Mutual Information (PMI), which has been reported to have achieved a large degree of correctness in the synonym questions in the TOEFL and SAT tests BIBREF11. Moreover, normalized knowledge-based measures, such as Leacock & Chodrow BIBREF12, Lesk ("how to tell a pine cone from an ice-cream cone" BIBREF13, or measures for the depth of two concepts (preferably vebs) in the Word-Net taxonomy BIBREF14 have experimentally proven to be efficient. Their accuracy converges to approximately 69%, Leacock & Chodrow and Lesk have showed the highest precision, and having them combined turns out to be the approximately optimal solution BIBREF11. The Spearman's Rank Correlation Coefficient Similarity Measure The main idea behind our proposed technique is to introduce uncertainty in the calculations of the similarity between TDs represented in a vector space model, based on the nonlinear properties of SRCC. Unlike PCC, which is only able to detect linear correlation, SRCC's nonlinear ability provides a convenient way of taking different ordering of terms into account. The Spearman's Rank Correlation Coefficient Similarity Measure ::: Spearman's Rank Correlation Coefficient The Spreaman's Rank Correlation Coefficient BIBREF3, denoted $\rho $, has a from which is very similar to PCC. Namely, for $n$ raw scores $U_i, V_i$ for $i=1,\dots ,n$ denoting TF-IDF values for two document vectors $\mathbf {U}, \mathbf {V}$, where $u_i$ and $v_i$ are the corresponding ranks of $U_i$ and $V_i$, for $i=0,\dots ,n-1$. A metric to assign the ranks of each of the TF-IDF values has to be determined beforehand. Each $U_i$ is assigned a rank value $u_i$, such that $u_i=0,1,\dots ,n-1$. It is important to note that the metric by which the TF-IDF values are ranked is essentially their sorting criteria. A convenient way of determining this criteria when dealing with TF-IDF values, which emphasize the importance of a term within a TD set, is to sort these values in an ascending order. Thus, the largest (or most important) TF-IDF value within a TD vector is assigned the rank value of $n-1$, and the least important is assigned a value of 0. The Spearman's Rank Correlation Coefficient Similarity Measure ::: Spearman's Rank Correlation Coefficient ::: An Illustration of the Ranking TF-IDF Vectors Consider two TDs $d_1$ and $d_2$, each containing a single sentence. Document 1: John had asked Mary to marry him before she left. Document 2: Before she left, Mary was asked by John to be his wife. Now consider these sentences lemmatized: Document 1: John have ask Mary marry before leave. Document 2: Before leave Mary ask John his wife. Let us now represent $d_1$ and $d_2$ as TF-IDF vectors for the vocabulary in our small corpus. The results in Table TABREF7 show that SRCC performs much better in knowledge extraction. The two documents' contents contain the same idea expressed by terms in a different order that John had asked Mary to marry him before she left. It is obvious that cosine similarity cannot recognize this association, but SRCC has successfully recognized it and produced a similarity value of -0.285714. SRCC is essentially conducive to semantic similarity. Rising the importance of a term in a TD will eventually rise its importance in another TD. But if the two TDs are of different size, the terms' importance values will also differ, by which a nonlinear association will emerge. This association will not be recognized by PCC at all (as it only detects linear association), but SRCC will definitely catch this detail and produce the desirable similarity value. The idea is to use SRCC to catch such terms which drive the semantic context of a TD, which will follow a nonlinear and lie on a polynomial curve, and not on the line $x=y$. In our approach, we use a non-standard measure of similarity in textual data with simple and common frequency values, such as TF-IDF, in contrast to the statement that simple frequencies are not enough for high-quality knowledge extraction BIBREF5. In the next section, we will present our experiments and discuss the results we have obtained. Experiments In order to test our proposed approach, we have conducted a series of experiments. In this section, we briefly discuss the outcome and provide a clear view of whether our approach is suitable for knowledge extraction from textual data in a semantic context. We have used a dataset of 14 TDs to conduct our experiments. There are several subjects on which their content is based: (aliens, stories, law, news) BIBREF15. Experiments ::: Comparison Between Similarity Measures In this part, we have compared the similarity values produced by each of the similarity measures CS, SRCC and PCC. We have picked a few notable results and they are summarized in Table TABREF9 below. In Table TABREF9 that SRCC mostly differs from CS and PCC, which also differ in some cases.For instance, $d_1$ refers to leadership in the nineties, while $d_5$ refers to the family and medical lead act of 1993. We have empirically observed that the general topics discussed in these two textual documents are very different. Namely, discusses different frameworks for leadership empowerment, while $d_5$ discusses medical treatment and self-care of employees. We have observed that the term employee is the only connection between $d_1$ and $d_5$. The similarity value of CS of 0.36 is very unreal in this case, while PCC (0.05), and especially SRCC (0.0018) provide a much more realistic view of the semantic knowledge aggregated in these documents. Another example are $d_8$ and $d_9$. The contents of these documents are very straightforward and very similar, because they discuss aliens seen by Boeing-747 pilots and $d_9$ discusses angels that were considered to be aliens. It is obvious that SRCC is able to detect this association as good as CS and PCC which are very good in such straightforward cases. We have observed that SRCC does not perform worse than any other of these similarity measures. It does not always produce the most suitable similarity value, but it indeed does perform at least equally good as other measures. The values in Table TABREF9 are very small, and suggest that SRCC performs well in extracting tiny associations in such cases. It is mostly a few times larger than CS and PCC when there actually exist associations between the documents. These results are visually summarized in Figure FIGREF10. The two above-described examples can be clearly seen as standing out. Experiments ::: Non-linearity of Documents In this part we will briefly present the nonlinear association between some of the TDs we have used in our experiments. Our purpose is to point out that $(d_6,d_{10})$ and $(d_7,d_{12})$ are the pairs where SRCC is the most appropriate measure for the observed content, and as such, it is able to detect the nonlinear association between them. This can be seen in Figure FIGREF12 below. The straightforward case of $d_8$ and $d_9$ also stands out here (SRCC can also detect it very well). The obtained results showed that our technique shows good performance on similarity computing, although it is not a perfect measure. But, it sure comes close to convenient and widely used similarity measures such as CS and PCC. The next section provides a conclusion of our research and suggestions for further work. Conclusion and Future Work In this paper we have presented a non-standard technique for computing the similarity between TF-IDF vectors. We have propagated our idea and contributed a portion of new knowledge in this field of text analysis. We have proposed a technique that is widely used in similar fields, and our goal is to provide starting information to other researches in this area. We consider our observations promising and they should be extensively researched. Our experiments have proved that our technique should be a subject for further research. Our future work will concentrate on the implementation of machine learning techniques, such as clustering and subsequent classification of textual data. We expect an information of good quality to be extracted. To summarize, the rapidly emerging area of big data and information retrieval is where our technique should reside and where it should be applied.

finite sequence of terms

13ca4bf76565564c8ec3238c0cbfacb0b41e14d2

13ca4bf76565564c8ec3238c0cbfacb0b41e14d2_0

Q: Which dataset(s) do they use? Text: Introduction Over the past few years, the term big data has become an important key point for research into data mining and information retrieval. Through the years, the quantity of data managed across enterprises has evolved from a simple and imperceptible task to an extent to which it has become the central performance improvement problem. In other words, it evolved to be the next frontier for innovation, competition and productivity BIBREF0. Extracting knowledge from data is now a very competitive environment. Many companies process vast amounts of customer/user data in order to improve the quality of experience (QoE) of their customers. For instance, a typical use-case scenario would be a book seller that performs an automatic extraction of the content of the books a customer has bought, and subsequently extracts knowledge of what customers prefer to read. The knowledge extracted could then be used to recommend other books. Book recommending systems are typical examples where data mining techniques should be considered as the primary tool for making future decisions BIBREF1. KE from TDs is an essential field of research in data mining and it certainly requires techniques that are reliable and accurate in order to neutralize (or even eliminate) uncertainty in future decisions. Grouping TDs based on their content and mutual key information is referred to as clustering. Clustering is mostly performed with respect to a measure of similarity between TDs, which must be represented as vectors in a vector space beforehand BIBREF2. News aggregation engines can be considered as a typical representative where such techniques are extensively applied as a sub-field of natural language processing (NLP). In this paper we present a new technique for measuring similarity between TDs, represented in a vector space, based on SRCC - "a statistical measure of association between two things" BIBREF3, which in this case things refer to TDs. The mathematical properties of SRCC (such as the ability to detect nonlinear correlation) make it compelling to be researched into. Our motivation is to provide a new technique of improving the quality of KE based on the well-known association measure SRCC, as opposed to other well-known TD similarity measures. The paper is organized as follows: Section SECREF2 gives a brief overview of the vector space representation of a TD and the corresponding similarity measures, in Section SECREF3 we address conducted research of the role of SRCC in data mining and trend prediction. Section SECREF4 is a detailed description of the proposed technique, and later, in Section SECREF5 we present clustering and classification experiments conducted on several sets of TDs, while Section SECREF6 summarizes our research and contribution to the broad area of statistical text analysis. Background In this section we provide a brief background of vector space representation of TDs and existing similarity measures that have been widely used in statistical text analysis. To begin with, we consider the representation of documents. Background ::: Document Representation A document $d$ can be defined as a finite sequence of terms (independent textual entities within a document, for example, words), namely $d=(t_1,t_2,\dots ,t_n)$. A general idea is to associate weight to each term $t_i$ within $d$, such that which has proven superior in prior extensive research BIBREF4. The most common weight measure is Term Frequency - Inverse Document Frequency (TF-IDF). TF is the frequency of a term within a single document, and IDF represents the importance, or uniqueness of a term within a set of documents $D=\lbrace d_1, d_2, \dots ,d_m\rbrace $. TF-IDF is defined as follows: where such that $f$ is the number of occurrences of $t$ in $d$ and $\log $ is used to avoid very small values close to zero. Having these measures defined, it becomes obvious that each $w_i$, for $i=1,\dots ,n$ is assigned the TF-IDF value of the corresponding term. It turns out that each document is represented as a vector of TF-IDF weights within a vector space model (VSM) with its properties BIBREF5. Background ::: Measures of Similarity Different ways of computing the similarity of two vector exist. There are two main approaches in similarity computation: Deterministic - similarity measures exploiting algebraic properties of vectors and their geometrical interpretation. These include, for instance, cosine similarity (CS), Jaccard coefficients (for binary representations), etc. Stochastic - similarity measures in which uncertainty is taken into account. These include, for instance, statistics such as Pearson's Correlation Coefficient (PCC) BIBREF6. Let $\mathbf {u}$ and $\mathbf {v}$ be the vector representations of two documents $d_1$ and $d_2$. Cosine similarity simply measures $cos\theta $, where $\theta $ is the angle between $\mathbf {u}$ and $\mathbf {v}$ (cosine similarity) (PCC) where All of the above measures are widely used and have proven efficient, but an important aspect is the lack of importance of the order of terms in textual data. It is easy for one to conclude that, two documents containing a single sentence each, but in a reverse order of terms, most deterministic methods fail to express that these are actually very similar. On the other hand, PCC detects only linear correlation, which constraints the diversity present in textual data. In the following section, we study relevant research in solving this problem, and then in Sections SECREF4 and SECREF5 we present our solution and results. Related Work A significant number of similarity measures have been proposed and this topic has been thoroughly elaborated. Its main application is considered to be clustering and classification of textual data organized in TDs. In this section, we provide an overview of relevant research on this topic, to which we can later compare our proposed technique for computing vector similarity. KE (also referred to as knowledge discovery) techniques are used to extract information from unstructured data, which can be subsequently used for applying supervised or unsupervised learning techniques, such as clustering and classification of the content BIBREF7. Text clustering should address several challenges such as vast amounts of data, very high dimensionality of more than 10,000 terms (dimensions), and most importantly - an understandable description of the clusters BIBREF8, which essentially implies the demand for high quality of extracted information. Regarding high quality KE and information accuracy, much effort has been put into improving similarity measurements. An improvement based on linear algebra, known as Singular Value Decomposition (SVD), is oriented towards word similarity, but instead, its main application is document similarity BIBREF9. Alluring is the fact that this measure takes the advantage of synonym recognition and has been used to achieve human-level scores on multiple-choice synonym questions from the Test of English as a Foreign Language (TOEFL) in a technique known as Latent Semantic Analysis (LSA) BIBREF10 BIBREF5. Other semantic term similarity measures have been also proposed, based on information exclusively derived from large corpora of words, such as Pointwise Mutual Information (PMI), which has been reported to have achieved a large degree of correctness in the synonym questions in the TOEFL and SAT tests BIBREF11. Moreover, normalized knowledge-based measures, such as Leacock & Chodrow BIBREF12, Lesk ("how to tell a pine cone from an ice-cream cone" BIBREF13, or measures for the depth of two concepts (preferably vebs) in the Word-Net taxonomy BIBREF14 have experimentally proven to be efficient. Their accuracy converges to approximately 69%, Leacock & Chodrow and Lesk have showed the highest precision, and having them combined turns out to be the approximately optimal solution BIBREF11. The Spearman's Rank Correlation Coefficient Similarity Measure The main idea behind our proposed technique is to introduce uncertainty in the calculations of the similarity between TDs represented in a vector space model, based on the nonlinear properties of SRCC. Unlike PCC, which is only able to detect linear correlation, SRCC's nonlinear ability provides a convenient way of taking different ordering of terms into account. The Spearman's Rank Correlation Coefficient Similarity Measure ::: Spearman's Rank Correlation Coefficient The Spreaman's Rank Correlation Coefficient BIBREF3, denoted $\rho $, has a from which is very similar to PCC. Namely, for $n$ raw scores $U_i, V_i$ for $i=1,\dots ,n$ denoting TF-IDF values for two document vectors $\mathbf {U}, \mathbf {V}$, where $u_i$ and $v_i$ are the corresponding ranks of $U_i$ and $V_i$, for $i=0,\dots ,n-1$. A metric to assign the ranks of each of the TF-IDF values has to be determined beforehand. Each $U_i$ is assigned a rank value $u_i$, such that $u_i=0,1,\dots ,n-1$. It is important to note that the metric by which the TF-IDF values are ranked is essentially their sorting criteria. A convenient way of determining this criteria when dealing with TF-IDF values, which emphasize the importance of a term within a TD set, is to sort these values in an ascending order. Thus, the largest (or most important) TF-IDF value within a TD vector is assigned the rank value of $n-1$, and the least important is assigned a value of 0. The Spearman's Rank Correlation Coefficient Similarity Measure ::: Spearman's Rank Correlation Coefficient ::: An Illustration of the Ranking TF-IDF Vectors Consider two TDs $d_1$ and $d_2$, each containing a single sentence. Document 1: John had asked Mary to marry him before she left. Document 2: Before she left, Mary was asked by John to be his wife. Now consider these sentences lemmatized: Document 1: John have ask Mary marry before leave. Document 2: Before leave Mary ask John his wife. Let us now represent $d_1$ and $d_2$ as TF-IDF vectors for the vocabulary in our small corpus. The results in Table TABREF7 show that SRCC performs much better in knowledge extraction. The two documents' contents contain the same idea expressed by terms in a different order that John had asked Mary to marry him before she left. It is obvious that cosine similarity cannot recognize this association, but SRCC has successfully recognized it and produced a similarity value of -0.285714. SRCC is essentially conducive to semantic similarity. Rising the importance of a term in a TD will eventually rise its importance in another TD. But if the two TDs are of different size, the terms' importance values will also differ, by which a nonlinear association will emerge. This association will not be recognized by PCC at all (as it only detects linear association), but SRCC will definitely catch this detail and produce the desirable similarity value. The idea is to use SRCC to catch such terms which drive the semantic context of a TD, which will follow a nonlinear and lie on a polynomial curve, and not on the line $x=y$. In our approach, we use a non-standard measure of similarity in textual data with simple and common frequency values, such as TF-IDF, in contrast to the statement that simple frequencies are not enough for high-quality knowledge extraction BIBREF5. In the next section, we will present our experiments and discuss the results we have obtained. Experiments In order to test our proposed approach, we have conducted a series of experiments. In this section, we briefly discuss the outcome and provide a clear view of whether our approach is suitable for knowledge extraction from textual data in a semantic context. We have used a dataset of 14 TDs to conduct our experiments. There are several subjects on which their content is based: (aliens, stories, law, news) BIBREF15. Experiments ::: Comparison Between Similarity Measures In this part, we have compared the similarity values produced by each of the similarity measures CS, SRCC and PCC. We have picked a few notable results and they are summarized in Table TABREF9 below. In Table TABREF9 that SRCC mostly differs from CS and PCC, which also differ in some cases.For instance, $d_1$ refers to leadership in the nineties, while $d_5$ refers to the family and medical lead act of 1993. We have empirically observed that the general topics discussed in these two textual documents are very different. Namely, discusses different frameworks for leadership empowerment, while $d_5$ discusses medical treatment and self-care of employees. We have observed that the term employee is the only connection between $d_1$ and $d_5$. The similarity value of CS of 0.36 is very unreal in this case, while PCC (0.05), and especially SRCC (0.0018) provide a much more realistic view of the semantic knowledge aggregated in these documents. Another example are $d_8$ and $d_9$. The contents of these documents are very straightforward and very similar, because they discuss aliens seen by Boeing-747 pilots and $d_9$ discusses angels that were considered to be aliens. It is obvious that SRCC is able to detect this association as good as CS and PCC which are very good in such straightforward cases. We have observed that SRCC does not perform worse than any other of these similarity measures. It does not always produce the most suitable similarity value, but it indeed does perform at least equally good as other measures. The values in Table TABREF9 are very small, and suggest that SRCC performs well in extracting tiny associations in such cases. It is mostly a few times larger than CS and PCC when there actually exist associations between the documents. These results are visually summarized in Figure FIGREF10. The two above-described examples can be clearly seen as standing out. Experiments ::: Non-linearity of Documents In this part we will briefly present the nonlinear association between some of the TDs we have used in our experiments. Our purpose is to point out that $(d_6,d_{10})$ and $(d_7,d_{12})$ are the pairs where SRCC is the most appropriate measure for the observed content, and as such, it is able to detect the nonlinear association between them. This can be seen in Figure FIGREF12 below. The straightforward case of $d_8$ and $d_9$ also stands out here (SRCC can also detect it very well). The obtained results showed that our technique shows good performance on similarity computing, although it is not a perfect measure. But, it sure comes close to convenient and widely used similarity measures such as CS and PCC. The next section provides a conclusion of our research and suggestions for further work. Conclusion and Future Work In this paper we have presented a non-standard technique for computing the similarity between TF-IDF vectors. We have propagated our idea and contributed a portion of new knowledge in this field of text analysis. We have proposed a technique that is widely used in similar fields, and our goal is to provide starting information to other researches in this area. We consider our observations promising and they should be extensively researched. Our experiments have proved that our technique should be a subject for further research. Our future work will concentrate on the implementation of machine learning techniques, such as clustering and subsequent classification of textual data. We expect an information of good quality to be extracted. To summarize, the rapidly emerging area of big data and information retrieval is where our technique should reside and where it should be applied.

14 TDs, BIBREF15

70797f66d96aa163a3bee2be30a328ba61c40a18

70797f66d96aa163a3bee2be30a328ba61c40a18_0

Q: How do they evaluate knowledge extraction performance? Text: Introduction Over the past few years, the term big data has become an important key point for research into data mining and information retrieval. Through the years, the quantity of data managed across enterprises has evolved from a simple and imperceptible task to an extent to which it has become the central performance improvement problem. In other words, it evolved to be the next frontier for innovation, competition and productivity BIBREF0. Extracting knowledge from data is now a very competitive environment. Many companies process vast amounts of customer/user data in order to improve the quality of experience (QoE) of their customers. For instance, a typical use-case scenario would be a book seller that performs an automatic extraction of the content of the books a customer has bought, and subsequently extracts knowledge of what customers prefer to read. The knowledge extracted could then be used to recommend other books. Book recommending systems are typical examples where data mining techniques should be considered as the primary tool for making future decisions BIBREF1. KE from TDs is an essential field of research in data mining and it certainly requires techniques that are reliable and accurate in order to neutralize (or even eliminate) uncertainty in future decisions. Grouping TDs based on their content and mutual key information is referred to as clustering. Clustering is mostly performed with respect to a measure of similarity between TDs, which must be represented as vectors in a vector space beforehand BIBREF2. News aggregation engines can be considered as a typical representative where such techniques are extensively applied as a sub-field of natural language processing (NLP). In this paper we present a new technique for measuring similarity between TDs, represented in a vector space, based on SRCC - "a statistical measure of association between two things" BIBREF3, which in this case things refer to TDs. The mathematical properties of SRCC (such as the ability to detect nonlinear correlation) make it compelling to be researched into. Our motivation is to provide a new technique of improving the quality of KE based on the well-known association measure SRCC, as opposed to other well-known TD similarity measures. The paper is organized as follows: Section SECREF2 gives a brief overview of the vector space representation of a TD and the corresponding similarity measures, in Section SECREF3 we address conducted research of the role of SRCC in data mining and trend prediction. Section SECREF4 is a detailed description of the proposed technique, and later, in Section SECREF5 we present clustering and classification experiments conducted on several sets of TDs, while Section SECREF6 summarizes our research and contribution to the broad area of statistical text analysis. Background In this section we provide a brief background of vector space representation of TDs and existing similarity measures that have been widely used in statistical text analysis. To begin with, we consider the representation of documents. Background ::: Document Representation A document $d$ can be defined as a finite sequence of terms (independent textual entities within a document, for example, words), namely $d=(t_1,t_2,\dots ,t_n)$. A general idea is to associate weight to each term $t_i$ within $d$, such that which has proven superior in prior extensive research BIBREF4. The most common weight measure is Term Frequency - Inverse Document Frequency (TF-IDF). TF is the frequency of a term within a single document, and IDF represents the importance, or uniqueness of a term within a set of documents $D=\lbrace d_1, d_2, \dots ,d_m\rbrace $. TF-IDF is defined as follows: where such that $f$ is the number of occurrences of $t$ in $d$ and $\log $ is used to avoid very small values close to zero. Having these measures defined, it becomes obvious that each $w_i$, for $i=1,\dots ,n$ is assigned the TF-IDF value of the corresponding term. It turns out that each document is represented as a vector of TF-IDF weights within a vector space model (VSM) with its properties BIBREF5. Background ::: Measures of Similarity Different ways of computing the similarity of two vector exist. There are two main approaches in similarity computation: Deterministic - similarity measures exploiting algebraic properties of vectors and their geometrical interpretation. These include, for instance, cosine similarity (CS), Jaccard coefficients (for binary representations), etc. Stochastic - similarity measures in which uncertainty is taken into account. These include, for instance, statistics such as Pearson's Correlation Coefficient (PCC) BIBREF6. Let $\mathbf {u}$ and $\mathbf {v}$ be the vector representations of two documents $d_1$ and $d_2$. Cosine similarity simply measures $cos\theta $, where $\theta $ is the angle between $\mathbf {u}$ and $\mathbf {v}$ (cosine similarity) (PCC) where All of the above measures are widely used and have proven efficient, but an important aspect is the lack of importance of the order of terms in textual data. It is easy for one to conclude that, two documents containing a single sentence each, but in a reverse order of terms, most deterministic methods fail to express that these are actually very similar. On the other hand, PCC detects only linear correlation, which constraints the diversity present in textual data. In the following section, we study relevant research in solving this problem, and then in Sections SECREF4 and SECREF5 we present our solution and results. Related Work A significant number of similarity measures have been proposed and this topic has been thoroughly elaborated. Its main application is considered to be clustering and classification of textual data organized in TDs. In this section, we provide an overview of relevant research on this topic, to which we can later compare our proposed technique for computing vector similarity. KE (also referred to as knowledge discovery) techniques are used to extract information from unstructured data, which can be subsequently used for applying supervised or unsupervised learning techniques, such as clustering and classification of the content BIBREF7. Text clustering should address several challenges such as vast amounts of data, very high dimensionality of more than 10,000 terms (dimensions), and most importantly - an understandable description of the clusters BIBREF8, which essentially implies the demand for high quality of extracted information. Regarding high quality KE and information accuracy, much effort has been put into improving similarity measurements. An improvement based on linear algebra, known as Singular Value Decomposition (SVD), is oriented towards word similarity, but instead, its main application is document similarity BIBREF9. Alluring is the fact that this measure takes the advantage of synonym recognition and has been used to achieve human-level scores on multiple-choice synonym questions from the Test of English as a Foreign Language (TOEFL) in a technique known as Latent Semantic Analysis (LSA) BIBREF10 BIBREF5. Other semantic term similarity measures have been also proposed, based on information exclusively derived from large corpora of words, such as Pointwise Mutual Information (PMI), which has been reported to have achieved a large degree of correctness in the synonym questions in the TOEFL and SAT tests BIBREF11. Moreover, normalized knowledge-based measures, such as Leacock & Chodrow BIBREF12, Lesk ("how to tell a pine cone from an ice-cream cone" BIBREF13, or measures for the depth of two concepts (preferably vebs) in the Word-Net taxonomy BIBREF14 have experimentally proven to be efficient. Their accuracy converges to approximately 69%, Leacock & Chodrow and Lesk have showed the highest precision, and having them combined turns out to be the approximately optimal solution BIBREF11. The Spearman's Rank Correlation Coefficient Similarity Measure The main idea behind our proposed technique is to introduce uncertainty in the calculations of the similarity between TDs represented in a vector space model, based on the nonlinear properties of SRCC. Unlike PCC, which is only able to detect linear correlation, SRCC's nonlinear ability provides a convenient way of taking different ordering of terms into account. The Spearman's Rank Correlation Coefficient Similarity Measure ::: Spearman's Rank Correlation Coefficient The Spreaman's Rank Correlation Coefficient BIBREF3, denoted $\rho $, has a from which is very similar to PCC. Namely, for $n$ raw scores $U_i, V_i$ for $i=1,\dots ,n$ denoting TF-IDF values for two document vectors $\mathbf {U}, \mathbf {V}$, where $u_i$ and $v_i$ are the corresponding ranks of $U_i$ and $V_i$, for $i=0,\dots ,n-1$. A metric to assign the ranks of each of the TF-IDF values has to be determined beforehand. Each $U_i$ is assigned a rank value $u_i$, such that $u_i=0,1,\dots ,n-1$. It is important to note that the metric by which the TF-IDF values are ranked is essentially their sorting criteria. A convenient way of determining this criteria when dealing with TF-IDF values, which emphasize the importance of a term within a TD set, is to sort these values in an ascending order. Thus, the largest (or most important) TF-IDF value within a TD vector is assigned the rank value of $n-1$, and the least important is assigned a value of 0. The Spearman's Rank Correlation Coefficient Similarity Measure ::: Spearman's Rank Correlation Coefficient ::: An Illustration of the Ranking TF-IDF Vectors Consider two TDs $d_1$ and $d_2$, each containing a single sentence. Document 1: John had asked Mary to marry him before she left. Document 2: Before she left, Mary was asked by John to be his wife. Now consider these sentences lemmatized: Document 1: John have ask Mary marry before leave. Document 2: Before leave Mary ask John his wife. Let us now represent $d_1$ and $d_2$ as TF-IDF vectors for the vocabulary in our small corpus. The results in Table TABREF7 show that SRCC performs much better in knowledge extraction. The two documents' contents contain the same idea expressed by terms in a different order that John had asked Mary to marry him before she left. It is obvious that cosine similarity cannot recognize this association, but SRCC has successfully recognized it and produced a similarity value of -0.285714. SRCC is essentially conducive to semantic similarity. Rising the importance of a term in a TD will eventually rise its importance in another TD. But if the two TDs are of different size, the terms' importance values will also differ, by which a nonlinear association will emerge. This association will not be recognized by PCC at all (as it only detects linear association), but SRCC will definitely catch this detail and produce the desirable similarity value. The idea is to use SRCC to catch such terms which drive the semantic context of a TD, which will follow a nonlinear and lie on a polynomial curve, and not on the line $x=y$. In our approach, we use a non-standard measure of similarity in textual data with simple and common frequency values, such as TF-IDF, in contrast to the statement that simple frequencies are not enough for high-quality knowledge extraction BIBREF5. In the next section, we will present our experiments and discuss the results we have obtained. Experiments In order to test our proposed approach, we have conducted a series of experiments. In this section, we briefly discuss the outcome and provide a clear view of whether our approach is suitable for knowledge extraction from textual data in a semantic context. We have used a dataset of 14 TDs to conduct our experiments. There are several subjects on which their content is based: (aliens, stories, law, news) BIBREF15. Experiments ::: Comparison Between Similarity Measures In this part, we have compared the similarity values produced by each of the similarity measures CS, SRCC and PCC. We have picked a few notable results and they are summarized in Table TABREF9 below. In Table TABREF9 that SRCC mostly differs from CS and PCC, which also differ in some cases.For instance, $d_1$ refers to leadership in the nineties, while $d_5$ refers to the family and medical lead act of 1993. We have empirically observed that the general topics discussed in these two textual documents are very different. Namely, discusses different frameworks for leadership empowerment, while $d_5$ discusses medical treatment and self-care of employees. We have observed that the term employee is the only connection between $d_1$ and $d_5$. The similarity value of CS of 0.36 is very unreal in this case, while PCC (0.05), and especially SRCC (0.0018) provide a much more realistic view of the semantic knowledge aggregated in these documents. Another example are $d_8$ and $d_9$. The contents of these documents are very straightforward and very similar, because they discuss aliens seen by Boeing-747 pilots and $d_9$ discusses angels that were considered to be aliens. It is obvious that SRCC is able to detect this association as good as CS and PCC which are very good in such straightforward cases. We have observed that SRCC does not perform worse than any other of these similarity measures. It does not always produce the most suitable similarity value, but it indeed does perform at least equally good as other measures. The values in Table TABREF9 are very small, and suggest that SRCC performs well in extracting tiny associations in such cases. It is mostly a few times larger than CS and PCC when there actually exist associations between the documents. These results are visually summarized in Figure FIGREF10. The two above-described examples can be clearly seen as standing out. Experiments ::: Non-linearity of Documents In this part we will briefly present the nonlinear association between some of the TDs we have used in our experiments. Our purpose is to point out that $(d_6,d_{10})$ and $(d_7,d_{12})$ are the pairs where SRCC is the most appropriate measure for the observed content, and as such, it is able to detect the nonlinear association between them. This can be seen in Figure FIGREF12 below. The straightforward case of $d_8$ and $d_9$ also stands out here (SRCC can also detect it very well). The obtained results showed that our technique shows good performance on similarity computing, although it is not a perfect measure. But, it sure comes close to convenient and widely used similarity measures such as CS and PCC. The next section provides a conclusion of our research and suggestions for further work. Conclusion and Future Work In this paper we have presented a non-standard technique for computing the similarity between TF-IDF vectors. We have propagated our idea and contributed a portion of new knowledge in this field of text analysis. We have proposed a technique that is widely used in similar fields, and our goal is to provide starting information to other researches in this area. We consider our observations promising and they should be extensively researched. Our experiments have proved that our technique should be a subject for further research. Our future work will concentrate on the implementation of machine learning techniques, such as clustering and subsequent classification of textual data. We expect an information of good quality to be extracted. To summarize, the rapidly emerging area of big data and information retrieval is where our technique should reside and where it should be applied.

SRCC

71f2b368228a748fd348f1abf540236568a61b07

71f2b368228a748fd348f1abf540236568a61b07_0

Q: What is CamemBERT trained on? Text: Introduction Pretrained word representations have a long history in Natural Language Processing (NLP), from non-neural methods BIBREF0, BIBREF1, BIBREF2 to neural word embeddings BIBREF3, BIBREF4 and to contextualised representations BIBREF5, BIBREF6. Approaches shifted more recently from using these representations as an input to task-specific architectures to replacing these architectures with large pretrained language models. These models are then fine-tuned to the task at hand with large improvements in performance over a wide range of tasks BIBREF7, BIBREF8, BIBREF9, BIBREF10. These transfer learning methods exhibit clear advantages over more traditional task-specific approaches, probably the most important being that they can be trained in an unsupervised manner. They nevertheless come with implementation challenges, namely the amount of data and computational resources needed for pretraining that can reach hundreds of gigabytes of uncompressed text and require hundreds of GPUs BIBREF11, BIBREF9. The latest transformer architecture has gone uses as much as 750GB of plain text and 1024 TPU v3 for pretraining BIBREF10. This has limited the availability of these state-of-the-art models to the English language, at least in the monolingual setting. Even though multilingual models give remarkable results, they are often larger and their results still lag behind their monolingual counterparts BIBREF12. This is particularly inconvenient as it hinders their practical use in NLP systems as well as the investigation of their language modeling capacity, something that remains to be investigated in the case of, for instance, morphologically rich languages. We take advantage of the newly available multilingual corpus OSCAR BIBREF13 and train a monolingual language model for French using the RoBERTa architecture. We pretrain the model - which we dub CamemBERT- and evaluate it in four different downstream tasks for French: part-of-speech (POS) tagging, dependency parsing, named entity recognition (NER) and natural language inference (NLI). CamemBERT improves the state of the art for most tasks over previous monolingual and multilingual approaches, which confirms the effectiveness of large pretrained language models for French. We summarise our contributions as follows: We train a monolingual BERT model on the French language using recent large-scale corpora. We evaluate our model on four downstream tasks (POS tagging, dependency parsing, NER and natural language inference (NLI)), achieving state-of-the-art results in most tasks, confirming the effectiveness of large BERT-based models for French. We release our model in a user-friendly format for popular open-source libraries so that it can serve as a strong baseline for future research and be useful for French NLP practitioners. Related Work ::: From non-contextual to contextual word embeddings The first neural word vector representations were non-contextualised word embeddings, most notably word2vec BIBREF3, GloVe BIBREF4 and fastText BIBREF14, which were designed to be used as input to task-specific neural architectures. Contextualised word representations such as ELMo BIBREF5 and flair BIBREF6, improved the expressivity of word embeddings by taking context into account. They improved the performance of downstream tasks when they replaced traditional word representations. This paved the way towards larger contextualised models that replaced downstream architectures in most tasks. These approaches, trained with language modeling objectives, range from LSTM-based architectures such as ULMFiT BIBREF15 to the successful transformer-based architectures such as GPT2 BIBREF8, BERT BIBREF7, RoBERTa BIBREF9 and more recently ALBERT BIBREF16 and T5 BIBREF10. Related Work ::: Non-contextual word embeddings for languages other than English Since the introduction of word2vec BIBREF3, many attempts have been made to create monolingual models for a wide range of languages. For non-contextual word embeddings, the first two attempts were by BIBREF17 and BIBREF18 who created word embeddings for a large number of languages using Wikipedia. Later BIBREF19 trained fastText word embeddings for 157 languages using Common Crawl and showed that using crawled data significantly increased the performance of the embeddings relatively to those trained only on Wikipedia. Related Work ::: Contextualised models for languages other than English Following the success of large pretrained language models, they were extended to the multilingual setting with multilingual BERT , a single multilingual model for 104 different languages trained on Wikipedia data, and later XLM BIBREF12, which greatly improved unsupervised machine translation. A few monolingual models have been released: ELMo models for Japanese, Portuguese, German and Basque and BERT for Simplified and Traditional Chinese and German. However, to the best of our knowledge, no particular effort has been made toward training models for languages other than English, at a scale similar to the latest English models (e.g. RoBERTa trained on more than 100GB of data). CamemBERT Our approach is based on RoBERTa BIBREF9, which replicates and improves the initial BERT by identifying key hyper-parameters for more robust performance. In this section, we describe the architecture, training objective, optimisation setup and pretraining data that was used for CamemBERT. CamemBERT differs from RoBERTa mainly with the addition of whole-word masking and the usage of SentencePiece tokenisation BIBREF20. CamemBERT ::: Architecture Similar to RoBERTa and BERT, CamemBERT is a multi-layer bidirectional Transformer BIBREF21. Given the widespread usage of Transformers, we do not describe them in detail here and refer the reader to BIBREF21. CamemBERT uses the original BERT $_{\small \textsc {BASE}}$ configuration: 12 layers, 768 hidden dimensions, 12 attention heads, which amounts to 110M parameters. CamemBERT ::: Pretraining objective We train our model on the Masked Language Modeling (MLM) task. Given an input text sequence composed of $N$ tokens $x_1, ..., x_N$, we select $15\%$ of tokens for possible replacement. Among those selected tokens, 80% are replaced with the special $<$mask$>$ token, 10% are left unchanged and 10% are replaced by a random token. The model is then trained to predict the initial masked tokens using cross-entropy loss. Following RoBERTa we dynamically mask tokens instead of fixing them statically for the whole dataset during preprocessing. This improves variability and makes the model more robust when training for multiple epochs. Since we segment the input sentence into subwords using SentencePiece, the input tokens to the models can be subwords. An upgraded version of BERT and BIBREF22 have shown that masking whole words instead of individual subwords leads to improved performance. Whole-word masking (WWM) makes the training task more difficult because the model has to predict a whole word instead of predicting only part of the word given the rest. As a result, we used WWM for CamemBERT by first randomly sampling 15% of the words in the sequence and then considering all subword tokens in each of these 15% words for candidate replacement. This amounts to a proportion of selected tokens that is close to the original 15%. These tokens are then either replaced by $<$mask$>$ tokens (80%), left unchanged (10%) or replaced by a random token. Subsequent work has shown that the next sentence prediction task (NSP) originally used in BERT does not improve downstream task performance BIBREF12, BIBREF9, we do not use NSP as a consequence. CamemBERT ::: Optimisation Following BIBREF9, we optimise the model using Adam BIBREF23 ($\beta _1 = 0.9$, $\beta _2 = 0.98$) for 100k steps. We use large batch sizes of 8192 sequences. Each sequence contains at most 512 tokens. We enforce each sequence to only contain complete sentences. Additionally, we used the DOC-SENTENCES scenario from BIBREF9, consisting of not mixing multiple documents in the same sequence, which showed slightly better results. CamemBERT ::: Segmentation into subword units We segment the input text into subword units using SentencePiece BIBREF20. SentencePiece is an extension of Byte-Pair encoding (BPE) BIBREF24 and WordPiece BIBREF25 that does not require pre-tokenisation (at the word or token level), thus removing the need for language-specific tokenisers. We use a vocabulary size of 32k subword tokens. These are learned on $10^7$ sentences sampled from the pretraining dataset. We do not use subword regularisation (i.e. sampling from multiple possible segmentations) in our implementation for simplicity. CamemBERT ::: Pretraining data Pretrained language models can be significantly improved by using more data BIBREF9, BIBREF10. Therefore we used French text extracted from Common Crawl, in particular, we use OSCAR BIBREF13 a pre-classified and pre-filtered version of the November 2018 Common Craw snapshot. OSCAR is a set of monolingual corpora extracted from Common Crawl, specifically from the plain text WET format distributed by Common Crawl, which removes all HTML tags and converts all text encodings to UTF-8. OSCAR follows the same approach as BIBREF19 by using a language classification model based on the fastText linear classifier BIBREF26, BIBREF27 pretrained on Wikipedia, Tatoeba and SETimes, which supports 176 different languages. OSCAR performs a deduplication step after language classification and without introducing a specialised filtering scheme, other than only keeping paragraphs containing 100 or more UTF-8 encoded characters, making OSCAR quite close to the original Crawled data. We use the unshuffled version of the French OSCAR corpus, which amounts to 138GB of uncompressed text and 32.7B SentencePiece tokens. Evaluation ::: Part-of-speech tagging and dependency parsing We fist evaluate CamemBERT on the two downstream tasks of part-of-speech (POS) tagging and dependency parsing. POS tagging is a low-level syntactic task, which consists in assigning to each word its corresponding grammatical category. Dependency parsing consists in predicting the labeled syntactic tree capturing the syntactic relations between words. We run our experiments using the Universal Dependencies (UD) paradigm and its corresponding UD POS tag set BIBREF28 and UD treebank collection version 2.2 BIBREF29, which was used for the CoNLL 2018 shared task. We perform our work on the four freely available French UD treebanks in UD v2.2: GSD, Sequoia, Spoken, and ParTUT. GSD BIBREF30 is the second-largest treebank available for French after the FTB (described in subsection SECREF25), it contains data from blogs, news articles, reviews, and Wikipedia. The Sequoia treebank BIBREF31, BIBREF32 comprises more than 3000 sentences, from the French Europarl, the regional newspaper L’Est Républicain, the French Wikipedia and documents from the European Medicines Agency. Spoken is a corpus converted automatically from the Rhapsodie treebank BIBREF33, BIBREF34 with manual corrections. It consists of 57 sound samples of spoken French with orthographic transcription and phonetic transcription aligned with sound (word boundaries, syllables, and phonemes), syntactic and prosodic annotations. Finally, ParTUT is a conversion of a multilingual parallel treebank developed at the University of Turin, and consisting of a variety of text genres, including talks, legal texts, and Wikipedia articles, among others; ParTUT data is derived from the already-existing parallel treebank Par(allel)TUT BIBREF35 . Table TABREF23 contains a summary comparing the sizes of the treebanks. We evaluate the performance of our models using the standard UPOS accuracy for POS tagging, and Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS) for dependency parsing. We assume gold tokenisation and gold word segmentation as provided in the UD treebanks. Evaluation ::: Part-of-speech tagging and dependency parsing ::: Baselines To demonstrate the value of building a dedicated version of BERT for French, we first compare CamemBERT to the multilingual cased version of BERT (designated as mBERT). We then compare our models to UDify BIBREF36. UDify is a multitask and multilingual model based on mBERT that is near state-of-the-art on all UD languages including French for both POS tagging and dependency parsing. It is relevant to compare CamemBERT to UDify on those tasks because UDify is the work that pushed the furthest the performance in fine-tuning end-to-end a BERT-based model on downstream POS tagging and dependency parsing. Finally, we compare our model to UDPipe Future BIBREF37, a model ranked 3rd in dependency parsing and 6th in POS tagging during the CoNLL 2018 shared task BIBREF38. UDPipe Future provides us a strong baseline that does not make use of any pretrained contextual embedding. We will compare to the more recent cross-lingual language model XLM BIBREF12, as well as the state-of-the-art CoNLL 2018 shared task results with predicted tokenisation and segmentation in an updated version of the paper. Evaluation ::: Named Entity Recognition Named Entity Recognition (NER) is a sequence labeling task that consists in predicting which words refer to real-world objects, such as people, locations, artifacts and organisations. We use the French Treebank (FTB) BIBREF39 in its 2008 version introduced by cc-clustering:09short and with NER annotations by sagot2012annotation. The NER-annotated FTB contains more than 12k sentences and more than 350k tokens extracted from articles of the newspaper Le Monde published between 1989 and 1995. In total, it contains 11,636 entity mentions distributed among 7 different types of entities, namely: 2025 mentions of “Person”, 3761 of “Location”, 2382 of “Organisation”, 3357 of “Company”, 67 of “Product”, 15 of “POI” (Point of Interest) and 29 of “Fictional Character”. A large proportion of the entity mentions in the treebank are multi-word entities. For NER we therefore report the 3 metrics that are commonly used to evaluate models: precision, recall, and F1 score. Here precision measures the percentage of entities found by the system that are correctly tagged, recall measures the percentage of named entities present in the corpus that are found and the F1 score combines both precision and recall measures giving a general idea of a model's performance. Evaluation ::: Named Entity Recognition ::: Baselines Most of the advances in NER haven been achieved on English, particularly focusing on the CoNLL 2003 BIBREF40 and the Ontonotes v5 BIBREF41, BIBREF42 English corpora. NER is a task that was traditionally tackled using Conditional Random Fields (CRF) BIBREF43 which are quite suited for NER; CRFs were later used as decoding layers for Bi-LSTM architectures BIBREF44, BIBREF45 showing considerable improvements over CRFs alone. These Bi-LSTM-CRF architectures were later enhanced with contextualised word embeddings which yet again brought major improvements to the task BIBREF5, BIBREF6. Finally, large pretrained architectures settled the current state of the art showing a small yet important improvement over previous NER-specific architectures BIBREF7, BIBREF46. In non-English NER the CoNLL 2002 shared task included NER corpora for Spanish and Dutch corpora BIBREF47 while the CoNLL 2003 included a German corpus BIBREF40. Here the recent efforts of BIBREF48 settled the state of the art for Spanish and Dutch, while BIBREF6 did it for German. In French, no extensive work has been done due to the limited availability of NER corpora. We compare our model with the strong baselines settled by BIBREF49, who trained both CRF and BiLSTM-CRF architectures on the FTB and enhanced them using heuristics and pretrained word embeddings. Evaluation ::: Natural Language Inference We also evaluate our model on the Natural Language Inference (NLI) task, using the French part of the XNLI dataset BIBREF50. NLI consists in predicting whether a hypothesis sentence is entailed, neutral or contradicts a premise sentence. The XNLI dataset is the extension of the Multi-Genre NLI (MultiNLI) corpus BIBREF51 to 15 languages by translating the validation and test sets manually into each of those languages. The English training set is also machine translated for all languages. The dataset is composed of 122k train, 2490 valid and 5010 test examples. As usual, NLI performance is evaluated using accuracy. To evaluate a model on a language other than English (such as French), we consider the two following settings: TRANSLATE-TEST: The French test set is machine translated into English, and then used with an English classification model. This setting provides a reasonable, although imperfect, way to circumvent the fact that no such data set exists for French, and results in very strong baseline scores. TRANSLATE-TRAIN: The French model is fine-tuned on the machine-translated English training set and then evaluated on the French test set. This is the setting that we used for CamemBERT. Evaluation ::: Natural Language Inference ::: Baselines For the TRANSLATE-TEST setting, we report results of the English RoBERTa to act as a reference. In the TRANSLATE-TRAIN setting, we report the best scores from previous literature along with ours. BiLSTM-max is the best model in the original XNLI paper, mBERT which has been reported in French in BIBREF52 and XLM (MLM+TLM) is the best-presented model from BIBREF50. Experiments In this section, we measure the performance of CamemBERT by evaluating it on the four aforementioned tasks: POS tagging, dependency parsing, NER and NLI. Experiments ::: Experimental Setup ::: Pretraining We use the RoBERTa implementation in the fairseq library BIBREF53. Our learning rate is warmed up for 10k steps up to a peak value of $0.0007$ instead of the original $0.0001$ given our large batch size (8192). The learning rate fades to zero with polynomial decay. We pretrain our model on 256 Nvidia V100 GPUs (32GB each) for 100k steps during 17h. Experiments ::: Experimental Setup ::: Fine-tuning For each task, we append the relevant predictive layer on top of CamemBERT's Transformer architecture. Following the work done on BERT BIBREF7, for sequence tagging and sequence labeling we append a linear layer respectively to the $<$s$>$ special token and to the first subword token of each word. For dependency parsing, we plug a bi-affine graph predictor head as inspired by BIBREF54 following the work done on multilingual parsing with BERT by BIBREF36. We refer the reader to these two articles for more details on this module. We fine-tune independently CamemBERT for each task and each dataset. We optimise the model using the Adam optimiser BIBREF23 with a fixed learning rate. We run a grid search on a combination of learning rates and batch sizes. We select the best model on the validation set out of the 30 first epochs. Although this might push the performances even further, for all tasks except NLI, we don't apply any regularisation techniques such as weight decay, learning rate warm-up or discriminative fine-tuning. We show that fine-tuning CamemBERT in a straight-forward manner leads to state-of-the-art results on most tasks and outperforms the existing BERT-based models in most cases. The POS tagging, dependency parsing, and NER experiments are run using hugging face's Transformer library extended to support CamemBERT and dependency parsing BIBREF55. The NLI experiments use the fairseq library following the RoBERTa implementation. Experiments ::: Results ::: Part-of-Speech tagging and dependency parsing For POS tagging and dependency parsing, we compare CamemBERT to three other near state-of-the-art models in Table TABREF32. CamemBERT outperforms UDPipe Future by a large margin for all treebanks and all metrics. Despite a much simpler optimisation process, CamemBERT beats UDify performances on all the available French treebanks. CamemBERT also demonstrates higher performances than mBERT on those tasks. We observe a larger error reduction for parsing than for tagging. For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT. For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT. Experiments ::: Results ::: Natural Language Inference: XNLI On the XNLI benchmark, CamemBERT obtains improved performance over multilingual language models on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM) while using less than half the parameters (110M vs. 250M). However, its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa. It should be noted that CamemBERT uses far fewer parameters than RoBERTa (110M vs. 355M parameters). Experiments ::: Results ::: Named-Entity Recognition For named entity recognition, our experiments show that CamemBERT achieves a slightly better precision than the traditional CRF-based SEM architectures described above in Section SECREF25 (CRF and Bi-LSTM+CRF), but shows a dramatic improvement in finding entity mentions, raising the recall score by 3.5 points. Both improvements result in a 2.36 point increase in the F1 score with respect to the best SEM architecture (BiLSTM-CRF), giving CamemBERT the state of the art for NER on the FTB. One other important finding is the results obtained by mBERT. Previous work with this model showed increased performance in NER for German, Dutch and Spanish when mBERT is used as contextualised word embedding for an NER-specific model BIBREF48, but our results suggest that the multilingual setting in which mBERT was trained is simply not enough to use it alone and fine-tune it for French NER, as it shows worse performance than even simple CRF models, suggesting that monolingual models could be better at NER. Experiments ::: Discussion CamemBERT displays improved performance compared to prior work for the 4 downstream tasks considered. This confirms the hypothesis that pretrained language models can be effectively fine-tuned for various downstream tasks, as observed for English in previous work. Moreover, our results also show that dedicated monolingual models still outperform multilingual ones. We explain this point in two ways. First, the scale of data is possibly essential to the performance of CamemBERT. Indeed, we use 138GB of uncompressed text vs. 57GB for mBERT. Second, with more data comes more diversity in the pretraining distribution. Reaching state-of-the-art performances on 4 different tasks and 6 different datasets requires robust pretrained models. Our results suggest that the variability in the downstream tasks and datasets considered is handled more efficiently by a general language model than by Wikipedia-pretrained models such as mBERT. Conclusion CamemBERT improves the state of the art for multiple downstream tasks in French. It is also lighter than other BERT-based approaches such as mBERT or XLM. By releasing our model, we hope that it can serve as a strong baseline for future research in French NLP, and expect our experiments to be reproduced in many other languages. We will publish an updated version in the near future where we will explore and release models trained for longer, with additional downstream tasks, baselines (e.g. XLM) and analysis, we will also train additional models with potentially cleaner corpora such as CCNet BIBREF56 for more accurate performance evaluation and more complete ablation. Acknowledgments This work was partly funded by three French National grants from the Agence Nationale de la Recherche, namely projects PARSITI (ANR-16-CE33-0021), SoSweet (ANR-15-CE38-0011) and BASNUM (ANR-18-CE38-0003), as well as by the last author's chair in the PRAIRIE institute. Appendix ::: Impact of Whole-Word Masking We analyze the addition of whole-word masking on the downstream performance of CamemBERT. As reported for English on other downstream tasks, whole word masking improves downstream performances for all tasks but NER as seen in Table TABREF46. NER is highly sensitive to capitalisation, prefixes, suffixes and other subword features that could be used by a model to correctly identify entity mentions. Thus the added information by learning the masking at a subword level rather than at whole-word level seems to have a detrimental effect on downstream NER results.

unshuffled version of the French OSCAR corpus

d3d4eef047aa01391e3e5d613a0f1f786ae7cfc7

d3d4eef047aa01391e3e5d613a0f1f786ae7cfc7_0

Q: Which tasks does CamemBERT not improve on? Text: Introduction Pretrained word representations have a long history in Natural Language Processing (NLP), from non-neural methods BIBREF0, BIBREF1, BIBREF2 to neural word embeddings BIBREF3, BIBREF4 and to contextualised representations BIBREF5, BIBREF6. Approaches shifted more recently from using these representations as an input to task-specific architectures to replacing these architectures with large pretrained language models. These models are then fine-tuned to the task at hand with large improvements in performance over a wide range of tasks BIBREF7, BIBREF8, BIBREF9, BIBREF10. These transfer learning methods exhibit clear advantages over more traditional task-specific approaches, probably the most important being that they can be trained in an unsupervised manner. They nevertheless come with implementation challenges, namely the amount of data and computational resources needed for pretraining that can reach hundreds of gigabytes of uncompressed text and require hundreds of GPUs BIBREF11, BIBREF9. The latest transformer architecture has gone uses as much as 750GB of plain text and 1024 TPU v3 for pretraining BIBREF10. This has limited the availability of these state-of-the-art models to the English language, at least in the monolingual setting. Even though multilingual models give remarkable results, they are often larger and their results still lag behind their monolingual counterparts BIBREF12. This is particularly inconvenient as it hinders their practical use in NLP systems as well as the investigation of their language modeling capacity, something that remains to be investigated in the case of, for instance, morphologically rich languages. We take advantage of the newly available multilingual corpus OSCAR BIBREF13 and train a monolingual language model for French using the RoBERTa architecture. We pretrain the model - which we dub CamemBERT- and evaluate it in four different downstream tasks for French: part-of-speech (POS) tagging, dependency parsing, named entity recognition (NER) and natural language inference (NLI). CamemBERT improves the state of the art for most tasks over previous monolingual and multilingual approaches, which confirms the effectiveness of large pretrained language models for French. We summarise our contributions as follows: We train a monolingual BERT model on the French language using recent large-scale corpora. We evaluate our model on four downstream tasks (POS tagging, dependency parsing, NER and natural language inference (NLI)), achieving state-of-the-art results in most tasks, confirming the effectiveness of large BERT-based models for French. We release our model in a user-friendly format for popular open-source libraries so that it can serve as a strong baseline for future research and be useful for French NLP practitioners. Related Work ::: From non-contextual to contextual word embeddings The first neural word vector representations were non-contextualised word embeddings, most notably word2vec BIBREF3, GloVe BIBREF4 and fastText BIBREF14, which were designed to be used as input to task-specific neural architectures. Contextualised word representations such as ELMo BIBREF5 and flair BIBREF6, improved the expressivity of word embeddings by taking context into account. They improved the performance of downstream tasks when they replaced traditional word representations. This paved the way towards larger contextualised models that replaced downstream architectures in most tasks. These approaches, trained with language modeling objectives, range from LSTM-based architectures such as ULMFiT BIBREF15 to the successful transformer-based architectures such as GPT2 BIBREF8, BERT BIBREF7, RoBERTa BIBREF9 and more recently ALBERT BIBREF16 and T5 BIBREF10. Related Work ::: Non-contextual word embeddings for languages other than English Since the introduction of word2vec BIBREF3, many attempts have been made to create monolingual models for a wide range of languages. For non-contextual word embeddings, the first two attempts were by BIBREF17 and BIBREF18 who created word embeddings for a large number of languages using Wikipedia. Later BIBREF19 trained fastText word embeddings for 157 languages using Common Crawl and showed that using crawled data significantly increased the performance of the embeddings relatively to those trained only on Wikipedia. Related Work ::: Contextualised models for languages other than English Following the success of large pretrained language models, they were extended to the multilingual setting with multilingual BERT , a single multilingual model for 104 different languages trained on Wikipedia data, and later XLM BIBREF12, which greatly improved unsupervised machine translation. A few monolingual models have been released: ELMo models for Japanese, Portuguese, German and Basque and BERT for Simplified and Traditional Chinese and German. However, to the best of our knowledge, no particular effort has been made toward training models for languages other than English, at a scale similar to the latest English models (e.g. RoBERTa trained on more than 100GB of data). CamemBERT Our approach is based on RoBERTa BIBREF9, which replicates and improves the initial BERT by identifying key hyper-parameters for more robust performance. In this section, we describe the architecture, training objective, optimisation setup and pretraining data that was used for CamemBERT. CamemBERT differs from RoBERTa mainly with the addition of whole-word masking and the usage of SentencePiece tokenisation BIBREF20. CamemBERT ::: Architecture Similar to RoBERTa and BERT, CamemBERT is a multi-layer bidirectional Transformer BIBREF21. Given the widespread usage of Transformers, we do not describe them in detail here and refer the reader to BIBREF21. CamemBERT uses the original BERT $_{\small \textsc {BASE}}$ configuration: 12 layers, 768 hidden dimensions, 12 attention heads, which amounts to 110M parameters. CamemBERT ::: Pretraining objective We train our model on the Masked Language Modeling (MLM) task. Given an input text sequence composed of $N$ tokens $x_1, ..., x_N$, we select $15\%$ of tokens for possible replacement. Among those selected tokens, 80% are replaced with the special $<$mask$>$ token, 10% are left unchanged and 10% are replaced by a random token. The model is then trained to predict the initial masked tokens using cross-entropy loss. Following RoBERTa we dynamically mask tokens instead of fixing them statically for the whole dataset during preprocessing. This improves variability and makes the model more robust when training for multiple epochs. Since we segment the input sentence into subwords using SentencePiece, the input tokens to the models can be subwords. An upgraded version of BERT and BIBREF22 have shown that masking whole words instead of individual subwords leads to improved performance. Whole-word masking (WWM) makes the training task more difficult because the model has to predict a whole word instead of predicting only part of the word given the rest. As a result, we used WWM for CamemBERT by first randomly sampling 15% of the words in the sequence and then considering all subword tokens in each of these 15% words for candidate replacement. This amounts to a proportion of selected tokens that is close to the original 15%. These tokens are then either replaced by $<$mask$>$ tokens (80%), left unchanged (10%) or replaced by a random token. Subsequent work has shown that the next sentence prediction task (NSP) originally used in BERT does not improve downstream task performance BIBREF12, BIBREF9, we do not use NSP as a consequence. CamemBERT ::: Optimisation Following BIBREF9, we optimise the model using Adam BIBREF23 ($\beta _1 = 0.9$, $\beta _2 = 0.98$) for 100k steps. We use large batch sizes of 8192 sequences. Each sequence contains at most 512 tokens. We enforce each sequence to only contain complete sentences. Additionally, we used the DOC-SENTENCES scenario from BIBREF9, consisting of not mixing multiple documents in the same sequence, which showed slightly better results. CamemBERT ::: Segmentation into subword units We segment the input text into subword units using SentencePiece BIBREF20. SentencePiece is an extension of Byte-Pair encoding (BPE) BIBREF24 and WordPiece BIBREF25 that does not require pre-tokenisation (at the word or token level), thus removing the need for language-specific tokenisers. We use a vocabulary size of 32k subword tokens. These are learned on $10^7$ sentences sampled from the pretraining dataset. We do not use subword regularisation (i.e. sampling from multiple possible segmentations) in our implementation for simplicity. CamemBERT ::: Pretraining data Pretrained language models can be significantly improved by using more data BIBREF9, BIBREF10. Therefore we used French text extracted from Common Crawl, in particular, we use OSCAR BIBREF13 a pre-classified and pre-filtered version of the November 2018 Common Craw snapshot. OSCAR is a set of monolingual corpora extracted from Common Crawl, specifically from the plain text WET format distributed by Common Crawl, which removes all HTML tags and converts all text encodings to UTF-8. OSCAR follows the same approach as BIBREF19 by using a language classification model based on the fastText linear classifier BIBREF26, BIBREF27 pretrained on Wikipedia, Tatoeba and SETimes, which supports 176 different languages. OSCAR performs a deduplication step after language classification and without introducing a specialised filtering scheme, other than only keeping paragraphs containing 100 or more UTF-8 encoded characters, making OSCAR quite close to the original Crawled data. We use the unshuffled version of the French OSCAR corpus, which amounts to 138GB of uncompressed text and 32.7B SentencePiece tokens. Evaluation ::: Part-of-speech tagging and dependency parsing We fist evaluate CamemBERT on the two downstream tasks of part-of-speech (POS) tagging and dependency parsing. POS tagging is a low-level syntactic task, which consists in assigning to each word its corresponding grammatical category. Dependency parsing consists in predicting the labeled syntactic tree capturing the syntactic relations between words. We run our experiments using the Universal Dependencies (UD) paradigm and its corresponding UD POS tag set BIBREF28 and UD treebank collection version 2.2 BIBREF29, which was used for the CoNLL 2018 shared task. We perform our work on the four freely available French UD treebanks in UD v2.2: GSD, Sequoia, Spoken, and ParTUT. GSD BIBREF30 is the second-largest treebank available for French after the FTB (described in subsection SECREF25), it contains data from blogs, news articles, reviews, and Wikipedia. The Sequoia treebank BIBREF31, BIBREF32 comprises more than 3000 sentences, from the French Europarl, the regional newspaper L’Est Républicain, the French Wikipedia and documents from the European Medicines Agency. Spoken is a corpus converted automatically from the Rhapsodie treebank BIBREF33, BIBREF34 with manual corrections. It consists of 57 sound samples of spoken French with orthographic transcription and phonetic transcription aligned with sound (word boundaries, syllables, and phonemes), syntactic and prosodic annotations. Finally, ParTUT is a conversion of a multilingual parallel treebank developed at the University of Turin, and consisting of a variety of text genres, including talks, legal texts, and Wikipedia articles, among others; ParTUT data is derived from the already-existing parallel treebank Par(allel)TUT BIBREF35 . Table TABREF23 contains a summary comparing the sizes of the treebanks. We evaluate the performance of our models using the standard UPOS accuracy for POS tagging, and Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS) for dependency parsing. We assume gold tokenisation and gold word segmentation as provided in the UD treebanks. Evaluation ::: Part-of-speech tagging and dependency parsing ::: Baselines To demonstrate the value of building a dedicated version of BERT for French, we first compare CamemBERT to the multilingual cased version of BERT (designated as mBERT). We then compare our models to UDify BIBREF36. UDify is a multitask and multilingual model based on mBERT that is near state-of-the-art on all UD languages including French for both POS tagging and dependency parsing. It is relevant to compare CamemBERT to UDify on those tasks because UDify is the work that pushed the furthest the performance in fine-tuning end-to-end a BERT-based model on downstream POS tagging and dependency parsing. Finally, we compare our model to UDPipe Future BIBREF37, a model ranked 3rd in dependency parsing and 6th in POS tagging during the CoNLL 2018 shared task BIBREF38. UDPipe Future provides us a strong baseline that does not make use of any pretrained contextual embedding. We will compare to the more recent cross-lingual language model XLM BIBREF12, as well as the state-of-the-art CoNLL 2018 shared task results with predicted tokenisation and segmentation in an updated version of the paper. Evaluation ::: Named Entity Recognition Named Entity Recognition (NER) is a sequence labeling task that consists in predicting which words refer to real-world objects, such as people, locations, artifacts and organisations. We use the French Treebank (FTB) BIBREF39 in its 2008 version introduced by cc-clustering:09short and with NER annotations by sagot2012annotation. The NER-annotated FTB contains more than 12k sentences and more than 350k tokens extracted from articles of the newspaper Le Monde published between 1989 and 1995. In total, it contains 11,636 entity mentions distributed among 7 different types of entities, namely: 2025 mentions of “Person”, 3761 of “Location”, 2382 of “Organisation”, 3357 of “Company”, 67 of “Product”, 15 of “POI” (Point of Interest) and 29 of “Fictional Character”. A large proportion of the entity mentions in the treebank are multi-word entities. For NER we therefore report the 3 metrics that are commonly used to evaluate models: precision, recall, and F1 score. Here precision measures the percentage of entities found by the system that are correctly tagged, recall measures the percentage of named entities present in the corpus that are found and the F1 score combines both precision and recall measures giving a general idea of a model's performance. Evaluation ::: Named Entity Recognition ::: Baselines Most of the advances in NER haven been achieved on English, particularly focusing on the CoNLL 2003 BIBREF40 and the Ontonotes v5 BIBREF41, BIBREF42 English corpora. NER is a task that was traditionally tackled using Conditional Random Fields (CRF) BIBREF43 which are quite suited for NER; CRFs were later used as decoding layers for Bi-LSTM architectures BIBREF44, BIBREF45 showing considerable improvements over CRFs alone. These Bi-LSTM-CRF architectures were later enhanced with contextualised word embeddings which yet again brought major improvements to the task BIBREF5, BIBREF6. Finally, large pretrained architectures settled the current state of the art showing a small yet important improvement over previous NER-specific architectures BIBREF7, BIBREF46. In non-English NER the CoNLL 2002 shared task included NER corpora for Spanish and Dutch corpora BIBREF47 while the CoNLL 2003 included a German corpus BIBREF40. Here the recent efforts of BIBREF48 settled the state of the art for Spanish and Dutch, while BIBREF6 did it for German. In French, no extensive work has been done due to the limited availability of NER corpora. We compare our model with the strong baselines settled by BIBREF49, who trained both CRF and BiLSTM-CRF architectures on the FTB and enhanced them using heuristics and pretrained word embeddings. Evaluation ::: Natural Language Inference We also evaluate our model on the Natural Language Inference (NLI) task, using the French part of the XNLI dataset BIBREF50. NLI consists in predicting whether a hypothesis sentence is entailed, neutral or contradicts a premise sentence. The XNLI dataset is the extension of the Multi-Genre NLI (MultiNLI) corpus BIBREF51 to 15 languages by translating the validation and test sets manually into each of those languages. The English training set is also machine translated for all languages. The dataset is composed of 122k train, 2490 valid and 5010 test examples. As usual, NLI performance is evaluated using accuracy. To evaluate a model on a language other than English (such as French), we consider the two following settings: TRANSLATE-TEST: The French test set is machine translated into English, and then used with an English classification model. This setting provides a reasonable, although imperfect, way to circumvent the fact that no such data set exists for French, and results in very strong baseline scores. TRANSLATE-TRAIN: The French model is fine-tuned on the machine-translated English training set and then evaluated on the French test set. This is the setting that we used for CamemBERT. Evaluation ::: Natural Language Inference ::: Baselines For the TRANSLATE-TEST setting, we report results of the English RoBERTa to act as a reference. In the TRANSLATE-TRAIN setting, we report the best scores from previous literature along with ours. BiLSTM-max is the best model in the original XNLI paper, mBERT which has been reported in French in BIBREF52 and XLM (MLM+TLM) is the best-presented model from BIBREF50. Experiments In this section, we measure the performance of CamemBERT by evaluating it on the four aforementioned tasks: POS tagging, dependency parsing, NER and NLI. Experiments ::: Experimental Setup ::: Pretraining We use the RoBERTa implementation in the fairseq library BIBREF53. Our learning rate is warmed up for 10k steps up to a peak value of $0.0007$ instead of the original $0.0001$ given our large batch size (8192). The learning rate fades to zero with polynomial decay. We pretrain our model on 256 Nvidia V100 GPUs (32GB each) for 100k steps during 17h. Experiments ::: Experimental Setup ::: Fine-tuning For each task, we append the relevant predictive layer on top of CamemBERT's Transformer architecture. Following the work done on BERT BIBREF7, for sequence tagging and sequence labeling we append a linear layer respectively to the $<$s$>$ special token and to the first subword token of each word. For dependency parsing, we plug a bi-affine graph predictor head as inspired by BIBREF54 following the work done on multilingual parsing with BERT by BIBREF36. We refer the reader to these two articles for more details on this module. We fine-tune independently CamemBERT for each task and each dataset. We optimise the model using the Adam optimiser BIBREF23 with a fixed learning rate. We run a grid search on a combination of learning rates and batch sizes. We select the best model on the validation set out of the 30 first epochs. Although this might push the performances even further, for all tasks except NLI, we don't apply any regularisation techniques such as weight decay, learning rate warm-up or discriminative fine-tuning. We show that fine-tuning CamemBERT in a straight-forward manner leads to state-of-the-art results on most tasks and outperforms the existing BERT-based models in most cases. The POS tagging, dependency parsing, and NER experiments are run using hugging face's Transformer library extended to support CamemBERT and dependency parsing BIBREF55. The NLI experiments use the fairseq library following the RoBERTa implementation. Experiments ::: Results ::: Part-of-Speech tagging and dependency parsing For POS tagging and dependency parsing, we compare CamemBERT to three other near state-of-the-art models in Table TABREF32. CamemBERT outperforms UDPipe Future by a large margin for all treebanks and all metrics. Despite a much simpler optimisation process, CamemBERT beats UDify performances on all the available French treebanks. CamemBERT also demonstrates higher performances than mBERT on those tasks. We observe a larger error reduction for parsing than for tagging. For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT. For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT. Experiments ::: Results ::: Natural Language Inference: XNLI On the XNLI benchmark, CamemBERT obtains improved performance over multilingual language models on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM) while using less than half the parameters (110M vs. 250M). However, its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa. It should be noted that CamemBERT uses far fewer parameters than RoBERTa (110M vs. 355M parameters). Experiments ::: Results ::: Named-Entity Recognition For named entity recognition, our experiments show that CamemBERT achieves a slightly better precision than the traditional CRF-based SEM architectures described above in Section SECREF25 (CRF and Bi-LSTM+CRF), but shows a dramatic improvement in finding entity mentions, raising the recall score by 3.5 points. Both improvements result in a 2.36 point increase in the F1 score with respect to the best SEM architecture (BiLSTM-CRF), giving CamemBERT the state of the art for NER on the FTB. One other important finding is the results obtained by mBERT. Previous work with this model showed increased performance in NER for German, Dutch and Spanish when mBERT is used as contextualised word embedding for an NER-specific model BIBREF48, but our results suggest that the multilingual setting in which mBERT was trained is simply not enough to use it alone and fine-tune it for French NER, as it shows worse performance than even simple CRF models, suggesting that monolingual models could be better at NER. Experiments ::: Discussion CamemBERT displays improved performance compared to prior work for the 4 downstream tasks considered. This confirms the hypothesis that pretrained language models can be effectively fine-tuned for various downstream tasks, as observed for English in previous work. Moreover, our results also show that dedicated monolingual models still outperform multilingual ones. We explain this point in two ways. First, the scale of data is possibly essential to the performance of CamemBERT. Indeed, we use 138GB of uncompressed text vs. 57GB for mBERT. Second, with more data comes more diversity in the pretraining distribution. Reaching state-of-the-art performances on 4 different tasks and 6 different datasets requires robust pretrained models. Our results suggest that the variability in the downstream tasks and datasets considered is handled more efficiently by a general language model than by Wikipedia-pretrained models such as mBERT. Conclusion CamemBERT improves the state of the art for multiple downstream tasks in French. It is also lighter than other BERT-based approaches such as mBERT or XLM. By releasing our model, we hope that it can serve as a strong baseline for future research in French NLP, and expect our experiments to be reproduced in many other languages. We will publish an updated version in the near future where we will explore and release models trained for longer, with additional downstream tasks, baselines (e.g. XLM) and analysis, we will also train additional models with potentially cleaner corpora such as CCNet BIBREF56 for more accurate performance evaluation and more complete ablation. Acknowledgments This work was partly funded by three French National grants from the Agence Nationale de la Recherche, namely projects PARSITI (ANR-16-CE33-0021), SoSweet (ANR-15-CE38-0011) and BASNUM (ANR-18-CE38-0003), as well as by the last author's chair in the PRAIRIE institute. Appendix ::: Impact of Whole-Word Masking We analyze the addition of whole-word masking on the downstream performance of CamemBERT. As reported for English on other downstream tasks, whole word masking improves downstream performances for all tasks but NER as seen in Table TABREF46. NER is highly sensitive to capitalisation, prefixes, suffixes and other subword features that could be used by a model to correctly identify entity mentions. Thus the added information by learning the masking at a subword level rather than at whole-word level seems to have a detrimental effect on downstream NER results.

its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa

63723c6b398100bba5dc21754451f503cb91c9b8

63723c6b398100bba5dc21754451f503cb91c9b8_0

Q: What is the state of the art? Text: Introduction Pretrained word representations have a long history in Natural Language Processing (NLP), from non-neural methods BIBREF0, BIBREF1, BIBREF2 to neural word embeddings BIBREF3, BIBREF4 and to contextualised representations BIBREF5, BIBREF6. Approaches shifted more recently from using these representations as an input to task-specific architectures to replacing these architectures with large pretrained language models. These models are then fine-tuned to the task at hand with large improvements in performance over a wide range of tasks BIBREF7, BIBREF8, BIBREF9, BIBREF10. These transfer learning methods exhibit clear advantages over more traditional task-specific approaches, probably the most important being that they can be trained in an unsupervised manner. They nevertheless come with implementation challenges, namely the amount of data and computational resources needed for pretraining that can reach hundreds of gigabytes of uncompressed text and require hundreds of GPUs BIBREF11, BIBREF9. The latest transformer architecture has gone uses as much as 750GB of plain text and 1024 TPU v3 for pretraining BIBREF10. This has limited the availability of these state-of-the-art models to the English language, at least in the monolingual setting. Even though multilingual models give remarkable results, they are often larger and their results still lag behind their monolingual counterparts BIBREF12. This is particularly inconvenient as it hinders their practical use in NLP systems as well as the investigation of their language modeling capacity, something that remains to be investigated in the case of, for instance, morphologically rich languages. We take advantage of the newly available multilingual corpus OSCAR BIBREF13 and train a monolingual language model for French using the RoBERTa architecture. We pretrain the model - which we dub CamemBERT- and evaluate it in four different downstream tasks for French: part-of-speech (POS) tagging, dependency parsing, named entity recognition (NER) and natural language inference (NLI). CamemBERT improves the state of the art for most tasks over previous monolingual and multilingual approaches, which confirms the effectiveness of large pretrained language models for French. We summarise our contributions as follows: We train a monolingual BERT model on the French language using recent large-scale corpora. We evaluate our model on four downstream tasks (POS tagging, dependency parsing, NER and natural language inference (NLI)), achieving state-of-the-art results in most tasks, confirming the effectiveness of large BERT-based models for French. We release our model in a user-friendly format for popular open-source libraries so that it can serve as a strong baseline for future research and be useful for French NLP practitioners. Related Work ::: From non-contextual to contextual word embeddings The first neural word vector representations were non-contextualised word embeddings, most notably word2vec BIBREF3, GloVe BIBREF4 and fastText BIBREF14, which were designed to be used as input to task-specific neural architectures. Contextualised word representations such as ELMo BIBREF5 and flair BIBREF6, improved the expressivity of word embeddings by taking context into account. They improved the performance of downstream tasks when they replaced traditional word representations. This paved the way towards larger contextualised models that replaced downstream architectures in most tasks. These approaches, trained with language modeling objectives, range from LSTM-based architectures such as ULMFiT BIBREF15 to the successful transformer-based architectures such as GPT2 BIBREF8, BERT BIBREF7, RoBERTa BIBREF9 and more recently ALBERT BIBREF16 and T5 BIBREF10. Related Work ::: Non-contextual word embeddings for languages other than English Since the introduction of word2vec BIBREF3, many attempts have been made to create monolingual models for a wide range of languages. For non-contextual word embeddings, the first two attempts were by BIBREF17 and BIBREF18 who created word embeddings for a large number of languages using Wikipedia. Later BIBREF19 trained fastText word embeddings for 157 languages using Common Crawl and showed that using crawled data significantly increased the performance of the embeddings relatively to those trained only on Wikipedia. Related Work ::: Contextualised models for languages other than English Following the success of large pretrained language models, they were extended to the multilingual setting with multilingual BERT , a single multilingual model for 104 different languages trained on Wikipedia data, and later XLM BIBREF12, which greatly improved unsupervised machine translation. A few monolingual models have been released: ELMo models for Japanese, Portuguese, German and Basque and BERT for Simplified and Traditional Chinese and German. However, to the best of our knowledge, no particular effort has been made toward training models for languages other than English, at a scale similar to the latest English models (e.g. RoBERTa trained on more than 100GB of data). CamemBERT Our approach is based on RoBERTa BIBREF9, which replicates and improves the initial BERT by identifying key hyper-parameters for more robust performance. In this section, we describe the architecture, training objective, optimisation setup and pretraining data that was used for CamemBERT. CamemBERT differs from RoBERTa mainly with the addition of whole-word masking and the usage of SentencePiece tokenisation BIBREF20. CamemBERT ::: Architecture Similar to RoBERTa and BERT, CamemBERT is a multi-layer bidirectional Transformer BIBREF21. Given the widespread usage of Transformers, we do not describe them in detail here and refer the reader to BIBREF21. CamemBERT uses the original BERT $_{\small \textsc {BASE}}$ configuration: 12 layers, 768 hidden dimensions, 12 attention heads, which amounts to 110M parameters. CamemBERT ::: Pretraining objective We train our model on the Masked Language Modeling (MLM) task. Given an input text sequence composed of $N$ tokens $x_1, ..., x_N$, we select $15\%$ of tokens for possible replacement. Among those selected tokens, 80% are replaced with the special $<$mask$>$ token, 10% are left unchanged and 10% are replaced by a random token. The model is then trained to predict the initial masked tokens using cross-entropy loss. Following RoBERTa we dynamically mask tokens instead of fixing them statically for the whole dataset during preprocessing. This improves variability and makes the model more robust when training for multiple epochs. Since we segment the input sentence into subwords using SentencePiece, the input tokens to the models can be subwords. An upgraded version of BERT and BIBREF22 have shown that masking whole words instead of individual subwords leads to improved performance. Whole-word masking (WWM) makes the training task more difficult because the model has to predict a whole word instead of predicting only part of the word given the rest. As a result, we used WWM for CamemBERT by first randomly sampling 15% of the words in the sequence and then considering all subword tokens in each of these 15% words for candidate replacement. This amounts to a proportion of selected tokens that is close to the original 15%. These tokens are then either replaced by $<$mask$>$ tokens (80%), left unchanged (10%) or replaced by a random token. Subsequent work has shown that the next sentence prediction task (NSP) originally used in BERT does not improve downstream task performance BIBREF12, BIBREF9, we do not use NSP as a consequence. CamemBERT ::: Optimisation Following BIBREF9, we optimise the model using Adam BIBREF23 ($\beta _1 = 0.9$, $\beta _2 = 0.98$) for 100k steps. We use large batch sizes of 8192 sequences. Each sequence contains at most 512 tokens. We enforce each sequence to only contain complete sentences. Additionally, we used the DOC-SENTENCES scenario from BIBREF9, consisting of not mixing multiple documents in the same sequence, which showed slightly better results. CamemBERT ::: Segmentation into subword units We segment the input text into subword units using SentencePiece BIBREF20. SentencePiece is an extension of Byte-Pair encoding (BPE) BIBREF24 and WordPiece BIBREF25 that does not require pre-tokenisation (at the word or token level), thus removing the need for language-specific tokenisers. We use a vocabulary size of 32k subword tokens. These are learned on $10^7$ sentences sampled from the pretraining dataset. We do not use subword regularisation (i.e. sampling from multiple possible segmentations) in our implementation for simplicity. CamemBERT ::: Pretraining data Pretrained language models can be significantly improved by using more data BIBREF9, BIBREF10. Therefore we used French text extracted from Common Crawl, in particular, we use OSCAR BIBREF13 a pre-classified and pre-filtered version of the November 2018 Common Craw snapshot. OSCAR is a set of monolingual corpora extracted from Common Crawl, specifically from the plain text WET format distributed by Common Crawl, which removes all HTML tags and converts all text encodings to UTF-8. OSCAR follows the same approach as BIBREF19 by using a language classification model based on the fastText linear classifier BIBREF26, BIBREF27 pretrained on Wikipedia, Tatoeba and SETimes, which supports 176 different languages. OSCAR performs a deduplication step after language classification and without introducing a specialised filtering scheme, other than only keeping paragraphs containing 100 or more UTF-8 encoded characters, making OSCAR quite close to the original Crawled data. We use the unshuffled version of the French OSCAR corpus, which amounts to 138GB of uncompressed text and 32.7B SentencePiece tokens. Evaluation ::: Part-of-speech tagging and dependency parsing We fist evaluate CamemBERT on the two downstream tasks of part-of-speech (POS) tagging and dependency parsing. POS tagging is a low-level syntactic task, which consists in assigning to each word its corresponding grammatical category. Dependency parsing consists in predicting the labeled syntactic tree capturing the syntactic relations between words. We run our experiments using the Universal Dependencies (UD) paradigm and its corresponding UD POS tag set BIBREF28 and UD treebank collection version 2.2 BIBREF29, which was used for the CoNLL 2018 shared task. We perform our work on the four freely available French UD treebanks in UD v2.2: GSD, Sequoia, Spoken, and ParTUT. GSD BIBREF30 is the second-largest treebank available for French after the FTB (described in subsection SECREF25), it contains data from blogs, news articles, reviews, and Wikipedia. The Sequoia treebank BIBREF31, BIBREF32 comprises more than 3000 sentences, from the French Europarl, the regional newspaper L’Est Républicain, the French Wikipedia and documents from the European Medicines Agency. Spoken is a corpus converted automatically from the Rhapsodie treebank BIBREF33, BIBREF34 with manual corrections. It consists of 57 sound samples of spoken French with orthographic transcription and phonetic transcription aligned with sound (word boundaries, syllables, and phonemes), syntactic and prosodic annotations. Finally, ParTUT is a conversion of a multilingual parallel treebank developed at the University of Turin, and consisting of a variety of text genres, including talks, legal texts, and Wikipedia articles, among others; ParTUT data is derived from the already-existing parallel treebank Par(allel)TUT BIBREF35 . Table TABREF23 contains a summary comparing the sizes of the treebanks. We evaluate the performance of our models using the standard UPOS accuracy for POS tagging, and Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS) for dependency parsing. We assume gold tokenisation and gold word segmentation as provided in the UD treebanks. Evaluation ::: Part-of-speech tagging and dependency parsing ::: Baselines To demonstrate the value of building a dedicated version of BERT for French, we first compare CamemBERT to the multilingual cased version of BERT (designated as mBERT). We then compare our models to UDify BIBREF36. UDify is a multitask and multilingual model based on mBERT that is near state-of-the-art on all UD languages including French for both POS tagging and dependency parsing. It is relevant to compare CamemBERT to UDify on those tasks because UDify is the work that pushed the furthest the performance in fine-tuning end-to-end a BERT-based model on downstream POS tagging and dependency parsing. Finally, we compare our model to UDPipe Future BIBREF37, a model ranked 3rd in dependency parsing and 6th in POS tagging during the CoNLL 2018 shared task BIBREF38. UDPipe Future provides us a strong baseline that does not make use of any pretrained contextual embedding. We will compare to the more recent cross-lingual language model XLM BIBREF12, as well as the state-of-the-art CoNLL 2018 shared task results with predicted tokenisation and segmentation in an updated version of the paper. Evaluation ::: Named Entity Recognition Named Entity Recognition (NER) is a sequence labeling task that consists in predicting which words refer to real-world objects, such as people, locations, artifacts and organisations. We use the French Treebank (FTB) BIBREF39 in its 2008 version introduced by cc-clustering:09short and with NER annotations by sagot2012annotation. The NER-annotated FTB contains more than 12k sentences and more than 350k tokens extracted from articles of the newspaper Le Monde published between 1989 and 1995. In total, it contains 11,636 entity mentions distributed among 7 different types of entities, namely: 2025 mentions of “Person”, 3761 of “Location”, 2382 of “Organisation”, 3357 of “Company”, 67 of “Product”, 15 of “POI” (Point of Interest) and 29 of “Fictional Character”. A large proportion of the entity mentions in the treebank are multi-word entities. For NER we therefore report the 3 metrics that are commonly used to evaluate models: precision, recall, and F1 score. Here precision measures the percentage of entities found by the system that are correctly tagged, recall measures the percentage of named entities present in the corpus that are found and the F1 score combines both precision and recall measures giving a general idea of a model's performance. Evaluation ::: Named Entity Recognition ::: Baselines Most of the advances in NER haven been achieved on English, particularly focusing on the CoNLL 2003 BIBREF40 and the Ontonotes v5 BIBREF41, BIBREF42 English corpora. NER is a task that was traditionally tackled using Conditional Random Fields (CRF) BIBREF43 which are quite suited for NER; CRFs were later used as decoding layers for Bi-LSTM architectures BIBREF44, BIBREF45 showing considerable improvements over CRFs alone. These Bi-LSTM-CRF architectures were later enhanced with contextualised word embeddings which yet again brought major improvements to the task BIBREF5, BIBREF6. Finally, large pretrained architectures settled the current state of the art showing a small yet important improvement over previous NER-specific architectures BIBREF7, BIBREF46. In non-English NER the CoNLL 2002 shared task included NER corpora for Spanish and Dutch corpora BIBREF47 while the CoNLL 2003 included a German corpus BIBREF40. Here the recent efforts of BIBREF48 settled the state of the art for Spanish and Dutch, while BIBREF6 did it for German. In French, no extensive work has been done due to the limited availability of NER corpora. We compare our model with the strong baselines settled by BIBREF49, who trained both CRF and BiLSTM-CRF architectures on the FTB and enhanced them using heuristics and pretrained word embeddings. Evaluation ::: Natural Language Inference We also evaluate our model on the Natural Language Inference (NLI) task, using the French part of the XNLI dataset BIBREF50. NLI consists in predicting whether a hypothesis sentence is entailed, neutral or contradicts a premise sentence. The XNLI dataset is the extension of the Multi-Genre NLI (MultiNLI) corpus BIBREF51 to 15 languages by translating the validation and test sets manually into each of those languages. The English training set is also machine translated for all languages. The dataset is composed of 122k train, 2490 valid and 5010 test examples. As usual, NLI performance is evaluated using accuracy. To evaluate a model on a language other than English (such as French), we consider the two following settings: TRANSLATE-TEST: The French test set is machine translated into English, and then used with an English classification model. This setting provides a reasonable, although imperfect, way to circumvent the fact that no such data set exists for French, and results in very strong baseline scores. TRANSLATE-TRAIN: The French model is fine-tuned on the machine-translated English training set and then evaluated on the French test set. This is the setting that we used for CamemBERT. Evaluation ::: Natural Language Inference ::: Baselines For the TRANSLATE-TEST setting, we report results of the English RoBERTa to act as a reference. In the TRANSLATE-TRAIN setting, we report the best scores from previous literature along with ours. BiLSTM-max is the best model in the original XNLI paper, mBERT which has been reported in French in BIBREF52 and XLM (MLM+TLM) is the best-presented model from BIBREF50. Experiments In this section, we measure the performance of CamemBERT by evaluating it on the four aforementioned tasks: POS tagging, dependency parsing, NER and NLI. Experiments ::: Experimental Setup ::: Pretraining We use the RoBERTa implementation in the fairseq library BIBREF53. Our learning rate is warmed up for 10k steps up to a peak value of $0.0007$ instead of the original $0.0001$ given our large batch size (8192). The learning rate fades to zero with polynomial decay. We pretrain our model on 256 Nvidia V100 GPUs (32GB each) for 100k steps during 17h. Experiments ::: Experimental Setup ::: Fine-tuning For each task, we append the relevant predictive layer on top of CamemBERT's Transformer architecture. Following the work done on BERT BIBREF7, for sequence tagging and sequence labeling we append a linear layer respectively to the $<$s$>$ special token and to the first subword token of each word. For dependency parsing, we plug a bi-affine graph predictor head as inspired by BIBREF54 following the work done on multilingual parsing with BERT by BIBREF36. We refer the reader to these two articles for more details on this module. We fine-tune independently CamemBERT for each task and each dataset. We optimise the model using the Adam optimiser BIBREF23 with a fixed learning rate. We run a grid search on a combination of learning rates and batch sizes. We select the best model on the validation set out of the 30 first epochs. Although this might push the performances even further, for all tasks except NLI, we don't apply any regularisation techniques such as weight decay, learning rate warm-up or discriminative fine-tuning. We show that fine-tuning CamemBERT in a straight-forward manner leads to state-of-the-art results on most tasks and outperforms the existing BERT-based models in most cases. The POS tagging, dependency parsing, and NER experiments are run using hugging face's Transformer library extended to support CamemBERT and dependency parsing BIBREF55. The NLI experiments use the fairseq library following the RoBERTa implementation. Experiments ::: Results ::: Part-of-Speech tagging and dependency parsing For POS tagging and dependency parsing, we compare CamemBERT to three other near state-of-the-art models in Table TABREF32. CamemBERT outperforms UDPipe Future by a large margin for all treebanks and all metrics. Despite a much simpler optimisation process, CamemBERT beats UDify performances on all the available French treebanks. CamemBERT also demonstrates higher performances than mBERT on those tasks. We observe a larger error reduction for parsing than for tagging. For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT. For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT. Experiments ::: Results ::: Natural Language Inference: XNLI On the XNLI benchmark, CamemBERT obtains improved performance over multilingual language models on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM) while using less than half the parameters (110M vs. 250M). However, its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa. It should be noted that CamemBERT uses far fewer parameters than RoBERTa (110M vs. 355M parameters). Experiments ::: Results ::: Named-Entity Recognition For named entity recognition, our experiments show that CamemBERT achieves a slightly better precision than the traditional CRF-based SEM architectures described above in Section SECREF25 (CRF and Bi-LSTM+CRF), but shows a dramatic improvement in finding entity mentions, raising the recall score by 3.5 points. Both improvements result in a 2.36 point increase in the F1 score with respect to the best SEM architecture (BiLSTM-CRF), giving CamemBERT the state of the art for NER on the FTB. One other important finding is the results obtained by mBERT. Previous work with this model showed increased performance in NER for German, Dutch and Spanish when mBERT is used as contextualised word embedding for an NER-specific model BIBREF48, but our results suggest that the multilingual setting in which mBERT was trained is simply not enough to use it alone and fine-tune it for French NER, as it shows worse performance than even simple CRF models, suggesting that monolingual models could be better at NER. Experiments ::: Discussion CamemBERT displays improved performance compared to prior work for the 4 downstream tasks considered. This confirms the hypothesis that pretrained language models can be effectively fine-tuned for various downstream tasks, as observed for English in previous work. Moreover, our results also show that dedicated monolingual models still outperform multilingual ones. We explain this point in two ways. First, the scale of data is possibly essential to the performance of CamemBERT. Indeed, we use 138GB of uncompressed text vs. 57GB for mBERT. Second, with more data comes more diversity in the pretraining distribution. Reaching state-of-the-art performances on 4 different tasks and 6 different datasets requires robust pretrained models. Our results suggest that the variability in the downstream tasks and datasets considered is handled more efficiently by a general language model than by Wikipedia-pretrained models such as mBERT. Conclusion CamemBERT improves the state of the art for multiple downstream tasks in French. It is also lighter than other BERT-based approaches such as mBERT or XLM. By releasing our model, we hope that it can serve as a strong baseline for future research in French NLP, and expect our experiments to be reproduced in many other languages. We will publish an updated version in the near future where we will explore and release models trained for longer, with additional downstream tasks, baselines (e.g. XLM) and analysis, we will also train additional models with potentially cleaner corpora such as CCNet BIBREF56 for more accurate performance evaluation and more complete ablation. Acknowledgments This work was partly funded by three French National grants from the Agence Nationale de la Recherche, namely projects PARSITI (ANR-16-CE33-0021), SoSweet (ANR-15-CE38-0011) and BASNUM (ANR-18-CE38-0003), as well as by the last author's chair in the PRAIRIE institute. Appendix ::: Impact of Whole-Word Masking We analyze the addition of whole-word masking on the downstream performance of CamemBERT. As reported for English on other downstream tasks, whole word masking improves downstream performances for all tasks but NER as seen in Table TABREF46. NER is highly sensitive to capitalisation, prefixes, suffixes and other subword features that could be used by a model to correctly identify entity mentions. Thus the added information by learning the masking at a subword level rather than at whole-word level seems to have a detrimental effect on downstream NER results.

POS and DP task: CONLL 2018 NER task: (no extensive work) Strong baselines CRF and BiLSTM-CRF NLI task: mBERT or XLM (not clear from text)

5471766ca7c995dd7f0f449407902b32ac9db269

5471766ca7c995dd7f0f449407902b32ac9db269_0

Q: How much better was results of CamemBERT than previous results on these tasks? Text: Introduction Pretrained word representations have a long history in Natural Language Processing (NLP), from non-neural methods BIBREF0, BIBREF1, BIBREF2 to neural word embeddings BIBREF3, BIBREF4 and to contextualised representations BIBREF5, BIBREF6. Approaches shifted more recently from using these representations as an input to task-specific architectures to replacing these architectures with large pretrained language models. These models are then fine-tuned to the task at hand with large improvements in performance over a wide range of tasks BIBREF7, BIBREF8, BIBREF9, BIBREF10. These transfer learning methods exhibit clear advantages over more traditional task-specific approaches, probably the most important being that they can be trained in an unsupervised manner. They nevertheless come with implementation challenges, namely the amount of data and computational resources needed for pretraining that can reach hundreds of gigabytes of uncompressed text and require hundreds of GPUs BIBREF11, BIBREF9. The latest transformer architecture has gone uses as much as 750GB of plain text and 1024 TPU v3 for pretraining BIBREF10. This has limited the availability of these state-of-the-art models to the English language, at least in the monolingual setting. Even though multilingual models give remarkable results, they are often larger and their results still lag behind their monolingual counterparts BIBREF12. This is particularly inconvenient as it hinders their practical use in NLP systems as well as the investigation of their language modeling capacity, something that remains to be investigated in the case of, for instance, morphologically rich languages. We take advantage of the newly available multilingual corpus OSCAR BIBREF13 and train a monolingual language model for French using the RoBERTa architecture. We pretrain the model - which we dub CamemBERT- and evaluate it in four different downstream tasks for French: part-of-speech (POS) tagging, dependency parsing, named entity recognition (NER) and natural language inference (NLI). CamemBERT improves the state of the art for most tasks over previous monolingual and multilingual approaches, which confirms the effectiveness of large pretrained language models for French. We summarise our contributions as follows: We train a monolingual BERT model on the French language using recent large-scale corpora. We evaluate our model on four downstream tasks (POS tagging, dependency parsing, NER and natural language inference (NLI)), achieving state-of-the-art results in most tasks, confirming the effectiveness of large BERT-based models for French. We release our model in a user-friendly format for popular open-source libraries so that it can serve as a strong baseline for future research and be useful for French NLP practitioners. Related Work ::: From non-contextual to contextual word embeddings The first neural word vector representations were non-contextualised word embeddings, most notably word2vec BIBREF3, GloVe BIBREF4 and fastText BIBREF14, which were designed to be used as input to task-specific neural architectures. Contextualised word representations such as ELMo BIBREF5 and flair BIBREF6, improved the expressivity of word embeddings by taking context into account. They improved the performance of downstream tasks when they replaced traditional word representations. This paved the way towards larger contextualised models that replaced downstream architectures in most tasks. These approaches, trained with language modeling objectives, range from LSTM-based architectures such as ULMFiT BIBREF15 to the successful transformer-based architectures such as GPT2 BIBREF8, BERT BIBREF7, RoBERTa BIBREF9 and more recently ALBERT BIBREF16 and T5 BIBREF10. Related Work ::: Non-contextual word embeddings for languages other than English Since the introduction of word2vec BIBREF3, many attempts have been made to create monolingual models for a wide range of languages. For non-contextual word embeddings, the first two attempts were by BIBREF17 and BIBREF18 who created word embeddings for a large number of languages using Wikipedia. Later BIBREF19 trained fastText word embeddings for 157 languages using Common Crawl and showed that using crawled data significantly increased the performance of the embeddings relatively to those trained only on Wikipedia. Related Work ::: Contextualised models for languages other than English Following the success of large pretrained language models, they were extended to the multilingual setting with multilingual BERT , a single multilingual model for 104 different languages trained on Wikipedia data, and later XLM BIBREF12, which greatly improved unsupervised machine translation. A few monolingual models have been released: ELMo models for Japanese, Portuguese, German and Basque and BERT for Simplified and Traditional Chinese and German. However, to the best of our knowledge, no particular effort has been made toward training models for languages other than English, at a scale similar to the latest English models (e.g. RoBERTa trained on more than 100GB of data). CamemBERT Our approach is based on RoBERTa BIBREF9, which replicates and improves the initial BERT by identifying key hyper-parameters for more robust performance. In this section, we describe the architecture, training objective, optimisation setup and pretraining data that was used for CamemBERT. CamemBERT differs from RoBERTa mainly with the addition of whole-word masking and the usage of SentencePiece tokenisation BIBREF20. CamemBERT ::: Architecture Similar to RoBERTa and BERT, CamemBERT is a multi-layer bidirectional Transformer BIBREF21. Given the widespread usage of Transformers, we do not describe them in detail here and refer the reader to BIBREF21. CamemBERT uses the original BERT $_{\small \textsc {BASE}}$ configuration: 12 layers, 768 hidden dimensions, 12 attention heads, which amounts to 110M parameters. CamemBERT ::: Pretraining objective We train our model on the Masked Language Modeling (MLM) task. Given an input text sequence composed of $N$ tokens $x_1, ..., x_N$, we select $15\%$ of tokens for possible replacement. Among those selected tokens, 80% are replaced with the special $<$mask$>$ token, 10% are left unchanged and 10% are replaced by a random token. The model is then trained to predict the initial masked tokens using cross-entropy loss. Following RoBERTa we dynamically mask tokens instead of fixing them statically for the whole dataset during preprocessing. This improves variability and makes the model more robust when training for multiple epochs. Since we segment the input sentence into subwords using SentencePiece, the input tokens to the models can be subwords. An upgraded version of BERT and BIBREF22 have shown that masking whole words instead of individual subwords leads to improved performance. Whole-word masking (WWM) makes the training task more difficult because the model has to predict a whole word instead of predicting only part of the word given the rest. As a result, we used WWM for CamemBERT by first randomly sampling 15% of the words in the sequence and then considering all subword tokens in each of these 15% words for candidate replacement. This amounts to a proportion of selected tokens that is close to the original 15%. These tokens are then either replaced by $<$mask$>$ tokens (80%), left unchanged (10%) or replaced by a random token. Subsequent work has shown that the next sentence prediction task (NSP) originally used in BERT does not improve downstream task performance BIBREF12, BIBREF9, we do not use NSP as a consequence. CamemBERT ::: Optimisation Following BIBREF9, we optimise the model using Adam BIBREF23 ($\beta _1 = 0.9$, $\beta _2 = 0.98$) for 100k steps. We use large batch sizes of 8192 sequences. Each sequence contains at most 512 tokens. We enforce each sequence to only contain complete sentences. Additionally, we used the DOC-SENTENCES scenario from BIBREF9, consisting of not mixing multiple documents in the same sequence, which showed slightly better results. CamemBERT ::: Segmentation into subword units We segment the input text into subword units using SentencePiece BIBREF20. SentencePiece is an extension of Byte-Pair encoding (BPE) BIBREF24 and WordPiece BIBREF25 that does not require pre-tokenisation (at the word or token level), thus removing the need for language-specific tokenisers. We use a vocabulary size of 32k subword tokens. These are learned on $10^7$ sentences sampled from the pretraining dataset. We do not use subword regularisation (i.e. sampling from multiple possible segmentations) in our implementation for simplicity. CamemBERT ::: Pretraining data Pretrained language models can be significantly improved by using more data BIBREF9, BIBREF10. Therefore we used French text extracted from Common Crawl, in particular, we use OSCAR BIBREF13 a pre-classified and pre-filtered version of the November 2018 Common Craw snapshot. OSCAR is a set of monolingual corpora extracted from Common Crawl, specifically from the plain text WET format distributed by Common Crawl, which removes all HTML tags and converts all text encodings to UTF-8. OSCAR follows the same approach as BIBREF19 by using a language classification model based on the fastText linear classifier BIBREF26, BIBREF27 pretrained on Wikipedia, Tatoeba and SETimes, which supports 176 different languages. OSCAR performs a deduplication step after language classification and without introducing a specialised filtering scheme, other than only keeping paragraphs containing 100 or more UTF-8 encoded characters, making OSCAR quite close to the original Crawled data. We use the unshuffled version of the French OSCAR corpus, which amounts to 138GB of uncompressed text and 32.7B SentencePiece tokens. Evaluation ::: Part-of-speech tagging and dependency parsing We fist evaluate CamemBERT on the two downstream tasks of part-of-speech (POS) tagging and dependency parsing. POS tagging is a low-level syntactic task, which consists in assigning to each word its corresponding grammatical category. Dependency parsing consists in predicting the labeled syntactic tree capturing the syntactic relations between words. We run our experiments using the Universal Dependencies (UD) paradigm and its corresponding UD POS tag set BIBREF28 and UD treebank collection version 2.2 BIBREF29, which was used for the CoNLL 2018 shared task. We perform our work on the four freely available French UD treebanks in UD v2.2: GSD, Sequoia, Spoken, and ParTUT. GSD BIBREF30 is the second-largest treebank available for French after the FTB (described in subsection SECREF25), it contains data from blogs, news articles, reviews, and Wikipedia. The Sequoia treebank BIBREF31, BIBREF32 comprises more than 3000 sentences, from the French Europarl, the regional newspaper L’Est Républicain, the French Wikipedia and documents from the European Medicines Agency. Spoken is a corpus converted automatically from the Rhapsodie treebank BIBREF33, BIBREF34 with manual corrections. It consists of 57 sound samples of spoken French with orthographic transcription and phonetic transcription aligned with sound (word boundaries, syllables, and phonemes), syntactic and prosodic annotations. Finally, ParTUT is a conversion of a multilingual parallel treebank developed at the University of Turin, and consisting of a variety of text genres, including talks, legal texts, and Wikipedia articles, among others; ParTUT data is derived from the already-existing parallel treebank Par(allel)TUT BIBREF35 . Table TABREF23 contains a summary comparing the sizes of the treebanks. We evaluate the performance of our models using the standard UPOS accuracy for POS tagging, and Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS) for dependency parsing. We assume gold tokenisation and gold word segmentation as provided in the UD treebanks. Evaluation ::: Part-of-speech tagging and dependency parsing ::: Baselines To demonstrate the value of building a dedicated version of BERT for French, we first compare CamemBERT to the multilingual cased version of BERT (designated as mBERT). We then compare our models to UDify BIBREF36. UDify is a multitask and multilingual model based on mBERT that is near state-of-the-art on all UD languages including French for both POS tagging and dependency parsing. It is relevant to compare CamemBERT to UDify on those tasks because UDify is the work that pushed the furthest the performance in fine-tuning end-to-end a BERT-based model on downstream POS tagging and dependency parsing. Finally, we compare our model to UDPipe Future BIBREF37, a model ranked 3rd in dependency parsing and 6th in POS tagging during the CoNLL 2018 shared task BIBREF38. UDPipe Future provides us a strong baseline that does not make use of any pretrained contextual embedding. We will compare to the more recent cross-lingual language model XLM BIBREF12, as well as the state-of-the-art CoNLL 2018 shared task results with predicted tokenisation and segmentation in an updated version of the paper. Evaluation ::: Named Entity Recognition Named Entity Recognition (NER) is a sequence labeling task that consists in predicting which words refer to real-world objects, such as people, locations, artifacts and organisations. We use the French Treebank (FTB) BIBREF39 in its 2008 version introduced by cc-clustering:09short and with NER annotations by sagot2012annotation. The NER-annotated FTB contains more than 12k sentences and more than 350k tokens extracted from articles of the newspaper Le Monde published between 1989 and 1995. In total, it contains 11,636 entity mentions distributed among 7 different types of entities, namely: 2025 mentions of “Person”, 3761 of “Location”, 2382 of “Organisation”, 3357 of “Company”, 67 of “Product”, 15 of “POI” (Point of Interest) and 29 of “Fictional Character”. A large proportion of the entity mentions in the treebank are multi-word entities. For NER we therefore report the 3 metrics that are commonly used to evaluate models: precision, recall, and F1 score. Here precision measures the percentage of entities found by the system that are correctly tagged, recall measures the percentage of named entities present in the corpus that are found and the F1 score combines both precision and recall measures giving a general idea of a model's performance. Evaluation ::: Named Entity Recognition ::: Baselines Most of the advances in NER haven been achieved on English, particularly focusing on the CoNLL 2003 BIBREF40 and the Ontonotes v5 BIBREF41, BIBREF42 English corpora. NER is a task that was traditionally tackled using Conditional Random Fields (CRF) BIBREF43 which are quite suited for NER; CRFs were later used as decoding layers for Bi-LSTM architectures BIBREF44, BIBREF45 showing considerable improvements over CRFs alone. These Bi-LSTM-CRF architectures were later enhanced with contextualised word embeddings which yet again brought major improvements to the task BIBREF5, BIBREF6. Finally, large pretrained architectures settled the current state of the art showing a small yet important improvement over previous NER-specific architectures BIBREF7, BIBREF46. In non-English NER the CoNLL 2002 shared task included NER corpora for Spanish and Dutch corpora BIBREF47 while the CoNLL 2003 included a German corpus BIBREF40. Here the recent efforts of BIBREF48 settled the state of the art for Spanish and Dutch, while BIBREF6 did it for German. In French, no extensive work has been done due to the limited availability of NER corpora. We compare our model with the strong baselines settled by BIBREF49, who trained both CRF and BiLSTM-CRF architectures on the FTB and enhanced them using heuristics and pretrained word embeddings. Evaluation ::: Natural Language Inference We also evaluate our model on the Natural Language Inference (NLI) task, using the French part of the XNLI dataset BIBREF50. NLI consists in predicting whether a hypothesis sentence is entailed, neutral or contradicts a premise sentence. The XNLI dataset is the extension of the Multi-Genre NLI (MultiNLI) corpus BIBREF51 to 15 languages by translating the validation and test sets manually into each of those languages. The English training set is also machine translated for all languages. The dataset is composed of 122k train, 2490 valid and 5010 test examples. As usual, NLI performance is evaluated using accuracy. To evaluate a model on a language other than English (such as French), we consider the two following settings: TRANSLATE-TEST: The French test set is machine translated into English, and then used with an English classification model. This setting provides a reasonable, although imperfect, way to circumvent the fact that no such data set exists for French, and results in very strong baseline scores. TRANSLATE-TRAIN: The French model is fine-tuned on the machine-translated English training set and then evaluated on the French test set. This is the setting that we used for CamemBERT. Evaluation ::: Natural Language Inference ::: Baselines For the TRANSLATE-TEST setting, we report results of the English RoBERTa to act as a reference. In the TRANSLATE-TRAIN setting, we report the best scores from previous literature along with ours. BiLSTM-max is the best model in the original XNLI paper, mBERT which has been reported in French in BIBREF52 and XLM (MLM+TLM) is the best-presented model from BIBREF50. Experiments In this section, we measure the performance of CamemBERT by evaluating it on the four aforementioned tasks: POS tagging, dependency parsing, NER and NLI. Experiments ::: Experimental Setup ::: Pretraining We use the RoBERTa implementation in the fairseq library BIBREF53. Our learning rate is warmed up for 10k steps up to a peak value of $0.0007$ instead of the original $0.0001$ given our large batch size (8192). The learning rate fades to zero with polynomial decay. We pretrain our model on 256 Nvidia V100 GPUs (32GB each) for 100k steps during 17h. Experiments ::: Experimental Setup ::: Fine-tuning For each task, we append the relevant predictive layer on top of CamemBERT's Transformer architecture. Following the work done on BERT BIBREF7, for sequence tagging and sequence labeling we append a linear layer respectively to the $<$s$>$ special token and to the first subword token of each word. For dependency parsing, we plug a bi-affine graph predictor head as inspired by BIBREF54 following the work done on multilingual parsing with BERT by BIBREF36. We refer the reader to these two articles for more details on this module. We fine-tune independently CamemBERT for each task and each dataset. We optimise the model using the Adam optimiser BIBREF23 with a fixed learning rate. We run a grid search on a combination of learning rates and batch sizes. We select the best model on the validation set out of the 30 first epochs. Although this might push the performances even further, for all tasks except NLI, we don't apply any regularisation techniques such as weight decay, learning rate warm-up or discriminative fine-tuning. We show that fine-tuning CamemBERT in a straight-forward manner leads to state-of-the-art results on most tasks and outperforms the existing BERT-based models in most cases. The POS tagging, dependency parsing, and NER experiments are run using hugging face's Transformer library extended to support CamemBERT and dependency parsing BIBREF55. The NLI experiments use the fairseq library following the RoBERTa implementation. Experiments ::: Results ::: Part-of-Speech tagging and dependency parsing For POS tagging and dependency parsing, we compare CamemBERT to three other near state-of-the-art models in Table TABREF32. CamemBERT outperforms UDPipe Future by a large margin for all treebanks and all metrics. Despite a much simpler optimisation process, CamemBERT beats UDify performances on all the available French treebanks. CamemBERT also demonstrates higher performances than mBERT on those tasks. We observe a larger error reduction for parsing than for tagging. For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT. For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT. Experiments ::: Results ::: Natural Language Inference: XNLI On the XNLI benchmark, CamemBERT obtains improved performance over multilingual language models on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM) while using less than half the parameters (110M vs. 250M). However, its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa. It should be noted that CamemBERT uses far fewer parameters than RoBERTa (110M vs. 355M parameters). Experiments ::: Results ::: Named-Entity Recognition For named entity recognition, our experiments show that CamemBERT achieves a slightly better precision than the traditional CRF-based SEM architectures described above in Section SECREF25 (CRF and Bi-LSTM+CRF), but shows a dramatic improvement in finding entity mentions, raising the recall score by 3.5 points. Both improvements result in a 2.36 point increase in the F1 score with respect to the best SEM architecture (BiLSTM-CRF), giving CamemBERT the state of the art for NER on the FTB. One other important finding is the results obtained by mBERT. Previous work with this model showed increased performance in NER for German, Dutch and Spanish when mBERT is used as contextualised word embedding for an NER-specific model BIBREF48, but our results suggest that the multilingual setting in which mBERT was trained is simply not enough to use it alone and fine-tune it for French NER, as it shows worse performance than even simple CRF models, suggesting that monolingual models could be better at NER. Experiments ::: Discussion CamemBERT displays improved performance compared to prior work for the 4 downstream tasks considered. This confirms the hypothesis that pretrained language models can be effectively fine-tuned for various downstream tasks, as observed for English in previous work. Moreover, our results also show that dedicated monolingual models still outperform multilingual ones. We explain this point in two ways. First, the scale of data is possibly essential to the performance of CamemBERT. Indeed, we use 138GB of uncompressed text vs. 57GB for mBERT. Second, with more data comes more diversity in the pretraining distribution. Reaching state-of-the-art performances on 4 different tasks and 6 different datasets requires robust pretrained models. Our results suggest that the variability in the downstream tasks and datasets considered is handled more efficiently by a general language model than by Wikipedia-pretrained models such as mBERT. Conclusion CamemBERT improves the state of the art for multiple downstream tasks in French. It is also lighter than other BERT-based approaches such as mBERT or XLM. By releasing our model, we hope that it can serve as a strong baseline for future research in French NLP, and expect our experiments to be reproduced in many other languages. We will publish an updated version in the near future where we will explore and release models trained for longer, with additional downstream tasks, baselines (e.g. XLM) and analysis, we will also train additional models with potentially cleaner corpora such as CCNet BIBREF56 for more accurate performance evaluation and more complete ablation. Acknowledgments This work was partly funded by three French National grants from the Agence Nationale de la Recherche, namely projects PARSITI (ANR-16-CE33-0021), SoSweet (ANR-15-CE38-0011) and BASNUM (ANR-18-CE38-0003), as well as by the last author's chair in the PRAIRIE institute. Appendix ::: Impact of Whole-Word Masking We analyze the addition of whole-word masking on the downstream performance of CamemBERT. As reported for English on other downstream tasks, whole word masking improves downstream performances for all tasks but NER as seen in Table TABREF46. NER is highly sensitive to capitalisation, prefixes, suffixes and other subword features that could be used by a model to correctly identify entity mentions. Thus the added information by learning the masking at a subword level rather than at whole-word level seems to have a detrimental effect on downstream NER results.

2.36 point increase in the F1 score with respect to the best SEM architecture, on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM), lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa, For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT, For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT

dc49746fc98647445599da9d17bc004bafdc4579

dc49746fc98647445599da9d17bc004bafdc4579_0

Q: Was CamemBERT compared against multilingual BERT on these tasks? Text: Introduction Pretrained word representations have a long history in Natural Language Processing (NLP), from non-neural methods BIBREF0, BIBREF1, BIBREF2 to neural word embeddings BIBREF3, BIBREF4 and to contextualised representations BIBREF5, BIBREF6. Approaches shifted more recently from using these representations as an input to task-specific architectures to replacing these architectures with large pretrained language models. These models are then fine-tuned to the task at hand with large improvements in performance over a wide range of tasks BIBREF7, BIBREF8, BIBREF9, BIBREF10. These transfer learning methods exhibit clear advantages over more traditional task-specific approaches, probably the most important being that they can be trained in an unsupervised manner. They nevertheless come with implementation challenges, namely the amount of data and computational resources needed for pretraining that can reach hundreds of gigabytes of uncompressed text and require hundreds of GPUs BIBREF11, BIBREF9. The latest transformer architecture has gone uses as much as 750GB of plain text and 1024 TPU v3 for pretraining BIBREF10. This has limited the availability of these state-of-the-art models to the English language, at least in the monolingual setting. Even though multilingual models give remarkable results, they are often larger and their results still lag behind their monolingual counterparts BIBREF12. This is particularly inconvenient as it hinders their practical use in NLP systems as well as the investigation of their language modeling capacity, something that remains to be investigated in the case of, for instance, morphologically rich languages. We take advantage of the newly available multilingual corpus OSCAR BIBREF13 and train a monolingual language model for French using the RoBERTa architecture. We pretrain the model - which we dub CamemBERT- and evaluate it in four different downstream tasks for French: part-of-speech (POS) tagging, dependency parsing, named entity recognition (NER) and natural language inference (NLI). CamemBERT improves the state of the art for most tasks over previous monolingual and multilingual approaches, which confirms the effectiveness of large pretrained language models for French. We summarise our contributions as follows: We train a monolingual BERT model on the French language using recent large-scale corpora. We evaluate our model on four downstream tasks (POS tagging, dependency parsing, NER and natural language inference (NLI)), achieving state-of-the-art results in most tasks, confirming the effectiveness of large BERT-based models for French. We release our model in a user-friendly format for popular open-source libraries so that it can serve as a strong baseline for future research and be useful for French NLP practitioners. Related Work ::: From non-contextual to contextual word embeddings The first neural word vector representations were non-contextualised word embeddings, most notably word2vec BIBREF3, GloVe BIBREF4 and fastText BIBREF14, which were designed to be used as input to task-specific neural architectures. Contextualised word representations such as ELMo BIBREF5 and flair BIBREF6, improved the expressivity of word embeddings by taking context into account. They improved the performance of downstream tasks when they replaced traditional word representations. This paved the way towards larger contextualised models that replaced downstream architectures in most tasks. These approaches, trained with language modeling objectives, range from LSTM-based architectures such as ULMFiT BIBREF15 to the successful transformer-based architectures such as GPT2 BIBREF8, BERT BIBREF7, RoBERTa BIBREF9 and more recently ALBERT BIBREF16 and T5 BIBREF10. Related Work ::: Non-contextual word embeddings for languages other than English Since the introduction of word2vec BIBREF3, many attempts have been made to create monolingual models for a wide range of languages. For non-contextual word embeddings, the first two attempts were by BIBREF17 and BIBREF18 who created word embeddings for a large number of languages using Wikipedia. Later BIBREF19 trained fastText word embeddings for 157 languages using Common Crawl and showed that using crawled data significantly increased the performance of the embeddings relatively to those trained only on Wikipedia. Related Work ::: Contextualised models for languages other than English Following the success of large pretrained language models, they were extended to the multilingual setting with multilingual BERT , a single multilingual model for 104 different languages trained on Wikipedia data, and later XLM BIBREF12, which greatly improved unsupervised machine translation. A few monolingual models have been released: ELMo models for Japanese, Portuguese, German and Basque and BERT for Simplified and Traditional Chinese and German. However, to the best of our knowledge, no particular effort has been made toward training models for languages other than English, at a scale similar to the latest English models (e.g. RoBERTa trained on more than 100GB of data). CamemBERT Our approach is based on RoBERTa BIBREF9, which replicates and improves the initial BERT by identifying key hyper-parameters for more robust performance. In this section, we describe the architecture, training objective, optimisation setup and pretraining data that was used for CamemBERT. CamemBERT differs from RoBERTa mainly with the addition of whole-word masking and the usage of SentencePiece tokenisation BIBREF20. CamemBERT ::: Architecture Similar to RoBERTa and BERT, CamemBERT is a multi-layer bidirectional Transformer BIBREF21. Given the widespread usage of Transformers, we do not describe them in detail here and refer the reader to BIBREF21. CamemBERT uses the original BERT $_{\small \textsc {BASE}}$ configuration: 12 layers, 768 hidden dimensions, 12 attention heads, which amounts to 110M parameters. CamemBERT ::: Pretraining objective We train our model on the Masked Language Modeling (MLM) task. Given an input text sequence composed of $N$ tokens $x_1, ..., x_N$, we select $15\%$ of tokens for possible replacement. Among those selected tokens, 80% are replaced with the special $<$mask$>$ token, 10% are left unchanged and 10% are replaced by a random token. The model is then trained to predict the initial masked tokens using cross-entropy loss. Following RoBERTa we dynamically mask tokens instead of fixing them statically for the whole dataset during preprocessing. This improves variability and makes the model more robust when training for multiple epochs. Since we segment the input sentence into subwords using SentencePiece, the input tokens to the models can be subwords. An upgraded version of BERT and BIBREF22 have shown that masking whole words instead of individual subwords leads to improved performance. Whole-word masking (WWM) makes the training task more difficult because the model has to predict a whole word instead of predicting only part of the word given the rest. As a result, we used WWM for CamemBERT by first randomly sampling 15% of the words in the sequence and then considering all subword tokens in each of these 15% words for candidate replacement. This amounts to a proportion of selected tokens that is close to the original 15%. These tokens are then either replaced by $<$mask$>$ tokens (80%), left unchanged (10%) or replaced by a random token. Subsequent work has shown that the next sentence prediction task (NSP) originally used in BERT does not improve downstream task performance BIBREF12, BIBREF9, we do not use NSP as a consequence. CamemBERT ::: Optimisation Following BIBREF9, we optimise the model using Adam BIBREF23 ($\beta _1 = 0.9$, $\beta _2 = 0.98$) for 100k steps. We use large batch sizes of 8192 sequences. Each sequence contains at most 512 tokens. We enforce each sequence to only contain complete sentences. Additionally, we used the DOC-SENTENCES scenario from BIBREF9, consisting of not mixing multiple documents in the same sequence, which showed slightly better results. CamemBERT ::: Segmentation into subword units We segment the input text into subword units using SentencePiece BIBREF20. SentencePiece is an extension of Byte-Pair encoding (BPE) BIBREF24 and WordPiece BIBREF25 that does not require pre-tokenisation (at the word or token level), thus removing the need for language-specific tokenisers. We use a vocabulary size of 32k subword tokens. These are learned on $10^7$ sentences sampled from the pretraining dataset. We do not use subword regularisation (i.e. sampling from multiple possible segmentations) in our implementation for simplicity. CamemBERT ::: Pretraining data Pretrained language models can be significantly improved by using more data BIBREF9, BIBREF10. Therefore we used French text extracted from Common Crawl, in particular, we use OSCAR BIBREF13 a pre-classified and pre-filtered version of the November 2018 Common Craw snapshot. OSCAR is a set of monolingual corpora extracted from Common Crawl, specifically from the plain text WET format distributed by Common Crawl, which removes all HTML tags and converts all text encodings to UTF-8. OSCAR follows the same approach as BIBREF19 by using a language classification model based on the fastText linear classifier BIBREF26, BIBREF27 pretrained on Wikipedia, Tatoeba and SETimes, which supports 176 different languages. OSCAR performs a deduplication step after language classification and without introducing a specialised filtering scheme, other than only keeping paragraphs containing 100 or more UTF-8 encoded characters, making OSCAR quite close to the original Crawled data. We use the unshuffled version of the French OSCAR corpus, which amounts to 138GB of uncompressed text and 32.7B SentencePiece tokens. Evaluation ::: Part-of-speech tagging and dependency parsing We fist evaluate CamemBERT on the two downstream tasks of part-of-speech (POS) tagging and dependency parsing. POS tagging is a low-level syntactic task, which consists in assigning to each word its corresponding grammatical category. Dependency parsing consists in predicting the labeled syntactic tree capturing the syntactic relations between words. We run our experiments using the Universal Dependencies (UD) paradigm and its corresponding UD POS tag set BIBREF28 and UD treebank collection version 2.2 BIBREF29, which was used for the CoNLL 2018 shared task. We perform our work on the four freely available French UD treebanks in UD v2.2: GSD, Sequoia, Spoken, and ParTUT. GSD BIBREF30 is the second-largest treebank available for French after the FTB (described in subsection SECREF25), it contains data from blogs, news articles, reviews, and Wikipedia. The Sequoia treebank BIBREF31, BIBREF32 comprises more than 3000 sentences, from the French Europarl, the regional newspaper L’Est Républicain, the French Wikipedia and documents from the European Medicines Agency. Spoken is a corpus converted automatically from the Rhapsodie treebank BIBREF33, BIBREF34 with manual corrections. It consists of 57 sound samples of spoken French with orthographic transcription and phonetic transcription aligned with sound (word boundaries, syllables, and phonemes), syntactic and prosodic annotations. Finally, ParTUT is a conversion of a multilingual parallel treebank developed at the University of Turin, and consisting of a variety of text genres, including talks, legal texts, and Wikipedia articles, among others; ParTUT data is derived from the already-existing parallel treebank Par(allel)TUT BIBREF35 . Table TABREF23 contains a summary comparing the sizes of the treebanks. We evaluate the performance of our models using the standard UPOS accuracy for POS tagging, and Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS) for dependency parsing. We assume gold tokenisation and gold word segmentation as provided in the UD treebanks. Evaluation ::: Part-of-speech tagging and dependency parsing ::: Baselines To demonstrate the value of building a dedicated version of BERT for French, we first compare CamemBERT to the multilingual cased version of BERT (designated as mBERT). We then compare our models to UDify BIBREF36. UDify is a multitask and multilingual model based on mBERT that is near state-of-the-art on all UD languages including French for both POS tagging and dependency parsing. It is relevant to compare CamemBERT to UDify on those tasks because UDify is the work that pushed the furthest the performance in fine-tuning end-to-end a BERT-based model on downstream POS tagging and dependency parsing. Finally, we compare our model to UDPipe Future BIBREF37, a model ranked 3rd in dependency parsing and 6th in POS tagging during the CoNLL 2018 shared task BIBREF38. UDPipe Future provides us a strong baseline that does not make use of any pretrained contextual embedding. We will compare to the more recent cross-lingual language model XLM BIBREF12, as well as the state-of-the-art CoNLL 2018 shared task results with predicted tokenisation and segmentation in an updated version of the paper. Evaluation ::: Named Entity Recognition Named Entity Recognition (NER) is a sequence labeling task that consists in predicting which words refer to real-world objects, such as people, locations, artifacts and organisations. We use the French Treebank (FTB) BIBREF39 in its 2008 version introduced by cc-clustering:09short and with NER annotations by sagot2012annotation. The NER-annotated FTB contains more than 12k sentences and more than 350k tokens extracted from articles of the newspaper Le Monde published between 1989 and 1995. In total, it contains 11,636 entity mentions distributed among 7 different types of entities, namely: 2025 mentions of “Person”, 3761 of “Location”, 2382 of “Organisation”, 3357 of “Company”, 67 of “Product”, 15 of “POI” (Point of Interest) and 29 of “Fictional Character”. A large proportion of the entity mentions in the treebank are multi-word entities. For NER we therefore report the 3 metrics that are commonly used to evaluate models: precision, recall, and F1 score. Here precision measures the percentage of entities found by the system that are correctly tagged, recall measures the percentage of named entities present in the corpus that are found and the F1 score combines both precision and recall measures giving a general idea of a model's performance. Evaluation ::: Named Entity Recognition ::: Baselines Most of the advances in NER haven been achieved on English, particularly focusing on the CoNLL 2003 BIBREF40 and the Ontonotes v5 BIBREF41, BIBREF42 English corpora. NER is a task that was traditionally tackled using Conditional Random Fields (CRF) BIBREF43 which are quite suited for NER; CRFs were later used as decoding layers for Bi-LSTM architectures BIBREF44, BIBREF45 showing considerable improvements over CRFs alone. These Bi-LSTM-CRF architectures were later enhanced with contextualised word embeddings which yet again brought major improvements to the task BIBREF5, BIBREF6. Finally, large pretrained architectures settled the current state of the art showing a small yet important improvement over previous NER-specific architectures BIBREF7, BIBREF46. In non-English NER the CoNLL 2002 shared task included NER corpora for Spanish and Dutch corpora BIBREF47 while the CoNLL 2003 included a German corpus BIBREF40. Here the recent efforts of BIBREF48 settled the state of the art for Spanish and Dutch, while BIBREF6 did it for German. In French, no extensive work has been done due to the limited availability of NER corpora. We compare our model with the strong baselines settled by BIBREF49, who trained both CRF and BiLSTM-CRF architectures on the FTB and enhanced them using heuristics and pretrained word embeddings. Evaluation ::: Natural Language Inference We also evaluate our model on the Natural Language Inference (NLI) task, using the French part of the XNLI dataset BIBREF50. NLI consists in predicting whether a hypothesis sentence is entailed, neutral or contradicts a premise sentence. The XNLI dataset is the extension of the Multi-Genre NLI (MultiNLI) corpus BIBREF51 to 15 languages by translating the validation and test sets manually into each of those languages. The English training set is also machine translated for all languages. The dataset is composed of 122k train, 2490 valid and 5010 test examples. As usual, NLI performance is evaluated using accuracy. To evaluate a model on a language other than English (such as French), we consider the two following settings: TRANSLATE-TEST: The French test set is machine translated into English, and then used with an English classification model. This setting provides a reasonable, although imperfect, way to circumvent the fact that no such data set exists for French, and results in very strong baseline scores. TRANSLATE-TRAIN: The French model is fine-tuned on the machine-translated English training set and then evaluated on the French test set. This is the setting that we used for CamemBERT. Evaluation ::: Natural Language Inference ::: Baselines For the TRANSLATE-TEST setting, we report results of the English RoBERTa to act as a reference. In the TRANSLATE-TRAIN setting, we report the best scores from previous literature along with ours. BiLSTM-max is the best model in the original XNLI paper, mBERT which has been reported in French in BIBREF52 and XLM (MLM+TLM) is the best-presented model from BIBREF50. Experiments In this section, we measure the performance of CamemBERT by evaluating it on the four aforementioned tasks: POS tagging, dependency parsing, NER and NLI. Experiments ::: Experimental Setup ::: Pretraining We use the RoBERTa implementation in the fairseq library BIBREF53. Our learning rate is warmed up for 10k steps up to a peak value of $0.0007$ instead of the original $0.0001$ given our large batch size (8192). The learning rate fades to zero with polynomial decay. We pretrain our model on 256 Nvidia V100 GPUs (32GB each) for 100k steps during 17h. Experiments ::: Experimental Setup ::: Fine-tuning For each task, we append the relevant predictive layer on top of CamemBERT's Transformer architecture. Following the work done on BERT BIBREF7, for sequence tagging and sequence labeling we append a linear layer respectively to the $<$s$>$ special token and to the first subword token of each word. For dependency parsing, we plug a bi-affine graph predictor head as inspired by BIBREF54 following the work done on multilingual parsing with BERT by BIBREF36. We refer the reader to these two articles for more details on this module. We fine-tune independently CamemBERT for each task and each dataset. We optimise the model using the Adam optimiser BIBREF23 with a fixed learning rate. We run a grid search on a combination of learning rates and batch sizes. We select the best model on the validation set out of the 30 first epochs. Although this might push the performances even further, for all tasks except NLI, we don't apply any regularisation techniques such as weight decay, learning rate warm-up or discriminative fine-tuning. We show that fine-tuning CamemBERT in a straight-forward manner leads to state-of-the-art results on most tasks and outperforms the existing BERT-based models in most cases. The POS tagging, dependency parsing, and NER experiments are run using hugging face's Transformer library extended to support CamemBERT and dependency parsing BIBREF55. The NLI experiments use the fairseq library following the RoBERTa implementation. Experiments ::: Results ::: Part-of-Speech tagging and dependency parsing For POS tagging and dependency parsing, we compare CamemBERT to three other near state-of-the-art models in Table TABREF32. CamemBERT outperforms UDPipe Future by a large margin for all treebanks and all metrics. Despite a much simpler optimisation process, CamemBERT beats UDify performances on all the available French treebanks. CamemBERT also demonstrates higher performances than mBERT on those tasks. We observe a larger error reduction for parsing than for tagging. For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT. For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT. Experiments ::: Results ::: Natural Language Inference: XNLI On the XNLI benchmark, CamemBERT obtains improved performance over multilingual language models on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM) while using less than half the parameters (110M vs. 250M). However, its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa. It should be noted that CamemBERT uses far fewer parameters than RoBERTa (110M vs. 355M parameters). Experiments ::: Results ::: Named-Entity Recognition For named entity recognition, our experiments show that CamemBERT achieves a slightly better precision than the traditional CRF-based SEM architectures described above in Section SECREF25 (CRF and Bi-LSTM+CRF), but shows a dramatic improvement in finding entity mentions, raising the recall score by 3.5 points. Both improvements result in a 2.36 point increase in the F1 score with respect to the best SEM architecture (BiLSTM-CRF), giving CamemBERT the state of the art for NER on the FTB. One other important finding is the results obtained by mBERT. Previous work with this model showed increased performance in NER for German, Dutch and Spanish when mBERT is used as contextualised word embedding for an NER-specific model BIBREF48, but our results suggest that the multilingual setting in which mBERT was trained is simply not enough to use it alone and fine-tune it for French NER, as it shows worse performance than even simple CRF models, suggesting that monolingual models could be better at NER. Experiments ::: Discussion CamemBERT displays improved performance compared to prior work for the 4 downstream tasks considered. This confirms the hypothesis that pretrained language models can be effectively fine-tuned for various downstream tasks, as observed for English in previous work. Moreover, our results also show that dedicated monolingual models still outperform multilingual ones. We explain this point in two ways. First, the scale of data is possibly essential to the performance of CamemBERT. Indeed, we use 138GB of uncompressed text vs. 57GB for mBERT. Second, with more data comes more diversity in the pretraining distribution. Reaching state-of-the-art performances on 4 different tasks and 6 different datasets requires robust pretrained models. Our results suggest that the variability in the downstream tasks and datasets considered is handled more efficiently by a general language model than by Wikipedia-pretrained models such as mBERT. Conclusion CamemBERT improves the state of the art for multiple downstream tasks in French. It is also lighter than other BERT-based approaches such as mBERT or XLM. By releasing our model, we hope that it can serve as a strong baseline for future research in French NLP, and expect our experiments to be reproduced in many other languages. We will publish an updated version in the near future where we will explore and release models trained for longer, with additional downstream tasks, baselines (e.g. XLM) and analysis, we will also train additional models with potentially cleaner corpora such as CCNet BIBREF56 for more accurate performance evaluation and more complete ablation. Acknowledgments This work was partly funded by three French National grants from the Agence Nationale de la Recherche, namely projects PARSITI (ANR-16-CE33-0021), SoSweet (ANR-15-CE38-0011) and BASNUM (ANR-18-CE38-0003), as well as by the last author's chair in the PRAIRIE institute. Appendix ::: Impact of Whole-Word Masking We analyze the addition of whole-word masking on the downstream performance of CamemBERT. As reported for English on other downstream tasks, whole word masking improves downstream performances for all tasks but NER as seen in Table TABREF46. NER is highly sensitive to capitalisation, prefixes, suffixes and other subword features that could be used by a model to correctly identify entity mentions. Thus the added information by learning the masking at a subword level rather than at whole-word level seems to have a detrimental effect on downstream NER results.

Yes

8720c096c8b990c7b19f956ee4930d5f2c019e2b

8720c096c8b990c7b19f956ee4930d5f2c019e2b_0

Q: How long was CamemBERT trained? Text: Introduction Pretrained word representations have a long history in Natural Language Processing (NLP), from non-neural methods BIBREF0, BIBREF1, BIBREF2 to neural word embeddings BIBREF3, BIBREF4 and to contextualised representations BIBREF5, BIBREF6. Approaches shifted more recently from using these representations as an input to task-specific architectures to replacing these architectures with large pretrained language models. These models are then fine-tuned to the task at hand with large improvements in performance over a wide range of tasks BIBREF7, BIBREF8, BIBREF9, BIBREF10. These transfer learning methods exhibit clear advantages over more traditional task-specific approaches, probably the most important being that they can be trained in an unsupervised manner. They nevertheless come with implementation challenges, namely the amount of data and computational resources needed for pretraining that can reach hundreds of gigabytes of uncompressed text and require hundreds of GPUs BIBREF11, BIBREF9. The latest transformer architecture has gone uses as much as 750GB of plain text and 1024 TPU v3 for pretraining BIBREF10. This has limited the availability of these state-of-the-art models to the English language, at least in the monolingual setting. Even though multilingual models give remarkable results, they are often larger and their results still lag behind their monolingual counterparts BIBREF12. This is particularly inconvenient as it hinders their practical use in NLP systems as well as the investigation of their language modeling capacity, something that remains to be investigated in the case of, for instance, morphologically rich languages. We take advantage of the newly available multilingual corpus OSCAR BIBREF13 and train a monolingual language model for French using the RoBERTa architecture. We pretrain the model - which we dub CamemBERT- and evaluate it in four different downstream tasks for French: part-of-speech (POS) tagging, dependency parsing, named entity recognition (NER) and natural language inference (NLI). CamemBERT improves the state of the art for most tasks over previous monolingual and multilingual approaches, which confirms the effectiveness of large pretrained language models for French. We summarise our contributions as follows: We train a monolingual BERT model on the French language using recent large-scale corpora. We evaluate our model on four downstream tasks (POS tagging, dependency parsing, NER and natural language inference (NLI)), achieving state-of-the-art results in most tasks, confirming the effectiveness of large BERT-based models for French. We release our model in a user-friendly format for popular open-source libraries so that it can serve as a strong baseline for future research and be useful for French NLP practitioners. Related Work ::: From non-contextual to contextual word embeddings The first neural word vector representations were non-contextualised word embeddings, most notably word2vec BIBREF3, GloVe BIBREF4 and fastText BIBREF14, which were designed to be used as input to task-specific neural architectures. Contextualised word representations such as ELMo BIBREF5 and flair BIBREF6, improved the expressivity of word embeddings by taking context into account. They improved the performance of downstream tasks when they replaced traditional word representations. This paved the way towards larger contextualised models that replaced downstream architectures in most tasks. These approaches, trained with language modeling objectives, range from LSTM-based architectures such as ULMFiT BIBREF15 to the successful transformer-based architectures such as GPT2 BIBREF8, BERT BIBREF7, RoBERTa BIBREF9 and more recently ALBERT BIBREF16 and T5 BIBREF10. Related Work ::: Non-contextual word embeddings for languages other than English Since the introduction of word2vec BIBREF3, many attempts have been made to create monolingual models for a wide range of languages. For non-contextual word embeddings, the first two attempts were by BIBREF17 and BIBREF18 who created word embeddings for a large number of languages using Wikipedia. Later BIBREF19 trained fastText word embeddings for 157 languages using Common Crawl and showed that using crawled data significantly increased the performance of the embeddings relatively to those trained only on Wikipedia. Related Work ::: Contextualised models for languages other than English Following the success of large pretrained language models, they were extended to the multilingual setting with multilingual BERT , a single multilingual model for 104 different languages trained on Wikipedia data, and later XLM BIBREF12, which greatly improved unsupervised machine translation. A few monolingual models have been released: ELMo models for Japanese, Portuguese, German and Basque and BERT for Simplified and Traditional Chinese and German. However, to the best of our knowledge, no particular effort has been made toward training models for languages other than English, at a scale similar to the latest English models (e.g. RoBERTa trained on more than 100GB of data). CamemBERT Our approach is based on RoBERTa BIBREF9, which replicates and improves the initial BERT by identifying key hyper-parameters for more robust performance. In this section, we describe the architecture, training objective, optimisation setup and pretraining data that was used for CamemBERT. CamemBERT differs from RoBERTa mainly with the addition of whole-word masking and the usage of SentencePiece tokenisation BIBREF20. CamemBERT ::: Architecture Similar to RoBERTa and BERT, CamemBERT is a multi-layer bidirectional Transformer BIBREF21. Given the widespread usage of Transformers, we do not describe them in detail here and refer the reader to BIBREF21. CamemBERT uses the original BERT $_{\small \textsc {BASE}}$ configuration: 12 layers, 768 hidden dimensions, 12 attention heads, which amounts to 110M parameters. CamemBERT ::: Pretraining objective We train our model on the Masked Language Modeling (MLM) task. Given an input text sequence composed of $N$ tokens $x_1, ..., x_N$, we select $15\%$ of tokens for possible replacement. Among those selected tokens, 80% are replaced with the special $<$mask$>$ token, 10% are left unchanged and 10% are replaced by a random token. The model is then trained to predict the initial masked tokens using cross-entropy loss. Following RoBERTa we dynamically mask tokens instead of fixing them statically for the whole dataset during preprocessing. This improves variability and makes the model more robust when training for multiple epochs. Since we segment the input sentence into subwords using SentencePiece, the input tokens to the models can be subwords. An upgraded version of BERT and BIBREF22 have shown that masking whole words instead of individual subwords leads to improved performance. Whole-word masking (WWM) makes the training task more difficult because the model has to predict a whole word instead of predicting only part of the word given the rest. As a result, we used WWM for CamemBERT by first randomly sampling 15% of the words in the sequence and then considering all subword tokens in each of these 15% words for candidate replacement. This amounts to a proportion of selected tokens that is close to the original 15%. These tokens are then either replaced by $<$mask$>$ tokens (80%), left unchanged (10%) or replaced by a random token. Subsequent work has shown that the next sentence prediction task (NSP) originally used in BERT does not improve downstream task performance BIBREF12, BIBREF9, we do not use NSP as a consequence. CamemBERT ::: Optimisation Following BIBREF9, we optimise the model using Adam BIBREF23 ($\beta _1 = 0.9$, $\beta _2 = 0.98$) for 100k steps. We use large batch sizes of 8192 sequences. Each sequence contains at most 512 tokens. We enforce each sequence to only contain complete sentences. Additionally, we used the DOC-SENTENCES scenario from BIBREF9, consisting of not mixing multiple documents in the same sequence, which showed slightly better results. CamemBERT ::: Segmentation into subword units We segment the input text into subword units using SentencePiece BIBREF20. SentencePiece is an extension of Byte-Pair encoding (BPE) BIBREF24 and WordPiece BIBREF25 that does not require pre-tokenisation (at the word or token level), thus removing the need for language-specific tokenisers. We use a vocabulary size of 32k subword tokens. These are learned on $10^7$ sentences sampled from the pretraining dataset. We do not use subword regularisation (i.e. sampling from multiple possible segmentations) in our implementation for simplicity. CamemBERT ::: Pretraining data Pretrained language models can be significantly improved by using more data BIBREF9, BIBREF10. Therefore we used French text extracted from Common Crawl, in particular, we use OSCAR BIBREF13 a pre-classified and pre-filtered version of the November 2018 Common Craw snapshot. OSCAR is a set of monolingual corpora extracted from Common Crawl, specifically from the plain text WET format distributed by Common Crawl, which removes all HTML tags and converts all text encodings to UTF-8. OSCAR follows the same approach as BIBREF19 by using a language classification model based on the fastText linear classifier BIBREF26, BIBREF27 pretrained on Wikipedia, Tatoeba and SETimes, which supports 176 different languages. OSCAR performs a deduplication step after language classification and without introducing a specialised filtering scheme, other than only keeping paragraphs containing 100 or more UTF-8 encoded characters, making OSCAR quite close to the original Crawled data. We use the unshuffled version of the French OSCAR corpus, which amounts to 138GB of uncompressed text and 32.7B SentencePiece tokens. Evaluation ::: Part-of-speech tagging and dependency parsing We fist evaluate CamemBERT on the two downstream tasks of part-of-speech (POS) tagging and dependency parsing. POS tagging is a low-level syntactic task, which consists in assigning to each word its corresponding grammatical category. Dependency parsing consists in predicting the labeled syntactic tree capturing the syntactic relations between words. We run our experiments using the Universal Dependencies (UD) paradigm and its corresponding UD POS tag set BIBREF28 and UD treebank collection version 2.2 BIBREF29, which was used for the CoNLL 2018 shared task. We perform our work on the four freely available French UD treebanks in UD v2.2: GSD, Sequoia, Spoken, and ParTUT. GSD BIBREF30 is the second-largest treebank available for French after the FTB (described in subsection SECREF25), it contains data from blogs, news articles, reviews, and Wikipedia. The Sequoia treebank BIBREF31, BIBREF32 comprises more than 3000 sentences, from the French Europarl, the regional newspaper L’Est Républicain, the French Wikipedia and documents from the European Medicines Agency. Spoken is a corpus converted automatically from the Rhapsodie treebank BIBREF33, BIBREF34 with manual corrections. It consists of 57 sound samples of spoken French with orthographic transcription and phonetic transcription aligned with sound (word boundaries, syllables, and phonemes), syntactic and prosodic annotations. Finally, ParTUT is a conversion of a multilingual parallel treebank developed at the University of Turin, and consisting of a variety of text genres, including talks, legal texts, and Wikipedia articles, among others; ParTUT data is derived from the already-existing parallel treebank Par(allel)TUT BIBREF35 . Table TABREF23 contains a summary comparing the sizes of the treebanks. We evaluate the performance of our models using the standard UPOS accuracy for POS tagging, and Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS) for dependency parsing. We assume gold tokenisation and gold word segmentation as provided in the UD treebanks. Evaluation ::: Part-of-speech tagging and dependency parsing ::: Baselines To demonstrate the value of building a dedicated version of BERT for French, we first compare CamemBERT to the multilingual cased version of BERT (designated as mBERT). We then compare our models to UDify BIBREF36. UDify is a multitask and multilingual model based on mBERT that is near state-of-the-art on all UD languages including French for both POS tagging and dependency parsing. It is relevant to compare CamemBERT to UDify on those tasks because UDify is the work that pushed the furthest the performance in fine-tuning end-to-end a BERT-based model on downstream POS tagging and dependency parsing. Finally, we compare our model to UDPipe Future BIBREF37, a model ranked 3rd in dependency parsing and 6th in POS tagging during the CoNLL 2018 shared task BIBREF38. UDPipe Future provides us a strong baseline that does not make use of any pretrained contextual embedding. We will compare to the more recent cross-lingual language model XLM BIBREF12, as well as the state-of-the-art CoNLL 2018 shared task results with predicted tokenisation and segmentation in an updated version of the paper. Evaluation ::: Named Entity Recognition Named Entity Recognition (NER) is a sequence labeling task that consists in predicting which words refer to real-world objects, such as people, locations, artifacts and organisations. We use the French Treebank (FTB) BIBREF39 in its 2008 version introduced by cc-clustering:09short and with NER annotations by sagot2012annotation. The NER-annotated FTB contains more than 12k sentences and more than 350k tokens extracted from articles of the newspaper Le Monde published between 1989 and 1995. In total, it contains 11,636 entity mentions distributed among 7 different types of entities, namely: 2025 mentions of “Person”, 3761 of “Location”, 2382 of “Organisation”, 3357 of “Company”, 67 of “Product”, 15 of “POI” (Point of Interest) and 29 of “Fictional Character”. A large proportion of the entity mentions in the treebank are multi-word entities. For NER we therefore report the 3 metrics that are commonly used to evaluate models: precision, recall, and F1 score. Here precision measures the percentage of entities found by the system that are correctly tagged, recall measures the percentage of named entities present in the corpus that are found and the F1 score combines both precision and recall measures giving a general idea of a model's performance. Evaluation ::: Named Entity Recognition ::: Baselines Most of the advances in NER haven been achieved on English, particularly focusing on the CoNLL 2003 BIBREF40 and the Ontonotes v5 BIBREF41, BIBREF42 English corpora. NER is a task that was traditionally tackled using Conditional Random Fields (CRF) BIBREF43 which are quite suited for NER; CRFs were later used as decoding layers for Bi-LSTM architectures BIBREF44, BIBREF45 showing considerable improvements over CRFs alone. These Bi-LSTM-CRF architectures were later enhanced with contextualised word embeddings which yet again brought major improvements to the task BIBREF5, BIBREF6. Finally, large pretrained architectures settled the current state of the art showing a small yet important improvement over previous NER-specific architectures BIBREF7, BIBREF46. In non-English NER the CoNLL 2002 shared task included NER corpora for Spanish and Dutch corpora BIBREF47 while the CoNLL 2003 included a German corpus BIBREF40. Here the recent efforts of BIBREF48 settled the state of the art for Spanish and Dutch, while BIBREF6 did it for German. In French, no extensive work has been done due to the limited availability of NER corpora. We compare our model with the strong baselines settled by BIBREF49, who trained both CRF and BiLSTM-CRF architectures on the FTB and enhanced them using heuristics and pretrained word embeddings. Evaluation ::: Natural Language Inference We also evaluate our model on the Natural Language Inference (NLI) task, using the French part of the XNLI dataset BIBREF50. NLI consists in predicting whether a hypothesis sentence is entailed, neutral or contradicts a premise sentence. The XNLI dataset is the extension of the Multi-Genre NLI (MultiNLI) corpus BIBREF51 to 15 languages by translating the validation and test sets manually into each of those languages. The English training set is also machine translated for all languages. The dataset is composed of 122k train, 2490 valid and 5010 test examples. As usual, NLI performance is evaluated using accuracy. To evaluate a model on a language other than English (such as French), we consider the two following settings: TRANSLATE-TEST: The French test set is machine translated into English, and then used with an English classification model. This setting provides a reasonable, although imperfect, way to circumvent the fact that no such data set exists for French, and results in very strong baseline scores. TRANSLATE-TRAIN: The French model is fine-tuned on the machine-translated English training set and then evaluated on the French test set. This is the setting that we used for CamemBERT. Evaluation ::: Natural Language Inference ::: Baselines For the TRANSLATE-TEST setting, we report results of the English RoBERTa to act as a reference. In the TRANSLATE-TRAIN setting, we report the best scores from previous literature along with ours. BiLSTM-max is the best model in the original XNLI paper, mBERT which has been reported in French in BIBREF52 and XLM (MLM+TLM) is the best-presented model from BIBREF50. Experiments In this section, we measure the performance of CamemBERT by evaluating it on the four aforementioned tasks: POS tagging, dependency parsing, NER and NLI. Experiments ::: Experimental Setup ::: Pretraining We use the RoBERTa implementation in the fairseq library BIBREF53. Our learning rate is warmed up for 10k steps up to a peak value of $0.0007$ instead of the original $0.0001$ given our large batch size (8192). The learning rate fades to zero with polynomial decay. We pretrain our model on 256 Nvidia V100 GPUs (32GB each) for 100k steps during 17h. Experiments ::: Experimental Setup ::: Fine-tuning For each task, we append the relevant predictive layer on top of CamemBERT's Transformer architecture. Following the work done on BERT BIBREF7, for sequence tagging and sequence labeling we append a linear layer respectively to the $<$s$>$ special token and to the first subword token of each word. For dependency parsing, we plug a bi-affine graph predictor head as inspired by BIBREF54 following the work done on multilingual parsing with BERT by BIBREF36. We refer the reader to these two articles for more details on this module. We fine-tune independently CamemBERT for each task and each dataset. We optimise the model using the Adam optimiser BIBREF23 with a fixed learning rate. We run a grid search on a combination of learning rates and batch sizes. We select the best model on the validation set out of the 30 first epochs. Although this might push the performances even further, for all tasks except NLI, we don't apply any regularisation techniques such as weight decay, learning rate warm-up or discriminative fine-tuning. We show that fine-tuning CamemBERT in a straight-forward manner leads to state-of-the-art results on most tasks and outperforms the existing BERT-based models in most cases. The POS tagging, dependency parsing, and NER experiments are run using hugging face's Transformer library extended to support CamemBERT and dependency parsing BIBREF55. The NLI experiments use the fairseq library following the RoBERTa implementation. Experiments ::: Results ::: Part-of-Speech tagging and dependency parsing For POS tagging and dependency parsing, we compare CamemBERT to three other near state-of-the-art models in Table TABREF32. CamemBERT outperforms UDPipe Future by a large margin for all treebanks and all metrics. Despite a much simpler optimisation process, CamemBERT beats UDify performances on all the available French treebanks. CamemBERT also demonstrates higher performances than mBERT on those tasks. We observe a larger error reduction for parsing than for tagging. For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT. For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT. Experiments ::: Results ::: Natural Language Inference: XNLI On the XNLI benchmark, CamemBERT obtains improved performance over multilingual language models on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM) while using less than half the parameters (110M vs. 250M). However, its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa. It should be noted that CamemBERT uses far fewer parameters than RoBERTa (110M vs. 355M parameters). Experiments ::: Results ::: Named-Entity Recognition For named entity recognition, our experiments show that CamemBERT achieves a slightly better precision than the traditional CRF-based SEM architectures described above in Section SECREF25 (CRF and Bi-LSTM+CRF), but shows a dramatic improvement in finding entity mentions, raising the recall score by 3.5 points. Both improvements result in a 2.36 point increase in the F1 score with respect to the best SEM architecture (BiLSTM-CRF), giving CamemBERT the state of the art for NER on the FTB. One other important finding is the results obtained by mBERT. Previous work with this model showed increased performance in NER for German, Dutch and Spanish when mBERT is used as contextualised word embedding for an NER-specific model BIBREF48, but our results suggest that the multilingual setting in which mBERT was trained is simply not enough to use it alone and fine-tune it for French NER, as it shows worse performance than even simple CRF models, suggesting that monolingual models could be better at NER. Experiments ::: Discussion CamemBERT displays improved performance compared to prior work for the 4 downstream tasks considered. This confirms the hypothesis that pretrained language models can be effectively fine-tuned for various downstream tasks, as observed for English in previous work. Moreover, our results also show that dedicated monolingual models still outperform multilingual ones. We explain this point in two ways. First, the scale of data is possibly essential to the performance of CamemBERT. Indeed, we use 138GB of uncompressed text vs. 57GB for mBERT. Second, with more data comes more diversity in the pretraining distribution. Reaching state-of-the-art performances on 4 different tasks and 6 different datasets requires robust pretrained models. Our results suggest that the variability in the downstream tasks and datasets considered is handled more efficiently by a general language model than by Wikipedia-pretrained models such as mBERT. Conclusion CamemBERT improves the state of the art for multiple downstream tasks in French. It is also lighter than other BERT-based approaches such as mBERT or XLM. By releasing our model, we hope that it can serve as a strong baseline for future research in French NLP, and expect our experiments to be reproduced in many other languages. We will publish an updated version in the near future where we will explore and release models trained for longer, with additional downstream tasks, baselines (e.g. XLM) and analysis, we will also train additional models with potentially cleaner corpora such as CCNet BIBREF56 for more accurate performance evaluation and more complete ablation. Acknowledgments This work was partly funded by three French National grants from the Agence Nationale de la Recherche, namely projects PARSITI (ANR-16-CE33-0021), SoSweet (ANR-15-CE38-0011) and BASNUM (ANR-18-CE38-0003), as well as by the last author's chair in the PRAIRIE institute. Appendix ::: Impact of Whole-Word Masking We analyze the addition of whole-word masking on the downstream performance of CamemBERT. As reported for English on other downstream tasks, whole word masking improves downstream performances for all tasks but NER as seen in Table TABREF46. NER is highly sensitive to capitalisation, prefixes, suffixes and other subword features that could be used by a model to correctly identify entity mentions. Thus the added information by learning the masking at a subword level rather than at whole-word level seems to have a detrimental effect on downstream NER results.

Unanswerable

b573b36936ffdf1d70e66f9b5567511c989b46b2

b573b36936ffdf1d70e66f9b5567511c989b46b2_0

Q: What data is used for training CamemBERT? Text: Introduction Pretrained word representations have a long history in Natural Language Processing (NLP), from non-neural methods BIBREF0, BIBREF1, BIBREF2 to neural word embeddings BIBREF3, BIBREF4 and to contextualised representations BIBREF5, BIBREF6. Approaches shifted more recently from using these representations as an input to task-specific architectures to replacing these architectures with large pretrained language models. These models are then fine-tuned to the task at hand with large improvements in performance over a wide range of tasks BIBREF7, BIBREF8, BIBREF9, BIBREF10. These transfer learning methods exhibit clear advantages over more traditional task-specific approaches, probably the most important being that they can be trained in an unsupervised manner. They nevertheless come with implementation challenges, namely the amount of data and computational resources needed for pretraining that can reach hundreds of gigabytes of uncompressed text and require hundreds of GPUs BIBREF11, BIBREF9. The latest transformer architecture has gone uses as much as 750GB of plain text and 1024 TPU v3 for pretraining BIBREF10. This has limited the availability of these state-of-the-art models to the English language, at least in the monolingual setting. Even though multilingual models give remarkable results, they are often larger and their results still lag behind their monolingual counterparts BIBREF12. This is particularly inconvenient as it hinders their practical use in NLP systems as well as the investigation of their language modeling capacity, something that remains to be investigated in the case of, for instance, morphologically rich languages. We take advantage of the newly available multilingual corpus OSCAR BIBREF13 and train a monolingual language model for French using the RoBERTa architecture. We pretrain the model - which we dub CamemBERT- and evaluate it in four different downstream tasks for French: part-of-speech (POS) tagging, dependency parsing, named entity recognition (NER) and natural language inference (NLI). CamemBERT improves the state of the art for most tasks over previous monolingual and multilingual approaches, which confirms the effectiveness of large pretrained language models for French. We summarise our contributions as follows: We train a monolingual BERT model on the French language using recent large-scale corpora. We evaluate our model on four downstream tasks (POS tagging, dependency parsing, NER and natural language inference (NLI)), achieving state-of-the-art results in most tasks, confirming the effectiveness of large BERT-based models for French. We release our model in a user-friendly format for popular open-source libraries so that it can serve as a strong baseline for future research and be useful for French NLP practitioners. Related Work ::: From non-contextual to contextual word embeddings The first neural word vector representations were non-contextualised word embeddings, most notably word2vec BIBREF3, GloVe BIBREF4 and fastText BIBREF14, which were designed to be used as input to task-specific neural architectures. Contextualised word representations such as ELMo BIBREF5 and flair BIBREF6, improved the expressivity of word embeddings by taking context into account. They improved the performance of downstream tasks when they replaced traditional word representations. This paved the way towards larger contextualised models that replaced downstream architectures in most tasks. These approaches, trained with language modeling objectives, range from LSTM-based architectures such as ULMFiT BIBREF15 to the successful transformer-based architectures such as GPT2 BIBREF8, BERT BIBREF7, RoBERTa BIBREF9 and more recently ALBERT BIBREF16 and T5 BIBREF10. Related Work ::: Non-contextual word embeddings for languages other than English Since the introduction of word2vec BIBREF3, many attempts have been made to create monolingual models for a wide range of languages. For non-contextual word embeddings, the first two attempts were by BIBREF17 and BIBREF18 who created word embeddings for a large number of languages using Wikipedia. Later BIBREF19 trained fastText word embeddings for 157 languages using Common Crawl and showed that using crawled data significantly increased the performance of the embeddings relatively to those trained only on Wikipedia. Related Work ::: Contextualised models for languages other than English Following the success of large pretrained language models, they were extended to the multilingual setting with multilingual BERT , a single multilingual model for 104 different languages trained on Wikipedia data, and later XLM BIBREF12, which greatly improved unsupervised machine translation. A few monolingual models have been released: ELMo models for Japanese, Portuguese, German and Basque and BERT for Simplified and Traditional Chinese and German. However, to the best of our knowledge, no particular effort has been made toward training models for languages other than English, at a scale similar to the latest English models (e.g. RoBERTa trained on more than 100GB of data). CamemBERT Our approach is based on RoBERTa BIBREF9, which replicates and improves the initial BERT by identifying key hyper-parameters for more robust performance. In this section, we describe the architecture, training objective, optimisation setup and pretraining data that was used for CamemBERT. CamemBERT differs from RoBERTa mainly with the addition of whole-word masking and the usage of SentencePiece tokenisation BIBREF20. CamemBERT ::: Architecture Similar to RoBERTa and BERT, CamemBERT is a multi-layer bidirectional Transformer BIBREF21. Given the widespread usage of Transformers, we do not describe them in detail here and refer the reader to BIBREF21. CamemBERT uses the original BERT $_{\small \textsc {BASE}}$ configuration: 12 layers, 768 hidden dimensions, 12 attention heads, which amounts to 110M parameters. CamemBERT ::: Pretraining objective We train our model on the Masked Language Modeling (MLM) task. Given an input text sequence composed of $N$ tokens $x_1, ..., x_N$, we select $15\%$ of tokens for possible replacement. Among those selected tokens, 80% are replaced with the special $<$mask$>$ token, 10% are left unchanged and 10% are replaced by a random token. The model is then trained to predict the initial masked tokens using cross-entropy loss. Following RoBERTa we dynamically mask tokens instead of fixing them statically for the whole dataset during preprocessing. This improves variability and makes the model more robust when training for multiple epochs. Since we segment the input sentence into subwords using SentencePiece, the input tokens to the models can be subwords. An upgraded version of BERT and BIBREF22 have shown that masking whole words instead of individual subwords leads to improved performance. Whole-word masking (WWM) makes the training task more difficult because the model has to predict a whole word instead of predicting only part of the word given the rest. As a result, we used WWM for CamemBERT by first randomly sampling 15% of the words in the sequence and then considering all subword tokens in each of these 15% words for candidate replacement. This amounts to a proportion of selected tokens that is close to the original 15%. These tokens are then either replaced by $<$mask$>$ tokens (80%), left unchanged (10%) or replaced by a random token. Subsequent work has shown that the next sentence prediction task (NSP) originally used in BERT does not improve downstream task performance BIBREF12, BIBREF9, we do not use NSP as a consequence. CamemBERT ::: Optimisation Following BIBREF9, we optimise the model using Adam BIBREF23 ($\beta _1 = 0.9$, $\beta _2 = 0.98$) for 100k steps. We use large batch sizes of 8192 sequences. Each sequence contains at most 512 tokens. We enforce each sequence to only contain complete sentences. Additionally, we used the DOC-SENTENCES scenario from BIBREF9, consisting of not mixing multiple documents in the same sequence, which showed slightly better results. CamemBERT ::: Segmentation into subword units We segment the input text into subword units using SentencePiece BIBREF20. SentencePiece is an extension of Byte-Pair encoding (BPE) BIBREF24 and WordPiece BIBREF25 that does not require pre-tokenisation (at the word or token level), thus removing the need for language-specific tokenisers. We use a vocabulary size of 32k subword tokens. These are learned on $10^7$ sentences sampled from the pretraining dataset. We do not use subword regularisation (i.e. sampling from multiple possible segmentations) in our implementation for simplicity. CamemBERT ::: Pretraining data Pretrained language models can be significantly improved by using more data BIBREF9, BIBREF10. Therefore we used French text extracted from Common Crawl, in particular, we use OSCAR BIBREF13 a pre-classified and pre-filtered version of the November 2018 Common Craw snapshot. OSCAR is a set of monolingual corpora extracted from Common Crawl, specifically from the plain text WET format distributed by Common Crawl, which removes all HTML tags and converts all text encodings to UTF-8. OSCAR follows the same approach as BIBREF19 by using a language classification model based on the fastText linear classifier BIBREF26, BIBREF27 pretrained on Wikipedia, Tatoeba and SETimes, which supports 176 different languages. OSCAR performs a deduplication step after language classification and without introducing a specialised filtering scheme, other than only keeping paragraphs containing 100 or more UTF-8 encoded characters, making OSCAR quite close to the original Crawled data. We use the unshuffled version of the French OSCAR corpus, which amounts to 138GB of uncompressed text and 32.7B SentencePiece tokens. Evaluation ::: Part-of-speech tagging and dependency parsing We fist evaluate CamemBERT on the two downstream tasks of part-of-speech (POS) tagging and dependency parsing. POS tagging is a low-level syntactic task, which consists in assigning to each word its corresponding grammatical category. Dependency parsing consists in predicting the labeled syntactic tree capturing the syntactic relations between words. We run our experiments using the Universal Dependencies (UD) paradigm and its corresponding UD POS tag set BIBREF28 and UD treebank collection version 2.2 BIBREF29, which was used for the CoNLL 2018 shared task. We perform our work on the four freely available French UD treebanks in UD v2.2: GSD, Sequoia, Spoken, and ParTUT. GSD BIBREF30 is the second-largest treebank available for French after the FTB (described in subsection SECREF25), it contains data from blogs, news articles, reviews, and Wikipedia. The Sequoia treebank BIBREF31, BIBREF32 comprises more than 3000 sentences, from the French Europarl, the regional newspaper L’Est Républicain, the French Wikipedia and documents from the European Medicines Agency. Spoken is a corpus converted automatically from the Rhapsodie treebank BIBREF33, BIBREF34 with manual corrections. It consists of 57 sound samples of spoken French with orthographic transcription and phonetic transcription aligned with sound (word boundaries, syllables, and phonemes), syntactic and prosodic annotations. Finally, ParTUT is a conversion of a multilingual parallel treebank developed at the University of Turin, and consisting of a variety of text genres, including talks, legal texts, and Wikipedia articles, among others; ParTUT data is derived from the already-existing parallel treebank Par(allel)TUT BIBREF35 . Table TABREF23 contains a summary comparing the sizes of the treebanks. We evaluate the performance of our models using the standard UPOS accuracy for POS tagging, and Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS) for dependency parsing. We assume gold tokenisation and gold word segmentation as provided in the UD treebanks. Evaluation ::: Part-of-speech tagging and dependency parsing ::: Baselines To demonstrate the value of building a dedicated version of BERT for French, we first compare CamemBERT to the multilingual cased version of BERT (designated as mBERT). We then compare our models to UDify BIBREF36. UDify is a multitask and multilingual model based on mBERT that is near state-of-the-art on all UD languages including French for both POS tagging and dependency parsing. It is relevant to compare CamemBERT to UDify on those tasks because UDify is the work that pushed the furthest the performance in fine-tuning end-to-end a BERT-based model on downstream POS tagging and dependency parsing. Finally, we compare our model to UDPipe Future BIBREF37, a model ranked 3rd in dependency parsing and 6th in POS tagging during the CoNLL 2018 shared task BIBREF38. UDPipe Future provides us a strong baseline that does not make use of any pretrained contextual embedding. We will compare to the more recent cross-lingual language model XLM BIBREF12, as well as the state-of-the-art CoNLL 2018 shared task results with predicted tokenisation and segmentation in an updated version of the paper. Evaluation ::: Named Entity Recognition Named Entity Recognition (NER) is a sequence labeling task that consists in predicting which words refer to real-world objects, such as people, locations, artifacts and organisations. We use the French Treebank (FTB) BIBREF39 in its 2008 version introduced by cc-clustering:09short and with NER annotations by sagot2012annotation. The NER-annotated FTB contains more than 12k sentences and more than 350k tokens extracted from articles of the newspaper Le Monde published between 1989 and 1995. In total, it contains 11,636 entity mentions distributed among 7 different types of entities, namely: 2025 mentions of “Person”, 3761 of “Location”, 2382 of “Organisation”, 3357 of “Company”, 67 of “Product”, 15 of “POI” (Point of Interest) and 29 of “Fictional Character”. A large proportion of the entity mentions in the treebank are multi-word entities. For NER we therefore report the 3 metrics that are commonly used to evaluate models: precision, recall, and F1 score. Here precision measures the percentage of entities found by the system that are correctly tagged, recall measures the percentage of named entities present in the corpus that are found and the F1 score combines both precision and recall measures giving a general idea of a model's performance. Evaluation ::: Named Entity Recognition ::: Baselines Most of the advances in NER haven been achieved on English, particularly focusing on the CoNLL 2003 BIBREF40 and the Ontonotes v5 BIBREF41, BIBREF42 English corpora. NER is a task that was traditionally tackled using Conditional Random Fields (CRF) BIBREF43 which are quite suited for NER; CRFs were later used as decoding layers for Bi-LSTM architectures BIBREF44, BIBREF45 showing considerable improvements over CRFs alone. These Bi-LSTM-CRF architectures were later enhanced with contextualised word embeddings which yet again brought major improvements to the task BIBREF5, BIBREF6. Finally, large pretrained architectures settled the current state of the art showing a small yet important improvement over previous NER-specific architectures BIBREF7, BIBREF46. In non-English NER the CoNLL 2002 shared task included NER corpora for Spanish and Dutch corpora BIBREF47 while the CoNLL 2003 included a German corpus BIBREF40. Here the recent efforts of BIBREF48 settled the state of the art for Spanish and Dutch, while BIBREF6 did it for German. In French, no extensive work has been done due to the limited availability of NER corpora. We compare our model with the strong baselines settled by BIBREF49, who trained both CRF and BiLSTM-CRF architectures on the FTB and enhanced them using heuristics and pretrained word embeddings. Evaluation ::: Natural Language Inference We also evaluate our model on the Natural Language Inference (NLI) task, using the French part of the XNLI dataset BIBREF50. NLI consists in predicting whether a hypothesis sentence is entailed, neutral or contradicts a premise sentence. The XNLI dataset is the extension of the Multi-Genre NLI (MultiNLI) corpus BIBREF51 to 15 languages by translating the validation and test sets manually into each of those languages. The English training set is also machine translated for all languages. The dataset is composed of 122k train, 2490 valid and 5010 test examples. As usual, NLI performance is evaluated using accuracy. To evaluate a model on a language other than English (such as French), we consider the two following settings: TRANSLATE-TEST: The French test set is machine translated into English, and then used with an English classification model. This setting provides a reasonable, although imperfect, way to circumvent the fact that no such data set exists for French, and results in very strong baseline scores. TRANSLATE-TRAIN: The French model is fine-tuned on the machine-translated English training set and then evaluated on the French test set. This is the setting that we used for CamemBERT. Evaluation ::: Natural Language Inference ::: Baselines For the TRANSLATE-TEST setting, we report results of the English RoBERTa to act as a reference. In the TRANSLATE-TRAIN setting, we report the best scores from previous literature along with ours. BiLSTM-max is the best model in the original XNLI paper, mBERT which has been reported in French in BIBREF52 and XLM (MLM+TLM) is the best-presented model from BIBREF50. Experiments In this section, we measure the performance of CamemBERT by evaluating it on the four aforementioned tasks: POS tagging, dependency parsing, NER and NLI. Experiments ::: Experimental Setup ::: Pretraining We use the RoBERTa implementation in the fairseq library BIBREF53. Our learning rate is warmed up for 10k steps up to a peak value of $0.0007$ instead of the original $0.0001$ given our large batch size (8192). The learning rate fades to zero with polynomial decay. We pretrain our model on 256 Nvidia V100 GPUs (32GB each) for 100k steps during 17h. Experiments ::: Experimental Setup ::: Fine-tuning For each task, we append the relevant predictive layer on top of CamemBERT's Transformer architecture. Following the work done on BERT BIBREF7, for sequence tagging and sequence labeling we append a linear layer respectively to the $<$s$>$ special token and to the first subword token of each word. For dependency parsing, we plug a bi-affine graph predictor head as inspired by BIBREF54 following the work done on multilingual parsing with BERT by BIBREF36. We refer the reader to these two articles for more details on this module. We fine-tune independently CamemBERT for each task and each dataset. We optimise the model using the Adam optimiser BIBREF23 with a fixed learning rate. We run a grid search on a combination of learning rates and batch sizes. We select the best model on the validation set out of the 30 first epochs. Although this might push the performances even further, for all tasks except NLI, we don't apply any regularisation techniques such as weight decay, learning rate warm-up or discriminative fine-tuning. We show that fine-tuning CamemBERT in a straight-forward manner leads to state-of-the-art results on most tasks and outperforms the existing BERT-based models in most cases. The POS tagging, dependency parsing, and NER experiments are run using hugging face's Transformer library extended to support CamemBERT and dependency parsing BIBREF55. The NLI experiments use the fairseq library following the RoBERTa implementation. Experiments ::: Results ::: Part-of-Speech tagging and dependency parsing For POS tagging and dependency parsing, we compare CamemBERT to three other near state-of-the-art models in Table TABREF32. CamemBERT outperforms UDPipe Future by a large margin for all treebanks and all metrics. Despite a much simpler optimisation process, CamemBERT beats UDify performances on all the available French treebanks. CamemBERT also demonstrates higher performances than mBERT on those tasks. We observe a larger error reduction for parsing than for tagging. For POS tagging, we observe error reductions of respectively 0.71% for GSD, 0.81% for Sequoia, 0.7% for Spoken and 0.28% for ParTUT. For parsing, we observe error reductions in LAS of 2.96% for GSD, 3.33% for Sequoia, 1.70% for Spoken and 1.65% for ParTUT. Experiments ::: Results ::: Natural Language Inference: XNLI On the XNLI benchmark, CamemBERT obtains improved performance over multilingual language models on the TRANSLATE-TRAIN setting (81.2 vs. 80.2 for XLM) while using less than half the parameters (110M vs. 250M). However, its performance still lags behind models trained on the original English training set in the TRANSLATE-TEST setting, 81.2 vs. 82.91 for RoBERTa. It should be noted that CamemBERT uses far fewer parameters than RoBERTa (110M vs. 355M parameters). Experiments ::: Results ::: Named-Entity Recognition For named entity recognition, our experiments show that CamemBERT achieves a slightly better precision than the traditional CRF-based SEM architectures described above in Section SECREF25 (CRF and Bi-LSTM+CRF), but shows a dramatic improvement in finding entity mentions, raising the recall score by 3.5 points. Both improvements result in a 2.36 point increase in the F1 score with respect to the best SEM architecture (BiLSTM-CRF), giving CamemBERT the state of the art for NER on the FTB. One other important finding is the results obtained by mBERT. Previous work with this model showed increased performance in NER for German, Dutch and Spanish when mBERT is used as contextualised word embedding for an NER-specific model BIBREF48, but our results suggest that the multilingual setting in which mBERT was trained is simply not enough to use it alone and fine-tune it for French NER, as it shows worse performance than even simple CRF models, suggesting that monolingual models could be better at NER. Experiments ::: Discussion CamemBERT displays improved performance compared to prior work for the 4 downstream tasks considered. This confirms the hypothesis that pretrained language models can be effectively fine-tuned for various downstream tasks, as observed for English in previous work. Moreover, our results also show that dedicated monolingual models still outperform multilingual ones. We explain this point in two ways. First, the scale of data is possibly essential to the performance of CamemBERT. Indeed, we use 138GB of uncompressed text vs. 57GB for mBERT. Second, with more data comes more diversity in the pretraining distribution. Reaching state-of-the-art performances on 4 different tasks and 6 different datasets requires robust pretrained models. Our results suggest that the variability in the downstream tasks and datasets considered is handled more efficiently by a general language model than by Wikipedia-pretrained models such as mBERT. Conclusion CamemBERT improves the state of the art for multiple downstream tasks in French. It is also lighter than other BERT-based approaches such as mBERT or XLM. By releasing our model, we hope that it can serve as a strong baseline for future research in French NLP, and expect our experiments to be reproduced in many other languages. We will publish an updated version in the near future where we will explore and release models trained for longer, with additional downstream tasks, baselines (e.g. XLM) and analysis, we will also train additional models with potentially cleaner corpora such as CCNet BIBREF56 for more accurate performance evaluation and more complete ablation. Acknowledgments This work was partly funded by three French National grants from the Agence Nationale de la Recherche, namely projects PARSITI (ANR-16-CE33-0021), SoSweet (ANR-15-CE38-0011) and BASNUM (ANR-18-CE38-0003), as well as by the last author's chair in the PRAIRIE institute. Appendix ::: Impact of Whole-Word Masking We analyze the addition of whole-word masking on the downstream performance of CamemBERT. As reported for English on other downstream tasks, whole word masking improves downstream performances for all tasks but NER as seen in Table TABREF46. NER is highly sensitive to capitalisation, prefixes, suffixes and other subword features that could be used by a model to correctly identify entity mentions. Thus the added information by learning the masking at a subword level rather than at whole-word level seems to have a detrimental effect on downstream NER results.

unshuffled version of the French OSCAR corpus

bf25a202ac713a34e09bf599b3601058d9cace46

bf25a202ac713a34e09bf599b3601058d9cace46_0

Q: What are the state of the art measures? Text: Introduction Controversy is a phenomenom with a high impact at various levels. It has been broadly studied from the perspective of different disciplines, ranging from the seminal analysis of the conflicts within the members of a karate club BIBREF0 to political issues in modern times BIBREF1, BIBREF2. The irruption of digital social networks BIBREF3 gave raise to new ways of intentionally intervening on them for taking some advantage BIBREF4, BIBREF5. Moreover highly contrasting points of view in some groups tend to provoke conflicts that lead to attacks from one community to the other by harassing, “brigading”, or “trolling” it BIBREF6. The existing literature shows different issues that controversy brings up such as splitting of communities, biased information, hateful discussions and attacks between groups, generally proposing ways to solve them. For example, Kumar, Srijan, et al. BIBREF6 analyze many techniques to defend us from attacks in Reddit while Stewart, et al. BIBREF4 insinuate that there was external interference in Twitter during the 2016 US presidential elections to benefit one candidate. Also, as shown in BIBREF7, detecting controversy could provide the basis to improve the “news diet" of readers, offering the possibility to connect users with different points of views by recommending them new content to read BIBREF8. Moreover, other studies on “bridging echo chambers” BIBREF9 and the positive effects of intergroup dialogue BIBREF10, BIBREF11 suggest that direct engagement could be effective for mitigating such conflicts. Therefore, easily and automatically identifying controversial topics could allow us to quickly implement different strategies for preventing miss-information, fights and bias. Quantifying the controversy is even more powerful, as it allows us to establish controversy levels, and in particular to classify controversial and non-controversial topics by establishing a threshold score that separates the two types of topics. With this aim, we propose in this work a systematic, language-agnostic method to quantify controversy on social networks taking tweet's content as root input. Our main contribution is a new vocabulary-based method that works in any language and equates the performance of state-of-the-art structure-based methods. Finally, controversy quantification through vocabulary analysis opens several research avenues to analyze whether polarization is being created, maintained or augmented by the ways of talking of each community. Having this in mind and if we draw from the premise that when a discussion has a high controversy it is in general due to the presence of two principal communities fighting each other (or, conversely, that when there is no controversy there is just one principal community the members of which share a common point of view), we can measure the controversy by detecting if the discussion has one or two principal jargons in use. Our method is tested on Twitter datasets. This microblogging platform has been widely used to analyze discussions and polarization BIBREF12, BIBREF13, BIBREF14, BIBREF15, BIBREF2. It is a natural choice for these kind of problems, as it represents one of the main fora for public debate in online social media BIBREF15, it is a common destination for affiliative expressions BIBREF16 and is often used to report and read news about current events BIBREF17. An extra advantage of Twitter for this kind of studies is the availability of real-time data generated by millions of users. Other social media platforms offer similar data-sharing services, but few can match the amount of data and the accompanied documentation provided by Twitter. One last asset of Twitter for our work is given by retweets, whom typically indicate endorsement BIBREF18 and hence become a useful concept to model discussions as we can set “who is with who". However, our method has a general approach and it could be used a priori in any social network. In this work we report excellent result tested on Twitter but in future work we are going to test it in other social networks. Our paper is organized as follows: in Section SECREF2, we review related work. Section SECREF3 contains the detailed explanation of the pipeline we use for quantifying controversy of a topic, and each of its stages. In Section SECREF4 we report the results of an extensive empirical evaluation of the proposed measure of controversy. Finally, Section SECREF5 is devoted to discuss possible improvements and directions for future work, as well as lessons learned. Related work Many previous works are dedicated to quantifying the polarization observed in online social networks and social media BIBREF1, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23. The main characteristic of those works is that the measures proposed are based on the structural characteristics of the underlying graph. Among them, we highlight the work of Garimella et al.BIBREF23 that presents an extensive comparison of controversy measures, different graph-building approaches, and data sources, achieving the best performance of all. In their research they propose different metrics to measure polarization on Twitter. Their techniques based on the structure of the endorsement graph can successfully detect whether a discussion (represented by a set of tweets), is controversial or not regardless of the context and most importantly, without the need of any domain expertise. They also consider two different methods to measure controversy based on the analysis of the posts contents, but both fail when used to create a measure of controversy. Matakos et al. BIBREF24 develop a polarization index. Their measure captures the tendency of opinions to concentrate in network communities, creating echo-chambers. They obtain a good performance at identifying controversy by taking into account both the network structure and the existing opinions of users. However, they model opinions as positive or negative with a real number between -1 and 1. Their performance is good, but although it is an opinion-based method it is not a text-related one.Other recent works BIBREF25, BIBREF26, BIBREF27 have shown that communities may express themselves with different terms or ways of speaking, use different jargon, which in turn can be detected with the use of text-related techniques. In his thesis BIBREF28, Jang explains controversy via generating a summary of two conflicting stances that make up the controversy. This work shows that a specific sub-set of tweets could represent the two opposite positions in a polarized debate. A good tool to see how communities interact is ForceAtlas2 BIBREF29, a force-directed layout widely used for visualization. This layout has been recently found to be very useful at visualizing community interactions BIBREF30, as this algorithm will draw groups with little communication between them in different areas, whereas, if they have many interactions they will be drawn closer to each other. Therefore, whenever there is controversy the layout will show two well separated groups and will tend to show only one big community otherwise. The method we propose to measure the controversy equates in accuracy the one developed by Garimella et al.BIBREF23 and improves considerably computing time and robustness wrt the amount of data needed to effectively apply it. Our method is also based on a graph approach but it has its main focus on the vocabulary. We first train an NLP classifier that estimates opinion polarity of main users, then we run label-propagation BIBREF31 on the endorsement graph to get polarity of the whole network. Finally we compute the controversy score through a computation inspired in Dipole Moment, a measure used in physics to estimate electric polarity on a system. In our experiments we use the same data-sets from other works BIBREF32, BIBREF23, BIBREF33 as well as other datasets that we collected by us using a similar criterion (described in Section SECREF4). Method Our approach to measuring controversy is based on a systematic way of characterizing social media activity through its content. We employ a pipeline with five stages, namely graph building, community identification, model training, predicting and controversy measure. The final output of the pipeline is a value that measures how controversial a topic is, with higher values corresponding to higher degrees of controversy. The method is based on analysing posts content through Fasttext BIBREF34, a library for efficient learning of word representations and sentence classification developed by Facebook Research team. In short, our method works as follows: through Fasttext we train a language-agnostic model which can predict the community of many users by their jargon. Then we take there predictions and compute a score based on the physic notion Dipole Moment using a language approach to identify core or characteristic users and set the polarity trough them. We provide a detailed description of each stage in the following. Graph Building This paragraph provides details about the approach used to build graphs from raw data. As we said in Section SECREF1, we extract our discussions from Twitter. Our purpose is to build a conversation graph that represents activity related to a single topic of discussion -a debate about a specific event. For each topic, we build a graph $G$ where we assign a vertex to each user who contributes to it and we add a directed edge from node $u$ to node $v$ whenever user $u$ retweets a tweet posted by $v$. Retweets typically indicate endorsement BIBREF18: users who retweet signal endorsement of the opinion expressed in the original tweet by propagating it further. Retweets are not constrained to occur only between users who are connected in Twitter's social network, but users are allowed to retweet posts generated by any other user. As many other works in literature BIBREF5, BIBREF35, BIBREF36, BIBREF37, BIBREF4, BIBREF2 we establish that one retweet among a pair of users are needed to define an edge between them. Community Identification To identify a community's jargon we need to be very accurate at defining its members. If we, in our will of finding two principal communities, force the partition of the graph in that precise number of communities, we may be adding noise in the jargon of the principal communities that are fighting each other. Because of that, we decide to cluster the graph trying two popular algorithms: Walktrap BIBREF38 and Louvain BIBREF39. Both are structure-based algorithms that have very good performance with respect to the Modularity Q measure. These techniques does not detect a fixed number of clusters; their output will depend on the Modularity Q optimization, resulting in less “noisy" communities. The main differences between the two methods, in what regards our work, are that Louvain is a much faster heuristic algorithm but produces clusters with worse Modularity Q. Therefore, in order to analyze the trade-off between computing time and quality we decide to test both methods. At this step we want to capture the tweets of the principal communities to create the model that could differentiate them. Therefore, we take the two communities identified by the cluster algorithm that have the maximum number of users, and use them for the following step of our method. Model Training After detecting the principal communities we create our training dataset to feed the model. To do that, we extract the tweets of each cluster, we sanitize and we subject them to some transformations. First, we remove duplicate tweets -e.g. retweets without additional text. Second, we remove from the text of the tweets user names, links, punctuation, tabs, leading and lagging blanks, general spaces and “RT" - the text that points that a tweet is in fact a retweet. As shown in previous works, emojis are correlated with sentiment BIBREF40. Moreover, as we think that communities will express different sentiment during discussion, it is forseeable that emojis will play an important role as separators of tweets that differentiate between the two sides. Accordingly, we decide to add them to the train-set by translating each emoji into a different word. For example, the emoji :) will be translated into happy and :( into sad. Relations between emojis and words are defined in the R library textclean. Finally, we group tweets by user concatenating them in one string and labeling them with the user's community, namely with tags C1 and C2, corresponding respectively to the biggest and second biggest groups. It is important to note that we take the same number of users of each community to prevent bias in the model. Thus, we use the number of users of the smallest principal community. The train-set built that way is used to feed the model. As we said, we use Fasttext BIBREF34 to do this training. To define the values of the hyper-parameters we use the findings of BIBREF41. In their work they investigate the best hyper-parameters to train word embedding models using Fasttext BIBREF34 and Twitter data. We also change the default value of the hyper-parameter epoch to 20 instead of 5 because we want more convergence preventing as much as possible the variance between different training. These values could change in other context or social networks where we have more text per user or different discussion dynamics. Predicting The next stage consists of identifying the characteristic users of each side the discussion. These are the users that better represent the jargon of each side. To do that, tweets of the users belonging to the largest connected component of the graph are sanitized and transformed exactly as in the Training step. We decide to restrict to the largest connected component because in all cases it contains more than 90% of the nodes. The remaining 10% of the users don't participate in the discussion from a collective point of view but rather in an isolated way and this kind of intervention does not add interesting information to our approach. Then, we remove from this component users with degree smaller or equal to 2 (i.e. users that were retweeted by another user or retweeted other person less than three times in total). Their participation in the discussion is marginal, consequently they are not relevant wrt controversy as they add more noise than information at measuring time. This step could be adjusted differently in a different social network. We name this result component root-graph. Finally, let's see how we do classification. Considering that Fasttext returns for each classification both the predicted tag and the probability of the prediction, we classify each user of the resulting component by his sanitized tweets with our trained model, and take users that were tagged with a probability greater or equal than 0.9. These are the characteristic users that will be used in next step to compute the controversy measure. Controversy Measure This section describes the controversy measures used in this work. This computation is inspired in the measure presented by Morales et al. BIBREF2, and is based on the notion of dipole moment that has its origin in physics. First, we assign to the characteristic users the probability returned by the model, negativizing them if the predicted tag was C2. Therefore, these users are assigned values in the set [-1,-0.9] $\cup $ [0.9,1]. Then, we set values for the rest of the users of the root-graph by label-propagation BIBREF31 - an iterative algorithm to propagate values through a graph by node's neighborhood. Let $n^{+}$ and $n^{-}$ be the number of vertices $V$ with positive and negative values, respectively, and $\Delta A = \dfrac{\mid n^{+} - n^{-}\mid }{\mid V \mid }$ the absolute difference of their normalized size. Moreover, let $gc^{+}$ ($gc^{-}$) be the average value among vertices $n^{+}$ ($n^{-}$) and set $\tau $ as half their absolute difference, $\tau = \dfrac{\mid gc^{+} - gc^{- }\mid }{2}$. The dipole moment content controversy measure is defined as: $\textit {DMC} = (1 -\Delta A)\tau $. The rationale for this measure is that if the two sides are well separated, then label propagation will assign different extreme values to the two partitions, where users from one community will have values near to 1 and users from the other to -1, leading to higher values of the DMC measure. Note also that larger differences in the size of the two partitions (reflected in the value of $\Delta A$) lead to smaller values for the measure, which takes values between zero and one. Experiments In this section we report the results obtained by running the above proposed method over different discussions. Experiments ::: Topic definition In the literature, a topic is often defined by a single hashtag. However, this might be too restrictive in many cases. In our approach, a topic is operationalized as an specific hashtags or key words. Sometimes a discussion in a particular moment could not have a defined hashtag but it could be around a certain keyword, i.e. a word or expression that is not specifically a hashtag but it is widely used in the topic. For example during the Brazilian presidential elections in 2018 we captured the discussion by the mentions to the word Bolsonaro, that is the principal candidate's surname. Thus, for each topic we retrieve all the tweets that contain one of its hashtags or the keyword and that are generated during the observation window. We also ensure that the selected topic is associated with a large enough volume of activity. Experiments ::: Datasets In this section we detail the discussions we use to test our metric and how we determine the ground truth (i.e. if the discussion is controversial or not). We use thirty different discussions that took place between March 2015 and June 2019, half of them with controversy and half without it. We considered discussions in four different languages: English, Portuguese, Spanish and French, occurring in five regions over the world: South and North America, Western Europe, Central and Southern Asia. We also studied these discussions taking first 140 characters and then 280 from each tweet to analyze the difference in performance and computing time wrt the length of the posts. To define the amount of data needed to run our method we established that the Fasttext model has to predict at least one user of each community with a probability greater or equal than 0.9 during ten different trainings. If that is not the case, we are not able to use DPC method. This decision made us consider only a subset of the datasets used in BIBREF23, because due to the time elapsed since their work, many tweets had been deleted and consequently the volume of the data was not enough for our framework. To enlarge our experiment base we added new debates, more detailed information about each one is shown in Table TABREF24 in UNKREF6. To select new discussions and to determine if they are controversial or not we looked for topics widely covered by mainstream media, and that have generated ample discussion, both online and offline. For non-controversy discussions we focused on “soft news" and entertainment, but also to events that, while being impactful and/or dramatic, did not generate large controversies. To validate that intuition, we manually checked a sample of tweets, being unable to identify any clear instance of controversy On the other side, for controversial debates we focused on political events such as elections, corruption cases or justice decisions. To furtherly establish the presence of absence of controversy of our datasets, we visualized the corresponding networks through ForceAtlas2 BIBREF29. Figures FIGREF9 and FIGREF9 show an example of how non-controversial and controversial discussions look like respectively with ForceAtlas2 layout. As we can see in these figures, in a controversial discussion this layout tends to show two well separated groups while in a non-controversial one it tends to be only one big group. More information on the discussions is given in Table TABREF24. To avoid potential overfitting, we use only twelve graphs as testbed during the development of the measures, half of them controversial (netanyahu, ukraine, @mauriciomacri 1-11 Jan, Kavanaugh 3 Oct, @mauriciomacri 11-18 Mar, Bolsonaro 27 Oct) and half non-controversial (sxsw, germanwings, onedirection, ultralive, nepal, mothersday). This procedure resembles a 40/60% train/test split in traditional machine learning applications. Some of the discussions we consider refer to the same topics but in different periods of time. We needed to split them because our computing infrastructure does not allow us to compute such an enormous amount of data. However, being able to estimate controversy with only a subset of the discussion is an advantage, because discussions could take many days or months and we want to identify controversy as soon as possible, without the need of downloading the whole discussion. Moreover, for very long lasting discussions in social networks gathering the whole data would be impractical for any method. Experiments ::: Results Training a Fasttext model is not a deterministic process, as different runs could yield different results even using the same training set in each one. To analyze if these differences are significant, we decide to compute 20 scores for each discussion. The standard deviations among these 20 scores were low in all cases, with mean 0.01 and maximum 0.05. Consequently, we decided to report in this paper the average between the 20 scores, in practice taking the average between 5 runs would be enough. Figure FIGREF18 reports the scores computed by our measure in each topic for the two cluster methods. The beanplot shows the estimated probability density function for a measure computed on the topics, the individual observations are shown as small white lines in a one-dimensional scatter plot, and the median as a longer black line. The beanplot is divided into two groups, one for controversial topics (left/dark) and one for non-controversial ones (right/light). Hence, the black group shows the score distribution over controversial discussions and the white group over non-controversial ones. A larger separation of the two distributions indicates that the measure is better at capturing the characteristics of controversial topics, because a good separation allows to establish a threshold in the score that separates controversial and non-controversial discussions. As we may see in the figure, the medians are well separated in both cases, with little overlapping. To better quantify this overlap we measure the sensitivity BIBREF42 of these predictions by measuring the area under the ROC curve (AUC ROC), obtaining a value of 0.98 for Walktrap clustering and 0.967 for Louvain (where 1 represents a perfect separation and 0.5 means that they are indistinguishable). As Garimella et al. BIBREF23 have made their code public , we reproduced their best method Randomwalk on our datasets and measured the AUC ROC, obtaining a score of 0.935. An interesting finding was that their method had a poor performance over their own datasets. This was due to the fact (already explained in Section SECREF4) that it was not possible to retrieve the complete discussions, moreover, in no case could we restore more than 50% of the tweets. So we decided to remove these discussions and measure again the AUC ROC of this method, obtaining a 0.99 value. Our hypothesis is that the performance of that method was seriously hurt by the incompleteness of the data. We also tested our method on these datasets, obtaining a 0.99 AUC ROC with Walktrap and 0.989 with Louvain clustering. We conclude that our method works better, as in practice both approaches show same performances -specially with Walktrap, but in presence of incomplete information our measure is more robust. The performance of Louvain is slightly worse but, as we mentioned in Section SECREF3, this method is much faster. Therefore, we decided to compare the running time of our method with both clustering techniques and also with the Randomwalk algorithm. In figure FIGREF18 we can see the distribution of running times of all techniques through box plots. Both versions of our method are faster than Randomwalk, while Louvain is faster than Walktrap. We now analyze the impact of the length of the considered text in our method. Figure FIGREF18 depicts the results of similar experiment as Figure FIGREF18, but considering only 140 characters per tweet. As we may see, here the overlapping is bigger, having an AUC of 0.88. As for the impact on computing time, we observe that despite of the results of BIBREF34 that reported a complexity of O(h $log_{2}$(k)) at training and test tasks, in practice we observed a linear growth. We measured the running times of the training and predicting phases (the two text-related phases of our method), the resulting times are reported in figure FIGREF18, which shows running time as a function of the text-size. We include also the best estimated function that approximate computing time as a function of text-set size. As it may be seen, time grows almost linearly, ranging from 30 seconds for a set of 111 KB to 84 seconds for a set of 11941 KB. Finally, we measured running times for the whole method over each dataset with 280 characters. Times were between 170 and 2467 seconds with a mean of 842, making it in practice a reasonable amount of time. Discussions The task we address in this work is certainly not an easy one, and our study has some limitations, which we discuss in this section. Our work lead us to some conclusions regarding the overall possibility of measuring controversy through text, and what aspects need to be considered to deepen our work. Discussions ::: Limitations As our approach to controversy is similar to that of Garimella et al. BIBREF23, we share some of their limitations with respect to several aspects: Evaluation -difficulties to establish ground-truth, Multisided controversies -controversy with more than two sides, Choice of data - manually pick topics, and Overfitting - small set of experiments. Although we have more discussions, it is still small set from statistical point of view. Apart from that, our language-based approach has other limitations which we mention in the following, together with their solutions or mitigation. Data-size. Training an NLP model that can predict tags with a probability greater or equal than 0.9 requires significant amount of text, therefore our method works only for “big" discussions. Most interesting controversies are those that have consequence at a society level, in general big enough for our method. Multi-language discussions. When multiple languages are participating in a discussion it is common that users tend to retweet more tweets in their own language, creating sub-communities. In this cases our model will tend to predict higher controversy scores. This is the case for example of #germanwings, where users tweet in English, German and Spanish and it has the highest score in no-controversial topics. However, the polarization that we tackle in this work is normally part of a society cell (a nation, a city, etc.), and thus developed in just one language. We think that limiting the effectiveness of our analysis to single-language discussions is not a serious limitation. Twitter only. Our findings are based on datasets coming from Twitter. While this is certainly a limitation, Twitter is one of the main venues for online public discussion, and one of the few for which data is available. Hence, Twitter is a natural choice. However, Twitter's characteristic limit of 280 characters per message (140 till short time ago) is an intrinsic limitation of that network. We think that in other social networks as Facebook or Reddit our method will work even better, as having more text per user could redound on a better NLP model as we verified comparing the results with 140 and 280 characters per post. Discussions ::: Conclusions In this article, we introduced the first large-scale systematic method for quantifying controversy in social media through content. We have shown that this method works on Spanish, English, French and Portuguese, it is context-agnostic and does not require the intervention of a domain expert. We have compared its performance with state-of-the-art structure-based controversy measures showing that they have same performance and it is more robust. We also have shown that more text implies better performance and without significantly increasing computing time, therefore, it could be used in other contexts such as other social networks like Reddit or Facebook and we are going to test it in future works. Training the model is not an expensive task since Fasttext has a good performance at this. However, the best performance for detecting principal communities is obtained by Walktrap. The complexity of that algorithm is O(m$n^2$)BIBREF38, where $m$ and $n$ are the number of edges and vertices respectively. This makes this method rather expensive to compute on big networks. Nevertheless, we have shown that with Louvain the method still obtains a very similar AUC ROC (0.99 with Walktrap and 0.989 with Louvain). With incomplete information its performance gets worse but it is still good (0.96) and better than previous state of the art. This work opens several avenues for future research. One is identifying what words, semantics/concepts or language expressions make differ one community from the other. There are various ways to do this, for instance through the word-embbedings that Fasttext returns after training BIBREF34. Also we could use interpretability techniques on machine learning models BIBREF43. Finally, we could try other techniques for measuring controversy through text, using another NLP model as pre-trained neural network BERT BIBREF44 or, in a completely different approach measuring the dispersion index of the discussions word-embbedings BIBREF25. We are currently starting to follow this direction. Details on the discussions F

Randomwalk, Walktrap, Louvain clustering

abebf9c8c9cf70ae222ecb1d3cabf8115b9fc8ac

abebf9c8c9cf70ae222ecb1d3cabf8115b9fc8ac_0

Q: What controversial topics are experimented with? Text: Introduction Controversy is a phenomenom with a high impact at various levels. It has been broadly studied from the perspective of different disciplines, ranging from the seminal analysis of the conflicts within the members of a karate club BIBREF0 to political issues in modern times BIBREF1, BIBREF2. The irruption of digital social networks BIBREF3 gave raise to new ways of intentionally intervening on them for taking some advantage BIBREF4, BIBREF5. Moreover highly contrasting points of view in some groups tend to provoke conflicts that lead to attacks from one community to the other by harassing, “brigading”, or “trolling” it BIBREF6. The existing literature shows different issues that controversy brings up such as splitting of communities, biased information, hateful discussions and attacks between groups, generally proposing ways to solve them. For example, Kumar, Srijan, et al. BIBREF6 analyze many techniques to defend us from attacks in Reddit while Stewart, et al. BIBREF4 insinuate that there was external interference in Twitter during the 2016 US presidential elections to benefit one candidate. Also, as shown in BIBREF7, detecting controversy could provide the basis to improve the “news diet" of readers, offering the possibility to connect users with different points of views by recommending them new content to read BIBREF8. Moreover, other studies on “bridging echo chambers” BIBREF9 and the positive effects of intergroup dialogue BIBREF10, BIBREF11 suggest that direct engagement could be effective for mitigating such conflicts. Therefore, easily and automatically identifying controversial topics could allow us to quickly implement different strategies for preventing miss-information, fights and bias. Quantifying the controversy is even more powerful, as it allows us to establish controversy levels, and in particular to classify controversial and non-controversial topics by establishing a threshold score that separates the two types of topics. With this aim, we propose in this work a systematic, language-agnostic method to quantify controversy on social networks taking tweet's content as root input. Our main contribution is a new vocabulary-based method that works in any language and equates the performance of state-of-the-art structure-based methods. Finally, controversy quantification through vocabulary analysis opens several research avenues to analyze whether polarization is being created, maintained or augmented by the ways of talking of each community. Having this in mind and if we draw from the premise that when a discussion has a high controversy it is in general due to the presence of two principal communities fighting each other (or, conversely, that when there is no controversy there is just one principal community the members of which share a common point of view), we can measure the controversy by detecting if the discussion has one or two principal jargons in use. Our method is tested on Twitter datasets. This microblogging platform has been widely used to analyze discussions and polarization BIBREF12, BIBREF13, BIBREF14, BIBREF15, BIBREF2. It is a natural choice for these kind of problems, as it represents one of the main fora for public debate in online social media BIBREF15, it is a common destination for affiliative expressions BIBREF16 and is often used to report and read news about current events BIBREF17. An extra advantage of Twitter for this kind of studies is the availability of real-time data generated by millions of users. Other social media platforms offer similar data-sharing services, but few can match the amount of data and the accompanied documentation provided by Twitter. One last asset of Twitter for our work is given by retweets, whom typically indicate endorsement BIBREF18 and hence become a useful concept to model discussions as we can set “who is with who". However, our method has a general approach and it could be used a priori in any social network. In this work we report excellent result tested on Twitter but in future work we are going to test it in other social networks. Our paper is organized as follows: in Section SECREF2, we review related work. Section SECREF3 contains the detailed explanation of the pipeline we use for quantifying controversy of a topic, and each of its stages. In Section SECREF4 we report the results of an extensive empirical evaluation of the proposed measure of controversy. Finally, Section SECREF5 is devoted to discuss possible improvements and directions for future work, as well as lessons learned. Related work Many previous works are dedicated to quantifying the polarization observed in online social networks and social media BIBREF1, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23. The main characteristic of those works is that the measures proposed are based on the structural characteristics of the underlying graph. Among them, we highlight the work of Garimella et al.BIBREF23 that presents an extensive comparison of controversy measures, different graph-building approaches, and data sources, achieving the best performance of all. In their research they propose different metrics to measure polarization on Twitter. Their techniques based on the structure of the endorsement graph can successfully detect whether a discussion (represented by a set of tweets), is controversial or not regardless of the context and most importantly, without the need of any domain expertise. They also consider two different methods to measure controversy based on the analysis of the posts contents, but both fail when used to create a measure of controversy. Matakos et al. BIBREF24 develop a polarization index. Their measure captures the tendency of opinions to concentrate in network communities, creating echo-chambers. They obtain a good performance at identifying controversy by taking into account both the network structure and the existing opinions of users. However, they model opinions as positive or negative with a real number between -1 and 1. Their performance is good, but although it is an opinion-based method it is not a text-related one.Other recent works BIBREF25, BIBREF26, BIBREF27 have shown that communities may express themselves with different terms or ways of speaking, use different jargon, which in turn can be detected with the use of text-related techniques. In his thesis BIBREF28, Jang explains controversy via generating a summary of two conflicting stances that make up the controversy. This work shows that a specific sub-set of tweets could represent the two opposite positions in a polarized debate. A good tool to see how communities interact is ForceAtlas2 BIBREF29, a force-directed layout widely used for visualization. This layout has been recently found to be very useful at visualizing community interactions BIBREF30, as this algorithm will draw groups with little communication between them in different areas, whereas, if they have many interactions they will be drawn closer to each other. Therefore, whenever there is controversy the layout will show two well separated groups and will tend to show only one big community otherwise. The method we propose to measure the controversy equates in accuracy the one developed by Garimella et al.BIBREF23 and improves considerably computing time and robustness wrt the amount of data needed to effectively apply it. Our method is also based on a graph approach but it has its main focus on the vocabulary. We first train an NLP classifier that estimates opinion polarity of main users, then we run label-propagation BIBREF31 on the endorsement graph to get polarity of the whole network. Finally we compute the controversy score through a computation inspired in Dipole Moment, a measure used in physics to estimate electric polarity on a system. In our experiments we use the same data-sets from other works BIBREF32, BIBREF23, BIBREF33 as well as other datasets that we collected by us using a similar criterion (described in Section SECREF4). Method Our approach to measuring controversy is based on a systematic way of characterizing social media activity through its content. We employ a pipeline with five stages, namely graph building, community identification, model training, predicting and controversy measure. The final output of the pipeline is a value that measures how controversial a topic is, with higher values corresponding to higher degrees of controversy. The method is based on analysing posts content through Fasttext BIBREF34, a library for efficient learning of word representations and sentence classification developed by Facebook Research team. In short, our method works as follows: through Fasttext we train a language-agnostic model which can predict the community of many users by their jargon. Then we take there predictions and compute a score based on the physic notion Dipole Moment using a language approach to identify core or characteristic users and set the polarity trough them. We provide a detailed description of each stage in the following. Graph Building This paragraph provides details about the approach used to build graphs from raw data. As we said in Section SECREF1, we extract our discussions from Twitter. Our purpose is to build a conversation graph that represents activity related to a single topic of discussion -a debate about a specific event. For each topic, we build a graph $G$ where we assign a vertex to each user who contributes to it and we add a directed edge from node $u$ to node $v$ whenever user $u$ retweets a tweet posted by $v$. Retweets typically indicate endorsement BIBREF18: users who retweet signal endorsement of the opinion expressed in the original tweet by propagating it further. Retweets are not constrained to occur only between users who are connected in Twitter's social network, but users are allowed to retweet posts generated by any other user. As many other works in literature BIBREF5, BIBREF35, BIBREF36, BIBREF37, BIBREF4, BIBREF2 we establish that one retweet among a pair of users are needed to define an edge between them. Community Identification To identify a community's jargon we need to be very accurate at defining its members. If we, in our will of finding two principal communities, force the partition of the graph in that precise number of communities, we may be adding noise in the jargon of the principal communities that are fighting each other. Because of that, we decide to cluster the graph trying two popular algorithms: Walktrap BIBREF38 and Louvain BIBREF39. Both are structure-based algorithms that have very good performance with respect to the Modularity Q measure. These techniques does not detect a fixed number of clusters; their output will depend on the Modularity Q optimization, resulting in less “noisy" communities. The main differences between the two methods, in what regards our work, are that Louvain is a much faster heuristic algorithm but produces clusters with worse Modularity Q. Therefore, in order to analyze the trade-off between computing time and quality we decide to test both methods. At this step we want to capture the tweets of the principal communities to create the model that could differentiate them. Therefore, we take the two communities identified by the cluster algorithm that have the maximum number of users, and use them for the following step of our method. Model Training After detecting the principal communities we create our training dataset to feed the model. To do that, we extract the tweets of each cluster, we sanitize and we subject them to some transformations. First, we remove duplicate tweets -e.g. retweets without additional text. Second, we remove from the text of the tweets user names, links, punctuation, tabs, leading and lagging blanks, general spaces and “RT" - the text that points that a tweet is in fact a retweet. As shown in previous works, emojis are correlated with sentiment BIBREF40. Moreover, as we think that communities will express different sentiment during discussion, it is forseeable that emojis will play an important role as separators of tweets that differentiate between the two sides. Accordingly, we decide to add them to the train-set by translating each emoji into a different word. For example, the emoji :) will be translated into happy and :( into sad. Relations between emojis and words are defined in the R library textclean. Finally, we group tweets by user concatenating them in one string and labeling them with the user's community, namely with tags C1 and C2, corresponding respectively to the biggest and second biggest groups. It is important to note that we take the same number of users of each community to prevent bias in the model. Thus, we use the number of users of the smallest principal community. The train-set built that way is used to feed the model. As we said, we use Fasttext BIBREF34 to do this training. To define the values of the hyper-parameters we use the findings of BIBREF41. In their work they investigate the best hyper-parameters to train word embedding models using Fasttext BIBREF34 and Twitter data. We also change the default value of the hyper-parameter epoch to 20 instead of 5 because we want more convergence preventing as much as possible the variance between different training. These values could change in other context or social networks where we have more text per user or different discussion dynamics. Predicting The next stage consists of identifying the characteristic users of each side the discussion. These are the users that better represent the jargon of each side. To do that, tweets of the users belonging to the largest connected component of the graph are sanitized and transformed exactly as in the Training step. We decide to restrict to the largest connected component because in all cases it contains more than 90% of the nodes. The remaining 10% of the users don't participate in the discussion from a collective point of view but rather in an isolated way and this kind of intervention does not add interesting information to our approach. Then, we remove from this component users with degree smaller or equal to 2 (i.e. users that were retweeted by another user or retweeted other person less than three times in total). Their participation in the discussion is marginal, consequently they are not relevant wrt controversy as they add more noise than information at measuring time. This step could be adjusted differently in a different social network. We name this result component root-graph. Finally, let's see how we do classification. Considering that Fasttext returns for each classification both the predicted tag and the probability of the prediction, we classify each user of the resulting component by his sanitized tweets with our trained model, and take users that were tagged with a probability greater or equal than 0.9. These are the characteristic users that will be used in next step to compute the controversy measure. Controversy Measure This section describes the controversy measures used in this work. This computation is inspired in the measure presented by Morales et al. BIBREF2, and is based on the notion of dipole moment that has its origin in physics. First, we assign to the characteristic users the probability returned by the model, negativizing them if the predicted tag was C2. Therefore, these users are assigned values in the set [-1,-0.9] $\cup $ [0.9,1]. Then, we set values for the rest of the users of the root-graph by label-propagation BIBREF31 - an iterative algorithm to propagate values through a graph by node's neighborhood. Let $n^{+}$ and $n^{-}$ be the number of vertices $V$ with positive and negative values, respectively, and $\Delta A = \dfrac{\mid n^{+} - n^{-}\mid }{\mid V \mid }$ the absolute difference of their normalized size. Moreover, let $gc^{+}$ ($gc^{-}$) be the average value among vertices $n^{+}$ ($n^{-}$) and set $\tau $ as half their absolute difference, $\tau = \dfrac{\mid gc^{+} - gc^{- }\mid }{2}$. The dipole moment content controversy measure is defined as: $\textit {DMC} = (1 -\Delta A)\tau $. The rationale for this measure is that if the two sides are well separated, then label propagation will assign different extreme values to the two partitions, where users from one community will have values near to 1 and users from the other to -1, leading to higher values of the DMC measure. Note also that larger differences in the size of the two partitions (reflected in the value of $\Delta A$) lead to smaller values for the measure, which takes values between zero and one. Experiments In this section we report the results obtained by running the above proposed method over different discussions. Experiments ::: Topic definition In the literature, a topic is often defined by a single hashtag. However, this might be too restrictive in many cases. In our approach, a topic is operationalized as an specific hashtags or key words. Sometimes a discussion in a particular moment could not have a defined hashtag but it could be around a certain keyword, i.e. a word or expression that is not specifically a hashtag but it is widely used in the topic. For example during the Brazilian presidential elections in 2018 we captured the discussion by the mentions to the word Bolsonaro, that is the principal candidate's surname. Thus, for each topic we retrieve all the tweets that contain one of its hashtags or the keyword and that are generated during the observation window. We also ensure that the selected topic is associated with a large enough volume of activity. Experiments ::: Datasets In this section we detail the discussions we use to test our metric and how we determine the ground truth (i.e. if the discussion is controversial or not). We use thirty different discussions that took place between March 2015 and June 2019, half of them with controversy and half without it. We considered discussions in four different languages: English, Portuguese, Spanish and French, occurring in five regions over the world: South and North America, Western Europe, Central and Southern Asia. We also studied these discussions taking first 140 characters and then 280 from each tweet to analyze the difference in performance and computing time wrt the length of the posts. To define the amount of data needed to run our method we established that the Fasttext model has to predict at least one user of each community with a probability greater or equal than 0.9 during ten different trainings. If that is not the case, we are not able to use DPC method. This decision made us consider only a subset of the datasets used in BIBREF23, because due to the time elapsed since their work, many tweets had been deleted and consequently the volume of the data was not enough for our framework. To enlarge our experiment base we added new debates, more detailed information about each one is shown in Table TABREF24 in UNKREF6. To select new discussions and to determine if they are controversial or not we looked for topics widely covered by mainstream media, and that have generated ample discussion, both online and offline. For non-controversy discussions we focused on “soft news" and entertainment, but also to events that, while being impactful and/or dramatic, did not generate large controversies. To validate that intuition, we manually checked a sample of tweets, being unable to identify any clear instance of controversy On the other side, for controversial debates we focused on political events such as elections, corruption cases or justice decisions. To furtherly establish the presence of absence of controversy of our datasets, we visualized the corresponding networks through ForceAtlas2 BIBREF29. Figures FIGREF9 and FIGREF9 show an example of how non-controversial and controversial discussions look like respectively with ForceAtlas2 layout. As we can see in these figures, in a controversial discussion this layout tends to show two well separated groups while in a non-controversial one it tends to be only one big group. More information on the discussions is given in Table TABREF24. To avoid potential overfitting, we use only twelve graphs as testbed during the development of the measures, half of them controversial (netanyahu, ukraine, @mauriciomacri 1-11 Jan, Kavanaugh 3 Oct, @mauriciomacri 11-18 Mar, Bolsonaro 27 Oct) and half non-controversial (sxsw, germanwings, onedirection, ultralive, nepal, mothersday). This procedure resembles a 40/60% train/test split in traditional machine learning applications. Some of the discussions we consider refer to the same topics but in different periods of time. We needed to split them because our computing infrastructure does not allow us to compute such an enormous amount of data. However, being able to estimate controversy with only a subset of the discussion is an advantage, because discussions could take many days or months and we want to identify controversy as soon as possible, without the need of downloading the whole discussion. Moreover, for very long lasting discussions in social networks gathering the whole data would be impractical for any method. Experiments ::: Results Training a Fasttext model is not a deterministic process, as different runs could yield different results even using the same training set in each one. To analyze if these differences are significant, we decide to compute 20 scores for each discussion. The standard deviations among these 20 scores were low in all cases, with mean 0.01 and maximum 0.05. Consequently, we decided to report in this paper the average between the 20 scores, in practice taking the average between 5 runs would be enough. Figure FIGREF18 reports the scores computed by our measure in each topic for the two cluster methods. The beanplot shows the estimated probability density function for a measure computed on the topics, the individual observations are shown as small white lines in a one-dimensional scatter plot, and the median as a longer black line. The beanplot is divided into two groups, one for controversial topics (left/dark) and one for non-controversial ones (right/light). Hence, the black group shows the score distribution over controversial discussions and the white group over non-controversial ones. A larger separation of the two distributions indicates that the measure is better at capturing the characteristics of controversial topics, because a good separation allows to establish a threshold in the score that separates controversial and non-controversial discussions. As we may see in the figure, the medians are well separated in both cases, with little overlapping. To better quantify this overlap we measure the sensitivity BIBREF42 of these predictions by measuring the area under the ROC curve (AUC ROC), obtaining a value of 0.98 for Walktrap clustering and 0.967 for Louvain (where 1 represents a perfect separation and 0.5 means that they are indistinguishable). As Garimella et al. BIBREF23 have made their code public , we reproduced their best method Randomwalk on our datasets and measured the AUC ROC, obtaining a score of 0.935. An interesting finding was that their method had a poor performance over their own datasets. This was due to the fact (already explained in Section SECREF4) that it was not possible to retrieve the complete discussions, moreover, in no case could we restore more than 50% of the tweets. So we decided to remove these discussions and measure again the AUC ROC of this method, obtaining a 0.99 value. Our hypothesis is that the performance of that method was seriously hurt by the incompleteness of the data. We also tested our method on these datasets, obtaining a 0.99 AUC ROC with Walktrap and 0.989 with Louvain clustering. We conclude that our method works better, as in practice both approaches show same performances -specially with Walktrap, but in presence of incomplete information our measure is more robust. The performance of Louvain is slightly worse but, as we mentioned in Section SECREF3, this method is much faster. Therefore, we decided to compare the running time of our method with both clustering techniques and also with the Randomwalk algorithm. In figure FIGREF18 we can see the distribution of running times of all techniques through box plots. Both versions of our method are faster than Randomwalk, while Louvain is faster than Walktrap. We now analyze the impact of the length of the considered text in our method. Figure FIGREF18 depicts the results of similar experiment as Figure FIGREF18, but considering only 140 characters per tweet. As we may see, here the overlapping is bigger, having an AUC of 0.88. As for the impact on computing time, we observe that despite of the results of BIBREF34 that reported a complexity of O(h $log_{2}$(k)) at training and test tasks, in practice we observed a linear growth. We measured the running times of the training and predicting phases (the two text-related phases of our method), the resulting times are reported in figure FIGREF18, which shows running time as a function of the text-size. We include also the best estimated function that approximate computing time as a function of text-set size. As it may be seen, time grows almost linearly, ranging from 30 seconds for a set of 111 KB to 84 seconds for a set of 11941 KB. Finally, we measured running times for the whole method over each dataset with 280 characters. Times were between 170 and 2467 seconds with a mean of 842, making it in practice a reasonable amount of time. Discussions The task we address in this work is certainly not an easy one, and our study has some limitations, which we discuss in this section. Our work lead us to some conclusions regarding the overall possibility of measuring controversy through text, and what aspects need to be considered to deepen our work. Discussions ::: Limitations As our approach to controversy is similar to that of Garimella et al. BIBREF23, we share some of their limitations with respect to several aspects: Evaluation -difficulties to establish ground-truth, Multisided controversies -controversy with more than two sides, Choice of data - manually pick topics, and Overfitting - small set of experiments. Although we have more discussions, it is still small set from statistical point of view. Apart from that, our language-based approach has other limitations which we mention in the following, together with their solutions or mitigation. Data-size. Training an NLP model that can predict tags with a probability greater or equal than 0.9 requires significant amount of text, therefore our method works only for “big" discussions. Most interesting controversies are those that have consequence at a society level, in general big enough for our method. Multi-language discussions. When multiple languages are participating in a discussion it is common that users tend to retweet more tweets in their own language, creating sub-communities. In this cases our model will tend to predict higher controversy scores. This is the case for example of #germanwings, where users tweet in English, German and Spanish and it has the highest score in no-controversial topics. However, the polarization that we tackle in this work is normally part of a society cell (a nation, a city, etc.), and thus developed in just one language. We think that limiting the effectiveness of our analysis to single-language discussions is not a serious limitation. Twitter only. Our findings are based on datasets coming from Twitter. While this is certainly a limitation, Twitter is one of the main venues for online public discussion, and one of the few for which data is available. Hence, Twitter is a natural choice. However, Twitter's characteristic limit of 280 characters per message (140 till short time ago) is an intrinsic limitation of that network. We think that in other social networks as Facebook or Reddit our method will work even better, as having more text per user could redound on a better NLP model as we verified comparing the results with 140 and 280 characters per post. Discussions ::: Conclusions In this article, we introduced the first large-scale systematic method for quantifying controversy in social media through content. We have shown that this method works on Spanish, English, French and Portuguese, it is context-agnostic and does not require the intervention of a domain expert. We have compared its performance with state-of-the-art structure-based controversy measures showing that they have same performance and it is more robust. We also have shown that more text implies better performance and without significantly increasing computing time, therefore, it could be used in other contexts such as other social networks like Reddit or Facebook and we are going to test it in future works. Training the model is not an expensive task since Fasttext has a good performance at this. However, the best performance for detecting principal communities is obtained by Walktrap. The complexity of that algorithm is O(m$n^2$)BIBREF38, where $m$ and $n$ are the number of edges and vertices respectively. This makes this method rather expensive to compute on big networks. Nevertheless, we have shown that with Louvain the method still obtains a very similar AUC ROC (0.99 with Walktrap and 0.989 with Louvain). With incomplete information its performance gets worse but it is still good (0.96) and better than previous state of the art. This work opens several avenues for future research. One is identifying what words, semantics/concepts or language expressions make differ one community from the other. There are various ways to do this, for instance through the word-embbedings that Fasttext returns after training BIBREF34. Also we could use interpretability techniques on machine learning models BIBREF43. Finally, we could try other techniques for measuring controversy through text, using another NLP model as pre-trained neural network BERT BIBREF44 or, in a completely different approach measuring the dispersion index of the discussions word-embbedings BIBREF25. We are currently starting to follow this direction. Details on the discussions F

political events such as elections, corruption cases or justice decisions

2df910c9806f0c379d7bb1bc2be2610438e487dc

2df910c9806f0c379d7bb1bc2be2610438e487dc_0

Q: What datasets did they use? Text: Introduction Controversy is a phenomenom with a high impact at various levels. It has been broadly studied from the perspective of different disciplines, ranging from the seminal analysis of the conflicts within the members of a karate club BIBREF0 to political issues in modern times BIBREF1, BIBREF2. The irruption of digital social networks BIBREF3 gave raise to new ways of intentionally intervening on them for taking some advantage BIBREF4, BIBREF5. Moreover highly contrasting points of view in some groups tend to provoke conflicts that lead to attacks from one community to the other by harassing, “brigading”, or “trolling” it BIBREF6. The existing literature shows different issues that controversy brings up such as splitting of communities, biased information, hateful discussions and attacks between groups, generally proposing ways to solve them. For example, Kumar, Srijan, et al. BIBREF6 analyze many techniques to defend us from attacks in Reddit while Stewart, et al. BIBREF4 insinuate that there was external interference in Twitter during the 2016 US presidential elections to benefit one candidate. Also, as shown in BIBREF7, detecting controversy could provide the basis to improve the “news diet" of readers, offering the possibility to connect users with different points of views by recommending them new content to read BIBREF8. Moreover, other studies on “bridging echo chambers” BIBREF9 and the positive effects of intergroup dialogue BIBREF10, BIBREF11 suggest that direct engagement could be effective for mitigating such conflicts. Therefore, easily and automatically identifying controversial topics could allow us to quickly implement different strategies for preventing miss-information, fights and bias. Quantifying the controversy is even more powerful, as it allows us to establish controversy levels, and in particular to classify controversial and non-controversial topics by establishing a threshold score that separates the two types of topics. With this aim, we propose in this work a systematic, language-agnostic method to quantify controversy on social networks taking tweet's content as root input. Our main contribution is a new vocabulary-based method that works in any language and equates the performance of state-of-the-art structure-based methods. Finally, controversy quantification through vocabulary analysis opens several research avenues to analyze whether polarization is being created, maintained or augmented by the ways of talking of each community. Having this in mind and if we draw from the premise that when a discussion has a high controversy it is in general due to the presence of two principal communities fighting each other (or, conversely, that when there is no controversy there is just one principal community the members of which share a common point of view), we can measure the controversy by detecting if the discussion has one or two principal jargons in use. Our method is tested on Twitter datasets. This microblogging platform has been widely used to analyze discussions and polarization BIBREF12, BIBREF13, BIBREF14, BIBREF15, BIBREF2. It is a natural choice for these kind of problems, as it represents one of the main fora for public debate in online social media BIBREF15, it is a common destination for affiliative expressions BIBREF16 and is often used to report and read news about current events BIBREF17. An extra advantage of Twitter for this kind of studies is the availability of real-time data generated by millions of users. Other social media platforms offer similar data-sharing services, but few can match the amount of data and the accompanied documentation provided by Twitter. One last asset of Twitter for our work is given by retweets, whom typically indicate endorsement BIBREF18 and hence become a useful concept to model discussions as we can set “who is with who". However, our method has a general approach and it could be used a priori in any social network. In this work we report excellent result tested on Twitter but in future work we are going to test it in other social networks. Our paper is organized as follows: in Section SECREF2, we review related work. Section SECREF3 contains the detailed explanation of the pipeline we use for quantifying controversy of a topic, and each of its stages. In Section SECREF4 we report the results of an extensive empirical evaluation of the proposed measure of controversy. Finally, Section SECREF5 is devoted to discuss possible improvements and directions for future work, as well as lessons learned. Related work Many previous works are dedicated to quantifying the polarization observed in online social networks and social media BIBREF1, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23. The main characteristic of those works is that the measures proposed are based on the structural characteristics of the underlying graph. Among them, we highlight the work of Garimella et al.BIBREF23 that presents an extensive comparison of controversy measures, different graph-building approaches, and data sources, achieving the best performance of all. In their research they propose different metrics to measure polarization on Twitter. Their techniques based on the structure of the endorsement graph can successfully detect whether a discussion (represented by a set of tweets), is controversial or not regardless of the context and most importantly, without the need of any domain expertise. They also consider two different methods to measure controversy based on the analysis of the posts contents, but both fail when used to create a measure of controversy. Matakos et al. BIBREF24 develop a polarization index. Their measure captures the tendency of opinions to concentrate in network communities, creating echo-chambers. They obtain a good performance at identifying controversy by taking into account both the network structure and the existing opinions of users. However, they model opinions as positive or negative with a real number between -1 and 1. Their performance is good, but although it is an opinion-based method it is not a text-related one.Other recent works BIBREF25, BIBREF26, BIBREF27 have shown that communities may express themselves with different terms or ways of speaking, use different jargon, which in turn can be detected with the use of text-related techniques. In his thesis BIBREF28, Jang explains controversy via generating a summary of two conflicting stances that make up the controversy. This work shows that a specific sub-set of tweets could represent the two opposite positions in a polarized debate. A good tool to see how communities interact is ForceAtlas2 BIBREF29, a force-directed layout widely used for visualization. This layout has been recently found to be very useful at visualizing community interactions BIBREF30, as this algorithm will draw groups with little communication between them in different areas, whereas, if they have many interactions they will be drawn closer to each other. Therefore, whenever there is controversy the layout will show two well separated groups and will tend to show only one big community otherwise. The method we propose to measure the controversy equates in accuracy the one developed by Garimella et al.BIBREF23 and improves considerably computing time and robustness wrt the amount of data needed to effectively apply it. Our method is also based on a graph approach but it has its main focus on the vocabulary. We first train an NLP classifier that estimates opinion polarity of main users, then we run label-propagation BIBREF31 on the endorsement graph to get polarity of the whole network. Finally we compute the controversy score through a computation inspired in Dipole Moment, a measure used in physics to estimate electric polarity on a system. In our experiments we use the same data-sets from other works BIBREF32, BIBREF23, BIBREF33 as well as other datasets that we collected by us using a similar criterion (described in Section SECREF4). Method Our approach to measuring controversy is based on a systematic way of characterizing social media activity through its content. We employ a pipeline with five stages, namely graph building, community identification, model training, predicting and controversy measure. The final output of the pipeline is a value that measures how controversial a topic is, with higher values corresponding to higher degrees of controversy. The method is based on analysing posts content through Fasttext BIBREF34, a library for efficient learning of word representations and sentence classification developed by Facebook Research team. In short, our method works as follows: through Fasttext we train a language-agnostic model which can predict the community of many users by their jargon. Then we take there predictions and compute a score based on the physic notion Dipole Moment using a language approach to identify core or characteristic users and set the polarity trough them. We provide a detailed description of each stage in the following. Graph Building This paragraph provides details about the approach used to build graphs from raw data. As we said in Section SECREF1, we extract our discussions from Twitter. Our purpose is to build a conversation graph that represents activity related to a single topic of discussion -a debate about a specific event. For each topic, we build a graph $G$ where we assign a vertex to each user who contributes to it and we add a directed edge from node $u$ to node $v$ whenever user $u$ retweets a tweet posted by $v$. Retweets typically indicate endorsement BIBREF18: users who retweet signal endorsement of the opinion expressed in the original tweet by propagating it further. Retweets are not constrained to occur only between users who are connected in Twitter's social network, but users are allowed to retweet posts generated by any other user. As many other works in literature BIBREF5, BIBREF35, BIBREF36, BIBREF37, BIBREF4, BIBREF2 we establish that one retweet among a pair of users are needed to define an edge between them. Community Identification To identify a community's jargon we need to be very accurate at defining its members. If we, in our will of finding two principal communities, force the partition of the graph in that precise number of communities, we may be adding noise in the jargon of the principal communities that are fighting each other. Because of that, we decide to cluster the graph trying two popular algorithms: Walktrap BIBREF38 and Louvain BIBREF39. Both are structure-based algorithms that have very good performance with respect to the Modularity Q measure. These techniques does not detect a fixed number of clusters; their output will depend on the Modularity Q optimization, resulting in less “noisy" communities. The main differences between the two methods, in what regards our work, are that Louvain is a much faster heuristic algorithm but produces clusters with worse Modularity Q. Therefore, in order to analyze the trade-off between computing time and quality we decide to test both methods. At this step we want to capture the tweets of the principal communities to create the model that could differentiate them. Therefore, we take the two communities identified by the cluster algorithm that have the maximum number of users, and use them for the following step of our method. Model Training After detecting the principal communities we create our training dataset to feed the model. To do that, we extract the tweets of each cluster, we sanitize and we subject them to some transformations. First, we remove duplicate tweets -e.g. retweets without additional text. Second, we remove from the text of the tweets user names, links, punctuation, tabs, leading and lagging blanks, general spaces and “RT" - the text that points that a tweet is in fact a retweet. As shown in previous works, emojis are correlated with sentiment BIBREF40. Moreover, as we think that communities will express different sentiment during discussion, it is forseeable that emojis will play an important role as separators of tweets that differentiate between the two sides. Accordingly, we decide to add them to the train-set by translating each emoji into a different word. For example, the emoji :) will be translated into happy and :( into sad. Relations between emojis and words are defined in the R library textclean. Finally, we group tweets by user concatenating them in one string and labeling them with the user's community, namely with tags C1 and C2, corresponding respectively to the biggest and second biggest groups. It is important to note that we take the same number of users of each community to prevent bias in the model. Thus, we use the number of users of the smallest principal community. The train-set built that way is used to feed the model. As we said, we use Fasttext BIBREF34 to do this training. To define the values of the hyper-parameters we use the findings of BIBREF41. In their work they investigate the best hyper-parameters to train word embedding models using Fasttext BIBREF34 and Twitter data. We also change the default value of the hyper-parameter epoch to 20 instead of 5 because we want more convergence preventing as much as possible the variance between different training. These values could change in other context or social networks where we have more text per user or different discussion dynamics. Predicting The next stage consists of identifying the characteristic users of each side the discussion. These are the users that better represent the jargon of each side. To do that, tweets of the users belonging to the largest connected component of the graph are sanitized and transformed exactly as in the Training step. We decide to restrict to the largest connected component because in all cases it contains more than 90% of the nodes. The remaining 10% of the users don't participate in the discussion from a collective point of view but rather in an isolated way and this kind of intervention does not add interesting information to our approach. Then, we remove from this component users with degree smaller or equal to 2 (i.e. users that were retweeted by another user or retweeted other person less than three times in total). Their participation in the discussion is marginal, consequently they are not relevant wrt controversy as they add more noise than information at measuring time. This step could be adjusted differently in a different social network. We name this result component root-graph. Finally, let's see how we do classification. Considering that Fasttext returns for each classification both the predicted tag and the probability of the prediction, we classify each user of the resulting component by his sanitized tweets with our trained model, and take users that were tagged with a probability greater or equal than 0.9. These are the characteristic users that will be used in next step to compute the controversy measure. Controversy Measure This section describes the controversy measures used in this work. This computation is inspired in the measure presented by Morales et al. BIBREF2, and is based on the notion of dipole moment that has its origin in physics. First, we assign to the characteristic users the probability returned by the model, negativizing them if the predicted tag was C2. Therefore, these users are assigned values in the set [-1,-0.9] $\cup $ [0.9,1]. Then, we set values for the rest of the users of the root-graph by label-propagation BIBREF31 - an iterative algorithm to propagate values through a graph by node's neighborhood. Let $n^{+}$ and $n^{-}$ be the number of vertices $V$ with positive and negative values, respectively, and $\Delta A = \dfrac{\mid n^{+} - n^{-}\mid }{\mid V \mid }$ the absolute difference of their normalized size. Moreover, let $gc^{+}$ ($gc^{-}$) be the average value among vertices $n^{+}$ ($n^{-}$) and set $\tau $ as half their absolute difference, $\tau = \dfrac{\mid gc^{+} - gc^{- }\mid }{2}$. The dipole moment content controversy measure is defined as: $\textit {DMC} = (1 -\Delta A)\tau $. The rationale for this measure is that if the two sides are well separated, then label propagation will assign different extreme values to the two partitions, where users from one community will have values near to 1 and users from the other to -1, leading to higher values of the DMC measure. Note also that larger differences in the size of the two partitions (reflected in the value of $\Delta A$) lead to smaller values for the measure, which takes values between zero and one. Experiments In this section we report the results obtained by running the above proposed method over different discussions. Experiments ::: Topic definition In the literature, a topic is often defined by a single hashtag. However, this might be too restrictive in many cases. In our approach, a topic is operationalized as an specific hashtags or key words. Sometimes a discussion in a particular moment could not have a defined hashtag but it could be around a certain keyword, i.e. a word or expression that is not specifically a hashtag but it is widely used in the topic. For example during the Brazilian presidential elections in 2018 we captured the discussion by the mentions to the word Bolsonaro, that is the principal candidate's surname. Thus, for each topic we retrieve all the tweets that contain one of its hashtags or the keyword and that are generated during the observation window. We also ensure that the selected topic is associated with a large enough volume of activity. Experiments ::: Datasets In this section we detail the discussions we use to test our metric and how we determine the ground truth (i.e. if the discussion is controversial or not). We use thirty different discussions that took place between March 2015 and June 2019, half of them with controversy and half without it. We considered discussions in four different languages: English, Portuguese, Spanish and French, occurring in five regions over the world: South and North America, Western Europe, Central and Southern Asia. We also studied these discussions taking first 140 characters and then 280 from each tweet to analyze the difference in performance and computing time wrt the length of the posts. To define the amount of data needed to run our method we established that the Fasttext model has to predict at least one user of each community with a probability greater or equal than 0.9 during ten different trainings. If that is not the case, we are not able to use DPC method. This decision made us consider only a subset of the datasets used in BIBREF23, because due to the time elapsed since their work, many tweets had been deleted and consequently the volume of the data was not enough for our framework. To enlarge our experiment base we added new debates, more detailed information about each one is shown in Table TABREF24 in UNKREF6. To select new discussions and to determine if they are controversial or not we looked for topics widely covered by mainstream media, and that have generated ample discussion, both online and offline. For non-controversy discussions we focused on “soft news" and entertainment, but also to events that, while being impactful and/or dramatic, did not generate large controversies. To validate that intuition, we manually checked a sample of tweets, being unable to identify any clear instance of controversy On the other side, for controversial debates we focused on political events such as elections, corruption cases or justice decisions. To furtherly establish the presence of absence of controversy of our datasets, we visualized the corresponding networks through ForceAtlas2 BIBREF29. Figures FIGREF9 and FIGREF9 show an example of how non-controversial and controversial discussions look like respectively with ForceAtlas2 layout. As we can see in these figures, in a controversial discussion this layout tends to show two well separated groups while in a non-controversial one it tends to be only one big group. More information on the discussions is given in Table TABREF24. To avoid potential overfitting, we use only twelve graphs as testbed during the development of the measures, half of them controversial (netanyahu, ukraine, @mauriciomacri 1-11 Jan, Kavanaugh 3 Oct, @mauriciomacri 11-18 Mar, Bolsonaro 27 Oct) and half non-controversial (sxsw, germanwings, onedirection, ultralive, nepal, mothersday). This procedure resembles a 40/60% train/test split in traditional machine learning applications. Some of the discussions we consider refer to the same topics but in different periods of time. We needed to split them because our computing infrastructure does not allow us to compute such an enormous amount of data. However, being able to estimate controversy with only a subset of the discussion is an advantage, because discussions could take many days or months and we want to identify controversy as soon as possible, without the need of downloading the whole discussion. Moreover, for very long lasting discussions in social networks gathering the whole data would be impractical for any method. Experiments ::: Results Training a Fasttext model is not a deterministic process, as different runs could yield different results even using the same training set in each one. To analyze if these differences are significant, we decide to compute 20 scores for each discussion. The standard deviations among these 20 scores were low in all cases, with mean 0.01 and maximum 0.05. Consequently, we decided to report in this paper the average between the 20 scores, in practice taking the average between 5 runs would be enough. Figure FIGREF18 reports the scores computed by our measure in each topic for the two cluster methods. The beanplot shows the estimated probability density function for a measure computed on the topics, the individual observations are shown as small white lines in a one-dimensional scatter plot, and the median as a longer black line. The beanplot is divided into two groups, one for controversial topics (left/dark) and one for non-controversial ones (right/light). Hence, the black group shows the score distribution over controversial discussions and the white group over non-controversial ones. A larger separation of the two distributions indicates that the measure is better at capturing the characteristics of controversial topics, because a good separation allows to establish a threshold in the score that separates controversial and non-controversial discussions. As we may see in the figure, the medians are well separated in both cases, with little overlapping. To better quantify this overlap we measure the sensitivity BIBREF42 of these predictions by measuring the area under the ROC curve (AUC ROC), obtaining a value of 0.98 for Walktrap clustering and 0.967 for Louvain (where 1 represents a perfect separation and 0.5 means that they are indistinguishable). As Garimella et al. BIBREF23 have made their code public , we reproduced their best method Randomwalk on our datasets and measured the AUC ROC, obtaining a score of 0.935. An interesting finding was that their method had a poor performance over their own datasets. This was due to the fact (already explained in Section SECREF4) that it was not possible to retrieve the complete discussions, moreover, in no case could we restore more than 50% of the tweets. So we decided to remove these discussions and measure again the AUC ROC of this method, obtaining a 0.99 value. Our hypothesis is that the performance of that method was seriously hurt by the incompleteness of the data. We also tested our method on these datasets, obtaining a 0.99 AUC ROC with Walktrap and 0.989 with Louvain clustering. We conclude that our method works better, as in practice both approaches show same performances -specially with Walktrap, but in presence of incomplete information our measure is more robust. The performance of Louvain is slightly worse but, as we mentioned in Section SECREF3, this method is much faster. Therefore, we decided to compare the running time of our method with both clustering techniques and also with the Randomwalk algorithm. In figure FIGREF18 we can see the distribution of running times of all techniques through box plots. Both versions of our method are faster than Randomwalk, while Louvain is faster than Walktrap. We now analyze the impact of the length of the considered text in our method. Figure FIGREF18 depicts the results of similar experiment as Figure FIGREF18, but considering only 140 characters per tweet. As we may see, here the overlapping is bigger, having an AUC of 0.88. As for the impact on computing time, we observe that despite of the results of BIBREF34 that reported a complexity of O(h $log_{2}$(k)) at training and test tasks, in practice we observed a linear growth. We measured the running times of the training and predicting phases (the two text-related phases of our method), the resulting times are reported in figure FIGREF18, which shows running time as a function of the text-size. We include also the best estimated function that approximate computing time as a function of text-set size. As it may be seen, time grows almost linearly, ranging from 30 seconds for a set of 111 KB to 84 seconds for a set of 11941 KB. Finally, we measured running times for the whole method over each dataset with 280 characters. Times were between 170 and 2467 seconds with a mean of 842, making it in practice a reasonable amount of time. Discussions The task we address in this work is certainly not an easy one, and our study has some limitations, which we discuss in this section. Our work lead us to some conclusions regarding the overall possibility of measuring controversy through text, and what aspects need to be considered to deepen our work. Discussions ::: Limitations As our approach to controversy is similar to that of Garimella et al. BIBREF23, we share some of their limitations with respect to several aspects: Evaluation -difficulties to establish ground-truth, Multisided controversies -controversy with more than two sides, Choice of data - manually pick topics, and Overfitting - small set of experiments. Although we have more discussions, it is still small set from statistical point of view. Apart from that, our language-based approach has other limitations which we mention in the following, together with their solutions or mitigation. Data-size. Training an NLP model that can predict tags with a probability greater or equal than 0.9 requires significant amount of text, therefore our method works only for “big" discussions. Most interesting controversies are those that have consequence at a society level, in general big enough for our method. Multi-language discussions. When multiple languages are participating in a discussion it is common that users tend to retweet more tweets in their own language, creating sub-communities. In this cases our model will tend to predict higher controversy scores. This is the case for example of #germanwings, where users tweet in English, German and Spanish and it has the highest score in no-controversial topics. However, the polarization that we tackle in this work is normally part of a society cell (a nation, a city, etc.), and thus developed in just one language. We think that limiting the effectiveness of our analysis to single-language discussions is not a serious limitation. Twitter only. Our findings are based on datasets coming from Twitter. While this is certainly a limitation, Twitter is one of the main venues for online public discussion, and one of the few for which data is available. Hence, Twitter is a natural choice. However, Twitter's characteristic limit of 280 characters per message (140 till short time ago) is an intrinsic limitation of that network. We think that in other social networks as Facebook or Reddit our method will work even better, as having more text per user could redound on a better NLP model as we verified comparing the results with 140 and 280 characters per post. Discussions ::: Conclusions In this article, we introduced the first large-scale systematic method for quantifying controversy in social media through content. We have shown that this method works on Spanish, English, French and Portuguese, it is context-agnostic and does not require the intervention of a domain expert. We have compared its performance with state-of-the-art structure-based controversy measures showing that they have same performance and it is more robust. We also have shown that more text implies better performance and without significantly increasing computing time, therefore, it could be used in other contexts such as other social networks like Reddit or Facebook and we are going to test it in future works. Training the model is not an expensive task since Fasttext has a good performance at this. However, the best performance for detecting principal communities is obtained by Walktrap. The complexity of that algorithm is O(m$n^2$)BIBREF38, where $m$ and $n$ are the number of edges and vertices respectively. This makes this method rather expensive to compute on big networks. Nevertheless, we have shown that with Louvain the method still obtains a very similar AUC ROC (0.99 with Walktrap and 0.989 with Louvain). With incomplete information its performance gets worse but it is still good (0.96) and better than previous state of the art. This work opens several avenues for future research. One is identifying what words, semantics/concepts or language expressions make differ one community from the other. There are various ways to do this, for instance through the word-embbedings that Fasttext returns after training BIBREF34. Also we could use interpretability techniques on machine learning models BIBREF43. Finally, we could try other techniques for measuring controversy through text, using another NLP model as pre-trained neural network BERT BIBREF44 or, in a completely different approach measuring the dispersion index of the discussions word-embbedings BIBREF25. We are currently starting to follow this direction. Details on the discussions F

BIBREF32, BIBREF23, BIBREF33, discussions in four different languages: English, Portuguese, Spanish and French, occurring in five regions over the world: South and North America, Western Europe, Central and Southern Asia.

a2a3af59f3f18a28eb2ca7055e1613948f395052

a2a3af59f3f18a28eb2ca7055e1613948f395052_0

Q: What social media platform is observed? Text: Introduction Controversy is a phenomenom with a high impact at various levels. It has been broadly studied from the perspective of different disciplines, ranging from the seminal analysis of the conflicts within the members of a karate club BIBREF0 to political issues in modern times BIBREF1, BIBREF2. The irruption of digital social networks BIBREF3 gave raise to new ways of intentionally intervening on them for taking some advantage BIBREF4, BIBREF5. Moreover highly contrasting points of view in some groups tend to provoke conflicts that lead to attacks from one community to the other by harassing, “brigading”, or “trolling” it BIBREF6. The existing literature shows different issues that controversy brings up such as splitting of communities, biased information, hateful discussions and attacks between groups, generally proposing ways to solve them. For example, Kumar, Srijan, et al. BIBREF6 analyze many techniques to defend us from attacks in Reddit while Stewart, et al. BIBREF4 insinuate that there was external interference in Twitter during the 2016 US presidential elections to benefit one candidate. Also, as shown in BIBREF7, detecting controversy could provide the basis to improve the “news diet" of readers, offering the possibility to connect users with different points of views by recommending them new content to read BIBREF8. Moreover, other studies on “bridging echo chambers” BIBREF9 and the positive effects of intergroup dialogue BIBREF10, BIBREF11 suggest that direct engagement could be effective for mitigating such conflicts. Therefore, easily and automatically identifying controversial topics could allow us to quickly implement different strategies for preventing miss-information, fights and bias. Quantifying the controversy is even more powerful, as it allows us to establish controversy levels, and in particular to classify controversial and non-controversial topics by establishing a threshold score that separates the two types of topics. With this aim, we propose in this work a systematic, language-agnostic method to quantify controversy on social networks taking tweet's content as root input. Our main contribution is a new vocabulary-based method that works in any language and equates the performance of state-of-the-art structure-based methods. Finally, controversy quantification through vocabulary analysis opens several research avenues to analyze whether polarization is being created, maintained or augmented by the ways of talking of each community. Having this in mind and if we draw from the premise that when a discussion has a high controversy it is in general due to the presence of two principal communities fighting each other (or, conversely, that when there is no controversy there is just one principal community the members of which share a common point of view), we can measure the controversy by detecting if the discussion has one or two principal jargons in use. Our method is tested on Twitter datasets. This microblogging platform has been widely used to analyze discussions and polarization BIBREF12, BIBREF13, BIBREF14, BIBREF15, BIBREF2. It is a natural choice for these kind of problems, as it represents one of the main fora for public debate in online social media BIBREF15, it is a common destination for affiliative expressions BIBREF16 and is often used to report and read news about current events BIBREF17. An extra advantage of Twitter for this kind of studies is the availability of real-time data generated by millions of users. Other social media platforms offer similar data-sharing services, but few can match the amount of data and the accompanied documentation provided by Twitter. One last asset of Twitter for our work is given by retweets, whom typically indicate endorsement BIBREF18 and hence become a useful concept to model discussions as we can set “who is with who". However, our method has a general approach and it could be used a priori in any social network. In this work we report excellent result tested on Twitter but in future work we are going to test it in other social networks. Our paper is organized as follows: in Section SECREF2, we review related work. Section SECREF3 contains the detailed explanation of the pipeline we use for quantifying controversy of a topic, and each of its stages. In Section SECREF4 we report the results of an extensive empirical evaluation of the proposed measure of controversy. Finally, Section SECREF5 is devoted to discuss possible improvements and directions for future work, as well as lessons learned. Related work Many previous works are dedicated to quantifying the polarization observed in online social networks and social media BIBREF1, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23. The main characteristic of those works is that the measures proposed are based on the structural characteristics of the underlying graph. Among them, we highlight the work of Garimella et al.BIBREF23 that presents an extensive comparison of controversy measures, different graph-building approaches, and data sources, achieving the best performance of all. In their research they propose different metrics to measure polarization on Twitter. Their techniques based on the structure of the endorsement graph can successfully detect whether a discussion (represented by a set of tweets), is controversial or not regardless of the context and most importantly, without the need of any domain expertise. They also consider two different methods to measure controversy based on the analysis of the posts contents, but both fail when used to create a measure of controversy. Matakos et al. BIBREF24 develop a polarization index. Their measure captures the tendency of opinions to concentrate in network communities, creating echo-chambers. They obtain a good performance at identifying controversy by taking into account both the network structure and the existing opinions of users. However, they model opinions as positive or negative with a real number between -1 and 1. Their performance is good, but although it is an opinion-based method it is not a text-related one.Other recent works BIBREF25, BIBREF26, BIBREF27 have shown that communities may express themselves with different terms or ways of speaking, use different jargon, which in turn can be detected with the use of text-related techniques. In his thesis BIBREF28, Jang explains controversy via generating a summary of two conflicting stances that make up the controversy. This work shows that a specific sub-set of tweets could represent the two opposite positions in a polarized debate. A good tool to see how communities interact is ForceAtlas2 BIBREF29, a force-directed layout widely used for visualization. This layout has been recently found to be very useful at visualizing community interactions BIBREF30, as this algorithm will draw groups with little communication between them in different areas, whereas, if they have many interactions they will be drawn closer to each other. Therefore, whenever there is controversy the layout will show two well separated groups and will tend to show only one big community otherwise. The method we propose to measure the controversy equates in accuracy the one developed by Garimella et al.BIBREF23 and improves considerably computing time and robustness wrt the amount of data needed to effectively apply it. Our method is also based on a graph approach but it has its main focus on the vocabulary. We first train an NLP classifier that estimates opinion polarity of main users, then we run label-propagation BIBREF31 on the endorsement graph to get polarity of the whole network. Finally we compute the controversy score through a computation inspired in Dipole Moment, a measure used in physics to estimate electric polarity on a system. In our experiments we use the same data-sets from other works BIBREF32, BIBREF23, BIBREF33 as well as other datasets that we collected by us using a similar criterion (described in Section SECREF4). Method Our approach to measuring controversy is based on a systematic way of characterizing social media activity through its content. We employ a pipeline with five stages, namely graph building, community identification, model training, predicting and controversy measure. The final output of the pipeline is a value that measures how controversial a topic is, with higher values corresponding to higher degrees of controversy. The method is based on analysing posts content through Fasttext BIBREF34, a library for efficient learning of word representations and sentence classification developed by Facebook Research team. In short, our method works as follows: through Fasttext we train a language-agnostic model which can predict the community of many users by their jargon. Then we take there predictions and compute a score based on the physic notion Dipole Moment using a language approach to identify core or characteristic users and set the polarity trough them. We provide a detailed description of each stage in the following. Graph Building This paragraph provides details about the approach used to build graphs from raw data. As we said in Section SECREF1, we extract our discussions from Twitter. Our purpose is to build a conversation graph that represents activity related to a single topic of discussion -a debate about a specific event. For each topic, we build a graph $G$ where we assign a vertex to each user who contributes to it and we add a directed edge from node $u$ to node $v$ whenever user $u$ retweets a tweet posted by $v$. Retweets typically indicate endorsement BIBREF18: users who retweet signal endorsement of the opinion expressed in the original tweet by propagating it further. Retweets are not constrained to occur only between users who are connected in Twitter's social network, but users are allowed to retweet posts generated by any other user. As many other works in literature BIBREF5, BIBREF35, BIBREF36, BIBREF37, BIBREF4, BIBREF2 we establish that one retweet among a pair of users are needed to define an edge between them. Community Identification To identify a community's jargon we need to be very accurate at defining its members. If we, in our will of finding two principal communities, force the partition of the graph in that precise number of communities, we may be adding noise in the jargon of the principal communities that are fighting each other. Because of that, we decide to cluster the graph trying two popular algorithms: Walktrap BIBREF38 and Louvain BIBREF39. Both are structure-based algorithms that have very good performance with respect to the Modularity Q measure. These techniques does not detect a fixed number of clusters; their output will depend on the Modularity Q optimization, resulting in less “noisy" communities. The main differences between the two methods, in what regards our work, are that Louvain is a much faster heuristic algorithm but produces clusters with worse Modularity Q. Therefore, in order to analyze the trade-off between computing time and quality we decide to test both methods. At this step we want to capture the tweets of the principal communities to create the model that could differentiate them. Therefore, we take the two communities identified by the cluster algorithm that have the maximum number of users, and use them for the following step of our method. Model Training After detecting the principal communities we create our training dataset to feed the model. To do that, we extract the tweets of each cluster, we sanitize and we subject them to some transformations. First, we remove duplicate tweets -e.g. retweets without additional text. Second, we remove from the text of the tweets user names, links, punctuation, tabs, leading and lagging blanks, general spaces and “RT" - the text that points that a tweet is in fact a retweet. As shown in previous works, emojis are correlated with sentiment BIBREF40. Moreover, as we think that communities will express different sentiment during discussion, it is forseeable that emojis will play an important role as separators of tweets that differentiate between the two sides. Accordingly, we decide to add them to the train-set by translating each emoji into a different word. For example, the emoji :) will be translated into happy and :( into sad. Relations between emojis and words are defined in the R library textclean. Finally, we group tweets by user concatenating them in one string and labeling them with the user's community, namely with tags C1 and C2, corresponding respectively to the biggest and second biggest groups. It is important to note that we take the same number of users of each community to prevent bias in the model. Thus, we use the number of users of the smallest principal community. The train-set built that way is used to feed the model. As we said, we use Fasttext BIBREF34 to do this training. To define the values of the hyper-parameters we use the findings of BIBREF41. In their work they investigate the best hyper-parameters to train word embedding models using Fasttext BIBREF34 and Twitter data. We also change the default value of the hyper-parameter epoch to 20 instead of 5 because we want more convergence preventing as much as possible the variance between different training. These values could change in other context or social networks where we have more text per user or different discussion dynamics. Predicting The next stage consists of identifying the characteristic users of each side the discussion. These are the users that better represent the jargon of each side. To do that, tweets of the users belonging to the largest connected component of the graph are sanitized and transformed exactly as in the Training step. We decide to restrict to the largest connected component because in all cases it contains more than 90% of the nodes. The remaining 10% of the users don't participate in the discussion from a collective point of view but rather in an isolated way and this kind of intervention does not add interesting information to our approach. Then, we remove from this component users with degree smaller or equal to 2 (i.e. users that were retweeted by another user or retweeted other person less than three times in total). Their participation in the discussion is marginal, consequently they are not relevant wrt controversy as they add more noise than information at measuring time. This step could be adjusted differently in a different social network. We name this result component root-graph. Finally, let's see how we do classification. Considering that Fasttext returns for each classification both the predicted tag and the probability of the prediction, we classify each user of the resulting component by his sanitized tweets with our trained model, and take users that were tagged with a probability greater or equal than 0.9. These are the characteristic users that will be used in next step to compute the controversy measure. Controversy Measure This section describes the controversy measures used in this work. This computation is inspired in the measure presented by Morales et al. BIBREF2, and is based on the notion of dipole moment that has its origin in physics. First, we assign to the characteristic users the probability returned by the model, negativizing them if the predicted tag was C2. Therefore, these users are assigned values in the set [-1,-0.9] $\cup $ [0.9,1]. Then, we set values for the rest of the users of the root-graph by label-propagation BIBREF31 - an iterative algorithm to propagate values through a graph by node's neighborhood. Let $n^{+}$ and $n^{-}$ be the number of vertices $V$ with positive and negative values, respectively, and $\Delta A = \dfrac{\mid n^{+} - n^{-}\mid }{\mid V \mid }$ the absolute difference of their normalized size. Moreover, let $gc^{+}$ ($gc^{-}$) be the average value among vertices $n^{+}$ ($n^{-}$) and set $\tau $ as half their absolute difference, $\tau = \dfrac{\mid gc^{+} - gc^{- }\mid }{2}$. The dipole moment content controversy measure is defined as: $\textit {DMC} = (1 -\Delta A)\tau $. The rationale for this measure is that if the two sides are well separated, then label propagation will assign different extreme values to the two partitions, where users from one community will have values near to 1 and users from the other to -1, leading to higher values of the DMC measure. Note also that larger differences in the size of the two partitions (reflected in the value of $\Delta A$) lead to smaller values for the measure, which takes values between zero and one. Experiments In this section we report the results obtained by running the above proposed method over different discussions. Experiments ::: Topic definition In the literature, a topic is often defined by a single hashtag. However, this might be too restrictive in many cases. In our approach, a topic is operationalized as an specific hashtags or key words. Sometimes a discussion in a particular moment could not have a defined hashtag but it could be around a certain keyword, i.e. a word or expression that is not specifically a hashtag but it is widely used in the topic. For example during the Brazilian presidential elections in 2018 we captured the discussion by the mentions to the word Bolsonaro, that is the principal candidate's surname. Thus, for each topic we retrieve all the tweets that contain one of its hashtags or the keyword and that are generated during the observation window. We also ensure that the selected topic is associated with a large enough volume of activity. Experiments ::: Datasets In this section we detail the discussions we use to test our metric and how we determine the ground truth (i.e. if the discussion is controversial or not). We use thirty different discussions that took place between March 2015 and June 2019, half of them with controversy and half without it. We considered discussions in four different languages: English, Portuguese, Spanish and French, occurring in five regions over the world: South and North America, Western Europe, Central and Southern Asia. We also studied these discussions taking first 140 characters and then 280 from each tweet to analyze the difference in performance and computing time wrt the length of the posts. To define the amount of data needed to run our method we established that the Fasttext model has to predict at least one user of each community with a probability greater or equal than 0.9 during ten different trainings. If that is not the case, we are not able to use DPC method. This decision made us consider only a subset of the datasets used in BIBREF23, because due to the time elapsed since their work, many tweets had been deleted and consequently the volume of the data was not enough for our framework. To enlarge our experiment base we added new debates, more detailed information about each one is shown in Table TABREF24 in UNKREF6. To select new discussions and to determine if they are controversial or not we looked for topics widely covered by mainstream media, and that have generated ample discussion, both online and offline. For non-controversy discussions we focused on “soft news" and entertainment, but also to events that, while being impactful and/or dramatic, did not generate large controversies. To validate that intuition, we manually checked a sample of tweets, being unable to identify any clear instance of controversy On the other side, for controversial debates we focused on political events such as elections, corruption cases or justice decisions. To furtherly establish the presence of absence of controversy of our datasets, we visualized the corresponding networks through ForceAtlas2 BIBREF29. Figures FIGREF9 and FIGREF9 show an example of how non-controversial and controversial discussions look like respectively with ForceAtlas2 layout. As we can see in these figures, in a controversial discussion this layout tends to show two well separated groups while in a non-controversial one it tends to be only one big group. More information on the discussions is given in Table TABREF24. To avoid potential overfitting, we use only twelve graphs as testbed during the development of the measures, half of them controversial (netanyahu, ukraine, @mauriciomacri 1-11 Jan, Kavanaugh 3 Oct, @mauriciomacri 11-18 Mar, Bolsonaro 27 Oct) and half non-controversial (sxsw, germanwings, onedirection, ultralive, nepal, mothersday). This procedure resembles a 40/60% train/test split in traditional machine learning applications. Some of the discussions we consider refer to the same topics but in different periods of time. We needed to split them because our computing infrastructure does not allow us to compute such an enormous amount of data. However, being able to estimate controversy with only a subset of the discussion is an advantage, because discussions could take many days or months and we want to identify controversy as soon as possible, without the need of downloading the whole discussion. Moreover, for very long lasting discussions in social networks gathering the whole data would be impractical for any method. Experiments ::: Results Training a Fasttext model is not a deterministic process, as different runs could yield different results even using the same training set in each one. To analyze if these differences are significant, we decide to compute 20 scores for each discussion. The standard deviations among these 20 scores were low in all cases, with mean 0.01 and maximum 0.05. Consequently, we decided to report in this paper the average between the 20 scores, in practice taking the average between 5 runs would be enough. Figure FIGREF18 reports the scores computed by our measure in each topic for the two cluster methods. The beanplot shows the estimated probability density function for a measure computed on the topics, the individual observations are shown as small white lines in a one-dimensional scatter plot, and the median as a longer black line. The beanplot is divided into two groups, one for controversial topics (left/dark) and one for non-controversial ones (right/light). Hence, the black group shows the score distribution over controversial discussions and the white group over non-controversial ones. A larger separation of the two distributions indicates that the measure is better at capturing the characteristics of controversial topics, because a good separation allows to establish a threshold in the score that separates controversial and non-controversial discussions. As we may see in the figure, the medians are well separated in both cases, with little overlapping. To better quantify this overlap we measure the sensitivity BIBREF42 of these predictions by measuring the area under the ROC curve (AUC ROC), obtaining a value of 0.98 for Walktrap clustering and 0.967 for Louvain (where 1 represents a perfect separation and 0.5 means that they are indistinguishable). As Garimella et al. BIBREF23 have made their code public , we reproduced their best method Randomwalk on our datasets and measured the AUC ROC, obtaining a score of 0.935. An interesting finding was that their method had a poor performance over their own datasets. This was due to the fact (already explained in Section SECREF4) that it was not possible to retrieve the complete discussions, moreover, in no case could we restore more than 50% of the tweets. So we decided to remove these discussions and measure again the AUC ROC of this method, obtaining a 0.99 value. Our hypothesis is that the performance of that method was seriously hurt by the incompleteness of the data. We also tested our method on these datasets, obtaining a 0.99 AUC ROC with Walktrap and 0.989 with Louvain clustering. We conclude that our method works better, as in practice both approaches show same performances -specially with Walktrap, but in presence of incomplete information our measure is more robust. The performance of Louvain is slightly worse but, as we mentioned in Section SECREF3, this method is much faster. Therefore, we decided to compare the running time of our method with both clustering techniques and also with the Randomwalk algorithm. In figure FIGREF18 we can see the distribution of running times of all techniques through box plots. Both versions of our method are faster than Randomwalk, while Louvain is faster than Walktrap. We now analyze the impact of the length of the considered text in our method. Figure FIGREF18 depicts the results of similar experiment as Figure FIGREF18, but considering only 140 characters per tweet. As we may see, here the overlapping is bigger, having an AUC of 0.88. As for the impact on computing time, we observe that despite of the results of BIBREF34 that reported a complexity of O(h $log_{2}$(k)) at training and test tasks, in practice we observed a linear growth. We measured the running times of the training and predicting phases (the two text-related phases of our method), the resulting times are reported in figure FIGREF18, which shows running time as a function of the text-size. We include also the best estimated function that approximate computing time as a function of text-set size. As it may be seen, time grows almost linearly, ranging from 30 seconds for a set of 111 KB to 84 seconds for a set of 11941 KB. Finally, we measured running times for the whole method over each dataset with 280 characters. Times were between 170 and 2467 seconds with a mean of 842, making it in practice a reasonable amount of time. Discussions The task we address in this work is certainly not an easy one, and our study has some limitations, which we discuss in this section. Our work lead us to some conclusions regarding the overall possibility of measuring controversy through text, and what aspects need to be considered to deepen our work. Discussions ::: Limitations As our approach to controversy is similar to that of Garimella et al. BIBREF23, we share some of their limitations with respect to several aspects: Evaluation -difficulties to establish ground-truth, Multisided controversies -controversy with more than two sides, Choice of data - manually pick topics, and Overfitting - small set of experiments. Although we have more discussions, it is still small set from statistical point of view. Apart from that, our language-based approach has other limitations which we mention in the following, together with their solutions or mitigation. Data-size. Training an NLP model that can predict tags with a probability greater or equal than 0.9 requires significant amount of text, therefore our method works only for “big" discussions. Most interesting controversies are those that have consequence at a society level, in general big enough for our method. Multi-language discussions. When multiple languages are participating in a discussion it is common that users tend to retweet more tweets in their own language, creating sub-communities. In this cases our model will tend to predict higher controversy scores. This is the case for example of #germanwings, where users tweet in English, German and Spanish and it has the highest score in no-controversial topics. However, the polarization that we tackle in this work is normally part of a society cell (a nation, a city, etc.), and thus developed in just one language. We think that limiting the effectiveness of our analysis to single-language discussions is not a serious limitation. Twitter only. Our findings are based on datasets coming from Twitter. While this is certainly a limitation, Twitter is one of the main venues for online public discussion, and one of the few for which data is available. Hence, Twitter is a natural choice. However, Twitter's characteristic limit of 280 characters per message (140 till short time ago) is an intrinsic limitation of that network. We think that in other social networks as Facebook or Reddit our method will work even better, as having more text per user could redound on a better NLP model as we verified comparing the results with 140 and 280 characters per post. Discussions ::: Conclusions In this article, we introduced the first large-scale systematic method for quantifying controversy in social media through content. We have shown that this method works on Spanish, English, French and Portuguese, it is context-agnostic and does not require the intervention of a domain expert. We have compared its performance with state-of-the-art structure-based controversy measures showing that they have same performance and it is more robust. We also have shown that more text implies better performance and without significantly increasing computing time, therefore, it could be used in other contexts such as other social networks like Reddit or Facebook and we are going to test it in future works. Training the model is not an expensive task since Fasttext has a good performance at this. However, the best performance for detecting principal communities is obtained by Walktrap. The complexity of that algorithm is O(m$n^2$)BIBREF38, where $m$ and $n$ are the number of edges and vertices respectively. This makes this method rather expensive to compute on big networks. Nevertheless, we have shown that with Louvain the method still obtains a very similar AUC ROC (0.99 with Walktrap and 0.989 with Louvain). With incomplete information its performance gets worse but it is still good (0.96) and better than previous state of the art. This work opens several avenues for future research. One is identifying what words, semantics/concepts or language expressions make differ one community from the other. There are various ways to do this, for instance through the word-embbedings that Fasttext returns after training BIBREF34. Also we could use interpretability techniques on machine learning models BIBREF43. Finally, we could try other techniques for measuring controversy through text, using another NLP model as pre-trained neural network BERT BIBREF44 or, in a completely different approach measuring the dispersion index of the discussions word-embbedings BIBREF25. We are currently starting to follow this direction. Details on the discussions F

Twitter

d92f1c15537b33b32bfc436e6d017ae7d9d6c29a

d92f1c15537b33b32bfc436e6d017ae7d9d6c29a_0

Q: How many languages do they experiment with? Text: Introduction Controversy is a phenomenom with a high impact at various levels. It has been broadly studied from the perspective of different disciplines, ranging from the seminal analysis of the conflicts within the members of a karate club BIBREF0 to political issues in modern times BIBREF1, BIBREF2. The irruption of digital social networks BIBREF3 gave raise to new ways of intentionally intervening on them for taking some advantage BIBREF4, BIBREF5. Moreover highly contrasting points of view in some groups tend to provoke conflicts that lead to attacks from one community to the other by harassing, “brigading”, or “trolling” it BIBREF6. The existing literature shows different issues that controversy brings up such as splitting of communities, biased information, hateful discussions and attacks between groups, generally proposing ways to solve them. For example, Kumar, Srijan, et al. BIBREF6 analyze many techniques to defend us from attacks in Reddit while Stewart, et al. BIBREF4 insinuate that there was external interference in Twitter during the 2016 US presidential elections to benefit one candidate. Also, as shown in BIBREF7, detecting controversy could provide the basis to improve the “news diet" of readers, offering the possibility to connect users with different points of views by recommending them new content to read BIBREF8. Moreover, other studies on “bridging echo chambers” BIBREF9 and the positive effects of intergroup dialogue BIBREF10, BIBREF11 suggest that direct engagement could be effective for mitigating such conflicts. Therefore, easily and automatically identifying controversial topics could allow us to quickly implement different strategies for preventing miss-information, fights and bias. Quantifying the controversy is even more powerful, as it allows us to establish controversy levels, and in particular to classify controversial and non-controversial topics by establishing a threshold score that separates the two types of topics. With this aim, we propose in this work a systematic, language-agnostic method to quantify controversy on social networks taking tweet's content as root input. Our main contribution is a new vocabulary-based method that works in any language and equates the performance of state-of-the-art structure-based methods. Finally, controversy quantification through vocabulary analysis opens several research avenues to analyze whether polarization is being created, maintained or augmented by the ways of talking of each community. Having this in mind and if we draw from the premise that when a discussion has a high controversy it is in general due to the presence of two principal communities fighting each other (or, conversely, that when there is no controversy there is just one principal community the members of which share a common point of view), we can measure the controversy by detecting if the discussion has one or two principal jargons in use. Our method is tested on Twitter datasets. This microblogging platform has been widely used to analyze discussions and polarization BIBREF12, BIBREF13, BIBREF14, BIBREF15, BIBREF2. It is a natural choice for these kind of problems, as it represents one of the main fora for public debate in online social media BIBREF15, it is a common destination for affiliative expressions BIBREF16 and is often used to report and read news about current events BIBREF17. An extra advantage of Twitter for this kind of studies is the availability of real-time data generated by millions of users. Other social media platforms offer similar data-sharing services, but few can match the amount of data and the accompanied documentation provided by Twitter. One last asset of Twitter for our work is given by retweets, whom typically indicate endorsement BIBREF18 and hence become a useful concept to model discussions as we can set “who is with who". However, our method has a general approach and it could be used a priori in any social network. In this work we report excellent result tested on Twitter but in future work we are going to test it in other social networks. Our paper is organized as follows: in Section SECREF2, we review related work. Section SECREF3 contains the detailed explanation of the pipeline we use for quantifying controversy of a topic, and each of its stages. In Section SECREF4 we report the results of an extensive empirical evaluation of the proposed measure of controversy. Finally, Section SECREF5 is devoted to discuss possible improvements and directions for future work, as well as lessons learned. Related work Many previous works are dedicated to quantifying the polarization observed in online social networks and social media BIBREF1, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23. The main characteristic of those works is that the measures proposed are based on the structural characteristics of the underlying graph. Among them, we highlight the work of Garimella et al.BIBREF23 that presents an extensive comparison of controversy measures, different graph-building approaches, and data sources, achieving the best performance of all. In their research they propose different metrics to measure polarization on Twitter. Their techniques based on the structure of the endorsement graph can successfully detect whether a discussion (represented by a set of tweets), is controversial or not regardless of the context and most importantly, without the need of any domain expertise. They also consider two different methods to measure controversy based on the analysis of the posts contents, but both fail when used to create a measure of controversy. Matakos et al. BIBREF24 develop a polarization index. Their measure captures the tendency of opinions to concentrate in network communities, creating echo-chambers. They obtain a good performance at identifying controversy by taking into account both the network structure and the existing opinions of users. However, they model opinions as positive or negative with a real number between -1 and 1. Their performance is good, but although it is an opinion-based method it is not a text-related one.Other recent works BIBREF25, BIBREF26, BIBREF27 have shown that communities may express themselves with different terms or ways of speaking, use different jargon, which in turn can be detected with the use of text-related techniques. In his thesis BIBREF28, Jang explains controversy via generating a summary of two conflicting stances that make up the controversy. This work shows that a specific sub-set of tweets could represent the two opposite positions in a polarized debate. A good tool to see how communities interact is ForceAtlas2 BIBREF29, a force-directed layout widely used for visualization. This layout has been recently found to be very useful at visualizing community interactions BIBREF30, as this algorithm will draw groups with little communication between them in different areas, whereas, if they have many interactions they will be drawn closer to each other. Therefore, whenever there is controversy the layout will show two well separated groups and will tend to show only one big community otherwise. The method we propose to measure the controversy equates in accuracy the one developed by Garimella et al.BIBREF23 and improves considerably computing time and robustness wrt the amount of data needed to effectively apply it. Our method is also based on a graph approach but it has its main focus on the vocabulary. We first train an NLP classifier that estimates opinion polarity of main users, then we run label-propagation BIBREF31 on the endorsement graph to get polarity of the whole network. Finally we compute the controversy score through a computation inspired in Dipole Moment, a measure used in physics to estimate electric polarity on a system. In our experiments we use the same data-sets from other works BIBREF32, BIBREF23, BIBREF33 as well as other datasets that we collected by us using a similar criterion (described in Section SECREF4). Method Our approach to measuring controversy is based on a systematic way of characterizing social media activity through its content. We employ a pipeline with five stages, namely graph building, community identification, model training, predicting and controversy measure. The final output of the pipeline is a value that measures how controversial a topic is, with higher values corresponding to higher degrees of controversy. The method is based on analysing posts content through Fasttext BIBREF34, a library for efficient learning of word representations and sentence classification developed by Facebook Research team. In short, our method works as follows: through Fasttext we train a language-agnostic model which can predict the community of many users by their jargon. Then we take there predictions and compute a score based on the physic notion Dipole Moment using a language approach to identify core or characteristic users and set the polarity trough them. We provide a detailed description of each stage in the following. Graph Building This paragraph provides details about the approach used to build graphs from raw data. As we said in Section SECREF1, we extract our discussions from Twitter. Our purpose is to build a conversation graph that represents activity related to a single topic of discussion -a debate about a specific event. For each topic, we build a graph $G$ where we assign a vertex to each user who contributes to it and we add a directed edge from node $u$ to node $v$ whenever user $u$ retweets a tweet posted by $v$. Retweets typically indicate endorsement BIBREF18: users who retweet signal endorsement of the opinion expressed in the original tweet by propagating it further. Retweets are not constrained to occur only between users who are connected in Twitter's social network, but users are allowed to retweet posts generated by any other user. As many other works in literature BIBREF5, BIBREF35, BIBREF36, BIBREF37, BIBREF4, BIBREF2 we establish that one retweet among a pair of users are needed to define an edge between them. Community Identification To identify a community's jargon we need to be very accurate at defining its members. If we, in our will of finding two principal communities, force the partition of the graph in that precise number of communities, we may be adding noise in the jargon of the principal communities that are fighting each other. Because of that, we decide to cluster the graph trying two popular algorithms: Walktrap BIBREF38 and Louvain BIBREF39. Both are structure-based algorithms that have very good performance with respect to the Modularity Q measure. These techniques does not detect a fixed number of clusters; their output will depend on the Modularity Q optimization, resulting in less “noisy" communities. The main differences between the two methods, in what regards our work, are that Louvain is a much faster heuristic algorithm but produces clusters with worse Modularity Q. Therefore, in order to analyze the trade-off between computing time and quality we decide to test both methods. At this step we want to capture the tweets of the principal communities to create the model that could differentiate them. Therefore, we take the two communities identified by the cluster algorithm that have the maximum number of users, and use them for the following step of our method. Model Training After detecting the principal communities we create our training dataset to feed the model. To do that, we extract the tweets of each cluster, we sanitize and we subject them to some transformations. First, we remove duplicate tweets -e.g. retweets without additional text. Second, we remove from the text of the tweets user names, links, punctuation, tabs, leading and lagging blanks, general spaces and “RT" - the text that points that a tweet is in fact a retweet. As shown in previous works, emojis are correlated with sentiment BIBREF40. Moreover, as we think that communities will express different sentiment during discussion, it is forseeable that emojis will play an important role as separators of tweets that differentiate between the two sides. Accordingly, we decide to add them to the train-set by translating each emoji into a different word. For example, the emoji :) will be translated into happy and :( into sad. Relations between emojis and words are defined in the R library textclean. Finally, we group tweets by user concatenating them in one string and labeling them with the user's community, namely with tags C1 and C2, corresponding respectively to the biggest and second biggest groups. It is important to note that we take the same number of users of each community to prevent bias in the model. Thus, we use the number of users of the smallest principal community. The train-set built that way is used to feed the model. As we said, we use Fasttext BIBREF34 to do this training. To define the values of the hyper-parameters we use the findings of BIBREF41. In their work they investigate the best hyper-parameters to train word embedding models using Fasttext BIBREF34 and Twitter data. We also change the default value of the hyper-parameter epoch to 20 instead of 5 because we want more convergence preventing as much as possible the variance between different training. These values could change in other context or social networks where we have more text per user or different discussion dynamics. Predicting The next stage consists of identifying the characteristic users of each side the discussion. These are the users that better represent the jargon of each side. To do that, tweets of the users belonging to the largest connected component of the graph are sanitized and transformed exactly as in the Training step. We decide to restrict to the largest connected component because in all cases it contains more than 90% of the nodes. The remaining 10% of the users don't participate in the discussion from a collective point of view but rather in an isolated way and this kind of intervention does not add interesting information to our approach. Then, we remove from this component users with degree smaller or equal to 2 (i.e. users that were retweeted by another user or retweeted other person less than three times in total). Their participation in the discussion is marginal, consequently they are not relevant wrt controversy as they add more noise than information at measuring time. This step could be adjusted differently in a different social network. We name this result component root-graph. Finally, let's see how we do classification. Considering that Fasttext returns for each classification both the predicted tag and the probability of the prediction, we classify each user of the resulting component by his sanitized tweets with our trained model, and take users that were tagged with a probability greater or equal than 0.9. These are the characteristic users that will be used in next step to compute the controversy measure. Controversy Measure This section describes the controversy measures used in this work. This computation is inspired in the measure presented by Morales et al. BIBREF2, and is based on the notion of dipole moment that has its origin in physics. First, we assign to the characteristic users the probability returned by the model, negativizing them if the predicted tag was C2. Therefore, these users are assigned values in the set [-1,-0.9] $\cup $ [0.9,1]. Then, we set values for the rest of the users of the root-graph by label-propagation BIBREF31 - an iterative algorithm to propagate values through a graph by node's neighborhood. Let $n^{+}$ and $n^{-}$ be the number of vertices $V$ with positive and negative values, respectively, and $\Delta A = \dfrac{\mid n^{+} - n^{-}\mid }{\mid V \mid }$ the absolute difference of their normalized size. Moreover, let $gc^{+}$ ($gc^{-}$) be the average value among vertices $n^{+}$ ($n^{-}$) and set $\tau $ as half their absolute difference, $\tau = \dfrac{\mid gc^{+} - gc^{- }\mid }{2}$. The dipole moment content controversy measure is defined as: $\textit {DMC} = (1 -\Delta A)\tau $. The rationale for this measure is that if the two sides are well separated, then label propagation will assign different extreme values to the two partitions, where users from one community will have values near to 1 and users from the other to -1, leading to higher values of the DMC measure. Note also that larger differences in the size of the two partitions (reflected in the value of $\Delta A$) lead to smaller values for the measure, which takes values between zero and one. Experiments In this section we report the results obtained by running the above proposed method over different discussions. Experiments ::: Topic definition In the literature, a topic is often defined by a single hashtag. However, this might be too restrictive in many cases. In our approach, a topic is operationalized as an specific hashtags or key words. Sometimes a discussion in a particular moment could not have a defined hashtag but it could be around a certain keyword, i.e. a word or expression that is not specifically a hashtag but it is widely used in the topic. For example during the Brazilian presidential elections in 2018 we captured the discussion by the mentions to the word Bolsonaro, that is the principal candidate's surname. Thus, for each topic we retrieve all the tweets that contain one of its hashtags or the keyword and that are generated during the observation window. We also ensure that the selected topic is associated with a large enough volume of activity. Experiments ::: Datasets In this section we detail the discussions we use to test our metric and how we determine the ground truth (i.e. if the discussion is controversial or not). We use thirty different discussions that took place between March 2015 and June 2019, half of them with controversy and half without it. We considered discussions in four different languages: English, Portuguese, Spanish and French, occurring in five regions over the world: South and North America, Western Europe, Central and Southern Asia. We also studied these discussions taking first 140 characters and then 280 from each tweet to analyze the difference in performance and computing time wrt the length of the posts. To define the amount of data needed to run our method we established that the Fasttext model has to predict at least one user of each community with a probability greater or equal than 0.9 during ten different trainings. If that is not the case, we are not able to use DPC method. This decision made us consider only a subset of the datasets used in BIBREF23, because due to the time elapsed since their work, many tweets had been deleted and consequently the volume of the data was not enough for our framework. To enlarge our experiment base we added new debates, more detailed information about each one is shown in Table TABREF24 in UNKREF6. To select new discussions and to determine if they are controversial or not we looked for topics widely covered by mainstream media, and that have generated ample discussion, both online and offline. For non-controversy discussions we focused on “soft news" and entertainment, but also to events that, while being impactful and/or dramatic, did not generate large controversies. To validate that intuition, we manually checked a sample of tweets, being unable to identify any clear instance of controversy On the other side, for controversial debates we focused on political events such as elections, corruption cases or justice decisions. To furtherly establish the presence of absence of controversy of our datasets, we visualized the corresponding networks through ForceAtlas2 BIBREF29. Figures FIGREF9 and FIGREF9 show an example of how non-controversial and controversial discussions look like respectively with ForceAtlas2 layout. As we can see in these figures, in a controversial discussion this layout tends to show two well separated groups while in a non-controversial one it tends to be only one big group. More information on the discussions is given in Table TABREF24. To avoid potential overfitting, we use only twelve graphs as testbed during the development of the measures, half of them controversial (netanyahu, ukraine, @mauriciomacri 1-11 Jan, Kavanaugh 3 Oct, @mauriciomacri 11-18 Mar, Bolsonaro 27 Oct) and half non-controversial (sxsw, germanwings, onedirection, ultralive, nepal, mothersday). This procedure resembles a 40/60% train/test split in traditional machine learning applications. Some of the discussions we consider refer to the same topics but in different periods of time. We needed to split them because our computing infrastructure does not allow us to compute such an enormous amount of data. However, being able to estimate controversy with only a subset of the discussion is an advantage, because discussions could take many days or months and we want to identify controversy as soon as possible, without the need of downloading the whole discussion. Moreover, for very long lasting discussions in social networks gathering the whole data would be impractical for any method. Experiments ::: Results Training a Fasttext model is not a deterministic process, as different runs could yield different results even using the same training set in each one. To analyze if these differences are significant, we decide to compute 20 scores for each discussion. The standard deviations among these 20 scores were low in all cases, with mean 0.01 and maximum 0.05. Consequently, we decided to report in this paper the average between the 20 scores, in practice taking the average between 5 runs would be enough. Figure FIGREF18 reports the scores computed by our measure in each topic for the two cluster methods. The beanplot shows the estimated probability density function for a measure computed on the topics, the individual observations are shown as small white lines in a one-dimensional scatter plot, and the median as a longer black line. The beanplot is divided into two groups, one for controversial topics (left/dark) and one for non-controversial ones (right/light). Hence, the black group shows the score distribution over controversial discussions and the white group over non-controversial ones. A larger separation of the two distributions indicates that the measure is better at capturing the characteristics of controversial topics, because a good separation allows to establish a threshold in the score that separates controversial and non-controversial discussions. As we may see in the figure, the medians are well separated in both cases, with little overlapping. To better quantify this overlap we measure the sensitivity BIBREF42 of these predictions by measuring the area under the ROC curve (AUC ROC), obtaining a value of 0.98 for Walktrap clustering and 0.967 for Louvain (where 1 represents a perfect separation and 0.5 means that they are indistinguishable). As Garimella et al. BIBREF23 have made their code public , we reproduced their best method Randomwalk on our datasets and measured the AUC ROC, obtaining a score of 0.935. An interesting finding was that their method had a poor performance over their own datasets. This was due to the fact (already explained in Section SECREF4) that it was not possible to retrieve the complete discussions, moreover, in no case could we restore more than 50% of the tweets. So we decided to remove these discussions and measure again the AUC ROC of this method, obtaining a 0.99 value. Our hypothesis is that the performance of that method was seriously hurt by the incompleteness of the data. We also tested our method on these datasets, obtaining a 0.99 AUC ROC with Walktrap and 0.989 with Louvain clustering. We conclude that our method works better, as in practice both approaches show same performances -specially with Walktrap, but in presence of incomplete information our measure is more robust. The performance of Louvain is slightly worse but, as we mentioned in Section SECREF3, this method is much faster. Therefore, we decided to compare the running time of our method with both clustering techniques and also with the Randomwalk algorithm. In figure FIGREF18 we can see the distribution of running times of all techniques through box plots. Both versions of our method are faster than Randomwalk, while Louvain is faster than Walktrap. We now analyze the impact of the length of the considered text in our method. Figure FIGREF18 depicts the results of similar experiment as Figure FIGREF18, but considering only 140 characters per tweet. As we may see, here the overlapping is bigger, having an AUC of 0.88. As for the impact on computing time, we observe that despite of the results of BIBREF34 that reported a complexity of O(h $log_{2}$(k)) at training and test tasks, in practice we observed a linear growth. We measured the running times of the training and predicting phases (the two text-related phases of our method), the resulting times are reported in figure FIGREF18, which shows running time as a function of the text-size. We include also the best estimated function that approximate computing time as a function of text-set size. As it may be seen, time grows almost linearly, ranging from 30 seconds for a set of 111 KB to 84 seconds for a set of 11941 KB. Finally, we measured running times for the whole method over each dataset with 280 characters. Times were between 170 and 2467 seconds with a mean of 842, making it in practice a reasonable amount of time. Discussions The task we address in this work is certainly not an easy one, and our study has some limitations, which we discuss in this section. Our work lead us to some conclusions regarding the overall possibility of measuring controversy through text, and what aspects need to be considered to deepen our work. Discussions ::: Limitations As our approach to controversy is similar to that of Garimella et al. BIBREF23, we share some of their limitations with respect to several aspects: Evaluation -difficulties to establish ground-truth, Multisided controversies -controversy with more than two sides, Choice of data - manually pick topics, and Overfitting - small set of experiments. Although we have more discussions, it is still small set from statistical point of view. Apart from that, our language-based approach has other limitations which we mention in the following, together with their solutions or mitigation. Data-size. Training an NLP model that can predict tags with a probability greater or equal than 0.9 requires significant amount of text, therefore our method works only for “big" discussions. Most interesting controversies are those that have consequence at a society level, in general big enough for our method. Multi-language discussions. When multiple languages are participating in a discussion it is common that users tend to retweet more tweets in their own language, creating sub-communities. In this cases our model will tend to predict higher controversy scores. This is the case for example of #germanwings, where users tweet in English, German and Spanish and it has the highest score in no-controversial topics. However, the polarization that we tackle in this work is normally part of a society cell (a nation, a city, etc.), and thus developed in just one language. We think that limiting the effectiveness of our analysis to single-language discussions is not a serious limitation. Twitter only. Our findings are based on datasets coming from Twitter. While this is certainly a limitation, Twitter is one of the main venues for online public discussion, and one of the few for which data is available. Hence, Twitter is a natural choice. However, Twitter's characteristic limit of 280 characters per message (140 till short time ago) is an intrinsic limitation of that network. We think that in other social networks as Facebook or Reddit our method will work even better, as having more text per user could redound on a better NLP model as we verified comparing the results with 140 and 280 characters per post. Discussions ::: Conclusions In this article, we introduced the first large-scale systematic method for quantifying controversy in social media through content. We have shown that this method works on Spanish, English, French and Portuguese, it is context-agnostic and does not require the intervention of a domain expert. We have compared its performance with state-of-the-art structure-based controversy measures showing that they have same performance and it is more robust. We also have shown that more text implies better performance and without significantly increasing computing time, therefore, it could be used in other contexts such as other social networks like Reddit or Facebook and we are going to test it in future works. Training the model is not an expensive task since Fasttext has a good performance at this. However, the best performance for detecting principal communities is obtained by Walktrap. The complexity of that algorithm is O(m$n^2$)BIBREF38, where $m$ and $n$ are the number of edges and vertices respectively. This makes this method rather expensive to compute on big networks. Nevertheless, we have shown that with Louvain the method still obtains a very similar AUC ROC (0.99 with Walktrap and 0.989 with Louvain). With incomplete information its performance gets worse but it is still good (0.96) and better than previous state of the art. This work opens several avenues for future research. One is identifying what words, semantics/concepts or language expressions make differ one community from the other. There are various ways to do this, for instance through the word-embbedings that Fasttext returns after training BIBREF34. Also we could use interpretability techniques on machine learning models BIBREF43. Finally, we could try other techniques for measuring controversy through text, using another NLP model as pre-trained neural network BERT BIBREF44 or, in a completely different approach measuring the dispersion index of the discussions word-embbedings BIBREF25. We are currently starting to follow this direction. Details on the discussions F

four different languages: English, Portuguese, Spanish and French

fa3663567c48c27703e09c42930e51bacfa54905

fa3663567c48c27703e09c42930e51bacfa54905_0

Q: What is the current SOTA for sentiment analysis on Twitter at the time of writing? Text: Synonyms Microblog sentiment analysis; Twitter opinion mining Glossary Sentiment Analysis: This is text analysis aiming to determine the attitude of a speaker or a writer with respect to some topic or the overall contextual polarity of a piece of text. Definition Sentiment analysis on Twitter is the use of natural language processing techniques to identify and categorize opinions expressed in a tweet, in order to determine the author's attitude toward a particular topic or in general. Typically, discrete labels such as positive, negative, neutral, and objective are used for this purpose, but it is also possible to use labels on an ordinal scale, or even continuous numerical values. Introduction Internet and the proliferation of smart mobile devices have changed the way information is created, shared, and spreads, e.g., microblogs such as Twitter, weblogs such as LiveJournal, social networks such as Facebook, and instant messengers such as Skype and WhatsApp are now commonly used to share thoughts and opinions about anything in the surrounding world. This has resulted in the proliferation of social media content, thus creating new opportunities to study public opinion at a scale that was never possible before. Naturally, this abundance of data has quickly attracted business and research interest from various fields including marketing, political science, and social studies, among many others, which are interested in questions like these: Do people like the new Apple Watch? What do they hate about iPhone6? Do Americans support ObamaCare? What do Europeans think of Pope's visit to Palestine? How do we recognize the emergence of health problems such as depression? Do Germans like how Angela Merkel is handling the refugee crisis in Europe? What do republican voters in USA like/hate about Donald Trump? How do Scottish feel about the Brexit? Answering these questions requires studying the sentiment of opinions people express in social media, which has given rise to the fast growth of the field of sentiment analysis in social media, with Twitter being especially popular for research due to its scale, representativeness, variety of topics discussed, as well as ease of public access to its messages BIBREF0 , BIBREF1 . Despite all these opportunities, the rise of social media has also presented new challenges for natural language processing (NLP) applications, which had largely relied on NLP tools tuned for formal text genres such as newswire, and thus were not readily applicable to the informal language and style of social media. That language proved to be quite challenging with its use of creative spelling and punctuation, misspellings, slang, new words, URLs, and genre-specific terminology and abbreviations, e.g., RT for re-tweet and #hashtags. In addition to the genre difference, there is also a difference in length: social media messages are generally short, often length-limited by design as in Twitter, i.e., a sentence or a headline rather than a full document. How to handle such challenges has only recently been the subject of thorough research BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 . Key Points Sentiment analysis has a wide number of applications in areas such as market research, political and social sciences, and for studying public opinion in general, and Twitter is one of the most commonly-used platforms for this. This is due to its streaming nature, which allows for real-time analysis, to its social aspect, which encourages people to share opinions, and to the short size of the tweets, which simplifies linguistic analysis. There are several formulations of the task of Sentiment Analysis on Twitter that look at different sizes of the target (e.g., at the level of words vs. phrases vs. tweets vs. sets of tweets), at different types of semantic targets (e.g., aspect vs. topic vs. overall tweet), at the explicitness of the target (e.g., sentiment vs. stance detection), at the scale of the expected label (2-point vs. 3-point vs. ordinal), etc. All these are explored at SemEval, the International Workshop on Semantic Evaluation, which has created a number of benchmark datasets and has enabled direct comparison between different systems and approaches, both as part of the competition and beyond. Traditionally, the task has been addressed using supervised and semi-supervised methods, as well as using distant supervision, with the most important resource being sentiment polarity lexicons, and with feature-rich approaches as the dominant research direction for years. With the recent rise of deep learning, which in many cases eliminates the need for any explicit feature modeling, the importance of both lexicons and features diminishes, while at the same time attention is shifting towards learning from large unlabeled data, which is needed to train the high number of parameters of such complex models. Finally, as methods for sentiment analysis mature, more attention is also being paid to linguistic structure and to multi-linguality and cross-linguality. Historical Background Sentiment analysis emerged as a popular research direction in the early 2000s. Initially, it was regarded as standard document classification into topics such as business, sport, and politics BIBREF10 . However, researchers soon realized that it was quite different from standard document classification BIBREF11 , and that it crucially needed external knowledge in the form of sentiment polarity lexicons. Around the same time, other researchers realized the importance of external sentiment lexicons, e.g., Turney BIBREF12 proposed an unsupervised approach to learn the sentiment orientation of words/phrases: positive vs. negative. Later work studied the linguistic aspects of expressing opinions, evaluations, and speculations BIBREF13 , the role of context in determining the sentiment orientation BIBREF14 , of deeper linguistic processing such as negation handling BIBREF15 , of finer-grained sentiment distinctions BIBREF16 , of positional information BIBREF17 , etc. Moreover, it was recognized that in many cases, it is crucial to know not just the polarity of the sentiment but also the topic toward which this sentiment is expressed BIBREF18 . Until the rise of social media, research on opinion mining and sentiment analysis had focused primarily on learning about the language of sentiment in general, meaning that it was either genre-agnostic BIBREF19 or focused on newswire texts BIBREF20 and customer reviews (e.g., from web forums), most notably about movies BIBREF10 and restaurants BIBREF21 but also about hotels, digital cameras, cell phones, MP3 and DVD players BIBREF22 , laptops BIBREF21 , etc. This has given rise to several resources, mostly word and phrase polarity lexicons, which have proven to be very valuable for their respective domains and types of texts, but less useful for short social media messages. Later, with the emergence of social media, sentiment analysis in Twitter became a hot research topic. Unfortunately, research in that direction was hindered by the unavailability of suitable datasets and lexicons for system training, development, and testing. While some Twitter-specific resources were developed, initially they were either small and proprietary, such as the i-sieve corpus BIBREF6 , were created only for Spanish like the TASS corpus BIBREF23 , or relied on noisy labels obtained automatically, e.g., based on emoticons and hashtags BIBREF24 , BIBREF25 , BIBREF10 . This situation changed with the shared task on Sentiment Analysis on Twitter, which was organized at SemEval, the International Workshop on Semantic Evaluation, a semantic evaluation forum previously known as SensEval. The task ran in 2013, 2014, 2015, and 2016, attracting over 40 participating teams in all four editions. While the focus was on general tweets, the task also featured out-of-domain testing on SMS messages, LiveJournal messages, as well as on sarcastic tweets. SemEval-2013 Task 2 BIBREF26 and SemEval-2014 Task 9 BIBREF27 focused on expression-level and message-level polarity. SemEval-2015 Task 10 BIBREF28 , BIBREF29 featured topic-based message polarity classification on detecting trends toward a topic and on determining the out-of-context (a priori) strength of association of Twitter terms with positive sentiment. SemEval-2016 Task 4 BIBREF30 introduced a 5-point scale, which is used for human review ratings on popular websites such as Amazon, TripAdvisor, Yelp, etc.; from a research perspective, this meant moving from classification to ordinal regression. Moreover, it focused on quantification, i.e., determining what proportion of a set of tweets on a given topic are positive/negative about it. It also featured a 5-point scale ordinal quantification subtask BIBREF31 . Other related tasks have explored aspect-based sentiment analysis BIBREF32 , BIBREF33 , BIBREF21 , sentiment analysis of figurative language on Twitter BIBREF34 , implicit event polarity BIBREF35 , stance in tweets BIBREF36 , out-of-context sentiment intensity of phrases BIBREF37 , and emotion detection BIBREF38 . Some of these tasks featured languages other than English. Variants of the Task at SemEval Tweet-level sentiment. The simplest and also the most popular task of sentiment analysis on Twitter is to determine the overall sentiment expressed by the author of a tweet BIBREF30 , BIBREF28 , BIBREF26 , BIBREF29 , BIBREF27 . Typically, this means choosing one of the following three classes to describe the sentiment: Positive, Negative, and Neutral. Here are some examples: Positive: @nokia lumia620 cute and small and pocket-size, and available in the brigh test colours of day! #lumiacaption Negative: I hate tweeting on my iPhone 5 it's so small :( Neutral: If you work as a security in a samsung store...Does that make you guardian of the galaxy?? Sentiment polarity lexicons. Naturally, the overall sentiment in a tweet can be determined based on the sentiment-bearing words and phrases it contains as well as based on emoticons such as ;) and:(. For this purpose, researchers have been using lexicons of sentiment-bearing words. For example, cute is a positive word, while hate is a negative one, and the occurrence of these words in (1) and (2) can help determine the overall polarity of the respective tweet. We will discuss these lexicons in more detail below. Prior sentiment polarity of multi-word phrases. Unfortunately, many sentiment-bearing words are not universally good or universally bad. For example, the polarity of an adjective could depend on the noun it modifies, e.g., hot coffee and unpredictable story express positive sentiment, while hot beer and unpredictable steering are negative. Thus, determining the out-of-context (a priori) strength of association of Twitter terms, especially multi-word terms, with positive/negative sentiment is an active research direction BIBREF28 , BIBREF29 . Phrase-level polarity in context. Even when the target noun is the same, the polarity of the modifying adjective could be different in different tweets, e.g., small is positive in (1) but negative in (2), even though they both refer to a phone. Thus, there has been research in determining the sentiment polarity of a term in the context of a tweet BIBREF26 , BIBREF29 , BIBREF27 . Sarcasm. Going back to tweet-level sentiment analysis, we should mention sarcastic tweets, which are particularly challenging as the sentiment they express is often the opposite of what the words they contain suggest BIBREF4 , BIBREF29 , BIBREF27 . For example, (4) and (5) express a negative sentiment even though they contain positive words and phrases such as thanks, love, and boosts my morale. Negative: Thanks manager for putting me on the schedule for Sunday Negative: I just love missing my train every single day. Really boosts my morale. Sentiment toward a topic. Even though tweets are short, as they are limited to 140 characters by design (even though this was relaxed a bit as of September 19, 2016, and now media attachments such as images, videos, polls, etc., and quoted tweets no longer reduce the character count), they are still long enough to allow the tweet's author to mention several topics and to express potentially different sentiment toward each of them. A topic can be anything that people express opinions about, e.g., a product (e.g., iPhone6), a political candidate (e.g., Donald Trump), a policy (e.g., Obamacare), an event (e.g., Brexit), etc. For example, in (6) the author is positive about Donald Trump but negative about Hillary Clinton. A political analyzer would not be interested so much in the overall sentiment expressed in the tweet (even though one could argue that here it is positive overall), but rather in the sentiment with respect to a topic of his/her interest of study. As a democrat I couldnt ethically support Hillary no matter who was running against her. Just so glad that its Trump, just love the guy! (topic: Hillary INLINEFORM0 Negative) (topic: Trump INLINEFORM0 Positive) Aspect-based sentiment analysis. Looking again at (1) and (2), we can say that the sentiment is not about the phone (lumia620 and iPhone 5, respectively), but rather about some specific aspect thereof, namely, size. Similarly, in (7) instead of sentiment toward the topic lasagna, we can see sentiment toward two aspects thereof: quality (Positive sentiment) and quantity (Negative sentiment). Aspect-based sentiment analysis is an active research area BIBREF32 , BIBREF33 , BIBREF21 . The lasagna is delicious but do not come here on an empty stomach. Stance detection. A task related to, but arguably different in some respect from sentiment analysis, is that of stance detection. The goal here is to determine whether the author of a piece of text is in favor of, against, or neutral toward a proposition or a target BIBREF36 . For example, in (8) the author has a negative stance toward the proposition w​omen have the right to abortion, even though the target is not mentioned at all. Similarly, in (9§) the author expresses a negative sentiment toward Mitt Romney, from which one can imply that s/he has a positive stance toward the target ​Barack Obama. A​ foetus has rights too! Make your voice heard. (Target: w​omen have the right to abortion INLINEFORM0 Against) A​ll Mitt Romney cares about is making money for the rich. (Target: ​Barack Obama INLINEFORM0 InFavor) Ordinal regression. The above tasks were offered in different granularities, e.g., 2-way (Positive, Negative), 3-way (Positive, Neutral, Negative), 4-way (Positive, Neutral, Negative, Objective), 5-way (HighlyPositive, Positive, Neutral, Negative, HighlyNegative), and sometimes even 11-way BIBREF34 . It is important to note that the 5-way and the 11-way scales are ordinal, i.e., the classes can be associated with numbers, e.g., INLINEFORM0 2, INLINEFORM1 1, 0, 1, and 2 for the 5-point scale. This changes the machine learning task as not all mistakes are equal anymore BIBREF16 . For example, misclassifying a HighlyNegative example as HighlyPositive is a bigger mistake than misclassifying it as Negative or as Neutral. From a machine learning perspective, this means moving from classification to ordinal regression. This also requires different evaluation measures BIBREF30 . Quantification. Practical applications are hardly ever interested in the sentiment expressed in a specific tweet. Rather, they look at estimating the prevalence of positive and negative tweets about a given topic in a set of tweets from some time interval. Most (if not all) tweet sentiment classification studies conducted within political science BIBREF39 , BIBREF40 , BIBREF41 , economics BIBREF42 , BIBREF7 , social science BIBREF43 , and market research BIBREF44 , BIBREF45 use Twitter with an interest in aggregate data and not in individual classifications. Thus, some tasks, such as SemEval-2016 Task 4 BIBREF30 , replace classification with class prevalence estimation, which is also known as quantification in data mining and related fields. Note that quantification is not a mere byproduct of classification, since a good classifier is not necessarily a good quantifier, and vice versa BIBREF46 . Finally, in case of multiple labels on an ordinal scale, we have yet another machine learning problem: ordinal quantification. Both versions of quantification require specific evaluation measures and machine learning algorithms. Features and Learning Pre-processing. Tweets are subject to standard preprocessing steps for text such as tokenization, stemming, lemmatization, stop-word removal, and part-of-speech tagging. Moreover, due to their noisy nature, they are also processed using some Twitter-specific techniques such as substitution/removal of URLs, of user mentions, of hashtags, and of emoticons, spelling correction, elongation normalization, abbreviation lookup, punctuation removal, detection of amplifiers and diminishers, negation scope detection, etc. For this, one typically uses Twitter-specific NLP tools such as part-of-speech and named entity taggers, syntactic parsers, etc. BIBREF47 , BIBREF48 , BIBREF49 . Negation handling. Special handling is also done for negation. The most popular approach to negation handling is to transform any word that appeared in a negation context by adding a suffix _NEG to it, e.g., good would become good_NEG BIBREF50 , BIBREF10 . A negated context is typically defined as a text span between a negation word, e.g., no, not, shouldn't, and a punctuation mark or the end of the message. Alternatively, one could flip the polarity of sentiment words, e.g., the positive word good would become negative when negated. It has also been argued BIBREF51 that negation affects different words differently, and thus it was also proposed to build and use special sentiment polarity lexicons for words in negation contexts BIBREF52 . Features. Traditionally, systems for Sentiment Analysis on Twitter have relied on handcrafted features derived from word-level (e.g., great, freshly roasted coffee, becoming president) and character-level INLINEFORM0 -grams (e.g., bec, beco, comin, oming), stems (e.g., becom), lemmata (e.g., become, roast), punctuation (e.g., exclamation and question marks), part-of-speech tags (e.g., adjectives, adverbs, verbs, nouns), word clusters (e.g., probably, probly, and maybe could be collapsed to the same word cluster), and Twitter-specific encodings such as emoticons (e.g., ;), :D), hashtags (#Brexit), user tags (e.g., @allenai_org), abbreviations (e.g., RT, BTW, F2F, OMG), elongated words (e.g., soooo, yaayyy), use of capitalization (e.g., proportion of ALL CAPS words), URLs, etc. Finally, the most important features are those based on the presence of words and phrases in sentiment polarity lexicons with positive/negative scores; examples of such features include number of positive terms, number of negative terms, ratio of the number of positive terms to the number of positive+negative terms, ratio of the number of negative terms to the number of positive+negative terms, sum of all positive scores, sum of all negative scores, sum of all scores, etc. Supervised learning. Traditionally, the above features were fed into classifiers such as Maximum Entropy (MaxEnt) and Support Vector Machines (SVM) with various kernels. However, observation over the SemEval Twitter sentiment task in recent years shows growing interest in, and by now clear dominance of methods based on deep learning. In particular, the best-performing systems at SemEval-2015 and SemEval-2016 used deep convolutional networks BIBREF53 , BIBREF54 . Conversely, kernel machines seem to be less frequently used than in the past, and the use of learning methods other than the ones mentioned above is at this point scarce. All these models are examples of supervised learning as they need labeled training data. Semi-supervised learning. We should note two things about the use of deep neural networks. First they can often do quite well without the need for explicit feature modeling, as they can learn the relevant features in their hidden layers starting from the raw text. Second, they have too many parameters, and thus they require a lot of training data, orders of magnitude more than it is realistic to have manually annotated. A popular way to solve this latter problem is to use self training, a form of semi-supervised learning, where first a system is trained on the available training data only, then this system is applied to make predictions on a large unannotated set of tweets, and finally it is trained for a few more iterations on its own predictions. This works because parts of the network, e.g., with convolution or with LSTMs BIBREF55 , BIBREF54 , BIBREF56 , need to learn something like a language model, i.e., which word is likely to follow which one. Training these parts needs no labels. While these parts can be also pre-trained, it is easier, and often better, to use self training. Distantly-supervised learning. Another way to make use of large unannotated datasets is to rely on distant supervision BIBREF41 . For example, one can annotate tweets for sentiment polarity based on whether they contain a positive or a negative emoticon. This results in noisy labels, which can be used to train a system BIBREF54 , to induce sentiment-specific word embeddings BIBREF57 , sentiment-polarity lexicons BIBREF25 , etc. Unsupervised learning. Fully unsupervised learning is not a popular method for addressing sentiment analysis tasks. Yet, some features used in sentiment analysis have been learned in an unsupervised way, e.g., Brown clusters to generalize over words BIBREF58 . Similarly, word embeddings are typically trained from raw tweets that have no annotation for sentiment (even though there is also work on sentiment-specific word embeddings BIBREF57 , which uses distant supervision). Sentiment Polarity Lexicons Despite the wide variety of knowledge sources explored so far in the literature, sentiment polarity lexicons remain the most commonly used resource for the task of sentiment analysis. Until recently, such sentiment polarity lexicons were manually crafted and were thus of small to moderate size, e.g., LIWC BIBREF59 has 2,300 words, the General Inquirer BIBREF60 contains 4,206 words, Bing Liu's lexicon BIBREF22 includes 6,786 words, and MPQA BIBREF14 has about 8,000 words. Early efforts toward building sentiment polarity lexicons automatically yielded lexicons of moderate sizes such as the SentiWordNet BIBREF19 , BIBREF61 . However, recent results have shown that automatically extracted large-scale lexicons (e.g., up to a million words and phrases) offer important performance advantages, as confirmed at shared tasks on Sentiment Analysis on Twitter at SemEval 2013-2016 BIBREF30 , BIBREF26 , BIBREF29 , BIBREF27 . Using such large-scale lexicons was crucial for the performance of the top-ranked systems. Similar observations were made in the related Aspect-Based Sentiment Analysis task at SemEval 2014 BIBREF21 . In both tasks, the winning systems benefitted from building and using massive sentiment polarity lexicons BIBREF25 , BIBREF62 . The two most popular large-scale lexicons were the Hashtag Sentiment Lexicon and the Sentiment140 lexicon, which were developed by the team of NRC Canada for their participation in the SemEval-2013 shared task on sentiment analysis on Twitter. Similar automatically induced lexicons proved useful for other SemEval tasks, e.g., for SemEval-2016 Task 3 on Community Question Answering BIBREF63 , BIBREF30 . The importance of building sentiment polarity lexicons has resulted in a special subtask BIBREF29 at SemEval-2015 (part of Task 4) and an entire task BIBREF37 at SemEval-2016 (namely, Task 7), on predicting the out-of-context sentiment intensity of words and phrases. Yet, we should note though that the utility of using sentiment polarity lexicons for sentiment analysis probably needs to be revisited, as the best system at SemEval-2016 Task 4 could win without using any lexicons BIBREF53 ; it relied on semi-supervised learning using a deep neural network. Various approaches have been proposed in the literature for bootstrapping sentiment polarity lexicons starting from a small set of seeds: positive and negative terms (words and phrases). The dominant approach is that of Turney BIBREF12 , who uses pointwise mutual information and bootstrapping to build a large lexicon and to estimate the semantic orientation of each word in that lexicon. He starts with a small set of seed positive (e.g., excellent) and negative words (e.g., bad), and then uses these words to induce sentiment polarity orientation for new words in a large unannotated set of texts (in his case, product reviews). The idea is that words that co-occur in the same text with positive seed words are likely to be positive, while those that tend to co-occur with negative words are likely to be negative. To quantify this intuition, Turney defines the notion of sentiment orientation (SO) for a term INLINEFORM0 as follows: INLINEFORM0 where PMI is the pointwise mutual information, INLINEFORM0 and INLINEFORM1 are placeholders standing for any of the seed positive and negative terms, respectively, and INLINEFORM2 is a target word/phrase from the large unannotated set of texts (here tweets). A positive/negative value for INLINEFORM0 indicates positive/negative polarity for the word INLINEFORM1 , and its magnitude shows the corresponding sentiment strength. In turn, INLINEFORM2 , where INLINEFORM3 is the probability to see INLINEFORM4 with any of the seed positive words in the same tweet, INLINEFORM5 is the probability to see INLINEFORM6 in any tweet, and INLINEFORM7 is the probability to see any of the seed positive words in a tweet; INLINEFORM8 is defined similarly. The pointwise mutual information is a notion from information theory: given two random variables INLINEFORM0 and INLINEFORM1 , the mutual information of INLINEFORM2 and INLINEFORM3 is the “amount of information” (in units such as bits) obtained about the random variable INLINEFORM4 , through the random variable INLINEFORM5 BIBREF64 . Let INLINEFORM0 and INLINEFORM1 be two values from the sample space of INLINEFORM2 and INLINEFORM3 , respectively. The pointwise mutual information between INLINEFORM4 and INLINEFORM5 is defined as follows: DISPLAYFORM0 INLINEFORM0 takes values between INLINEFORM1 , which happens when INLINEFORM2 = 0, and INLINEFORM3 if INLINEFORM4 . In his experiments, Turney BIBREF12 used five positive and five negative words as seeds. His PMI-based approach further served as the basis for the creation of the two above-mentioned large-scale automatic lexicons for sentiment analysis in Twitter for English, initially developed by NRC for their participation in SemEval-2013 BIBREF25 . The Hashtag Sentiment Lexicon uses as seeds hashtags containing 32 positive and 36 negative words, e.g., #happy and #sad. Similarly, the Sentiment140 lexicon uses smileys as seed indicators for positive and negative sentiment, e.g., :), :-), and :)) as positive seeds, and :( and :-( as negative ones. An alternative approach to lexicon induction has been proposed BIBREF65 , which, instead of using PMI, assigns positive/negative labels to the unlabeled tweets (based on the seeds), and then trains an SVM classifier on them, using word INLINEFORM0 -grams as features. These INLINEFORM1 -grams are then used as lexicon entries (words and phrases) with the learned classifier weights as polarity scores. Finally, it has been shown that sizable further performance gains can be obtained by starting with mid-sized seeds, i.e., hundreds of words and phrases BIBREF66 . Key Applications Sentiment analysis on Twitter has applications in a number of areas, including political science BIBREF39 , BIBREF40 , BIBREF41 , economics BIBREF42 , BIBREF7 , social science BIBREF43 , and market research BIBREF44 , BIBREF45 . It is used to study company reputation online BIBREF45 , to measure customer satisfaction, to identify detractors and promoters, to forecast market growth BIBREF42 , to predict the future income from newly-released movies, to forecast the outcome of upcoming elections BIBREF41 , BIBREF7 , to study political polarization BIBREF39 , BIBREF9 , etc. Future Directions We expect the quest for more interesting formulations of the general sentiment analysis task to continue. We see competitions such as those at SemEval as the engine of this innovation, as they not only perform head-to-head comparisons, but also create databases and tools that enable follow-up research for many years afterward. In terms of methods, we believe that deep learning BIBREF55 , BIBREF54 , BIBREF56 , together with semi-supervised and distantly-supervised methods BIBREF67 , BIBREF57 , will be the main focus of future research. We also expect more attention to be paid to linguistic structure and sentiment compositionality BIBREF68 , BIBREF69 . Moreover, we forecast more interest for languages other than English, and for cross-lingual methods BIBREF40 , BIBREF70 , BIBREF71 , which will allow leveraging on the rich resources that are already available for English. Last, but not least, the increase in opinion spam on Twitter will make it important to study astroturfing BIBREF72 and troll detection BIBREF73 , BIBREF74 , BIBREF75 . Cross-References Microblog Sentiment Analysis 100590 Multi-classifier System for Sentiment Analysis and Opinion Mining 351 Sentiment Analysis in Social Media 120 Sentiment Analysis of Microblogging Data 110168 Sentiment Analysis of Reviews 110169 Sentiment Analysis, Basics of 110159 Sentiment Quantification of User-Generated Content 110170 Social Media Analysis for Monitoring Political Sentiment 110172 Twitter Microblog Sentiment Analysis 265 User Sentiment and Opinion Analysis 192 Recommended Reading For general research on sentiment analysis, we recommend the following surveys: BIBREF76 and BIBREF15 . For sentiment analysis on Twitter, we recommend the overview article on Sentiment Analysis on Twitter about the SemEval task BIBREF28 as well as the task description papers for different editions of the task BIBREF30 , BIBREF26 , BIBREF29 , BIBREF27 .

deep convolutional networks BIBREF53 , BIBREF54

7997b9971f864a504014110a708f215c84815941

7997b9971f864a504014110a708f215c84815941_0

Q: What difficulties does sentiment analysis on Twitter have, compared to sentiment analysis in other domains? Text: Synonyms Microblog sentiment analysis; Twitter opinion mining Glossary Sentiment Analysis: This is text analysis aiming to determine the attitude of a speaker or a writer with respect to some topic or the overall contextual polarity of a piece of text. Definition Sentiment analysis on Twitter is the use of natural language processing techniques to identify and categorize opinions expressed in a tweet, in order to determine the author's attitude toward a particular topic or in general. Typically, discrete labels such as positive, negative, neutral, and objective are used for this purpose, but it is also possible to use labels on an ordinal scale, or even continuous numerical values. Introduction Internet and the proliferation of smart mobile devices have changed the way information is created, shared, and spreads, e.g., microblogs such as Twitter, weblogs such as LiveJournal, social networks such as Facebook, and instant messengers such as Skype and WhatsApp are now commonly used to share thoughts and opinions about anything in the surrounding world. This has resulted in the proliferation of social media content, thus creating new opportunities to study public opinion at a scale that was never possible before. Naturally, this abundance of data has quickly attracted business and research interest from various fields including marketing, political science, and social studies, among many others, which are interested in questions like these: Do people like the new Apple Watch? What do they hate about iPhone6? Do Americans support ObamaCare? What do Europeans think of Pope's visit to Palestine? How do we recognize the emergence of health problems such as depression? Do Germans like how Angela Merkel is handling the refugee crisis in Europe? What do republican voters in USA like/hate about Donald Trump? How do Scottish feel about the Brexit? Answering these questions requires studying the sentiment of opinions people express in social media, which has given rise to the fast growth of the field of sentiment analysis in social media, with Twitter being especially popular for research due to its scale, representativeness, variety of topics discussed, as well as ease of public access to its messages BIBREF0 , BIBREF1 . Despite all these opportunities, the rise of social media has also presented new challenges for natural language processing (NLP) applications, which had largely relied on NLP tools tuned for formal text genres such as newswire, and thus were not readily applicable to the informal language and style of social media. That language proved to be quite challenging with its use of creative spelling and punctuation, misspellings, slang, new words, URLs, and genre-specific terminology and abbreviations, e.g., RT for re-tweet and #hashtags. In addition to the genre difference, there is also a difference in length: social media messages are generally short, often length-limited by design as in Twitter, i.e., a sentence or a headline rather than a full document. How to handle such challenges has only recently been the subject of thorough research BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 . Key Points Sentiment analysis has a wide number of applications in areas such as market research, political and social sciences, and for studying public opinion in general, and Twitter is one of the most commonly-used platforms for this. This is due to its streaming nature, which allows for real-time analysis, to its social aspect, which encourages people to share opinions, and to the short size of the tweets, which simplifies linguistic analysis. There are several formulations of the task of Sentiment Analysis on Twitter that look at different sizes of the target (e.g., at the level of words vs. phrases vs. tweets vs. sets of tweets), at different types of semantic targets (e.g., aspect vs. topic vs. overall tweet), at the explicitness of the target (e.g., sentiment vs. stance detection), at the scale of the expected label (2-point vs. 3-point vs. ordinal), etc. All these are explored at SemEval, the International Workshop on Semantic Evaluation, which has created a number of benchmark datasets and has enabled direct comparison between different systems and approaches, both as part of the competition and beyond. Traditionally, the task has been addressed using supervised and semi-supervised methods, as well as using distant supervision, with the most important resource being sentiment polarity lexicons, and with feature-rich approaches as the dominant research direction for years. With the recent rise of deep learning, which in many cases eliminates the need for any explicit feature modeling, the importance of both lexicons and features diminishes, while at the same time attention is shifting towards learning from large unlabeled data, which is needed to train the high number of parameters of such complex models. Finally, as methods for sentiment analysis mature, more attention is also being paid to linguistic structure and to multi-linguality and cross-linguality. Historical Background Sentiment analysis emerged as a popular research direction in the early 2000s. Initially, it was regarded as standard document classification into topics such as business, sport, and politics BIBREF10 . However, researchers soon realized that it was quite different from standard document classification BIBREF11 , and that it crucially needed external knowledge in the form of sentiment polarity lexicons. Around the same time, other researchers realized the importance of external sentiment lexicons, e.g., Turney BIBREF12 proposed an unsupervised approach to learn the sentiment orientation of words/phrases: positive vs. negative. Later work studied the linguistic aspects of expressing opinions, evaluations, and speculations BIBREF13 , the role of context in determining the sentiment orientation BIBREF14 , of deeper linguistic processing such as negation handling BIBREF15 , of finer-grained sentiment distinctions BIBREF16 , of positional information BIBREF17 , etc. Moreover, it was recognized that in many cases, it is crucial to know not just the polarity of the sentiment but also the topic toward which this sentiment is expressed BIBREF18 . Until the rise of social media, research on opinion mining and sentiment analysis had focused primarily on learning about the language of sentiment in general, meaning that it was either genre-agnostic BIBREF19 or focused on newswire texts BIBREF20 and customer reviews (e.g., from web forums), most notably about movies BIBREF10 and restaurants BIBREF21 but also about hotels, digital cameras, cell phones, MP3 and DVD players BIBREF22 , laptops BIBREF21 , etc. This has given rise to several resources, mostly word and phrase polarity lexicons, which have proven to be very valuable for their respective domains and types of texts, but less useful for short social media messages. Later, with the emergence of social media, sentiment analysis in Twitter became a hot research topic. Unfortunately, research in that direction was hindered by the unavailability of suitable datasets and lexicons for system training, development, and testing. While some Twitter-specific resources were developed, initially they were either small and proprietary, such as the i-sieve corpus BIBREF6 , were created only for Spanish like the TASS corpus BIBREF23 , or relied on noisy labels obtained automatically, e.g., based on emoticons and hashtags BIBREF24 , BIBREF25 , BIBREF10 . This situation changed with the shared task on Sentiment Analysis on Twitter, which was organized at SemEval, the International Workshop on Semantic Evaluation, a semantic evaluation forum previously known as SensEval. The task ran in 2013, 2014, 2015, and 2016, attracting over 40 participating teams in all four editions. While the focus was on general tweets, the task also featured out-of-domain testing on SMS messages, LiveJournal messages, as well as on sarcastic tweets. SemEval-2013 Task 2 BIBREF26 and SemEval-2014 Task 9 BIBREF27 focused on expression-level and message-level polarity. SemEval-2015 Task 10 BIBREF28 , BIBREF29 featured topic-based message polarity classification on detecting trends toward a topic and on determining the out-of-context (a priori) strength of association of Twitter terms with positive sentiment. SemEval-2016 Task 4 BIBREF30 introduced a 5-point scale, which is used for human review ratings on popular websites such as Amazon, TripAdvisor, Yelp, etc.; from a research perspective, this meant moving from classification to ordinal regression. Moreover, it focused on quantification, i.e., determining what proportion of a set of tweets on a given topic are positive/negative about it. It also featured a 5-point scale ordinal quantification subtask BIBREF31 . Other related tasks have explored aspect-based sentiment analysis BIBREF32 , BIBREF33 , BIBREF21 , sentiment analysis of figurative language on Twitter BIBREF34 , implicit event polarity BIBREF35 , stance in tweets BIBREF36 , out-of-context sentiment intensity of phrases BIBREF37 , and emotion detection BIBREF38 . Some of these tasks featured languages other than English. Variants of the Task at SemEval Tweet-level sentiment. The simplest and also the most popular task of sentiment analysis on Twitter is to determine the overall sentiment expressed by the author of a tweet BIBREF30 , BIBREF28 , BIBREF26 , BIBREF29 , BIBREF27 . Typically, this means choosing one of the following three classes to describe the sentiment: Positive, Negative, and Neutral. Here are some examples: Positive: @nokia lumia620 cute and small and pocket-size, and available in the brigh test colours of day! #lumiacaption Negative: I hate tweeting on my iPhone 5 it's so small :( Neutral: If you work as a security in a samsung store...Does that make you guardian of the galaxy?? Sentiment polarity lexicons. Naturally, the overall sentiment in a tweet can be determined based on the sentiment-bearing words and phrases it contains as well as based on emoticons such as ;) and:(. For this purpose, researchers have been using lexicons of sentiment-bearing words. For example, cute is a positive word, while hate is a negative one, and the occurrence of these words in (1) and (2) can help determine the overall polarity of the respective tweet. We will discuss these lexicons in more detail below. Prior sentiment polarity of multi-word phrases. Unfortunately, many sentiment-bearing words are not universally good or universally bad. For example, the polarity of an adjective could depend on the noun it modifies, e.g., hot coffee and unpredictable story express positive sentiment, while hot beer and unpredictable steering are negative. Thus, determining the out-of-context (a priori) strength of association of Twitter terms, especially multi-word terms, with positive/negative sentiment is an active research direction BIBREF28 , BIBREF29 . Phrase-level polarity in context. Even when the target noun is the same, the polarity of the modifying adjective could be different in different tweets, e.g., small is positive in (1) but negative in (2), even though they both refer to a phone. Thus, there has been research in determining the sentiment polarity of a term in the context of a tweet BIBREF26 , BIBREF29 , BIBREF27 . Sarcasm. Going back to tweet-level sentiment analysis, we should mention sarcastic tweets, which are particularly challenging as the sentiment they express is often the opposite of what the words they contain suggest BIBREF4 , BIBREF29 , BIBREF27 . For example, (4) and (5) express a negative sentiment even though they contain positive words and phrases such as thanks, love, and boosts my morale. Negative: Thanks manager for putting me on the schedule for Sunday Negative: I just love missing my train every single day. Really boosts my morale. Sentiment toward a topic. Even though tweets are short, as they are limited to 140 characters by design (even though this was relaxed a bit as of September 19, 2016, and now media attachments such as images, videos, polls, etc., and quoted tweets no longer reduce the character count), they are still long enough to allow the tweet's author to mention several topics and to express potentially different sentiment toward each of them. A topic can be anything that people express opinions about, e.g., a product (e.g., iPhone6), a political candidate (e.g., Donald Trump), a policy (e.g., Obamacare), an event (e.g., Brexit), etc. For example, in (6) the author is positive about Donald Trump but negative about Hillary Clinton. A political analyzer would not be interested so much in the overall sentiment expressed in the tweet (even though one could argue that here it is positive overall), but rather in the sentiment with respect to a topic of his/her interest of study. As a democrat I couldnt ethically support Hillary no matter who was running against her. Just so glad that its Trump, just love the guy! (topic: Hillary INLINEFORM0 Negative) (topic: Trump INLINEFORM0 Positive) Aspect-based sentiment analysis. Looking again at (1) and (2), we can say that the sentiment is not about the phone (lumia620 and iPhone 5, respectively), but rather about some specific aspect thereof, namely, size. Similarly, in (7) instead of sentiment toward the topic lasagna, we can see sentiment toward two aspects thereof: quality (Positive sentiment) and quantity (Negative sentiment). Aspect-based sentiment analysis is an active research area BIBREF32 , BIBREF33 , BIBREF21 . The lasagna is delicious but do not come here on an empty stomach. Stance detection. A task related to, but arguably different in some respect from sentiment analysis, is that of stance detection. The goal here is to determine whether the author of a piece of text is in favor of, against, or neutral toward a proposition or a target BIBREF36 . For example, in (8) the author has a negative stance toward the proposition w​omen have the right to abortion, even though the target is not mentioned at all. Similarly, in (9§) the author expresses a negative sentiment toward Mitt Romney, from which one can imply that s/he has a positive stance toward the target ​Barack Obama. A​ foetus has rights too! Make your voice heard. (Target: w​omen have the right to abortion INLINEFORM0 Against) A​ll Mitt Romney cares about is making money for the rich. (Target: ​Barack Obama INLINEFORM0 InFavor) Ordinal regression. The above tasks were offered in different granularities, e.g., 2-way (Positive, Negative), 3-way (Positive, Neutral, Negative), 4-way (Positive, Neutral, Negative, Objective), 5-way (HighlyPositive, Positive, Neutral, Negative, HighlyNegative), and sometimes even 11-way BIBREF34 . It is important to note that the 5-way and the 11-way scales are ordinal, i.e., the classes can be associated with numbers, e.g., INLINEFORM0 2, INLINEFORM1 1, 0, 1, and 2 for the 5-point scale. This changes the machine learning task as not all mistakes are equal anymore BIBREF16 . For example, misclassifying a HighlyNegative example as HighlyPositive is a bigger mistake than misclassifying it as Negative or as Neutral. From a machine learning perspective, this means moving from classification to ordinal regression. This also requires different evaluation measures BIBREF30 . Quantification. Practical applications are hardly ever interested in the sentiment expressed in a specific tweet. Rather, they look at estimating the prevalence of positive and negative tweets about a given topic in a set of tweets from some time interval. Most (if not all) tweet sentiment classification studies conducted within political science BIBREF39 , BIBREF40 , BIBREF41 , economics BIBREF42 , BIBREF7 , social science BIBREF43 , and market research BIBREF44 , BIBREF45 use Twitter with an interest in aggregate data and not in individual classifications. Thus, some tasks, such as SemEval-2016 Task 4 BIBREF30 , replace classification with class prevalence estimation, which is also known as quantification in data mining and related fields. Note that quantification is not a mere byproduct of classification, since a good classifier is not necessarily a good quantifier, and vice versa BIBREF46 . Finally, in case of multiple labels on an ordinal scale, we have yet another machine learning problem: ordinal quantification. Both versions of quantification require specific evaluation measures and machine learning algorithms. Features and Learning Pre-processing. Tweets are subject to standard preprocessing steps for text such as tokenization, stemming, lemmatization, stop-word removal, and part-of-speech tagging. Moreover, due to their noisy nature, they are also processed using some Twitter-specific techniques such as substitution/removal of URLs, of user mentions, of hashtags, and of emoticons, spelling correction, elongation normalization, abbreviation lookup, punctuation removal, detection of amplifiers and diminishers, negation scope detection, etc. For this, one typically uses Twitter-specific NLP tools such as part-of-speech and named entity taggers, syntactic parsers, etc. BIBREF47 , BIBREF48 , BIBREF49 . Negation handling. Special handling is also done for negation. The most popular approach to negation handling is to transform any word that appeared in a negation context by adding a suffix _NEG to it, e.g., good would become good_NEG BIBREF50 , BIBREF10 . A negated context is typically defined as a text span between a negation word, e.g., no, not, shouldn't, and a punctuation mark or the end of the message. Alternatively, one could flip the polarity of sentiment words, e.g., the positive word good would become negative when negated. It has also been argued BIBREF51 that negation affects different words differently, and thus it was also proposed to build and use special sentiment polarity lexicons for words in negation contexts BIBREF52 . Features. Traditionally, systems for Sentiment Analysis on Twitter have relied on handcrafted features derived from word-level (e.g., great, freshly roasted coffee, becoming president) and character-level INLINEFORM0 -grams (e.g., bec, beco, comin, oming), stems (e.g., becom), lemmata (e.g., become, roast), punctuation (e.g., exclamation and question marks), part-of-speech tags (e.g., adjectives, adverbs, verbs, nouns), word clusters (e.g., probably, probly, and maybe could be collapsed to the same word cluster), and Twitter-specific encodings such as emoticons (e.g., ;), :D), hashtags (#Brexit), user tags (e.g., @allenai_org), abbreviations (e.g., RT, BTW, F2F, OMG), elongated words (e.g., soooo, yaayyy), use of capitalization (e.g., proportion of ALL CAPS words), URLs, etc. Finally, the most important features are those based on the presence of words and phrases in sentiment polarity lexicons with positive/negative scores; examples of such features include number of positive terms, number of negative terms, ratio of the number of positive terms to the number of positive+negative terms, ratio of the number of negative terms to the number of positive+negative terms, sum of all positive scores, sum of all negative scores, sum of all scores, etc. Supervised learning. Traditionally, the above features were fed into classifiers such as Maximum Entropy (MaxEnt) and Support Vector Machines (SVM) with various kernels. However, observation over the SemEval Twitter sentiment task in recent years shows growing interest in, and by now clear dominance of methods based on deep learning. In particular, the best-performing systems at SemEval-2015 and SemEval-2016 used deep convolutional networks BIBREF53 , BIBREF54 . Conversely, kernel machines seem to be less frequently used than in the past, and the use of learning methods other than the ones mentioned above is at this point scarce. All these models are examples of supervised learning as they need labeled training data. Semi-supervised learning. We should note two things about the use of deep neural networks. First they can often do quite well without the need for explicit feature modeling, as they can learn the relevant features in their hidden layers starting from the raw text. Second, they have too many parameters, and thus they require a lot of training data, orders of magnitude more than it is realistic to have manually annotated. A popular way to solve this latter problem is to use self training, a form of semi-supervised learning, where first a system is trained on the available training data only, then this system is applied to make predictions on a large unannotated set of tweets, and finally it is trained for a few more iterations on its own predictions. This works because parts of the network, e.g., with convolution or with LSTMs BIBREF55 , BIBREF54 , BIBREF56 , need to learn something like a language model, i.e., which word is likely to follow which one. Training these parts needs no labels. While these parts can be also pre-trained, it is easier, and often better, to use self training. Distantly-supervised learning. Another way to make use of large unannotated datasets is to rely on distant supervision BIBREF41 . For example, one can annotate tweets for sentiment polarity based on whether they contain a positive or a negative emoticon. This results in noisy labels, which can be used to train a system BIBREF54 , to induce sentiment-specific word embeddings BIBREF57 , sentiment-polarity lexicons BIBREF25 , etc. Unsupervised learning. Fully unsupervised learning is not a popular method for addressing sentiment analysis tasks. Yet, some features used in sentiment analysis have been learned in an unsupervised way, e.g., Brown clusters to generalize over words BIBREF58 . Similarly, word embeddings are typically trained from raw tweets that have no annotation for sentiment (even though there is also work on sentiment-specific word embeddings BIBREF57 , which uses distant supervision). Sentiment Polarity Lexicons Despite the wide variety of knowledge sources explored so far in the literature, sentiment polarity lexicons remain the most commonly used resource for the task of sentiment analysis. Until recently, such sentiment polarity lexicons were manually crafted and were thus of small to moderate size, e.g., LIWC BIBREF59 has 2,300 words, the General Inquirer BIBREF60 contains 4,206 words, Bing Liu's lexicon BIBREF22 includes 6,786 words, and MPQA BIBREF14 has about 8,000 words. Early efforts toward building sentiment polarity lexicons automatically yielded lexicons of moderate sizes such as the SentiWordNet BIBREF19 , BIBREF61 . However, recent results have shown that automatically extracted large-scale lexicons (e.g., up to a million words and phrases) offer important performance advantages, as confirmed at shared tasks on Sentiment Analysis on Twitter at SemEval 2013-2016 BIBREF30 , BIBREF26 , BIBREF29 , BIBREF27 . Using such large-scale lexicons was crucial for the performance of the top-ranked systems. Similar observations were made in the related Aspect-Based Sentiment Analysis task at SemEval 2014 BIBREF21 . In both tasks, the winning systems benefitted from building and using massive sentiment polarity lexicons BIBREF25 , BIBREF62 . The two most popular large-scale lexicons were the Hashtag Sentiment Lexicon and the Sentiment140 lexicon, which were developed by the team of NRC Canada for their participation in the SemEval-2013 shared task on sentiment analysis on Twitter. Similar automatically induced lexicons proved useful for other SemEval tasks, e.g., for SemEval-2016 Task 3 on Community Question Answering BIBREF63 , BIBREF30 . The importance of building sentiment polarity lexicons has resulted in a special subtask BIBREF29 at SemEval-2015 (part of Task 4) and an entire task BIBREF37 at SemEval-2016 (namely, Task 7), on predicting the out-of-context sentiment intensity of words and phrases. Yet, we should note though that the utility of using sentiment polarity lexicons for sentiment analysis probably needs to be revisited, as the best system at SemEval-2016 Task 4 could win without using any lexicons BIBREF53 ; it relied on semi-supervised learning using a deep neural network. Various approaches have been proposed in the literature for bootstrapping sentiment polarity lexicons starting from a small set of seeds: positive and negative terms (words and phrases). The dominant approach is that of Turney BIBREF12 , who uses pointwise mutual information and bootstrapping to build a large lexicon and to estimate the semantic orientation of each word in that lexicon. He starts with a small set of seed positive (e.g., excellent) and negative words (e.g., bad), and then uses these words to induce sentiment polarity orientation for new words in a large unannotated set of texts (in his case, product reviews). The idea is that words that co-occur in the same text with positive seed words are likely to be positive, while those that tend to co-occur with negative words are likely to be negative. To quantify this intuition, Turney defines the notion of sentiment orientation (SO) for a term INLINEFORM0 as follows: INLINEFORM0 where PMI is the pointwise mutual information, INLINEFORM0 and INLINEFORM1 are placeholders standing for any of the seed positive and negative terms, respectively, and INLINEFORM2 is a target word/phrase from the large unannotated set of texts (here tweets). A positive/negative value for INLINEFORM0 indicates positive/negative polarity for the word INLINEFORM1 , and its magnitude shows the corresponding sentiment strength. In turn, INLINEFORM2 , where INLINEFORM3 is the probability to see INLINEFORM4 with any of the seed positive words in the same tweet, INLINEFORM5 is the probability to see INLINEFORM6 in any tweet, and INLINEFORM7 is the probability to see any of the seed positive words in a tweet; INLINEFORM8 is defined similarly. The pointwise mutual information is a notion from information theory: given two random variables INLINEFORM0 and INLINEFORM1 , the mutual information of INLINEFORM2 and INLINEFORM3 is the “amount of information” (in units such as bits) obtained about the random variable INLINEFORM4 , through the random variable INLINEFORM5 BIBREF64 . Let INLINEFORM0 and INLINEFORM1 be two values from the sample space of INLINEFORM2 and INLINEFORM3 , respectively. The pointwise mutual information between INLINEFORM4 and INLINEFORM5 is defined as follows: DISPLAYFORM0 INLINEFORM0 takes values between INLINEFORM1 , which happens when INLINEFORM2 = 0, and INLINEFORM3 if INLINEFORM4 . In his experiments, Turney BIBREF12 used five positive and five negative words as seeds. His PMI-based approach further served as the basis for the creation of the two above-mentioned large-scale automatic lexicons for sentiment analysis in Twitter for English, initially developed by NRC for their participation in SemEval-2013 BIBREF25 . The Hashtag Sentiment Lexicon uses as seeds hashtags containing 32 positive and 36 negative words, e.g., #happy and #sad. Similarly, the Sentiment140 lexicon uses smileys as seed indicators for positive and negative sentiment, e.g., :), :-), and :)) as positive seeds, and :( and :-( as negative ones. An alternative approach to lexicon induction has been proposed BIBREF65 , which, instead of using PMI, assigns positive/negative labels to the unlabeled tweets (based on the seeds), and then trains an SVM classifier on them, using word INLINEFORM0 -grams as features. These INLINEFORM1 -grams are then used as lexicon entries (words and phrases) with the learned classifier weights as polarity scores. Finally, it has been shown that sizable further performance gains can be obtained by starting with mid-sized seeds, i.e., hundreds of words and phrases BIBREF66 . Key Applications Sentiment analysis on Twitter has applications in a number of areas, including political science BIBREF39 , BIBREF40 , BIBREF41 , economics BIBREF42 , BIBREF7 , social science BIBREF43 , and market research BIBREF44 , BIBREF45 . It is used to study company reputation online BIBREF45 , to measure customer satisfaction, to identify detractors and promoters, to forecast market growth BIBREF42 , to predict the future income from newly-released movies, to forecast the outcome of upcoming elections BIBREF41 , BIBREF7 , to study political polarization BIBREF39 , BIBREF9 , etc. Future Directions We expect the quest for more interesting formulations of the general sentiment analysis task to continue. We see competitions such as those at SemEval as the engine of this innovation, as they not only perform head-to-head comparisons, but also create databases and tools that enable follow-up research for many years afterward. In terms of methods, we believe that deep learning BIBREF55 , BIBREF54 , BIBREF56 , together with semi-supervised and distantly-supervised methods BIBREF67 , BIBREF57 , will be the main focus of future research. We also expect more attention to be paid to linguistic structure and sentiment compositionality BIBREF68 , BIBREF69 . Moreover, we forecast more interest for languages other than English, and for cross-lingual methods BIBREF40 , BIBREF70 , BIBREF71 , which will allow leveraging on the rich resources that are already available for English. Last, but not least, the increase in opinion spam on Twitter will make it important to study astroturfing BIBREF72 and troll detection BIBREF73 , BIBREF74 , BIBREF75 . Cross-References Microblog Sentiment Analysis 100590 Multi-classifier System for Sentiment Analysis and Opinion Mining 351 Sentiment Analysis in Social Media 120 Sentiment Analysis of Microblogging Data 110168 Sentiment Analysis of Reviews 110169 Sentiment Analysis, Basics of 110159 Sentiment Quantification of User-Generated Content 110170 Social Media Analysis for Monitoring Political Sentiment 110172 Twitter Microblog Sentiment Analysis 265 User Sentiment and Opinion Analysis 192 Recommended Reading For general research on sentiment analysis, we recommend the following surveys: BIBREF76 and BIBREF15 . For sentiment analysis on Twitter, we recommend the overview article on Sentiment Analysis on Twitter about the SemEval task BIBREF28 as well as the task description papers for different editions of the task BIBREF30 , BIBREF26 , BIBREF29 , BIBREF27 .

Tweets noisy nature, use of creative spelling and punctuation, misspellings, slang, new words, URLs, and genre-specific terminology and abbreviations, short (length limited) text

0d1408744651c3847469c4a005e4a9dccbd89cf1

0d1408744651c3847469c4a005e4a9dccbd89cf1_0

Q: What are the metrics to evaluate sentiment analysis on Twitter? Text: Synonyms Microblog sentiment analysis; Twitter opinion mining Glossary Sentiment Analysis: This is text analysis aiming to determine the attitude of a speaker or a writer with respect to some topic or the overall contextual polarity of a piece of text. Definition Sentiment analysis on Twitter is the use of natural language processing techniques to identify and categorize opinions expressed in a tweet, in order to determine the author's attitude toward a particular topic or in general. Typically, discrete labels such as positive, negative, neutral, and objective are used for this purpose, but it is also possible to use labels on an ordinal scale, or even continuous numerical values. Introduction Internet and the proliferation of smart mobile devices have changed the way information is created, shared, and spreads, e.g., microblogs such as Twitter, weblogs such as LiveJournal, social networks such as Facebook, and instant messengers such as Skype and WhatsApp are now commonly used to share thoughts and opinions about anything in the surrounding world. This has resulted in the proliferation of social media content, thus creating new opportunities to study public opinion at a scale that was never possible before. Naturally, this abundance of data has quickly attracted business and research interest from various fields including marketing, political science, and social studies, among many others, which are interested in questions like these: Do people like the new Apple Watch? What do they hate about iPhone6? Do Americans support ObamaCare? What do Europeans think of Pope's visit to Palestine? How do we recognize the emergence of health problems such as depression? Do Germans like how Angela Merkel is handling the refugee crisis in Europe? What do republican voters in USA like/hate about Donald Trump? How do Scottish feel about the Brexit? Answering these questions requires studying the sentiment of opinions people express in social media, which has given rise to the fast growth of the field of sentiment analysis in social media, with Twitter being especially popular for research due to its scale, representativeness, variety of topics discussed, as well as ease of public access to its messages BIBREF0 , BIBREF1 . Despite all these opportunities, the rise of social media has also presented new challenges for natural language processing (NLP) applications, which had largely relied on NLP tools tuned for formal text genres such as newswire, and thus were not readily applicable to the informal language and style of social media. That language proved to be quite challenging with its use of creative spelling and punctuation, misspellings, slang, new words, URLs, and genre-specific terminology and abbreviations, e.g., RT for re-tweet and #hashtags. In addition to the genre difference, there is also a difference in length: social media messages are generally short, often length-limited by design as in Twitter, i.e., a sentence or a headline rather than a full document. How to handle such challenges has only recently been the subject of thorough research BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 . Key Points Sentiment analysis has a wide number of applications in areas such as market research, political and social sciences, and for studying public opinion in general, and Twitter is one of the most commonly-used platforms for this. This is due to its streaming nature, which allows for real-time analysis, to its social aspect, which encourages people to share opinions, and to the short size of the tweets, which simplifies linguistic analysis. There are several formulations of the task of Sentiment Analysis on Twitter that look at different sizes of the target (e.g., at the level of words vs. phrases vs. tweets vs. sets of tweets), at different types of semantic targets (e.g., aspect vs. topic vs. overall tweet), at the explicitness of the target (e.g., sentiment vs. stance detection), at the scale of the expected label (2-point vs. 3-point vs. ordinal), etc. All these are explored at SemEval, the International Workshop on Semantic Evaluation, which has created a number of benchmark datasets and has enabled direct comparison between different systems and approaches, both as part of the competition and beyond. Traditionally, the task has been addressed using supervised and semi-supervised methods, as well as using distant supervision, with the most important resource being sentiment polarity lexicons, and with feature-rich approaches as the dominant research direction for years. With the recent rise of deep learning, which in many cases eliminates the need for any explicit feature modeling, the importance of both lexicons and features diminishes, while at the same time attention is shifting towards learning from large unlabeled data, which is needed to train the high number of parameters of such complex models. Finally, as methods for sentiment analysis mature, more attention is also being paid to linguistic structure and to multi-linguality and cross-linguality. Historical Background Sentiment analysis emerged as a popular research direction in the early 2000s. Initially, it was regarded as standard document classification into topics such as business, sport, and politics BIBREF10 . However, researchers soon realized that it was quite different from standard document classification BIBREF11 , and that it crucially needed external knowledge in the form of sentiment polarity lexicons. Around the same time, other researchers realized the importance of external sentiment lexicons, e.g., Turney BIBREF12 proposed an unsupervised approach to learn the sentiment orientation of words/phrases: positive vs. negative. Later work studied the linguistic aspects of expressing opinions, evaluations, and speculations BIBREF13 , the role of context in determining the sentiment orientation BIBREF14 , of deeper linguistic processing such as negation handling BIBREF15 , of finer-grained sentiment distinctions BIBREF16 , of positional information BIBREF17 , etc. Moreover, it was recognized that in many cases, it is crucial to know not just the polarity of the sentiment but also the topic toward which this sentiment is expressed BIBREF18 . Until the rise of social media, research on opinion mining and sentiment analysis had focused primarily on learning about the language of sentiment in general, meaning that it was either genre-agnostic BIBREF19 or focused on newswire texts BIBREF20 and customer reviews (e.g., from web forums), most notably about movies BIBREF10 and restaurants BIBREF21 but also about hotels, digital cameras, cell phones, MP3 and DVD players BIBREF22 , laptops BIBREF21 , etc. This has given rise to several resources, mostly word and phrase polarity lexicons, which have proven to be very valuable for their respective domains and types of texts, but less useful for short social media messages. Later, with the emergence of social media, sentiment analysis in Twitter became a hot research topic. Unfortunately, research in that direction was hindered by the unavailability of suitable datasets and lexicons for system training, development, and testing. While some Twitter-specific resources were developed, initially they were either small and proprietary, such as the i-sieve corpus BIBREF6 , were created only for Spanish like the TASS corpus BIBREF23 , or relied on noisy labels obtained automatically, e.g., based on emoticons and hashtags BIBREF24 , BIBREF25 , BIBREF10 . This situation changed with the shared task on Sentiment Analysis on Twitter, which was organized at SemEval, the International Workshop on Semantic Evaluation, a semantic evaluation forum previously known as SensEval. The task ran in 2013, 2014, 2015, and 2016, attracting over 40 participating teams in all four editions. While the focus was on general tweets, the task also featured out-of-domain testing on SMS messages, LiveJournal messages, as well as on sarcastic tweets. SemEval-2013 Task 2 BIBREF26 and SemEval-2014 Task 9 BIBREF27 focused on expression-level and message-level polarity. SemEval-2015 Task 10 BIBREF28 , BIBREF29 featured topic-based message polarity classification on detecting trends toward a topic and on determining the out-of-context (a priori) strength of association of Twitter terms with positive sentiment. SemEval-2016 Task 4 BIBREF30 introduced a 5-point scale, which is used for human review ratings on popular websites such as Amazon, TripAdvisor, Yelp, etc.; from a research perspective, this meant moving from classification to ordinal regression. Moreover, it focused on quantification, i.e., determining what proportion of a set of tweets on a given topic are positive/negative about it. It also featured a 5-point scale ordinal quantification subtask BIBREF31 . Other related tasks have explored aspect-based sentiment analysis BIBREF32 , BIBREF33 , BIBREF21 , sentiment analysis of figurative language on Twitter BIBREF34 , implicit event polarity BIBREF35 , stance in tweets BIBREF36 , out-of-context sentiment intensity of phrases BIBREF37 , and emotion detection BIBREF38 . Some of these tasks featured languages other than English. Variants of the Task at SemEval Tweet-level sentiment. The simplest and also the most popular task of sentiment analysis on Twitter is to determine the overall sentiment expressed by the author of a tweet BIBREF30 , BIBREF28 , BIBREF26 , BIBREF29 , BIBREF27 . Typically, this means choosing one of the following three classes to describe the sentiment: Positive, Negative, and Neutral. Here are some examples: Positive: @nokia lumia620 cute and small and pocket-size, and available in the brigh test colours of day! #lumiacaption Negative: I hate tweeting on my iPhone 5 it's so small :( Neutral: If you work as a security in a samsung store...Does that make you guardian of the galaxy?? Sentiment polarity lexicons. Naturally, the overall sentiment in a tweet can be determined based on the sentiment-bearing words and phrases it contains as well as based on emoticons such as ;) and:(. For this purpose, researchers have been using lexicons of sentiment-bearing words. For example, cute is a positive word, while hate is a negative one, and the occurrence of these words in (1) and (2) can help determine the overall polarity of the respective tweet. We will discuss these lexicons in more detail below. Prior sentiment polarity of multi-word phrases. Unfortunately, many sentiment-bearing words are not universally good or universally bad. For example, the polarity of an adjective could depend on the noun it modifies, e.g., hot coffee and unpredictable story express positive sentiment, while hot beer and unpredictable steering are negative. Thus, determining the out-of-context (a priori) strength of association of Twitter terms, especially multi-word terms, with positive/negative sentiment is an active research direction BIBREF28 , BIBREF29 . Phrase-level polarity in context. Even when the target noun is the same, the polarity of the modifying adjective could be different in different tweets, e.g., small is positive in (1) but negative in (2), even though they both refer to a phone. Thus, there has been research in determining the sentiment polarity of a term in the context of a tweet BIBREF26 , BIBREF29 , BIBREF27 . Sarcasm. Going back to tweet-level sentiment analysis, we should mention sarcastic tweets, which are particularly challenging as the sentiment they express is often the opposite of what the words they contain suggest BIBREF4 , BIBREF29 , BIBREF27 . For example, (4) and (5) express a negative sentiment even though they contain positive words and phrases such as thanks, love, and boosts my morale. Negative: Thanks manager for putting me on the schedule for Sunday Negative: I just love missing my train every single day. Really boosts my morale. Sentiment toward a topic. Even though tweets are short, as they are limited to 140 characters by design (even though this was relaxed a bit as of September 19, 2016, and now media attachments such as images, videos, polls, etc., and quoted tweets no longer reduce the character count), they are still long enough to allow the tweet's author to mention several topics and to express potentially different sentiment toward each of them. A topic can be anything that people express opinions about, e.g., a product (e.g., iPhone6), a political candidate (e.g., Donald Trump), a policy (e.g., Obamacare), an event (e.g., Brexit), etc. For example, in (6) the author is positive about Donald Trump but negative about Hillary Clinton. A political analyzer would not be interested so much in the overall sentiment expressed in the tweet (even though one could argue that here it is positive overall), but rather in the sentiment with respect to a topic of his/her interest of study. As a democrat I couldnt ethically support Hillary no matter who was running against her. Just so glad that its Trump, just love the guy! (topic: Hillary INLINEFORM0 Negative) (topic: Trump INLINEFORM0 Positive) Aspect-based sentiment analysis. Looking again at (1) and (2), we can say that the sentiment is not about the phone (lumia620 and iPhone 5, respectively), but rather about some specific aspect thereof, namely, size. Similarly, in (7) instead of sentiment toward the topic lasagna, we can see sentiment toward two aspects thereof: quality (Positive sentiment) and quantity (Negative sentiment). Aspect-based sentiment analysis is an active research area BIBREF32 , BIBREF33 , BIBREF21 . The lasagna is delicious but do not come here on an empty stomach. Stance detection. A task related to, but arguably different in some respect from sentiment analysis, is that of stance detection. The goal here is to determine whether the author of a piece of text is in favor of, against, or neutral toward a proposition or a target BIBREF36 . For example, in (8) the author has a negative stance toward the proposition w​omen have the right to abortion, even though the target is not mentioned at all. Similarly, in (9§) the author expresses a negative sentiment toward Mitt Romney, from which one can imply that s/he has a positive stance toward the target ​Barack Obama. A​ foetus has rights too! Make your voice heard. (Target: w​omen have the right to abortion INLINEFORM0 Against) A​ll Mitt Romney cares about is making money for the rich. (Target: ​Barack Obama INLINEFORM0 InFavor) Ordinal regression. The above tasks were offered in different granularities, e.g., 2-way (Positive, Negative), 3-way (Positive, Neutral, Negative), 4-way (Positive, Neutral, Negative, Objective), 5-way (HighlyPositive, Positive, Neutral, Negative, HighlyNegative), and sometimes even 11-way BIBREF34 . It is important to note that the 5-way and the 11-way scales are ordinal, i.e., the classes can be associated with numbers, e.g., INLINEFORM0 2, INLINEFORM1 1, 0, 1, and 2 for the 5-point scale. This changes the machine learning task as not all mistakes are equal anymore BIBREF16 . For example, misclassifying a HighlyNegative example as HighlyPositive is a bigger mistake than misclassifying it as Negative or as Neutral. From a machine learning perspective, this means moving from classification to ordinal regression. This also requires different evaluation measures BIBREF30 . Quantification. Practical applications are hardly ever interested in the sentiment expressed in a specific tweet. Rather, they look at estimating the prevalence of positive and negative tweets about a given topic in a set of tweets from some time interval. Most (if not all) tweet sentiment classification studies conducted within political science BIBREF39 , BIBREF40 , BIBREF41 , economics BIBREF42 , BIBREF7 , social science BIBREF43 , and market research BIBREF44 , BIBREF45 use Twitter with an interest in aggregate data and not in individual classifications. Thus, some tasks, such as SemEval-2016 Task 4 BIBREF30 , replace classification with class prevalence estimation, which is also known as quantification in data mining and related fields. Note that quantification is not a mere byproduct of classification, since a good classifier is not necessarily a good quantifier, and vice versa BIBREF46 . Finally, in case of multiple labels on an ordinal scale, we have yet another machine learning problem: ordinal quantification. Both versions of quantification require specific evaluation measures and machine learning algorithms. Features and Learning Pre-processing. Tweets are subject to standard preprocessing steps for text such as tokenization, stemming, lemmatization, stop-word removal, and part-of-speech tagging. Moreover, due to their noisy nature, they are also processed using some Twitter-specific techniques such as substitution/removal of URLs, of user mentions, of hashtags, and of emoticons, spelling correction, elongation normalization, abbreviation lookup, punctuation removal, detection of amplifiers and diminishers, negation scope detection, etc. For this, one typically uses Twitter-specific NLP tools such as part-of-speech and named entity taggers, syntactic parsers, etc. BIBREF47 , BIBREF48 , BIBREF49 . Negation handling. Special handling is also done for negation. The most popular approach to negation handling is to transform any word that appeared in a negation context by adding a suffix _NEG to it, e.g., good would become good_NEG BIBREF50 , BIBREF10 . A negated context is typically defined as a text span between a negation word, e.g., no, not, shouldn't, and a punctuation mark or the end of the message. Alternatively, one could flip the polarity of sentiment words, e.g., the positive word good would become negative when negated. It has also been argued BIBREF51 that negation affects different words differently, and thus it was also proposed to build and use special sentiment polarity lexicons for words in negation contexts BIBREF52 . Features. Traditionally, systems for Sentiment Analysis on Twitter have relied on handcrafted features derived from word-level (e.g., great, freshly roasted coffee, becoming president) and character-level INLINEFORM0 -grams (e.g., bec, beco, comin, oming), stems (e.g., becom), lemmata (e.g., become, roast), punctuation (e.g., exclamation and question marks), part-of-speech tags (e.g., adjectives, adverbs, verbs, nouns), word clusters (e.g., probably, probly, and maybe could be collapsed to the same word cluster), and Twitter-specific encodings such as emoticons (e.g., ;), :D), hashtags (#Brexit), user tags (e.g., @allenai_org), abbreviations (e.g., RT, BTW, F2F, OMG), elongated words (e.g., soooo, yaayyy), use of capitalization (e.g., proportion of ALL CAPS words), URLs, etc. Finally, the most important features are those based on the presence of words and phrases in sentiment polarity lexicons with positive/negative scores; examples of such features include number of positive terms, number of negative terms, ratio of the number of positive terms to the number of positive+negative terms, ratio of the number of negative terms to the number of positive+negative terms, sum of all positive scores, sum of all negative scores, sum of all scores, etc. Supervised learning. Traditionally, the above features were fed into classifiers such as Maximum Entropy (MaxEnt) and Support Vector Machines (SVM) with various kernels. However, observation over the SemEval Twitter sentiment task in recent years shows growing interest in, and by now clear dominance of methods based on deep learning. In particular, the best-performing systems at SemEval-2015 and SemEval-2016 used deep convolutional networks BIBREF53 , BIBREF54 . Conversely, kernel machines seem to be less frequently used than in the past, and the use of learning methods other than the ones mentioned above is at this point scarce. All these models are examples of supervised learning as they need labeled training data. Semi-supervised learning. We should note two things about the use of deep neural networks. First they can often do quite well without the need for explicit feature modeling, as they can learn the relevant features in their hidden layers starting from the raw text. Second, they have too many parameters, and thus they require a lot of training data, orders of magnitude more than it is realistic to have manually annotated. A popular way to solve this latter problem is to use self training, a form of semi-supervised learning, where first a system is trained on the available training data only, then this system is applied to make predictions on a large unannotated set of tweets, and finally it is trained for a few more iterations on its own predictions. This works because parts of the network, e.g., with convolution or with LSTMs BIBREF55 , BIBREF54 , BIBREF56 , need to learn something like a language model, i.e., which word is likely to follow which one. Training these parts needs no labels. While these parts can be also pre-trained, it is easier, and often better, to use self training. Distantly-supervised learning. Another way to make use of large unannotated datasets is to rely on distant supervision BIBREF41 . For example, one can annotate tweets for sentiment polarity based on whether they contain a positive or a negative emoticon. This results in noisy labels, which can be used to train a system BIBREF54 , to induce sentiment-specific word embeddings BIBREF57 , sentiment-polarity lexicons BIBREF25 , etc. Unsupervised learning. Fully unsupervised learning is not a popular method for addressing sentiment analysis tasks. Yet, some features used in sentiment analysis have been learned in an unsupervised way, e.g., Brown clusters to generalize over words BIBREF58 . Similarly, word embeddings are typically trained from raw tweets that have no annotation for sentiment (even though there is also work on sentiment-specific word embeddings BIBREF57 , which uses distant supervision). Sentiment Polarity Lexicons Despite the wide variety of knowledge sources explored so far in the literature, sentiment polarity lexicons remain the most commonly used resource for the task of sentiment analysis. Until recently, such sentiment polarity lexicons were manually crafted and were thus of small to moderate size, e.g., LIWC BIBREF59 has 2,300 words, the General Inquirer BIBREF60 contains 4,206 words, Bing Liu's lexicon BIBREF22 includes 6,786 words, and MPQA BIBREF14 has about 8,000 words. Early efforts toward building sentiment polarity lexicons automatically yielded lexicons of moderate sizes such as the SentiWordNet BIBREF19 , BIBREF61 . However, recent results have shown that automatically extracted large-scale lexicons (e.g., up to a million words and phrases) offer important performance advantages, as confirmed at shared tasks on Sentiment Analysis on Twitter at SemEval 2013-2016 BIBREF30 , BIBREF26 , BIBREF29 , BIBREF27 . Using such large-scale lexicons was crucial for the performance of the top-ranked systems. Similar observations were made in the related Aspect-Based Sentiment Analysis task at SemEval 2014 BIBREF21 . In both tasks, the winning systems benefitted from building and using massive sentiment polarity lexicons BIBREF25 , BIBREF62 . The two most popular large-scale lexicons were the Hashtag Sentiment Lexicon and the Sentiment140 lexicon, which were developed by the team of NRC Canada for their participation in the SemEval-2013 shared task on sentiment analysis on Twitter. Similar automatically induced lexicons proved useful for other SemEval tasks, e.g., for SemEval-2016 Task 3 on Community Question Answering BIBREF63 , BIBREF30 . The importance of building sentiment polarity lexicons has resulted in a special subtask BIBREF29 at SemEval-2015 (part of Task 4) and an entire task BIBREF37 at SemEval-2016 (namely, Task 7), on predicting the out-of-context sentiment intensity of words and phrases. Yet, we should note though that the utility of using sentiment polarity lexicons for sentiment analysis probably needs to be revisited, as the best system at SemEval-2016 Task 4 could win without using any lexicons BIBREF53 ; it relied on semi-supervised learning using a deep neural network. Various approaches have been proposed in the literature for bootstrapping sentiment polarity lexicons starting from a small set of seeds: positive and negative terms (words and phrases). The dominant approach is that of Turney BIBREF12 , who uses pointwise mutual information and bootstrapping to build a large lexicon and to estimate the semantic orientation of each word in that lexicon. He starts with a small set of seed positive (e.g., excellent) and negative words (e.g., bad), and then uses these words to induce sentiment polarity orientation for new words in a large unannotated set of texts (in his case, product reviews). The idea is that words that co-occur in the same text with positive seed words are likely to be positive, while those that tend to co-occur with negative words are likely to be negative. To quantify this intuition, Turney defines the notion of sentiment orientation (SO) for a term INLINEFORM0 as follows: INLINEFORM0 where PMI is the pointwise mutual information, INLINEFORM0 and INLINEFORM1 are placeholders standing for any of the seed positive and negative terms, respectively, and INLINEFORM2 is a target word/phrase from the large unannotated set of texts (here tweets). A positive/negative value for INLINEFORM0 indicates positive/negative polarity for the word INLINEFORM1 , and its magnitude shows the corresponding sentiment strength. In turn, INLINEFORM2 , where INLINEFORM3 is the probability to see INLINEFORM4 with any of the seed positive words in the same tweet, INLINEFORM5 is the probability to see INLINEFORM6 in any tweet, and INLINEFORM7 is the probability to see any of the seed positive words in a tweet; INLINEFORM8 is defined similarly. The pointwise mutual information is a notion from information theory: given two random variables INLINEFORM0 and INLINEFORM1 , the mutual information of INLINEFORM2 and INLINEFORM3 is the “amount of information” (in units such as bits) obtained about the random variable INLINEFORM4 , through the random variable INLINEFORM5 BIBREF64 . Let INLINEFORM0 and INLINEFORM1 be two values from the sample space of INLINEFORM2 and INLINEFORM3 , respectively. The pointwise mutual information between INLINEFORM4 and INLINEFORM5 is defined as follows: DISPLAYFORM0 INLINEFORM0 takes values between INLINEFORM1 , which happens when INLINEFORM2 = 0, and INLINEFORM3 if INLINEFORM4 . In his experiments, Turney BIBREF12 used five positive and five negative words as seeds. His PMI-based approach further served as the basis for the creation of the two above-mentioned large-scale automatic lexicons for sentiment analysis in Twitter for English, initially developed by NRC for their participation in SemEval-2013 BIBREF25 . The Hashtag Sentiment Lexicon uses as seeds hashtags containing 32 positive and 36 negative words, e.g., #happy and #sad. Similarly, the Sentiment140 lexicon uses smileys as seed indicators for positive and negative sentiment, e.g., :), :-), and :)) as positive seeds, and :( and :-( as negative ones. An alternative approach to lexicon induction has been proposed BIBREF65 , which, instead of using PMI, assigns positive/negative labels to the unlabeled tweets (based on the seeds), and then trains an SVM classifier on them, using word INLINEFORM0 -grams as features. These INLINEFORM1 -grams are then used as lexicon entries (words and phrases) with the learned classifier weights as polarity scores. Finally, it has been shown that sizable further performance gains can be obtained by starting with mid-sized seeds, i.e., hundreds of words and phrases BIBREF66 . Key Applications Sentiment analysis on Twitter has applications in a number of areas, including political science BIBREF39 , BIBREF40 , BIBREF41 , economics BIBREF42 , BIBREF7 , social science BIBREF43 , and market research BIBREF44 , BIBREF45 . It is used to study company reputation online BIBREF45 , to measure customer satisfaction, to identify detractors and promoters, to forecast market growth BIBREF42 , to predict the future income from newly-released movies, to forecast the outcome of upcoming elections BIBREF41 , BIBREF7 , to study political polarization BIBREF39 , BIBREF9 , etc. Future Directions We expect the quest for more interesting formulations of the general sentiment analysis task to continue. We see competitions such as those at SemEval as the engine of this innovation, as they not only perform head-to-head comparisons, but also create databases and tools that enable follow-up research for many years afterward. In terms of methods, we believe that deep learning BIBREF55 , BIBREF54 , BIBREF56 , together with semi-supervised and distantly-supervised methods BIBREF67 , BIBREF57 , will be the main focus of future research. We also expect more attention to be paid to linguistic structure and sentiment compositionality BIBREF68 , BIBREF69 . Moreover, we forecast more interest for languages other than English, and for cross-lingual methods BIBREF40 , BIBREF70 , BIBREF71 , which will allow leveraging on the rich resources that are already available for English. Last, but not least, the increase in opinion spam on Twitter will make it important to study astroturfing BIBREF72 and troll detection BIBREF73 , BIBREF74 , BIBREF75 . Cross-References Microblog Sentiment Analysis 100590 Multi-classifier System for Sentiment Analysis and Opinion Mining 351 Sentiment Analysis in Social Media 120 Sentiment Analysis of Microblogging Data 110168 Sentiment Analysis of Reviews 110169 Sentiment Analysis, Basics of 110159 Sentiment Quantification of User-Generated Content 110170 Social Media Analysis for Monitoring Political Sentiment 110172 Twitter Microblog Sentiment Analysis 265 User Sentiment and Opinion Analysis 192 Recommended Reading For general research on sentiment analysis, we recommend the following surveys: BIBREF76 and BIBREF15 . For sentiment analysis on Twitter, we recommend the overview article on Sentiment Analysis on Twitter about the SemEval task BIBREF28 as well as the task description papers for different editions of the task BIBREF30 , BIBREF26 , BIBREF29 , BIBREF27 .

Unanswerable

a3d83c2a1b98060d609e7ff63e00112d36ce2607

a3d83c2a1b98060d609e7ff63e00112d36ce2607_0

Q: How many sentence transformations on average are available per unique sentence in dataset? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

27.41 transformation on average of single seed sentence is available in dataset.

aeda22ae760de7f5c0212dad048e4984cd613162

aeda22ae760de7f5c0212dad048e4984cd613162_0

Q: What annotations are available in the dataset? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

For each source sentence, transformation sentences that are transformed according to some criteria (paraphrase, minimal change etc.)

d5fa26a2b7506733f3fa0973e2fe3fc1bbd1a12d

d5fa26a2b7506733f3fa0973e2fe3fc1bbd1a12d_0

Q: How are possible sentence transformations represented in dataset, as new sentences? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

Yes, as new sentences.

2d536961c6e1aec9f8491e41e383dc0aac700e0a

2d536961c6e1aec9f8491e41e383dc0aac700e0a_0

Q: What are all 15 types of modifications ilustrated in the dataset? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

- paraphrase 1 - paraphrase 2 - different meaning - opposite meaning - nonsense - minimal change - generalization - gossip - formal sentence - non-standard sentence - simple sentence - possibility - ban - future - past

18482658e0756d69e39a77f8fcb5912545a72b9b

18482658e0756d69e39a77f8fcb5912545a72b9b_0

Q: Is this dataset publicly available? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

Yes

9d336c4c725e390b6eba8bb8fe148997135ee981

9d336c4c725e390b6eba8bb8fe148997135ee981_0

Q: Are some baseline models trained on this dataset? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

Yes

016b59daa84269a93ce821070f4f5c1a71752a8a

016b59daa84269a93ce821070f4f5c1a71752a8a_0

Q: Do they do any analysis of of how the modifications changed the starting set of sentences? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

Yes

771b373d09e6eb50a74fffbf72d059ad44e73ab0

771b373d09e6eb50a74fffbf72d059ad44e73ab0_0

Q: How do they introduce language variation? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

we were looking for original and uncommon sentence change suggestions

efb52bda7366d2b96545cf927f38de27de3b5b77

efb52bda7366d2b96545cf927f38de27de3b5b77_0

Q: Do they use external resources to make modifications to sentences? Text: Introduction Vector representations are becoming truly essential in majority of natural language processing tasks. Word embeddings became widely popular with the introduction of word2vec BIBREF0 and GloVe BIBREF1 and their properties have been analyzed in length from various aspects. Studies of word embeddings range from word similarity BIBREF2, BIBREF3, over the ability to capture derivational relations BIBREF4, linear superposition of multiple senses BIBREF5, the ability to predict semantic hierarchies BIBREF6 or POS tags BIBREF7 up to data efficiency BIBREF8. Several studies BIBREF9, BIBREF10, BIBREF11, BIBREF12 show that word vector representations are capable of capturing meaningful syntactic and semantic regularities. These include, for example, male/female relation demonstrated by the pairs “man:woman”, “king:queen” and the country/capital relation (“Russia:Moscow”, “Japan:Tokyo”). These regularities correspond to simple arithmetic operations in the vector space. Sentence embeddings are becoming equally ubiquitous in NLP, with novel representations appearing almost every other week. With an overwhelming number of methods to compute sentence vector representations, the study of their general properties becomes difficult. Furthermore, it is not so clear in which way the embeddings should be evaluated. In an attempt to bring together more traditional representations of sentence meanings and the emerging vector representations, bojar:etal:jnle:representations:2019 introduce a number of aspects or desirable properties of sentence embeddings. One of them is denoted as “relatability”, which highlights the correspondence between meaningful differences between sentences and geometrical relations between their respective embeddings in the highly dimensional continuous vector space. If such a correspondence could be found, we could use geometrical operations in the space to induce meaningful changes in sentences. In this work, we present COSTRA, a new dataset of COmplex Sentence TRAnsformations. In its first version, the dataset is limited to sample sentences in Czech. The goal is to support studies of semantic and syntactic relations between sentences in the continuous space. Our dataset is the prerequisite for one of possible ways of exploring sentence meaning relatability: we envision that the continuous space of sentences induced by an ideal embedding method would exhibit topological similarity to the graph of sentence variations. For instance, one could argue that a subset of sentences could be organized along a linear scale reflecting the formalness of the language used. Another set of sentences could form a partially ordered set of gradually less and less concrete statements. And yet another set, intersecting both of the previous ones in multiple sentences could be partially or linearly ordered according to the strength of the speakers confidence in the claim. Our long term goal is to search for an embedding method which exhibits this behaviour, i.e. that the topological map of the embedding space corresponds to meaningful operations or changes in the set of sentences of a language (or more languages at once). We prefer this behaviour to emerge, as it happened for word vector operations, but regardless if the behaviour is emergent or trained, we need a dataset of sentences illustrating these patterns. If large enough, such a dataset could serve for training. If it will be smaller, it will provide a test set. In either case, these sentences could provide a “skeleton” to the continuous space of sentence embeddings. The paper is structured as follows: related summarizes existing methods of sentence embeddings evaluation and related work. annotation describes our methodology for constructing our dataset. data details the obtained dataset and some first observations. We conclude and provide the link to the dataset in conclusion Background As hinted above, there are many methods of converting a sequence of words into a vector in a highly dimensional space. To name a few: BiLSTM with the max-pooling trained for natural language inference BIBREF13, masked language modeling and next sentence prediction using bidirectional Transformer BIBREF14, max-pooling last states of neural machine translation among many languages BIBREF15 or the encoder final state in attentionless neural machine translation BIBREF16. The most common way of evaluating methods of sentence embeddings is extrinsic, using so called `transfer tasks', i.e. comparing embeddings via the performance in downstream tasks such as paraphrasing, entailment, sentence sentiment analysis, natural language inference and other assignments. However, even simple bag-of-words (BOW) approaches achieve often competitive results on such tasks BIBREF17. Adi16 introduce intrinsic evaluation by measuring the ability of models to encode basic linguistic properties of a sentence such as its length, word order, and word occurrences. These so called `probing tasks' are further extended by a depth of the syntactic tree, top constituent or verb tense by DBLP:journals/corr/abs-1805-01070. Both transfer and probing tasks are integrated in SentEval BIBREF18 framework for sentence vector representations. Later, Perone2018 applied SentEval to eleven different encoding methods revealing that there is no consistently well performing method across all tasks. SentEval was further criticized for pitfalls such as comparing different embedding sizes or correlation between tasks BIBREF19, BIBREF20. shi-etal-2016-string show that NMT encoder is able to capture syntactic information about the source sentence. DBLP:journals/corr/BelinkovDDSG17 examine the ability of NMT to learn morphology through POS and morphological tagging. Still, very little is known about semantic properties of sentence embeddings. Interestingly, cifka:bojar:meanings:2018 observe that the better self-attention embeddings serve in NMT, the worse they perform in most of SentEval tasks. zhu-etal-2018-exploring generate automatically sentence variations such as: Original sentence: A rooster pecked grain. Synonym Substitution: A cock pecked grain. Not-Negation: A rooster didn't peck grain. Quantifier-Negation: There was no rooster pecking grain. and compare their triplets by examining distances between their embeddings, i.e. distance between (1) and (2) should be smaller than distances between (1) and (3), (2) and (3), similarly, (3) and (4) should be closer together than (1)–(3) or (1)–(4). In our previous study BIBREF21, we examined the effect of small sentence alternations in sentence vector spaces. We used sentence pairs automatically extracted from datasets for natural language inference BIBREF22, BIBREF23 and observed, that the simple vector difference, familiar from word embeddings, serves reasonably well also in sentence embedding spaces. The examined relations were however very simple: a change of gender, number, addition of an adjective, etc. The structure of the sentence and its wording remained almost identical. We would like to move to more interesting non-trivial sentence comparison, beyond those in zhu-etal-2018-exploring or BaBo2019, such as change of style of a sentence, the introduction of a small modification that drastically changes the meaning of a sentence or reshuffling of words in a sentence that alters its meaning. Unfortunately, such a dataset cannot be generated automatically and it is not available to our best knowledge. We try to start filling this gap with COSTRA 1.0. Annotation We acquired the data in two rounds of annotation. In the first one, we were looking for original and uncommon sentence change suggestions. In the second one, we collected sentence alternations using ideas from the first round. The first and second rounds of annotation could be broadly called as collecting ideas and collecting data, respectively. Annotation ::: First Round: Collecting Ideas We manually selected 15 newspaper headlines. Eleven annotators were asked to modify each headline up to 20 times and describe the modification with a short name. They were given an example sentence and several of its possible alternations, see tab:firstroundexamples. Unfortunately, these examples turned out to be highly influential on the annotators' decisions and they correspond to almost two thirds of all of modifications gathered in the first round. Other very common transformations include change of a word order or transformation into a interrogative/imperative sentence. Other interesting modification were also proposed such as change into a fairy-tale style, excessive use of diminutives/vulgarisms or dadaism—a swap of roles in the sentence so that the resulting sentence is grammatically correct but nonsensical in our world. Of these suggestions, we selected only the dadaistic swap of roles for the current exploration (see nonsense in Table TABREF7). In total, we collected 984 sentences with 269 described unique changes. We use them as an inspiration for second round of annotation. Annotation ::: Second Round: Collecting Data ::: Sentence Transformations We selected 15 modifications types to collect COSTRA 1.0. They are presented in annotationinstructions. We asked for two distinct paraphrases of each sentence because we believe that a good sentence embedding should put paraphrases close together in vector space. Several modification types were specifically selected to constitute a thorough test of embeddings. In different meaning, the annotators should create a sentence with some other meaning using the same words as the original sentence. Other transformations which should be difficult for embeddings include minimal change, in which the sentence meaning should be significantly changed by using only very small modification, or nonsense, in which words of the source sentence should be shuffled so that it is grammatically correct, but without any sense. Annotation ::: Second Round: Collecting Data ::: Seed Data The source sentences for annotations were selected from Czech data of Global Voices BIBREF24 and OpenSubtitles BIBREF25. We used two sources in order to have different styles of seed sentences, both journalistic and common spoken language. We considered only sentences with more than 5 and less than 15 words and we manually selected 150 of them for further annotation. This step was necessary to remove sentences that are: too unreal, out of this world, such as: Jedno fotonový torpédo a je z tebe vesmírná topinka. “One photon torpedo and you're a space toast.” photo captions (i.e. incomplete sentences), e.g.: Zvláštní ekvádorský případ Correa vs. Crudo “Specific Ecuadorian case Correa vs. Crudo” too vague, overly dependent on the context: Běž tam a mluv na ni. “Go there and speak to her.” Many of the intended sentence transformations would be impossible to apply to such sentences and annotators' time would be wasted. Even after such filtering, it was still quite possible that a desired sentence modification could not be achieved for a sentence. For such a case, we gave the annotators the option to enter the keyword IMPOSSIBLE instead of the particular (impossible) modification. This option allowed to explicitly state that no such transformation is possible. At the same time most of the transformations are likely to lead to a large number possible outcomes. As documented in scratching2013, Czech sentence might have hundreds of thousand of paraphrases. To support some minimal exploration of this possible diversity, most of sentences were assigned to several annotators. Annotation ::: Second Round: Collecting Data ::: Spell-Checking The annotation is a challenging task and the annotators naturally make mistakes. Unfortunately, a single typo can significantly influence the resulting embedding BIBREF26. After collecting all the sentence variations, we applied the statistical spellchecker and grammar checker Korektor BIBREF27 in order to minimize influence of typos to performance of embedding methods. We manually inspected 519 errors identified by Korektor and fixed 129, which were identified correctly. Dataset Description In the second round, we collected 293 annotations from 12 annotators. After Korektor, there are 4262 unique sentences (including 150 seed sentences) that form the COSTRA 1.0 dataset. Statistics of individual annotators are available in tab:statistics. The time needed to carry out one piece of annotation (i.e. to provide one seed sentence with all 15 transformations) was on average almost 20 minutes but some annotators easily needed even half an hour. Out of the 4262 distinct sentences, only 188 was recorded more than once. In other words, the chance of two annotators producing the same output string is quite low. The most repeated transformations are by far past, future and ban. The least repeated is paraphrase with only single one repeated. multiple-annots documents this in another way. The 293 annotations are split into groups depending on how many annotators saw the same input sentence: 30 annotations were annotated by one person only, 30 annotations by two different persons etc. The last column shows the number of unique outputs obtained in that group. Across all cases, 96.8% of produced strings were unique. In line with instructions, the annotators were using the IMPOSSIBLE option scarcely (95 times, i.e. only 2%). It was also a case of 7 annotators only; the remaining 5 annotators were capable of producing all requested transformations. The top three transformations considered unfeasible were different meaning (using the same set of words), past (esp. for sentences already in the past tense) and simple sentence. Dataset Description ::: First Observations We embedded COSTRA sentences with LASER BIBREF15, the method that performed very well in revealing linear relations in BaBo2019. Having browsed a number of 2D visualizations (PCA and t-SNE) of the space, we have to conclude that visually, LASER space does not seem to exhibit any of the desired topological properties discussed above, see fig:pca for one example. The lack of semantic relations in the LASER space is also reflected in vector similarities, summarized in similarities. The minimal change operation substantially changed the meaning of the sentence, and yet the embedding of the transformation lies very closely to the original sentence (average similarity of 0.930). Tense changes and some form of negation or banning also keep the vectors very similar. The lowest average similarity was observed for generalization (0.739) and simplification (0.781), which is not any bad sign. However the fact that paraphrases have much smaller similarity (0.826) than opposite meaning (0.902) documents that the vector space lacks in terms of “relatability”. Conclusion and Future Work We presented COSTRA 1.0, a small corpus of complex transformations of Czech sentences. We plan to use this corpus to analyze a wide spectrum sentence embeddings methods to see to what extent the continuous space they induce reflects semantic relations between sentences in our corpus. The very first analysis using LASER embeddings indicates lack of “meaning relatability”, i.e. the ability to move along a trajectory in the space in order to reach desired sentence transformations. Actually, not even paraphrases are found in close neighbourhoods of embedded sentences. More “semantic” sentence embeddings methods are thus to be sought for. The corpus is freely available at the following link: http://hdl.handle.net/11234/1-3123 Aside from extending the corpus in Czech and adding other language variants, we are also considering to wrap COSTRA 1.0 into an API such as SentEval, so that it is very easy for researchers to evaluate their sentence embeddings in terms of “relatability”.

No

1a7d28c25bb7e7202230e1b70a885a46dac8a384

1a7d28c25bb7e7202230e1b70a885a46dac8a384_0

Q: How big is dataset domain-specific embedding are trained on? Text: Introduction Today's increasing flood of information on the web creates a need for automated multi-document summarization systems that produce high quality summaries. However, producing summaries in a multi-document setting is difficult, as the language used to display the same information in a sentence can vary significantly, making it difficult for summarization models to capture. Given the complexity of the task and the lack of datasets, most researchers use extractive summarization, where the final summary is composed of existing sentences in the input documents. More specifically, extractive summarization systems output summaries in two steps : via sentence ranking, where an importance score is assigned to each sentence, and via the subsequent sentence selection, where the most appropriate sentence is chosen, by considering 1) their importance and 2) their frequency among all documents. Due to data sparcity, models heavily rely on well-designed features at the word level BIBREF0, BIBREF1, BIBREF2, BIBREF3 or take advantage of other large, manually annotated datasets and then apply transfer learning BIBREF4. Additionally, most of the time, all sentences in the same collection of documents are processed independently and therefore, their relationships are lost. In realistic scenarios, features are hard to craft, gathering additional annotated data is costly, and the large variety in expressing the same fact cannot be handled by the use of word-based features only, as is often the case. In this paper, we address these obstacles by proposing to simultaneously leverage two types of sentence embeddings, namely embeddings pre-trained on a large corpus that capture a variety of meanings and domain-specific embeddings learned during training. The former is typically trained on an unrelated corpus composed of high quality texts, allowing to cover additional contexts for each encountered word and sentence. Hereby, we build on the assumption that sentence embeddings capture both the syntactic and semantic content of sentences. We hypothesize that using two types of sentence embeddings, general and domain-specific, is beneficial for the task of multi-document summarization, as the former captures the most common semantic structures from a large, general corpus, while the latter captures the aspects related to the domain. We present SemSentSum (Figure FIGREF3), a fully data-driven summarization system, which does not depend on hand-crafted features, nor additional data, and is thus domain-independent. It first makes use of general sentence embedding knowledge to build a sentenc semantic relation graph that captures sentence similarities (Section SECREF4). In a second step, it trains genre-specific sentence embeddings related to the domains of the collection of documents, by utilizing a sentence encoder (Section SECREF5). Both representations are afterwards merged, by using a graph convolutional network BIBREF5 (Section SECREF6). Then, it employs a linear layer to project high-level hidden features for individual sentences to salience scores (Section SECREF8). Finally, it greedily produces relevant and non-redundant summaries by using sentence embeddings to detect similarities between candidate sentences and the current summary (Section SECREF11). The main contributions of this work are as follows : We aggregate two types of sentences embeddings using a graph representation. They share different properties and are consequently complementary. The first one is trained on a large unrelated corpus to model general semantics among sentences, whereas the second is domain-specific to the dataset and learned during training. Together, they enable a model to be domain-independent as it can be applied easily on other domains. Moreover, it could be used for other tasks including detecting information cascades, query-focused summarization, keyphrase extraction and information retrieval. We devise a competitive multi-document summarization system, which does not need hand-crafted features nor additional annotated data. Moreover, the results are competitive for 665-byte and 100-word summaries. Usually, models are compared in one of the two settings but not both and thus lack comparability. Method Let $C$ denote a collection of related documents composed of a set of documents $\lbrace D_i|i \in [1,N]\rbrace $ where $N$ is the number of documents. Moreover, each document $D_i$ consists of a set of sentences $\lbrace S_{i,j}|j \in [1,M]\rbrace $, $M$ being the number of sentences in $D_i$. Given a collection of related documents $C$, our goal is to produce a summary $Sum$ using a subset of these in the input documents ordered in some way, such that $Sum = (S_{i_1,j_1},S_{i_2,j_2},...,S_{i_n,j_m})$. In this section, we describe how SemSentSum estimates the salience score of each sentence and how it selects a subset of these to create the final summary. The architecture of SemSentSum is depicted in Figure FIGREF3. In order to perform sentence selection, we first build our sentence semantic relation graph, where each vertex is a sentence and edges capture the semantic similarity among them. At the same time, each sentence is fed into a recurrent neural network, as a sentence encoder, to generate sentence embeddings using the last hidden states. A single-layer graph convolutional neural network is then applied on top, where the sentence semantic relation graph is the adjacency matrix and the sentence embeddings are the node features. Afterward, a linear layer is used to project high-level hidden features for individual sentences to salience scores, representing how salient a sentence is with respect to the final summary. Finally, based on this, we devise an innovative greedy method that leverages sentence embeddings to detect redundant sentences and select sentences until reaching the summary length limit. Method ::: Sentence Semantic Relation Graph We model the semantic relationship among sentences using a graph representation. In this graph, each vertex is a sentence $S_{i,j}$ ($j$'th sentence of document $D_i$) from the collection documents $C$ and an undirected edge between $S_{i_u,j_u}$ and $S_{i_v,j_v}$ indicates their degree of similarity. In order to compute the semantic similarity, we use the model of BIBREF6 trained on the English Wikipedia corpus. In this manner, we incorporate general knowledge (i.e. not domain-specific) that will complete the specialized sentence embeddings obtained during training (see Section SECREF5). We process sentences by their model and compute the cosine similarity between every sentence pair, resulting in a complete graph. However, having a complete graph alone does not allow the model to leverage the semantic structure across sentences significantly, as every sentence pair is connected, and likewise, a sparse graph does not contain enough information to exploit semantic similarities. Furthermore, all edges have a weight above zero, since it is very unlikely that two sentence embeddings are completely orthogonal. To overcome this problem, we introduce an edge-removal-method, where every edge below a certain threshold $t_{sim}^g$ is removed in order to emphasize high sentence similarity. Nonetheless, $t_{sim}^g$ should not be too large, as we otherwise found the model to be prone to overfitting. After removing edges below $t_{sim}^g$, our sentence semantic relation graph is used as the adjacency matrix $A$. The impact of $t_{sim}^g$ with different values is shown in Section SECREF26. Based on our aforementioned hypothesis that a combination of general and genre-specific sentence embeddings is beneficial for the task of multi-document summarization, we further incorporate general sentence embeddings, pre-trained on Wikipedia entries, into edges between sentences. Additionally, we compute specialised sentence embeddings, which are related to the domains of the documents (see Section SECREF35). Note that 1) the pre-trained sentence embeddings are only used to compute the weights of the edges and are not used by the summarization model (as others are produced by the sentence encoder) and 2) the edge weights are static and do not change during training. Method ::: Sentence Encoder Given a list of documents $C$, we encode each document's sentence $S_{i,j}$, where each has at most $L$ words $(w_{i,j,1}, w_{i,j,2}, ..., w_{i,j,L})$. In our experiments, all words are kept and converted into word embeddings, which are then fed to the sentence encoder in order to compute specialized sentence embeddings $S^{\prime }_{i,j}$. We employ a single-layer forward recurrent neural network, using Long Short-Term Memory (LSTM) of BIBREF7 as sentence encoder, where the sentence embeddings are extracted from the last hidden states. We then concatenate all sentence embeddings into a matrix $X$ which constitutes the input node features that will be used by the graph convolutional network. Method ::: Graph Convolutional Network After having computed all sentence embeddings and the sentence semantic relation graph, we apply a single-layer Graph Convolutional Network (GCN) from BIBREF5, in order to capture high-level hidden features for each sentence, encapsulating sentence information as well as the graph structure. We believe that our sentence semantic relation graph contains information not present in the data (via universal embeddings) and thus, we leverage this information by running a graph convolution on the first order neighborhood. The GCN model takes as input the node features matrix $X$ and a squared adjacency matrix $A$. The former contains all sentence embeddings of the collection of documents, while the latter is our underlying sentence semantic relation graph. It outputs hidden representations for each node that encode both local graph structure and nodes's features. In order to take into account the sentences themselves during the information propagation, we add self-connections (i.e. the identity matrix) to $A$ such that $\tilde{A} = A + I$. Subsequently, we obtain our sentence hidden features by using Equation DISPLAY_FORM7. where $W_i$ is the weight matrix of the $i$'th graph convolution layer and $b_i$ the bias vector. We choose the Exponential Linear Unit (ELU) activation function from BIBREF8 due to its ability to handle the vanishing gradient problem, by pushing the mean unit activations close to zero and consequently facilitating the backpropagation. By using only one hidden layer, as we only have one input-to-hidden layer and one hidden-to-output layer, we limit the information propagation to the first order neighborhood. Method ::: Saliency Estimation We use a simple linear layer to estimate a salience score for each sentence and then normalize the scores via softmax and obtain our estimated salience score $S^s_{i,j}$. Method ::: Training Our model SemSentSum is trained in an end-to-end manner and minimizes the cross-entropy loss of Equation DISPLAY_FORM10 between the salience score prediction and the ROUGE-1 $F_1$ score for each sentence. $F_1(S)$ is computed as the ROUGE-1 $F_1$ score, unlike the common practice in the area of single and multi-document summarization as recall favors longer sentences whereas $F_1$ prevents this tendency. The scores are normalized via softmax. Method ::: Summary Generation Process While our model SemSentSum provides estimated saliency scores, we use a greedy strategy to construct an informative and non-redundant summary $Sum$. We first discard sentences having less than 9 words, as in BIBREF9, and then sort them in descending order of their estimated salience scores. We iteratively dequeue the sentence having the highest score and append it to the current summary $Sum$ if it is non-redundant with respect to the current content of $Sum$. We iterate until reaching the summary length limit. To determine the similarity of a candidate sentence with the current summary, a sentence is considered as dissimilar if and only if the cosine similarity between its sentence embeddings and the embeddings of the current summary is below a certain threshold $t_{sim}^s$. We use the pre-trained model of BIBREF6 to compute sentence as well as summary embeddings, similarly to the sentence semantic relation graph construction. Our approach is novel, since it focuses on the semantic sentence structures and captures similarity between sentence meanings, instead of focusing on word similarities only, like previous TF-IDF approaches ( BIBREF0, BIBREF1, BIBREF3, BIBREF4). Experiments ::: Datasets We conduct experiments on the most commonly used datasets for multi-document summarization from the Document Understanding Conferences (DUC). We use DUC 2001, 2002, 2003 and 2004 as the tasks of generic multi-document summarization, because they have been carried out during these years. We use DUC 2001, 2002, 2003 and 2004 for generic multi-document summarization, where DUC 2001/2002 are used for training, DUC 2003 for validation and finally, DUC 2004 for testing, following the common practice. Experiments ::: Evaluation Metric For the evaluation, we use ROUGE BIBREF10 with the official parameters of the DUC tasks and also truncate the summaries to 100 words for DUC 2001/2002/2003 and to 665 bytes for DUC 2004. Notably, we take ROUGE-1 and ROUGE-2 recall scores as the main metrics for comparison between produced summaries and golden ones as proposed by BIBREF11. The goal of the ROUGE-N metric is to compute the ratio of the number of N-grams from the generated summary matching these of the human reference summaries. Experiments ::: Model Settings To define the edge weights of our sentence semantic relation graph, we employ the 600-dimensional pre-trained unigram model of BIBREF6, using English Wikipedia as source corpus. We keep only edges having a weight larger than $t_{sim}^g = 0.5$ (tuned on the validation set). For word embeddings, the 300-dimensional pre-trained GloVe embeddings BIBREF12 are used and fixed during training. The output dimension of the sentence embeddings produced by the sentence encoder is the same as that of the word embeddings, i.e. 300. For the graph convolutional network, the number of hidden units is 128 and the size of the generated hidden feature vectors is also 300. We use a batch size of 1, a learning rate of $0.0075$ using Adam optimizer BIBREF13 with $\beta _1=0.9, \beta _2=0.999$ and $\epsilon =10^{-8}$. In order to make SemSentSum generalize better, we use dropout BIBREF14 of $0.2$, batch normalization BIBREF15, clip the gradient norm at $1.0$ if higher, add L2-norm regularizer with a regularization factor of $10^{-12}$ and train using early stopping with a patience of 10 iterations. Finally, the similarity threshold $t_{sim}^s$ in the summary generation process is $0.8$ (tuned on the validation set). Experiments ::: Summarization Performance We train our model SemSentSum on DUC 2001/2002, tune it on DUC 2003 and assess the performance on DUC 2004. In order to fairly compare SemSentSum with other models available in the literature, experiments are conducted with summaries truncated to 665 bytes (official summary length in the DUC competition), but also with summaries with a length constraint of 100 words. To the best of our knowledge, we are the first to conduct experiments on both summary lengths and compare our model with other systems producing either 100 words or 665 bytes summaries. Experiments ::: Sentence Semantic Relation Graph Construction We investigate different methods to build our sentence semantic relation graph and vary the value of $t_{sim}^g$ from $0.0$ to $0.75$ to study the impact of the threshold cut-off. Among these are : Cosine : Using cosine similarity ; Tf-idf : Considering a node as the query and another as document. The weight corresponds to the cosine similarity between the query and the document ; TextRank BIBREF16 : A weighted graph is created where nodes are sentences and edges defined by a similarity measure based on word overlap. Afterward, an algorithm similar to PageRank BIBREF17 is used to compute sentence importance and refined edge weights ; LexRank BIBREF9 : An unsupervised multi-document summarizer based on the concept of eigenvector centrality in a graph of sentences to set up the edge weights ; Approximate Discourse Graph (ADG) BIBREF2 : Approximation of a discourse graph where nodes are sentences and edges $(S_u,S_v)$ indicates sentence $S_v$ can be placed after $S_u$ in a coherent summary ; Personalized ADG (PADG) BIBREF3 : Normalized version of ADG where sentence nodes are normalized over all edges. Experiments ::: Ablation Study In order to quantify the contribution of the different components of SemSentSum, we try variations of our model by removing different modules one at a time. Our two main elements are the sentence encoder (Sent) and the graph convolutional neural network (GCN). When we omit Sent, we substitute it with the pre-trained sentence embeddings used to build our sentence semantic relation graph. Experiments ::: Results and Discussion Three dimensions are used to evaluate our model SemSentSum : 1) the summarization performance, to assess its capability 2) the impact of the sentence semantic relation graph generation using various methods and different thresholds $t_{sim}^g$ 3) an ablation study to analyze the importance of each component of SemSentSum. Experiments ::: Results and Discussion ::: Summarization Performance We compare the results of SemSentSum for both settings : 665 bytes and 100 words summaries. We only include models using the same parameters to compute the ROUGE-1/ROUGE-2 score and recall as metrics. The results for 665 bytes summaries are reported in Table TABREF28. We compare SemSentSum with three types of model relying on either 1) sentence or document embeddings 2) various hand-crafted features or 3) additional data. For the first category, we significantly outperform MMR BIBREF18, PV-DBOW+BS BIBREF19 and PG-MMR BIBREF20. Although their methods are based on embeddings to represent the meaning, it shows that using only various distance metrics or encoder-decoder architecture on these is not efficient for the task of multi-document summarization (as also shown in the Ablation Study). We hypothesize that SemSentSum performs better by leveraging pre-trained sentence embeddings and hence lowering the effects of data scarcity. Systems based on hand-crafted features include a widely-used learning-based summarization method, built on support vector regression SVR BIBREF21 ; a graph-based method based on approximating discourse graph G-Flow BIBREF2 ; Peer 65 which is the best peer systems participating in DUC evaluations ; and the recursive neural network R2N2 of BIBREF1 that learns automatically combinations of hand-crafted features. As can be seen, among these models completely dependent on hand-crafted features, SemSentSum achieves highest performance on both ROUGE scores. This denotes that using different linguistic and word-based features might not be enough to capture the semantic structures, in addition to being cumbersome to craft. The last type of model is shown in TCSum BIBREF4 and uses transfer learning from a text classifier model, based on a domain-related dataset of $30\,000$ documents from New York Times (sharing the same topics of the DUC datasets). In terms of ROUGE-1, SemSentSum significantly outperforms TCSum and performs similarly on ROUGE-2 score. This demonstrates that collecting more manually annotated data and training two models is unnecessary, in addition to being difficult to use in other domains, whereas SemSentSum is fully data driven, domain-independent and usable in realistic scenarios. Table TABREF32 depicts models producing 100 words summaries, all depending on hand-crafted features. We use as baselines FreqSum BIBREF22 ; TsSum BIBREF23 ; traditional graph-based approaches such as Cont. LexRank BIBREF9 ; Centroid BIBREF24 ; CLASSY04 BIBREF25 ; its improved version CLASSY11 BIBREF26 and the greedy model GreedyKL BIBREF27. All of these models are significantly underperforming compared to SemSentSum. In addition, we include state-of-the-art models : RegSum BIBREF0 and GCN+PADG BIBREF3. We outperform both in terms of ROUGE-1. For ROUGE-2 scores we achieve better results than GCN+PADG but without any use of domain-specific hand-crafted features and a much smaller and simpler model. Finally, RegSum achieves a similar ROUGE-2 score but computes sentence saliences based on word scores, incorporating a rich set of word-level and domain-specific features. Nonetheless, our model is competitive and does not depend on hand-crafted features due to its full data-driven nature and thus, it is not limited to a single domain. Consequently, the experiments show that achieving good performance for multi-document summarization without hand-crafted features or additional data is clearly feasible and SemSentSum produces competitive results without depending on these, is domain independent, fast to train and thus usable in real scenarios. Experiments ::: Results and Discussion ::: Sentence Semantic Relation Graph Table TABREF34 shows the results of different methods to create the sentence semantic relation graph with various thresholds $t_{sim}^g$ for 665 bytes summaries (we obtain similar results for 100 words). A first observation is that using cosine similarity with sentence embeddings significantly outperforms all other methods for ROUGE-1 and ROUGE-2 scores, mainly because it relies on the semantic of sentences instead of their individual words. A second is that different methods evolve similarly : PADG, Textrank, Tf-idf behave similarly to an U-shaped curve for both ROUGE scores while Cosine is the only one having an inverted U-shaped curve. The reason for this behavior is a consequence of its distribution being similar to a normal distribution because it relies on the semantic instead of words, while the others are more skewed towards zero. This confirms our hypothesis that 1) having a complete graph does not allow the model to leverage much the semantic 2) a sparse graph might not contain enough information to exploit similarities. Finally, Lexrank and ADG have different trends between both ROUGE scores. Experiments ::: Results and Discussion ::: Ablation Study We quantify the contribution of each module of SemSentSum in Table TABREF36 for 665 bytes summaries (we obtain similar results for 100 words). Removing the sentence encoder produces slightly lower results. This shows that the sentence semantic relation graph captures semantic attributes well, while the fine-tuned sentence embeddings obtained via the encoder help boost the performance, making these methods complementary. By disabling only the graph convolutional layer, a drastic drop in terms of performance is observed, which emphasizes that the relationship among sentences is indeed important and not present in the data itself. Therefore, our sentence semantic relation graph is able to capture sentence similarities by analyzing the semantic structures. Interestingly, if we remove the sentence encoder in addition to the graph convolutional layer, similar results are achieved. This confirms that alone, the sentence encoder is not able to compute an efficient representation of sentences for the task of multi-document summarization, probably due to the poor size of the DUC datasets. Finally, we can observe that the use of sentence embeddings only results in similar performance to the baselines, which rely on sentence or document embeddings BIBREF18, BIBREF19. Related Work The idea of using multiple embeddings has been employed at the word level. BIBREF28 use an attention mechanism to combine the embeddings for each word for the task of natural language inference. BIBREF29, BIBREF30 concatenate the embeddings of each word into a vector before feeding a neural network for the tasks of aspect extraction and sentiment analysis. To our knowledge, we are the first to combine multiple types of sentence embeddings. Extractive multi-document summarization has been addressed by a large range of approaches. Several of them employ graph-based methods. BIBREF31 introduced a cross-document structure theory, as a basis for multi-document summarization. BIBREF9 proposed LexRank, an unsupervised multi-document summarizer based on the concept of eigenvector centrality in a graph of sentences. Other works exploit shallow or deep features from the graph's topology BIBREF32, BIBREF33. BIBREF34 pairs graph-based methods (e.g. random walk) with clustering. BIBREF35 improved results by using a reinforced random walk model to rank sentences and keep non-redundant ones. The system by BIBREF2 does sentence selection, while balancing coherence and salience and by building a graph that approximates discourse relations across sentences BIBREF36. Besides graph-based methods, other viable approaches include Maximum Marginal Relevance BIBREF37, which uses a greedy approach to select sentences and considers the tradeoff between relevance and redundancy ; support vector regression BIBREF21 ; conditional random field BIBREF38 ; or hidden markov model BIBREF25. Yet other approaches rely on n-grams regression as in BIBREF39. More recently, BIBREF1 built a recursive neural network, which tries to automatically detect combination of hand-crafted features. BIBREF4 employ a neural model for text classification on a large manually annotated dataset and apply transfer learning for multi-document summarization afterward. The work most closely related to ours is BIBREF3. They create a normalized version of the approximate discourse graph BIBREF2, based on hand-crafted features, where sentence nodes are normalized over all the incoming edges. They then employ a deep neural network, composed of a sentence encoder, three graph convolutional layers, one document encoder and an attention mechanism. Afterward, they greedily select sentences using TF-IDF similarity to detect redundant sentences. Our model differs in four ways : 1) we build our sentence semantic relation graph by using pre-trained sentence embeddings with cosine similarity, where neither heavy preprocessing, nor hand-crafted features are necessary. Thus, our model is fully data-driven and domain-independent unlike other systems. In addition, the sentence semantic relation graph could be used for other tasks than multi-document summarization, such as detecting information cascades, query-focused summarization, keyphrase extraction or information retrieval, as it is not composed of hand-crafted features. 2) SemSentSum is much smaller and consequently has fewer parameters as it only uses a sentence encoder and a single convolutional layer. 3) The loss function is based on ROUGE-1 $F_1$ score instead of recall to prevent the tendency of choosing longer sentences. 4) Our method for summary generation is also different and novel as we leverage sentence embeddings to compute the similarity between a candidate sentence and the current summary instead of TF-IDF based approaches. Conclusion In this work, we propose a method to combine two types of sentence embeddings : 1) universal embeddings, pre-trained on a large corpus such as Wikipedia and incorporating general semantic structures across sentences and 2) domain-specific embeddings, learned during training. We merge them together by using a graph convolutional network that eliminates the need of hand-crafted features or additional annotated data. We introduce a fully data-driven model SemSentSum that achieves competitive results for multi-document summarization on both kind of summary length (665 bytes and 100 words summaries), without requiring hand-crafted features or additional annotated data. As SemSentSum is domain-independent, we believe that our sentence semantic relation graph and model can be used for other tasks including detecting information cascades, query-focused summarization, keyphrase extraction and information retrieval. In addition, we plan to leave the weights of the sentence semantic relation graph dynamic during training, and to integrate an attention mechanism directly into the graph. Acknowledgments We thank Michaela Benk for proofreading and helpful advice.

Unanswerable

6bc45d4f908672945192390642da5a2760971c40

6bc45d4f908672945192390642da5a2760971c40_0

Q: How big is unrelated corpus universal embedding is traned on? Text: Introduction Today's increasing flood of information on the web creates a need for automated multi-document summarization systems that produce high quality summaries. However, producing summaries in a multi-document setting is difficult, as the language used to display the same information in a sentence can vary significantly, making it difficult for summarization models to capture. Given the complexity of the task and the lack of datasets, most researchers use extractive summarization, where the final summary is composed of existing sentences in the input documents. More specifically, extractive summarization systems output summaries in two steps : via sentence ranking, where an importance score is assigned to each sentence, and via the subsequent sentence selection, where the most appropriate sentence is chosen, by considering 1) their importance and 2) their frequency among all documents. Due to data sparcity, models heavily rely on well-designed features at the word level BIBREF0, BIBREF1, BIBREF2, BIBREF3 or take advantage of other large, manually annotated datasets and then apply transfer learning BIBREF4. Additionally, most of the time, all sentences in the same collection of documents are processed independently and therefore, their relationships are lost. In realistic scenarios, features are hard to craft, gathering additional annotated data is costly, and the large variety in expressing the same fact cannot be handled by the use of word-based features only, as is often the case. In this paper, we address these obstacles by proposing to simultaneously leverage two types of sentence embeddings, namely embeddings pre-trained on a large corpus that capture a variety of meanings and domain-specific embeddings learned during training. The former is typically trained on an unrelated corpus composed of high quality texts, allowing to cover additional contexts for each encountered word and sentence. Hereby, we build on the assumption that sentence embeddings capture both the syntactic and semantic content of sentences. We hypothesize that using two types of sentence embeddings, general and domain-specific, is beneficial for the task of multi-document summarization, as the former captures the most common semantic structures from a large, general corpus, while the latter captures the aspects related to the domain. We present SemSentSum (Figure FIGREF3), a fully data-driven summarization system, which does not depend on hand-crafted features, nor additional data, and is thus domain-independent. It first makes use of general sentence embedding knowledge to build a sentenc semantic relation graph that captures sentence similarities (Section SECREF4). In a second step, it trains genre-specific sentence embeddings related to the domains of the collection of documents, by utilizing a sentence encoder (Section SECREF5). Both representations are afterwards merged, by using a graph convolutional network BIBREF5 (Section SECREF6). Then, it employs a linear layer to project high-level hidden features for individual sentences to salience scores (Section SECREF8). Finally, it greedily produces relevant and non-redundant summaries by using sentence embeddings to detect similarities between candidate sentences and the current summary (Section SECREF11). The main contributions of this work are as follows : We aggregate two types of sentences embeddings using a graph representation. They share different properties and are consequently complementary. The first one is trained on a large unrelated corpus to model general semantics among sentences, whereas the second is domain-specific to the dataset and learned during training. Together, they enable a model to be domain-independent as it can be applied easily on other domains. Moreover, it could be used for other tasks including detecting information cascades, query-focused summarization, keyphrase extraction and information retrieval. We devise a competitive multi-document summarization system, which does not need hand-crafted features nor additional annotated data. Moreover, the results are competitive for 665-byte and 100-word summaries. Usually, models are compared in one of the two settings but not both and thus lack comparability. Method Let $C$ denote a collection of related documents composed of a set of documents $\lbrace D_i|i \in [1,N]\rbrace $ where $N$ is the number of documents. Moreover, each document $D_i$ consists of a set of sentences $\lbrace S_{i,j}|j \in [1,M]\rbrace $, $M$ being the number of sentences in $D_i$. Given a collection of related documents $C$, our goal is to produce a summary $Sum$ using a subset of these in the input documents ordered in some way, such that $Sum = (S_{i_1,j_1},S_{i_2,j_2},...,S_{i_n,j_m})$. In this section, we describe how SemSentSum estimates the salience score of each sentence and how it selects a subset of these to create the final summary. The architecture of SemSentSum is depicted in Figure FIGREF3. In order to perform sentence selection, we first build our sentence semantic relation graph, where each vertex is a sentence and edges capture the semantic similarity among them. At the same time, each sentence is fed into a recurrent neural network, as a sentence encoder, to generate sentence embeddings using the last hidden states. A single-layer graph convolutional neural network is then applied on top, where the sentence semantic relation graph is the adjacency matrix and the sentence embeddings are the node features. Afterward, a linear layer is used to project high-level hidden features for individual sentences to salience scores, representing how salient a sentence is with respect to the final summary. Finally, based on this, we devise an innovative greedy method that leverages sentence embeddings to detect redundant sentences and select sentences until reaching the summary length limit. Method ::: Sentence Semantic Relation Graph We model the semantic relationship among sentences using a graph representation. In this graph, each vertex is a sentence $S_{i,j}$ ($j$'th sentence of document $D_i$) from the collection documents $C$ and an undirected edge between $S_{i_u,j_u}$ and $S_{i_v,j_v}$ indicates their degree of similarity. In order to compute the semantic similarity, we use the model of BIBREF6 trained on the English Wikipedia corpus. In this manner, we incorporate general knowledge (i.e. not domain-specific) that will complete the specialized sentence embeddings obtained during training (see Section SECREF5). We process sentences by their model and compute the cosine similarity between every sentence pair, resulting in a complete graph. However, having a complete graph alone does not allow the model to leverage the semantic structure across sentences significantly, as every sentence pair is connected, and likewise, a sparse graph does not contain enough information to exploit semantic similarities. Furthermore, all edges have a weight above zero, since it is very unlikely that two sentence embeddings are completely orthogonal. To overcome this problem, we introduce an edge-removal-method, where every edge below a certain threshold $t_{sim}^g$ is removed in order to emphasize high sentence similarity. Nonetheless, $t_{sim}^g$ should not be too large, as we otherwise found the model to be prone to overfitting. After removing edges below $t_{sim}^g$, our sentence semantic relation graph is used as the adjacency matrix $A$. The impact of $t_{sim}^g$ with different values is shown in Section SECREF26. Based on our aforementioned hypothesis that a combination of general and genre-specific sentence embeddings is beneficial for the task of multi-document summarization, we further incorporate general sentence embeddings, pre-trained on Wikipedia entries, into edges between sentences. Additionally, we compute specialised sentence embeddings, which are related to the domains of the documents (see Section SECREF35). Note that 1) the pre-trained sentence embeddings are only used to compute the weights of the edges and are not used by the summarization model (as others are produced by the sentence encoder) and 2) the edge weights are static and do not change during training. Method ::: Sentence Encoder Given a list of documents $C$, we encode each document's sentence $S_{i,j}$, where each has at most $L$ words $(w_{i,j,1}, w_{i,j,2}, ..., w_{i,j,L})$. In our experiments, all words are kept and converted into word embeddings, which are then fed to the sentence encoder in order to compute specialized sentence embeddings $S^{\prime }_{i,j}$. We employ a single-layer forward recurrent neural network, using Long Short-Term Memory (LSTM) of BIBREF7 as sentence encoder, where the sentence embeddings are extracted from the last hidden states. We then concatenate all sentence embeddings into a matrix $X$ which constitutes the input node features that will be used by the graph convolutional network. Method ::: Graph Convolutional Network After having computed all sentence embeddings and the sentence semantic relation graph, we apply a single-layer Graph Convolutional Network (GCN) from BIBREF5, in order to capture high-level hidden features for each sentence, encapsulating sentence information as well as the graph structure. We believe that our sentence semantic relation graph contains information not present in the data (via universal embeddings) and thus, we leverage this information by running a graph convolution on the first order neighborhood. The GCN model takes as input the node features matrix $X$ and a squared adjacency matrix $A$. The former contains all sentence embeddings of the collection of documents, while the latter is our underlying sentence semantic relation graph. It outputs hidden representations for each node that encode both local graph structure and nodes's features. In order to take into account the sentences themselves during the information propagation, we add self-connections (i.e. the identity matrix) to $A$ such that $\tilde{A} = A + I$. Subsequently, we obtain our sentence hidden features by using Equation DISPLAY_FORM7. where $W_i$ is the weight matrix of the $i$'th graph convolution layer and $b_i$ the bias vector. We choose the Exponential Linear Unit (ELU) activation function from BIBREF8 due to its ability to handle the vanishing gradient problem, by pushing the mean unit activations close to zero and consequently facilitating the backpropagation. By using only one hidden layer, as we only have one input-to-hidden layer and one hidden-to-output layer, we limit the information propagation to the first order neighborhood. Method ::: Saliency Estimation We use a simple linear layer to estimate a salience score for each sentence and then normalize the scores via softmax and obtain our estimated salience score $S^s_{i,j}$. Method ::: Training Our model SemSentSum is trained in an end-to-end manner and minimizes the cross-entropy loss of Equation DISPLAY_FORM10 between the salience score prediction and the ROUGE-1 $F_1$ score for each sentence. $F_1(S)$ is computed as the ROUGE-1 $F_1$ score, unlike the common practice in the area of single and multi-document summarization as recall favors longer sentences whereas $F_1$ prevents this tendency. The scores are normalized via softmax. Method ::: Summary Generation Process While our model SemSentSum provides estimated saliency scores, we use a greedy strategy to construct an informative and non-redundant summary $Sum$. We first discard sentences having less than 9 words, as in BIBREF9, and then sort them in descending order of their estimated salience scores. We iteratively dequeue the sentence having the highest score and append it to the current summary $Sum$ if it is non-redundant with respect to the current content of $Sum$. We iterate until reaching the summary length limit. To determine the similarity of a candidate sentence with the current summary, a sentence is considered as dissimilar if and only if the cosine similarity between its sentence embeddings and the embeddings of the current summary is below a certain threshold $t_{sim}^s$. We use the pre-trained model of BIBREF6 to compute sentence as well as summary embeddings, similarly to the sentence semantic relation graph construction. Our approach is novel, since it focuses on the semantic sentence structures and captures similarity between sentence meanings, instead of focusing on word similarities only, like previous TF-IDF approaches ( BIBREF0, BIBREF1, BIBREF3, BIBREF4). Experiments ::: Datasets We conduct experiments on the most commonly used datasets for multi-document summarization from the Document Understanding Conferences (DUC). We use DUC 2001, 2002, 2003 and 2004 as the tasks of generic multi-document summarization, because they have been carried out during these years. We use DUC 2001, 2002, 2003 and 2004 for generic multi-document summarization, where DUC 2001/2002 are used for training, DUC 2003 for validation and finally, DUC 2004 for testing, following the common practice. Experiments ::: Evaluation Metric For the evaluation, we use ROUGE BIBREF10 with the official parameters of the DUC tasks and also truncate the summaries to 100 words for DUC 2001/2002/2003 and to 665 bytes for DUC 2004. Notably, we take ROUGE-1 and ROUGE-2 recall scores as the main metrics for comparison between produced summaries and golden ones as proposed by BIBREF11. The goal of the ROUGE-N metric is to compute the ratio of the number of N-grams from the generated summary matching these of the human reference summaries. Experiments ::: Model Settings To define the edge weights of our sentence semantic relation graph, we employ the 600-dimensional pre-trained unigram model of BIBREF6, using English Wikipedia as source corpus. We keep only edges having a weight larger than $t_{sim}^g = 0.5$ (tuned on the validation set). For word embeddings, the 300-dimensional pre-trained GloVe embeddings BIBREF12 are used and fixed during training. The output dimension of the sentence embeddings produced by the sentence encoder is the same as that of the word embeddings, i.e. 300. For the graph convolutional network, the number of hidden units is 128 and the size of the generated hidden feature vectors is also 300. We use a batch size of 1, a learning rate of $0.0075$ using Adam optimizer BIBREF13 with $\beta _1=0.9, \beta _2=0.999$ and $\epsilon =10^{-8}$. In order to make SemSentSum generalize better, we use dropout BIBREF14 of $0.2$, batch normalization BIBREF15, clip the gradient norm at $1.0$ if higher, add L2-norm regularizer with a regularization factor of $10^{-12}$ and train using early stopping with a patience of 10 iterations. Finally, the similarity threshold $t_{sim}^s$ in the summary generation process is $0.8$ (tuned on the validation set). Experiments ::: Summarization Performance We train our model SemSentSum on DUC 2001/2002, tune it on DUC 2003 and assess the performance on DUC 2004. In order to fairly compare SemSentSum with other models available in the literature, experiments are conducted with summaries truncated to 665 bytes (official summary length in the DUC competition), but also with summaries with a length constraint of 100 words. To the best of our knowledge, we are the first to conduct experiments on both summary lengths and compare our model with other systems producing either 100 words or 665 bytes summaries. Experiments ::: Sentence Semantic Relation Graph Construction We investigate different methods to build our sentence semantic relation graph and vary the value of $t_{sim}^g$ from $0.0$ to $0.75$ to study the impact of the threshold cut-off. Among these are : Cosine : Using cosine similarity ; Tf-idf : Considering a node as the query and another as document. The weight corresponds to the cosine similarity between the query and the document ; TextRank BIBREF16 : A weighted graph is created where nodes are sentences and edges defined by a similarity measure based on word overlap. Afterward, an algorithm similar to PageRank BIBREF17 is used to compute sentence importance and refined edge weights ; LexRank BIBREF9 : An unsupervised multi-document summarizer based on the concept of eigenvector centrality in a graph of sentences to set up the edge weights ; Approximate Discourse Graph (ADG) BIBREF2 : Approximation of a discourse graph where nodes are sentences and edges $(S_u,S_v)$ indicates sentence $S_v$ can be placed after $S_u$ in a coherent summary ; Personalized ADG (PADG) BIBREF3 : Normalized version of ADG where sentence nodes are normalized over all edges. Experiments ::: Ablation Study In order to quantify the contribution of the different components of SemSentSum, we try variations of our model by removing different modules one at a time. Our two main elements are the sentence encoder (Sent) and the graph convolutional neural network (GCN). When we omit Sent, we substitute it with the pre-trained sentence embeddings used to build our sentence semantic relation graph. Experiments ::: Results and Discussion Three dimensions are used to evaluate our model SemSentSum : 1) the summarization performance, to assess its capability 2) the impact of the sentence semantic relation graph generation using various methods and different thresholds $t_{sim}^g$ 3) an ablation study to analyze the importance of each component of SemSentSum. Experiments ::: Results and Discussion ::: Summarization Performance We compare the results of SemSentSum for both settings : 665 bytes and 100 words summaries. We only include models using the same parameters to compute the ROUGE-1/ROUGE-2 score and recall as metrics. The results for 665 bytes summaries are reported in Table TABREF28. We compare SemSentSum with three types of model relying on either 1) sentence or document embeddings 2) various hand-crafted features or 3) additional data. For the first category, we significantly outperform MMR BIBREF18, PV-DBOW+BS BIBREF19 and PG-MMR BIBREF20. Although their methods are based on embeddings to represent the meaning, it shows that using only various distance metrics or encoder-decoder architecture on these is not efficient for the task of multi-document summarization (as also shown in the Ablation Study). We hypothesize that SemSentSum performs better by leveraging pre-trained sentence embeddings and hence lowering the effects of data scarcity. Systems based on hand-crafted features include a widely-used learning-based summarization method, built on support vector regression SVR BIBREF21 ; a graph-based method based on approximating discourse graph G-Flow BIBREF2 ; Peer 65 which is the best peer systems participating in DUC evaluations ; and the recursive neural network R2N2 of BIBREF1 that learns automatically combinations of hand-crafted features. As can be seen, among these models completely dependent on hand-crafted features, SemSentSum achieves highest performance on both ROUGE scores. This denotes that using different linguistic and word-based features might not be enough to capture the semantic structures, in addition to being cumbersome to craft. The last type of model is shown in TCSum BIBREF4 and uses transfer learning from a text classifier model, based on a domain-related dataset of $30\,000$ documents from New York Times (sharing the same topics of the DUC datasets). In terms of ROUGE-1, SemSentSum significantly outperforms TCSum and performs similarly on ROUGE-2 score. This demonstrates that collecting more manually annotated data and training two models is unnecessary, in addition to being difficult to use in other domains, whereas SemSentSum is fully data driven, domain-independent and usable in realistic scenarios. Table TABREF32 depicts models producing 100 words summaries, all depending on hand-crafted features. We use as baselines FreqSum BIBREF22 ; TsSum BIBREF23 ; traditional graph-based approaches such as Cont. LexRank BIBREF9 ; Centroid BIBREF24 ; CLASSY04 BIBREF25 ; its improved version CLASSY11 BIBREF26 and the greedy model GreedyKL BIBREF27. All of these models are significantly underperforming compared to SemSentSum. In addition, we include state-of-the-art models : RegSum BIBREF0 and GCN+PADG BIBREF3. We outperform both in terms of ROUGE-1. For ROUGE-2 scores we achieve better results than GCN+PADG but without any use of domain-specific hand-crafted features and a much smaller and simpler model. Finally, RegSum achieves a similar ROUGE-2 score but computes sentence saliences based on word scores, incorporating a rich set of word-level and domain-specific features. Nonetheless, our model is competitive and does not depend on hand-crafted features due to its full data-driven nature and thus, it is not limited to a single domain. Consequently, the experiments show that achieving good performance for multi-document summarization without hand-crafted features or additional data is clearly feasible and SemSentSum produces competitive results without depending on these, is domain independent, fast to train and thus usable in real scenarios. Experiments ::: Results and Discussion ::: Sentence Semantic Relation Graph Table TABREF34 shows the results of different methods to create the sentence semantic relation graph with various thresholds $t_{sim}^g$ for 665 bytes summaries (we obtain similar results for 100 words). A first observation is that using cosine similarity with sentence embeddings significantly outperforms all other methods for ROUGE-1 and ROUGE-2 scores, mainly because it relies on the semantic of sentences instead of their individual words. A second is that different methods evolve similarly : PADG, Textrank, Tf-idf behave similarly to an U-shaped curve for both ROUGE scores while Cosine is the only one having an inverted U-shaped curve. The reason for this behavior is a consequence of its distribution being similar to a normal distribution because it relies on the semantic instead of words, while the others are more skewed towards zero. This confirms our hypothesis that 1) having a complete graph does not allow the model to leverage much the semantic 2) a sparse graph might not contain enough information to exploit similarities. Finally, Lexrank and ADG have different trends between both ROUGE scores. Experiments ::: Results and Discussion ::: Ablation Study We quantify the contribution of each module of SemSentSum in Table TABREF36 for 665 bytes summaries (we obtain similar results for 100 words). Removing the sentence encoder produces slightly lower results. This shows that the sentence semantic relation graph captures semantic attributes well, while the fine-tuned sentence embeddings obtained via the encoder help boost the performance, making these methods complementary. By disabling only the graph convolutional layer, a drastic drop in terms of performance is observed, which emphasizes that the relationship among sentences is indeed important and not present in the data itself. Therefore, our sentence semantic relation graph is able to capture sentence similarities by analyzing the semantic structures. Interestingly, if we remove the sentence encoder in addition to the graph convolutional layer, similar results are achieved. This confirms that alone, the sentence encoder is not able to compute an efficient representation of sentences for the task of multi-document summarization, probably due to the poor size of the DUC datasets. Finally, we can observe that the use of sentence embeddings only results in similar performance to the baselines, which rely on sentence or document embeddings BIBREF18, BIBREF19. Related Work The idea of using multiple embeddings has been employed at the word level. BIBREF28 use an attention mechanism to combine the embeddings for each word for the task of natural language inference. BIBREF29, BIBREF30 concatenate the embeddings of each word into a vector before feeding a neural network for the tasks of aspect extraction and sentiment analysis. To our knowledge, we are the first to combine multiple types of sentence embeddings. Extractive multi-document summarization has been addressed by a large range of approaches. Several of them employ graph-based methods. BIBREF31 introduced a cross-document structure theory, as a basis for multi-document summarization. BIBREF9 proposed LexRank, an unsupervised multi-document summarizer based on the concept of eigenvector centrality in a graph of sentences. Other works exploit shallow or deep features from the graph's topology BIBREF32, BIBREF33. BIBREF34 pairs graph-based methods (e.g. random walk) with clustering. BIBREF35 improved results by using a reinforced random walk model to rank sentences and keep non-redundant ones. The system by BIBREF2 does sentence selection, while balancing coherence and salience and by building a graph that approximates discourse relations across sentences BIBREF36. Besides graph-based methods, other viable approaches include Maximum Marginal Relevance BIBREF37, which uses a greedy approach to select sentences and considers the tradeoff between relevance and redundancy ; support vector regression BIBREF21 ; conditional random field BIBREF38 ; or hidden markov model BIBREF25. Yet other approaches rely on n-grams regression as in BIBREF39. More recently, BIBREF1 built a recursive neural network, which tries to automatically detect combination of hand-crafted features. BIBREF4 employ a neural model for text classification on a large manually annotated dataset and apply transfer learning for multi-document summarization afterward. The work most closely related to ours is BIBREF3. They create a normalized version of the approximate discourse graph BIBREF2, based on hand-crafted features, where sentence nodes are normalized over all the incoming edges. They then employ a deep neural network, composed of a sentence encoder, three graph convolutional layers, one document encoder and an attention mechanism. Afterward, they greedily select sentences using TF-IDF similarity to detect redundant sentences. Our model differs in four ways : 1) we build our sentence semantic relation graph by using pre-trained sentence embeddings with cosine similarity, where neither heavy preprocessing, nor hand-crafted features are necessary. Thus, our model is fully data-driven and domain-independent unlike other systems. In addition, the sentence semantic relation graph could be used for other tasks than multi-document summarization, such as detecting information cascades, query-focused summarization, keyphrase extraction or information retrieval, as it is not composed of hand-crafted features. 2) SemSentSum is much smaller and consequently has fewer parameters as it only uses a sentence encoder and a single convolutional layer. 3) The loss function is based on ROUGE-1 $F_1$ score instead of recall to prevent the tendency of choosing longer sentences. 4) Our method for summary generation is also different and novel as we leverage sentence embeddings to compute the similarity between a candidate sentence and the current summary instead of TF-IDF based approaches. Conclusion In this work, we propose a method to combine two types of sentence embeddings : 1) universal embeddings, pre-trained on a large corpus such as Wikipedia and incorporating general semantic structures across sentences and 2) domain-specific embeddings, learned during training. We merge them together by using a graph convolutional network that eliminates the need of hand-crafted features or additional annotated data. We introduce a fully data-driven model SemSentSum that achieves competitive results for multi-document summarization on both kind of summary length (665 bytes and 100 words summaries), without requiring hand-crafted features or additional annotated data. As SemSentSum is domain-independent, we believe that our sentence semantic relation graph and model can be used for other tasks including detecting information cascades, query-focused summarization, keyphrase extraction and information retrieval. In addition, we plan to leave the weights of the sentence semantic relation graph dynamic during training, and to integrate an attention mechanism directly into the graph. Acknowledgments We thank Michaela Benk for proofreading and helpful advice.

Unanswerable

48cc41c372d44b69a477998be449f8b81384786b

48cc41c372d44b69a477998be449f8b81384786b_0

Q: How better are state-of-the-art results than this model? Text: Introduction Today's increasing flood of information on the web creates a need for automated multi-document summarization systems that produce high quality summaries. However, producing summaries in a multi-document setting is difficult, as the language used to display the same information in a sentence can vary significantly, making it difficult for summarization models to capture. Given the complexity of the task and the lack of datasets, most researchers use extractive summarization, where the final summary is composed of existing sentences in the input documents. More specifically, extractive summarization systems output summaries in two steps : via sentence ranking, where an importance score is assigned to each sentence, and via the subsequent sentence selection, where the most appropriate sentence is chosen, by considering 1) their importance and 2) their frequency among all documents. Due to data sparcity, models heavily rely on well-designed features at the word level BIBREF0, BIBREF1, BIBREF2, BIBREF3 or take advantage of other large, manually annotated datasets and then apply transfer learning BIBREF4. Additionally, most of the time, all sentences in the same collection of documents are processed independently and therefore, their relationships are lost. In realistic scenarios, features are hard to craft, gathering additional annotated data is costly, and the large variety in expressing the same fact cannot be handled by the use of word-based features only, as is often the case. In this paper, we address these obstacles by proposing to simultaneously leverage two types of sentence embeddings, namely embeddings pre-trained on a large corpus that capture a variety of meanings and domain-specific embeddings learned during training. The former is typically trained on an unrelated corpus composed of high quality texts, allowing to cover additional contexts for each encountered word and sentence. Hereby, we build on the assumption that sentence embeddings capture both the syntactic and semantic content of sentences. We hypothesize that using two types of sentence embeddings, general and domain-specific, is beneficial for the task of multi-document summarization, as the former captures the most common semantic structures from a large, general corpus, while the latter captures the aspects related to the domain. We present SemSentSum (Figure FIGREF3), a fully data-driven summarization system, which does not depend on hand-crafted features, nor additional data, and is thus domain-independent. It first makes use of general sentence embedding knowledge to build a sentenc semantic relation graph that captures sentence similarities (Section SECREF4). In a second step, it trains genre-specific sentence embeddings related to the domains of the collection of documents, by utilizing a sentence encoder (Section SECREF5). Both representations are afterwards merged, by using a graph convolutional network BIBREF5 (Section SECREF6). Then, it employs a linear layer to project high-level hidden features for individual sentences to salience scores (Section SECREF8). Finally, it greedily produces relevant and non-redundant summaries by using sentence embeddings to detect similarities between candidate sentences and the current summary (Section SECREF11). The main contributions of this work are as follows : We aggregate two types of sentences embeddings using a graph representation. They share different properties and are consequently complementary. The first one is trained on a large unrelated corpus to model general semantics among sentences, whereas the second is domain-specific to the dataset and learned during training. Together, they enable a model to be domain-independent as it can be applied easily on other domains. Moreover, it could be used for other tasks including detecting information cascades, query-focused summarization, keyphrase extraction and information retrieval. We devise a competitive multi-document summarization system, which does not need hand-crafted features nor additional annotated data. Moreover, the results are competitive for 665-byte and 100-word summaries. Usually, models are compared in one of the two settings but not both and thus lack comparability. Method Let $C$ denote a collection of related documents composed of a set of documents $\lbrace D_i|i \in [1,N]\rbrace $ where $N$ is the number of documents. Moreover, each document $D_i$ consists of a set of sentences $\lbrace S_{i,j}|j \in [1,M]\rbrace $, $M$ being the number of sentences in $D_i$. Given a collection of related documents $C$, our goal is to produce a summary $Sum$ using a subset of these in the input documents ordered in some way, such that $Sum = (S_{i_1,j_1},S_{i_2,j_2},...,S_{i_n,j_m})$. In this section, we describe how SemSentSum estimates the salience score of each sentence and how it selects a subset of these to create the final summary. The architecture of SemSentSum is depicted in Figure FIGREF3. In order to perform sentence selection, we first build our sentence semantic relation graph, where each vertex is a sentence and edges capture the semantic similarity among them. At the same time, each sentence is fed into a recurrent neural network, as a sentence encoder, to generate sentence embeddings using the last hidden states. A single-layer graph convolutional neural network is then applied on top, where the sentence semantic relation graph is the adjacency matrix and the sentence embeddings are the node features. Afterward, a linear layer is used to project high-level hidden features for individual sentences to salience scores, representing how salient a sentence is with respect to the final summary. Finally, based on this, we devise an innovative greedy method that leverages sentence embeddings to detect redundant sentences and select sentences until reaching the summary length limit. Method ::: Sentence Semantic Relation Graph We model the semantic relationship among sentences using a graph representation. In this graph, each vertex is a sentence $S_{i,j}$ ($j$'th sentence of document $D_i$) from the collection documents $C$ and an undirected edge between $S_{i_u,j_u}$ and $S_{i_v,j_v}$ indicates their degree of similarity. In order to compute the semantic similarity, we use the model of BIBREF6 trained on the English Wikipedia corpus. In this manner, we incorporate general knowledge (i.e. not domain-specific) that will complete the specialized sentence embeddings obtained during training (see Section SECREF5). We process sentences by their model and compute the cosine similarity between every sentence pair, resulting in a complete graph. However, having a complete graph alone does not allow the model to leverage the semantic structure across sentences significantly, as every sentence pair is connected, and likewise, a sparse graph does not contain enough information to exploit semantic similarities. Furthermore, all edges have a weight above zero, since it is very unlikely that two sentence embeddings are completely orthogonal. To overcome this problem, we introduce an edge-removal-method, where every edge below a certain threshold $t_{sim}^g$ is removed in order to emphasize high sentence similarity. Nonetheless, $t_{sim}^g$ should not be too large, as we otherwise found the model to be prone to overfitting. After removing edges below $t_{sim}^g$, our sentence semantic relation graph is used as the adjacency matrix $A$. The impact of $t_{sim}^g$ with different values is shown in Section SECREF26. Based on our aforementioned hypothesis that a combination of general and genre-specific sentence embeddings is beneficial for the task of multi-document summarization, we further incorporate general sentence embeddings, pre-trained on Wikipedia entries, into edges between sentences. Additionally, we compute specialised sentence embeddings, which are related to the domains of the documents (see Section SECREF35). Note that 1) the pre-trained sentence embeddings are only used to compute the weights of the edges and are not used by the summarization model (as others are produced by the sentence encoder) and 2) the edge weights are static and do not change during training. Method ::: Sentence Encoder Given a list of documents $C$, we encode each document's sentence $S_{i,j}$, where each has at most $L$ words $(w_{i,j,1}, w_{i,j,2}, ..., w_{i,j,L})$. In our experiments, all words are kept and converted into word embeddings, which are then fed to the sentence encoder in order to compute specialized sentence embeddings $S^{\prime }_{i,j}$. We employ a single-layer forward recurrent neural network, using Long Short-Term Memory (LSTM) of BIBREF7 as sentence encoder, where the sentence embeddings are extracted from the last hidden states. We then concatenate all sentence embeddings into a matrix $X$ which constitutes the input node features that will be used by the graph convolutional network. Method ::: Graph Convolutional Network After having computed all sentence embeddings and the sentence semantic relation graph, we apply a single-layer Graph Convolutional Network (GCN) from BIBREF5, in order to capture high-level hidden features for each sentence, encapsulating sentence information as well as the graph structure. We believe that our sentence semantic relation graph contains information not present in the data (via universal embeddings) and thus, we leverage this information by running a graph convolution on the first order neighborhood. The GCN model takes as input the node features matrix $X$ and a squared adjacency matrix $A$. The former contains all sentence embeddings of the collection of documents, while the latter is our underlying sentence semantic relation graph. It outputs hidden representations for each node that encode both local graph structure and nodes's features. In order to take into account the sentences themselves during the information propagation, we add self-connections (i.e. the identity matrix) to $A$ such that $\tilde{A} = A + I$. Subsequently, we obtain our sentence hidden features by using Equation DISPLAY_FORM7. where $W_i$ is the weight matrix of the $i$'th graph convolution layer and $b_i$ the bias vector. We choose the Exponential Linear Unit (ELU) activation function from BIBREF8 due to its ability to handle the vanishing gradient problem, by pushing the mean unit activations close to zero and consequently facilitating the backpropagation. By using only one hidden layer, as we only have one input-to-hidden layer and one hidden-to-output layer, we limit the information propagation to the first order neighborhood. Method ::: Saliency Estimation We use a simple linear layer to estimate a salience score for each sentence and then normalize the scores via softmax and obtain our estimated salience score $S^s_{i,j}$. Method ::: Training Our model SemSentSum is trained in an end-to-end manner and minimizes the cross-entropy loss of Equation DISPLAY_FORM10 between the salience score prediction and the ROUGE-1 $F_1$ score for each sentence. $F_1(S)$ is computed as the ROUGE-1 $F_1$ score, unlike the common practice in the area of single and multi-document summarization as recall favors longer sentences whereas $F_1$ prevents this tendency. The scores are normalized via softmax. Method ::: Summary Generation Process While our model SemSentSum provides estimated saliency scores, we use a greedy strategy to construct an informative and non-redundant summary $Sum$. We first discard sentences having less than 9 words, as in BIBREF9, and then sort them in descending order of their estimated salience scores. We iteratively dequeue the sentence having the highest score and append it to the current summary $Sum$ if it is non-redundant with respect to the current content of $Sum$. We iterate until reaching the summary length limit. To determine the similarity of a candidate sentence with the current summary, a sentence is considered as dissimilar if and only if the cosine similarity between its sentence embeddings and the embeddings of the current summary is below a certain threshold $t_{sim}^s$. We use the pre-trained model of BIBREF6 to compute sentence as well as summary embeddings, similarly to the sentence semantic relation graph construction. Our approach is novel, since it focuses on the semantic sentence structures and captures similarity between sentence meanings, instead of focusing on word similarities only, like previous TF-IDF approaches ( BIBREF0, BIBREF1, BIBREF3, BIBREF4). Experiments ::: Datasets We conduct experiments on the most commonly used datasets for multi-document summarization from the Document Understanding Conferences (DUC). We use DUC 2001, 2002, 2003 and 2004 as the tasks of generic multi-document summarization, because they have been carried out during these years. We use DUC 2001, 2002, 2003 and 2004 for generic multi-document summarization, where DUC 2001/2002 are used for training, DUC 2003 for validation and finally, DUC 2004 for testing, following the common practice. Experiments ::: Evaluation Metric For the evaluation, we use ROUGE BIBREF10 with the official parameters of the DUC tasks and also truncate the summaries to 100 words for DUC 2001/2002/2003 and to 665 bytes for DUC 2004. Notably, we take ROUGE-1 and ROUGE-2 recall scores as the main metrics for comparison between produced summaries and golden ones as proposed by BIBREF11. The goal of the ROUGE-N metric is to compute the ratio of the number of N-grams from the generated summary matching these of the human reference summaries. Experiments ::: Model Settings To define the edge weights of our sentence semantic relation graph, we employ the 600-dimensional pre-trained unigram model of BIBREF6, using English Wikipedia as source corpus. We keep only edges having a weight larger than $t_{sim}^g = 0.5$ (tuned on the validation set). For word embeddings, the 300-dimensional pre-trained GloVe embeddings BIBREF12 are used and fixed during training. The output dimension of the sentence embeddings produced by the sentence encoder is the same as that of the word embeddings, i.e. 300. For the graph convolutional network, the number of hidden units is 128 and the size of the generated hidden feature vectors is also 300. We use a batch size of 1, a learning rate of $0.0075$ using Adam optimizer BIBREF13 with $\beta _1=0.9, \beta _2=0.999$ and $\epsilon =10^{-8}$. In order to make SemSentSum generalize better, we use dropout BIBREF14 of $0.2$, batch normalization BIBREF15, clip the gradient norm at $1.0$ if higher, add L2-norm regularizer with a regularization factor of $10^{-12}$ and train using early stopping with a patience of 10 iterations. Finally, the similarity threshold $t_{sim}^s$ in the summary generation process is $0.8$ (tuned on the validation set). Experiments ::: Summarization Performance We train our model SemSentSum on DUC 2001/2002, tune it on DUC 2003 and assess the performance on DUC 2004. In order to fairly compare SemSentSum with other models available in the literature, experiments are conducted with summaries truncated to 665 bytes (official summary length in the DUC competition), but also with summaries with a length constraint of 100 words. To the best of our knowledge, we are the first to conduct experiments on both summary lengths and compare our model with other systems producing either 100 words or 665 bytes summaries. Experiments ::: Sentence Semantic Relation Graph Construction We investigate different methods to build our sentence semantic relation graph and vary the value of $t_{sim}^g$ from $0.0$ to $0.75$ to study the impact of the threshold cut-off. Among these are : Cosine : Using cosine similarity ; Tf-idf : Considering a node as the query and another as document. The weight corresponds to the cosine similarity between the query and the document ; TextRank BIBREF16 : A weighted graph is created where nodes are sentences and edges defined by a similarity measure based on word overlap. Afterward, an algorithm similar to PageRank BIBREF17 is used to compute sentence importance and refined edge weights ; LexRank BIBREF9 : An unsupervised multi-document summarizer based on the concept of eigenvector centrality in a graph of sentences to set up the edge weights ; Approximate Discourse Graph (ADG) BIBREF2 : Approximation of a discourse graph where nodes are sentences and edges $(S_u,S_v)$ indicates sentence $S_v$ can be placed after $S_u$ in a coherent summary ; Personalized ADG (PADG) BIBREF3 : Normalized version of ADG where sentence nodes are normalized over all edges. Experiments ::: Ablation Study In order to quantify the contribution of the different components of SemSentSum, we try variations of our model by removing different modules one at a time. Our two main elements are the sentence encoder (Sent) and the graph convolutional neural network (GCN). When we omit Sent, we substitute it with the pre-trained sentence embeddings used to build our sentence semantic relation graph. Experiments ::: Results and Discussion Three dimensions are used to evaluate our model SemSentSum : 1) the summarization performance, to assess its capability 2) the impact of the sentence semantic relation graph generation using various methods and different thresholds $t_{sim}^g$ 3) an ablation study to analyze the importance of each component of SemSentSum. Experiments ::: Results and Discussion ::: Summarization Performance We compare the results of SemSentSum for both settings : 665 bytes and 100 words summaries. We only include models using the same parameters to compute the ROUGE-1/ROUGE-2 score and recall as metrics. The results for 665 bytes summaries are reported in Table TABREF28. We compare SemSentSum with three types of model relying on either 1) sentence or document embeddings 2) various hand-crafted features or 3) additional data. For the first category, we significantly outperform MMR BIBREF18, PV-DBOW+BS BIBREF19 and PG-MMR BIBREF20. Although their methods are based on embeddings to represent the meaning, it shows that using only various distance metrics or encoder-decoder architecture on these is not efficient for the task of multi-document summarization (as also shown in the Ablation Study). We hypothesize that SemSentSum performs better by leveraging pre-trained sentence embeddings and hence lowering the effects of data scarcity. Systems based on hand-crafted features include a widely-used learning-based summarization method, built on support vector regression SVR BIBREF21 ; a graph-based method based on approximating discourse graph G-Flow BIBREF2 ; Peer 65 which is the best peer systems participating in DUC evaluations ; and the recursive neural network R2N2 of BIBREF1 that learns automatically combinations of hand-crafted features. As can be seen, among these models completely dependent on hand-crafted features, SemSentSum achieves highest performance on both ROUGE scores. This denotes that using different linguistic and word-based features might not be enough to capture the semantic structures, in addition to being cumbersome to craft. The last type of model is shown in TCSum BIBREF4 and uses transfer learning from a text classifier model, based on a domain-related dataset of $30\,000$ documents from New York Times (sharing the same topics of the DUC datasets). In terms of ROUGE-1, SemSentSum significantly outperforms TCSum and performs similarly on ROUGE-2 score. This demonstrates that collecting more manually annotated data and training two models is unnecessary, in addition to being difficult to use in other domains, whereas SemSentSum is fully data driven, domain-independent and usable in realistic scenarios. Table TABREF32 depicts models producing 100 words summaries, all depending on hand-crafted features. We use as baselines FreqSum BIBREF22 ; TsSum BIBREF23 ; traditional graph-based approaches such as Cont. LexRank BIBREF9 ; Centroid BIBREF24 ; CLASSY04 BIBREF25 ; its improved version CLASSY11 BIBREF26 and the greedy model GreedyKL BIBREF27. All of these models are significantly underperforming compared to SemSentSum. In addition, we include state-of-the-art models : RegSum BIBREF0 and GCN+PADG BIBREF3. We outperform both in terms of ROUGE-1. For ROUGE-2 scores we achieve better results than GCN+PADG but without any use of domain-specific hand-crafted features and a much smaller and simpler model. Finally, RegSum achieves a similar ROUGE-2 score but computes sentence saliences based on word scores, incorporating a rich set of word-level and domain-specific features. Nonetheless, our model is competitive and does not depend on hand-crafted features due to its full data-driven nature and thus, it is not limited to a single domain. Consequently, the experiments show that achieving good performance for multi-document summarization without hand-crafted features or additional data is clearly feasible and SemSentSum produces competitive results without depending on these, is domain independent, fast to train and thus usable in real scenarios. Experiments ::: Results and Discussion ::: Sentence Semantic Relation Graph Table TABREF34 shows the results of different methods to create the sentence semantic relation graph with various thresholds $t_{sim}^g$ for 665 bytes summaries (we obtain similar results for 100 words). A first observation is that using cosine similarity with sentence embeddings significantly outperforms all other methods for ROUGE-1 and ROUGE-2 scores, mainly because it relies on the semantic of sentences instead of their individual words. A second is that different methods evolve similarly : PADG, Textrank, Tf-idf behave similarly to an U-shaped curve for both ROUGE scores while Cosine is the only one having an inverted U-shaped curve. The reason for this behavior is a consequence of its distribution being similar to a normal distribution because it relies on the semantic instead of words, while the others are more skewed towards zero. This confirms our hypothesis that 1) having a complete graph does not allow the model to leverage much the semantic 2) a sparse graph might not contain enough information to exploit similarities. Finally, Lexrank and ADG have different trends between both ROUGE scores. Experiments ::: Results and Discussion ::: Ablation Study We quantify the contribution of each module of SemSentSum in Table TABREF36 for 665 bytes summaries (we obtain similar results for 100 words). Removing the sentence encoder produces slightly lower results. This shows that the sentence semantic relation graph captures semantic attributes well, while the fine-tuned sentence embeddings obtained via the encoder help boost the performance, making these methods complementary. By disabling only the graph convolutional layer, a drastic drop in terms of performance is observed, which emphasizes that the relationship among sentences is indeed important and not present in the data itself. Therefore, our sentence semantic relation graph is able to capture sentence similarities by analyzing the semantic structures. Interestingly, if we remove the sentence encoder in addition to the graph convolutional layer, similar results are achieved. This confirms that alone, the sentence encoder is not able to compute an efficient representation of sentences for the task of multi-document summarization, probably due to the poor size of the DUC datasets. Finally, we can observe that the use of sentence embeddings only results in similar performance to the baselines, which rely on sentence or document embeddings BIBREF18, BIBREF19. Related Work The idea of using multiple embeddings has been employed at the word level. BIBREF28 use an attention mechanism to combine the embeddings for each word for the task of natural language inference. BIBREF29, BIBREF30 concatenate the embeddings of each word into a vector before feeding a neural network for the tasks of aspect extraction and sentiment analysis. To our knowledge, we are the first to combine multiple types of sentence embeddings. Extractive multi-document summarization has been addressed by a large range of approaches. Several of them employ graph-based methods. BIBREF31 introduced a cross-document structure theory, as a basis for multi-document summarization. BIBREF9 proposed LexRank, an unsupervised multi-document summarizer based on the concept of eigenvector centrality in a graph of sentences. Other works exploit shallow or deep features from the graph's topology BIBREF32, BIBREF33. BIBREF34 pairs graph-based methods (e.g. random walk) with clustering. BIBREF35 improved results by using a reinforced random walk model to rank sentences and keep non-redundant ones. The system by BIBREF2 does sentence selection, while balancing coherence and salience and by building a graph that approximates discourse relations across sentences BIBREF36. Besides graph-based methods, other viable approaches include Maximum Marginal Relevance BIBREF37, which uses a greedy approach to select sentences and considers the tradeoff between relevance and redundancy ; support vector regression BIBREF21 ; conditional random field BIBREF38 ; or hidden markov model BIBREF25. Yet other approaches rely on n-grams regression as in BIBREF39. More recently, BIBREF1 built a recursive neural network, which tries to automatically detect combination of hand-crafted features. BIBREF4 employ a neural model for text classification on a large manually annotated dataset and apply transfer learning for multi-document summarization afterward. The work most closely related to ours is BIBREF3. They create a normalized version of the approximate discourse graph BIBREF2, based on hand-crafted features, where sentence nodes are normalized over all the incoming edges. They then employ a deep neural network, composed of a sentence encoder, three graph convolutional layers, one document encoder and an attention mechanism. Afterward, they greedily select sentences using TF-IDF similarity to detect redundant sentences. Our model differs in four ways : 1) we build our sentence semantic relation graph by using pre-trained sentence embeddings with cosine similarity, where neither heavy preprocessing, nor hand-crafted features are necessary. Thus, our model is fully data-driven and domain-independent unlike other systems. In addition, the sentence semantic relation graph could be used for other tasks than multi-document summarization, such as detecting information cascades, query-focused summarization, keyphrase extraction or information retrieval, as it is not composed of hand-crafted features. 2) SemSentSum is much smaller and consequently has fewer parameters as it only uses a sentence encoder and a single convolutional layer. 3) The loss function is based on ROUGE-1 $F_1$ score instead of recall to prevent the tendency of choosing longer sentences. 4) Our method for summary generation is also different and novel as we leverage sentence embeddings to compute the similarity between a candidate sentence and the current summary instead of TF-IDF based approaches. Conclusion In this work, we propose a method to combine two types of sentence embeddings : 1) universal embeddings, pre-trained on a large corpus such as Wikipedia and incorporating general semantic structures across sentences and 2) domain-specific embeddings, learned during training. We merge them together by using a graph convolutional network that eliminates the need of hand-crafted features or additional annotated data. We introduce a fully data-driven model SemSentSum that achieves competitive results for multi-document summarization on both kind of summary length (665 bytes and 100 words summaries), without requiring hand-crafted features or additional annotated data. As SemSentSum is domain-independent, we believe that our sentence semantic relation graph and model can be used for other tasks including detecting information cascades, query-focused summarization, keyphrase extraction and information retrieval. In addition, we plan to leave the weights of the sentence semantic relation graph dynamic during training, and to integrate an attention mechanism directly into the graph. Acknowledgments We thank Michaela Benk for proofreading and helpful advice.

we achieve better results than GCN+PADG but without any use of domain-specific hand-crafted features, RegSum achieves a similar ROUGE-2 score

efb3a87845460655c53bd7365bcb8393c99358ec

efb3a87845460655c53bd7365bcb8393c99358ec_0

Q: What were their results on the three datasets? Text: Introduction Twitter sentiment classification have intensively researched in recent years BIBREF0 BIBREF1 . Different approaches were developed for Twitter sentiment classification by using machine learning such as Support Vector Machine (SVM) with rule-based features BIBREF2 and the combination of SVMs and Naive Bayes (NB) BIBREF3 . In addition, hybrid approaches combining lexicon-based and machine learning methods also achieved high performance described in BIBREF4 . However, a problem of traditional machine learning is how to define a feature extractor for a specific domain in order to extract important features. Deep learning models are different from traditional machine learning methods in that a deep learning model does not depend on feature extractors because features are extracted during training progress. The use of deep learning methods becomes to achieve remarkable results for sentiment analysis BIBREF5 BIBREF6 BIBREF7 . Some researchers used Convolutional Neural Network (CNN) for sentiment classification. CNN models have been shown to be effective for NLP. For example, BIBREF6 proposed various kinds of CNN to learn sentiment-bearing sentence vectors, BIBREF5 adopted two CNNs in character-level to sentence-level representation for sentiment analysis. BIBREF7 constructs experiments on a character-level CNN for several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is another state-of-the-art semantic composition model for sentiment classification with many variants described in BIBREF8 . The studies reveal that using a CNN is useful in extracting information and finding feature detectors from texts. In addition, a LSTM can be good in maintaining word order and the context of words. However, in some important aspects, the use of CNN or LSTM separately may not capture enough information. Inspired by the models above, the goal of this research is using a Deep Convolutional Neural Network (DeepCNN) to exploit the information of characters of words in order to support word-level embedding. A Bi-LSTM produces a sentence-wide feature representation based on these embeddings. The Bi-LSTM is a version of BIBREF9 with Full Gradient described in BIBREF10 . In addition, the rules-based approach also effects classification accuracy by focusing on important sub-sentences expressing the main sentiment of a tweet while removing unnecessary parts of a tweet. The paper makes the following contributions: The organization of the present paper is as follows: In section 2, we describe the model architecture which introduces the structure of the model. We explain the basic idea of model and the way of constructing the model. Section 3 show results and analysis and section 4 summarize this paper. Basic idea Our proposed model consists of a deep learning classifier and a tweet processor. The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The tweet processor standardizes tweets and then applies semantic rules on datasets. We construct a framework that treats the deep learning classifier and the tweet processor as two distinct components. We believe that standardizing data is an important step to achieve high accuracy. To formulate our problem in increasing the accuracy of the classifier, we illustrate our model in Figure. FIGREF4 as follows: Tweets are firstly considered via a processor based on preprocessing steps BIBREF0 and the semantic rules-based method BIBREF11 in order to standardize tweets and capture only important information containing the main sentiment of a tweet. We use DeepCNN with Wide convolution for character-level embeddings. A wide convolution can learn to recognize specific n-grams at every position in a word that allows features to be extracted independently of these positions in the word. These features maintain the order and relative positions of characters. A DeepCNN is constructed by two wide convolution layers and the need of multiple wide convolution layers is widely accepted that a model constructing by multiple processing layers have the ability to learn representations of data with higher levels of abstraction BIBREF12 . Therefore, we use DeepCNN for character-level embeddings to support morphological and shape information for a word. The DeepCNN produces INLINEFORM0 global fixed-sized feature vectors for INLINEFORM1 words. A combination of the global fixed-size feature vectors and word-level embedding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level representation by maintaining the order of words. Our work is philosophically similar to BIBREF5 . However, our model is distinguished with their approaches in two aspects: Using DeepCNN with two wide convolution layers to increase representation with multiple levels of abstraction. Integrating global character fixed-sized feature vectors with word-level embedding to extract a sentence-wide feature set via Bi-LSTM. This deals with three main problems: (i) Sentences have any different size; (ii) The semantic and the syntactic of words in a sentence are captured in order to increase information for a word; (iii) Important information of characters that can appear at any position in a word are extracted. In sub-section B, we introduce various kinds of dataset. The modules of our model are constructed in other sub-sections. Data Preparation Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains 1,600K training tweets collected by a crawler from BIBREF0 . BIBREF0 constructed a test set manually with 177 negative and 182 positive tweets. The Stanford test set is small. However, it has been widely used in different evaluation tasks BIBREF0 BIBREF5 BIBREF13 . Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified tweets collected by using search terms: INLINEFORM0 , #google, #microsoft and #twitter. We construct the dataset as BIBREF14 for binary classification. Health Care Reform (HCR): This dataset was constructed by crawling tweets containing the hashtag #hcr BIBREF15 . Task is to predict positive/negative tweets BIBREF14 . Preprocessing We firstly take unique properties of Twitter in order to reduce the feature space such as Username, Usage of links, None, URLs and Repeated Letters. We then process retweets, stop words, links, URLs, mentions, punctuation and accentuation. For emoticons, BIBREF0 revealed that the training process makes the use of emoticons as noisy labels and they stripped the emoticons out from their training dataset because BIBREF0 believed that if we consider the emoticons, there is a negative impact on the accuracies of classifiers. In addition, removing emoticons makes the classifiers learns from other features (e.g. unigrams and bi-grams) presented in tweets and the classifiers only use these non-emoticon features to predict the sentiment of tweets. However, there is a problem is that if the test set contains emoticons, they do not influence the classifiers because emoticon features do not contain in its training data. This is a limitation of BIBREF0 , because the emoticon features would be useful when classifying test data. Therefore, we keep emoticon features in the datasets because deep learning models can capture more information from emoticon features for increasing classification accuracy. Semantic Rules (SR) In Twitter social networking, people express their opinions containing sub-sentences. These sub-sentences using specific PoS particles (Conjunction and Conjunctive adverbs), like "but, while, however, despite, however" have different polarities. However, the overall sentiment of tweets often focus on certain sub-sentences. For example: @lonedog bwahahah...you are amazing! However, it was quite the letdown. @kirstiealley my dentist is great but she's expensive...=( In two tweets above, the overall sentiment is negative. However, the main sentiment is only in the sub-sentences following but and however. This inspires a processing step to remove unessential parts in a tweet. Rule-based approach can assists these problems in handling negation and dealing with specific PoS particles led to effectively affect the final output of classification BIBREF11 BIBREF16 . BIBREF11 summarized a full presentation of their semantic rules approach and devised ten semantic rules in their hybrid approach based on the presentation of BIBREF16 . We use five rules in the semantic rules set because other five rules are only used to compute polarity of words after POS tagging or Parsing steps. We follow the same naming convention for rules utilized by BIBREF11 to represent the rules utilized in our proposed method. The rules utilized in the proposed method are displayed in Table TABREF15 in which is included examples from STS Corpus and output after using the rules. Table TABREF16 illustrates the number of processed sentences on each dataset. Representation Levels To construct embedding inputs for our model, we use a fixed-sized word vocabulary INLINEFORM0 and a fixed-sized character vocabulary INLINEFORM1 . Given a word INLINEFORM2 is composed from characters INLINEFORM3 , the character-level embeddings are encoded by column vectors INLINEFORM4 in the embedding matrix INLINEFORM5 , where INLINEFORM6 is the size of the character vocabulary. For word-level embedding INLINEFORM7 , we use a pre-trained word-level embedding with dimension 200 or 300. A pre-trained word-level embedding can capture the syntactic and semantic information of words BIBREF17 . We build every word INLINEFORM8 into an embedding INLINEFORM9 which is constructed by two sub-vectors: the word-level embedding INLINEFORM10 and the character fixed-size feature vector INLINEFORM11 of INLINEFORM12 where INLINEFORM13 is the length of the filter of wide convolutions. We have INLINEFORM14 character fixed-size feature vectors corresponding to word-level embedding in a sentence. Deep Learning Module DeepCNN in the deep learning module is illustrated in Figure. FIGREF22 . The DeepCNN has two wide convolution layers. The first layer extract local features around each character windows of the given word and using a max pooling over character windows to produce a global fixed-sized feature vector for the word. The second layer retrieves important context characters and transforms the representation at previous level into a representation at higher abstract level. We have INLINEFORM0 global character fixed-sized feature vectors for INLINEFORM1 words. In the next step of Figure. FIGREF4 , we construct the vector INLINEFORM0 by concatenating the word-level embedding with the global character fixed-size feature vectors. The input of Bi-LSTM is a sequence of embeddings INLINEFORM1 . The use of the global character fixed-size feature vectors increases the relationship of words in the word-level embedding. The purpose of this Bi-LSTM is to capture the context of words in a sentence and maintain the order of words toward to extract sentence-level representation. The top of the model is a softmax function to predict sentiment label. We describe in detail the kinds of CNN and LSTM that we use in next sub-part 1 and 2. The one-dimensional convolution called time-delay neural net has a filter vector INLINEFORM0 and take the dot product of filter INLINEFORM1 with each m-grams in the sequence of characters INLINEFORM2 of a word in order to obtain a sequence INLINEFORM3 : DISPLAYFORM0 Based on Equation 1, we have two types of convolutions that depend on the range of the index INLINEFORM0 . The narrow type of convolution requires that INLINEFORM1 and produce a sequence INLINEFORM2 . The wide type of convolution does not require on INLINEFORM3 or INLINEFORM4 and produce a sequence INLINEFORM5 . Out-of-range input values INLINEFORM6 where INLINEFORM7 or INLINEFORM8 are taken to be zero. We use wide convolution for our model. Given a word INLINEFORM0 composed of INLINEFORM1 characters INLINEFORM2 , we take a character embedding INLINEFORM3 for each character INLINEFORM4 and construct a character matrix INLINEFORM5 as following Equation. 2: DISPLAYFORM0 The values of the embeddings INLINEFORM0 are parameters that are optimized during training. The trained weights in the filter INLINEFORM1 correspond to a feature detector which learns to recognize a specific class of n-grams. The n-grams have size INLINEFORM2 . The use of a wide convolution has some advantages more than a narrow convolution because a wide convolution ensures that all weights of filter reach the whole characters of a word at the margins. The resulting matrix has dimension INLINEFORM3 . Long Short-Term Memory networks usually called LSTMs are a improved version of RNN. The core idea behind LSTMs is the cell state which can maintain its state over time, and non-linear gating units which regulate the information flow into and out of the cell. The LSTM architecture that we used in our proposed model is described in BIBREF9 . A single LSTM memory cell is implemented by the following composite function: DISPLAYFORM0 DISPLAYFORM1 where INLINEFORM0 is the logistic sigmoid function, INLINEFORM1 and INLINEFORM2 are the input gate, forget gate, output gate, cell and cell input activation vectors respectively. All of them have a same size as the hidden vector INLINEFORM3 . INLINEFORM4 is the hidden-input gate matrix, INLINEFORM5 is the input-output gate matrix. The bias terms which are added to INLINEFORM6 and INLINEFORM7 have been omitted for clarity. In addition, we also use the full gradient for calculating with full backpropagation through time (BPTT) described in BIBREF10 . A LSTM gradients using finite differences could be checked and making practical implementations more reliable. Regularization For regularization, we use a constraint on INLINEFORM0 of the weight vectors BIBREF18 . Experimental setups For the Stanford Twitter Sentiment Corpus, we use the number of samples as BIBREF5 . The training data is selected 80K tweets for a training data and 16K tweets for the development set randomly from the training data of BIBREF0 . We conduct a binary prediction for STS Corpus. For Sander dataset, we use standard 10-fold cross validation as BIBREF14 . We construct the development set by selecting 10% randomly from 9-fold training data. In Health Care Reform Corpus, we also select 10% randomly for the development set in a training set and construct as BIBREF14 for comparison. We describe the summary of datasets in Table III. for all datasets, the filter window size ( INLINEFORM0 ) is 7 with 6 feature maps each for the first wide convolution layer, the second wide convolution layer has a filter window size of 5 with 14 feature maps each. Dropout rate ( INLINEFORM1 ) is 0.5, INLINEFORM2 constraint, learning rate is 0.1 and momentum of 0.9. Mini-batch size for STS Corpus is 100 and others are 4. In addition, training is done through stochastic gradient descent over shuffled mini-batches with Adadelta update rule BIBREF19 . we use the publicly available Word2Vec trained from 100 billion words from Google and TwitterGlove of Stanford is performed on aggregated global word-word co-occurrence statistics from a corpus. Word2Vec has dimensionality of 300 and Twitter Glove have dimensionality of 200. Words that do not present in the set of pre-train words are initialized randomly. Experimental results Table IV shows the result of our model for sentiment classification against other models. We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN. As can be seen, 86.63 is the best prediction accuracy of our model so far for the STS Corpus. For Sanders and HCR datasets, we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets. Our models outperform the model of BIBREF14 for the Sanders dataset and HCR dataset. Analysis As can be seen, the models with SR outperforms the model with no SR. Semantic rules is effective in order to increase classification accuracy. We evaluate the efficiency of SR for the model in Table V of our full paper . We also conduct two experiments on two separate models: DeepCNN and Bi-LSTM in order to show the effectiveness of combination of DeepCNN and Bi-LSTM. In addition, the model using TwitterGlove outperform the model using GoogleW2V because TwitterGlove captures more information in Twitter than GoogleW2V. These results show that the character-level information and SR have a great impact on Twitter Data. The pre-train word vectors are good, universal feature extractors. The difference between our model and other approaches is the ability of our model to capture important features by using SR and combine these features at high benefit. The use of DeepCNN can learn a representation of words in higher abstract level. The combination of global character fixed-sized feature vectors and a word embedding helps the model to find important detectors for particles such as 'not' that negate sentiment and potentiate sentiment such as 'too', 'so' standing beside expected features. The model not only learns to recognize single n-grams, but also patterns in n-grams lead to form a structure significance of a sentence. Conclusions In the present work, we have pointed out that the use of character embeddings through a DeepCNN to enhance information for word embeddings built on top of Word2Vec or TwitterGlove improves classification accuracy in Tweet sentiment classification. Our results add to the well-establish evidence that character vectors are an important ingredient for word-level in deep learning for NLP. In addition, semantic rules contribute handling non-essential sub-tweets in order to improve classification accuracy.

accuracy of 86.63 on STS, 85.14 on Sanders and 80.9 on HCR

0619fc797730a3e59ac146a5a4575c81517cc618

0619fc797730a3e59ac146a5a4575c81517cc618_0

Q: What was the baseline? Text: Introduction Twitter sentiment classification have intensively researched in recent years BIBREF0 BIBREF1 . Different approaches were developed for Twitter sentiment classification by using machine learning such as Support Vector Machine (SVM) with rule-based features BIBREF2 and the combination of SVMs and Naive Bayes (NB) BIBREF3 . In addition, hybrid approaches combining lexicon-based and machine learning methods also achieved high performance described in BIBREF4 . However, a problem of traditional machine learning is how to define a feature extractor for a specific domain in order to extract important features. Deep learning models are different from traditional machine learning methods in that a deep learning model does not depend on feature extractors because features are extracted during training progress. The use of deep learning methods becomes to achieve remarkable results for sentiment analysis BIBREF5 BIBREF6 BIBREF7 . Some researchers used Convolutional Neural Network (CNN) for sentiment classification. CNN models have been shown to be effective for NLP. For example, BIBREF6 proposed various kinds of CNN to learn sentiment-bearing sentence vectors, BIBREF5 adopted two CNNs in character-level to sentence-level representation for sentiment analysis. BIBREF7 constructs experiments on a character-level CNN for several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is another state-of-the-art semantic composition model for sentiment classification with many variants described in BIBREF8 . The studies reveal that using a CNN is useful in extracting information and finding feature detectors from texts. In addition, a LSTM can be good in maintaining word order and the context of words. However, in some important aspects, the use of CNN or LSTM separately may not capture enough information. Inspired by the models above, the goal of this research is using a Deep Convolutional Neural Network (DeepCNN) to exploit the information of characters of words in order to support word-level embedding. A Bi-LSTM produces a sentence-wide feature representation based on these embeddings. The Bi-LSTM is a version of BIBREF9 with Full Gradient described in BIBREF10 . In addition, the rules-based approach also effects classification accuracy by focusing on important sub-sentences expressing the main sentiment of a tweet while removing unnecessary parts of a tweet. The paper makes the following contributions: The organization of the present paper is as follows: In section 2, we describe the model architecture which introduces the structure of the model. We explain the basic idea of model and the way of constructing the model. Section 3 show results and analysis and section 4 summarize this paper. Basic idea Our proposed model consists of a deep learning classifier and a tweet processor. The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The tweet processor standardizes tweets and then applies semantic rules on datasets. We construct a framework that treats the deep learning classifier and the tweet processor as two distinct components. We believe that standardizing data is an important step to achieve high accuracy. To formulate our problem in increasing the accuracy of the classifier, we illustrate our model in Figure. FIGREF4 as follows: Tweets are firstly considered via a processor based on preprocessing steps BIBREF0 and the semantic rules-based method BIBREF11 in order to standardize tweets and capture only important information containing the main sentiment of a tweet. We use DeepCNN with Wide convolution for character-level embeddings. A wide convolution can learn to recognize specific n-grams at every position in a word that allows features to be extracted independently of these positions in the word. These features maintain the order and relative positions of characters. A DeepCNN is constructed by two wide convolution layers and the need of multiple wide convolution layers is widely accepted that a model constructing by multiple processing layers have the ability to learn representations of data with higher levels of abstraction BIBREF12 . Therefore, we use DeepCNN for character-level embeddings to support morphological and shape information for a word. The DeepCNN produces INLINEFORM0 global fixed-sized feature vectors for INLINEFORM1 words. A combination of the global fixed-size feature vectors and word-level embedding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level representation by maintaining the order of words. Our work is philosophically similar to BIBREF5 . However, our model is distinguished with their approaches in two aspects: Using DeepCNN with two wide convolution layers to increase representation with multiple levels of abstraction. Integrating global character fixed-sized feature vectors with word-level embedding to extract a sentence-wide feature set via Bi-LSTM. This deals with three main problems: (i) Sentences have any different size; (ii) The semantic and the syntactic of words in a sentence are captured in order to increase information for a word; (iii) Important information of characters that can appear at any position in a word are extracted. In sub-section B, we introduce various kinds of dataset. The modules of our model are constructed in other sub-sections. Data Preparation Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains 1,600K training tweets collected by a crawler from BIBREF0 . BIBREF0 constructed a test set manually with 177 negative and 182 positive tweets. The Stanford test set is small. However, it has been widely used in different evaluation tasks BIBREF0 BIBREF5 BIBREF13 . Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified tweets collected by using search terms: INLINEFORM0 , #google, #microsoft and #twitter. We construct the dataset as BIBREF14 for binary classification. Health Care Reform (HCR): This dataset was constructed by crawling tweets containing the hashtag #hcr BIBREF15 . Task is to predict positive/negative tweets BIBREF14 . Preprocessing We firstly take unique properties of Twitter in order to reduce the feature space such as Username, Usage of links, None, URLs and Repeated Letters. We then process retweets, stop words, links, URLs, mentions, punctuation and accentuation. For emoticons, BIBREF0 revealed that the training process makes the use of emoticons as noisy labels and they stripped the emoticons out from their training dataset because BIBREF0 believed that if we consider the emoticons, there is a negative impact on the accuracies of classifiers. In addition, removing emoticons makes the classifiers learns from other features (e.g. unigrams and bi-grams) presented in tweets and the classifiers only use these non-emoticon features to predict the sentiment of tweets. However, there is a problem is that if the test set contains emoticons, they do not influence the classifiers because emoticon features do not contain in its training data. This is a limitation of BIBREF0 , because the emoticon features would be useful when classifying test data. Therefore, we keep emoticon features in the datasets because deep learning models can capture more information from emoticon features for increasing classification accuracy. Semantic Rules (SR) In Twitter social networking, people express their opinions containing sub-sentences. These sub-sentences using specific PoS particles (Conjunction and Conjunctive adverbs), like "but, while, however, despite, however" have different polarities. However, the overall sentiment of tweets often focus on certain sub-sentences. For example: @lonedog bwahahah...you are amazing! However, it was quite the letdown. @kirstiealley my dentist is great but she's expensive...=( In two tweets above, the overall sentiment is negative. However, the main sentiment is only in the sub-sentences following but and however. This inspires a processing step to remove unessential parts in a tweet. Rule-based approach can assists these problems in handling negation and dealing with specific PoS particles led to effectively affect the final output of classification BIBREF11 BIBREF16 . BIBREF11 summarized a full presentation of their semantic rules approach and devised ten semantic rules in their hybrid approach based on the presentation of BIBREF16 . We use five rules in the semantic rules set because other five rules are only used to compute polarity of words after POS tagging or Parsing steps. We follow the same naming convention for rules utilized by BIBREF11 to represent the rules utilized in our proposed method. The rules utilized in the proposed method are displayed in Table TABREF15 in which is included examples from STS Corpus and output after using the rules. Table TABREF16 illustrates the number of processed sentences on each dataset. Representation Levels To construct embedding inputs for our model, we use a fixed-sized word vocabulary INLINEFORM0 and a fixed-sized character vocabulary INLINEFORM1 . Given a word INLINEFORM2 is composed from characters INLINEFORM3 , the character-level embeddings are encoded by column vectors INLINEFORM4 in the embedding matrix INLINEFORM5 , where INLINEFORM6 is the size of the character vocabulary. For word-level embedding INLINEFORM7 , we use a pre-trained word-level embedding with dimension 200 or 300. A pre-trained word-level embedding can capture the syntactic and semantic information of words BIBREF17 . We build every word INLINEFORM8 into an embedding INLINEFORM9 which is constructed by two sub-vectors: the word-level embedding INLINEFORM10 and the character fixed-size feature vector INLINEFORM11 of INLINEFORM12 where INLINEFORM13 is the length of the filter of wide convolutions. We have INLINEFORM14 character fixed-size feature vectors corresponding to word-level embedding in a sentence. Deep Learning Module DeepCNN in the deep learning module is illustrated in Figure. FIGREF22 . The DeepCNN has two wide convolution layers. The first layer extract local features around each character windows of the given word and using a max pooling over character windows to produce a global fixed-sized feature vector for the word. The second layer retrieves important context characters and transforms the representation at previous level into a representation at higher abstract level. We have INLINEFORM0 global character fixed-sized feature vectors for INLINEFORM1 words. In the next step of Figure. FIGREF4 , we construct the vector INLINEFORM0 by concatenating the word-level embedding with the global character fixed-size feature vectors. The input of Bi-LSTM is a sequence of embeddings INLINEFORM1 . The use of the global character fixed-size feature vectors increases the relationship of words in the word-level embedding. The purpose of this Bi-LSTM is to capture the context of words in a sentence and maintain the order of words toward to extract sentence-level representation. The top of the model is a softmax function to predict sentiment label. We describe in detail the kinds of CNN and LSTM that we use in next sub-part 1 and 2. The one-dimensional convolution called time-delay neural net has a filter vector INLINEFORM0 and take the dot product of filter INLINEFORM1 with each m-grams in the sequence of characters INLINEFORM2 of a word in order to obtain a sequence INLINEFORM3 : DISPLAYFORM0 Based on Equation 1, we have two types of convolutions that depend on the range of the index INLINEFORM0 . The narrow type of convolution requires that INLINEFORM1 and produce a sequence INLINEFORM2 . The wide type of convolution does not require on INLINEFORM3 or INLINEFORM4 and produce a sequence INLINEFORM5 . Out-of-range input values INLINEFORM6 where INLINEFORM7 or INLINEFORM8 are taken to be zero. We use wide convolution for our model. Given a word INLINEFORM0 composed of INLINEFORM1 characters INLINEFORM2 , we take a character embedding INLINEFORM3 for each character INLINEFORM4 and construct a character matrix INLINEFORM5 as following Equation. 2: DISPLAYFORM0 The values of the embeddings INLINEFORM0 are parameters that are optimized during training. The trained weights in the filter INLINEFORM1 correspond to a feature detector which learns to recognize a specific class of n-grams. The n-grams have size INLINEFORM2 . The use of a wide convolution has some advantages more than a narrow convolution because a wide convolution ensures that all weights of filter reach the whole characters of a word at the margins. The resulting matrix has dimension INLINEFORM3 . Long Short-Term Memory networks usually called LSTMs are a improved version of RNN. The core idea behind LSTMs is the cell state which can maintain its state over time, and non-linear gating units which regulate the information flow into and out of the cell. The LSTM architecture that we used in our proposed model is described in BIBREF9 . A single LSTM memory cell is implemented by the following composite function: DISPLAYFORM0 DISPLAYFORM1 where INLINEFORM0 is the logistic sigmoid function, INLINEFORM1 and INLINEFORM2 are the input gate, forget gate, output gate, cell and cell input activation vectors respectively. All of them have a same size as the hidden vector INLINEFORM3 . INLINEFORM4 is the hidden-input gate matrix, INLINEFORM5 is the input-output gate matrix. The bias terms which are added to INLINEFORM6 and INLINEFORM7 have been omitted for clarity. In addition, we also use the full gradient for calculating with full backpropagation through time (BPTT) described in BIBREF10 . A LSTM gradients using finite differences could be checked and making practical implementations more reliable. Regularization For regularization, we use a constraint on INLINEFORM0 of the weight vectors BIBREF18 . Experimental setups For the Stanford Twitter Sentiment Corpus, we use the number of samples as BIBREF5 . The training data is selected 80K tweets for a training data and 16K tweets for the development set randomly from the training data of BIBREF0 . We conduct a binary prediction for STS Corpus. For Sander dataset, we use standard 10-fold cross validation as BIBREF14 . We construct the development set by selecting 10% randomly from 9-fold training data. In Health Care Reform Corpus, we also select 10% randomly for the development set in a training set and construct as BIBREF14 for comparison. We describe the summary of datasets in Table III. for all datasets, the filter window size ( INLINEFORM0 ) is 7 with 6 feature maps each for the first wide convolution layer, the second wide convolution layer has a filter window size of 5 with 14 feature maps each. Dropout rate ( INLINEFORM1 ) is 0.5, INLINEFORM2 constraint, learning rate is 0.1 and momentum of 0.9. Mini-batch size for STS Corpus is 100 and others are 4. In addition, training is done through stochastic gradient descent over shuffled mini-batches with Adadelta update rule BIBREF19 . we use the publicly available Word2Vec trained from 100 billion words from Google and TwitterGlove of Stanford is performed on aggregated global word-word co-occurrence statistics from a corpus. Word2Vec has dimensionality of 300 and Twitter Glove have dimensionality of 200. Words that do not present in the set of pre-train words are initialized randomly. Experimental results Table IV shows the result of our model for sentiment classification against other models. We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN. As can be seen, 86.63 is the best prediction accuracy of our model so far for the STS Corpus. For Sanders and HCR datasets, we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets. Our models outperform the model of BIBREF14 for the Sanders dataset and HCR dataset. Analysis As can be seen, the models with SR outperforms the model with no SR. Semantic rules is effective in order to increase classification accuracy. We evaluate the efficiency of SR for the model in Table V of our full paper . We also conduct two experiments on two separate models: DeepCNN and Bi-LSTM in order to show the effectiveness of combination of DeepCNN and Bi-LSTM. In addition, the model using TwitterGlove outperform the model using GoogleW2V because TwitterGlove captures more information in Twitter than GoogleW2V. These results show that the character-level information and SR have a great impact on Twitter Data. The pre-train word vectors are good, universal feature extractors. The difference between our model and other approaches is the ability of our model to capture important features by using SR and combine these features at high benefit. The use of DeepCNN can learn a representation of words in higher abstract level. The combination of global character fixed-sized feature vectors and a word embedding helps the model to find important detectors for particles such as 'not' that negate sentiment and potentiate sentiment such as 'too', 'so' standing beside expected features. The model not only learns to recognize single n-grams, but also patterns in n-grams lead to form a structure significance of a sentence. Conclusions In the present work, we have pointed out that the use of character embeddings through a DeepCNN to enhance information for word embeddings built on top of Word2Vec or TwitterGlove improves classification accuracy in Tweet sentiment classification. Our results add to the well-establish evidence that character vectors are an important ingredient for word-level in deep learning for NLP. In addition, semantic rules contribute handling non-essential sub-tweets in order to improve classification accuracy.

We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN., we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets.

846a1992d66d955fa1747bca9a139141c19908e8

846a1992d66d955fa1747bca9a139141c19908e8_0

Q: Which datasets did they use? Text: Introduction Twitter sentiment classification have intensively researched in recent years BIBREF0 BIBREF1 . Different approaches were developed for Twitter sentiment classification by using machine learning such as Support Vector Machine (SVM) with rule-based features BIBREF2 and the combination of SVMs and Naive Bayes (NB) BIBREF3 . In addition, hybrid approaches combining lexicon-based and machine learning methods also achieved high performance described in BIBREF4 . However, a problem of traditional machine learning is how to define a feature extractor for a specific domain in order to extract important features. Deep learning models are different from traditional machine learning methods in that a deep learning model does not depend on feature extractors because features are extracted during training progress. The use of deep learning methods becomes to achieve remarkable results for sentiment analysis BIBREF5 BIBREF6 BIBREF7 . Some researchers used Convolutional Neural Network (CNN) for sentiment classification. CNN models have been shown to be effective for NLP. For example, BIBREF6 proposed various kinds of CNN to learn sentiment-bearing sentence vectors, BIBREF5 adopted two CNNs in character-level to sentence-level representation for sentiment analysis. BIBREF7 constructs experiments on a character-level CNN for several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is another state-of-the-art semantic composition model for sentiment classification with many variants described in BIBREF8 . The studies reveal that using a CNN is useful in extracting information and finding feature detectors from texts. In addition, a LSTM can be good in maintaining word order and the context of words. However, in some important aspects, the use of CNN or LSTM separately may not capture enough information. Inspired by the models above, the goal of this research is using a Deep Convolutional Neural Network (DeepCNN) to exploit the information of characters of words in order to support word-level embedding. A Bi-LSTM produces a sentence-wide feature representation based on these embeddings. The Bi-LSTM is a version of BIBREF9 with Full Gradient described in BIBREF10 . In addition, the rules-based approach also effects classification accuracy by focusing on important sub-sentences expressing the main sentiment of a tweet while removing unnecessary parts of a tweet. The paper makes the following contributions: The organization of the present paper is as follows: In section 2, we describe the model architecture which introduces the structure of the model. We explain the basic idea of model and the way of constructing the model. Section 3 show results and analysis and section 4 summarize this paper. Basic idea Our proposed model consists of a deep learning classifier and a tweet processor. The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The tweet processor standardizes tweets and then applies semantic rules on datasets. We construct a framework that treats the deep learning classifier and the tweet processor as two distinct components. We believe that standardizing data is an important step to achieve high accuracy. To formulate our problem in increasing the accuracy of the classifier, we illustrate our model in Figure. FIGREF4 as follows: Tweets are firstly considered via a processor based on preprocessing steps BIBREF0 and the semantic rules-based method BIBREF11 in order to standardize tweets and capture only important information containing the main sentiment of a tweet. We use DeepCNN with Wide convolution for character-level embeddings. A wide convolution can learn to recognize specific n-grams at every position in a word that allows features to be extracted independently of these positions in the word. These features maintain the order and relative positions of characters. A DeepCNN is constructed by two wide convolution layers and the need of multiple wide convolution layers is widely accepted that a model constructing by multiple processing layers have the ability to learn representations of data with higher levels of abstraction BIBREF12 . Therefore, we use DeepCNN for character-level embeddings to support morphological and shape information for a word. The DeepCNN produces INLINEFORM0 global fixed-sized feature vectors for INLINEFORM1 words. A combination of the global fixed-size feature vectors and word-level embedding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level representation by maintaining the order of words. Our work is philosophically similar to BIBREF5 . However, our model is distinguished with their approaches in two aspects: Using DeepCNN with two wide convolution layers to increase representation with multiple levels of abstraction. Integrating global character fixed-sized feature vectors with word-level embedding to extract a sentence-wide feature set via Bi-LSTM. This deals with three main problems: (i) Sentences have any different size; (ii) The semantic and the syntactic of words in a sentence are captured in order to increase information for a word; (iii) Important information of characters that can appear at any position in a word are extracted. In sub-section B, we introduce various kinds of dataset. The modules of our model are constructed in other sub-sections. Data Preparation Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains 1,600K training tweets collected by a crawler from BIBREF0 . BIBREF0 constructed a test set manually with 177 negative and 182 positive tweets. The Stanford test set is small. However, it has been widely used in different evaluation tasks BIBREF0 BIBREF5 BIBREF13 . Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified tweets collected by using search terms: INLINEFORM0 , #google, #microsoft and #twitter. We construct the dataset as BIBREF14 for binary classification. Health Care Reform (HCR): This dataset was constructed by crawling tweets containing the hashtag #hcr BIBREF15 . Task is to predict positive/negative tweets BIBREF14 . Preprocessing We firstly take unique properties of Twitter in order to reduce the feature space such as Username, Usage of links, None, URLs and Repeated Letters. We then process retweets, stop words, links, URLs, mentions, punctuation and accentuation. For emoticons, BIBREF0 revealed that the training process makes the use of emoticons as noisy labels and they stripped the emoticons out from their training dataset because BIBREF0 believed that if we consider the emoticons, there is a negative impact on the accuracies of classifiers. In addition, removing emoticons makes the classifiers learns from other features (e.g. unigrams and bi-grams) presented in tweets and the classifiers only use these non-emoticon features to predict the sentiment of tweets. However, there is a problem is that if the test set contains emoticons, they do not influence the classifiers because emoticon features do not contain in its training data. This is a limitation of BIBREF0 , because the emoticon features would be useful when classifying test data. Therefore, we keep emoticon features in the datasets because deep learning models can capture more information from emoticon features for increasing classification accuracy. Semantic Rules (SR) In Twitter social networking, people express their opinions containing sub-sentences. These sub-sentences using specific PoS particles (Conjunction and Conjunctive adverbs), like "but, while, however, despite, however" have different polarities. However, the overall sentiment of tweets often focus on certain sub-sentences. For example: @lonedog bwahahah...you are amazing! However, it was quite the letdown. @kirstiealley my dentist is great but she's expensive...=( In two tweets above, the overall sentiment is negative. However, the main sentiment is only in the sub-sentences following but and however. This inspires a processing step to remove unessential parts in a tweet. Rule-based approach can assists these problems in handling negation and dealing with specific PoS particles led to effectively affect the final output of classification BIBREF11 BIBREF16 . BIBREF11 summarized a full presentation of their semantic rules approach and devised ten semantic rules in their hybrid approach based on the presentation of BIBREF16 . We use five rules in the semantic rules set because other five rules are only used to compute polarity of words after POS tagging or Parsing steps. We follow the same naming convention for rules utilized by BIBREF11 to represent the rules utilized in our proposed method. The rules utilized in the proposed method are displayed in Table TABREF15 in which is included examples from STS Corpus and output after using the rules. Table TABREF16 illustrates the number of processed sentences on each dataset. Representation Levels To construct embedding inputs for our model, we use a fixed-sized word vocabulary INLINEFORM0 and a fixed-sized character vocabulary INLINEFORM1 . Given a word INLINEFORM2 is composed from characters INLINEFORM3 , the character-level embeddings are encoded by column vectors INLINEFORM4 in the embedding matrix INLINEFORM5 , where INLINEFORM6 is the size of the character vocabulary. For word-level embedding INLINEFORM7 , we use a pre-trained word-level embedding with dimension 200 or 300. A pre-trained word-level embedding can capture the syntactic and semantic information of words BIBREF17 . We build every word INLINEFORM8 into an embedding INLINEFORM9 which is constructed by two sub-vectors: the word-level embedding INLINEFORM10 and the character fixed-size feature vector INLINEFORM11 of INLINEFORM12 where INLINEFORM13 is the length of the filter of wide convolutions. We have INLINEFORM14 character fixed-size feature vectors corresponding to word-level embedding in a sentence. Deep Learning Module DeepCNN in the deep learning module is illustrated in Figure. FIGREF22 . The DeepCNN has two wide convolution layers. The first layer extract local features around each character windows of the given word and using a max pooling over character windows to produce a global fixed-sized feature vector for the word. The second layer retrieves important context characters and transforms the representation at previous level into a representation at higher abstract level. We have INLINEFORM0 global character fixed-sized feature vectors for INLINEFORM1 words. In the next step of Figure. FIGREF4 , we construct the vector INLINEFORM0 by concatenating the word-level embedding with the global character fixed-size feature vectors. The input of Bi-LSTM is a sequence of embeddings INLINEFORM1 . The use of the global character fixed-size feature vectors increases the relationship of words in the word-level embedding. The purpose of this Bi-LSTM is to capture the context of words in a sentence and maintain the order of words toward to extract sentence-level representation. The top of the model is a softmax function to predict sentiment label. We describe in detail the kinds of CNN and LSTM that we use in next sub-part 1 and 2. The one-dimensional convolution called time-delay neural net has a filter vector INLINEFORM0 and take the dot product of filter INLINEFORM1 with each m-grams in the sequence of characters INLINEFORM2 of a word in order to obtain a sequence INLINEFORM3 : DISPLAYFORM0 Based on Equation 1, we have two types of convolutions that depend on the range of the index INLINEFORM0 . The narrow type of convolution requires that INLINEFORM1 and produce a sequence INLINEFORM2 . The wide type of convolution does not require on INLINEFORM3 or INLINEFORM4 and produce a sequence INLINEFORM5 . Out-of-range input values INLINEFORM6 where INLINEFORM7 or INLINEFORM8 are taken to be zero. We use wide convolution for our model. Given a word INLINEFORM0 composed of INLINEFORM1 characters INLINEFORM2 , we take a character embedding INLINEFORM3 for each character INLINEFORM4 and construct a character matrix INLINEFORM5 as following Equation. 2: DISPLAYFORM0 The values of the embeddings INLINEFORM0 are parameters that are optimized during training. The trained weights in the filter INLINEFORM1 correspond to a feature detector which learns to recognize a specific class of n-grams. The n-grams have size INLINEFORM2 . The use of a wide convolution has some advantages more than a narrow convolution because a wide convolution ensures that all weights of filter reach the whole characters of a word at the margins. The resulting matrix has dimension INLINEFORM3 . Long Short-Term Memory networks usually called LSTMs are a improved version of RNN. The core idea behind LSTMs is the cell state which can maintain its state over time, and non-linear gating units which regulate the information flow into and out of the cell. The LSTM architecture that we used in our proposed model is described in BIBREF9 . A single LSTM memory cell is implemented by the following composite function: DISPLAYFORM0 DISPLAYFORM1 where INLINEFORM0 is the logistic sigmoid function, INLINEFORM1 and INLINEFORM2 are the input gate, forget gate, output gate, cell and cell input activation vectors respectively. All of them have a same size as the hidden vector INLINEFORM3 . INLINEFORM4 is the hidden-input gate matrix, INLINEFORM5 is the input-output gate matrix. The bias terms which are added to INLINEFORM6 and INLINEFORM7 have been omitted for clarity. In addition, we also use the full gradient for calculating with full backpropagation through time (BPTT) described in BIBREF10 . A LSTM gradients using finite differences could be checked and making practical implementations more reliable. Regularization For regularization, we use a constraint on INLINEFORM0 of the weight vectors BIBREF18 . Experimental setups For the Stanford Twitter Sentiment Corpus, we use the number of samples as BIBREF5 . The training data is selected 80K tweets for a training data and 16K tweets for the development set randomly from the training data of BIBREF0 . We conduct a binary prediction for STS Corpus. For Sander dataset, we use standard 10-fold cross validation as BIBREF14 . We construct the development set by selecting 10% randomly from 9-fold training data. In Health Care Reform Corpus, we also select 10% randomly for the development set in a training set and construct as BIBREF14 for comparison. We describe the summary of datasets in Table III. for all datasets, the filter window size ( INLINEFORM0 ) is 7 with 6 feature maps each for the first wide convolution layer, the second wide convolution layer has a filter window size of 5 with 14 feature maps each. Dropout rate ( INLINEFORM1 ) is 0.5, INLINEFORM2 constraint, learning rate is 0.1 and momentum of 0.9. Mini-batch size for STS Corpus is 100 and others are 4. In addition, training is done through stochastic gradient descent over shuffled mini-batches with Adadelta update rule BIBREF19 . we use the publicly available Word2Vec trained from 100 billion words from Google and TwitterGlove of Stanford is performed on aggregated global word-word co-occurrence statistics from a corpus. Word2Vec has dimensionality of 300 and Twitter Glove have dimensionality of 200. Words that do not present in the set of pre-train words are initialized randomly. Experimental results Table IV shows the result of our model for sentiment classification against other models. We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN. As can be seen, 86.63 is the best prediction accuracy of our model so far for the STS Corpus. For Sanders and HCR datasets, we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets. Our models outperform the model of BIBREF14 for the Sanders dataset and HCR dataset. Analysis As can be seen, the models with SR outperforms the model with no SR. Semantic rules is effective in order to increase classification accuracy. We evaluate the efficiency of SR for the model in Table V of our full paper . We also conduct two experiments on two separate models: DeepCNN and Bi-LSTM in order to show the effectiveness of combination of DeepCNN and Bi-LSTM. In addition, the model using TwitterGlove outperform the model using GoogleW2V because TwitterGlove captures more information in Twitter than GoogleW2V. These results show that the character-level information and SR have a great impact on Twitter Data. The pre-train word vectors are good, universal feature extractors. The difference between our model and other approaches is the ability of our model to capture important features by using SR and combine these features at high benefit. The use of DeepCNN can learn a representation of words in higher abstract level. The combination of global character fixed-sized feature vectors and a word embedding helps the model to find important detectors for particles such as 'not' that negate sentiment and potentiate sentiment such as 'too', 'so' standing beside expected features. The model not only learns to recognize single n-grams, but also patterns in n-grams lead to form a structure significance of a sentence. Conclusions In the present work, we have pointed out that the use of character embeddings through a DeepCNN to enhance information for word embeddings built on top of Word2Vec or TwitterGlove improves classification accuracy in Tweet sentiment classification. Our results add to the well-establish evidence that character vectors are an important ingredient for word-level in deep learning for NLP. In addition, semantic rules contribute handling non-essential sub-tweets in order to improve classification accuracy.

Stanford - Twitter Sentiment Corpus (STS Corpus), Sanders - Twitter Sentiment Corpus, Health Care Reform (HCR)

1ef8d1cb1199e1504b6b0daea52f2e4bd2ef7023

1ef8d1cb1199e1504b6b0daea52f2e4bd2ef7023_0

Q: Are results reported only on English datasets? Text: Introduction Twitter sentiment classification have intensively researched in recent years BIBREF0 BIBREF1 . Different approaches were developed for Twitter sentiment classification by using machine learning such as Support Vector Machine (SVM) with rule-based features BIBREF2 and the combination of SVMs and Naive Bayes (NB) BIBREF3 . In addition, hybrid approaches combining lexicon-based and machine learning methods also achieved high performance described in BIBREF4 . However, a problem of traditional machine learning is how to define a feature extractor for a specific domain in order to extract important features. Deep learning models are different from traditional machine learning methods in that a deep learning model does not depend on feature extractors because features are extracted during training progress. The use of deep learning methods becomes to achieve remarkable results for sentiment analysis BIBREF5 BIBREF6 BIBREF7 . Some researchers used Convolutional Neural Network (CNN) for sentiment classification. CNN models have been shown to be effective for NLP. For example, BIBREF6 proposed various kinds of CNN to learn sentiment-bearing sentence vectors, BIBREF5 adopted two CNNs in character-level to sentence-level representation for sentiment analysis. BIBREF7 constructs experiments on a character-level CNN for several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is another state-of-the-art semantic composition model for sentiment classification with many variants described in BIBREF8 . The studies reveal that using a CNN is useful in extracting information and finding feature detectors from texts. In addition, a LSTM can be good in maintaining word order and the context of words. However, in some important aspects, the use of CNN or LSTM separately may not capture enough information. Inspired by the models above, the goal of this research is using a Deep Convolutional Neural Network (DeepCNN) to exploit the information of characters of words in order to support word-level embedding. A Bi-LSTM produces a sentence-wide feature representation based on these embeddings. The Bi-LSTM is a version of BIBREF9 with Full Gradient described in BIBREF10 . In addition, the rules-based approach also effects classification accuracy by focusing on important sub-sentences expressing the main sentiment of a tweet while removing unnecessary parts of a tweet. The paper makes the following contributions: The organization of the present paper is as follows: In section 2, we describe the model architecture which introduces the structure of the model. We explain the basic idea of model and the way of constructing the model. Section 3 show results and analysis and section 4 summarize this paper. Basic idea Our proposed model consists of a deep learning classifier and a tweet processor. The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The tweet processor standardizes tweets and then applies semantic rules on datasets. We construct a framework that treats the deep learning classifier and the tweet processor as two distinct components. We believe that standardizing data is an important step to achieve high accuracy. To formulate our problem in increasing the accuracy of the classifier, we illustrate our model in Figure. FIGREF4 as follows: Tweets are firstly considered via a processor based on preprocessing steps BIBREF0 and the semantic rules-based method BIBREF11 in order to standardize tweets and capture only important information containing the main sentiment of a tweet. We use DeepCNN with Wide convolution for character-level embeddings. A wide convolution can learn to recognize specific n-grams at every position in a word that allows features to be extracted independently of these positions in the word. These features maintain the order and relative positions of characters. A DeepCNN is constructed by two wide convolution layers and the need of multiple wide convolution layers is widely accepted that a model constructing by multiple processing layers have the ability to learn representations of data with higher levels of abstraction BIBREF12 . Therefore, we use DeepCNN for character-level embeddings to support morphological and shape information for a word. The DeepCNN produces INLINEFORM0 global fixed-sized feature vectors for INLINEFORM1 words. A combination of the global fixed-size feature vectors and word-level embedding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level representation by maintaining the order of words. Our work is philosophically similar to BIBREF5 . However, our model is distinguished with their approaches in two aspects: Using DeepCNN with two wide convolution layers to increase representation with multiple levels of abstraction. Integrating global character fixed-sized feature vectors with word-level embedding to extract a sentence-wide feature set via Bi-LSTM. This deals with three main problems: (i) Sentences have any different size; (ii) The semantic and the syntactic of words in a sentence are captured in order to increase information for a word; (iii) Important information of characters that can appear at any position in a word are extracted. In sub-section B, we introduce various kinds of dataset. The modules of our model are constructed in other sub-sections. Data Preparation Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains 1,600K training tweets collected by a crawler from BIBREF0 . BIBREF0 constructed a test set manually with 177 negative and 182 positive tweets. The Stanford test set is small. However, it has been widely used in different evaluation tasks BIBREF0 BIBREF5 BIBREF13 . Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified tweets collected by using search terms: INLINEFORM0 , #google, #microsoft and #twitter. We construct the dataset as BIBREF14 for binary classification. Health Care Reform (HCR): This dataset was constructed by crawling tweets containing the hashtag #hcr BIBREF15 . Task is to predict positive/negative tweets BIBREF14 . Preprocessing We firstly take unique properties of Twitter in order to reduce the feature space such as Username, Usage of links, None, URLs and Repeated Letters. We then process retweets, stop words, links, URLs, mentions, punctuation and accentuation. For emoticons, BIBREF0 revealed that the training process makes the use of emoticons as noisy labels and they stripped the emoticons out from their training dataset because BIBREF0 believed that if we consider the emoticons, there is a negative impact on the accuracies of classifiers. In addition, removing emoticons makes the classifiers learns from other features (e.g. unigrams and bi-grams) presented in tweets and the classifiers only use these non-emoticon features to predict the sentiment of tweets. However, there is a problem is that if the test set contains emoticons, they do not influence the classifiers because emoticon features do not contain in its training data. This is a limitation of BIBREF0 , because the emoticon features would be useful when classifying test data. Therefore, we keep emoticon features in the datasets because deep learning models can capture more information from emoticon features for increasing classification accuracy. Semantic Rules (SR) In Twitter social networking, people express their opinions containing sub-sentences. These sub-sentences using specific PoS particles (Conjunction and Conjunctive adverbs), like "but, while, however, despite, however" have different polarities. However, the overall sentiment of tweets often focus on certain sub-sentences. For example: @lonedog bwahahah...you are amazing! However, it was quite the letdown. @kirstiealley my dentist is great but she's expensive...=( In two tweets above, the overall sentiment is negative. However, the main sentiment is only in the sub-sentences following but and however. This inspires a processing step to remove unessential parts in a tweet. Rule-based approach can assists these problems in handling negation and dealing with specific PoS particles led to effectively affect the final output of classification BIBREF11 BIBREF16 . BIBREF11 summarized a full presentation of their semantic rules approach and devised ten semantic rules in their hybrid approach based on the presentation of BIBREF16 . We use five rules in the semantic rules set because other five rules are only used to compute polarity of words after POS tagging or Parsing steps. We follow the same naming convention for rules utilized by BIBREF11 to represent the rules utilized in our proposed method. The rules utilized in the proposed method are displayed in Table TABREF15 in which is included examples from STS Corpus and output after using the rules. Table TABREF16 illustrates the number of processed sentences on each dataset. Representation Levels To construct embedding inputs for our model, we use a fixed-sized word vocabulary INLINEFORM0 and a fixed-sized character vocabulary INLINEFORM1 . Given a word INLINEFORM2 is composed from characters INLINEFORM3 , the character-level embeddings are encoded by column vectors INLINEFORM4 in the embedding matrix INLINEFORM5 , where INLINEFORM6 is the size of the character vocabulary. For word-level embedding INLINEFORM7 , we use a pre-trained word-level embedding with dimension 200 or 300. A pre-trained word-level embedding can capture the syntactic and semantic information of words BIBREF17 . We build every word INLINEFORM8 into an embedding INLINEFORM9 which is constructed by two sub-vectors: the word-level embedding INLINEFORM10 and the character fixed-size feature vector INLINEFORM11 of INLINEFORM12 where INLINEFORM13 is the length of the filter of wide convolutions. We have INLINEFORM14 character fixed-size feature vectors corresponding to word-level embedding in a sentence. Deep Learning Module DeepCNN in the deep learning module is illustrated in Figure. FIGREF22 . The DeepCNN has two wide convolution layers. The first layer extract local features around each character windows of the given word and using a max pooling over character windows to produce a global fixed-sized feature vector for the word. The second layer retrieves important context characters and transforms the representation at previous level into a representation at higher abstract level. We have INLINEFORM0 global character fixed-sized feature vectors for INLINEFORM1 words. In the next step of Figure. FIGREF4 , we construct the vector INLINEFORM0 by concatenating the word-level embedding with the global character fixed-size feature vectors. The input of Bi-LSTM is a sequence of embeddings INLINEFORM1 . The use of the global character fixed-size feature vectors increases the relationship of words in the word-level embedding. The purpose of this Bi-LSTM is to capture the context of words in a sentence and maintain the order of words toward to extract sentence-level representation. The top of the model is a softmax function to predict sentiment label. We describe in detail the kinds of CNN and LSTM that we use in next sub-part 1 and 2. The one-dimensional convolution called time-delay neural net has a filter vector INLINEFORM0 and take the dot product of filter INLINEFORM1 with each m-grams in the sequence of characters INLINEFORM2 of a word in order to obtain a sequence INLINEFORM3 : DISPLAYFORM0 Based on Equation 1, we have two types of convolutions that depend on the range of the index INLINEFORM0 . The narrow type of convolution requires that INLINEFORM1 and produce a sequence INLINEFORM2 . The wide type of convolution does not require on INLINEFORM3 or INLINEFORM4 and produce a sequence INLINEFORM5 . Out-of-range input values INLINEFORM6 where INLINEFORM7 or INLINEFORM8 are taken to be zero. We use wide convolution for our model. Given a word INLINEFORM0 composed of INLINEFORM1 characters INLINEFORM2 , we take a character embedding INLINEFORM3 for each character INLINEFORM4 and construct a character matrix INLINEFORM5 as following Equation. 2: DISPLAYFORM0 The values of the embeddings INLINEFORM0 are parameters that are optimized during training. The trained weights in the filter INLINEFORM1 correspond to a feature detector which learns to recognize a specific class of n-grams. The n-grams have size INLINEFORM2 . The use of a wide convolution has some advantages more than a narrow convolution because a wide convolution ensures that all weights of filter reach the whole characters of a word at the margins. The resulting matrix has dimension INLINEFORM3 . Long Short-Term Memory networks usually called LSTMs are a improved version of RNN. The core idea behind LSTMs is the cell state which can maintain its state over time, and non-linear gating units which regulate the information flow into and out of the cell. The LSTM architecture that we used in our proposed model is described in BIBREF9 . A single LSTM memory cell is implemented by the following composite function: DISPLAYFORM0 DISPLAYFORM1 where INLINEFORM0 is the logistic sigmoid function, INLINEFORM1 and INLINEFORM2 are the input gate, forget gate, output gate, cell and cell input activation vectors respectively. All of them have a same size as the hidden vector INLINEFORM3 . INLINEFORM4 is the hidden-input gate matrix, INLINEFORM5 is the input-output gate matrix. The bias terms which are added to INLINEFORM6 and INLINEFORM7 have been omitted for clarity. In addition, we also use the full gradient for calculating with full backpropagation through time (BPTT) described in BIBREF10 . A LSTM gradients using finite differences could be checked and making practical implementations more reliable. Regularization For regularization, we use a constraint on INLINEFORM0 of the weight vectors BIBREF18 . Experimental setups For the Stanford Twitter Sentiment Corpus, we use the number of samples as BIBREF5 . The training data is selected 80K tweets for a training data and 16K tweets for the development set randomly from the training data of BIBREF0 . We conduct a binary prediction for STS Corpus. For Sander dataset, we use standard 10-fold cross validation as BIBREF14 . We construct the development set by selecting 10% randomly from 9-fold training data. In Health Care Reform Corpus, we also select 10% randomly for the development set in a training set and construct as BIBREF14 for comparison. We describe the summary of datasets in Table III. for all datasets, the filter window size ( INLINEFORM0 ) is 7 with 6 feature maps each for the first wide convolution layer, the second wide convolution layer has a filter window size of 5 with 14 feature maps each. Dropout rate ( INLINEFORM1 ) is 0.5, INLINEFORM2 constraint, learning rate is 0.1 and momentum of 0.9. Mini-batch size for STS Corpus is 100 and others are 4. In addition, training is done through stochastic gradient descent over shuffled mini-batches with Adadelta update rule BIBREF19 . we use the publicly available Word2Vec trained from 100 billion words from Google and TwitterGlove of Stanford is performed on aggregated global word-word co-occurrence statistics from a corpus. Word2Vec has dimensionality of 300 and Twitter Glove have dimensionality of 200. Words that do not present in the set of pre-train words are initialized randomly. Experimental results Table IV shows the result of our model for sentiment classification against other models. We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN. As can be seen, 86.63 is the best prediction accuracy of our model so far for the STS Corpus. For Sanders and HCR datasets, we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets. Our models outperform the model of BIBREF14 for the Sanders dataset and HCR dataset. Analysis As can be seen, the models with SR outperforms the model with no SR. Semantic rules is effective in order to increase classification accuracy. We evaluate the efficiency of SR for the model in Table V of our full paper . We also conduct two experiments on two separate models: DeepCNN and Bi-LSTM in order to show the effectiveness of combination of DeepCNN and Bi-LSTM. In addition, the model using TwitterGlove outperform the model using GoogleW2V because TwitterGlove captures more information in Twitter than GoogleW2V. These results show that the character-level information and SR have a great impact on Twitter Data. The pre-train word vectors are good, universal feature extractors. The difference between our model and other approaches is the ability of our model to capture important features by using SR and combine these features at high benefit. The use of DeepCNN can learn a representation of words in higher abstract level. The combination of global character fixed-sized feature vectors and a word embedding helps the model to find important detectors for particles such as 'not' that negate sentiment and potentiate sentiment such as 'too', 'so' standing beside expected features. The model not only learns to recognize single n-grams, but also patterns in n-grams lead to form a structure significance of a sentence. Conclusions In the present work, we have pointed out that the use of character embeddings through a DeepCNN to enhance information for word embeddings built on top of Word2Vec or TwitterGlove improves classification accuracy in Tweet sentiment classification. Our results add to the well-establish evidence that character vectors are an important ingredient for word-level in deep learning for NLP. In addition, semantic rules contribute handling non-essential sub-tweets in order to improve classification accuracy.

Yes

12d77ac09c659d2e04b5e3955a283101c3ad1058

12d77ac09c659d2e04b5e3955a283101c3ad1058_0

Q: Which three Twitter sentiment classification datasets are used for experiments? Text: Introduction Twitter sentiment classification have intensively researched in recent years BIBREF0 BIBREF1 . Different approaches were developed for Twitter sentiment classification by using machine learning such as Support Vector Machine (SVM) with rule-based features BIBREF2 and the combination of SVMs and Naive Bayes (NB) BIBREF3 . In addition, hybrid approaches combining lexicon-based and machine learning methods also achieved high performance described in BIBREF4 . However, a problem of traditional machine learning is how to define a feature extractor for a specific domain in order to extract important features. Deep learning models are different from traditional machine learning methods in that a deep learning model does not depend on feature extractors because features are extracted during training progress. The use of deep learning methods becomes to achieve remarkable results for sentiment analysis BIBREF5 BIBREF6 BIBREF7 . Some researchers used Convolutional Neural Network (CNN) for sentiment classification. CNN models have been shown to be effective for NLP. For example, BIBREF6 proposed various kinds of CNN to learn sentiment-bearing sentence vectors, BIBREF5 adopted two CNNs in character-level to sentence-level representation for sentiment analysis. BIBREF7 constructs experiments on a character-level CNN for several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is another state-of-the-art semantic composition model for sentiment classification with many variants described in BIBREF8 . The studies reveal that using a CNN is useful in extracting information and finding feature detectors from texts. In addition, a LSTM can be good in maintaining word order and the context of words. However, in some important aspects, the use of CNN or LSTM separately may not capture enough information. Inspired by the models above, the goal of this research is using a Deep Convolutional Neural Network (DeepCNN) to exploit the information of characters of words in order to support word-level embedding. A Bi-LSTM produces a sentence-wide feature representation based on these embeddings. The Bi-LSTM is a version of BIBREF9 with Full Gradient described in BIBREF10 . In addition, the rules-based approach also effects classification accuracy by focusing on important sub-sentences expressing the main sentiment of a tweet while removing unnecessary parts of a tweet. The paper makes the following contributions: The organization of the present paper is as follows: In section 2, we describe the model architecture which introduces the structure of the model. We explain the basic idea of model and the way of constructing the model. Section 3 show results and analysis and section 4 summarize this paper. Basic idea Our proposed model consists of a deep learning classifier and a tweet processor. The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The tweet processor standardizes tweets and then applies semantic rules on datasets. We construct a framework that treats the deep learning classifier and the tweet processor as two distinct components. We believe that standardizing data is an important step to achieve high accuracy. To formulate our problem in increasing the accuracy of the classifier, we illustrate our model in Figure. FIGREF4 as follows: Tweets are firstly considered via a processor based on preprocessing steps BIBREF0 and the semantic rules-based method BIBREF11 in order to standardize tweets and capture only important information containing the main sentiment of a tweet. We use DeepCNN with Wide convolution for character-level embeddings. A wide convolution can learn to recognize specific n-grams at every position in a word that allows features to be extracted independently of these positions in the word. These features maintain the order and relative positions of characters. A DeepCNN is constructed by two wide convolution layers and the need of multiple wide convolution layers is widely accepted that a model constructing by multiple processing layers have the ability to learn representations of data with higher levels of abstraction BIBREF12 . Therefore, we use DeepCNN for character-level embeddings to support morphological and shape information for a word. The DeepCNN produces INLINEFORM0 global fixed-sized feature vectors for INLINEFORM1 words. A combination of the global fixed-size feature vectors and word-level embedding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level representation by maintaining the order of words. Our work is philosophically similar to BIBREF5 . However, our model is distinguished with their approaches in two aspects: Using DeepCNN with two wide convolution layers to increase representation with multiple levels of abstraction. Integrating global character fixed-sized feature vectors with word-level embedding to extract a sentence-wide feature set via Bi-LSTM. This deals with three main problems: (i) Sentences have any different size; (ii) The semantic and the syntactic of words in a sentence are captured in order to increase information for a word; (iii) Important information of characters that can appear at any position in a word are extracted. In sub-section B, we introduce various kinds of dataset. The modules of our model are constructed in other sub-sections. Data Preparation Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains 1,600K training tweets collected by a crawler from BIBREF0 . BIBREF0 constructed a test set manually with 177 negative and 182 positive tweets. The Stanford test set is small. However, it has been widely used in different evaluation tasks BIBREF0 BIBREF5 BIBREF13 . Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified tweets collected by using search terms: INLINEFORM0 , #google, #microsoft and #twitter. We construct the dataset as BIBREF14 for binary classification. Health Care Reform (HCR): This dataset was constructed by crawling tweets containing the hashtag #hcr BIBREF15 . Task is to predict positive/negative tweets BIBREF14 . Preprocessing We firstly take unique properties of Twitter in order to reduce the feature space such as Username, Usage of links, None, URLs and Repeated Letters. We then process retweets, stop words, links, URLs, mentions, punctuation and accentuation. For emoticons, BIBREF0 revealed that the training process makes the use of emoticons as noisy labels and they stripped the emoticons out from their training dataset because BIBREF0 believed that if we consider the emoticons, there is a negative impact on the accuracies of classifiers. In addition, removing emoticons makes the classifiers learns from other features (e.g. unigrams and bi-grams) presented in tweets and the classifiers only use these non-emoticon features to predict the sentiment of tweets. However, there is a problem is that if the test set contains emoticons, they do not influence the classifiers because emoticon features do not contain in its training data. This is a limitation of BIBREF0 , because the emoticon features would be useful when classifying test data. Therefore, we keep emoticon features in the datasets because deep learning models can capture more information from emoticon features for increasing classification accuracy. Semantic Rules (SR) In Twitter social networking, people express their opinions containing sub-sentences. These sub-sentences using specific PoS particles (Conjunction and Conjunctive adverbs), like "but, while, however, despite, however" have different polarities. However, the overall sentiment of tweets often focus on certain sub-sentences. For example: @lonedog bwahahah...you are amazing! However, it was quite the letdown. @kirstiealley my dentist is great but she's expensive...=( In two tweets above, the overall sentiment is negative. However, the main sentiment is only in the sub-sentences following but and however. This inspires a processing step to remove unessential parts in a tweet. Rule-based approach can assists these problems in handling negation and dealing with specific PoS particles led to effectively affect the final output of classification BIBREF11 BIBREF16 . BIBREF11 summarized a full presentation of their semantic rules approach and devised ten semantic rules in their hybrid approach based on the presentation of BIBREF16 . We use five rules in the semantic rules set because other five rules are only used to compute polarity of words after POS tagging or Parsing steps. We follow the same naming convention for rules utilized by BIBREF11 to represent the rules utilized in our proposed method. The rules utilized in the proposed method are displayed in Table TABREF15 in which is included examples from STS Corpus and output after using the rules. Table TABREF16 illustrates the number of processed sentences on each dataset. Representation Levels To construct embedding inputs for our model, we use a fixed-sized word vocabulary INLINEFORM0 and a fixed-sized character vocabulary INLINEFORM1 . Given a word INLINEFORM2 is composed from characters INLINEFORM3 , the character-level embeddings are encoded by column vectors INLINEFORM4 in the embedding matrix INLINEFORM5 , where INLINEFORM6 is the size of the character vocabulary. For word-level embedding INLINEFORM7 , we use a pre-trained word-level embedding with dimension 200 or 300. A pre-trained word-level embedding can capture the syntactic and semantic information of words BIBREF17 . We build every word INLINEFORM8 into an embedding INLINEFORM9 which is constructed by two sub-vectors: the word-level embedding INLINEFORM10 and the character fixed-size feature vector INLINEFORM11 of INLINEFORM12 where INLINEFORM13 is the length of the filter of wide convolutions. We have INLINEFORM14 character fixed-size feature vectors corresponding to word-level embedding in a sentence. Deep Learning Module DeepCNN in the deep learning module is illustrated in Figure. FIGREF22 . The DeepCNN has two wide convolution layers. The first layer extract local features around each character windows of the given word and using a max pooling over character windows to produce a global fixed-sized feature vector for the word. The second layer retrieves important context characters and transforms the representation at previous level into a representation at higher abstract level. We have INLINEFORM0 global character fixed-sized feature vectors for INLINEFORM1 words. In the next step of Figure. FIGREF4 , we construct the vector INLINEFORM0 by concatenating the word-level embedding with the global character fixed-size feature vectors. The input of Bi-LSTM is a sequence of embeddings INLINEFORM1 . The use of the global character fixed-size feature vectors increases the relationship of words in the word-level embedding. The purpose of this Bi-LSTM is to capture the context of words in a sentence and maintain the order of words toward to extract sentence-level representation. The top of the model is a softmax function to predict sentiment label. We describe in detail the kinds of CNN and LSTM that we use in next sub-part 1 and 2. The one-dimensional convolution called time-delay neural net has a filter vector INLINEFORM0 and take the dot product of filter INLINEFORM1 with each m-grams in the sequence of characters INLINEFORM2 of a word in order to obtain a sequence INLINEFORM3 : DISPLAYFORM0 Based on Equation 1, we have two types of convolutions that depend on the range of the index INLINEFORM0 . The narrow type of convolution requires that INLINEFORM1 and produce a sequence INLINEFORM2 . The wide type of convolution does not require on INLINEFORM3 or INLINEFORM4 and produce a sequence INLINEFORM5 . Out-of-range input values INLINEFORM6 where INLINEFORM7 or INLINEFORM8 are taken to be zero. We use wide convolution for our model. Given a word INLINEFORM0 composed of INLINEFORM1 characters INLINEFORM2 , we take a character embedding INLINEFORM3 for each character INLINEFORM4 and construct a character matrix INLINEFORM5 as following Equation. 2: DISPLAYFORM0 The values of the embeddings INLINEFORM0 are parameters that are optimized during training. The trained weights in the filter INLINEFORM1 correspond to a feature detector which learns to recognize a specific class of n-grams. The n-grams have size INLINEFORM2 . The use of a wide convolution has some advantages more than a narrow convolution because a wide convolution ensures that all weights of filter reach the whole characters of a word at the margins. The resulting matrix has dimension INLINEFORM3 . Long Short-Term Memory networks usually called LSTMs are a improved version of RNN. The core idea behind LSTMs is the cell state which can maintain its state over time, and non-linear gating units which regulate the information flow into and out of the cell. The LSTM architecture that we used in our proposed model is described in BIBREF9 . A single LSTM memory cell is implemented by the following composite function: DISPLAYFORM0 DISPLAYFORM1 where INLINEFORM0 is the logistic sigmoid function, INLINEFORM1 and INLINEFORM2 are the input gate, forget gate, output gate, cell and cell input activation vectors respectively. All of them have a same size as the hidden vector INLINEFORM3 . INLINEFORM4 is the hidden-input gate matrix, INLINEFORM5 is the input-output gate matrix. The bias terms which are added to INLINEFORM6 and INLINEFORM7 have been omitted for clarity. In addition, we also use the full gradient for calculating with full backpropagation through time (BPTT) described in BIBREF10 . A LSTM gradients using finite differences could be checked and making practical implementations more reliable. Regularization For regularization, we use a constraint on INLINEFORM0 of the weight vectors BIBREF18 . Experimental setups For the Stanford Twitter Sentiment Corpus, we use the number of samples as BIBREF5 . The training data is selected 80K tweets for a training data and 16K tweets for the development set randomly from the training data of BIBREF0 . We conduct a binary prediction for STS Corpus. For Sander dataset, we use standard 10-fold cross validation as BIBREF14 . We construct the development set by selecting 10% randomly from 9-fold training data. In Health Care Reform Corpus, we also select 10% randomly for the development set in a training set and construct as BIBREF14 for comparison. We describe the summary of datasets in Table III. for all datasets, the filter window size ( INLINEFORM0 ) is 7 with 6 feature maps each for the first wide convolution layer, the second wide convolution layer has a filter window size of 5 with 14 feature maps each. Dropout rate ( INLINEFORM1 ) is 0.5, INLINEFORM2 constraint, learning rate is 0.1 and momentum of 0.9. Mini-batch size for STS Corpus is 100 and others are 4. In addition, training is done through stochastic gradient descent over shuffled mini-batches with Adadelta update rule BIBREF19 . we use the publicly available Word2Vec trained from 100 billion words from Google and TwitterGlove of Stanford is performed on aggregated global word-word co-occurrence statistics from a corpus. Word2Vec has dimensionality of 300 and Twitter Glove have dimensionality of 200. Words that do not present in the set of pre-train words are initialized randomly. Experimental results Table IV shows the result of our model for sentiment classification against other models. We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN. As can be seen, 86.63 is the best prediction accuracy of our model so far for the STS Corpus. For Sanders and HCR datasets, we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets. Our models outperform the model of BIBREF14 for the Sanders dataset and HCR dataset. Analysis As can be seen, the models with SR outperforms the model with no SR. Semantic rules is effective in order to increase classification accuracy. We evaluate the efficiency of SR for the model in Table V of our full paper . We also conduct two experiments on two separate models: DeepCNN and Bi-LSTM in order to show the effectiveness of combination of DeepCNN and Bi-LSTM. In addition, the model using TwitterGlove outperform the model using GoogleW2V because TwitterGlove captures more information in Twitter than GoogleW2V. These results show that the character-level information and SR have a great impact on Twitter Data. The pre-train word vectors are good, universal feature extractors. The difference between our model and other approaches is the ability of our model to capture important features by using SR and combine these features at high benefit. The use of DeepCNN can learn a representation of words in higher abstract level. The combination of global character fixed-sized feature vectors and a word embedding helps the model to find important detectors for particles such as 'not' that negate sentiment and potentiate sentiment such as 'too', 'so' standing beside expected features. The model not only learns to recognize single n-grams, but also patterns in n-grams lead to form a structure significance of a sentence. Conclusions In the present work, we have pointed out that the use of character embeddings through a DeepCNN to enhance information for word embeddings built on top of Word2Vec or TwitterGlove improves classification accuracy in Tweet sentiment classification. Our results add to the well-establish evidence that character vectors are an important ingredient for word-level in deep learning for NLP. In addition, semantic rules contribute handling non-essential sub-tweets in order to improve classification accuracy.

Stanford - Twitter Sentiment Corpus (STS Corpus), Sanders - Twitter Sentiment Corpus, Health Care Reform (HCR)

d60a3887a0d434abc0861637bbcd9ad0c596caf4

d60a3887a0d434abc0861637bbcd9ad0c596caf4_0

Q: What semantic rules are proposed? Text: Introduction Twitter sentiment classification have intensively researched in recent years BIBREF0 BIBREF1 . Different approaches were developed for Twitter sentiment classification by using machine learning such as Support Vector Machine (SVM) with rule-based features BIBREF2 and the combination of SVMs and Naive Bayes (NB) BIBREF3 . In addition, hybrid approaches combining lexicon-based and machine learning methods also achieved high performance described in BIBREF4 . However, a problem of traditional machine learning is how to define a feature extractor for a specific domain in order to extract important features. Deep learning models are different from traditional machine learning methods in that a deep learning model does not depend on feature extractors because features are extracted during training progress. The use of deep learning methods becomes to achieve remarkable results for sentiment analysis BIBREF5 BIBREF6 BIBREF7 . Some researchers used Convolutional Neural Network (CNN) for sentiment classification. CNN models have been shown to be effective for NLP. For example, BIBREF6 proposed various kinds of CNN to learn sentiment-bearing sentence vectors, BIBREF5 adopted two CNNs in character-level to sentence-level representation for sentiment analysis. BIBREF7 constructs experiments on a character-level CNN for several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is another state-of-the-art semantic composition model for sentiment classification with many variants described in BIBREF8 . The studies reveal that using a CNN is useful in extracting information and finding feature detectors from texts. In addition, a LSTM can be good in maintaining word order and the context of words. However, in some important aspects, the use of CNN or LSTM separately may not capture enough information. Inspired by the models above, the goal of this research is using a Deep Convolutional Neural Network (DeepCNN) to exploit the information of characters of words in order to support word-level embedding. A Bi-LSTM produces a sentence-wide feature representation based on these embeddings. The Bi-LSTM is a version of BIBREF9 with Full Gradient described in BIBREF10 . In addition, the rules-based approach also effects classification accuracy by focusing on important sub-sentences expressing the main sentiment of a tweet while removing unnecessary parts of a tweet. The paper makes the following contributions: The organization of the present paper is as follows: In section 2, we describe the model architecture which introduces the structure of the model. We explain the basic idea of model and the way of constructing the model. Section 3 show results and analysis and section 4 summarize this paper. Basic idea Our proposed model consists of a deep learning classifier and a tweet processor. The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The tweet processor standardizes tweets and then applies semantic rules on datasets. We construct a framework that treats the deep learning classifier and the tweet processor as two distinct components. We believe that standardizing data is an important step to achieve high accuracy. To formulate our problem in increasing the accuracy of the classifier, we illustrate our model in Figure. FIGREF4 as follows: Tweets are firstly considered via a processor based on preprocessing steps BIBREF0 and the semantic rules-based method BIBREF11 in order to standardize tweets and capture only important information containing the main sentiment of a tweet. We use DeepCNN with Wide convolution for character-level embeddings. A wide convolution can learn to recognize specific n-grams at every position in a word that allows features to be extracted independently of these positions in the word. These features maintain the order and relative positions of characters. A DeepCNN is constructed by two wide convolution layers and the need of multiple wide convolution layers is widely accepted that a model constructing by multiple processing layers have the ability to learn representations of data with higher levels of abstraction BIBREF12 . Therefore, we use DeepCNN for character-level embeddings to support morphological and shape information for a word. The DeepCNN produces INLINEFORM0 global fixed-sized feature vectors for INLINEFORM1 words. A combination of the global fixed-size feature vectors and word-level embedding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level representation by maintaining the order of words. Our work is philosophically similar to BIBREF5 . However, our model is distinguished with their approaches in two aspects: Using DeepCNN with two wide convolution layers to increase representation with multiple levels of abstraction. Integrating global character fixed-sized feature vectors with word-level embedding to extract a sentence-wide feature set via Bi-LSTM. This deals with three main problems: (i) Sentences have any different size; (ii) The semantic and the syntactic of words in a sentence are captured in order to increase information for a word; (iii) Important information of characters that can appear at any position in a word are extracted. In sub-section B, we introduce various kinds of dataset. The modules of our model are constructed in other sub-sections. Data Preparation Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains 1,600K training tweets collected by a crawler from BIBREF0 . BIBREF0 constructed a test set manually with 177 negative and 182 positive tweets. The Stanford test set is small. However, it has been widely used in different evaluation tasks BIBREF0 BIBREF5 BIBREF13 . Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified tweets collected by using search terms: INLINEFORM0 , #google, #microsoft and #twitter. We construct the dataset as BIBREF14 for binary classification. Health Care Reform (HCR): This dataset was constructed by crawling tweets containing the hashtag #hcr BIBREF15 . Task is to predict positive/negative tweets BIBREF14 . Preprocessing We firstly take unique properties of Twitter in order to reduce the feature space such as Username, Usage of links, None, URLs and Repeated Letters. We then process retweets, stop words, links, URLs, mentions, punctuation and accentuation. For emoticons, BIBREF0 revealed that the training process makes the use of emoticons as noisy labels and they stripped the emoticons out from their training dataset because BIBREF0 believed that if we consider the emoticons, there is a negative impact on the accuracies of classifiers. In addition, removing emoticons makes the classifiers learns from other features (e.g. unigrams and bi-grams) presented in tweets and the classifiers only use these non-emoticon features to predict the sentiment of tweets. However, there is a problem is that if the test set contains emoticons, they do not influence the classifiers because emoticon features do not contain in its training data. This is a limitation of BIBREF0 , because the emoticon features would be useful when classifying test data. Therefore, we keep emoticon features in the datasets because deep learning models can capture more information from emoticon features for increasing classification accuracy. Semantic Rules (SR) In Twitter social networking, people express their opinions containing sub-sentences. These sub-sentences using specific PoS particles (Conjunction and Conjunctive adverbs), like "but, while, however, despite, however" have different polarities. However, the overall sentiment of tweets often focus on certain sub-sentences. For example: @lonedog bwahahah...you are amazing! However, it was quite the letdown. @kirstiealley my dentist is great but she's expensive...=( In two tweets above, the overall sentiment is negative. However, the main sentiment is only in the sub-sentences following but and however. This inspires a processing step to remove unessential parts in a tweet. Rule-based approach can assists these problems in handling negation and dealing with specific PoS particles led to effectively affect the final output of classification BIBREF11 BIBREF16 . BIBREF11 summarized a full presentation of their semantic rules approach and devised ten semantic rules in their hybrid approach based on the presentation of BIBREF16 . We use five rules in the semantic rules set because other five rules are only used to compute polarity of words after POS tagging or Parsing steps. We follow the same naming convention for rules utilized by BIBREF11 to represent the rules utilized in our proposed method. The rules utilized in the proposed method are displayed in Table TABREF15 in which is included examples from STS Corpus and output after using the rules. Table TABREF16 illustrates the number of processed sentences on each dataset. Representation Levels To construct embedding inputs for our model, we use a fixed-sized word vocabulary INLINEFORM0 and a fixed-sized character vocabulary INLINEFORM1 . Given a word INLINEFORM2 is composed from characters INLINEFORM3 , the character-level embeddings are encoded by column vectors INLINEFORM4 in the embedding matrix INLINEFORM5 , where INLINEFORM6 is the size of the character vocabulary. For word-level embedding INLINEFORM7 , we use a pre-trained word-level embedding with dimension 200 or 300. A pre-trained word-level embedding can capture the syntactic and semantic information of words BIBREF17 . We build every word INLINEFORM8 into an embedding INLINEFORM9 which is constructed by two sub-vectors: the word-level embedding INLINEFORM10 and the character fixed-size feature vector INLINEFORM11 of INLINEFORM12 where INLINEFORM13 is the length of the filter of wide convolutions. We have INLINEFORM14 character fixed-size feature vectors corresponding to word-level embedding in a sentence. Deep Learning Module DeepCNN in the deep learning module is illustrated in Figure. FIGREF22 . The DeepCNN has two wide convolution layers. The first layer extract local features around each character windows of the given word and using a max pooling over character windows to produce a global fixed-sized feature vector for the word. The second layer retrieves important context characters and transforms the representation at previous level into a representation at higher abstract level. We have INLINEFORM0 global character fixed-sized feature vectors for INLINEFORM1 words. In the next step of Figure. FIGREF4 , we construct the vector INLINEFORM0 by concatenating the word-level embedding with the global character fixed-size feature vectors. The input of Bi-LSTM is a sequence of embeddings INLINEFORM1 . The use of the global character fixed-size feature vectors increases the relationship of words in the word-level embedding. The purpose of this Bi-LSTM is to capture the context of words in a sentence and maintain the order of words toward to extract sentence-level representation. The top of the model is a softmax function to predict sentiment label. We describe in detail the kinds of CNN and LSTM that we use in next sub-part 1 and 2. The one-dimensional convolution called time-delay neural net has a filter vector INLINEFORM0 and take the dot product of filter INLINEFORM1 with each m-grams in the sequence of characters INLINEFORM2 of a word in order to obtain a sequence INLINEFORM3 : DISPLAYFORM0 Based on Equation 1, we have two types of convolutions that depend on the range of the index INLINEFORM0 . The narrow type of convolution requires that INLINEFORM1 and produce a sequence INLINEFORM2 . The wide type of convolution does not require on INLINEFORM3 or INLINEFORM4 and produce a sequence INLINEFORM5 . Out-of-range input values INLINEFORM6 where INLINEFORM7 or INLINEFORM8 are taken to be zero. We use wide convolution for our model. Given a word INLINEFORM0 composed of INLINEFORM1 characters INLINEFORM2 , we take a character embedding INLINEFORM3 for each character INLINEFORM4 and construct a character matrix INLINEFORM5 as following Equation. 2: DISPLAYFORM0 The values of the embeddings INLINEFORM0 are parameters that are optimized during training. The trained weights in the filter INLINEFORM1 correspond to a feature detector which learns to recognize a specific class of n-grams. The n-grams have size INLINEFORM2 . The use of a wide convolution has some advantages more than a narrow convolution because a wide convolution ensures that all weights of filter reach the whole characters of a word at the margins. The resulting matrix has dimension INLINEFORM3 . Long Short-Term Memory networks usually called LSTMs are a improved version of RNN. The core idea behind LSTMs is the cell state which can maintain its state over time, and non-linear gating units which regulate the information flow into and out of the cell. The LSTM architecture that we used in our proposed model is described in BIBREF9 . A single LSTM memory cell is implemented by the following composite function: DISPLAYFORM0 DISPLAYFORM1 where INLINEFORM0 is the logistic sigmoid function, INLINEFORM1 and INLINEFORM2 are the input gate, forget gate, output gate, cell and cell input activation vectors respectively. All of them have a same size as the hidden vector INLINEFORM3 . INLINEFORM4 is the hidden-input gate matrix, INLINEFORM5 is the input-output gate matrix. The bias terms which are added to INLINEFORM6 and INLINEFORM7 have been omitted for clarity. In addition, we also use the full gradient for calculating with full backpropagation through time (BPTT) described in BIBREF10 . A LSTM gradients using finite differences could be checked and making practical implementations more reliable. Regularization For regularization, we use a constraint on INLINEFORM0 of the weight vectors BIBREF18 . Experimental setups For the Stanford Twitter Sentiment Corpus, we use the number of samples as BIBREF5 . The training data is selected 80K tweets for a training data and 16K tweets for the development set randomly from the training data of BIBREF0 . We conduct a binary prediction for STS Corpus. For Sander dataset, we use standard 10-fold cross validation as BIBREF14 . We construct the development set by selecting 10% randomly from 9-fold training data. In Health Care Reform Corpus, we also select 10% randomly for the development set in a training set and construct as BIBREF14 for comparison. We describe the summary of datasets in Table III. for all datasets, the filter window size ( INLINEFORM0 ) is 7 with 6 feature maps each for the first wide convolution layer, the second wide convolution layer has a filter window size of 5 with 14 feature maps each. Dropout rate ( INLINEFORM1 ) is 0.5, INLINEFORM2 constraint, learning rate is 0.1 and momentum of 0.9. Mini-batch size for STS Corpus is 100 and others are 4. In addition, training is done through stochastic gradient descent over shuffled mini-batches with Adadelta update rule BIBREF19 . we use the publicly available Word2Vec trained from 100 billion words from Google and TwitterGlove of Stanford is performed on aggregated global word-word co-occurrence statistics from a corpus. Word2Vec has dimensionality of 300 and Twitter Glove have dimensionality of 200. Words that do not present in the set of pre-train words are initialized randomly. Experimental results Table IV shows the result of our model for sentiment classification against other models. We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN. As can be seen, 86.63 is the best prediction accuracy of our model so far for the STS Corpus. For Sanders and HCR datasets, we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets. Our models outperform the model of BIBREF14 for the Sanders dataset and HCR dataset. Analysis As can be seen, the models with SR outperforms the model with no SR. Semantic rules is effective in order to increase classification accuracy. We evaluate the efficiency of SR for the model in Table V of our full paper . We also conduct two experiments on two separate models: DeepCNN and Bi-LSTM in order to show the effectiveness of combination of DeepCNN and Bi-LSTM. In addition, the model using TwitterGlove outperform the model using GoogleW2V because TwitterGlove captures more information in Twitter than GoogleW2V. These results show that the character-level information and SR have a great impact on Twitter Data. The pre-train word vectors are good, universal feature extractors. The difference between our model and other approaches is the ability of our model to capture important features by using SR and combine these features at high benefit. The use of DeepCNN can learn a representation of words in higher abstract level. The combination of global character fixed-sized feature vectors and a word embedding helps the model to find important detectors for particles such as 'not' that negate sentiment and potentiate sentiment such as 'too', 'so' standing beside expected features. The model not only learns to recognize single n-grams, but also patterns in n-grams lead to form a structure significance of a sentence. Conclusions In the present work, we have pointed out that the use of character embeddings through a DeepCNN to enhance information for word embeddings built on top of Word2Vec or TwitterGlove improves classification accuracy in Tweet sentiment classification. Our results add to the well-establish evidence that character vectors are an important ingredient for word-level in deep learning for NLP. In addition, semantic rules contribute handling non-essential sub-tweets in order to improve classification accuracy.

rules that compute polarity of words after POS tagging or parsing steps

69a7a6675c59a4c5fb70006523b9fe0f01ca415c

69a7a6675c59a4c5fb70006523b9fe0f01ca415c_0

Q: Which knowledge graph completion tasks do they experiment with? Text: Introduction Knowledge graphs (KGs) such as Freebase BIBREF0 , DBpedia BIBREF1 , and YAGO BIBREF2 play a critical role in various NLP tasks, including question answering BIBREF3 , information retrieval BIBREF4 , and personalized recommendation BIBREF5 . A typical KG consists of numerous facts about a predefined set of entities. Each fact is in the form of a triplet INLINEFORM0 (or INLINEFORM1 for short), where INLINEFORM2 and INLINEFORM3 are two entities and INLINEFORM4 is a relation the fact describes. Due to the discrete and incomplete natures of KGs, various KG embedding models are proposed to facilitate KG completion tasks, e.g., link prediction and triplet classification. After vectorizing entities and relations in a low-dimensional space, those models predict missing facts by manipulating the involved entity and relation embeddings. Although proving successful in previous studies, traditional KG embedding models simply ignore the evolving nature of KGs. They require all entities to be present when training the embeddings. However, BIBREF6 shi2018open suggest that, on DBpedia, 200 new entities emerge on a daily basis between late 2015 and early 2016. Given the infeasibility of retraining embeddings from scratch whenever new entities come, missing facts about emerging entities are, unfortunately, not guaranteed to be inferred in time. By transforming realistic networks, e.g., citation graphs, social networks, and protein interaction graphs, to simple graphs with single-typed and undirected edges, recent explorations BIBREF7 shed light on the evolution issue for homogeneous graphs. While learning embeddings for existing nodes, they inductively learn a neighborhood aggregator that represents a node by aggregating its neighbors' embeddings. The embeddings of unseen nodes can then be obtained by applying the aggregator on their existing neighbors. It is well received that KGs differ from homogeneous graphs by their multi-relational structure BIBREF8 . Despite the difference, it seems promising to generalize the neighborhood aggregating scheme to embed emerging KG entities in an inductive manner. For example, in Figure FIGREF1 , a news article may describe an emerging entity (marked gray) as well as some facts involving existing entities. By generalizing structural information in the underlying KG, e.g., other entities residing in a similar neighborhood or involving similar relations, to the current entity's neighborhood, we can infer that it may probably live in Chicago. Inspired by the above example, the inductive KG embedding problem boils down to designing a KG-specific neighborhood aggregator to capture essential neighborhood information. Intuitively, an ideal aggregator should have the following desired properties: This paper concentrates on KG-specific neighborhood aggregators, which is of practical importance but only received limited focus BIBREF9 . To the best of our knowledge, neither conventional aggregators for homogeneous graphs nor those for KGs satisfy all the above three properties. In this regard, we employ the attention mechanism BIBREF10 and propose an aggregator called Logic Attention Network (LAN). Aggregating neighbors by a weighted combination of their transformed embeddings, LAN is inherently permutation invariant. To estimate the attention weights in LAN, we adopt two mechanisms to model relation- and neighbor-level information in a coarse-to-fine manner, At both levels, LAN is made aware of both neighborhood redundancy and query relation. To summarize, our contributions are: (1) We propose three desired properties that decent neighborhood aggregators for KGs should possess. (2) We propose a novel aggregator, i.e., Logic Attention Network, to facilitate inductive KG embedding. (3) We conduct extensive comparisons with conventional aggregators on two KG completions tasks. The results validate the superiority of LAN w.r.t. the three properties. Transductive Embedding Models In recent years, representation learning problems on KGs have received much attention due to the wide applications of the resultant entity and relation embeddings. Typical KG embedding models include TransE BIBREF11 , Distmult BIBREF12 , Complex BIBREF13 , Analogy BIBREF14 , to name a few. For more explorations, we refer readers to an extensive survey BIBREF15 . However, conventional approaches on KG embedding work in a transductive manner. They require that all entities should be seen during training. Such limitation hinders them from efficiently generalizing to emerging entities. Inductive Embedding Models To relieve the issue of emerging entities, several inductive KG embedding models are proposed, including BIBREF16 xie2016representation, BIBREF6 shi2018open and BIBREF17 xie2016image which use description text or images as inputs. Although the resultant embeddings may be utilized for KG completion, it is not clear whether the embeddings are powerful enough to infer implicit or new facts beyond those expressed in the text/image. Moreover, when domain experts are recruited to introduce new entities via partial facts rather than text or images, those approaches may not help much. In light of the above scenario, existing neighbors of an emerging entity are considered as another type of input for inductive models. In BIBREF9 ijcai2017-250, the authors propose applying Graph Neural Network (GNN) on the KG, which generates the embedding of a new entity by aggregating all its known neighbors. However, their model aggregates the neighbors via simple pooling functions, which neglects the difference among the neighbors. Other works like BIBREF18 fu2017hin2vec and BIBREF19 tang2015pte aim at embedding nodes for node classification given the entire graph and thus are inapplicable for inductive KG-specific tasks. BIBREF20 schlichtkrull2017modeling and BIBREF21 xiong2018one also rely on neighborhood structures to embed entities, but they either work transductively or focus on emerging relations. Finally, we note another related line of studies on node representation learning for homogeneous graphs. Similar to text- or image-based inductive models for KGs, BIBREF22 duran2017learning, BIBREF23 yang2016revisiting, BIBREF24 velivckovic2017graph and BIBREF25 rossi2018deep exploit additional node attributes to embed unseen nodes. Another work more related to ours is BIBREF26 hamilton2017inductive. They tackle inductive node embedding by the neighborhood aggregation scheme. Their aggregators either trivially treat neighbors equally or unnecessarily require them to be ordered. Moreover, like all embedding models for homogeneous graphs, their model cannot be directly applied to KGs with multi-relational edges. Notations Let INLINEFORM0 and INLINEFORM1 be two sets of entities and relations of size INLINEFORM2 and INLINEFORM3 , respectively. A knowledge graph is composed of a set of triplet facts, namely DISPLAYFORM0 For each INLINEFORM0 , we denote the reverse of INLINEFORM1 by INLINEFORM2 , and add an additional triplet INLINEFORM3 to INLINEFORM4 . For an entity INLINEFORM0 , we denote by INLINEFORM1 its neighborhood in INLINEFORM2 , i.e., all related entities with the involved relations. Formally, DISPLAYFORM0 We denote the projection of INLINEFORM0 on INLINEFORM1 and INLINEFORM2 by INLINEFORM3 and INLINEFORM4 , respectively. Here INLINEFORM5 are neighbors and INLINEFORM6 are neighboring relations. When the context is clear, we simplify the INLINEFORM7 -th entity INLINEFORM8 by its subscript INLINEFORM9 . We denote vectors by bold lower letters, and matrices or sets of vectors by bold upper letters. Given a knowledge graph INLINEFORM0 , we would like to learn a neighborhood aggregator INLINEFORM1 that acts as follows: For an entity INLINEFORM0 on INLINEFORM1 , INLINEFORM2 depends on INLINEFORM3 's neighborhood INLINEFORM4 to embed INLINEFORM5 as a low-dimensional vector INLINEFORM6 ; For an unknown triplet INLINEFORM0 , the embeddings of INLINEFORM1 and INLINEFORM2 output by INLINEFORM3 suggest the plausibility of the triplet. When a new entity emerges with some triplets involving INLINEFORM0 and INLINEFORM1 , we could apply such an aggregator INLINEFORM2 on its newly established neighborhood, and use the output embedding to infer new facts about it. Framework To obtain such a neighborhood aggregator INLINEFORM0 , we adopt an encoder-decoder framework as illustrated by Figure FIGREF12 . Given a training triplet, the encoder INLINEFORM1 encodes INLINEFORM2 and INLINEFORM3 into two embeddings with INLINEFORM4 . The decoder measures the plausibility of the triplet, and provides feedbacks to the encoder to adjust the parameters of INLINEFORM5 . In the remainder of this section, we describe general configurations of the two components. As specified in Figure FIGREF12 , for an entity INLINEFORM0 on focus, the encoder works on a collection of input neighbor embeddings, and output INLINEFORM1 's embedding. To differentiate between input and output embeddings, we use superscripts INLINEFORM2 and INLINEFORM3 on the respective vectors. Let INLINEFORM4 , which is obtained from an embedding matrix INLINEFORM5 , be the embedding of a neighbor INLINEFORM6 , where INLINEFORM7 . To reflect the impact of relation INLINEFORM8 on INLINEFORM9 , we apply a relation-specific transforming function INLINEFORM10 on INLINEFORM11 as follows, DISPLAYFORM0 where INLINEFORM0 is the transforming vector for relation INLINEFORM1 and is restricted as a unit vector. We adopt this transformation from BIBREF27 wang2014knowledge since it does not involve matrix product operations and is of low computation complexity. After neighbor embeddings are transformed, these transformed embeddings are fed to the aggregator INLINEFORM0 to output an embedding INLINEFORM1 for the target entity INLINEFORM2 , i.e., DISPLAYFORM0 By definition, an aggregator INLINEFORM0 essentially takes as input a collection of vectors INLINEFORM1 ( INLINEFORM2 ) and maps them to a single vector. With this observation, the following two types of functions seem to be natural choices for neighborhood aggregators, and have been adopted previously: Pooling Functions. A typical pooling function is mean-pooling, which is defined by INLINEFORM0 . Besides mean-pooling, other previously adopted choices include sum- and max-pooling BIBREF9 . Due to their simple forms, pooling functions are permutation-invariant, but consider the neighbors equally. It is aware of neither potential redundancy in the neighborhood nor the query relations. Recurrent Neural Networks (RNNs). In various natural language processing tasks, RNNs prove effective in modeling sequential dependencies. In BIBREF26 , the authors adopt an RNN variant LSTM BIBREF28 as neighborhood aggregator, i.e., INLINEFORM0 . To train and apply the LSTM-based aggregator, they have to randomly permute the neighbors, which violates the permutation variance property. Given the subject and object embeddings INLINEFORM0 and INLINEFORM1 output by the encoder, the decoder is required to measure the plausibility of the training triplet. To avoid potential mixture with relations INLINEFORM2 in the neighborhood, we refer to the relation in the training triplet by query relation, and denote it by INLINEFORM3 instead. After looking up INLINEFORM4 's representation INLINEFORM5 from an embedding matrix INLINEFORM6 , the decoder scores the training triplet INLINEFORM7 with a scoring function INLINEFORM8 . Following BIBREF9 ijcai2017-250, we mainly investigate a scoring function based on TransE BIBREF11 defined by DISPLAYFORM0 where INLINEFORM0 denotes the L1 norm. To test whether the studied aggregators generalize among different scoring function, we will also consider several alternatives in experiments. Logic Attention Network As discussed above, traditional neighborhood aggregators do not preserve all desired properties. In this section, we describe a novel aggregator, namely Logic Attention Network (LAN), which addresses all three properties. We also provide details in training the LAN aggregator. Incorporating Neighborhood Attention Traditional neighborhood aggregators only depend on collections of transformed embeddings. They neglect other useful information in the neighborhood INLINEFORM0 and the query relation INLINEFORM1 , which may facilitate more effective aggregation of the transformed embeddings. To this end, we propose generalizing the aggregators from INLINEFORM2 to INLINEFORM3 . Specifically, for an entity INLINEFORM0 , its neighbors INLINEFORM1 should contribute differently to INLINEFORM2 according to its importance in representing INLINEFORM3 . To consider the different contribution while preserving the permutation invariance property, we employ a weighted or attention-based aggregating approach on the transformed embeddings. The additional information in INLINEFORM4 and INLINEFORM5 is then exploited when estimating the attention weights. Formally, we obtain INLINEFORM6 by DISPLAYFORM0 Here INLINEFORM0 is the attention weight specified for each neighbor INLINEFORM1 given INLINEFORM2 and the query relation INLINEFORM3 . To assign larger weights INLINEFORM0 to more important neighbors, from the perspective of INLINEFORM1 , we ask ourselves two questions at progressive levels: 1) What types of neighboring relations may lead us to potentially important neighbors? 2) Following those relations, which specific neighbor (in transformed embedding) may contain important information? Inspired by the two questions, we adopt the following two mechanisms to estimate INLINEFORM2 . Relations in a KG are simply not independent of each other. For an entity INLINEFORM0 , one neighboring relation INLINEFORM1 may imply the existence of another neighboring relation INLINEFORM2 , though they may not necessarily connect INLINEFORM3 to the same neighbor. For example, a neighboring relation play_for may suggest the home city, i.e., live_in, of the current athlete entity. Following notations in logics, we denote potential dependency between INLINEFORM4 and INLINEFORM5 by a “logic rule” INLINEFORM6 . To measure the extent of such dependency, we define the confidence of a logic rule INLINEFORM7 as follows: DISPLAYFORM0 Here the function INLINEFORM0 equals 1 when INLINEFORM1 is true and 0 otherwise. As an empirical statistic over the entire KG, INLINEFORM2 is larger if more entities with neighboring relation INLINEFORM3 also have INLINEFORM4 as a neighboring relation. With the confidence scores INLINEFORM0 between all relation pairs at hand, we are ready to characterize neighboring relations INLINEFORM1 that lead to important neighbors. On one hand, such a relation INLINEFORM2 should have a large INLINEFORM3 , i.e., it is statistically relevant to INLINEFORM4 . Following the above example, play_for should be consulted to if the query relation is live_in. On the other hand, INLINEFORM5 should not be implied by other relations in the neighborhood. For example, no matter whether the query relation is live_in or not, the neighboring relation work_as should not be assigned too much weight, because sufficient information is already provided by play_for. Following the above intuitions, we implement the logic rule mechanism of measuring neighboring relations' usefulness as follow: DISPLAYFORM0 We note that INLINEFORM0 promotes relations INLINEFORM1 strongly implying INLINEFORM2 (the numerator) and demotes those implied by some other relation in the same neighborhood (the denominator). In this manner, our logic rule mechanism addresses both query relation awareness and neighborhood redundancy awareness. With global statistics about relations, the logic rule mechanism guides the attention weight to be distributed at a coarse granularity of relations. However, it may be insufficient not to consult finer-grained information hidden in the transformed neighbor embeddings to determine which neighbor is important indeed. To take the transformed embeddings into consideration, we adopt an attention network BIBREF10 . Specifically, given a query relation INLINEFORM0 , the importance of an entity INLINEFORM1 's neighbor INLINEFORM2 is measured by DISPLAYFORM0 Here the unnormalized attention weight INLINEFORM0 is given by an attention neural network as DISPLAYFORM0 In this equation, INLINEFORM0 and INLINEFORM1 are global attention parameters, while INLINEFORM2 is a relation-specific attention parameter for the query relation INLINEFORM3 . All those attention parameters are regarded as parameters of the encoder, and learned directly from the data. Note that, unlike the logic rule mechanism at relation level, the computation of INLINEFORM0 concentrates more on the neighbor INLINEFORM1 itself. This is useful when the neighbor entity INLINEFORM2 is also helpful to explain the current training triplet. For example, in Figure FIGREF12 , the neighbor Chicago_Bulls could help to imply the object of live_in since there are other athletes playing for Chicago_Bulls while living in Chicago. Although working at the neighbor level, the dependency on transformed neighbor embeddings INLINEFORM3 and the relation-specific parameter INLINEFORM4 make INLINEFORM5 aware of both neighborhood redundancy and the query relation. Finally, to incorporate these two weighting mechanisms together in measuring the importance of neighbors, we employ a double-view attention and reformulate Eq. ( EQREF22 ) as DISPLAYFORM0 Training Objective To train the entire model in Figure FIGREF12 , we need both positive triplets and negative ones. All triplets INLINEFORM0 from the knowledge graph naturally serve as positive triplets, which we denote by INLINEFORM1 . To make up for the absence of negative triplets, for each INLINEFORM2 , we randomly corrupt the object or subject (but not both) by another entity in INLINEFORM3 , and denote the corresponding negative triplets by INLINEFORM4 . Formally, DISPLAYFORM0 To encourage the decoder to give high scores for positive triplets and low scores for negative ones, we apply a margin-based ranking loss on each triplet INLINEFORM0 , i.e., DISPLAYFORM0 Here INLINEFORM0 denotes the positive part of x, and INLINEFORM1 is a hyper-parameter for the margin. Finally, the training objective is defined by DISPLAYFORM0 The above training objective only optimizes the output of the aggregator, i.e., the output entity embeddings INLINEFORM0 . The input entity embeddings INLINEFORM1 , however, are not directly aware of the structure of the entire KG. To make the input embeddings and thus the aggregation more meaningful, we set up a subtask for LAN. First, we define a second scoring function, which is similar to Eq. ( EQREF20 ) except that input embeddings INLINEFORM0 from INLINEFORM1 are used to represent the subject and object, i.e., DISPLAYFORM0 The embedding of query relation INLINEFORM0 is obtained from the same embedding matrix INLINEFORM1 as in the first scoring function. Then a similar margin-based ranking loss INLINEFORM2 as Eq. ( EQREF32 ) is defined for the subtask. Finally, we combine the subtask with the main task, and reformulate the overall training objective of LAN as DISPLAYFORM0 Experimental Configurations We evaluate the effectiveness of our LAN model on two typical knowledge graph completion tasks, i.e., link prediction and triplet classification. We compare our LAN with two baseline aggregators, MEAN and LSTM, as described in the Encoder section. MEAN is used on behalf of pooling functions since it leads to the best performance in BIBREF9 ijcai2017-250. LSTM is used due to its large expressive capability BIBREF26 . Data Construction In both tasks, we need datasets whose test sets contain new entities unseen during training. For the task of triplet classification, we directly use the datasets released by BIBREF9 ijcai2017-250 which are based on WordNet11 BIBREF29 . Since they do not conduct experiments on the link prediction task, we construct the required datasets based on FB15K BIBREF11 following a similar protocol used in BIBREF9 ijcai2017-250 as follows. Sampling unseen entities. Firstly, we randomly sample INLINEFORM0 of the original testing triplets to form a new test set INLINEFORM1 for our inductive scenario ( BIBREF9 ijcai2017-250 samples INLINEFORM2 testing triplets). Then two different strategies are used to construct the candidate unseen entities INLINEFORM6 . One is called Subject, where only entities appearing as the subjects in INLINEFORM7 are added to INLINEFORM8 . Another is called Object, where only objects in INLINEFORM9 are added to INLINEFORM10 . For an entity INLINEFORM11 , if it does not have any neighbor in the original training set, such an entity is filtered out, yielding the final unseen entity set INLINEFORM12 . For a triplet INLINEFORM13 , if INLINEFORM14 or INLINEFORM15 , it is removed from INLINEFORM16 . Filtering and splitting data sets. The second step is to ensure that unseen entities would not appear in final training set or validation set. We split the original training set into two data sets, the new training set and auxiliary set. For a triplet INLINEFORM0 in original training set, if INLINEFORM1 , it is added to the new training set. If INLINEFORM2 or INLINEFORM3 , it is added to the auxiliary set, which serves as existing neighbors for unseen entities in INLINEFORM4 . Finally, for a triplet INLINEFORM0 in the original validation set, if INLINEFORM1 or INLINEFORM2 , it is removed from the validation set. The statistics for the resulting INLINEFORM0 datasets using Subject and Object strategies are in Table TABREF34 . Experiments on Triplet Classification Triplet classification aims at classifying a fact triplet INLINEFORM0 as true or false. In the dataset of BIBREF9 ijcai2017-250, triplets in the validation and testing sets are labeled as true or false, while triplets in the training set are all true ones. To tackle this task, we preset a threshold INLINEFORM0 for each relation r. If INLINEFORM1 , the triplet is classified as positive, otherwise it is negative. We determine the optimal INLINEFORM2 by maximizing classification accuracy on the validation set. Experimental Setup Since this task is also conducted in BIBREF9 ijcai2017-250, we use the same configurations with learning rate INLINEFORM0 , embedding dimension INLINEFORM1 , and margin INLINEFORM2 for all datasets. We randomly sample 64 neighbors for each entity. Zero padding is used when the number of neighbors is less than 64. L2-regularization is applied on the parameters of LAN. The regularization rate is INLINEFORM3 . We search the best hyper-parameters of all models according to the performance on validation set. In detail, we search learning rate INLINEFORM0 in INLINEFORM1 , embedding dimension for neighbors INLINEFORM2 in INLINEFORM3 , and margin INLINEFORM4 in INLINEFORM5 . The optimal configurations are INLINEFORM6 for all the datasets. Evaluation Results The results are reported in Table TABREF42 . Since we did not achieve the same results for MEAN as reported in BIBREF9 ijcai2017-250 with either our implementation or their released source code, the best results from their original paper are reported. From the table, we observe that, on one hand, LSTM results in poorer performance compared with MEAN, which involves fewer parameters though. This demonstrates the necessity of the permutation invariance for designing neighborhood aggregators for KGs. On the other hand, our LAN model consistently achieves the best results on all datasets, demonstrating the effectiveness of LAN on this KBC task. Experiments on Link Prediction Link prediction in the inductive setting aims at reasoning the missing part “?” in a triplet when given INLINEFORM0 or INLINEFORM1 with emerging entities INLINEFORM2 or INLINEFORM3 respectively. To tackle the task, we firstly hide the object (subject) of each testing triplet in Subject-R (Object-R) to produce a missing part. Then we replace the missing part with all entities in the entity set INLINEFORM4 to construct candidate triplets. We compute the scoring function INLINEFORM5 defined in Eq. ( EQREF20 ) for all candidate triplets, and rank them in descending order. Finally, we evaluate whether the ground-truth entities are ranked ahead of other entities. We use traditional evaluation metrics as in the KG completion literature, i.e., Mean Rank (MR), Mean Reciprocal Rank (MRR), and the proportion of ground truth entities ranked top-k (Hits@k, INLINEFORM6 ). Since certain candidate triplets might also be true, we follow previous works and filter out these fake negatives before ranking. Experimental Results The results on Subject-10 and Object-10 are reported in Table TABREF43 . The results on other datasets are similar and we summarize them later in Figure FIGREF50 . From Table TABREF43 , we still observe consistent results for all the models as in the triplet classification task. Firstly, LSTM results in the poorest performance on all datasets. Secondly, our LAN model outperforms all the other baselines significantly, especially on the Hit@k metrics. The improvement on the MR metric of LAN might not be considerable. This is due to the flaw of the MR metric since it is more sensitive to lower positions of the ranking, which is actually of less importance. The MRR metric is proposed for this reason, where we could observe consistent improvements brought by LAN. The effectiveness of LAN on link prediction validates LAN's superiority to other aggregators and the necessities to treat the neighbors differently in a permutation invariant way. To analyze whether LAN outperforms the others for expected reasons and generalizes to other configurations, we conduct the following studies. In this experiment, we would like to confirm that it's necessary for the aggregator to be aware of the query relation. Specifically, we investigate the attention neural network and design two degenerated baselines. One is referred to as Query-Attention and is simply an attention network as in LAN except that the logic rule mechanism is removed. The other is referred to as Global-Attention, which is also an attention network except that the query relation embedding INLINEFORM0 in Eq. ( EQREF28 ) is masked by a zero vector. The results are reported in Table TABREF46 . We observe that although superior to MEAN, Global-Attention is outperformed by Query-Attention, demonstrating the necessity of query relation awareness. The superiority of Global-Attention over MEAN could be attributed to the fact that the attention mechanism is effective to identify the neighbors which are globally important regardless of the query. We find that the logic rules greatly help to improve the attention network in LAN. We confirm this point by conducting further experiments where the logic rule mechanism is isolated as a single model (referred to as Logic Rules Only). The results are also demonstrated in Table TABREF46 , from which we find that Query-Attention outperforms MEAN by a limited margin. Meanwhile, Logic Rules Only outperforms both MEAN and Query-Attention by significant margins. These results demonstrate the effectiveness of logic rules in assigning meaningful weights to the neighbors. Specifically, in order to generate representations for unseen entities, it is crucial to incorporate the logic rules to train the aggregator, instead of depending solely on neural networks to learn from the data. By combining the logic rules and neural networks, LAN takes a step further in outperforming all the other models. To find out whether the superiority of LAN to the baselines can generalize to other scoring functions, we replace the scoring function in Eq. ( EQREF20 ) and Eq. ( EQREF36 ) by three typical scoring functions mentioned in Related Works. We omit the results of LSTM, for it is still inferior to MEAN. The results are listed in Table TABREF48 , from which we observe that with different scoring functions, LAN outperforms MEAN consistently by a large margin on all the evaluation metrics. Note that TransE leads to the best results on MEAN and LAN. It's reasonable to suppose that when the ratio of the unseen entities over the training entities increases (namely the observed knowledge graph becomes sparser), all models' performance would deteriorate. To figure out whether our LAN could suffer less on sparse knowledge graphs, we conduct link prediction on datasets with different sample rates INLINEFORM0 as described in Step 1 of the Data Construction section. The results are displayed in Figure FIGREF50 . We observe that the increasing proportion of unseen entities certainly has a negative impact on all models. However, the performance of LAN does not decrease as drastically as that of MEAN and LSTM, indicating that LAN is more robust on sparse KGs. Case Studies on Neighbors' Weights In order to visualize how LAN specifies weights to neighbors, we sample some cases from the Subject-10 testing set. From Table FIGREF50 , we have the following observations. First, with the query relation, LAN could attribute higher weights to neighbors with more relevant relations. In the first case, when the query is origin, the top two neighbors are involved by place_lived and breed_origin, which are helpful to imply origin. In addition, in all three cases, neighbors with relation gender gain the lowest weights since they imply nothing about the query relation. Second, LAN could attribute higher weights to neighbor entities that are more informative. When the query relation is profession, the neighbors Aristotle, Metaphysics and Aesthetics are all relevant to the answer Philosopher. In the third case, we also observe similar situations. Here, the neighbor with the highest weight is (institution, University_of_Calgary) since the query relation place_lived helps the aggregator to focus on the neighboring relation institution, then the neighbor entity University_of_Calgary assists in locating the answer Calgary. Conclusion In this paper, we address inductive KG embedding, which helps embed emerging entities efficiently. We formulate three characteristics required for effective neighborhood aggregators. To meet the three characteristics, we propose LAN, which attributes different weights to an entity's neighbors in a permutation invariant manner, considering both the redundancy of neighbors and the query relation. The weights are estimated from data with logic rules at a coarse relation level, and neural attention network at a fine neighbor level. Experiments show that LAN outperforms baseline models significantly on two typical KG completion tasks. Acknowledgements We thank the three anonymous authors for their constructive comments. This work is supported by the National Natural Science Foundation of China (61472453, U1401256, U1501252, U1611264, U1711261, U1711262).

link prediction , triplet classification

60cb756d382b3594d9e1f4a5e2366db407e378ae

60cb756d382b3594d9e1f4a5e2366db407e378ae_0

Q: Apart from using desired properties, do they evaluate their LAN approach in some other way? Text: Introduction Knowledge graphs (KGs) such as Freebase BIBREF0 , DBpedia BIBREF1 , and YAGO BIBREF2 play a critical role in various NLP tasks, including question answering BIBREF3 , information retrieval BIBREF4 , and personalized recommendation BIBREF5 . A typical KG consists of numerous facts about a predefined set of entities. Each fact is in the form of a triplet INLINEFORM0 (or INLINEFORM1 for short), where INLINEFORM2 and INLINEFORM3 are two entities and INLINEFORM4 is a relation the fact describes. Due to the discrete and incomplete natures of KGs, various KG embedding models are proposed to facilitate KG completion tasks, e.g., link prediction and triplet classification. After vectorizing entities and relations in a low-dimensional space, those models predict missing facts by manipulating the involved entity and relation embeddings. Although proving successful in previous studies, traditional KG embedding models simply ignore the evolving nature of KGs. They require all entities to be present when training the embeddings. However, BIBREF6 shi2018open suggest that, on DBpedia, 200 new entities emerge on a daily basis between late 2015 and early 2016. Given the infeasibility of retraining embeddings from scratch whenever new entities come, missing facts about emerging entities are, unfortunately, not guaranteed to be inferred in time. By transforming realistic networks, e.g., citation graphs, social networks, and protein interaction graphs, to simple graphs with single-typed and undirected edges, recent explorations BIBREF7 shed light on the evolution issue for homogeneous graphs. While learning embeddings for existing nodes, they inductively learn a neighborhood aggregator that represents a node by aggregating its neighbors' embeddings. The embeddings of unseen nodes can then be obtained by applying the aggregator on their existing neighbors. It is well received that KGs differ from homogeneous graphs by their multi-relational structure BIBREF8 . Despite the difference, it seems promising to generalize the neighborhood aggregating scheme to embed emerging KG entities in an inductive manner. For example, in Figure FIGREF1 , a news article may describe an emerging entity (marked gray) as well as some facts involving existing entities. By generalizing structural information in the underlying KG, e.g., other entities residing in a similar neighborhood or involving similar relations, to the current entity's neighborhood, we can infer that it may probably live in Chicago. Inspired by the above example, the inductive KG embedding problem boils down to designing a KG-specific neighborhood aggregator to capture essential neighborhood information. Intuitively, an ideal aggregator should have the following desired properties: This paper concentrates on KG-specific neighborhood aggregators, which is of practical importance but only received limited focus BIBREF9 . To the best of our knowledge, neither conventional aggregators for homogeneous graphs nor those for KGs satisfy all the above three properties. In this regard, we employ the attention mechanism BIBREF10 and propose an aggregator called Logic Attention Network (LAN). Aggregating neighbors by a weighted combination of their transformed embeddings, LAN is inherently permutation invariant. To estimate the attention weights in LAN, we adopt two mechanisms to model relation- and neighbor-level information in a coarse-to-fine manner, At both levels, LAN is made aware of both neighborhood redundancy and query relation. To summarize, our contributions are: (1) We propose three desired properties that decent neighborhood aggregators for KGs should possess. (2) We propose a novel aggregator, i.e., Logic Attention Network, to facilitate inductive KG embedding. (3) We conduct extensive comparisons with conventional aggregators on two KG completions tasks. The results validate the superiority of LAN w.r.t. the three properties. Transductive Embedding Models In recent years, representation learning problems on KGs have received much attention due to the wide applications of the resultant entity and relation embeddings. Typical KG embedding models include TransE BIBREF11 , Distmult BIBREF12 , Complex BIBREF13 , Analogy BIBREF14 , to name a few. For more explorations, we refer readers to an extensive survey BIBREF15 . However, conventional approaches on KG embedding work in a transductive manner. They require that all entities should be seen during training. Such limitation hinders them from efficiently generalizing to emerging entities. Inductive Embedding Models To relieve the issue of emerging entities, several inductive KG embedding models are proposed, including BIBREF16 xie2016representation, BIBREF6 shi2018open and BIBREF17 xie2016image which use description text or images as inputs. Although the resultant embeddings may be utilized for KG completion, it is not clear whether the embeddings are powerful enough to infer implicit or new facts beyond those expressed in the text/image. Moreover, when domain experts are recruited to introduce new entities via partial facts rather than text or images, those approaches may not help much. In light of the above scenario, existing neighbors of an emerging entity are considered as another type of input for inductive models. In BIBREF9 ijcai2017-250, the authors propose applying Graph Neural Network (GNN) on the KG, which generates the embedding of a new entity by aggregating all its known neighbors. However, their model aggregates the neighbors via simple pooling functions, which neglects the difference among the neighbors. Other works like BIBREF18 fu2017hin2vec and BIBREF19 tang2015pte aim at embedding nodes for node classification given the entire graph and thus are inapplicable for inductive KG-specific tasks. BIBREF20 schlichtkrull2017modeling and BIBREF21 xiong2018one also rely on neighborhood structures to embed entities, but they either work transductively or focus on emerging relations. Finally, we note another related line of studies on node representation learning for homogeneous graphs. Similar to text- or image-based inductive models for KGs, BIBREF22 duran2017learning, BIBREF23 yang2016revisiting, BIBREF24 velivckovic2017graph and BIBREF25 rossi2018deep exploit additional node attributes to embed unseen nodes. Another work more related to ours is BIBREF26 hamilton2017inductive. They tackle inductive node embedding by the neighborhood aggregation scheme. Their aggregators either trivially treat neighbors equally or unnecessarily require them to be ordered. Moreover, like all embedding models for homogeneous graphs, their model cannot be directly applied to KGs with multi-relational edges. Notations Let INLINEFORM0 and INLINEFORM1 be two sets of entities and relations of size INLINEFORM2 and INLINEFORM3 , respectively. A knowledge graph is composed of a set of triplet facts, namely DISPLAYFORM0 For each INLINEFORM0 , we denote the reverse of INLINEFORM1 by INLINEFORM2 , and add an additional triplet INLINEFORM3 to INLINEFORM4 . For an entity INLINEFORM0 , we denote by INLINEFORM1 its neighborhood in INLINEFORM2 , i.e., all related entities with the involved relations. Formally, DISPLAYFORM0 We denote the projection of INLINEFORM0 on INLINEFORM1 and INLINEFORM2 by INLINEFORM3 and INLINEFORM4 , respectively. Here INLINEFORM5 are neighbors and INLINEFORM6 are neighboring relations. When the context is clear, we simplify the INLINEFORM7 -th entity INLINEFORM8 by its subscript INLINEFORM9 . We denote vectors by bold lower letters, and matrices or sets of vectors by bold upper letters. Given a knowledge graph INLINEFORM0 , we would like to learn a neighborhood aggregator INLINEFORM1 that acts as follows: For an entity INLINEFORM0 on INLINEFORM1 , INLINEFORM2 depends on INLINEFORM3 's neighborhood INLINEFORM4 to embed INLINEFORM5 as a low-dimensional vector INLINEFORM6 ; For an unknown triplet INLINEFORM0 , the embeddings of INLINEFORM1 and INLINEFORM2 output by INLINEFORM3 suggest the plausibility of the triplet. When a new entity emerges with some triplets involving INLINEFORM0 and INLINEFORM1 , we could apply such an aggregator INLINEFORM2 on its newly established neighborhood, and use the output embedding to infer new facts about it. Framework To obtain such a neighborhood aggregator INLINEFORM0 , we adopt an encoder-decoder framework as illustrated by Figure FIGREF12 . Given a training triplet, the encoder INLINEFORM1 encodes INLINEFORM2 and INLINEFORM3 into two embeddings with INLINEFORM4 . The decoder measures the plausibility of the triplet, and provides feedbacks to the encoder to adjust the parameters of INLINEFORM5 . In the remainder of this section, we describe general configurations of the two components. As specified in Figure FIGREF12 , for an entity INLINEFORM0 on focus, the encoder works on a collection of input neighbor embeddings, and output INLINEFORM1 's embedding. To differentiate between input and output embeddings, we use superscripts INLINEFORM2 and INLINEFORM3 on the respective vectors. Let INLINEFORM4 , which is obtained from an embedding matrix INLINEFORM5 , be the embedding of a neighbor INLINEFORM6 , where INLINEFORM7 . To reflect the impact of relation INLINEFORM8 on INLINEFORM9 , we apply a relation-specific transforming function INLINEFORM10 on INLINEFORM11 as follows, DISPLAYFORM0 where INLINEFORM0 is the transforming vector for relation INLINEFORM1 and is restricted as a unit vector. We adopt this transformation from BIBREF27 wang2014knowledge since it does not involve matrix product operations and is of low computation complexity. After neighbor embeddings are transformed, these transformed embeddings are fed to the aggregator INLINEFORM0 to output an embedding INLINEFORM1 for the target entity INLINEFORM2 , i.e., DISPLAYFORM0 By definition, an aggregator INLINEFORM0 essentially takes as input a collection of vectors INLINEFORM1 ( INLINEFORM2 ) and maps them to a single vector. With this observation, the following two types of functions seem to be natural choices for neighborhood aggregators, and have been adopted previously: Pooling Functions. A typical pooling function is mean-pooling, which is defined by INLINEFORM0 . Besides mean-pooling, other previously adopted choices include sum- and max-pooling BIBREF9 . Due to their simple forms, pooling functions are permutation-invariant, but consider the neighbors equally. It is aware of neither potential redundancy in the neighborhood nor the query relations. Recurrent Neural Networks (RNNs). In various natural language processing tasks, RNNs prove effective in modeling sequential dependencies. In BIBREF26 , the authors adopt an RNN variant LSTM BIBREF28 as neighborhood aggregator, i.e., INLINEFORM0 . To train and apply the LSTM-based aggregator, they have to randomly permute the neighbors, which violates the permutation variance property. Given the subject and object embeddings INLINEFORM0 and INLINEFORM1 output by the encoder, the decoder is required to measure the plausibility of the training triplet. To avoid potential mixture with relations INLINEFORM2 in the neighborhood, we refer to the relation in the training triplet by query relation, and denote it by INLINEFORM3 instead. After looking up INLINEFORM4 's representation INLINEFORM5 from an embedding matrix INLINEFORM6 , the decoder scores the training triplet INLINEFORM7 with a scoring function INLINEFORM8 . Following BIBREF9 ijcai2017-250, we mainly investigate a scoring function based on TransE BIBREF11 defined by DISPLAYFORM0 where INLINEFORM0 denotes the L1 norm. To test whether the studied aggregators generalize among different scoring function, we will also consider several alternatives in experiments. Logic Attention Network As discussed above, traditional neighborhood aggregators do not preserve all desired properties. In this section, we describe a novel aggregator, namely Logic Attention Network (LAN), which addresses all three properties. We also provide details in training the LAN aggregator. Incorporating Neighborhood Attention Traditional neighborhood aggregators only depend on collections of transformed embeddings. They neglect other useful information in the neighborhood INLINEFORM0 and the query relation INLINEFORM1 , which may facilitate more effective aggregation of the transformed embeddings. To this end, we propose generalizing the aggregators from INLINEFORM2 to INLINEFORM3 . Specifically, for an entity INLINEFORM0 , its neighbors INLINEFORM1 should contribute differently to INLINEFORM2 according to its importance in representing INLINEFORM3 . To consider the different contribution while preserving the permutation invariance property, we employ a weighted or attention-based aggregating approach on the transformed embeddings. The additional information in INLINEFORM4 and INLINEFORM5 is then exploited when estimating the attention weights. Formally, we obtain INLINEFORM6 by DISPLAYFORM0 Here INLINEFORM0 is the attention weight specified for each neighbor INLINEFORM1 given INLINEFORM2 and the query relation INLINEFORM3 . To assign larger weights INLINEFORM0 to more important neighbors, from the perspective of INLINEFORM1 , we ask ourselves two questions at progressive levels: 1) What types of neighboring relations may lead us to potentially important neighbors? 2) Following those relations, which specific neighbor (in transformed embedding) may contain important information? Inspired by the two questions, we adopt the following two mechanisms to estimate INLINEFORM2 . Relations in a KG are simply not independent of each other. For an entity INLINEFORM0 , one neighboring relation INLINEFORM1 may imply the existence of another neighboring relation INLINEFORM2 , though they may not necessarily connect INLINEFORM3 to the same neighbor. For example, a neighboring relation play_for may suggest the home city, i.e., live_in, of the current athlete entity. Following notations in logics, we denote potential dependency between INLINEFORM4 and INLINEFORM5 by a “logic rule” INLINEFORM6 . To measure the extent of such dependency, we define the confidence of a logic rule INLINEFORM7 as follows: DISPLAYFORM0 Here the function INLINEFORM0 equals 1 when INLINEFORM1 is true and 0 otherwise. As an empirical statistic over the entire KG, INLINEFORM2 is larger if more entities with neighboring relation INLINEFORM3 also have INLINEFORM4 as a neighboring relation. With the confidence scores INLINEFORM0 between all relation pairs at hand, we are ready to characterize neighboring relations INLINEFORM1 that lead to important neighbors. On one hand, such a relation INLINEFORM2 should have a large INLINEFORM3 , i.e., it is statistically relevant to INLINEFORM4 . Following the above example, play_for should be consulted to if the query relation is live_in. On the other hand, INLINEFORM5 should not be implied by other relations in the neighborhood. For example, no matter whether the query relation is live_in or not, the neighboring relation work_as should not be assigned too much weight, because sufficient information is already provided by play_for. Following the above intuitions, we implement the logic rule mechanism of measuring neighboring relations' usefulness as follow: DISPLAYFORM0 We note that INLINEFORM0 promotes relations INLINEFORM1 strongly implying INLINEFORM2 (the numerator) and demotes those implied by some other relation in the same neighborhood (the denominator). In this manner, our logic rule mechanism addresses both query relation awareness and neighborhood redundancy awareness. With global statistics about relations, the logic rule mechanism guides the attention weight to be distributed at a coarse granularity of relations. However, it may be insufficient not to consult finer-grained information hidden in the transformed neighbor embeddings to determine which neighbor is important indeed. To take the transformed embeddings into consideration, we adopt an attention network BIBREF10 . Specifically, given a query relation INLINEFORM0 , the importance of an entity INLINEFORM1 's neighbor INLINEFORM2 is measured by DISPLAYFORM0 Here the unnormalized attention weight INLINEFORM0 is given by an attention neural network as DISPLAYFORM0 In this equation, INLINEFORM0 and INLINEFORM1 are global attention parameters, while INLINEFORM2 is a relation-specific attention parameter for the query relation INLINEFORM3 . All those attention parameters are regarded as parameters of the encoder, and learned directly from the data. Note that, unlike the logic rule mechanism at relation level, the computation of INLINEFORM0 concentrates more on the neighbor INLINEFORM1 itself. This is useful when the neighbor entity INLINEFORM2 is also helpful to explain the current training triplet. For example, in Figure FIGREF12 , the neighbor Chicago_Bulls could help to imply the object of live_in since there are other athletes playing for Chicago_Bulls while living in Chicago. Although working at the neighbor level, the dependency on transformed neighbor embeddings INLINEFORM3 and the relation-specific parameter INLINEFORM4 make INLINEFORM5 aware of both neighborhood redundancy and the query relation. Finally, to incorporate these two weighting mechanisms together in measuring the importance of neighbors, we employ a double-view attention and reformulate Eq. ( EQREF22 ) as DISPLAYFORM0 Training Objective To train the entire model in Figure FIGREF12 , we need both positive triplets and negative ones. All triplets INLINEFORM0 from the knowledge graph naturally serve as positive triplets, which we denote by INLINEFORM1 . To make up for the absence of negative triplets, for each INLINEFORM2 , we randomly corrupt the object or subject (but not both) by another entity in INLINEFORM3 , and denote the corresponding negative triplets by INLINEFORM4 . Formally, DISPLAYFORM0 To encourage the decoder to give high scores for positive triplets and low scores for negative ones, we apply a margin-based ranking loss on each triplet INLINEFORM0 , i.e., DISPLAYFORM0 Here INLINEFORM0 denotes the positive part of x, and INLINEFORM1 is a hyper-parameter for the margin. Finally, the training objective is defined by DISPLAYFORM0 The above training objective only optimizes the output of the aggregator, i.e., the output entity embeddings INLINEFORM0 . The input entity embeddings INLINEFORM1 , however, are not directly aware of the structure of the entire KG. To make the input embeddings and thus the aggregation more meaningful, we set up a subtask for LAN. First, we define a second scoring function, which is similar to Eq. ( EQREF20 ) except that input embeddings INLINEFORM0 from INLINEFORM1 are used to represent the subject and object, i.e., DISPLAYFORM0 The embedding of query relation INLINEFORM0 is obtained from the same embedding matrix INLINEFORM1 as in the first scoring function. Then a similar margin-based ranking loss INLINEFORM2 as Eq. ( EQREF32 ) is defined for the subtask. Finally, we combine the subtask with the main task, and reformulate the overall training objective of LAN as DISPLAYFORM0 Experimental Configurations We evaluate the effectiveness of our LAN model on two typical knowledge graph completion tasks, i.e., link prediction and triplet classification. We compare our LAN with two baseline aggregators, MEAN and LSTM, as described in the Encoder section. MEAN is used on behalf of pooling functions since it leads to the best performance in BIBREF9 ijcai2017-250. LSTM is used due to its large expressive capability BIBREF26 . Data Construction In both tasks, we need datasets whose test sets contain new entities unseen during training. For the task of triplet classification, we directly use the datasets released by BIBREF9 ijcai2017-250 which are based on WordNet11 BIBREF29 . Since they do not conduct experiments on the link prediction task, we construct the required datasets based on FB15K BIBREF11 following a similar protocol used in BIBREF9 ijcai2017-250 as follows. Sampling unseen entities. Firstly, we randomly sample INLINEFORM0 of the original testing triplets to form a new test set INLINEFORM1 for our inductive scenario ( BIBREF9 ijcai2017-250 samples INLINEFORM2 testing triplets). Then two different strategies are used to construct the candidate unseen entities INLINEFORM6 . One is called Subject, where only entities appearing as the subjects in INLINEFORM7 are added to INLINEFORM8 . Another is called Object, where only objects in INLINEFORM9 are added to INLINEFORM10 . For an entity INLINEFORM11 , if it does not have any neighbor in the original training set, such an entity is filtered out, yielding the final unseen entity set INLINEFORM12 . For a triplet INLINEFORM13 , if INLINEFORM14 or INLINEFORM15 , it is removed from INLINEFORM16 . Filtering and splitting data sets. The second step is to ensure that unseen entities would not appear in final training set or validation set. We split the original training set into two data sets, the new training set and auxiliary set. For a triplet INLINEFORM0 in original training set, if INLINEFORM1 , it is added to the new training set. If INLINEFORM2 or INLINEFORM3 , it is added to the auxiliary set, which serves as existing neighbors for unseen entities in INLINEFORM4 . Finally, for a triplet INLINEFORM0 in the original validation set, if INLINEFORM1 or INLINEFORM2 , it is removed from the validation set. The statistics for the resulting INLINEFORM0 datasets using Subject and Object strategies are in Table TABREF34 . Experiments on Triplet Classification Triplet classification aims at classifying a fact triplet INLINEFORM0 as true or false. In the dataset of BIBREF9 ijcai2017-250, triplets in the validation and testing sets are labeled as true or false, while triplets in the training set are all true ones. To tackle this task, we preset a threshold INLINEFORM0 for each relation r. If INLINEFORM1 , the triplet is classified as positive, otherwise it is negative. We determine the optimal INLINEFORM2 by maximizing classification accuracy on the validation set. Experimental Setup Since this task is also conducted in BIBREF9 ijcai2017-250, we use the same configurations with learning rate INLINEFORM0 , embedding dimension INLINEFORM1 , and margin INLINEFORM2 for all datasets. We randomly sample 64 neighbors for each entity. Zero padding is used when the number of neighbors is less than 64. L2-regularization is applied on the parameters of LAN. The regularization rate is INLINEFORM3 . We search the best hyper-parameters of all models according to the performance on validation set. In detail, we search learning rate INLINEFORM0 in INLINEFORM1 , embedding dimension for neighbors INLINEFORM2 in INLINEFORM3 , and margin INLINEFORM4 in INLINEFORM5 . The optimal configurations are INLINEFORM6 for all the datasets. Evaluation Results The results are reported in Table TABREF42 . Since we did not achieve the same results for MEAN as reported in BIBREF9 ijcai2017-250 with either our implementation or their released source code, the best results from their original paper are reported. From the table, we observe that, on one hand, LSTM results in poorer performance compared with MEAN, which involves fewer parameters though. This demonstrates the necessity of the permutation invariance for designing neighborhood aggregators for KGs. On the other hand, our LAN model consistently achieves the best results on all datasets, demonstrating the effectiveness of LAN on this KBC task. Experiments on Link Prediction Link prediction in the inductive setting aims at reasoning the missing part “?” in a triplet when given INLINEFORM0 or INLINEFORM1 with emerging entities INLINEFORM2 or INLINEFORM3 respectively. To tackle the task, we firstly hide the object (subject) of each testing triplet in Subject-R (Object-R) to produce a missing part. Then we replace the missing part with all entities in the entity set INLINEFORM4 to construct candidate triplets. We compute the scoring function INLINEFORM5 defined in Eq. ( EQREF20 ) for all candidate triplets, and rank them in descending order. Finally, we evaluate whether the ground-truth entities are ranked ahead of other entities. We use traditional evaluation metrics as in the KG completion literature, i.e., Mean Rank (MR), Mean Reciprocal Rank (MRR), and the proportion of ground truth entities ranked top-k (Hits@k, INLINEFORM6 ). Since certain candidate triplets might also be true, we follow previous works and filter out these fake negatives before ranking. Experimental Results The results on Subject-10 and Object-10 are reported in Table TABREF43 . The results on other datasets are similar and we summarize them later in Figure FIGREF50 . From Table TABREF43 , we still observe consistent results for all the models as in the triplet classification task. Firstly, LSTM results in the poorest performance on all datasets. Secondly, our LAN model outperforms all the other baselines significantly, especially on the Hit@k metrics. The improvement on the MR metric of LAN might not be considerable. This is due to the flaw of the MR metric since it is more sensitive to lower positions of the ranking, which is actually of less importance. The MRR metric is proposed for this reason, where we could observe consistent improvements brought by LAN. The effectiveness of LAN on link prediction validates LAN's superiority to other aggregators and the necessities to treat the neighbors differently in a permutation invariant way. To analyze whether LAN outperforms the others for expected reasons and generalizes to other configurations, we conduct the following studies. In this experiment, we would like to confirm that it's necessary for the aggregator to be aware of the query relation. Specifically, we investigate the attention neural network and design two degenerated baselines. One is referred to as Query-Attention and is simply an attention network as in LAN except that the logic rule mechanism is removed. The other is referred to as Global-Attention, which is also an attention network except that the query relation embedding INLINEFORM0 in Eq. ( EQREF28 ) is masked by a zero vector. The results are reported in Table TABREF46 . We observe that although superior to MEAN, Global-Attention is outperformed by Query-Attention, demonstrating the necessity of query relation awareness. The superiority of Global-Attention over MEAN could be attributed to the fact that the attention mechanism is effective to identify the neighbors which are globally important regardless of the query. We find that the logic rules greatly help to improve the attention network in LAN. We confirm this point by conducting further experiments where the logic rule mechanism is isolated as a single model (referred to as Logic Rules Only). The results are also demonstrated in Table TABREF46 , from which we find that Query-Attention outperforms MEAN by a limited margin. Meanwhile, Logic Rules Only outperforms both MEAN and Query-Attention by significant margins. These results demonstrate the effectiveness of logic rules in assigning meaningful weights to the neighbors. Specifically, in order to generate representations for unseen entities, it is crucial to incorporate the logic rules to train the aggregator, instead of depending solely on neural networks to learn from the data. By combining the logic rules and neural networks, LAN takes a step further in outperforming all the other models. To find out whether the superiority of LAN to the baselines can generalize to other scoring functions, we replace the scoring function in Eq. ( EQREF20 ) and Eq. ( EQREF36 ) by three typical scoring functions mentioned in Related Works. We omit the results of LSTM, for it is still inferior to MEAN. The results are listed in Table TABREF48 , from which we observe that with different scoring functions, LAN outperforms MEAN consistently by a large margin on all the evaluation metrics. Note that TransE leads to the best results on MEAN and LAN. It's reasonable to suppose that when the ratio of the unseen entities over the training entities increases (namely the observed knowledge graph becomes sparser), all models' performance would deteriorate. To figure out whether our LAN could suffer less on sparse knowledge graphs, we conduct link prediction on datasets with different sample rates INLINEFORM0 as described in Step 1 of the Data Construction section. The results are displayed in Figure FIGREF50 . We observe that the increasing proportion of unseen entities certainly has a negative impact on all models. However, the performance of LAN does not decrease as drastically as that of MEAN and LSTM, indicating that LAN is more robust on sparse KGs. Case Studies on Neighbors' Weights In order to visualize how LAN specifies weights to neighbors, we sample some cases from the Subject-10 testing set. From Table FIGREF50 , we have the following observations. First, with the query relation, LAN could attribute higher weights to neighbors with more relevant relations. In the first case, when the query is origin, the top two neighbors are involved by place_lived and breed_origin, which are helpful to imply origin. In addition, in all three cases, neighbors with relation gender gain the lowest weights since they imply nothing about the query relation. Second, LAN could attribute higher weights to neighbor entities that are more informative. When the query relation is profession, the neighbors Aristotle, Metaphysics and Aesthetics are all relevant to the answer Philosopher. In the third case, we also observe similar situations. Here, the neighbor with the highest weight is (institution, University_of_Calgary) since the query relation place_lived helps the aggregator to focus on the neighboring relation institution, then the neighbor entity University_of_Calgary assists in locating the answer Calgary. Conclusion In this paper, we address inductive KG embedding, which helps embed emerging entities efficiently. We formulate three characteristics required for effective neighborhood aggregators. To meet the three characteristics, we propose LAN, which attributes different weights to an entity's neighbors in a permutation invariant manner, considering both the redundancy of neighbors and the query relation. The weights are estimated from data with logic rules at a coarse relation level, and neural attention network at a fine neighbor level. Experiments show that LAN outperforms baseline models significantly on two typical KG completion tasks. Acknowledgements We thank the three anonymous authors for their constructive comments. This work is supported by the National Natural Science Foundation of China (61472453, U1401256, U1501252, U1611264, U1711261, U1711262).

No

352a1bf734b2d7f0618e9e2b0dbed4a3f1787160

352a1bf734b2d7f0618e9e2b0dbed4a3f1787160_0

Q: Do they evaluate existing methods in terms of desired properties? Text: Introduction Knowledge graphs (KGs) such as Freebase BIBREF0 , DBpedia BIBREF1 , and YAGO BIBREF2 play a critical role in various NLP tasks, including question answering BIBREF3 , information retrieval BIBREF4 , and personalized recommendation BIBREF5 . A typical KG consists of numerous facts about a predefined set of entities. Each fact is in the form of a triplet INLINEFORM0 (or INLINEFORM1 for short), where INLINEFORM2 and INLINEFORM3 are two entities and INLINEFORM4 is a relation the fact describes. Due to the discrete and incomplete natures of KGs, various KG embedding models are proposed to facilitate KG completion tasks, e.g., link prediction and triplet classification. After vectorizing entities and relations in a low-dimensional space, those models predict missing facts by manipulating the involved entity and relation embeddings. Although proving successful in previous studies, traditional KG embedding models simply ignore the evolving nature of KGs. They require all entities to be present when training the embeddings. However, BIBREF6 shi2018open suggest that, on DBpedia, 200 new entities emerge on a daily basis between late 2015 and early 2016. Given the infeasibility of retraining embeddings from scratch whenever new entities come, missing facts about emerging entities are, unfortunately, not guaranteed to be inferred in time. By transforming realistic networks, e.g., citation graphs, social networks, and protein interaction graphs, to simple graphs with single-typed and undirected edges, recent explorations BIBREF7 shed light on the evolution issue for homogeneous graphs. While learning embeddings for existing nodes, they inductively learn a neighborhood aggregator that represents a node by aggregating its neighbors' embeddings. The embeddings of unseen nodes can then be obtained by applying the aggregator on their existing neighbors. It is well received that KGs differ from homogeneous graphs by their multi-relational structure BIBREF8 . Despite the difference, it seems promising to generalize the neighborhood aggregating scheme to embed emerging KG entities in an inductive manner. For example, in Figure FIGREF1 , a news article may describe an emerging entity (marked gray) as well as some facts involving existing entities. By generalizing structural information in the underlying KG, e.g., other entities residing in a similar neighborhood or involving similar relations, to the current entity's neighborhood, we can infer that it may probably live in Chicago. Inspired by the above example, the inductive KG embedding problem boils down to designing a KG-specific neighborhood aggregator to capture essential neighborhood information. Intuitively, an ideal aggregator should have the following desired properties: This paper concentrates on KG-specific neighborhood aggregators, which is of practical importance but only received limited focus BIBREF9 . To the best of our knowledge, neither conventional aggregators for homogeneous graphs nor those for KGs satisfy all the above three properties. In this regard, we employ the attention mechanism BIBREF10 and propose an aggregator called Logic Attention Network (LAN). Aggregating neighbors by a weighted combination of their transformed embeddings, LAN is inherently permutation invariant. To estimate the attention weights in LAN, we adopt two mechanisms to model relation- and neighbor-level information in a coarse-to-fine manner, At both levels, LAN is made aware of both neighborhood redundancy and query relation. To summarize, our contributions are: (1) We propose three desired properties that decent neighborhood aggregators for KGs should possess. (2) We propose a novel aggregator, i.e., Logic Attention Network, to facilitate inductive KG embedding. (3) We conduct extensive comparisons with conventional aggregators on two KG completions tasks. The results validate the superiority of LAN w.r.t. the three properties. Transductive Embedding Models In recent years, representation learning problems on KGs have received much attention due to the wide applications of the resultant entity and relation embeddings. Typical KG embedding models include TransE BIBREF11 , Distmult BIBREF12 , Complex BIBREF13 , Analogy BIBREF14 , to name a few. For more explorations, we refer readers to an extensive survey BIBREF15 . However, conventional approaches on KG embedding work in a transductive manner. They require that all entities should be seen during training. Such limitation hinders them from efficiently generalizing to emerging entities. Inductive Embedding Models To relieve the issue of emerging entities, several inductive KG embedding models are proposed, including BIBREF16 xie2016representation, BIBREF6 shi2018open and BIBREF17 xie2016image which use description text or images as inputs. Although the resultant embeddings may be utilized for KG completion, it is not clear whether the embeddings are powerful enough to infer implicit or new facts beyond those expressed in the text/image. Moreover, when domain experts are recruited to introduce new entities via partial facts rather than text or images, those approaches may not help much. In light of the above scenario, existing neighbors of an emerging entity are considered as another type of input for inductive models. In BIBREF9 ijcai2017-250, the authors propose applying Graph Neural Network (GNN) on the KG, which generates the embedding of a new entity by aggregating all its known neighbors. However, their model aggregates the neighbors via simple pooling functions, which neglects the difference among the neighbors. Other works like BIBREF18 fu2017hin2vec and BIBREF19 tang2015pte aim at embedding nodes for node classification given the entire graph and thus are inapplicable for inductive KG-specific tasks. BIBREF20 schlichtkrull2017modeling and BIBREF21 xiong2018one also rely on neighborhood structures to embed entities, but they either work transductively or focus on emerging relations. Finally, we note another related line of studies on node representation learning for homogeneous graphs. Similar to text- or image-based inductive models for KGs, BIBREF22 duran2017learning, BIBREF23 yang2016revisiting, BIBREF24 velivckovic2017graph and BIBREF25 rossi2018deep exploit additional node attributes to embed unseen nodes. Another work more related to ours is BIBREF26 hamilton2017inductive. They tackle inductive node embedding by the neighborhood aggregation scheme. Their aggregators either trivially treat neighbors equally or unnecessarily require them to be ordered. Moreover, like all embedding models for homogeneous graphs, their model cannot be directly applied to KGs with multi-relational edges. Notations Let INLINEFORM0 and INLINEFORM1 be two sets of entities and relations of size INLINEFORM2 and INLINEFORM3 , respectively. A knowledge graph is composed of a set of triplet facts, namely DISPLAYFORM0 For each INLINEFORM0 , we denote the reverse of INLINEFORM1 by INLINEFORM2 , and add an additional triplet INLINEFORM3 to INLINEFORM4 . For an entity INLINEFORM0 , we denote by INLINEFORM1 its neighborhood in INLINEFORM2 , i.e., all related entities with the involved relations. Formally, DISPLAYFORM0 We denote the projection of INLINEFORM0 on INLINEFORM1 and INLINEFORM2 by INLINEFORM3 and INLINEFORM4 , respectively. Here INLINEFORM5 are neighbors and INLINEFORM6 are neighboring relations. When the context is clear, we simplify the INLINEFORM7 -th entity INLINEFORM8 by its subscript INLINEFORM9 . We denote vectors by bold lower letters, and matrices or sets of vectors by bold upper letters. Given a knowledge graph INLINEFORM0 , we would like to learn a neighborhood aggregator INLINEFORM1 that acts as follows: For an entity INLINEFORM0 on INLINEFORM1 , INLINEFORM2 depends on INLINEFORM3 's neighborhood INLINEFORM4 to embed INLINEFORM5 as a low-dimensional vector INLINEFORM6 ; For an unknown triplet INLINEFORM0 , the embeddings of INLINEFORM1 and INLINEFORM2 output by INLINEFORM3 suggest the plausibility of the triplet. When a new entity emerges with some triplets involving INLINEFORM0 and INLINEFORM1 , we could apply such an aggregator INLINEFORM2 on its newly established neighborhood, and use the output embedding to infer new facts about it. Framework To obtain such a neighborhood aggregator INLINEFORM0 , we adopt an encoder-decoder framework as illustrated by Figure FIGREF12 . Given a training triplet, the encoder INLINEFORM1 encodes INLINEFORM2 and INLINEFORM3 into two embeddings with INLINEFORM4 . The decoder measures the plausibility of the triplet, and provides feedbacks to the encoder to adjust the parameters of INLINEFORM5 . In the remainder of this section, we describe general configurations of the two components. As specified in Figure FIGREF12 , for an entity INLINEFORM0 on focus, the encoder works on a collection of input neighbor embeddings, and output INLINEFORM1 's embedding. To differentiate between input and output embeddings, we use superscripts INLINEFORM2 and INLINEFORM3 on the respective vectors. Let INLINEFORM4 , which is obtained from an embedding matrix INLINEFORM5 , be the embedding of a neighbor INLINEFORM6 , where INLINEFORM7 . To reflect the impact of relation INLINEFORM8 on INLINEFORM9 , we apply a relation-specific transforming function INLINEFORM10 on INLINEFORM11 as follows, DISPLAYFORM0 where INLINEFORM0 is the transforming vector for relation INLINEFORM1 and is restricted as a unit vector. We adopt this transformation from BIBREF27 wang2014knowledge since it does not involve matrix product operations and is of low computation complexity. After neighbor embeddings are transformed, these transformed embeddings are fed to the aggregator INLINEFORM0 to output an embedding INLINEFORM1 for the target entity INLINEFORM2 , i.e., DISPLAYFORM0 By definition, an aggregator INLINEFORM0 essentially takes as input a collection of vectors INLINEFORM1 ( INLINEFORM2 ) and maps them to a single vector. With this observation, the following two types of functions seem to be natural choices for neighborhood aggregators, and have been adopted previously: Pooling Functions. A typical pooling function is mean-pooling, which is defined by INLINEFORM0 . Besides mean-pooling, other previously adopted choices include sum- and max-pooling BIBREF9 . Due to their simple forms, pooling functions are permutation-invariant, but consider the neighbors equally. It is aware of neither potential redundancy in the neighborhood nor the query relations. Recurrent Neural Networks (RNNs). In various natural language processing tasks, RNNs prove effective in modeling sequential dependencies. In BIBREF26 , the authors adopt an RNN variant LSTM BIBREF28 as neighborhood aggregator, i.e., INLINEFORM0 . To train and apply the LSTM-based aggregator, they have to randomly permute the neighbors, which violates the permutation variance property. Given the subject and object embeddings INLINEFORM0 and INLINEFORM1 output by the encoder, the decoder is required to measure the plausibility of the training triplet. To avoid potential mixture with relations INLINEFORM2 in the neighborhood, we refer to the relation in the training triplet by query relation, and denote it by INLINEFORM3 instead. After looking up INLINEFORM4 's representation INLINEFORM5 from an embedding matrix INLINEFORM6 , the decoder scores the training triplet INLINEFORM7 with a scoring function INLINEFORM8 . Following BIBREF9 ijcai2017-250, we mainly investigate a scoring function based on TransE BIBREF11 defined by DISPLAYFORM0 where INLINEFORM0 denotes the L1 norm. To test whether the studied aggregators generalize among different scoring function, we will also consider several alternatives in experiments. Logic Attention Network As discussed above, traditional neighborhood aggregators do not preserve all desired properties. In this section, we describe a novel aggregator, namely Logic Attention Network (LAN), which addresses all three properties. We also provide details in training the LAN aggregator. Incorporating Neighborhood Attention Traditional neighborhood aggregators only depend on collections of transformed embeddings. They neglect other useful information in the neighborhood INLINEFORM0 and the query relation INLINEFORM1 , which may facilitate more effective aggregation of the transformed embeddings. To this end, we propose generalizing the aggregators from INLINEFORM2 to INLINEFORM3 . Specifically, for an entity INLINEFORM0 , its neighbors INLINEFORM1 should contribute differently to INLINEFORM2 according to its importance in representing INLINEFORM3 . To consider the different contribution while preserving the permutation invariance property, we employ a weighted or attention-based aggregating approach on the transformed embeddings. The additional information in INLINEFORM4 and INLINEFORM5 is then exploited when estimating the attention weights. Formally, we obtain INLINEFORM6 by DISPLAYFORM0 Here INLINEFORM0 is the attention weight specified for each neighbor INLINEFORM1 given INLINEFORM2 and the query relation INLINEFORM3 . To assign larger weights INLINEFORM0 to more important neighbors, from the perspective of INLINEFORM1 , we ask ourselves two questions at progressive levels: 1) What types of neighboring relations may lead us to potentially important neighbors? 2) Following those relations, which specific neighbor (in transformed embedding) may contain important information? Inspired by the two questions, we adopt the following two mechanisms to estimate INLINEFORM2 . Relations in a KG are simply not independent of each other. For an entity INLINEFORM0 , one neighboring relation INLINEFORM1 may imply the existence of another neighboring relation INLINEFORM2 , though they may not necessarily connect INLINEFORM3 to the same neighbor. For example, a neighboring relation play_for may suggest the home city, i.e., live_in, of the current athlete entity. Following notations in logics, we denote potential dependency between INLINEFORM4 and INLINEFORM5 by a “logic rule” INLINEFORM6 . To measure the extent of such dependency, we define the confidence of a logic rule INLINEFORM7 as follows: DISPLAYFORM0 Here the function INLINEFORM0 equals 1 when INLINEFORM1 is true and 0 otherwise. As an empirical statistic over the entire KG, INLINEFORM2 is larger if more entities with neighboring relation INLINEFORM3 also have INLINEFORM4 as a neighboring relation. With the confidence scores INLINEFORM0 between all relation pairs at hand, we are ready to characterize neighboring relations INLINEFORM1 that lead to important neighbors. On one hand, such a relation INLINEFORM2 should have a large INLINEFORM3 , i.e., it is statistically relevant to INLINEFORM4 . Following the above example, play_for should be consulted to if the query relation is live_in. On the other hand, INLINEFORM5 should not be implied by other relations in the neighborhood. For example, no matter whether the query relation is live_in or not, the neighboring relation work_as should not be assigned too much weight, because sufficient information is already provided by play_for. Following the above intuitions, we implement the logic rule mechanism of measuring neighboring relations' usefulness as follow: DISPLAYFORM0 We note that INLINEFORM0 promotes relations INLINEFORM1 strongly implying INLINEFORM2 (the numerator) and demotes those implied by some other relation in the same neighborhood (the denominator). In this manner, our logic rule mechanism addresses both query relation awareness and neighborhood redundancy awareness. With global statistics about relations, the logic rule mechanism guides the attention weight to be distributed at a coarse granularity of relations. However, it may be insufficient not to consult finer-grained information hidden in the transformed neighbor embeddings to determine which neighbor is important indeed. To take the transformed embeddings into consideration, we adopt an attention network BIBREF10 . Specifically, given a query relation INLINEFORM0 , the importance of an entity INLINEFORM1 's neighbor INLINEFORM2 is measured by DISPLAYFORM0 Here the unnormalized attention weight INLINEFORM0 is given by an attention neural network as DISPLAYFORM0 In this equation, INLINEFORM0 and INLINEFORM1 are global attention parameters, while INLINEFORM2 is a relation-specific attention parameter for the query relation INLINEFORM3 . All those attention parameters are regarded as parameters of the encoder, and learned directly from the data. Note that, unlike the logic rule mechanism at relation level, the computation of INLINEFORM0 concentrates more on the neighbor INLINEFORM1 itself. This is useful when the neighbor entity INLINEFORM2 is also helpful to explain the current training triplet. For example, in Figure FIGREF12 , the neighbor Chicago_Bulls could help to imply the object of live_in since there are other athletes playing for Chicago_Bulls while living in Chicago. Although working at the neighbor level, the dependency on transformed neighbor embeddings INLINEFORM3 and the relation-specific parameter INLINEFORM4 make INLINEFORM5 aware of both neighborhood redundancy and the query relation. Finally, to incorporate these two weighting mechanisms together in measuring the importance of neighbors, we employ a double-view attention and reformulate Eq. ( EQREF22 ) as DISPLAYFORM0 Training Objective To train the entire model in Figure FIGREF12 , we need both positive triplets and negative ones. All triplets INLINEFORM0 from the knowledge graph naturally serve as positive triplets, which we denote by INLINEFORM1 . To make up for the absence of negative triplets, for each INLINEFORM2 , we randomly corrupt the object or subject (but not both) by another entity in INLINEFORM3 , and denote the corresponding negative triplets by INLINEFORM4 . Formally, DISPLAYFORM0 To encourage the decoder to give high scores for positive triplets and low scores for negative ones, we apply a margin-based ranking loss on each triplet INLINEFORM0 , i.e., DISPLAYFORM0 Here INLINEFORM0 denotes the positive part of x, and INLINEFORM1 is a hyper-parameter for the margin. Finally, the training objective is defined by DISPLAYFORM0 The above training objective only optimizes the output of the aggregator, i.e., the output entity embeddings INLINEFORM0 . The input entity embeddings INLINEFORM1 , however, are not directly aware of the structure of the entire KG. To make the input embeddings and thus the aggregation more meaningful, we set up a subtask for LAN. First, we define a second scoring function, which is similar to Eq. ( EQREF20 ) except that input embeddings INLINEFORM0 from INLINEFORM1 are used to represent the subject and object, i.e., DISPLAYFORM0 The embedding of query relation INLINEFORM0 is obtained from the same embedding matrix INLINEFORM1 as in the first scoring function. Then a similar margin-based ranking loss INLINEFORM2 as Eq. ( EQREF32 ) is defined for the subtask. Finally, we combine the subtask with the main task, and reformulate the overall training objective of LAN as DISPLAYFORM0 Experimental Configurations We evaluate the effectiveness of our LAN model on two typical knowledge graph completion tasks, i.e., link prediction and triplet classification. We compare our LAN with two baseline aggregators, MEAN and LSTM, as described in the Encoder section. MEAN is used on behalf of pooling functions since it leads to the best performance in BIBREF9 ijcai2017-250. LSTM is used due to its large expressive capability BIBREF26 . Data Construction In both tasks, we need datasets whose test sets contain new entities unseen during training. For the task of triplet classification, we directly use the datasets released by BIBREF9 ijcai2017-250 which are based on WordNet11 BIBREF29 . Since they do not conduct experiments on the link prediction task, we construct the required datasets based on FB15K BIBREF11 following a similar protocol used in BIBREF9 ijcai2017-250 as follows. Sampling unseen entities. Firstly, we randomly sample INLINEFORM0 of the original testing triplets to form a new test set INLINEFORM1 for our inductive scenario ( BIBREF9 ijcai2017-250 samples INLINEFORM2 testing triplets). Then two different strategies are used to construct the candidate unseen entities INLINEFORM6 . One is called Subject, where only entities appearing as the subjects in INLINEFORM7 are added to INLINEFORM8 . Another is called Object, where only objects in INLINEFORM9 are added to INLINEFORM10 . For an entity INLINEFORM11 , if it does not have any neighbor in the original training set, such an entity is filtered out, yielding the final unseen entity set INLINEFORM12 . For a triplet INLINEFORM13 , if INLINEFORM14 or INLINEFORM15 , it is removed from INLINEFORM16 . Filtering and splitting data sets. The second step is to ensure that unseen entities would not appear in final training set or validation set. We split the original training set into two data sets, the new training set and auxiliary set. For a triplet INLINEFORM0 in original training set, if INLINEFORM1 , it is added to the new training set. If INLINEFORM2 or INLINEFORM3 , it is added to the auxiliary set, which serves as existing neighbors for unseen entities in INLINEFORM4 . Finally, for a triplet INLINEFORM0 in the original validation set, if INLINEFORM1 or INLINEFORM2 , it is removed from the validation set. The statistics for the resulting INLINEFORM0 datasets using Subject and Object strategies are in Table TABREF34 . Experiments on Triplet Classification Triplet classification aims at classifying a fact triplet INLINEFORM0 as true or false. In the dataset of BIBREF9 ijcai2017-250, triplets in the validation and testing sets are labeled as true or false, while triplets in the training set are all true ones. To tackle this task, we preset a threshold INLINEFORM0 for each relation r. If INLINEFORM1 , the triplet is classified as positive, otherwise it is negative. We determine the optimal INLINEFORM2 by maximizing classification accuracy on the validation set. Experimental Setup Since this task is also conducted in BIBREF9 ijcai2017-250, we use the same configurations with learning rate INLINEFORM0 , embedding dimension INLINEFORM1 , and margin INLINEFORM2 for all datasets. We randomly sample 64 neighbors for each entity. Zero padding is used when the number of neighbors is less than 64. L2-regularization is applied on the parameters of LAN. The regularization rate is INLINEFORM3 . We search the best hyper-parameters of all models according to the performance on validation set. In detail, we search learning rate INLINEFORM0 in INLINEFORM1 , embedding dimension for neighbors INLINEFORM2 in INLINEFORM3 , and margin INLINEFORM4 in INLINEFORM5 . The optimal configurations are INLINEFORM6 for all the datasets. Evaluation Results The results are reported in Table TABREF42 . Since we did not achieve the same results for MEAN as reported in BIBREF9 ijcai2017-250 with either our implementation or their released source code, the best results from their original paper are reported. From the table, we observe that, on one hand, LSTM results in poorer performance compared with MEAN, which involves fewer parameters though. This demonstrates the necessity of the permutation invariance for designing neighborhood aggregators for KGs. On the other hand, our LAN model consistently achieves the best results on all datasets, demonstrating the effectiveness of LAN on this KBC task. Experiments on Link Prediction Link prediction in the inductive setting aims at reasoning the missing part “?” in a triplet when given INLINEFORM0 or INLINEFORM1 with emerging entities INLINEFORM2 or INLINEFORM3 respectively. To tackle the task, we firstly hide the object (subject) of each testing triplet in Subject-R (Object-R) to produce a missing part. Then we replace the missing part with all entities in the entity set INLINEFORM4 to construct candidate triplets. We compute the scoring function INLINEFORM5 defined in Eq. ( EQREF20 ) for all candidate triplets, and rank them in descending order. Finally, we evaluate whether the ground-truth entities are ranked ahead of other entities. We use traditional evaluation metrics as in the KG completion literature, i.e., Mean Rank (MR), Mean Reciprocal Rank (MRR), and the proportion of ground truth entities ranked top-k (Hits@k, INLINEFORM6 ). Since certain candidate triplets might also be true, we follow previous works and filter out these fake negatives before ranking. Experimental Results The results on Subject-10 and Object-10 are reported in Table TABREF43 . The results on other datasets are similar and we summarize them later in Figure FIGREF50 . From Table TABREF43 , we still observe consistent results for all the models as in the triplet classification task. Firstly, LSTM results in the poorest performance on all datasets. Secondly, our LAN model outperforms all the other baselines significantly, especially on the Hit@k metrics. The improvement on the MR metric of LAN might not be considerable. This is due to the flaw of the MR metric since it is more sensitive to lower positions of the ranking, which is actually of less importance. The MRR metric is proposed for this reason, where we could observe consistent improvements brought by LAN. The effectiveness of LAN on link prediction validates LAN's superiority to other aggregators and the necessities to treat the neighbors differently in a permutation invariant way. To analyze whether LAN outperforms the others for expected reasons and generalizes to other configurations, we conduct the following studies. In this experiment, we would like to confirm that it's necessary for the aggregator to be aware of the query relation. Specifically, we investigate the attention neural network and design two degenerated baselines. One is referred to as Query-Attention and is simply an attention network as in LAN except that the logic rule mechanism is removed. The other is referred to as Global-Attention, which is also an attention network except that the query relation embedding INLINEFORM0 in Eq. ( EQREF28 ) is masked by a zero vector. The results are reported in Table TABREF46 . We observe that although superior to MEAN, Global-Attention is outperformed by Query-Attention, demonstrating the necessity of query relation awareness. The superiority of Global-Attention over MEAN could be attributed to the fact that the attention mechanism is effective to identify the neighbors which are globally important regardless of the query. We find that the logic rules greatly help to improve the attention network in LAN. We confirm this point by conducting further experiments where the logic rule mechanism is isolated as a single model (referred to as Logic Rules Only). The results are also demonstrated in Table TABREF46 , from which we find that Query-Attention outperforms MEAN by a limited margin. Meanwhile, Logic Rules Only outperforms both MEAN and Query-Attention by significant margins. These results demonstrate the effectiveness of logic rules in assigning meaningful weights to the neighbors. Specifically, in order to generate representations for unseen entities, it is crucial to incorporate the logic rules to train the aggregator, instead of depending solely on neural networks to learn from the data. By combining the logic rules and neural networks, LAN takes a step further in outperforming all the other models. To find out whether the superiority of LAN to the baselines can generalize to other scoring functions, we replace the scoring function in Eq. ( EQREF20 ) and Eq. ( EQREF36 ) by three typical scoring functions mentioned in Related Works. We omit the results of LSTM, for it is still inferior to MEAN. The results are listed in Table TABREF48 , from which we observe that with different scoring functions, LAN outperforms MEAN consistently by a large margin on all the evaluation metrics. Note that TransE leads to the best results on MEAN and LAN. It's reasonable to suppose that when the ratio of the unseen entities over the training entities increases (namely the observed knowledge graph becomes sparser), all models' performance would deteriorate. To figure out whether our LAN could suffer less on sparse knowledge graphs, we conduct link prediction on datasets with different sample rates INLINEFORM0 as described in Step 1 of the Data Construction section. The results are displayed in Figure FIGREF50 . We observe that the increasing proportion of unseen entities certainly has a negative impact on all models. However, the performance of LAN does not decrease as drastically as that of MEAN and LSTM, indicating that LAN is more robust on sparse KGs. Case Studies on Neighbors' Weights In order to visualize how LAN specifies weights to neighbors, we sample some cases from the Subject-10 testing set. From Table FIGREF50 , we have the following observations. First, with the query relation, LAN could attribute higher weights to neighbors with more relevant relations. In the first case, when the query is origin, the top two neighbors are involved by place_lived and breed_origin, which are helpful to imply origin. In addition, in all three cases, neighbors with relation gender gain the lowest weights since they imply nothing about the query relation. Second, LAN could attribute higher weights to neighbor entities that are more informative. When the query relation is profession, the neighbors Aristotle, Metaphysics and Aesthetics are all relevant to the answer Philosopher. In the third case, we also observe similar situations. Here, the neighbor with the highest weight is (institution, University_of_Calgary) since the query relation place_lived helps the aggregator to focus on the neighboring relation institution, then the neighbor entity University_of_Calgary assists in locating the answer Calgary. Conclusion In this paper, we address inductive KG embedding, which helps embed emerging entities efficiently. We formulate three characteristics required for effective neighborhood aggregators. To meet the three characteristics, we propose LAN, which attributes different weights to an entity's neighbors in a permutation invariant manner, considering both the redundancy of neighbors and the query relation. The weights are estimated from data with logic rules at a coarse relation level, and neural attention network at a fine neighbor level. Experiments show that LAN outperforms baseline models significantly on two typical KG completion tasks. Acknowledgements We thank the three anonymous authors for their constructive comments. This work is supported by the National Natural Science Foundation of China (61472453, U1401256, U1501252, U1611264, U1711261, U1711262).

Yes

045dbdbda5d96a672e5c69442e30dbf21917a1ee

045dbdbda5d96a672e5c69442e30dbf21917a1ee_0

Q: How does the model differ from Generative Adversarial Networks? Text: Introduction It is well known that sentiment annotation or labeling is subjective BIBREF0. Annotators often have many disagreements. This is especially so for crowd-workers who are not well trained. That is why one always feels that there are many errors in an annotated dataset. In this paper, we study whether it is possible to build accurate sentiment classifiers even with noisy-labeled training data. Sentiment classification aims to classify a piece of text according to the polarity of the sentiment expressed in the text, e.g., positive or negative BIBREF1, BIBREF0, BIBREF2. In this work, we focus on sentence-level sentiment classification (SSC) with labeling errors. As we will see in the experiment section, noisy labels in the training data can be highly damaging, especially for DNNs because they easily fit the training data and memorize their labels even when training data are corrupted with noisy labels BIBREF3. Collecting datasets annotated with clean labels is costly and time-consuming as DNN based models usually require a large number of training examples. Researchers and practitioners typically have to resort to crowdsourcing. However, as mentioned above, the crowdsourced annotations can be quite noisy. Research on learning with noisy labels dates back to 1980s BIBREF4. It is still vibrant today BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11, BIBREF12 as it is highly challenging. We will discuss the related work in the next section. This paper studies the problem of learning with noisy labels for SSC. Formally, we study the following problem. Problem Definition: Given noisy labeled training sentences $S=\lbrace (x_1,y_1),...,(x_n,y_n)\rbrace $, where $x_i|_{i=1}^n$ is the $i$-th sentence and $y_i\in \lbrace 1,...,c\rbrace $ is the sentiment label of this sentence, the noisy labeled sentences are used to train a DNN model for a SSC task. The trained model is then used to classify sentences with clean labels to one of the $c$ sentiment labels. In this paper, we propose a convolutional neural Network with Ab-networks (NetAb) to deal with noisy labels during training, as shown in Figure FIGREF2. We will introduce the details in the subsequent sections. Basically, NetAb consists of two convolutional neural networks (CNNs) (see Figure FIGREF2), one for learning sentiment scores to predict `clean' labels and the other for learning a noise transition matrix to handle input noisy labels. We call the two CNNs A-network and Ab-network, respectively. The fundamental here is that (1) DNNs memorize easy instances first and gradually adapt to hard instances as training epochs increase BIBREF3, BIBREF13; and (2) noisy labels are theoretically flipped from the clean/true labels by a noise transition matrix BIBREF14, BIBREF15, BIBREF16, BIBREF17. We motivate and propose a CNN model with a transition layer to estimate the noise transition matrix for the input noisy labels, while exploiting another CNN to predict `clean' labels for the input training (and test) sentences. In training, we pre-train A-network in early epochs and then train Ab-network and A-network with their own loss functions in an alternating manner. To our knowledge, this is the first work that addresses the noisy label problem in sentence-level sentiment analysis. Our experimental results show that the proposed model outperforms the state-of-the-art methods. Related Work Our work is related to sentence sentiment classification (SSC). SSC has been studied extensively BIBREF18, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23, BIBREF24, BIBREF25, BIBREF26, BIBREF27, BIBREF28. None of them can handle noisy labels. Since many social media datasets are noisy, researchers have tried to build robust models BIBREF29, BIBREF30, BIBREF31. However, they treat noisy data as additional information and don't specifically handle noisy labels. A noise-aware classification model in BIBREF12 trains using data annotated with multiple labels. BIBREF32 exploited the connection of users and noisy labels of sentiments in social networks. Since the two works use multiple-labeled data or users' information (we only use single-labeled data, and we do not use any additional information), they have different settings than ours. Our work is closely related to DNNs based approaches to learning with noisy labels. DNNs based approaches explored three main directions: (1) training DNNs on selected samples BIBREF33, BIBREF34, BIBREF35, BIBREF17, (2) modifying the loss function of DNNs with regularization biases BIBREF5, BIBREF36, BIBREF37, BIBREF38, BIBREF39, BIBREF40, and (3) plugging an extra layer into DNNs BIBREF14, BIBREF41, BIBREF15, BIBREF16. All these approaches were proposed for image classification where training images were corrupted with noisy labels. Some of them require noise rate to be known a priori in order to tune their models during training BIBREF37, BIBREF17. Our approach combines direction (1) and direction (3), and trains two networks jointly without knowing the noise rate. We have used five latest existing methods in our experiments for SSC. The experimental results show that they are inferior to our proposed method. In addition, BIBREF42, BIBREF43, BIBREF44, BIBREF45, BIBREF46, and BIBREF47 studied weakly-supervised DNNs or semi-supervised DNNs. But they still need some clean-labeled training data. We use no clean-labeled data. Proposed Model Our model builds on CNN BIBREF25. The key idea is to train two CNNs alternately, one for addressing the input noisy labels and the other for predicting `clean' labels. The overall architecture of the proposed model is given in Figure FIGREF2. Before going further, we first introduce a proposition, a property, and an assumption below. Proposition 1 Noisy labels are flipped from clean labels by an unknown noise transition matrix. Proposition UNKREF3 is reformulated from BIBREF16 and has been investigated in BIBREF14, BIBREF15, BIBREF41. This proposition shows that if we know the noise transition matrix, we can use it to recover the clean labels. In other words, we can put noise transition matrix on clean labels to deal with noisy labels. Given these, we ask the following question: How to estimate such an unknown noise transition matrix? Below we give a solution to this question based on the following property of DNNs. Property 1 DNNs tend to prioritize memorization of simple instances first and then gradually memorize hard instances BIBREF3. BIBREF13 further investigated this property of DNNs. Our setting is that simple instances are sentences of clean labels and hard instances are those with noisy labels. We also have the following assumption. Assumption 1 The noise rate of the training data is less than $50\%$. This assumption is usually satisfied in practice because without it, it is hard to tackle the input noisy labels during training. Based on the above preliminaries, we need to estimate the noisy transition matrix $Q\in \mathbb {R}^{c\times c}$ ($c=2$ in our case, i.e., positive and negative), and train two classifiers $\ddot{y}\sim P(\ddot{y}|x,\theta )$ and $\widehat{y}\sim \ P(\widehat{y}|x,\vartheta )$, where $x$ is an input sentence, $\ddot{y}$ is its noisy label, $\widehat{y}$ is its `clean' label, $\theta $ and $\vartheta $ are the parameters of two classifiers. Note that both $\ddot{y}$ and $\widehat{y}$ here are the prediction results from our model, not the input labels. We propose to formulate the probability of the sentence $x$ labeled as $j$ with where $P(\ddot{y}=j|\widehat{y}=i)$ is an item (the $ji$-th item) in the noisy transition matrix $Q$. We can see that the noisy transition matrix $Q$ is exploited on the `clean' scores $P(\widehat{y}|x,\vartheta )$ to tackle noisy labels. We now present our model NetAb and introduce how NetAb performs Eq. (DISPLAY_FORM6). As shown in Figure FIGREF2, NetAb consists of two CNNs. The intuition here is that we use one CNN to perform $P(\widehat{y}=i|x,\vartheta )$ and use another CNN to perform $P(\ddot{y}=j|x,\theta )$. Meanwhile, the CNN performing $P(\ddot{y}=j|x,\theta )$ estimates the noise transition matrix $Q$ to deal with noisy labels. Thus we add a transition layer into this CNN. More precisely, in Figure FIGREF2, the CNN with a clean loss performs $P(\widehat{y}=i|x,\vartheta )$. We call this CNN the A-network. The other CNN with a noisy loss performs $P(\ddot{y}=j|x,\theta )$. We call this CNN the Ab-network. Ab-network shares all the parameters of A-network except the parameters from the Gate unit and the clean loss. In addition, Ab-network has a transition layer to estimate the noisy transition matrix $Q$. In such a way, A-network predict `clean' labels, and Ab-network handles the input noisy labels. We use cross-entropy with the predicted labels $\ddot{y}$ and the input labels $y$ (given in the dataset) to compute the noisy loss, formulated as below where $\mathbb {I}$ is the indicator function (if $y\!==\!i$, $\mathbb {I}\!=\!1$; otherwise, $\mathbb {I}\!=\!0$), and $|\ddot{S}|$ is the number of sentences to train Ab-network in each batch. Similarly, we use cross-entropy with the predicted labels $\widehat{y}$ and the input labels $y$ to compute the clean loss, formulated as where $|\widehat{S}|$ is the number of sentences to train A-network in each batch. Next we introduce how our model learns the parameters ($\vartheta $, $\theta $ and $Q$). An embedding matrix $v$ is produced for each sentence $x$ by looking up a pre-trained word embedding database (e.g., GloVe.840B BIBREF48). Then an encoding vector $h\!=\!CNN(v)$ (and $u\!=\!CNN(v)$) is produced for each embedding matrix $v$ in A-network (and Ab-network). A sofmax classifier gives us $P(\hat{y}\!=\!i|x,\vartheta )$ (i.e., `clean' sentiment scores) on the learned encoding vector $h$. As the noise transition matrix $Q$ indicates the transition values from clean labels to noisy labels, we compute $Q$ as follows where $W_i$ is a trainable parameter matrix, $b_i$ and $f_i$ are two trainable parameter vectors. They are trained in the Ab-network. Finally, $P(\ddot{y}=j|x,\theta )$ is computed by Eq. (DISPLAY_FORM6). In training, NetAb is trained end-to-end. Based on Proposition UNKREF3 and Property UNKREF4, we pre-train A-network in early epochs (e.g., 5 epochs). Then we train Ab-network and A-network in an alternating manner. The two networks are trained using their respective cross-entropy loss. Given a batch of sentences, we first train Ab-network. Then we use the scores predicted from A-network to select some possibly clean sentences from this batch and train A-network on the selected sentences. Specifically speaking, we use the predicted scores to compute sentiment labels by $\arg \max _i \lbrace \ddot{y}=i|\ddot{y}\sim P(\ddot{y}|x,\theta )\rbrace $. Then we select the sentences whose resulting sentiment label equals to the input label. The selection process is marked by a Gate unit in Figure FIGREF2. When testing a sentence, we use A-network to produce the final classification result. Experiments In this section, we evaluate the performance of the proposed NetAb model. we conduct two types of experiments. (1) We corrupt clean-labeled datasets to produce noisy-labeled datasets to show the impact of noises on sentiment classification accuracy. (2) We collect some real noisy data and use them to train models to evaluate the performance of NetAb. Clean-labeled Datasets. We use three clean labeled datasets. The first one is the movie sentence polarity dataset from BIBREF19. The other two datasets are laptop and restaurant datasets collected from SemEval-2016 . The former consists of laptop review sentences and the latter consists of restaurant review sentences. The original datasets (i.e., Laptop and Restaurant) were annotated with aspect polarity in each sentence. We used all sentences with only one polarity (positive or negative) for their aspects. That is, we only used sentences with aspects having the same sentiment label in each sentence. Thus, the sentiment of each aspect gives the ground-truth as the sentiments of all aspects are the same. For each clean-labeled dataset, the sentences are randomly partitioned into training set and test set with $80\%$ and $20\%$, respectively. Following BIBREF25, We also randomly select $10\%$ of the test data for validation to check the model during training. Summary statistics of the training, validation, and test data are shown in Table TABREF9. Noisy-labeled Training Datasets. For the above three domains (movie, laptop, and restaurant), we collected 2,000 reviews for each domain from the same review source. We extracted sentences from each review and assigned review's label to its sentences. Like previous work, we treat 4 or 5 stars as positive and 1 or 2 stars as negative. The data is noisy because a positive (negative) review can contain negative (positive) sentences, and there are also neutral sentences. This gives us three noisy-labeled training datasets. We still use the same test sets as those for the clean-labeled datasets. Summary statistics of all the datasets are shown in Table TABREF9. Experiment 1: Here we use the clean-labeled data (i.e., the last three columns in Table TABREF9). We corrupt the clean training data by switching the labels of some random instances based on a noise rate parameter. Then we use the corrupted data to train NetAb and CNN BIBREF25. The test accuracy curves with the noise rates [0, $0.1$, $0.2$, $0.3$, $0.4$, $0.5$] are shown in Figure FIGREF13. From the figure, we can see that the test accuracy drops from around 0.8 to 0.5 when the noise rate increases from 0 to 0.5, but our NetAb outperforms CNN. The results clearly show that the performance of the CNN drops quite a lot with the noise rate increasing. Experiment 2: Here we use the real noisy-labeled training data to train our model and the baselines, and then test on the test data in Table TABREF9. Our goal is two fold. First, we want to evaluate NetAb using real noisy data. Second, we want to see whether sentences with review level labels can be used to build effective SSC models. Baselines. We use one strong non-DNN baseline, NBSVM (with unigrams or bigrams features) BIBREF23 and six DNN baselines. The first DNN baseline is CNN BIBREF25, which does not handle noisy labels. The other five were designed to handle noisy labels. The comparison results are shown in Table TABREF12. From the results, we can make the following observations. (1) Our NetAb model achieves the best ACC and F1 on all datasets except for F1 of negative class on Laptop. The results demonstrate the superiority of NetAb. (2) NetAb outperforms the baselines designed for learning with noisy labels. These baselines are inferior to ours as they were tailored for image classification. Note that we found no existing method to deal with noisy labels for SSC. Training Details. We use the publicly available pre-trained embedding GloVe.840B BIBREF48 to initialize the word vectors and the embedding dimension is 300. For each baseline, we obtain the system from its author and use its default parameters. As the DNN baselines (except CNN) were proposed for image classification, we change the input channels from 3 to 1. For our NetAb, we follow BIBREF25 to use window sizes of 3, 4 and 5 words with 100 feature maps per window size, resulting in 300-dimensional encoding vectors. The input length of sentence is set to 40. The network parameters are updated using the Adam optimizer BIBREF49 with a learning rate of 0.001. The learning rate is clipped gradually using a norm of 0.96 in performing the Adam optimization. The dropout rate is 0.5 in the input layer. The number of epochs is 200 and batch size is 50. Conclusions This paper proposed a novel CNN based model for sentence-level sentiment classification learning for data with noisy labels. The proposed model learns to handle noisy labels during training by training two networks alternately. The learned noisy transition matrices are used to tackle noisy labels. Experimental results showed that the proposed model outperforms a wide range of baselines markedly. We believe that learning with noisy labels is a promising direction as it is often easy to collect noisy-labeled training data. Acknowledgments Hao Wang and Yan Yang's work was partially supported by a grant from the National Natural Science Foundation of China (No. 61572407).

Unanswerable

c20b012ad31da46642c553ce462bc0aad56912db

c20b012ad31da46642c553ce462bc0aad56912db_0

Q: What is the dataset used to train the model? Text: Introduction It is well known that sentiment annotation or labeling is subjective BIBREF0. Annotators often have many disagreements. This is especially so for crowd-workers who are not well trained. That is why one always feels that there are many errors in an annotated dataset. In this paper, we study whether it is possible to build accurate sentiment classifiers even with noisy-labeled training data. Sentiment classification aims to classify a piece of text according to the polarity of the sentiment expressed in the text, e.g., positive or negative BIBREF1, BIBREF0, BIBREF2. In this work, we focus on sentence-level sentiment classification (SSC) with labeling errors. As we will see in the experiment section, noisy labels in the training data can be highly damaging, especially for DNNs because they easily fit the training data and memorize their labels even when training data are corrupted with noisy labels BIBREF3. Collecting datasets annotated with clean labels is costly and time-consuming as DNN based models usually require a large number of training examples. Researchers and practitioners typically have to resort to crowdsourcing. However, as mentioned above, the crowdsourced annotations can be quite noisy. Research on learning with noisy labels dates back to 1980s BIBREF4. It is still vibrant today BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11, BIBREF12 as it is highly challenging. We will discuss the related work in the next section. This paper studies the problem of learning with noisy labels for SSC. Formally, we study the following problem. Problem Definition: Given noisy labeled training sentences $S=\lbrace (x_1,y_1),...,(x_n,y_n)\rbrace $, where $x_i|_{i=1}^n$ is the $i$-th sentence and $y_i\in \lbrace 1,...,c\rbrace $ is the sentiment label of this sentence, the noisy labeled sentences are used to train a DNN model for a SSC task. The trained model is then used to classify sentences with clean labels to one of the $c$ sentiment labels. In this paper, we propose a convolutional neural Network with Ab-networks (NetAb) to deal with noisy labels during training, as shown in Figure FIGREF2. We will introduce the details in the subsequent sections. Basically, NetAb consists of two convolutional neural networks (CNNs) (see Figure FIGREF2), one for learning sentiment scores to predict `clean' labels and the other for learning a noise transition matrix to handle input noisy labels. We call the two CNNs A-network and Ab-network, respectively. The fundamental here is that (1) DNNs memorize easy instances first and gradually adapt to hard instances as training epochs increase BIBREF3, BIBREF13; and (2) noisy labels are theoretically flipped from the clean/true labels by a noise transition matrix BIBREF14, BIBREF15, BIBREF16, BIBREF17. We motivate and propose a CNN model with a transition layer to estimate the noise transition matrix for the input noisy labels, while exploiting another CNN to predict `clean' labels for the input training (and test) sentences. In training, we pre-train A-network in early epochs and then train Ab-network and A-network with their own loss functions in an alternating manner. To our knowledge, this is the first work that addresses the noisy label problem in sentence-level sentiment analysis. Our experimental results show that the proposed model outperforms the state-of-the-art methods. Related Work Our work is related to sentence sentiment classification (SSC). SSC has been studied extensively BIBREF18, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23, BIBREF24, BIBREF25, BIBREF26, BIBREF27, BIBREF28. None of them can handle noisy labels. Since many social media datasets are noisy, researchers have tried to build robust models BIBREF29, BIBREF30, BIBREF31. However, they treat noisy data as additional information and don't specifically handle noisy labels. A noise-aware classification model in BIBREF12 trains using data annotated with multiple labels. BIBREF32 exploited the connection of users and noisy labels of sentiments in social networks. Since the two works use multiple-labeled data or users' information (we only use single-labeled data, and we do not use any additional information), they have different settings than ours. Our work is closely related to DNNs based approaches to learning with noisy labels. DNNs based approaches explored three main directions: (1) training DNNs on selected samples BIBREF33, BIBREF34, BIBREF35, BIBREF17, (2) modifying the loss function of DNNs with regularization biases BIBREF5, BIBREF36, BIBREF37, BIBREF38, BIBREF39, BIBREF40, and (3) plugging an extra layer into DNNs BIBREF14, BIBREF41, BIBREF15, BIBREF16. All these approaches were proposed for image classification where training images were corrupted with noisy labels. Some of them require noise rate to be known a priori in order to tune their models during training BIBREF37, BIBREF17. Our approach combines direction (1) and direction (3), and trains two networks jointly without knowing the noise rate. We have used five latest existing methods in our experiments for SSC. The experimental results show that they are inferior to our proposed method. In addition, BIBREF42, BIBREF43, BIBREF44, BIBREF45, BIBREF46, and BIBREF47 studied weakly-supervised DNNs or semi-supervised DNNs. But they still need some clean-labeled training data. We use no clean-labeled data. Proposed Model Our model builds on CNN BIBREF25. The key idea is to train two CNNs alternately, one for addressing the input noisy labels and the other for predicting `clean' labels. The overall architecture of the proposed model is given in Figure FIGREF2. Before going further, we first introduce a proposition, a property, and an assumption below. Proposition 1 Noisy labels are flipped from clean labels by an unknown noise transition matrix. Proposition UNKREF3 is reformulated from BIBREF16 and has been investigated in BIBREF14, BIBREF15, BIBREF41. This proposition shows that if we know the noise transition matrix, we can use it to recover the clean labels. In other words, we can put noise transition matrix on clean labels to deal with noisy labels. Given these, we ask the following question: How to estimate such an unknown noise transition matrix? Below we give a solution to this question based on the following property of DNNs. Property 1 DNNs tend to prioritize memorization of simple instances first and then gradually memorize hard instances BIBREF3. BIBREF13 further investigated this property of DNNs. Our setting is that simple instances are sentences of clean labels and hard instances are those with noisy labels. We also have the following assumption. Assumption 1 The noise rate of the training data is less than $50\%$. This assumption is usually satisfied in practice because without it, it is hard to tackle the input noisy labels during training. Based on the above preliminaries, we need to estimate the noisy transition matrix $Q\in \mathbb {R}^{c\times c}$ ($c=2$ in our case, i.e., positive and negative), and train two classifiers $\ddot{y}\sim P(\ddot{y}|x,\theta )$ and $\widehat{y}\sim \ P(\widehat{y}|x,\vartheta )$, where $x$ is an input sentence, $\ddot{y}$ is its noisy label, $\widehat{y}$ is its `clean' label, $\theta $ and $\vartheta $ are the parameters of two classifiers. Note that both $\ddot{y}$ and $\widehat{y}$ here are the prediction results from our model, not the input labels. We propose to formulate the probability of the sentence $x$ labeled as $j$ with where $P(\ddot{y}=j|\widehat{y}=i)$ is an item (the $ji$-th item) in the noisy transition matrix $Q$. We can see that the noisy transition matrix $Q$ is exploited on the `clean' scores $P(\widehat{y}|x,\vartheta )$ to tackle noisy labels. We now present our model NetAb and introduce how NetAb performs Eq. (DISPLAY_FORM6). As shown in Figure FIGREF2, NetAb consists of two CNNs. The intuition here is that we use one CNN to perform $P(\widehat{y}=i|x,\vartheta )$ and use another CNN to perform $P(\ddot{y}=j|x,\theta )$. Meanwhile, the CNN performing $P(\ddot{y}=j|x,\theta )$ estimates the noise transition matrix $Q$ to deal with noisy labels. Thus we add a transition layer into this CNN. More precisely, in Figure FIGREF2, the CNN with a clean loss performs $P(\widehat{y}=i|x,\vartheta )$. We call this CNN the A-network. The other CNN with a noisy loss performs $P(\ddot{y}=j|x,\theta )$. We call this CNN the Ab-network. Ab-network shares all the parameters of A-network except the parameters from the Gate unit and the clean loss. In addition, Ab-network has a transition layer to estimate the noisy transition matrix $Q$. In such a way, A-network predict `clean' labels, and Ab-network handles the input noisy labels. We use cross-entropy with the predicted labels $\ddot{y}$ and the input labels $y$ (given in the dataset) to compute the noisy loss, formulated as below where $\mathbb {I}$ is the indicator function (if $y\!==\!i$, $\mathbb {I}\!=\!1$; otherwise, $\mathbb {I}\!=\!0$), and $|\ddot{S}|$ is the number of sentences to train Ab-network in each batch. Similarly, we use cross-entropy with the predicted labels $\widehat{y}$ and the input labels $y$ to compute the clean loss, formulated as where $|\widehat{S}|$ is the number of sentences to train A-network in each batch. Next we introduce how our model learns the parameters ($\vartheta $, $\theta $ and $Q$). An embedding matrix $v$ is produced for each sentence $x$ by looking up a pre-trained word embedding database (e.g., GloVe.840B BIBREF48). Then an encoding vector $h\!=\!CNN(v)$ (and $u\!=\!CNN(v)$) is produced for each embedding matrix $v$ in A-network (and Ab-network). A sofmax classifier gives us $P(\hat{y}\!=\!i|x,\vartheta )$ (i.e., `clean' sentiment scores) on the learned encoding vector $h$. As the noise transition matrix $Q$ indicates the transition values from clean labels to noisy labels, we compute $Q$ as follows where $W_i$ is a trainable parameter matrix, $b_i$ and $f_i$ are two trainable parameter vectors. They are trained in the Ab-network. Finally, $P(\ddot{y}=j|x,\theta )$ is computed by Eq. (DISPLAY_FORM6). In training, NetAb is trained end-to-end. Based on Proposition UNKREF3 and Property UNKREF4, we pre-train A-network in early epochs (e.g., 5 epochs). Then we train Ab-network and A-network in an alternating manner. The two networks are trained using their respective cross-entropy loss. Given a batch of sentences, we first train Ab-network. Then we use the scores predicted from A-network to select some possibly clean sentences from this batch and train A-network on the selected sentences. Specifically speaking, we use the predicted scores to compute sentiment labels by $\arg \max _i \lbrace \ddot{y}=i|\ddot{y}\sim P(\ddot{y}|x,\theta )\rbrace $. Then we select the sentences whose resulting sentiment label equals to the input label. The selection process is marked by a Gate unit in Figure FIGREF2. When testing a sentence, we use A-network to produce the final classification result. Experiments In this section, we evaluate the performance of the proposed NetAb model. we conduct two types of experiments. (1) We corrupt clean-labeled datasets to produce noisy-labeled datasets to show the impact of noises on sentiment classification accuracy. (2) We collect some real noisy data and use them to train models to evaluate the performance of NetAb. Clean-labeled Datasets. We use three clean labeled datasets. The first one is the movie sentence polarity dataset from BIBREF19. The other two datasets are laptop and restaurant datasets collected from SemEval-2016 . The former consists of laptop review sentences and the latter consists of restaurant review sentences. The original datasets (i.e., Laptop and Restaurant) were annotated with aspect polarity in each sentence. We used all sentences with only one polarity (positive or negative) for their aspects. That is, we only used sentences with aspects having the same sentiment label in each sentence. Thus, the sentiment of each aspect gives the ground-truth as the sentiments of all aspects are the same. For each clean-labeled dataset, the sentences are randomly partitioned into training set and test set with $80\%$ and $20\%$, respectively. Following BIBREF25, We also randomly select $10\%$ of the test data for validation to check the model during training. Summary statistics of the training, validation, and test data are shown in Table TABREF9. Noisy-labeled Training Datasets. For the above three domains (movie, laptop, and restaurant), we collected 2,000 reviews for each domain from the same review source. We extracted sentences from each review and assigned review's label to its sentences. Like previous work, we treat 4 or 5 stars as positive and 1 or 2 stars as negative. The data is noisy because a positive (negative) review can contain negative (positive) sentences, and there are also neutral sentences. This gives us three noisy-labeled training datasets. We still use the same test sets as those for the clean-labeled datasets. Summary statistics of all the datasets are shown in Table TABREF9. Experiment 1: Here we use the clean-labeled data (i.e., the last three columns in Table TABREF9). We corrupt the clean training data by switching the labels of some random instances based on a noise rate parameter. Then we use the corrupted data to train NetAb and CNN BIBREF25. The test accuracy curves with the noise rates [0, $0.1$, $0.2$, $0.3$, $0.4$, $0.5$] are shown in Figure FIGREF13. From the figure, we can see that the test accuracy drops from around 0.8 to 0.5 when the noise rate increases from 0 to 0.5, but our NetAb outperforms CNN. The results clearly show that the performance of the CNN drops quite a lot with the noise rate increasing. Experiment 2: Here we use the real noisy-labeled training data to train our model and the baselines, and then test on the test data in Table TABREF9. Our goal is two fold. First, we want to evaluate NetAb using real noisy data. Second, we want to see whether sentences with review level labels can be used to build effective SSC models. Baselines. We use one strong non-DNN baseline, NBSVM (with unigrams or bigrams features) BIBREF23 and six DNN baselines. The first DNN baseline is CNN BIBREF25, which does not handle noisy labels. The other five were designed to handle noisy labels. The comparison results are shown in Table TABREF12. From the results, we can make the following observations. (1) Our NetAb model achieves the best ACC and F1 on all datasets except for F1 of negative class on Laptop. The results demonstrate the superiority of NetAb. (2) NetAb outperforms the baselines designed for learning with noisy labels. These baselines are inferior to ours as they were tailored for image classification. Note that we found no existing method to deal with noisy labels for SSC. Training Details. We use the publicly available pre-trained embedding GloVe.840B BIBREF48 to initialize the word vectors and the embedding dimension is 300. For each baseline, we obtain the system from its author and use its default parameters. As the DNN baselines (except CNN) were proposed for image classification, we change the input channels from 3 to 1. For our NetAb, we follow BIBREF25 to use window sizes of 3, 4 and 5 words with 100 feature maps per window size, resulting in 300-dimensional encoding vectors. The input length of sentence is set to 40. The network parameters are updated using the Adam optimizer BIBREF49 with a learning rate of 0.001. The learning rate is clipped gradually using a norm of 0.96 in performing the Adam optimization. The dropout rate is 0.5 in the input layer. The number of epochs is 200 and batch size is 50. Conclusions This paper proposed a novel CNN based model for sentence-level sentiment classification learning for data with noisy labels. The proposed model learns to handle noisy labels during training by training two networks alternately. The learned noisy transition matrices are used to tackle noisy labels. Experimental results showed that the proposed model outperforms a wide range of baselines markedly. We believe that learning with noisy labels is a promising direction as it is often easy to collect noisy-labeled training data. Acknowledgments Hao Wang and Yan Yang's work was partially supported by a grant from the National Natural Science Foundation of China (No. 61572407).

movie sentence polarity dataset from BIBREF19, laptop and restaurant datasets collected from SemEval-201, we collected 2,000 reviews for each domain from the same review source

13e87f6d68f7217fd14f4f9a008a65dd2a0ba91c

13e87f6d68f7217fd14f4f9a008a65dd2a0ba91c_0

Q: What is the performance of the model? Text: Introduction It is well known that sentiment annotation or labeling is subjective BIBREF0. Annotators often have many disagreements. This is especially so for crowd-workers who are not well trained. That is why one always feels that there are many errors in an annotated dataset. In this paper, we study whether it is possible to build accurate sentiment classifiers even with noisy-labeled training data. Sentiment classification aims to classify a piece of text according to the polarity of the sentiment expressed in the text, e.g., positive or negative BIBREF1, BIBREF0, BIBREF2. In this work, we focus on sentence-level sentiment classification (SSC) with labeling errors. As we will see in the experiment section, noisy labels in the training data can be highly damaging, especially for DNNs because they easily fit the training data and memorize their labels even when training data are corrupted with noisy labels BIBREF3. Collecting datasets annotated with clean labels is costly and time-consuming as DNN based models usually require a large number of training examples. Researchers and practitioners typically have to resort to crowdsourcing. However, as mentioned above, the crowdsourced annotations can be quite noisy. Research on learning with noisy labels dates back to 1980s BIBREF4. It is still vibrant today BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11, BIBREF12 as it is highly challenging. We will discuss the related work in the next section. This paper studies the problem of learning with noisy labels for SSC. Formally, we study the following problem. Problem Definition: Given noisy labeled training sentences $S=\lbrace (x_1,y_1),...,(x_n,y_n)\rbrace $, where $x_i|_{i=1}^n$ is the $i$-th sentence and $y_i\in \lbrace 1,...,c\rbrace $ is the sentiment label of this sentence, the noisy labeled sentences are used to train a DNN model for a SSC task. The trained model is then used to classify sentences with clean labels to one of the $c$ sentiment labels. In this paper, we propose a convolutional neural Network with Ab-networks (NetAb) to deal with noisy labels during training, as shown in Figure FIGREF2. We will introduce the details in the subsequent sections. Basically, NetAb consists of two convolutional neural networks (CNNs) (see Figure FIGREF2), one for learning sentiment scores to predict `clean' labels and the other for learning a noise transition matrix to handle input noisy labels. We call the two CNNs A-network and Ab-network, respectively. The fundamental here is that (1) DNNs memorize easy instances first and gradually adapt to hard instances as training epochs increase BIBREF3, BIBREF13; and (2) noisy labels are theoretically flipped from the clean/true labels by a noise transition matrix BIBREF14, BIBREF15, BIBREF16, BIBREF17. We motivate and propose a CNN model with a transition layer to estimate the noise transition matrix for the input noisy labels, while exploiting another CNN to predict `clean' labels for the input training (and test) sentences. In training, we pre-train A-network in early epochs and then train Ab-network and A-network with their own loss functions in an alternating manner. To our knowledge, this is the first work that addresses the noisy label problem in sentence-level sentiment analysis. Our experimental results show that the proposed model outperforms the state-of-the-art methods. Related Work Our work is related to sentence sentiment classification (SSC). SSC has been studied extensively BIBREF18, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23, BIBREF24, BIBREF25, BIBREF26, BIBREF27, BIBREF28. None of them can handle noisy labels. Since many social media datasets are noisy, researchers have tried to build robust models BIBREF29, BIBREF30, BIBREF31. However, they treat noisy data as additional information and don't specifically handle noisy labels. A noise-aware classification model in BIBREF12 trains using data annotated with multiple labels. BIBREF32 exploited the connection of users and noisy labels of sentiments in social networks. Since the two works use multiple-labeled data or users' information (we only use single-labeled data, and we do not use any additional information), they have different settings than ours. Our work is closely related to DNNs based approaches to learning with noisy labels. DNNs based approaches explored three main directions: (1) training DNNs on selected samples BIBREF33, BIBREF34, BIBREF35, BIBREF17, (2) modifying the loss function of DNNs with regularization biases BIBREF5, BIBREF36, BIBREF37, BIBREF38, BIBREF39, BIBREF40, and (3) plugging an extra layer into DNNs BIBREF14, BIBREF41, BIBREF15, BIBREF16. All these approaches were proposed for image classification where training images were corrupted with noisy labels. Some of them require noise rate to be known a priori in order to tune their models during training BIBREF37, BIBREF17. Our approach combines direction (1) and direction (3), and trains two networks jointly without knowing the noise rate. We have used five latest existing methods in our experiments for SSC. The experimental results show that they are inferior to our proposed method. In addition, BIBREF42, BIBREF43, BIBREF44, BIBREF45, BIBREF46, and BIBREF47 studied weakly-supervised DNNs or semi-supervised DNNs. But they still need some clean-labeled training data. We use no clean-labeled data. Proposed Model Our model builds on CNN BIBREF25. The key idea is to train two CNNs alternately, one for addressing the input noisy labels and the other for predicting `clean' labels. The overall architecture of the proposed model is given in Figure FIGREF2. Before going further, we first introduce a proposition, a property, and an assumption below. Proposition 1 Noisy labels are flipped from clean labels by an unknown noise transition matrix. Proposition UNKREF3 is reformulated from BIBREF16 and has been investigated in BIBREF14, BIBREF15, BIBREF41. This proposition shows that if we know the noise transition matrix, we can use it to recover the clean labels. In other words, we can put noise transition matrix on clean labels to deal with noisy labels. Given these, we ask the following question: How to estimate such an unknown noise transition matrix? Below we give a solution to this question based on the following property of DNNs. Property 1 DNNs tend to prioritize memorization of simple instances first and then gradually memorize hard instances BIBREF3. BIBREF13 further investigated this property of DNNs. Our setting is that simple instances are sentences of clean labels and hard instances are those with noisy labels. We also have the following assumption. Assumption 1 The noise rate of the training data is less than $50\%$. This assumption is usually satisfied in practice because without it, it is hard to tackle the input noisy labels during training. Based on the above preliminaries, we need to estimate the noisy transition matrix $Q\in \mathbb {R}^{c\times c}$ ($c=2$ in our case, i.e., positive and negative), and train two classifiers $\ddot{y}\sim P(\ddot{y}|x,\theta )$ and $\widehat{y}\sim \ P(\widehat{y}|x,\vartheta )$, where $x$ is an input sentence, $\ddot{y}$ is its noisy label, $\widehat{y}$ is its `clean' label, $\theta $ and $\vartheta $ are the parameters of two classifiers. Note that both $\ddot{y}$ and $\widehat{y}$ here are the prediction results from our model, not the input labels. We propose to formulate the probability of the sentence $x$ labeled as $j$ with where $P(\ddot{y}=j|\widehat{y}=i)$ is an item (the $ji$-th item) in the noisy transition matrix $Q$. We can see that the noisy transition matrix $Q$ is exploited on the `clean' scores $P(\widehat{y}|x,\vartheta )$ to tackle noisy labels. We now present our model NetAb and introduce how NetAb performs Eq. (DISPLAY_FORM6). As shown in Figure FIGREF2, NetAb consists of two CNNs. The intuition here is that we use one CNN to perform $P(\widehat{y}=i|x,\vartheta )$ and use another CNN to perform $P(\ddot{y}=j|x,\theta )$. Meanwhile, the CNN performing $P(\ddot{y}=j|x,\theta )$ estimates the noise transition matrix $Q$ to deal with noisy labels. Thus we add a transition layer into this CNN. More precisely, in Figure FIGREF2, the CNN with a clean loss performs $P(\widehat{y}=i|x,\vartheta )$. We call this CNN the A-network. The other CNN with a noisy loss performs $P(\ddot{y}=j|x,\theta )$. We call this CNN the Ab-network. Ab-network shares all the parameters of A-network except the parameters from the Gate unit and the clean loss. In addition, Ab-network has a transition layer to estimate the noisy transition matrix $Q$. In such a way, A-network predict `clean' labels, and Ab-network handles the input noisy labels. We use cross-entropy with the predicted labels $\ddot{y}$ and the input labels $y$ (given in the dataset) to compute the noisy loss, formulated as below where $\mathbb {I}$ is the indicator function (if $y\!==\!i$, $\mathbb {I}\!=\!1$; otherwise, $\mathbb {I}\!=\!0$), and $|\ddot{S}|$ is the number of sentences to train Ab-network in each batch. Similarly, we use cross-entropy with the predicted labels $\widehat{y}$ and the input labels $y$ to compute the clean loss, formulated as where $|\widehat{S}|$ is the number of sentences to train A-network in each batch. Next we introduce how our model learns the parameters ($\vartheta $, $\theta $ and $Q$). An embedding matrix $v$ is produced for each sentence $x$ by looking up a pre-trained word embedding database (e.g., GloVe.840B BIBREF48). Then an encoding vector $h\!=\!CNN(v)$ (and $u\!=\!CNN(v)$) is produced for each embedding matrix $v$ in A-network (and Ab-network). A sofmax classifier gives us $P(\hat{y}\!=\!i|x,\vartheta )$ (i.e., `clean' sentiment scores) on the learned encoding vector $h$. As the noise transition matrix $Q$ indicates the transition values from clean labels to noisy labels, we compute $Q$ as follows where $W_i$ is a trainable parameter matrix, $b_i$ and $f_i$ are two trainable parameter vectors. They are trained in the Ab-network. Finally, $P(\ddot{y}=j|x,\theta )$ is computed by Eq. (DISPLAY_FORM6). In training, NetAb is trained end-to-end. Based on Proposition UNKREF3 and Property UNKREF4, we pre-train A-network in early epochs (e.g., 5 epochs). Then we train Ab-network and A-network in an alternating manner. The two networks are trained using their respective cross-entropy loss. Given a batch of sentences, we first train Ab-network. Then we use the scores predicted from A-network to select some possibly clean sentences from this batch and train A-network on the selected sentences. Specifically speaking, we use the predicted scores to compute sentiment labels by $\arg \max _i \lbrace \ddot{y}=i|\ddot{y}\sim P(\ddot{y}|x,\theta )\rbrace $. Then we select the sentences whose resulting sentiment label equals to the input label. The selection process is marked by a Gate unit in Figure FIGREF2. When testing a sentence, we use A-network to produce the final classification result. Experiments In this section, we evaluate the performance of the proposed NetAb model. we conduct two types of experiments. (1) We corrupt clean-labeled datasets to produce noisy-labeled datasets to show the impact of noises on sentiment classification accuracy. (2) We collect some real noisy data and use them to train models to evaluate the performance of NetAb. Clean-labeled Datasets. We use three clean labeled datasets. The first one is the movie sentence polarity dataset from BIBREF19. The other two datasets are laptop and restaurant datasets collected from SemEval-2016 . The former consists of laptop review sentences and the latter consists of restaurant review sentences. The original datasets (i.e., Laptop and Restaurant) were annotated with aspect polarity in each sentence. We used all sentences with only one polarity (positive or negative) for their aspects. That is, we only used sentences with aspects having the same sentiment label in each sentence. Thus, the sentiment of each aspect gives the ground-truth as the sentiments of all aspects are the same. For each clean-labeled dataset, the sentences are randomly partitioned into training set and test set with $80\%$ and $20\%$, respectively. Following BIBREF25, We also randomly select $10\%$ of the test data for validation to check the model during training. Summary statistics of the training, validation, and test data are shown in Table TABREF9. Noisy-labeled Training Datasets. For the above three domains (movie, laptop, and restaurant), we collected 2,000 reviews for each domain from the same review source. We extracted sentences from each review and assigned review's label to its sentences. Like previous work, we treat 4 or 5 stars as positive and 1 or 2 stars as negative. The data is noisy because a positive (negative) review can contain negative (positive) sentences, and there are also neutral sentences. This gives us three noisy-labeled training datasets. We still use the same test sets as those for the clean-labeled datasets. Summary statistics of all the datasets are shown in Table TABREF9. Experiment 1: Here we use the clean-labeled data (i.e., the last three columns in Table TABREF9). We corrupt the clean training data by switching the labels of some random instances based on a noise rate parameter. Then we use the corrupted data to train NetAb and CNN BIBREF25. The test accuracy curves with the noise rates [0, $0.1$, $0.2$, $0.3$, $0.4$, $0.5$] are shown in Figure FIGREF13. From the figure, we can see that the test accuracy drops from around 0.8 to 0.5 when the noise rate increases from 0 to 0.5, but our NetAb outperforms CNN. The results clearly show that the performance of the CNN drops quite a lot with the noise rate increasing. Experiment 2: Here we use the real noisy-labeled training data to train our model and the baselines, and then test on the test data in Table TABREF9. Our goal is two fold. First, we want to evaluate NetAb using real noisy data. Second, we want to see whether sentences with review level labels can be used to build effective SSC models. Baselines. We use one strong non-DNN baseline, NBSVM (with unigrams or bigrams features) BIBREF23 and six DNN baselines. The first DNN baseline is CNN BIBREF25, which does not handle noisy labels. The other five were designed to handle noisy labels. The comparison results are shown in Table TABREF12. From the results, we can make the following observations. (1) Our NetAb model achieves the best ACC and F1 on all datasets except for F1 of negative class on Laptop. The results demonstrate the superiority of NetAb. (2) NetAb outperforms the baselines designed for learning with noisy labels. These baselines are inferior to ours as they were tailored for image classification. Note that we found no existing method to deal with noisy labels for SSC. Training Details. We use the publicly available pre-trained embedding GloVe.840B BIBREF48 to initialize the word vectors and the embedding dimension is 300. For each baseline, we obtain the system from its author and use its default parameters. As the DNN baselines (except CNN) were proposed for image classification, we change the input channels from 3 to 1. For our NetAb, we follow BIBREF25 to use window sizes of 3, 4 and 5 words with 100 feature maps per window size, resulting in 300-dimensional encoding vectors. The input length of sentence is set to 40. The network parameters are updated using the Adam optimizer BIBREF49 with a learning rate of 0.001. The learning rate is clipped gradually using a norm of 0.96 in performing the Adam optimization. The dropout rate is 0.5 in the input layer. The number of epochs is 200 and batch size is 50. Conclusions This paper proposed a novel CNN based model for sentence-level sentiment classification learning for data with noisy labels. The proposed model learns to handle noisy labels during training by training two networks alternately. The learned noisy transition matrices are used to tackle noisy labels. Experimental results showed that the proposed model outperforms a wide range of baselines markedly. We believe that learning with noisy labels is a promising direction as it is often easy to collect noisy-labeled training data. Acknowledgments Hao Wang and Yan Yang's work was partially supported by a grant from the National Natural Science Foundation of China (No. 61572407).

Experiment 1: ACC around 0.5 with 50% noise rate in worst case - clearly higher than baselines for all noise rates Experiment 2: ACC on real noisy datasets: 0.7 on Movie, 0.79 on Laptop, 0.86 on Restaurant (clearly higher than baselines in almost all cases)

89b9a2389166b992c42ca19939d750d88c5fa79b

89b9a2389166b992c42ca19939d750d88c5fa79b_0

Q: Is the model evaluated against a CNN baseline? Text: Introduction It is well known that sentiment annotation or labeling is subjective BIBREF0. Annotators often have many disagreements. This is especially so for crowd-workers who are not well trained. That is why one always feels that there are many errors in an annotated dataset. In this paper, we study whether it is possible to build accurate sentiment classifiers even with noisy-labeled training data. Sentiment classification aims to classify a piece of text according to the polarity of the sentiment expressed in the text, e.g., positive or negative BIBREF1, BIBREF0, BIBREF2. In this work, we focus on sentence-level sentiment classification (SSC) with labeling errors. As we will see in the experiment section, noisy labels in the training data can be highly damaging, especially for DNNs because they easily fit the training data and memorize their labels even when training data are corrupted with noisy labels BIBREF3. Collecting datasets annotated with clean labels is costly and time-consuming as DNN based models usually require a large number of training examples. Researchers and practitioners typically have to resort to crowdsourcing. However, as mentioned above, the crowdsourced annotations can be quite noisy. Research on learning with noisy labels dates back to 1980s BIBREF4. It is still vibrant today BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11, BIBREF12 as it is highly challenging. We will discuss the related work in the next section. This paper studies the problem of learning with noisy labels for SSC. Formally, we study the following problem. Problem Definition: Given noisy labeled training sentences $S=\lbrace (x_1,y_1),...,(x_n,y_n)\rbrace $, where $x_i|_{i=1}^n$ is the $i$-th sentence and $y_i\in \lbrace 1,...,c\rbrace $ is the sentiment label of this sentence, the noisy labeled sentences are used to train a DNN model for a SSC task. The trained model is then used to classify sentences with clean labels to one of the $c$ sentiment labels. In this paper, we propose a convolutional neural Network with Ab-networks (NetAb) to deal with noisy labels during training, as shown in Figure FIGREF2. We will introduce the details in the subsequent sections. Basically, NetAb consists of two convolutional neural networks (CNNs) (see Figure FIGREF2), one for learning sentiment scores to predict `clean' labels and the other for learning a noise transition matrix to handle input noisy labels. We call the two CNNs A-network and Ab-network, respectively. The fundamental here is that (1) DNNs memorize easy instances first and gradually adapt to hard instances as training epochs increase BIBREF3, BIBREF13; and (2) noisy labels are theoretically flipped from the clean/true labels by a noise transition matrix BIBREF14, BIBREF15, BIBREF16, BIBREF17. We motivate and propose a CNN model with a transition layer to estimate the noise transition matrix for the input noisy labels, while exploiting another CNN to predict `clean' labels for the input training (and test) sentences. In training, we pre-train A-network in early epochs and then train Ab-network and A-network with their own loss functions in an alternating manner. To our knowledge, this is the first work that addresses the noisy label problem in sentence-level sentiment analysis. Our experimental results show that the proposed model outperforms the state-of-the-art methods. Related Work Our work is related to sentence sentiment classification (SSC). SSC has been studied extensively BIBREF18, BIBREF19, BIBREF20, BIBREF21, BIBREF22, BIBREF23, BIBREF24, BIBREF25, BIBREF26, BIBREF27, BIBREF28. None of them can handle noisy labels. Since many social media datasets are noisy, researchers have tried to build robust models BIBREF29, BIBREF30, BIBREF31. However, they treat noisy data as additional information and don't specifically handle noisy labels. A noise-aware classification model in BIBREF12 trains using data annotated with multiple labels. BIBREF32 exploited the connection of users and noisy labels of sentiments in social networks. Since the two works use multiple-labeled data or users' information (we only use single-labeled data, and we do not use any additional information), they have different settings than ours. Our work is closely related to DNNs based approaches to learning with noisy labels. DNNs based approaches explored three main directions: (1) training DNNs on selected samples BIBREF33, BIBREF34, BIBREF35, BIBREF17, (2) modifying the loss function of DNNs with regularization biases BIBREF5, BIBREF36, BIBREF37, BIBREF38, BIBREF39, BIBREF40, and (3) plugging an extra layer into DNNs BIBREF14, BIBREF41, BIBREF15, BIBREF16. All these approaches were proposed for image classification where training images were corrupted with noisy labels. Some of them require noise rate to be known a priori in order to tune their models during training BIBREF37, BIBREF17. Our approach combines direction (1) and direction (3), and trains two networks jointly without knowing the noise rate. We have used five latest existing methods in our experiments for SSC. The experimental results show that they are inferior to our proposed method. In addition, BIBREF42, BIBREF43, BIBREF44, BIBREF45, BIBREF46, and BIBREF47 studied weakly-supervised DNNs or semi-supervised DNNs. But they still need some clean-labeled training data. We use no clean-labeled data. Proposed Model Our model builds on CNN BIBREF25. The key idea is to train two CNNs alternately, one for addressing the input noisy labels and the other for predicting `clean' labels. The overall architecture of the proposed model is given in Figure FIGREF2. Before going further, we first introduce a proposition, a property, and an assumption below. Proposition 1 Noisy labels are flipped from clean labels by an unknown noise transition matrix. Proposition UNKREF3 is reformulated from BIBREF16 and has been investigated in BIBREF14, BIBREF15, BIBREF41. This proposition shows that if we know the noise transition matrix, we can use it to recover the clean labels. In other words, we can put noise transition matrix on clean labels to deal with noisy labels. Given these, we ask the following question: How to estimate such an unknown noise transition matrix? Below we give a solution to this question based on the following property of DNNs. Property 1 DNNs tend to prioritize memorization of simple instances first and then gradually memorize hard instances BIBREF3. BIBREF13 further investigated this property of DNNs. Our setting is that simple instances are sentences of clean labels and hard instances are those with noisy labels. We also have the following assumption. Assumption 1 The noise rate of the training data is less than $50\%$. This assumption is usually satisfied in practice because without it, it is hard to tackle the input noisy labels during training. Based on the above preliminaries, we need to estimate the noisy transition matrix $Q\in \mathbb {R}^{c\times c}$ ($c=2$ in our case, i.e., positive and negative), and train two classifiers $\ddot{y}\sim P(\ddot{y}|x,\theta )$ and $\widehat{y}\sim \ P(\widehat{y}|x,\vartheta )$, where $x$ is an input sentence, $\ddot{y}$ is its noisy label, $\widehat{y}$ is its `clean' label, $\theta $ and $\vartheta $ are the parameters of two classifiers. Note that both $\ddot{y}$ and $\widehat{y}$ here are the prediction results from our model, not the input labels. We propose to formulate the probability of the sentence $x$ labeled as $j$ with where $P(\ddot{y}=j|\widehat{y}=i)$ is an item (the $ji$-th item) in the noisy transition matrix $Q$. We can see that the noisy transition matrix $Q$ is exploited on the `clean' scores $P(\widehat{y}|x,\vartheta )$ to tackle noisy labels. We now present our model NetAb and introduce how NetAb performs Eq. (DISPLAY_FORM6). As shown in Figure FIGREF2, NetAb consists of two CNNs. The intuition here is that we use one CNN to perform $P(\widehat{y}=i|x,\vartheta )$ and use another CNN to perform $P(\ddot{y}=j|x,\theta )$. Meanwhile, the CNN performing $P(\ddot{y}=j|x,\theta )$ estimates the noise transition matrix $Q$ to deal with noisy labels. Thus we add a transition layer into this CNN. More precisely, in Figure FIGREF2, the CNN with a clean loss performs $P(\widehat{y}=i|x,\vartheta )$. We call this CNN the A-network. The other CNN with a noisy loss performs $P(\ddot{y}=j|x,\theta )$. We call this CNN the Ab-network. Ab-network shares all the parameters of A-network except the parameters from the Gate unit and the clean loss. In addition, Ab-network has a transition layer to estimate the noisy transition matrix $Q$. In such a way, A-network predict `clean' labels, and Ab-network handles the input noisy labels. We use cross-entropy with the predicted labels $\ddot{y}$ and the input labels $y$ (given in the dataset) to compute the noisy loss, formulated as below where $\mathbb {I}$ is the indicator function (if $y\!==\!i$, $\mathbb {I}\!=\!1$; otherwise, $\mathbb {I}\!=\!0$), and $|\ddot{S}|$ is the number of sentences to train Ab-network in each batch. Similarly, we use cross-entropy with the predicted labels $\widehat{y}$ and the input labels $y$ to compute the clean loss, formulated as where $|\widehat{S}|$ is the number of sentences to train A-network in each batch. Next we introduce how our model learns the parameters ($\vartheta $, $\theta $ and $Q$). An embedding matrix $v$ is produced for each sentence $x$ by looking up a pre-trained word embedding database (e.g., GloVe.840B BIBREF48). Then an encoding vector $h\!=\!CNN(v)$ (and $u\!=\!CNN(v)$) is produced for each embedding matrix $v$ in A-network (and Ab-network). A sofmax classifier gives us $P(\hat{y}\!=\!i|x,\vartheta )$ (i.e., `clean' sentiment scores) on the learned encoding vector $h$. As the noise transition matrix $Q$ indicates the transition values from clean labels to noisy labels, we compute $Q$ as follows where $W_i$ is a trainable parameter matrix, $b_i$ and $f_i$ are two trainable parameter vectors. They are trained in the Ab-network. Finally, $P(\ddot{y}=j|x,\theta )$ is computed by Eq. (DISPLAY_FORM6). In training, NetAb is trained end-to-end. Based on Proposition UNKREF3 and Property UNKREF4, we pre-train A-network in early epochs (e.g., 5 epochs). Then we train Ab-network and A-network in an alternating manner. The two networks are trained using their respective cross-entropy loss. Given a batch of sentences, we first train Ab-network. Then we use the scores predicted from A-network to select some possibly clean sentences from this batch and train A-network on the selected sentences. Specifically speaking, we use the predicted scores to compute sentiment labels by $\arg \max _i \lbrace \ddot{y}=i|\ddot{y}\sim P(\ddot{y}|x,\theta )\rbrace $. Then we select the sentences whose resulting sentiment label equals to the input label. The selection process is marked by a Gate unit in Figure FIGREF2. When testing a sentence, we use A-network to produce the final classification result. Experiments In this section, we evaluate the performance of the proposed NetAb model. we conduct two types of experiments. (1) We corrupt clean-labeled datasets to produce noisy-labeled datasets to show the impact of noises on sentiment classification accuracy. (2) We collect some real noisy data and use them to train models to evaluate the performance of NetAb. Clean-labeled Datasets. We use three clean labeled datasets. The first one is the movie sentence polarity dataset from BIBREF19. The other two datasets are laptop and restaurant datasets collected from SemEval-2016 . The former consists of laptop review sentences and the latter consists of restaurant review sentences. The original datasets (i.e., Laptop and Restaurant) were annotated with aspect polarity in each sentence. We used all sentences with only one polarity (positive or negative) for their aspects. That is, we only used sentences with aspects having the same sentiment label in each sentence. Thus, the sentiment of each aspect gives the ground-truth as the sentiments of all aspects are the same. For each clean-labeled dataset, the sentences are randomly partitioned into training set and test set with $80\%$ and $20\%$, respectively. Following BIBREF25, We also randomly select $10\%$ of the test data for validation to check the model during training. Summary statistics of the training, validation, and test data are shown in Table TABREF9. Noisy-labeled Training Datasets. For the above three domains (movie, laptop, and restaurant), we collected 2,000 reviews for each domain from the same review source. We extracted sentences from each review and assigned review's label to its sentences. Like previous work, we treat 4 or 5 stars as positive and 1 or 2 stars as negative. The data is noisy because a positive (negative) review can contain negative (positive) sentences, and there are also neutral sentences. This gives us three noisy-labeled training datasets. We still use the same test sets as those for the clean-labeled datasets. Summary statistics of all the datasets are shown in Table TABREF9. Experiment 1: Here we use the clean-labeled data (i.e., the last three columns in Table TABREF9). We corrupt the clean training data by switching the labels of some random instances based on a noise rate parameter. Then we use the corrupted data to train NetAb and CNN BIBREF25. The test accuracy curves with the noise rates [0, $0.1$, $0.2$, $0.3$, $0.4$, $0.5$] are shown in Figure FIGREF13. From the figure, we can see that the test accuracy drops from around 0.8 to 0.5 when the noise rate increases from 0 to 0.5, but our NetAb outperforms CNN. The results clearly show that the performance of the CNN drops quite a lot with the noise rate increasing. Experiment 2: Here we use the real noisy-labeled training data to train our model and the baselines, and then test on the test data in Table TABREF9. Our goal is two fold. First, we want to evaluate NetAb using real noisy data. Second, we want to see whether sentences with review level labels can be used to build effective SSC models. Baselines. We use one strong non-DNN baseline, NBSVM (with unigrams or bigrams features) BIBREF23 and six DNN baselines. The first DNN baseline is CNN BIBREF25, which does not handle noisy labels. The other five were designed to handle noisy labels. The comparison results are shown in Table TABREF12. From the results, we can make the following observations. (1) Our NetAb model achieves the best ACC and F1 on all datasets except for F1 of negative class on Laptop. The results demonstrate the superiority of NetAb. (2) NetAb outperforms the baselines designed for learning with noisy labels. These baselines are inferior to ours as they were tailored for image classification. Note that we found no existing method to deal with noisy labels for SSC. Training Details. We use the publicly available pre-trained embedding GloVe.840B BIBREF48 to initialize the word vectors and the embedding dimension is 300. For each baseline, we obtain the system from its author and use its default parameters. As the DNN baselines (except CNN) were proposed for image classification, we change the input channels from 3 to 1. For our NetAb, we follow BIBREF25 to use window sizes of 3, 4 and 5 words with 100 feature maps per window size, resulting in 300-dimensional encoding vectors. The input length of sentence is set to 40. The network parameters are updated using the Adam optimizer BIBREF49 with a learning rate of 0.001. The learning rate is clipped gradually using a norm of 0.96 in performing the Adam optimization. The dropout rate is 0.5 in the input layer. The number of epochs is 200 and batch size is 50. Conclusions This paper proposed a novel CNN based model for sentence-level sentiment classification learning for data with noisy labels. The proposed model learns to handle noisy labels during training by training two networks alternately. The learned noisy transition matrices are used to tackle noisy labels. Experimental results showed that the proposed model outperforms a wide range of baselines markedly. We believe that learning with noisy labels is a promising direction as it is often easy to collect noisy-labeled training data. Acknowledgments Hao Wang and Yan Yang's work was partially supported by a grant from the National Natural Science Foundation of China (No. 61572407).

Yes

dccc3b182861fd19ccce5bd00ce9c3f40451ed6e

dccc3b182861fd19ccce5bd00ce9c3f40451ed6e_0

Q: Does the model proposed beat the baseline models for all the values of the masking parameter tested? Text: Introduction There has been significant research on style transfer, with the goal of changing the style of text while preserving its semantic content. The alternative where semantics are adjusted while keeping style intact, which we call semantic text exchange (STE), has not been investigated to the best of our knowledge. Consider the following example, where the replacement entity defines the new semantic context: Original Text: It is sunny outside! Ugh, that means I must wear sunscreen. I hate being sweaty and sticky all over. Replacement Entity: weather = rainy Desired Text: It is rainy outside! Ugh, that means I must bring an umbrella. I hate being wet and having to carry it around. The weather within the original text is sunny, whereas the actual weather may be rainy. Not only is the word sunny replaced with rainy, but the rest of the text's content is changed while preserving its negative sentiment and fluency. With the rise of natural language processing (NLP) has come an increased demand for massive amounts of text data. Manually collecting and scraping data requires a significant amount of time and effort, and data augmentation techniques for NLP are limited compared to fields such as computer vision. STE can be used for text data augmentation by producing various modifications of a piece of text that differ in semantic content. Another use of STE is in building emotionally aligned chatbots and virtual assistants. This is useful for reasons such as marketing, overall enjoyment of interaction, and mental health therapy. However, due to limited data with emotional content in specific semantic contexts, the generated text may contain incorrect semantic content. STE can adjust text semantics (e.g. to align with reality or a specific task) while preserving emotions. One specific example is the development of virtual assistants with adjustable socio-emotional personalities in the effort to construct assistive technologies for persons with cognitive disabilities. Adjusting the emotional delivery of text in subtle ways can have a strong effect on the adoption of the technologies BIBREF0. It is challenging to transfer style this subtly due to lack of datasets on specific topics with consistent emotions. Instead, large datasets of emotionally consistent interactions not confined to specific topics exist. Hence, it is effective to generate text with a particular emotion and then adjust its semantics. We propose a pipeline called SMERTI (pronounced `smarty') for STE. Combining entity replacement (ER), similarity masking (SM), and text infilling (TI), SMERTI can modify the semantic content of text. We define a metric called the Semantic Text Exchange Score (STES) that evaluates the overall ability of a model to perform STE, and an adjustable parameter masking (replacement) rate threshold (MRT/RRT) that can be used to control the amount of semantic change. We evaluate on three datasets: Yelp and Amazon reviews BIBREF1, and Kaggle news headlines BIBREF2. We implement three baseline models for comparison: Noun WordNet Semantic Text Exchange Model (NWN-STEM), General WordNet Semantic Text Exchange Model (GWN-STEM), and Word2Vec Semantic Text Exchange Model (W2V-STEM). We illustrate the STE performance of two SMERTI variations on the datasets, demonstrating outperformance of the baselines and pipeline stability. We also run a human evaluation supporting our results. We analyze the results in detail and investigate relationships between the semantic change, fluency, sentiment, and MRT/RRT. Our major contributions can be summarized as: We define a new task called semantic text exchange (STE) with increasing importance in NLP applications that modifies text semantics while preserving other aspects such as sentiment. We propose a pipeline SMERTI capable of multi-word entity replacement and text infilling, and demonstrate its outperformance of baselines. We define an evaluation metric for overall performance on semantic text exchange called the Semantic Text Exchange Score (STES). Related Work ::: Word and Sentence-level Embeddings Word2Vec BIBREF3, BIBREF4 allows for analogy representation through vector arithmetic. We implement a baseline (W2V-STEM) using this technique. The Universal Sentence Encoder (USE) BIBREF5 encodes sentences and is trained on a variety of web sources and the Stanford Natural Language Inference corpus BIBREF6. Flair embeddings BIBREF7 are based on architectures such as BERT BIBREF8. We use USE for SMERTI as it is designed for transfer learning and shows higher performance on textual similarity tasks compared to other models BIBREF9. Related Work ::: Text Infilling Text infilling is the task of filling in missing parts of sentences called masks. MaskGAN BIBREF10 is restricted to a single word per mask token, while SMERTI is capable of variable length infilling for more flexible output. BIBREF11 uses a transformer-based architecture. They fill in random masks, while SMERTI fills in masks guided by semantic similarity, resulting in more natural infilling and fulfillment of the STE task. Related Work ::: Style and Sentiment Transfer Notable works in style/sentiment transfer include BIBREF12, BIBREF13, BIBREF14, BIBREF15. They attempt to learn latent representations of various text aspects such as its context and attributes, or separate style from content and encode them into hidden representations. They then use an RNN decoder to generate a new sentence given a targeted sentiment attribute. Related Work ::: Review Generation BIBREF16 generates fake reviews from scratch using language models. BIBREF17, BIBREF18, BIBREF19 generate reviews from scratch given auxiliary information (e.g. the item category and star rating). BIBREF20 generates reviews using RNNs with two components: generation from scratch and review customization (Algorithm 2 in BIBREF20). They define review customization as modifying the generated review to fit a new topic or context, such as from a Japanese restaurant to an Italian one. They condition on a keyword identifying the desired context, and replace similar nouns with others using WordNet BIBREF21. They require a “reference dataset" (required to be “on topic"; easy enough for restaurant reviews, but less so for arbitrary conversational agents). As noted by BIBREF19, the method of BIBREF20 may also replace words independently of context. We implement their review customization algorithm (NWN-STEM) and a modified version (GWN-STEM) as baseline models. SMERTI ::: Overview The task is to transform a corpus $C$ of lines of text $S_i$ and associated replacement entities $RE_i:C = \lbrace (S_1,RE_1),(S_2,RE_2),\ldots , (S_n, RE_n)\rbrace $ to a modified corpus $\hat{C} = \lbrace \hat{S}_1,\hat{S}_2,\ldots ,\hat{S}_n\rbrace $, where $\hat{S}_i$ are the original text lines $S_i$ replaced with $RE_i$ and overall semantics adjusted. SMERTI consists of the following modules, shown in Figure FIGREF15: Entity Replacement Module (ERM): Identify which word(s) within the original text are best replaced with the $RE$, which we call the Original Entity ($OE$). We replace $OE$ in $S$ with $RE$. We call this modified text $S^{\prime }$. Similarity Masking Module (SMM): Identify words/phrases in $S^{\prime }$ similar to $OE$ and replace them with a [mask]. Group adjacent [mask]s into a single one so we can fill a variable length of text into each. We call this masked text $S^{\prime \prime }$. Text Infilling Module (TIM): Fill in [mask] tokens with text that better suits the $RE$. This will modify semantics in the rest of the text. This final output text is called $\hat{S}$. SMERTI ::: Entity Replacement Module (ERM) For entity replacement, we use a combination of the Universal Sentence Encoder BIBREF5 and Stanford Parser BIBREF22. SMERTI ::: Entity Replacement Module (ERM) ::: Stanford Parser The Stanford Parser is a constituency parser that determines the grammatical structure of sentences, including phrases and part-of-speech (POS) labelling. By feeding our $RE$ through the parser, we are able to determine its parse-tree. Iterating through the parse-tree and its sub-trees, we can obtain a list of constituent tags for the $RE$. We then feed our input text $S$ through the parser, and through a similar process, we can obtain a list of leaves (where leaves under a single label are concatenated) that are equal or similar to any of the $RE$ constituent tags. This generates a list of entities having the same (or similar) grammatical structure as the $RE$, and are likely candidates for the $OE$. We then feed these entities along with the $RE$ into the Universal Sentence Encoder (USE). SMERTI ::: Entity Replacement Module (ERM) ::: Universal Sentence Encoder (USE) The USE is a sentence-level embedding model that comes with a deep averaging network (DAN) and transformer model BIBREF5. We choose the transformer model as these embeddings take context into account, and the exact same word/phrase will have a different embedding depending on its context and surrounding words. We compute the semantic similarity between two embeddings $u$ and $v$: $sim(u,v)$, using the angular (cosine) distance, defined as: $\cos (\theta _{u,v}) = (u\cdot v)/(||u|| ||v||)$, such that $sim(u,v) = 1-\frac{1}{\pi }arccos(\cos (\theta _{u,v}))$. Results are in $[0,1]$, with higher values representing greater similarity. Using USE and the above equation, we can identify words/phrases within the input text $S$ which are most similar to $RE$. To assist with this, we use the Stanford Parser as described above to obtain a list of candidate entities. In the rare case that this list is empty, we feed in each word of $S$ into USE, and identify which word is the most similar to $RE$. We then replace the most similar entity or word ($OE$) with the $RE$ and generate $S^{\prime }$. An example of this entity replacement process is in Figure FIGREF18. Two parse-trees are shown: for $RE$ (a) and $S$ (b) and (c). Figure FIGREF18(d) is a semantic similarity heat-map generated from the USE embeddings of the candidate $OE$s and $RE$, where values are similarity scores in the range $[0,1]$. As seen in Figure FIGREF18(d), we calculate semantic similarities between $RE$ and entities within $S$ which have noun constituency tags. Looking at the row for our $RE$ restaurant, the most similar entity (excluding itself) is hotel. We can then generate: $S^{\prime }$ = i love this restaurant ! the beds are comfortable and the service is great ! SMERTI ::: Similarity Masking Module (SMM) Next, we mask words similar to $OE$ to generate $S^{\prime \prime }$ using USE. We look at semantic similarities between every word in $S$ and $OE$, along with semantic similarities between $OE$ and the candidate entities determined in the previous ERM step to broaden the range of phrases our module can mask. We ignore $RE$, $OE$, and any entities or phrases containing $OE$ (for example, `this hotel'). After determining words similar to the $OE$ (discussed below), we replace each of them with a [mask] token. Next, we replace [mask] tokens adjacent to each other with a single [mask]. We set a base similarity threshold (ST) that selects a subset of words to mask. We compare the actual fraction of masked words to the masking rate threshold (MRT), as defined by the user, and increase ST in intervals of $0.05$ until the actual masking rate falls below the MRT. Some sample masked outputs ($S^{\prime \prime }$) using various MRT-ST combinations for the previous example are shown in Table TABREF21 (more examples in Appendix A). The MRT is similar to the temperature parameter used to control the “novelty” of generated text in works such as BIBREF20. A high MRT means the user wants to generate text very semantically dissimilar to the original, and may be desired in cases such as creating a lively chatbot or correcting text that is heavily incorrect semantically. A low MRT means the user wants to generate text semantically similar to the original, and may be desired in cases such as text recovery, grammar correction, or correcting a minor semantic error in text. By varying the MRT, various pieces of text that differ semantically in subtle ways can be generated, assisting greatly with text data augmentation. The MRT also affects sentiment and fluency, as we show in Section SECREF59. SMERTI ::: Text Infilling Module (TIM) We use two seq2seq models for our TIM: an RNN (recurrent neural network) model BIBREF23 (called SMERTI-RNN), and a transformer model (called SMERTI-Transformer). SMERTI ::: Text Infilling Module (TIM) ::: Bidirectional RNN with Attention We use a bidirectional variant of the GRU BIBREF24, and hence two RNNs for the encoder: one reads the input sequence in standard sequential order, and the other is fed this sequence in reverse. The outputs are summed at each time step, giving us the ability to encode information from both past and future context. The decoder generates the output in a sequential token-by-token manner. To combat information loss, we implement the attention mechanism BIBREF25. We use a Luong attention layer BIBREF26 which uses global attention, where all the encoder's hidden states are considered, and use the decoder's current time-step hidden state to calculate attention weights. We use the dot score function for attention, where $h_t$ is the current target decoder state and $\bar{h}_s$ is all encoder states: $score(h_t,\bar{h}_s)=h_t^T\bar{h}_s$. SMERTI ::: Text Infilling Module (TIM) ::: Transformer Our second model makes use of the transformer architecture, and our implementation replicates BIBREF27. We use an encoder-decoder structure with a multi-head self-attention token decoder to condition on information from both past and future context. It maps a query and set of key-value pairs to an output. The queries and keys are of dimension $d_k$, and values of dimension $d_v$. To compute the attention, we pack a set of queries, keys, and values into matrices $Q$, $K$, and $V$, respectively. The matrix of outputs is computed as: Multi-head attention allows the model to jointly attend to information from different positions. The decoder can make use of both local and global semantic information while filling in each [mask]. Experiment ::: Datasets We train our two TIMs on the three datasets. The Amazon dataset BIBREF1 contains over 83 million user reviews on products, with duplicate reviews removed. The Yelp dataset includes over six million user reviews on businesses. The news headlines dataset from Kaggle contains approximately $200,000$ news headlines from 2012 to 2018 obtained from HuffPost BIBREF2. We filter the text to obtain reviews and headlines which are English, do not contain hyperlinks and other obvious noise, and are less than 20 words long. We found that many longer than twenty words ramble on and are too verbose for our purposes. Rather than filtering by individual sentences we keep each text in its entirety so SMERTI can learn to generate multiple sentences at once. We preprocess the text by lowercasing and removing rare/duplicate punctuation and space. For Amazon and Yelp, we treat reviews greater than three stars as containing positive sentiment, equal to three stars as neutral, and less than three stars as negative. For each training and testing set, we include an equal number of randomly selected positive and negative reviews, and half as many neutral reviews. This is because neutral reviews only occupy one out of five stars compared to positive and negative which occupy two each. Our dataset statistics can be found in Appendix B. Experiment ::: Experiment Details To set up our training and testing data for text infilling, we mask the text. We use a tiered masking approach: for each dataset, we randomly mask 15% of the words in one-third of the lines, 30% of the words in another one-third, and 45% in the remaining one-third. These masked texts serve as the inputs, while the original texts serve as the ground-truth. This allows our TIM models to learn relationships between masked words and relationships between masked and unmasked words. The bidirectional RNN decoder fills in blanks one by one, with the objective of minimizing the cross entropy loss between its output and the ground-truth. We use a hidden size of 500, two layers for the encoder and decoder, teacher-forcing ratio of 1.0, learning rate of 0.0001, dropout of 0.1, batch size of 64, and train for up to 40 epochs. For the transformer, we use scaled dot-product attention and the same hyperparameters as BIBREF27. We use the Adam optimizer BIBREF28 with $\beta _1 = 0.9, \beta _2 = 0.98$, and $\epsilon = 10^{-9}$. As in BIBREF27, we increase the $learning\_rate$ linearly for the first $warmup\_steps$ training steps, and then decrease the $learning\_rate$ proportionally to the inverse square root of the step number. We set $factor=1$ and use $warmup\_steps = 2000$. We use a batch size of 4096, and we train for up to 40 epochs. Experiment ::: Baseline Models We implement three models to benchmark against. First is NWN-STEM (Algorithm 2 from BIBREF20). We use the training sets as the “reference review sets" to extract similar nouns to the $RE$ (using MINsim = 0.1). We then replace nouns in the text similar to the $RE$ with nouns extracted from the associated reference review set. Secondly, we modify NWN-STEM to work for verbs and adjectives, and call this GWN-STEM. From the reference review sets, we extract similar nouns, verbs, and adjectives to the $RE$ (using MINsim = 0.1), where the $RE$ is now not restricted to being a noun. We replace nouns, verbs, and adjectives in the text similar to the $RE$ with those extracted from the associated reference review set. Lastly, we implement W2V-STEM using Gensim BIBREF29. We train uni-gram Word2Vec models for single word $RE$s, and four-gram models for phrases. Models are trained on the training sets. We use cosine similarity to determine the most similar word/phrase in the input text to $RE$, which is the replaced $OE$. For all other words/phrases, we calculate $w_{i}^{\prime } = w_{i} - w_{OE} + w_{RE}$, where $w_{i}$ is the original word/phrase's embedding vector, $w_{OE}$ is the $OE$'s, $w_{RE}$ is the $RE$'s, and $w_{i}^{\prime }$ is the resulting embedding vector. The replacement word/phrase is $w_{i}^{\prime }$'s nearest neighbour. We use similarity thresholds to adjust replacement rates (RR) and produce text under various replacement rate thresholds (RRT). Evaluation ::: Evaluation Setup We manually select 10 nouns, 10 verbs, 10 adjectives, and 5 phrases from the top 10% most frequent words/phrases in each test set as our evaluation $RE$s. We filter the verbs and adjectives through a list of sentiment words BIBREF30 to ensure we do not choose $RE$s that would obviously significantly alter the text's sentiment. For each evaluation $RE$, we choose one-hundred lines from the corresponding test set that does not already contain $RE$. We choose lines with at least five words, as many with less carry little semantic meaning (e.g. `Great!', `It is okay'). For Amazon and Yelp, we choose 50 positive and 50 negative lines per $RE$. We repeat this process three times, resulting in three sets of 1000 lines per dataset per POS (excluding phrases), and three sets of 500 lines per dataset for phrases. Our final results are averaged metrics over these three sets. For SMERTI-Transformer, SMERTI-RNN, and W2V-STEM, we generate four outputs per text for MRT/RRT of 20%, 40%, 60%, and 80%, which represent upper-bounds on the percentage of the input that can be masked and/or replaced. Note that NWN-STEM and GWN-STEM can only evaluate on limited POS and their maximum replacement rates are limited. We select MINsim values of 0.075 and 0 for nouns and 0.1 and 0 for verbs, as these result in replacement rates approximately equal to the actual MR/RR of the other models' outputs for 20% and 40% MRT/RRT, respectively. Evaluation ::: Key Evaluation Metrics Fluency (SLOR) We use syntactic log-odds ratio (SLOR) BIBREF31 for sentence level fluency and modify from their word-level formula to character-level ($SLOR_{c}$). We use Flair perplexity values from a language model trained on the One Billion Words corpus BIBREF32: where $|S|$ and $|w|$ are the character lengths of the input text $S$ and the word $w$, respectively, $p_M(S)$ and $p_M(w)$ are the probabilities of $S$ and $w$ under the language model $M$, respectively, and $PPL_S$ and $PPL_w$ are the character-level perplexities of $S$ and $w$, respectively. SLOR (from hereon we refer to character-level SLOR as simply SLOR) measures aspects of text fluency such as grammaticality. Higher values represent higher fluency. We rescale resulting SLOR values to the interval [0,1] by first fitting and normalizing a Gaussian distribution. We then truncate normalized data points outside [-3,3], which shifts approximately 0.69% of total data. Finally, we divide each data point by six and add 0.5 to each result. Sentiment Preservation Accuracy (SPA) is defined as the percentage of outputs that carry the same sentiment as the input. We use VADER BIBREF33 to evaluate sentiment as positive, negative, or neutral. It handles typos, emojis, and other aspects of online text. Content Similarity Score (CSS) ranges from 0 to 1 and indicates the semantic similarity between generated text and the $RE$. A value closer to 1 indicates stronger semantic exchange, as the output is closer in semantic content to the $RE$. We also use the USE for this due to its design and strong performance as previously mentioned. Evaluation ::: Semantic Text Exchange Score (STES) We come up with a single score to evaluate overall performance of a model on STE that combines the key evaluation metrics. It uses the harmonic mean, similar to the F1 score (or F-score) BIBREF34, BIBREF35, and we call it the Semantic Text Exchange Score (STES): where $A$ is SPA, $B$ is SLOR, and $C$ is CSS. STES ranges between 0 and 1, with scores closer to 1 representing higher overall performance. Like the F1 score, STES penalizes models which perform very poorly in one or more metrics, and favors balanced models achieving strong results in all three. Evaluation ::: Automatic Evaluation Results Table TABREF38 shows overall average results by model. Table TABREF41 shows outputs for a Yelp example. As observed from Table TABREF41 (see also Appendix F), SMERTI is able to generate high quality output text similar to the $RE$ while flowing better than other models' outputs. It can replace entire phrases and sentences due to its variable length infilling. Note that for nouns, the outputs from GWN-STEM and NWN-STEM are equivalent. Evaluation ::: Human Evaluation Setup We conduct a human evaluation with eight participants, 6 males and 2 females, that are affiliated project researchers aged 20-39 at the University of Waterloo. We randomly choose one evaluation line for a randomly selected word or phrase for each POS per dataset. The input text and each model's output (for 40% MRT/RRT - chosen as a good middle ground) for each line is presented to participants, resulting in a total of 54 pieces of text, and rated on the following criteria from 1-5: RE Match: “How related is the entire text to the concept of [X]", where [X] is a word or phrase (1 - not at all related, 3 - somewhat related, 5 - very related). Note here that [X] is a given $RE$. Fluency: “Does the text make sense and flow well?" (1 - not at all, 3 - somewhat, 5 - very) Sentiment: “How do you think the author of the text was feeling?" (1 - very negative, 3 - neutral, 5 - very positive) Each participant evaluates every piece of text. They are presented with a single piece of text at a time, with the order of models, POS, and datasets completely randomized. Evaluation ::: Human Evaluation Results Average human evaluation scores are displayed in Table TABREF50. Sentiment Preservation (between 0 and 1) is calculated by comparing the average Sentiment rating for each model's output text to the Sentiment rating of the input text, and if both are less than 2.5 (negative), between 2.5 and 3.5 inclusive (neutral), or greater than 3.5 (positive), this is counted as a valid case of Sentiment Preservation. We repeat this for every evaluation line to calculate the final values per model. Harmonic means of all three metrics (using rescaled 0-1 values of RE Match and Fluency) are also displayed. Analysis ::: Performance by Model As seen in Table TABREF38, both SMERTI variations achieve higher STES and outperform the other models overall, with the WordNet models performing the worst. SMERTI excels especially on fluency and content similarity. The transformer variation achieves slightly higher SLOR, while the RNN variation achieves slightly higher CSS. The WordNet models perform strongest in sentiment preservation (SPA), likely because they modify little of the text and only verbs and nouns. They achieve by far the lowest CSS, likely in part due to this limited text replacement. They also do not account for context, and many words (e.g. proper nouns) do not exist in WordNet. Overall, the WordNet models are not very effective at STE. W2V-STEM achieves the lowest SLOR, especially for higher RRT, as supported by the example in Table TABREF41 (see also Appendix F). W2V-STEM and WordNet models output grammatically incorrect text that flows poorly. In many cases, words are repeated multiple times. We analyze the average Type Token Ratio (TTR) values of each model's outputs, which is the ratio of unique divided by total words. As shown in Table TABREF52, the SMERTI variations achieve the highest TTR, while W2V-STEM and NWN-STEM the lowest. Note that while W2V-STEM achieves lower CSS than SMERTI, it performs comparably in this aspect. This is likely due to its vector arithmetic operations algorithm, which replaces each word with one more similar to the RE. This is also supported by the lower TTR, as W2V-STEM frequently outputs the same words multiple times. Analysis ::: Performance By Model - Human Results As seen in Table TABREF50, the SMERTI variations outperform all baseline models overall, particularly in RE Match. SMERTI-Transformer performs the best, with SMERTI-RNN second. The WordNet models achieve high Sentiment Preservation, but much lower on RE Match. W2V-STEM achieves comparably high RE Match, but lowest Fluency. These results correspond well with our automatic evaluation results in Table TABREF38. We look at the Pearson correlation values between RE Match, Fluency, and Sentiment Preservation with CSS, SLOR, and SPA, respectively. These are 0.9952, 0.9327, and 0.8768, respectively, demonstrating that our automatic metrics are highly effective and correspond well with human ratings. Analysis ::: SMERTI's Performance By POS As seen from Table TABREF55 , SMERTI's SPA values are highest for nouns, likely because they typically carry little sentiment, and lowest for adjectives, likely because they typically carry the most. SLOR is lowest for adjectives and highest for phrases and nouns. Adjectives typically carry less semantic meaning and SMERTI likely has more trouble figuring out how best to infill the text. In contrast, nouns typically carry more, and phrases the most (since they consist of multiple words). SMERTI's CSS is highest for phrases then nouns, likely due to phrases and nouns carrying more semantic meaning, making it easier to generate semantically similar text. Both SMERTI's and the input text's CSS are lowest for adjectives, likely because they carry little semantic meaning. Overall, SMERTI appears to be more effective on nouns and phrases than verbs and adjectives. Analysis ::: SMERTI's Performance By Dataset As seen in Table TABREF58, SMERTI's SPA is lowest for news headlines. Amazon and Yelp reviews naturally carry stronger sentiment, likely making it easier to generate text with similar sentiment. Both SMERTI's and the input text's SLOR appear to be lower for Yelp reviews. This may be due to many reasons, such as more typos and emojis within the original reviews, and so forth. SMERTI's CSS values are slightly higher for news headlines. This may be due to them typically being shorter and carrying more semantic meaning as they are designed to be attention grabbers. Overall, it seems that using datasets which inherently carry more sentiment will lead to better sentiment preservation. Further, the quality of the dataset's original text, unsurprisingly, influences the ability of SMERTI to generate fluent text. Analysis ::: SMERTI's Performance By MRT/RRT From Table TABREF60, it can be seen that as MRT/RRT increases, SMERTI's SPA and SLOR decrease while CSS increases. These relationships are very strong as supported by the Pearson correlation values of -0.9972, -0.9183, and 0.9078, respectively. When SMERTI can alter more text, it has the opportunity to replace more related to sentiment while producing more of semantic similarity to the $RE$. Further, SMERTI generates more of the text itself, becoming less similar to the human-written input, resulting in lower fluency. To further demonstrate this, we look at average SMERTI BLEU BIBREF36 scores against MRT/RRT, shown in Table TABREF60. BLEU generally indicates how close two pieces of text are in content and structure, with higher values indicating greater similarity. We report our final BLEU scores as the average scores of 1 to 4-grams. As expected, BLEU decreases as MRT/RRT increases, and this relationship is very strong as supported by the Pearson correlation value of -0.9960. It is clear that MRT/RRT represents a trade-off between CSS against SPA and SLOR. It is thus an adjustable parameter that can be used to control the generated text, and balance semantic exchange against fluency and sentiment preservation. Conclusion and Future Work We introduced the task of semantic text exchange (STE), demonstrated that our pipeline SMERTI performs well on STE, and proposed an STES metric for evaluating overall STE performance. SMERTI outperformed other models and was the most balanced overall. We also showed a trade-off between semantic exchange against fluency and sentiment preservation, which can be controlled by the masking (replacement) rate threshold. Potential directions for future work include adding specific methods to control sentiment, and fine-tuning SMERTI for preservation of persona or personality. Experimenting with other text infilling models (e.g. fine-tuning BERT BIBREF8) is also an area of exploration. Lastly, our human evaluation is limited in size and a larger and more diverse participant pool is needed. We conclude by addressing potential ethical misuses of STE, including assisting in the generation of spam and fake-reviews/news. These risks come with any intelligent chatbot work, but we feel that the benefits, including usage in the detection of misuse such as fake-news, greatly outweigh the risks and help progress NLP and AI research. Acknowledgments We thank our anonymous reviewers, study participants, and Huawei Technologies Co., Ltd. for financial support.

No

98ba7a7aae388b1a77dd6cab890977251d906359

98ba7a7aae388b1a77dd6cab890977251d906359_0

Q: Has STES been previously used in the literature to evaluate similar tasks? Text: Introduction There has been significant research on style transfer, with the goal of changing the style of text while preserving its semantic content. The alternative where semantics are adjusted while keeping style intact, which we call semantic text exchange (STE), has not been investigated to the best of our knowledge. Consider the following example, where the replacement entity defines the new semantic context: Original Text: It is sunny outside! Ugh, that means I must wear sunscreen. I hate being sweaty and sticky all over. Replacement Entity: weather = rainy Desired Text: It is rainy outside! Ugh, that means I must bring an umbrella. I hate being wet and having to carry it around. The weather within the original text is sunny, whereas the actual weather may be rainy. Not only is the word sunny replaced with rainy, but the rest of the text's content is changed while preserving its negative sentiment and fluency. With the rise of natural language processing (NLP) has come an increased demand for massive amounts of text data. Manually collecting and scraping data requires a significant amount of time and effort, and data augmentation techniques for NLP are limited compared to fields such as computer vision. STE can be used for text data augmentation by producing various modifications of a piece of text that differ in semantic content. Another use of STE is in building emotionally aligned chatbots and virtual assistants. This is useful for reasons such as marketing, overall enjoyment of interaction, and mental health therapy. However, due to limited data with emotional content in specific semantic contexts, the generated text may contain incorrect semantic content. STE can adjust text semantics (e.g. to align with reality or a specific task) while preserving emotions. One specific example is the development of virtual assistants with adjustable socio-emotional personalities in the effort to construct assistive technologies for persons with cognitive disabilities. Adjusting the emotional delivery of text in subtle ways can have a strong effect on the adoption of the technologies BIBREF0. It is challenging to transfer style this subtly due to lack of datasets on specific topics with consistent emotions. Instead, large datasets of emotionally consistent interactions not confined to specific topics exist. Hence, it is effective to generate text with a particular emotion and then adjust its semantics. We propose a pipeline called SMERTI (pronounced `smarty') for STE. Combining entity replacement (ER), similarity masking (SM), and text infilling (TI), SMERTI can modify the semantic content of text. We define a metric called the Semantic Text Exchange Score (STES) that evaluates the overall ability of a model to perform STE, and an adjustable parameter masking (replacement) rate threshold (MRT/RRT) that can be used to control the amount of semantic change. We evaluate on three datasets: Yelp and Amazon reviews BIBREF1, and Kaggle news headlines BIBREF2. We implement three baseline models for comparison: Noun WordNet Semantic Text Exchange Model (NWN-STEM), General WordNet Semantic Text Exchange Model (GWN-STEM), and Word2Vec Semantic Text Exchange Model (W2V-STEM). We illustrate the STE performance of two SMERTI variations on the datasets, demonstrating outperformance of the baselines and pipeline stability. We also run a human evaluation supporting our results. We analyze the results in detail and investigate relationships between the semantic change, fluency, sentiment, and MRT/RRT. Our major contributions can be summarized as: We define a new task called semantic text exchange (STE) with increasing importance in NLP applications that modifies text semantics while preserving other aspects such as sentiment. We propose a pipeline SMERTI capable of multi-word entity replacement and text infilling, and demonstrate its outperformance of baselines. We define an evaluation metric for overall performance on semantic text exchange called the Semantic Text Exchange Score (STES). Related Work ::: Word and Sentence-level Embeddings Word2Vec BIBREF3, BIBREF4 allows for analogy representation through vector arithmetic. We implement a baseline (W2V-STEM) using this technique. The Universal Sentence Encoder (USE) BIBREF5 encodes sentences and is trained on a variety of web sources and the Stanford Natural Language Inference corpus BIBREF6. Flair embeddings BIBREF7 are based on architectures such as BERT BIBREF8. We use USE for SMERTI as it is designed for transfer learning and shows higher performance on textual similarity tasks compared to other models BIBREF9. Related Work ::: Text Infilling Text infilling is the task of filling in missing parts of sentences called masks. MaskGAN BIBREF10 is restricted to a single word per mask token, while SMERTI is capable of variable length infilling for more flexible output. BIBREF11 uses a transformer-based architecture. They fill in random masks, while SMERTI fills in masks guided by semantic similarity, resulting in more natural infilling and fulfillment of the STE task. Related Work ::: Style and Sentiment Transfer Notable works in style/sentiment transfer include BIBREF12, BIBREF13, BIBREF14, BIBREF15. They attempt to learn latent representations of various text aspects such as its context and attributes, or separate style from content and encode them into hidden representations. They then use an RNN decoder to generate a new sentence given a targeted sentiment attribute. Related Work ::: Review Generation BIBREF16 generates fake reviews from scratch using language models. BIBREF17, BIBREF18, BIBREF19 generate reviews from scratch given auxiliary information (e.g. the item category and star rating). BIBREF20 generates reviews using RNNs with two components: generation from scratch and review customization (Algorithm 2 in BIBREF20). They define review customization as modifying the generated review to fit a new topic or context, such as from a Japanese restaurant to an Italian one. They condition on a keyword identifying the desired context, and replace similar nouns with others using WordNet BIBREF21. They require a “reference dataset" (required to be “on topic"; easy enough for restaurant reviews, but less so for arbitrary conversational agents). As noted by BIBREF19, the method of BIBREF20 may also replace words independently of context. We implement their review customization algorithm (NWN-STEM) and a modified version (GWN-STEM) as baseline models. SMERTI ::: Overview The task is to transform a corpus $C$ of lines of text $S_i$ and associated replacement entities $RE_i:C = \lbrace (S_1,RE_1),(S_2,RE_2),\ldots , (S_n, RE_n)\rbrace $ to a modified corpus $\hat{C} = \lbrace \hat{S}_1,\hat{S}_2,\ldots ,\hat{S}_n\rbrace $, where $\hat{S}_i$ are the original text lines $S_i$ replaced with $RE_i$ and overall semantics adjusted. SMERTI consists of the following modules, shown in Figure FIGREF15: Entity Replacement Module (ERM): Identify which word(s) within the original text are best replaced with the $RE$, which we call the Original Entity ($OE$). We replace $OE$ in $S$ with $RE$. We call this modified text $S^{\prime }$. Similarity Masking Module (SMM): Identify words/phrases in $S^{\prime }$ similar to $OE$ and replace them with a [mask]. Group adjacent [mask]s into a single one so we can fill a variable length of text into each. We call this masked text $S^{\prime \prime }$. Text Infilling Module (TIM): Fill in [mask] tokens with text that better suits the $RE$. This will modify semantics in the rest of the text. This final output text is called $\hat{S}$. SMERTI ::: Entity Replacement Module (ERM) For entity replacement, we use a combination of the Universal Sentence Encoder BIBREF5 and Stanford Parser BIBREF22. SMERTI ::: Entity Replacement Module (ERM) ::: Stanford Parser The Stanford Parser is a constituency parser that determines the grammatical structure of sentences, including phrases and part-of-speech (POS) labelling. By feeding our $RE$ through the parser, we are able to determine its parse-tree. Iterating through the parse-tree and its sub-trees, we can obtain a list of constituent tags for the $RE$. We then feed our input text $S$ through the parser, and through a similar process, we can obtain a list of leaves (where leaves under a single label are concatenated) that are equal or similar to any of the $RE$ constituent tags. This generates a list of entities having the same (or similar) grammatical structure as the $RE$, and are likely candidates for the $OE$. We then feed these entities along with the $RE$ into the Universal Sentence Encoder (USE). SMERTI ::: Entity Replacement Module (ERM) ::: Universal Sentence Encoder (USE) The USE is a sentence-level embedding model that comes with a deep averaging network (DAN) and transformer model BIBREF5. We choose the transformer model as these embeddings take context into account, and the exact same word/phrase will have a different embedding depending on its context and surrounding words. We compute the semantic similarity between two embeddings $u$ and $v$: $sim(u,v)$, using the angular (cosine) distance, defined as: $\cos (\theta _{u,v}) = (u\cdot v)/(||u|| ||v||)$, such that $sim(u,v) = 1-\frac{1}{\pi }arccos(\cos (\theta _{u,v}))$. Results are in $[0,1]$, with higher values representing greater similarity. Using USE and the above equation, we can identify words/phrases within the input text $S$ which are most similar to $RE$. To assist with this, we use the Stanford Parser as described above to obtain a list of candidate entities. In the rare case that this list is empty, we feed in each word of $S$ into USE, and identify which word is the most similar to $RE$. We then replace the most similar entity or word ($OE$) with the $RE$ and generate $S^{\prime }$. An example of this entity replacement process is in Figure FIGREF18. Two parse-trees are shown: for $RE$ (a) and $S$ (b) and (c). Figure FIGREF18(d) is a semantic similarity heat-map generated from the USE embeddings of the candidate $OE$s and $RE$, where values are similarity scores in the range $[0,1]$. As seen in Figure FIGREF18(d), we calculate semantic similarities between $RE$ and entities within $S$ which have noun constituency tags. Looking at the row for our $RE$ restaurant, the most similar entity (excluding itself) is hotel. We can then generate: $S^{\prime }$ = i love this restaurant ! the beds are comfortable and the service is great ! SMERTI ::: Similarity Masking Module (SMM) Next, we mask words similar to $OE$ to generate $S^{\prime \prime }$ using USE. We look at semantic similarities between every word in $S$ and $OE$, along with semantic similarities between $OE$ and the candidate entities determined in the previous ERM step to broaden the range of phrases our module can mask. We ignore $RE$, $OE$, and any entities or phrases containing $OE$ (for example, `this hotel'). After determining words similar to the $OE$ (discussed below), we replace each of them with a [mask] token. Next, we replace [mask] tokens adjacent to each other with a single [mask]. We set a base similarity threshold (ST) that selects a subset of words to mask. We compare the actual fraction of masked words to the masking rate threshold (MRT), as defined by the user, and increase ST in intervals of $0.05$ until the actual masking rate falls below the MRT. Some sample masked outputs ($S^{\prime \prime }$) using various MRT-ST combinations for the previous example are shown in Table TABREF21 (more examples in Appendix A). The MRT is similar to the temperature parameter used to control the “novelty” of generated text in works such as BIBREF20. A high MRT means the user wants to generate text very semantically dissimilar to the original, and may be desired in cases such as creating a lively chatbot or correcting text that is heavily incorrect semantically. A low MRT means the user wants to generate text semantically similar to the original, and may be desired in cases such as text recovery, grammar correction, or correcting a minor semantic error in text. By varying the MRT, various pieces of text that differ semantically in subtle ways can be generated, assisting greatly with text data augmentation. The MRT also affects sentiment and fluency, as we show in Section SECREF59. SMERTI ::: Text Infilling Module (TIM) We use two seq2seq models for our TIM: an RNN (recurrent neural network) model BIBREF23 (called SMERTI-RNN), and a transformer model (called SMERTI-Transformer). SMERTI ::: Text Infilling Module (TIM) ::: Bidirectional RNN with Attention We use a bidirectional variant of the GRU BIBREF24, and hence two RNNs for the encoder: one reads the input sequence in standard sequential order, and the other is fed this sequence in reverse. The outputs are summed at each time step, giving us the ability to encode information from both past and future context. The decoder generates the output in a sequential token-by-token manner. To combat information loss, we implement the attention mechanism BIBREF25. We use a Luong attention layer BIBREF26 which uses global attention, where all the encoder's hidden states are considered, and use the decoder's current time-step hidden state to calculate attention weights. We use the dot score function for attention, where $h_t$ is the current target decoder state and $\bar{h}_s$ is all encoder states: $score(h_t,\bar{h}_s)=h_t^T\bar{h}_s$. SMERTI ::: Text Infilling Module (TIM) ::: Transformer Our second model makes use of the transformer architecture, and our implementation replicates BIBREF27. We use an encoder-decoder structure with a multi-head self-attention token decoder to condition on information from both past and future context. It maps a query and set of key-value pairs to an output. The queries and keys are of dimension $d_k$, and values of dimension $d_v$. To compute the attention, we pack a set of queries, keys, and values into matrices $Q$, $K$, and $V$, respectively. The matrix of outputs is computed as: Multi-head attention allows the model to jointly attend to information from different positions. The decoder can make use of both local and global semantic information while filling in each [mask]. Experiment ::: Datasets We train our two TIMs on the three datasets. The Amazon dataset BIBREF1 contains over 83 million user reviews on products, with duplicate reviews removed. The Yelp dataset includes over six million user reviews on businesses. The news headlines dataset from Kaggle contains approximately $200,000$ news headlines from 2012 to 2018 obtained from HuffPost BIBREF2. We filter the text to obtain reviews and headlines which are English, do not contain hyperlinks and other obvious noise, and are less than 20 words long. We found that many longer than twenty words ramble on and are too verbose for our purposes. Rather than filtering by individual sentences we keep each text in its entirety so SMERTI can learn to generate multiple sentences at once. We preprocess the text by lowercasing and removing rare/duplicate punctuation and space. For Amazon and Yelp, we treat reviews greater than three stars as containing positive sentiment, equal to three stars as neutral, and less than three stars as negative. For each training and testing set, we include an equal number of randomly selected positive and negative reviews, and half as many neutral reviews. This is because neutral reviews only occupy one out of five stars compared to positive and negative which occupy two each. Our dataset statistics can be found in Appendix B. Experiment ::: Experiment Details To set up our training and testing data for text infilling, we mask the text. We use a tiered masking approach: for each dataset, we randomly mask 15% of the words in one-third of the lines, 30% of the words in another one-third, and 45% in the remaining one-third. These masked texts serve as the inputs, while the original texts serve as the ground-truth. This allows our TIM models to learn relationships between masked words and relationships between masked and unmasked words. The bidirectional RNN decoder fills in blanks one by one, with the objective of minimizing the cross entropy loss between its output and the ground-truth. We use a hidden size of 500, two layers for the encoder and decoder, teacher-forcing ratio of 1.0, learning rate of 0.0001, dropout of 0.1, batch size of 64, and train for up to 40 epochs. For the transformer, we use scaled dot-product attention and the same hyperparameters as BIBREF27. We use the Adam optimizer BIBREF28 with $\beta _1 = 0.9, \beta _2 = 0.98$, and $\epsilon = 10^{-9}$. As in BIBREF27, we increase the $learning\_rate$ linearly for the first $warmup\_steps$ training steps, and then decrease the $learning\_rate$ proportionally to the inverse square root of the step number. We set $factor=1$ and use $warmup\_steps = 2000$. We use a batch size of 4096, and we train for up to 40 epochs. Experiment ::: Baseline Models We implement three models to benchmark against. First is NWN-STEM (Algorithm 2 from BIBREF20). We use the training sets as the “reference review sets" to extract similar nouns to the $RE$ (using MINsim = 0.1). We then replace nouns in the text similar to the $RE$ with nouns extracted from the associated reference review set. Secondly, we modify NWN-STEM to work for verbs and adjectives, and call this GWN-STEM. From the reference review sets, we extract similar nouns, verbs, and adjectives to the $RE$ (using MINsim = 0.1), where the $RE$ is now not restricted to being a noun. We replace nouns, verbs, and adjectives in the text similar to the $RE$ with those extracted from the associated reference review set. Lastly, we implement W2V-STEM using Gensim BIBREF29. We train uni-gram Word2Vec models for single word $RE$s, and four-gram models for phrases. Models are trained on the training sets. We use cosine similarity to determine the most similar word/phrase in the input text to $RE$, which is the replaced $OE$. For all other words/phrases, we calculate $w_{i}^{\prime } = w_{i} - w_{OE} + w_{RE}$, where $w_{i}$ is the original word/phrase's embedding vector, $w_{OE}$ is the $OE$'s, $w_{RE}$ is the $RE$'s, and $w_{i}^{\prime }$ is the resulting embedding vector. The replacement word/phrase is $w_{i}^{\prime }$'s nearest neighbour. We use similarity thresholds to adjust replacement rates (RR) and produce text under various replacement rate thresholds (RRT). Evaluation ::: Evaluation Setup We manually select 10 nouns, 10 verbs, 10 adjectives, and 5 phrases from the top 10% most frequent words/phrases in each test set as our evaluation $RE$s. We filter the verbs and adjectives through a list of sentiment words BIBREF30 to ensure we do not choose $RE$s that would obviously significantly alter the text's sentiment. For each evaluation $RE$, we choose one-hundred lines from the corresponding test set that does not already contain $RE$. We choose lines with at least five words, as many with less carry little semantic meaning (e.g. `Great!', `It is okay'). For Amazon and Yelp, we choose 50 positive and 50 negative lines per $RE$. We repeat this process three times, resulting in three sets of 1000 lines per dataset per POS (excluding phrases), and three sets of 500 lines per dataset for phrases. Our final results are averaged metrics over these three sets. For SMERTI-Transformer, SMERTI-RNN, and W2V-STEM, we generate four outputs per text for MRT/RRT of 20%, 40%, 60%, and 80%, which represent upper-bounds on the percentage of the input that can be masked and/or replaced. Note that NWN-STEM and GWN-STEM can only evaluate on limited POS and their maximum replacement rates are limited. We select MINsim values of 0.075 and 0 for nouns and 0.1 and 0 for verbs, as these result in replacement rates approximately equal to the actual MR/RR of the other models' outputs for 20% and 40% MRT/RRT, respectively. Evaluation ::: Key Evaluation Metrics Fluency (SLOR) We use syntactic log-odds ratio (SLOR) BIBREF31 for sentence level fluency and modify from their word-level formula to character-level ($SLOR_{c}$). We use Flair perplexity values from a language model trained on the One Billion Words corpus BIBREF32: where $|S|$ and $|w|$ are the character lengths of the input text $S$ and the word $w$, respectively, $p_M(S)$ and $p_M(w)$ are the probabilities of $S$ and $w$ under the language model $M$, respectively, and $PPL_S$ and $PPL_w$ are the character-level perplexities of $S$ and $w$, respectively. SLOR (from hereon we refer to character-level SLOR as simply SLOR) measures aspects of text fluency such as grammaticality. Higher values represent higher fluency. We rescale resulting SLOR values to the interval [0,1] by first fitting and normalizing a Gaussian distribution. We then truncate normalized data points outside [-3,3], which shifts approximately 0.69% of total data. Finally, we divide each data point by six and add 0.5 to each result. Sentiment Preservation Accuracy (SPA) is defined as the percentage of outputs that carry the same sentiment as the input. We use VADER BIBREF33 to evaluate sentiment as positive, negative, or neutral. It handles typos, emojis, and other aspects of online text. Content Similarity Score (CSS) ranges from 0 to 1 and indicates the semantic similarity between generated text and the $RE$. A value closer to 1 indicates stronger semantic exchange, as the output is closer in semantic content to the $RE$. We also use the USE for this due to its design and strong performance as previously mentioned. Evaluation ::: Semantic Text Exchange Score (STES) We come up with a single score to evaluate overall performance of a model on STE that combines the key evaluation metrics. It uses the harmonic mean, similar to the F1 score (or F-score) BIBREF34, BIBREF35, and we call it the Semantic Text Exchange Score (STES): where $A$ is SPA, $B$ is SLOR, and $C$ is CSS. STES ranges between 0 and 1, with scores closer to 1 representing higher overall performance. Like the F1 score, STES penalizes models which perform very poorly in one or more metrics, and favors balanced models achieving strong results in all three. Evaluation ::: Automatic Evaluation Results Table TABREF38 shows overall average results by model. Table TABREF41 shows outputs for a Yelp example. As observed from Table TABREF41 (see also Appendix F), SMERTI is able to generate high quality output text similar to the $RE$ while flowing better than other models' outputs. It can replace entire phrases and sentences due to its variable length infilling. Note that for nouns, the outputs from GWN-STEM and NWN-STEM are equivalent. Evaluation ::: Human Evaluation Setup We conduct a human evaluation with eight participants, 6 males and 2 females, that are affiliated project researchers aged 20-39 at the University of Waterloo. We randomly choose one evaluation line for a randomly selected word or phrase for each POS per dataset. The input text and each model's output (for 40% MRT/RRT - chosen as a good middle ground) for each line is presented to participants, resulting in a total of 54 pieces of text, and rated on the following criteria from 1-5: RE Match: “How related is the entire text to the concept of [X]", where [X] is a word or phrase (1 - not at all related, 3 - somewhat related, 5 - very related). Note here that [X] is a given $RE$. Fluency: “Does the text make sense and flow well?" (1 - not at all, 3 - somewhat, 5 - very) Sentiment: “How do you think the author of the text was feeling?" (1 - very negative, 3 - neutral, 5 - very positive) Each participant evaluates every piece of text. They are presented with a single piece of text at a time, with the order of models, POS, and datasets completely randomized. Evaluation ::: Human Evaluation Results Average human evaluation scores are displayed in Table TABREF50. Sentiment Preservation (between 0 and 1) is calculated by comparing the average Sentiment rating for each model's output text to the Sentiment rating of the input text, and if both are less than 2.5 (negative), between 2.5 and 3.5 inclusive (neutral), or greater than 3.5 (positive), this is counted as a valid case of Sentiment Preservation. We repeat this for every evaluation line to calculate the final values per model. Harmonic means of all three metrics (using rescaled 0-1 values of RE Match and Fluency) are also displayed. Analysis ::: Performance by Model As seen in Table TABREF38, both SMERTI variations achieve higher STES and outperform the other models overall, with the WordNet models performing the worst. SMERTI excels especially on fluency and content similarity. The transformer variation achieves slightly higher SLOR, while the RNN variation achieves slightly higher CSS. The WordNet models perform strongest in sentiment preservation (SPA), likely because they modify little of the text and only verbs and nouns. They achieve by far the lowest CSS, likely in part due to this limited text replacement. They also do not account for context, and many words (e.g. proper nouns) do not exist in WordNet. Overall, the WordNet models are not very effective at STE. W2V-STEM achieves the lowest SLOR, especially for higher RRT, as supported by the example in Table TABREF41 (see also Appendix F). W2V-STEM and WordNet models output grammatically incorrect text that flows poorly. In many cases, words are repeated multiple times. We analyze the average Type Token Ratio (TTR) values of each model's outputs, which is the ratio of unique divided by total words. As shown in Table TABREF52, the SMERTI variations achieve the highest TTR, while W2V-STEM and NWN-STEM the lowest. Note that while W2V-STEM achieves lower CSS than SMERTI, it performs comparably in this aspect. This is likely due to its vector arithmetic operations algorithm, which replaces each word with one more similar to the RE. This is also supported by the lower TTR, as W2V-STEM frequently outputs the same words multiple times. Analysis ::: Performance By Model - Human Results As seen in Table TABREF50, the SMERTI variations outperform all baseline models overall, particularly in RE Match. SMERTI-Transformer performs the best, with SMERTI-RNN second. The WordNet models achieve high Sentiment Preservation, but much lower on RE Match. W2V-STEM achieves comparably high RE Match, but lowest Fluency. These results correspond well with our automatic evaluation results in Table TABREF38. We look at the Pearson correlation values between RE Match, Fluency, and Sentiment Preservation with CSS, SLOR, and SPA, respectively. These are 0.9952, 0.9327, and 0.8768, respectively, demonstrating that our automatic metrics are highly effective and correspond well with human ratings. Analysis ::: SMERTI's Performance By POS As seen from Table TABREF55 , SMERTI's SPA values are highest for nouns, likely because they typically carry little sentiment, and lowest for adjectives, likely because they typically carry the most. SLOR is lowest for adjectives and highest for phrases and nouns. Adjectives typically carry less semantic meaning and SMERTI likely has more trouble figuring out how best to infill the text. In contrast, nouns typically carry more, and phrases the most (since they consist of multiple words). SMERTI's CSS is highest for phrases then nouns, likely due to phrases and nouns carrying more semantic meaning, making it easier to generate semantically similar text. Both SMERTI's and the input text's CSS are lowest for adjectives, likely because they carry little semantic meaning. Overall, SMERTI appears to be more effective on nouns and phrases than verbs and adjectives. Analysis ::: SMERTI's Performance By Dataset As seen in Table TABREF58, SMERTI's SPA is lowest for news headlines. Amazon and Yelp reviews naturally carry stronger sentiment, likely making it easier to generate text with similar sentiment. Both SMERTI's and the input text's SLOR appear to be lower for Yelp reviews. This may be due to many reasons, such as more typos and emojis within the original reviews, and so forth. SMERTI's CSS values are slightly higher for news headlines. This may be due to them typically being shorter and carrying more semantic meaning as they are designed to be attention grabbers. Overall, it seems that using datasets which inherently carry more sentiment will lead to better sentiment preservation. Further, the quality of the dataset's original text, unsurprisingly, influences the ability of SMERTI to generate fluent text. Analysis ::: SMERTI's Performance By MRT/RRT From Table TABREF60, it can be seen that as MRT/RRT increases, SMERTI's SPA and SLOR decrease while CSS increases. These relationships are very strong as supported by the Pearson correlation values of -0.9972, -0.9183, and 0.9078, respectively. When SMERTI can alter more text, it has the opportunity to replace more related to sentiment while producing more of semantic similarity to the $RE$. Further, SMERTI generates more of the text itself, becoming less similar to the human-written input, resulting in lower fluency. To further demonstrate this, we look at average SMERTI BLEU BIBREF36 scores against MRT/RRT, shown in Table TABREF60. BLEU generally indicates how close two pieces of text are in content and structure, with higher values indicating greater similarity. We report our final BLEU scores as the average scores of 1 to 4-grams. As expected, BLEU decreases as MRT/RRT increases, and this relationship is very strong as supported by the Pearson correlation value of -0.9960. It is clear that MRT/RRT represents a trade-off between CSS against SPA and SLOR. It is thus an adjustable parameter that can be used to control the generated text, and balance semantic exchange against fluency and sentiment preservation. Conclusion and Future Work We introduced the task of semantic text exchange (STE), demonstrated that our pipeline SMERTI performs well on STE, and proposed an STES metric for evaluating overall STE performance. SMERTI outperformed other models and was the most balanced overall. We also showed a trade-off between semantic exchange against fluency and sentiment preservation, which can be controlled by the masking (replacement) rate threshold. Potential directions for future work include adding specific methods to control sentiment, and fine-tuning SMERTI for preservation of persona or personality. Experimenting with other text infilling models (e.g. fine-tuning BERT BIBREF8) is also an area of exploration. Lastly, our human evaluation is limited in size and a larger and more diverse participant pool is needed. We conclude by addressing potential ethical misuses of STE, including assisting in the generation of spam and fake-reviews/news. These risks come with any intelligent chatbot work, but we feel that the benefits, including usage in the detection of misuse such as fake-news, greatly outweigh the risks and help progress NLP and AI research. Acknowledgments We thank our anonymous reviewers, study participants, and Huawei Technologies Co., Ltd. for financial support.

No

3da9a861dfa25ed486cff0ef657d398fdebf8a93

3da9a861dfa25ed486cff0ef657d398fdebf8a93_0

Q: What are the baseline models mentioned in the paper? Text: Introduction There has been significant research on style transfer, with the goal of changing the style of text while preserving its semantic content. The alternative where semantics are adjusted while keeping style intact, which we call semantic text exchange (STE), has not been investigated to the best of our knowledge. Consider the following example, where the replacement entity defines the new semantic context: Original Text: It is sunny outside! Ugh, that means I must wear sunscreen. I hate being sweaty and sticky all over. Replacement Entity: weather = rainy Desired Text: It is rainy outside! Ugh, that means I must bring an umbrella. I hate being wet and having to carry it around. The weather within the original text is sunny, whereas the actual weather may be rainy. Not only is the word sunny replaced with rainy, but the rest of the text's content is changed while preserving its negative sentiment and fluency. With the rise of natural language processing (NLP) has come an increased demand for massive amounts of text data. Manually collecting and scraping data requires a significant amount of time and effort, and data augmentation techniques for NLP are limited compared to fields such as computer vision. STE can be used for text data augmentation by producing various modifications of a piece of text that differ in semantic content. Another use of STE is in building emotionally aligned chatbots and virtual assistants. This is useful for reasons such as marketing, overall enjoyment of interaction, and mental health therapy. However, due to limited data with emotional content in specific semantic contexts, the generated text may contain incorrect semantic content. STE can adjust text semantics (e.g. to align with reality or a specific task) while preserving emotions. One specific example is the development of virtual assistants with adjustable socio-emotional personalities in the effort to construct assistive technologies for persons with cognitive disabilities. Adjusting the emotional delivery of text in subtle ways can have a strong effect on the adoption of the technologies BIBREF0. It is challenging to transfer style this subtly due to lack of datasets on specific topics with consistent emotions. Instead, large datasets of emotionally consistent interactions not confined to specific topics exist. Hence, it is effective to generate text with a particular emotion and then adjust its semantics. We propose a pipeline called SMERTI (pronounced `smarty') for STE. Combining entity replacement (ER), similarity masking (SM), and text infilling (TI), SMERTI can modify the semantic content of text. We define a metric called the Semantic Text Exchange Score (STES) that evaluates the overall ability of a model to perform STE, and an adjustable parameter masking (replacement) rate threshold (MRT/RRT) that can be used to control the amount of semantic change. We evaluate on three datasets: Yelp and Amazon reviews BIBREF1, and Kaggle news headlines BIBREF2. We implement three baseline models for comparison: Noun WordNet Semantic Text Exchange Model (NWN-STEM), General WordNet Semantic Text Exchange Model (GWN-STEM), and Word2Vec Semantic Text Exchange Model (W2V-STEM). We illustrate the STE performance of two SMERTI variations on the datasets, demonstrating outperformance of the baselines and pipeline stability. We also run a human evaluation supporting our results. We analyze the results in detail and investigate relationships between the semantic change, fluency, sentiment, and MRT/RRT. Our major contributions can be summarized as: We define a new task called semantic text exchange (STE) with increasing importance in NLP applications that modifies text semantics while preserving other aspects such as sentiment. We propose a pipeline SMERTI capable of multi-word entity replacement and text infilling, and demonstrate its outperformance of baselines. We define an evaluation metric for overall performance on semantic text exchange called the Semantic Text Exchange Score (STES). Related Work ::: Word and Sentence-level Embeddings Word2Vec BIBREF3, BIBREF4 allows for analogy representation through vector arithmetic. We implement a baseline (W2V-STEM) using this technique. The Universal Sentence Encoder (USE) BIBREF5 encodes sentences and is trained on a variety of web sources and the Stanford Natural Language Inference corpus BIBREF6. Flair embeddings BIBREF7 are based on architectures such as BERT BIBREF8. We use USE for SMERTI as it is designed for transfer learning and shows higher performance on textual similarity tasks compared to other models BIBREF9. Related Work ::: Text Infilling Text infilling is the task of filling in missing parts of sentences called masks. MaskGAN BIBREF10 is restricted to a single word per mask token, while SMERTI is capable of variable length infilling for more flexible output. BIBREF11 uses a transformer-based architecture. They fill in random masks, while SMERTI fills in masks guided by semantic similarity, resulting in more natural infilling and fulfillment of the STE task. Related Work ::: Style and Sentiment Transfer Notable works in style/sentiment transfer include BIBREF12, BIBREF13, BIBREF14, BIBREF15. They attempt to learn latent representations of various text aspects such as its context and attributes, or separate style from content and encode them into hidden representations. They then use an RNN decoder to generate a new sentence given a targeted sentiment attribute. Related Work ::: Review Generation BIBREF16 generates fake reviews from scratch using language models. BIBREF17, BIBREF18, BIBREF19 generate reviews from scratch given auxiliary information (e.g. the item category and star rating). BIBREF20 generates reviews using RNNs with two components: generation from scratch and review customization (Algorithm 2 in BIBREF20). They define review customization as modifying the generated review to fit a new topic or context, such as from a Japanese restaurant to an Italian one. They condition on a keyword identifying the desired context, and replace similar nouns with others using WordNet BIBREF21. They require a “reference dataset" (required to be “on topic"; easy enough for restaurant reviews, but less so for arbitrary conversational agents). As noted by BIBREF19, the method of BIBREF20 may also replace words independently of context. We implement their review customization algorithm (NWN-STEM) and a modified version (GWN-STEM) as baseline models. SMERTI ::: Overview The task is to transform a corpus $C$ of lines of text $S_i$ and associated replacement entities $RE_i:C = \lbrace (S_1,RE_1),(S_2,RE_2),\ldots , (S_n, RE_n)\rbrace $ to a modified corpus $\hat{C} = \lbrace \hat{S}_1,\hat{S}_2,\ldots ,\hat{S}_n\rbrace $, where $\hat{S}_i$ are the original text lines $S_i$ replaced with $RE_i$ and overall semantics adjusted. SMERTI consists of the following modules, shown in Figure FIGREF15: Entity Replacement Module (ERM): Identify which word(s) within the original text are best replaced with the $RE$, which we call the Original Entity ($OE$). We replace $OE$ in $S$ with $RE$. We call this modified text $S^{\prime }$. Similarity Masking Module (SMM): Identify words/phrases in $S^{\prime }$ similar to $OE$ and replace them with a [mask]. Group adjacent [mask]s into a single one so we can fill a variable length of text into each. We call this masked text $S^{\prime \prime }$. Text Infilling Module (TIM): Fill in [mask] tokens with text that better suits the $RE$. This will modify semantics in the rest of the text. This final output text is called $\hat{S}$. SMERTI ::: Entity Replacement Module (ERM) For entity replacement, we use a combination of the Universal Sentence Encoder BIBREF5 and Stanford Parser BIBREF22. SMERTI ::: Entity Replacement Module (ERM) ::: Stanford Parser The Stanford Parser is a constituency parser that determines the grammatical structure of sentences, including phrases and part-of-speech (POS) labelling. By feeding our $RE$ through the parser, we are able to determine its parse-tree. Iterating through the parse-tree and its sub-trees, we can obtain a list of constituent tags for the $RE$. We then feed our input text $S$ through the parser, and through a similar process, we can obtain a list of leaves (where leaves under a single label are concatenated) that are equal or similar to any of the $RE$ constituent tags. This generates a list of entities having the same (or similar) grammatical structure as the $RE$, and are likely candidates for the $OE$. We then feed these entities along with the $RE$ into the Universal Sentence Encoder (USE). SMERTI ::: Entity Replacement Module (ERM) ::: Universal Sentence Encoder (USE) The USE is a sentence-level embedding model that comes with a deep averaging network (DAN) and transformer model BIBREF5. We choose the transformer model as these embeddings take context into account, and the exact same word/phrase will have a different embedding depending on its context and surrounding words. We compute the semantic similarity between two embeddings $u$ and $v$: $sim(u,v)$, using the angular (cosine) distance, defined as: $\cos (\theta _{u,v}) = (u\cdot v)/(||u|| ||v||)$, such that $sim(u,v) = 1-\frac{1}{\pi }arccos(\cos (\theta _{u,v}))$. Results are in $[0,1]$, with higher values representing greater similarity. Using USE and the above equation, we can identify words/phrases within the input text $S$ which are most similar to $RE$. To assist with this, we use the Stanford Parser as described above to obtain a list of candidate entities. In the rare case that this list is empty, we feed in each word of $S$ into USE, and identify which word is the most similar to $RE$. We then replace the most similar entity or word ($OE$) with the $RE$ and generate $S^{\prime }$. An example of this entity replacement process is in Figure FIGREF18. Two parse-trees are shown: for $RE$ (a) and $S$ (b) and (c). Figure FIGREF18(d) is a semantic similarity heat-map generated from the USE embeddings of the candidate $OE$s and $RE$, where values are similarity scores in the range $[0,1]$. As seen in Figure FIGREF18(d), we calculate semantic similarities between $RE$ and entities within $S$ which have noun constituency tags. Looking at the row for our $RE$ restaurant, the most similar entity (excluding itself) is hotel. We can then generate: $S^{\prime }$ = i love this restaurant ! the beds are comfortable and the service is great ! SMERTI ::: Similarity Masking Module (SMM) Next, we mask words similar to $OE$ to generate $S^{\prime \prime }$ using USE. We look at semantic similarities between every word in $S$ and $OE$, along with semantic similarities between $OE$ and the candidate entities determined in the previous ERM step to broaden the range of phrases our module can mask. We ignore $RE$, $OE$, and any entities or phrases containing $OE$ (for example, `this hotel'). After determining words similar to the $OE$ (discussed below), we replace each of them with a [mask] token. Next, we replace [mask] tokens adjacent to each other with a single [mask]. We set a base similarity threshold (ST) that selects a subset of words to mask. We compare the actual fraction of masked words to the masking rate threshold (MRT), as defined by the user, and increase ST in intervals of $0.05$ until the actual masking rate falls below the MRT. Some sample masked outputs ($S^{\prime \prime }$) using various MRT-ST combinations for the previous example are shown in Table TABREF21 (more examples in Appendix A). The MRT is similar to the temperature parameter used to control the “novelty” of generated text in works such as BIBREF20. A high MRT means the user wants to generate text very semantically dissimilar to the original, and may be desired in cases such as creating a lively chatbot or correcting text that is heavily incorrect semantically. A low MRT means the user wants to generate text semantically similar to the original, and may be desired in cases such as text recovery, grammar correction, or correcting a minor semantic error in text. By varying the MRT, various pieces of text that differ semantically in subtle ways can be generated, assisting greatly with text data augmentation. The MRT also affects sentiment and fluency, as we show in Section SECREF59. SMERTI ::: Text Infilling Module (TIM) We use two seq2seq models for our TIM: an RNN (recurrent neural network) model BIBREF23 (called SMERTI-RNN), and a transformer model (called SMERTI-Transformer). SMERTI ::: Text Infilling Module (TIM) ::: Bidirectional RNN with Attention We use a bidirectional variant of the GRU BIBREF24, and hence two RNNs for the encoder: one reads the input sequence in standard sequential order, and the other is fed this sequence in reverse. The outputs are summed at each time step, giving us the ability to encode information from both past and future context. The decoder generates the output in a sequential token-by-token manner. To combat information loss, we implement the attention mechanism BIBREF25. We use a Luong attention layer BIBREF26 which uses global attention, where all the encoder's hidden states are considered, and use the decoder's current time-step hidden state to calculate attention weights. We use the dot score function for attention, where $h_t$ is the current target decoder state and $\bar{h}_s$ is all encoder states: $score(h_t,\bar{h}_s)=h_t^T\bar{h}_s$. SMERTI ::: Text Infilling Module (TIM) ::: Transformer Our second model makes use of the transformer architecture, and our implementation replicates BIBREF27. We use an encoder-decoder structure with a multi-head self-attention token decoder to condition on information from both past and future context. It maps a query and set of key-value pairs to an output. The queries and keys are of dimension $d_k$, and values of dimension $d_v$. To compute the attention, we pack a set of queries, keys, and values into matrices $Q$, $K$, and $V$, respectively. The matrix of outputs is computed as: Multi-head attention allows the model to jointly attend to information from different positions. The decoder can make use of both local and global semantic information while filling in each [mask]. Experiment ::: Datasets We train our two TIMs on the three datasets. The Amazon dataset BIBREF1 contains over 83 million user reviews on products, with duplicate reviews removed. The Yelp dataset includes over six million user reviews on businesses. The news headlines dataset from Kaggle contains approximately $200,000$ news headlines from 2012 to 2018 obtained from HuffPost BIBREF2. We filter the text to obtain reviews and headlines which are English, do not contain hyperlinks and other obvious noise, and are less than 20 words long. We found that many longer than twenty words ramble on and are too verbose for our purposes. Rather than filtering by individual sentences we keep each text in its entirety so SMERTI can learn to generate multiple sentences at once. We preprocess the text by lowercasing and removing rare/duplicate punctuation and space. For Amazon and Yelp, we treat reviews greater than three stars as containing positive sentiment, equal to three stars as neutral, and less than three stars as negative. For each training and testing set, we include an equal number of randomly selected positive and negative reviews, and half as many neutral reviews. This is because neutral reviews only occupy one out of five stars compared to positive and negative which occupy two each. Our dataset statistics can be found in Appendix B. Experiment ::: Experiment Details To set up our training and testing data for text infilling, we mask the text. We use a tiered masking approach: for each dataset, we randomly mask 15% of the words in one-third of the lines, 30% of the words in another one-third, and 45% in the remaining one-third. These masked texts serve as the inputs, while the original texts serve as the ground-truth. This allows our TIM models to learn relationships between masked words and relationships between masked and unmasked words. The bidirectional RNN decoder fills in blanks one by one, with the objective of minimizing the cross entropy loss between its output and the ground-truth. We use a hidden size of 500, two layers for the encoder and decoder, teacher-forcing ratio of 1.0, learning rate of 0.0001, dropout of 0.1, batch size of 64, and train for up to 40 epochs. For the transformer, we use scaled dot-product attention and the same hyperparameters as BIBREF27. We use the Adam optimizer BIBREF28 with $\beta _1 = 0.9, \beta _2 = 0.98$, and $\epsilon = 10^{-9}$. As in BIBREF27, we increase the $learning\_rate$ linearly for the first $warmup\_steps$ training steps, and then decrease the $learning\_rate$ proportionally to the inverse square root of the step number. We set $factor=1$ and use $warmup\_steps = 2000$. We use a batch size of 4096, and we train for up to 40 epochs. Experiment ::: Baseline Models We implement three models to benchmark against. First is NWN-STEM (Algorithm 2 from BIBREF20). We use the training sets as the “reference review sets" to extract similar nouns to the $RE$ (using MINsim = 0.1). We then replace nouns in the text similar to the $RE$ with nouns extracted from the associated reference review set. Secondly, we modify NWN-STEM to work for verbs and adjectives, and call this GWN-STEM. From the reference review sets, we extract similar nouns, verbs, and adjectives to the $RE$ (using MINsim = 0.1), where the $RE$ is now not restricted to being a noun. We replace nouns, verbs, and adjectives in the text similar to the $RE$ with those extracted from the associated reference review set. Lastly, we implement W2V-STEM using Gensim BIBREF29. We train uni-gram Word2Vec models for single word $RE$s, and four-gram models for phrases. Models are trained on the training sets. We use cosine similarity to determine the most similar word/phrase in the input text to $RE$, which is the replaced $OE$. For all other words/phrases, we calculate $w_{i}^{\prime } = w_{i} - w_{OE} + w_{RE}$, where $w_{i}$ is the original word/phrase's embedding vector, $w_{OE}$ is the $OE$'s, $w_{RE}$ is the $RE$'s, and $w_{i}^{\prime }$ is the resulting embedding vector. The replacement word/phrase is $w_{i}^{\prime }$'s nearest neighbour. We use similarity thresholds to adjust replacement rates (RR) and produce text under various replacement rate thresholds (RRT). Evaluation ::: Evaluation Setup We manually select 10 nouns, 10 verbs, 10 adjectives, and 5 phrases from the top 10% most frequent words/phrases in each test set as our evaluation $RE$s. We filter the verbs and adjectives through a list of sentiment words BIBREF30 to ensure we do not choose $RE$s that would obviously significantly alter the text's sentiment. For each evaluation $RE$, we choose one-hundred lines from the corresponding test set that does not already contain $RE$. We choose lines with at least five words, as many with less carry little semantic meaning (e.g. `Great!', `It is okay'). For Amazon and Yelp, we choose 50 positive and 50 negative lines per $RE$. We repeat this process three times, resulting in three sets of 1000 lines per dataset per POS (excluding phrases), and three sets of 500 lines per dataset for phrases. Our final results are averaged metrics over these three sets. For SMERTI-Transformer, SMERTI-RNN, and W2V-STEM, we generate four outputs per text for MRT/RRT of 20%, 40%, 60%, and 80%, which represent upper-bounds on the percentage of the input that can be masked and/or replaced. Note that NWN-STEM and GWN-STEM can only evaluate on limited POS and their maximum replacement rates are limited. We select MINsim values of 0.075 and 0 for nouns and 0.1 and 0 for verbs, as these result in replacement rates approximately equal to the actual MR/RR of the other models' outputs for 20% and 40% MRT/RRT, respectively. Evaluation ::: Key Evaluation Metrics Fluency (SLOR) We use syntactic log-odds ratio (SLOR) BIBREF31 for sentence level fluency and modify from their word-level formula to character-level ($SLOR_{c}$). We use Flair perplexity values from a language model trained on the One Billion Words corpus BIBREF32: where $|S|$ and $|w|$ are the character lengths of the input text $S$ and the word $w$, respectively, $p_M(S)$ and $p_M(w)$ are the probabilities of $S$ and $w$ under the language model $M$, respectively, and $PPL_S$ and $PPL_w$ are the character-level perplexities of $S$ and $w$, respectively. SLOR (from hereon we refer to character-level SLOR as simply SLOR) measures aspects of text fluency such as grammaticality. Higher values represent higher fluency. We rescale resulting SLOR values to the interval [0,1] by first fitting and normalizing a Gaussian distribution. We then truncate normalized data points outside [-3,3], which shifts approximately 0.69% of total data. Finally, we divide each data point by six and add 0.5 to each result. Sentiment Preservation Accuracy (SPA) is defined as the percentage of outputs that carry the same sentiment as the input. We use VADER BIBREF33 to evaluate sentiment as positive, negative, or neutral. It handles typos, emojis, and other aspects of online text. Content Similarity Score (CSS) ranges from 0 to 1 and indicates the semantic similarity between generated text and the $RE$. A value closer to 1 indicates stronger semantic exchange, as the output is closer in semantic content to the $RE$. We also use the USE for this due to its design and strong performance as previously mentioned. Evaluation ::: Semantic Text Exchange Score (STES) We come up with a single score to evaluate overall performance of a model on STE that combines the key evaluation metrics. It uses the harmonic mean, similar to the F1 score (or F-score) BIBREF34, BIBREF35, and we call it the Semantic Text Exchange Score (STES): where $A$ is SPA, $B$ is SLOR, and $C$ is CSS. STES ranges between 0 and 1, with scores closer to 1 representing higher overall performance. Like the F1 score, STES penalizes models which perform very poorly in one or more metrics, and favors balanced models achieving strong results in all three. Evaluation ::: Automatic Evaluation Results Table TABREF38 shows overall average results by model. Table TABREF41 shows outputs for a Yelp example. As observed from Table TABREF41 (see also Appendix F), SMERTI is able to generate high quality output text similar to the $RE$ while flowing better than other models' outputs. It can replace entire phrases and sentences due to its variable length infilling. Note that for nouns, the outputs from GWN-STEM and NWN-STEM are equivalent. Evaluation ::: Human Evaluation Setup We conduct a human evaluation with eight participants, 6 males and 2 females, that are affiliated project researchers aged 20-39 at the University of Waterloo. We randomly choose one evaluation line for a randomly selected word or phrase for each POS per dataset. The input text and each model's output (for 40% MRT/RRT - chosen as a good middle ground) for each line is presented to participants, resulting in a total of 54 pieces of text, and rated on the following criteria from 1-5: RE Match: “How related is the entire text to the concept of [X]", where [X] is a word or phrase (1 - not at all related, 3 - somewhat related, 5 - very related). Note here that [X] is a given $RE$. Fluency: “Does the text make sense and flow well?" (1 - not at all, 3 - somewhat, 5 - very) Sentiment: “How do you think the author of the text was feeling?" (1 - very negative, 3 - neutral, 5 - very positive) Each participant evaluates every piece of text. They are presented with a single piece of text at a time, with the order of models, POS, and datasets completely randomized. Evaluation ::: Human Evaluation Results Average human evaluation scores are displayed in Table TABREF50. Sentiment Preservation (between 0 and 1) is calculated by comparing the average Sentiment rating for each model's output text to the Sentiment rating of the input text, and if both are less than 2.5 (negative), between 2.5 and 3.5 inclusive (neutral), or greater than 3.5 (positive), this is counted as a valid case of Sentiment Preservation. We repeat this for every evaluation line to calculate the final values per model. Harmonic means of all three metrics (using rescaled 0-1 values of RE Match and Fluency) are also displayed. Analysis ::: Performance by Model As seen in Table TABREF38, both SMERTI variations achieve higher STES and outperform the other models overall, with the WordNet models performing the worst. SMERTI excels especially on fluency and content similarity. The transformer variation achieves slightly higher SLOR, while the RNN variation achieves slightly higher CSS. The WordNet models perform strongest in sentiment preservation (SPA), likely because they modify little of the text and only verbs and nouns. They achieve by far the lowest CSS, likely in part due to this limited text replacement. They also do not account for context, and many words (e.g. proper nouns) do not exist in WordNet. Overall, the WordNet models are not very effective at STE. W2V-STEM achieves the lowest SLOR, especially for higher RRT, as supported by the example in Table TABREF41 (see also Appendix F). W2V-STEM and WordNet models output grammatically incorrect text that flows poorly. In many cases, words are repeated multiple times. We analyze the average Type Token Ratio (TTR) values of each model's outputs, which is the ratio of unique divided by total words. As shown in Table TABREF52, the SMERTI variations achieve the highest TTR, while W2V-STEM and NWN-STEM the lowest. Note that while W2V-STEM achieves lower CSS than SMERTI, it performs comparably in this aspect. This is likely due to its vector arithmetic operations algorithm, which replaces each word with one more similar to the RE. This is also supported by the lower TTR, as W2V-STEM frequently outputs the same words multiple times. Analysis ::: Performance By Model - Human Results As seen in Table TABREF50, the SMERTI variations outperform all baseline models overall, particularly in RE Match. SMERTI-Transformer performs the best, with SMERTI-RNN second. The WordNet models achieve high Sentiment Preservation, but much lower on RE Match. W2V-STEM achieves comparably high RE Match, but lowest Fluency. These results correspond well with our automatic evaluation results in Table TABREF38. We look at the Pearson correlation values between RE Match, Fluency, and Sentiment Preservation with CSS, SLOR, and SPA, respectively. These are 0.9952, 0.9327, and 0.8768, respectively, demonstrating that our automatic metrics are highly effective and correspond well with human ratings. Analysis ::: SMERTI's Performance By POS As seen from Table TABREF55 , SMERTI's SPA values are highest for nouns, likely because they typically carry little sentiment, and lowest for adjectives, likely because they typically carry the most. SLOR is lowest for adjectives and highest for phrases and nouns. Adjectives typically carry less semantic meaning and SMERTI likely has more trouble figuring out how best to infill the text. In contrast, nouns typically carry more, and phrases the most (since they consist of multiple words). SMERTI's CSS is highest for phrases then nouns, likely due to phrases and nouns carrying more semantic meaning, making it easier to generate semantically similar text. Both SMERTI's and the input text's CSS are lowest for adjectives, likely because they carry little semantic meaning. Overall, SMERTI appears to be more effective on nouns and phrases than verbs and adjectives. Analysis ::: SMERTI's Performance By Dataset As seen in Table TABREF58, SMERTI's SPA is lowest for news headlines. Amazon and Yelp reviews naturally carry stronger sentiment, likely making it easier to generate text with similar sentiment. Both SMERTI's and the input text's SLOR appear to be lower for Yelp reviews. This may be due to many reasons, such as more typos and emojis within the original reviews, and so forth. SMERTI's CSS values are slightly higher for news headlines. This may be due to them typically being shorter and carrying more semantic meaning as they are designed to be attention grabbers. Overall, it seems that using datasets which inherently carry more sentiment will lead to better sentiment preservation. Further, the quality of the dataset's original text, unsurprisingly, influences the ability of SMERTI to generate fluent text. Analysis ::: SMERTI's Performance By MRT/RRT From Table TABREF60, it can be seen that as MRT/RRT increases, SMERTI's SPA and SLOR decrease while CSS increases. These relationships are very strong as supported by the Pearson correlation values of -0.9972, -0.9183, and 0.9078, respectively. When SMERTI can alter more text, it has the opportunity to replace more related to sentiment while producing more of semantic similarity to the $RE$. Further, SMERTI generates more of the text itself, becoming less similar to the human-written input, resulting in lower fluency. To further demonstrate this, we look at average SMERTI BLEU BIBREF36 scores against MRT/RRT, shown in Table TABREF60. BLEU generally indicates how close two pieces of text are in content and structure, with higher values indicating greater similarity. We report our final BLEU scores as the average scores of 1 to 4-grams. As expected, BLEU decreases as MRT/RRT increases, and this relationship is very strong as supported by the Pearson correlation value of -0.9960. It is clear that MRT/RRT represents a trade-off between CSS against SPA and SLOR. It is thus an adjustable parameter that can be used to control the generated text, and balance semantic exchange against fluency and sentiment preservation. Conclusion and Future Work We introduced the task of semantic text exchange (STE), demonstrated that our pipeline SMERTI performs well on STE, and proposed an STES metric for evaluating overall STE performance. SMERTI outperformed other models and was the most balanced overall. We also showed a trade-off between semantic exchange against fluency and sentiment preservation, which can be controlled by the masking (replacement) rate threshold. Potential directions for future work include adding specific methods to control sentiment, and fine-tuning SMERTI for preservation of persona or personality. Experimenting with other text infilling models (e.g. fine-tuning BERT BIBREF8) is also an area of exploration. Lastly, our human evaluation is limited in size and a larger and more diverse participant pool is needed. We conclude by addressing potential ethical misuses of STE, including assisting in the generation of spam and fake-reviews/news. These risks come with any intelligent chatbot work, but we feel that the benefits, including usage in the detection of misuse such as fake-news, greatly outweigh the risks and help progress NLP and AI research. Acknowledgments We thank our anonymous reviewers, study participants, and Huawei Technologies Co., Ltd. for financial support.

Noun WordNet Semantic Text Exchange Model (NWN-STEM), General WordNet Semantic Text Exchange Model (GWN-STEM), Word2Vec Semantic Text Exchange Model (W2V-STEM)

8c0a0747a970f6ea607ff9b18cfeb738502d9a95

8c0a0747a970f6ea607ff9b18cfeb738502d9a95_0

Q: What was the performance of both approaches on their dataset? Text: Introduction Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1. Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions. The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions. To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection. In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features: CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons. CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging. CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check. The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper. The CN-Celeb dataset ::: Data description The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check. In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances. The CN-Celeb dataset ::: Challenges with CN-Celeb Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows. Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh. A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres. Most of speakers have different genres of utterances, which results in significant variation in speaking styles. The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems. Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision. The CN-Celeb dataset ::: Collection pipeline CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process. Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision. The detailed steps of the collection process are summarized as follows. STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered. STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres. STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams. STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system. STEP 5. Double check by speaker recognition. Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely. STEP 6. Human check. The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours. Experiments on speaker recognition In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets. Experiments on speaker recognition ::: Data VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW). CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set. Experiments on speaker recognition ::: Settings Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10. For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4. For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details. Experiments on speaker recognition ::: Basic results We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E). The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW). Experiments on speaker recognition ::: Further comparison To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training. The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results. Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb. Conclusions We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought.

ERR of 19.05 with i-vectors and 15.52 with x-vectors

529dabe7b4a8a01b20ee099701834b60fb0c43b0

529dabe7b4a8a01b20ee099701834b60fb0c43b0_0

Q: What kind of settings do the utterances come from? Text: Introduction Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1. Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions. The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions. To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection. In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features: CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons. CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging. CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check. The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper. The CN-Celeb dataset ::: Data description The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check. In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances. The CN-Celeb dataset ::: Challenges with CN-Celeb Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows. Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh. A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres. Most of speakers have different genres of utterances, which results in significant variation in speaking styles. The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems. Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision. The CN-Celeb dataset ::: Collection pipeline CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process. Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision. The detailed steps of the collection process are summarized as follows. STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered. STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres. STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams. STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system. STEP 5. Double check by speaker recognition. Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely. STEP 6. Human check. The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours. Experiments on speaker recognition In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets. Experiments on speaker recognition ::: Data VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW). CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set. Experiments on speaker recognition ::: Settings Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10. For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4. For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details. Experiments on speaker recognition ::: Basic results We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E). The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW). Experiments on speaker recognition ::: Further comparison To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training. The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results. Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb. Conclusions We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought.

entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement

a2be2bd84e5ae85de2ab9968147b3d49c84dfb7f

a2be2bd84e5ae85de2ab9968147b3d49c84dfb7f_0

Q: What genres are covered? Text: Introduction Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1. Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions. The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions. To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection. In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features: CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons. CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging. CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check. The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper. The CN-Celeb dataset ::: Data description The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check. In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances. The CN-Celeb dataset ::: Challenges with CN-Celeb Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows. Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh. A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres. Most of speakers have different genres of utterances, which results in significant variation in speaking styles. The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems. Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision. The CN-Celeb dataset ::: Collection pipeline CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process. Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision. The detailed steps of the collection process are summarized as follows. STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered. STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres. STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams. STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system. STEP 5. Double check by speaker recognition. Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely. STEP 6. Human check. The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours. Experiments on speaker recognition In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets. Experiments on speaker recognition ::: Data VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW). CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set. Experiments on speaker recognition ::: Settings Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10. For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4. For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details. Experiments on speaker recognition ::: Basic results We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E). The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW). Experiments on speaker recognition ::: Further comparison To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training. The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results. Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb. Conclusions We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought.

genre, entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement

5699996a7a2bb62c68c1e62e730cabf1e3186eef

5699996a7a2bb62c68c1e62e730cabf1e3186eef_0

Q: Do they experiment with cross-genre setups? Text: Introduction Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1. Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions. The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions. To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection. In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features: CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons. CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging. CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check. The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper. The CN-Celeb dataset ::: Data description The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check. In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances. The CN-Celeb dataset ::: Challenges with CN-Celeb Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows. Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh. A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres. Most of speakers have different genres of utterances, which results in significant variation in speaking styles. The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems. Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision. The CN-Celeb dataset ::: Collection pipeline CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process. Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision. The detailed steps of the collection process are summarized as follows. STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered. STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres. STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams. STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system. STEP 5. Double check by speaker recognition. Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely. STEP 6. Human check. The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours. Experiments on speaker recognition In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets. Experiments on speaker recognition ::: Data VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW). CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set. Experiments on speaker recognition ::: Settings Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10. For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4. For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details. Experiments on speaker recognition ::: Basic results We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E). The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW). Experiments on speaker recognition ::: Further comparison To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training. The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results. Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb. Conclusions We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought.

No

944d5dbe0cfc64bf41ea36c11b1d378c408d40b8

944d5dbe0cfc64bf41ea36c11b1d378c408d40b8_0

Q: Which of the two speech recognition models works better overall on CN-Celeb? Text: Introduction Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1. Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions. The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions. To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection. In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features: CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons. CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging. CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check. The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper. The CN-Celeb dataset ::: Data description The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check. In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances. The CN-Celeb dataset ::: Challenges with CN-Celeb Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows. Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh. A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres. Most of speakers have different genres of utterances, which results in significant variation in speaking styles. The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems. Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision. The CN-Celeb dataset ::: Collection pipeline CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process. Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision. The detailed steps of the collection process are summarized as follows. STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered. STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres. STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams. STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system. STEP 5. Double check by speaker recognition. Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely. STEP 6. Human check. The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours. Experiments on speaker recognition In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets. Experiments on speaker recognition ::: Data VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW). CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set. Experiments on speaker recognition ::: Settings Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10. For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4. For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details. Experiments on speaker recognition ::: Basic results We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E). The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW). Experiments on speaker recognition ::: Further comparison To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training. The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results. Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb. Conclusions We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought.

x-vector

327e6c6609fbd4c6ae76284ca639951f03eb4a4c

327e6c6609fbd4c6ae76284ca639951f03eb4a4c_0

Q: By how much is performance on CN-Celeb inferior to performance on VoxCeleb? Text: Introduction Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1. Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions. The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions. To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection. In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features: CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons. CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging. CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check. The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper. The CN-Celeb dataset ::: Data description The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check. In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances. The CN-Celeb dataset ::: Challenges with CN-Celeb Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows. Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh. A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres. Most of speakers have different genres of utterances, which results in significant variation in speaking styles. The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems. Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision. The CN-Celeb dataset ::: Collection pipeline CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process. Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision. The detailed steps of the collection process are summarized as follows. STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered. STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres. STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams. STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system. STEP 5. Double check by speaker recognition. Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely. STEP 6. Human check. The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours. Experiments on speaker recognition In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets. Experiments on speaker recognition ::: Data VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW). CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set. Experiments on speaker recognition ::: Settings Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10. For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4. For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details. Experiments on speaker recognition ::: Basic results We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E). The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW). Experiments on speaker recognition ::: Further comparison To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training. The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results. Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb. Conclusions We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought.

For i-vector system, performances are 11.75% inferior to voxceleb. For x-vector system, performances are 10.74% inferior to voxceleb

df8cc1f395486a12db98df805248eb37c087458b

df8cc1f395486a12db98df805248eb37c087458b_0

Q: On what datasets is the new model evaluated on? Text: Introduction Deep neural network-based models are easy to overfit and result in losing their generalization due to limited size of training data. In order to address the issue, data augmentation methods are often applied to generate more training samples. Recent years have witnessed great success in applying data augmentation in the field of speech area BIBREF0 , BIBREF1 and computer vision BIBREF2 , BIBREF3 , BIBREF4 . Data augmentation in these areas can be easily performed by transformations like resizing, mirroring, random cropping, and color shifting. However, applying these universal transformations to texts is largely randomized and uncontrollable, which makes it impossible to ensure the semantic invariance and label correctness. For example, given a movie review “The actors is good", by mirroring we get “doog si srotca ehT", or by random cropping we get “actors is", both of which are meaningless. Existing data augmentation methods for text are often loss of generality, which are developed with handcrafted rules or pipelines for specific domains. A general approach for text data augmentation is replacement-based method, which generates new sentences by replacing the words in the sentences with relevant words (e.g. synonyms). However, words with synonyms from a handcrafted lexical database likes WordNet BIBREF5 are very limited , and the replacement-based augmentation with synonyms can only produce limited diverse patterns from the original texts. To address the limitation of replacement-based methods, Kobayashi BIBREF6 proposed contextual augmentation for labeled sentences by offering a wide range of substitute words, which are predicted by a label-conditional bidirectional language model according to the context. But contextual augmentation suffers from two shortages: the bidirectional language model is simply shallow concatenation of a forward and backward model, and the usage of LSTM models restricts their prediction ability to a short range. BERT, which stands for Bidirectional Encoder Representations from Transformers, pre-trained deep bidirectional representations by jointly conditioning on both left and right context in all layers. BERT addressed the unidirectional constraint by proposing a “masked language model" (MLM) objective by masking some percentage of the input tokens at random, and predicting the masked words based on its context. This is very similar to how contextual augmentation predict the replacement words. But BERT was proposed to pre-train text representations, so MLM task is performed in an unsupervised way, taking no label variance into consideration. This paper focuses on the replacement-based methods, by proposing a novel data augmentation method called conditional BERT contextual augmentation. The method applies contextual augmentation by conditional BERT, which is fine-tuned on BERT. We adopt BERT as our pre-trained language model with two reasons. First, BERT is based on Transformer. Transformer provides us with a more structured memory for handling long-term dependencies in text. Second, BERT, as a deep bidirectional model, is strictly more powerful than the shallow concatenation of a left-to-right and right-to left model. So we apply BERT to contextual augmentation for labeled sentences, by offering a wider range of substitute words predicted by the masked language model task. However, the masked language model predicts the masked word based only on its context, so the predicted word maybe incompatible with the annotated labels of the original sentences. In order to address this issue, we introduce a new fine-tuning objective: the "conditional masked language model"(C-MLM). The conditional masked language model randomly masks some of the tokens from an input, and the objective is to predict a label-compatible word based on both its context and sentence label. Unlike Kobayashi's work, the C-MLM objective allows a deep bidirectional representations by jointly conditioning on both left and right context in all layers. In order to evaluate how well our augmentation method improves performance of deep neural network models, following Kobayashi BIBREF6 , we experiment it on two most common neural network structures, LSTM-RNN and CNN, on text classification tasks. Through the experiments on six various different text classification tasks, we demonstrate that the proposed conditional BERT model augments sentence better than baselines, and conditional BERT contextual augmentation method can be easily applied to both convolutional or recurrent neural networks classifier. We further explore our conditional MLM task’s connection with style transfer task and demonstrate that our conditional BERT can also be applied to style transfer too. Our contributions are concluded as follows: To our best knowledge, this is the first attempt to alter BERT to a conditional BERT or apply BERT on text generation tasks. Fine-tuning on Pre-trained Language Model Language model pre-training has attracted wide attention and fine-tuning on pre-trained language model has shown to be effective for improving many downstream natural language processing tasks. Dai BIBREF7 pre-trained unlabeled data to improve Sequence Learning with recurrent networks. Howard BIBREF8 proposed a general transfer learning method, Universal Language Model Fine-tuning (ULMFiT), with the key techniques for fine-tuning a language model. Radford BIBREF9 proposed that by generative pre-training of a language model on a diverse corpus of unlabeled text, large gains on a diverse range of tasks could be realized. Radford BIBREF9 achieved large improvements on many sentence-level tasks from the GLUE benchmark BIBREF10 . BERT BIBREF11 obtained new state-of-the-art results on a broad range of diverse tasks. BERT pre-trained deep bidirectional representations which jointly conditioned on both left and right context in all layers, following by discriminative fine-tuning on each specific task. Unlike previous works fine-tuning pre-trained language model to perform discriminative tasks, we aim to apply pre-trained BERT on generative tasks by perform the masked language model(MLM) task. To generate sentences that are compatible with given labels, we retrofit BERT to conditional BERT, by introducing a conditional masked language model task and fine-tuning BERT on the task. Text Data Augmentation Text data augmentation has been extensively studied in natural language processing. Sample-based methods includes downsampling from the majority classes and oversampling from the minority class, both of which perform weakly in practice. Generation-based methods employ deep generative models such as GANs BIBREF12 or VAEs BIBREF13 , BIBREF14 , trying to generate sentences from a continuous space with desired attributes of sentiment and tense. However, sentences generated in these methods are very hard to guarantee the quality both in label compatibility and sentence readability. In some specific areas BIBREF15 , BIBREF16 , BIBREF17 . word replacement augmentation was applied. Wang BIBREF18 proposed the use of neighboring words in continuous representations to create new instances for every word in a tweet to augment the training dataset. Zhang BIBREF19 extracted all replaceable words from the given text and randomly choose $r$ of them to be replaced, then substituted the replaceable words with synonyms from WordNet BIBREF5 . Kolomiyets BIBREF20 replaced only the headwords under a task-specific assumption that temporal trigger words usually occur as headwords. Kolomiyets BIBREF20 selected substitute words with top- $K$ scores given by the Latent Words LM BIBREF21 , which is a LM based on fixed length contexts. Fadaee BIBREF22 focused on the rare word problem in machine translation, replacing words in a source sentence with only rare words. A word in the translated sentence is also replaced using a word alignment method and a rightward LM. The work most similar to our research is Kobayashi BIBREF6 . Kobayashi used a fill-in-the-blank context for data augmentation by replacing every words in the sentence with language model. In order to prevent the generated words from reversing the information related to the labels of the sentences, Kobayashi BIBREF6 introduced a conditional constraint to control the replacement of words. Unlike previous works, we adopt a deep bidirectional language model to apply replacement, and the attention mechanism within our model allows a more structured memory for handling long-term dependencies in text, which resulting in more general and robust improvement on various downstream tasks. Preliminary: Masked Language Model Task In general, the language model(LM) models the probability of generating natural language sentences or documents. Given a sequence $\textbf {\textit {S}}$ of N tokens, $<t_1,t_2,...,t_N>$ , a forward language model allows us to predict the probability of the sequence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_1, t_2,..., t_{i-1}).$$ (Eq. 8) Similarly, a backward language model allows us to predict the probability of the sentence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_{i+1}, t_{i+2},..., t_N).$$ (Eq. 9) Traditionally, a bidirectional language model a shallow concatenation of independently trained forward and backward LMs. In order to train a deep bidirectional language model, BERT proposed Masked Language Model (MLM) task, which was also referred to Cloze Task BIBREF23 . MLM task randomly masks some percentage of the input tokens, and then predicts only those masked tokens according to their context. Given a masked token ${t_i}$ , the context is the tokens surrounding token ${t_i}$ in the sequence $\textbf {\textit {S}}$ , i.e. cloze sentence ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ . The final hidden vectors corresponding to the mask tokens are fed into an output softmax over the vocabulary to produce words with a probability distribution ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . MLM task only predicts the masked words rather than reconstructing the entire input, which suggests that more pre-training steps are required for the model to converge. Pre-trained BERT can augment sentences through MLM task, by predicting new words in masked positions according to their context. Conditional BERT As shown in Fig 1 , our conditional BERT shares the same model architecture with the original BERT. The differences are the input representation and training procedure. The input embeddings of BERT are the sum of the token embeddings, the segmentation embeddings and the position embeddings. For the segmentation embeddings in BERT, a learned sentence A embedding is added to every token of the first sentence, and if a second sentence exists, a sentence B embedding will be added to every token of the second sentence. However, the segmentation embeddings has no connection to the actual annotated labels of a sentence, like sense, sentiment or subjectivity, so predicted word is not always compatible with annotated labels. For example, given a positive movie remark “this actor is good", we have the word “good" masked. Through the Masked Language Model task by BERT, the predicted word in the masked position has potential to be negative word likes "bad" or "boring". Such new generated sentences by substituting masked words are implausible with respect to their original labels, which will be harmful if added to the corpus to apply augmentation. In order to address this issue, we propose a new task: “conditional masked language model". The conditional masked language model randomly masks some of the tokens from the labeled sentence, and the objective is to predict the original vocabulary index of the masked word based on both its context and its label. Given a masked token ${t_i}$ , the context ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ and label ${y}$ are both considered, aiming to calculate ${p(\cdot |y,\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ , instead of calculating ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . Unlike MLM pre-training, the conditional MLM objective allows the representation to fuse the context information and the label information, which allows us to further train a label-conditional deep bidirectional representations. To perform conditional MLM task, we fine-tune on pre-trained BERT. We alter the segmentation embeddings to label embeddings, which are learned corresponding to their annotated labels on labeled datasets. Note that the BERT are designed with segmentation embedding being embedding A or embedding B, so when a downstream task dataset with more than two labels, we have to adapt the size of embedding to label size compatible. We train conditional BERT using conditional MLM task on labeled dataset. After the model has converged, it is expected to be able to predict words in masked position both considering the context and the label. Conditional BERT Contextual Augmentation After the conditional BERT is well-trained, we utilize it to augment sentences. Given a labeled sentence from the corpus, we randomly mask a few words in the sentence. Through conditional BERT, various words compatibly with the label of the sentence are predicted by conditional BERT. After substituting the masked words with predicted words, a new sentences is generated, which shares similar context and same label with original sentence. Then new sentences are added to original corpus. We elaborate the entire process in algorithm "Conditional BERT Contextual Augmentation" . Conditional BERT contextual augmentation algorithm. Fine-tuning on the pre-trained BERT , we retrofit BERT to conditional BERT using conditional MLM task on labeled dataset. After the model converged, we utilize it to augment sentences. New sentences are added into dataset to augment the dataset. [1] Alter the segmentation embeddings to label embeddings Fine-tune the pre-trained BERT using conditional MLM task on labeled dataset D until convergence each iteration i=1,2,...,M Sample a sentence $s$ from D Randomly mask $k$ words Using fine-tuned conditional BERT to predict label-compatible words on masked positions to generate a new sentence $S^{\prime }$ Add new sentences into dataset $D$ to get augmented dataset $D^{\prime }$ Perform downstream task on augmented dataset $D^{\prime }$ Experiment In this section, we present conditional BERT parameter settings and, following Kobayashi BIBREF6 , we apply different augmentation methods on two types of neural models through six text classification tasks. The pre-trained BERT model we used in our experiment is BERT $_{BASE}$ , with number of layers (i.e., Transformer blocks) $L = 12$ , the hidden size $ H = 768$ , and the number of self-attention heads $A = 12$ , total parameters $= 110M$ . Detailed pre-train parameters setting can be found in original paper BIBREF11 . For each task, we perform the following steps independently. First, we evaluate the augmentation ability of original BERT model pre-trained on MLM task. We use pre-trained BERT to augment dataset, by predicted masked words only condition on context for each sentence. Second, we fine-tune the original BERT model to a conditional BERT. Well-trained conditional BERT augments each sentence in dataset by predicted masked words condition on both context and label. Third, we compare the performance of the two methods with Kobayashi's BIBREF6 contextual augmentation results. Note that the original BERT’s segmentation embeddings layer is compatible with two-label dataset. When the task-specific dataset is with more than two different labels, we should re-train a label size compatible label embeddings layer instead of directly fine-tuning the pre-trained one. Datasets Six benchmark classification datasets are listed in table 1 . Following Kim BIBREF24 , for a dataset without validation data, we use 10% of its training set for the validation set. Summary statistics of six classification datasets are shown in table 1. SST BIBREF25 SST (Stanford Sentiment Treebank) is a dataset for sentiment classification on movie reviews, which are annotated with five labels (SST5: very positive, positive, neutral, negative, or very negative) or two labels (SST2: positive or negative). Subj BIBREF26 Subj (Subjectivity dataset) is annotated with whether a sentence is subjective or objective. MPQA BIBREF27 MPQA Opinion Corpus is an opinion polarity detection dataset of short phrases rather than sentences, which contains news articles from a wide variety of news sources manually annotated for opinions and other private states (i.e., beliefs, emotions, sentiments, speculations, etc.). RT BIBREF28 RT is another movie review sentiment dataset contains a collection of short review excerpts from Rotten Tomatoes collected by Bo Pang and Lillian Lee. TREC BIBREF29 TREC is a dataset for classification of the six question types (whether the question is about person, location, numeric information, etc.). Text classification We evaluate the performance improvement brought by conditional BERT contextual augmentation on sentence classification tasks, so we need to prepare two common sentence classifiers beforehand. For comparison, following Kobayashi BIBREF6 , we adopt two typical classifier architectures: CNN or LSTM-RNN. The CNN-based classifier BIBREF24 has convolutional filters of size 3, 4, 5 and word embeddings. All outputs of each filter are concatenated before applied with a max-pooling over time, then fed into a two-layer feed-forward network with ReLU, followed by the softmax function. An RNN-based classifier has a single layer LSTM and word embeddings, whose output is fed into an output affine layer with the softmax function. For both the architectures, dropout BIBREF30 and Adam optimization BIBREF31 are applied during training. The train process is finish by early stopping with validation at each epoch. Sentence classifier hyper-parameters including learning rate, embedding dimension, unit or filter size, and dropout ratio, are selected using grid-search for each task-specific dataset. We refer to Kobayashi's implementation in the released code. For BERT, all hyper-parameters are kept the same as Devlin BIBREF11 , codes in Tensorflow and PyTorch are all available on github and pre-trained BERT model can also be downloaded. The number of conditional BERT training epochs ranges in [1-50] and number of masked words ranges in [1-2]. We compare the performance improvements obtained by our proposed method with the following baseline methods, “w/" means “with": w/synonym: Words are randomly replaced with synonyms from WordNet BIBREF5 . w/context: Proposed by Kobayashi BIBREF6 , which used a bidirectional language model to apply contextual augmentation, each word was replaced with a probability. w/context+label: Kobayashi’s contextual augmentation method BIBREF6 in a label-conditional LM architecture. Table 2 lists the accuracies of the all methods on two classifier architectures. The results show that, for various datasets on different classifier architectures, our conditional BERT contextual augmentation improves the model performances most. BERT can also augments sentences to some extent, but not as much as conditional BERT does. For we masked words randomly, the masked words may be label-sensitive or label-insensitive. If label-insensitive words are masked, words predicted through BERT may not be compatible with original labels. The improvement over all benchmark datasets also shows that conditional BERT is a general augmentation method for multi-labels sentence classification tasks. We also explore the effect of number of training steps to the performance of conditional BERT data augmentation. The fine-tuning epoch setting ranges in [1-50], we list the fine-tuning epoch of conditional BERT to outperform BERT for various benchmarks in table 3 . The results show that our conditional BERT contextual augmentation can achieve obvious performance improvement after only a few fine-tuning epochs, which is very convenient to apply to downstream tasks. Connection to Style Transfer In this section, we further deep into the connection to style transfer and apply our well trained conditional BERT to style transfer task. Style transfer is defined as the task of rephrasing the text to contain specific stylistic properties without changing the intent or affect within the context BIBREF32 . Our conditional MLM task changes words in the text condition on given label without changing the context. View from this point, the two tasks are very close. So in order to apply conditional BERT to style transfer task, given a specific stylistic sentence, we break it into two steps: first, we find the words relevant to the style; second, we mask the style-relevant words, then use conditional BERT to predict new substitutes with sentence context and target style property. In order to find style-relevant words in a sentence, we refer to Xu BIBREF33 , which proposed an attention-based method to extract the contribution of each word to the sentence sentimental label. For example, given a positive movie remark “This movie is funny and interesting", we filter out the words contributes largely to the label and mask them. Then through our conditional BERT contextual augmentation method, we fill in the masked position by predicting words conditioning on opposite label and sentence context, resulting in “This movie is boring and dull". The words “boring" and “dull" contribute to the new sentence being labeled as negative style. We sample some sentences from dataset SST2, transferring them to the opposite label, as listed in table 4 . Conclusions and Future Work In this paper, we fine-tune BERT to conditional BERT by introducing a novel conditional MLM task. After being well trained, the conditional BERT can be applied to data augmentation for sentence classification tasks. Experiment results show that our model outperforms several baseline methods obviously. Furthermore, we demonstrate that our conditional BERT can also be applied to style transfer task. In the future, (1)We will explore how to perform text data augmentation on imbalanced datasets with pre-trained language model, (2) we believe the idea of conditional BERT contextual augmentation is universal and will be applied to paragraph or document level data augmentation.

SST (Stanford Sentiment Treebank), Subj (Subjectivity dataset), MPQA Opinion Corpus, RT is another movie review sentiment dataset, TREC is a dataset for classification of the six question types

6e97c06f998f09256be752fa75c24ba853b0db24

6e97c06f998f09256be752fa75c24ba853b0db24_0

Q: How do the authors measure performance? Text: Introduction Deep neural network-based models are easy to overfit and result in losing their generalization due to limited size of training data. In order to address the issue, data augmentation methods are often applied to generate more training samples. Recent years have witnessed great success in applying data augmentation in the field of speech area BIBREF0 , BIBREF1 and computer vision BIBREF2 , BIBREF3 , BIBREF4 . Data augmentation in these areas can be easily performed by transformations like resizing, mirroring, random cropping, and color shifting. However, applying these universal transformations to texts is largely randomized and uncontrollable, which makes it impossible to ensure the semantic invariance and label correctness. For example, given a movie review “The actors is good", by mirroring we get “doog si srotca ehT", or by random cropping we get “actors is", both of which are meaningless. Existing data augmentation methods for text are often loss of generality, which are developed with handcrafted rules or pipelines for specific domains. A general approach for text data augmentation is replacement-based method, which generates new sentences by replacing the words in the sentences with relevant words (e.g. synonyms). However, words with synonyms from a handcrafted lexical database likes WordNet BIBREF5 are very limited , and the replacement-based augmentation with synonyms can only produce limited diverse patterns from the original texts. To address the limitation of replacement-based methods, Kobayashi BIBREF6 proposed contextual augmentation for labeled sentences by offering a wide range of substitute words, which are predicted by a label-conditional bidirectional language model according to the context. But contextual augmentation suffers from two shortages: the bidirectional language model is simply shallow concatenation of a forward and backward model, and the usage of LSTM models restricts their prediction ability to a short range. BERT, which stands for Bidirectional Encoder Representations from Transformers, pre-trained deep bidirectional representations by jointly conditioning on both left and right context in all layers. BERT addressed the unidirectional constraint by proposing a “masked language model" (MLM) objective by masking some percentage of the input tokens at random, and predicting the masked words based on its context. This is very similar to how contextual augmentation predict the replacement words. But BERT was proposed to pre-train text representations, so MLM task is performed in an unsupervised way, taking no label variance into consideration. This paper focuses on the replacement-based methods, by proposing a novel data augmentation method called conditional BERT contextual augmentation. The method applies contextual augmentation by conditional BERT, which is fine-tuned on BERT. We adopt BERT as our pre-trained language model with two reasons. First, BERT is based on Transformer. Transformer provides us with a more structured memory for handling long-term dependencies in text. Second, BERT, as a deep bidirectional model, is strictly more powerful than the shallow concatenation of a left-to-right and right-to left model. So we apply BERT to contextual augmentation for labeled sentences, by offering a wider range of substitute words predicted by the masked language model task. However, the masked language model predicts the masked word based only on its context, so the predicted word maybe incompatible with the annotated labels of the original sentences. In order to address this issue, we introduce a new fine-tuning objective: the "conditional masked language model"(C-MLM). The conditional masked language model randomly masks some of the tokens from an input, and the objective is to predict a label-compatible word based on both its context and sentence label. Unlike Kobayashi's work, the C-MLM objective allows a deep bidirectional representations by jointly conditioning on both left and right context in all layers. In order to evaluate how well our augmentation method improves performance of deep neural network models, following Kobayashi BIBREF6 , we experiment it on two most common neural network structures, LSTM-RNN and CNN, on text classification tasks. Through the experiments on six various different text classification tasks, we demonstrate that the proposed conditional BERT model augments sentence better than baselines, and conditional BERT contextual augmentation method can be easily applied to both convolutional or recurrent neural networks classifier. We further explore our conditional MLM task’s connection with style transfer task and demonstrate that our conditional BERT can also be applied to style transfer too. Our contributions are concluded as follows: To our best knowledge, this is the first attempt to alter BERT to a conditional BERT or apply BERT on text generation tasks. Fine-tuning on Pre-trained Language Model Language model pre-training has attracted wide attention and fine-tuning on pre-trained language model has shown to be effective for improving many downstream natural language processing tasks. Dai BIBREF7 pre-trained unlabeled data to improve Sequence Learning with recurrent networks. Howard BIBREF8 proposed a general transfer learning method, Universal Language Model Fine-tuning (ULMFiT), with the key techniques for fine-tuning a language model. Radford BIBREF9 proposed that by generative pre-training of a language model on a diverse corpus of unlabeled text, large gains on a diverse range of tasks could be realized. Radford BIBREF9 achieved large improvements on many sentence-level tasks from the GLUE benchmark BIBREF10 . BERT BIBREF11 obtained new state-of-the-art results on a broad range of diverse tasks. BERT pre-trained deep bidirectional representations which jointly conditioned on both left and right context in all layers, following by discriminative fine-tuning on each specific task. Unlike previous works fine-tuning pre-trained language model to perform discriminative tasks, we aim to apply pre-trained BERT on generative tasks by perform the masked language model(MLM) task. To generate sentences that are compatible with given labels, we retrofit BERT to conditional BERT, by introducing a conditional masked language model task and fine-tuning BERT on the task. Text Data Augmentation Text data augmentation has been extensively studied in natural language processing. Sample-based methods includes downsampling from the majority classes and oversampling from the minority class, both of which perform weakly in practice. Generation-based methods employ deep generative models such as GANs BIBREF12 or VAEs BIBREF13 , BIBREF14 , trying to generate sentences from a continuous space with desired attributes of sentiment and tense. However, sentences generated in these methods are very hard to guarantee the quality both in label compatibility and sentence readability. In some specific areas BIBREF15 , BIBREF16 , BIBREF17 . word replacement augmentation was applied. Wang BIBREF18 proposed the use of neighboring words in continuous representations to create new instances for every word in a tweet to augment the training dataset. Zhang BIBREF19 extracted all replaceable words from the given text and randomly choose $r$ of them to be replaced, then substituted the replaceable words with synonyms from WordNet BIBREF5 . Kolomiyets BIBREF20 replaced only the headwords under a task-specific assumption that temporal trigger words usually occur as headwords. Kolomiyets BIBREF20 selected substitute words with top- $K$ scores given by the Latent Words LM BIBREF21 , which is a LM based on fixed length contexts. Fadaee BIBREF22 focused on the rare word problem in machine translation, replacing words in a source sentence with only rare words. A word in the translated sentence is also replaced using a word alignment method and a rightward LM. The work most similar to our research is Kobayashi BIBREF6 . Kobayashi used a fill-in-the-blank context for data augmentation by replacing every words in the sentence with language model. In order to prevent the generated words from reversing the information related to the labels of the sentences, Kobayashi BIBREF6 introduced a conditional constraint to control the replacement of words. Unlike previous works, we adopt a deep bidirectional language model to apply replacement, and the attention mechanism within our model allows a more structured memory for handling long-term dependencies in text, which resulting in more general and robust improvement on various downstream tasks. Preliminary: Masked Language Model Task In general, the language model(LM) models the probability of generating natural language sentences or documents. Given a sequence $\textbf {\textit {S}}$ of N tokens, $<t_1,t_2,...,t_N>$ , a forward language model allows us to predict the probability of the sequence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_1, t_2,..., t_{i-1}).$$ (Eq. 8) Similarly, a backward language model allows us to predict the probability of the sentence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_{i+1}, t_{i+2},..., t_N).$$ (Eq. 9) Traditionally, a bidirectional language model a shallow concatenation of independently trained forward and backward LMs. In order to train a deep bidirectional language model, BERT proposed Masked Language Model (MLM) task, which was also referred to Cloze Task BIBREF23 . MLM task randomly masks some percentage of the input tokens, and then predicts only those masked tokens according to their context. Given a masked token ${t_i}$ , the context is the tokens surrounding token ${t_i}$ in the sequence $\textbf {\textit {S}}$ , i.e. cloze sentence ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ . The final hidden vectors corresponding to the mask tokens are fed into an output softmax over the vocabulary to produce words with a probability distribution ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . MLM task only predicts the masked words rather than reconstructing the entire input, which suggests that more pre-training steps are required for the model to converge. Pre-trained BERT can augment sentences through MLM task, by predicting new words in masked positions according to their context. Conditional BERT As shown in Fig 1 , our conditional BERT shares the same model architecture with the original BERT. The differences are the input representation and training procedure. The input embeddings of BERT are the sum of the token embeddings, the segmentation embeddings and the position embeddings. For the segmentation embeddings in BERT, a learned sentence A embedding is added to every token of the first sentence, and if a second sentence exists, a sentence B embedding will be added to every token of the second sentence. However, the segmentation embeddings has no connection to the actual annotated labels of a sentence, like sense, sentiment or subjectivity, so predicted word is not always compatible with annotated labels. For example, given a positive movie remark “this actor is good", we have the word “good" masked. Through the Masked Language Model task by BERT, the predicted word in the masked position has potential to be negative word likes "bad" or "boring". Such new generated sentences by substituting masked words are implausible with respect to their original labels, which will be harmful if added to the corpus to apply augmentation. In order to address this issue, we propose a new task: “conditional masked language model". The conditional masked language model randomly masks some of the tokens from the labeled sentence, and the objective is to predict the original vocabulary index of the masked word based on both its context and its label. Given a masked token ${t_i}$ , the context ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ and label ${y}$ are both considered, aiming to calculate ${p(\cdot |y,\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ , instead of calculating ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . Unlike MLM pre-training, the conditional MLM objective allows the representation to fuse the context information and the label information, which allows us to further train a label-conditional deep bidirectional representations. To perform conditional MLM task, we fine-tune on pre-trained BERT. We alter the segmentation embeddings to label embeddings, which are learned corresponding to their annotated labels on labeled datasets. Note that the BERT are designed with segmentation embedding being embedding A or embedding B, so when a downstream task dataset with more than two labels, we have to adapt the size of embedding to label size compatible. We train conditional BERT using conditional MLM task on labeled dataset. After the model has converged, it is expected to be able to predict words in masked position both considering the context and the label. Conditional BERT Contextual Augmentation After the conditional BERT is well-trained, we utilize it to augment sentences. Given a labeled sentence from the corpus, we randomly mask a few words in the sentence. Through conditional BERT, various words compatibly with the label of the sentence are predicted by conditional BERT. After substituting the masked words with predicted words, a new sentences is generated, which shares similar context and same label with original sentence. Then new sentences are added to original corpus. We elaborate the entire process in algorithm "Conditional BERT Contextual Augmentation" . Conditional BERT contextual augmentation algorithm. Fine-tuning on the pre-trained BERT , we retrofit BERT to conditional BERT using conditional MLM task on labeled dataset. After the model converged, we utilize it to augment sentences. New sentences are added into dataset to augment the dataset. [1] Alter the segmentation embeddings to label embeddings Fine-tune the pre-trained BERT using conditional MLM task on labeled dataset D until convergence each iteration i=1,2,...,M Sample a sentence $s$ from D Randomly mask $k$ words Using fine-tuned conditional BERT to predict label-compatible words on masked positions to generate a new sentence $S^{\prime }$ Add new sentences into dataset $D$ to get augmented dataset $D^{\prime }$ Perform downstream task on augmented dataset $D^{\prime }$ Experiment In this section, we present conditional BERT parameter settings and, following Kobayashi BIBREF6 , we apply different augmentation methods on two types of neural models through six text classification tasks. The pre-trained BERT model we used in our experiment is BERT $_{BASE}$ , with number of layers (i.e., Transformer blocks) $L = 12$ , the hidden size $ H = 768$ , and the number of self-attention heads $A = 12$ , total parameters $= 110M$ . Detailed pre-train parameters setting can be found in original paper BIBREF11 . For each task, we perform the following steps independently. First, we evaluate the augmentation ability of original BERT model pre-trained on MLM task. We use pre-trained BERT to augment dataset, by predicted masked words only condition on context for each sentence. Second, we fine-tune the original BERT model to a conditional BERT. Well-trained conditional BERT augments each sentence in dataset by predicted masked words condition on both context and label. Third, we compare the performance of the two methods with Kobayashi's BIBREF6 contextual augmentation results. Note that the original BERT’s segmentation embeddings layer is compatible with two-label dataset. When the task-specific dataset is with more than two different labels, we should re-train a label size compatible label embeddings layer instead of directly fine-tuning the pre-trained one. Datasets Six benchmark classification datasets are listed in table 1 . Following Kim BIBREF24 , for a dataset without validation data, we use 10% of its training set for the validation set. Summary statistics of six classification datasets are shown in table 1. SST BIBREF25 SST (Stanford Sentiment Treebank) is a dataset for sentiment classification on movie reviews, which are annotated with five labels (SST5: very positive, positive, neutral, negative, or very negative) or two labels (SST2: positive or negative). Subj BIBREF26 Subj (Subjectivity dataset) is annotated with whether a sentence is subjective or objective. MPQA BIBREF27 MPQA Opinion Corpus is an opinion polarity detection dataset of short phrases rather than sentences, which contains news articles from a wide variety of news sources manually annotated for opinions and other private states (i.e., beliefs, emotions, sentiments, speculations, etc.). RT BIBREF28 RT is another movie review sentiment dataset contains a collection of short review excerpts from Rotten Tomatoes collected by Bo Pang and Lillian Lee. TREC BIBREF29 TREC is a dataset for classification of the six question types (whether the question is about person, location, numeric information, etc.). Text classification We evaluate the performance improvement brought by conditional BERT contextual augmentation on sentence classification tasks, so we need to prepare two common sentence classifiers beforehand. For comparison, following Kobayashi BIBREF6 , we adopt two typical classifier architectures: CNN or LSTM-RNN. The CNN-based classifier BIBREF24 has convolutional filters of size 3, 4, 5 and word embeddings. All outputs of each filter are concatenated before applied with a max-pooling over time, then fed into a two-layer feed-forward network with ReLU, followed by the softmax function. An RNN-based classifier has a single layer LSTM and word embeddings, whose output is fed into an output affine layer with the softmax function. For both the architectures, dropout BIBREF30 and Adam optimization BIBREF31 are applied during training. The train process is finish by early stopping with validation at each epoch. Sentence classifier hyper-parameters including learning rate, embedding dimension, unit or filter size, and dropout ratio, are selected using grid-search for each task-specific dataset. We refer to Kobayashi's implementation in the released code. For BERT, all hyper-parameters are kept the same as Devlin BIBREF11 , codes in Tensorflow and PyTorch are all available on github and pre-trained BERT model can also be downloaded. The number of conditional BERT training epochs ranges in [1-50] and number of masked words ranges in [1-2]. We compare the performance improvements obtained by our proposed method with the following baseline methods, “w/" means “with": w/synonym: Words are randomly replaced with synonyms from WordNet BIBREF5 . w/context: Proposed by Kobayashi BIBREF6 , which used a bidirectional language model to apply contextual augmentation, each word was replaced with a probability. w/context+label: Kobayashi’s contextual augmentation method BIBREF6 in a label-conditional LM architecture. Table 2 lists the accuracies of the all methods on two classifier architectures. The results show that, for various datasets on different classifier architectures, our conditional BERT contextual augmentation improves the model performances most. BERT can also augments sentences to some extent, but not as much as conditional BERT does. For we masked words randomly, the masked words may be label-sensitive or label-insensitive. If label-insensitive words are masked, words predicted through BERT may not be compatible with original labels. The improvement over all benchmark datasets also shows that conditional BERT is a general augmentation method for multi-labels sentence classification tasks. We also explore the effect of number of training steps to the performance of conditional BERT data augmentation. The fine-tuning epoch setting ranges in [1-50], we list the fine-tuning epoch of conditional BERT to outperform BERT for various benchmarks in table 3 . The results show that our conditional BERT contextual augmentation can achieve obvious performance improvement after only a few fine-tuning epochs, which is very convenient to apply to downstream tasks. Connection to Style Transfer In this section, we further deep into the connection to style transfer and apply our well trained conditional BERT to style transfer task. Style transfer is defined as the task of rephrasing the text to contain specific stylistic properties without changing the intent or affect within the context BIBREF32 . Our conditional MLM task changes words in the text condition on given label without changing the context. View from this point, the two tasks are very close. So in order to apply conditional BERT to style transfer task, given a specific stylistic sentence, we break it into two steps: first, we find the words relevant to the style; second, we mask the style-relevant words, then use conditional BERT to predict new substitutes with sentence context and target style property. In order to find style-relevant words in a sentence, we refer to Xu BIBREF33 , which proposed an attention-based method to extract the contribution of each word to the sentence sentimental label. For example, given a positive movie remark “This movie is funny and interesting", we filter out the words contributes largely to the label and mask them. Then through our conditional BERT contextual augmentation method, we fill in the masked position by predicting words conditioning on opposite label and sentence context, resulting in “This movie is boring and dull". The words “boring" and “dull" contribute to the new sentence being labeled as negative style. We sample some sentences from dataset SST2, transferring them to the opposite label, as listed in table 4 . Conclusions and Future Work In this paper, we fine-tune BERT to conditional BERT by introducing a novel conditional MLM task. After being well trained, the conditional BERT can be applied to data augmentation for sentence classification tasks. Experiment results show that our model outperforms several baseline methods obviously. Furthermore, we demonstrate that our conditional BERT can also be applied to style transfer task. In the future, (1)We will explore how to perform text data augmentation on imbalanced datasets with pre-trained language model, (2) we believe the idea of conditional BERT contextual augmentation is universal and will be applied to paragraph or document level data augmentation.

Accuracy across six datasets

de2d33760dc05f9d28e9dabc13bab2b3264cadb7

de2d33760dc05f9d28e9dabc13bab2b3264cadb7_0

Q: Does the new objective perform better than the original objective bert is trained on? Text: Introduction Deep neural network-based models are easy to overfit and result in losing their generalization due to limited size of training data. In order to address the issue, data augmentation methods are often applied to generate more training samples. Recent years have witnessed great success in applying data augmentation in the field of speech area BIBREF0 , BIBREF1 and computer vision BIBREF2 , BIBREF3 , BIBREF4 . Data augmentation in these areas can be easily performed by transformations like resizing, mirroring, random cropping, and color shifting. However, applying these universal transformations to texts is largely randomized and uncontrollable, which makes it impossible to ensure the semantic invariance and label correctness. For example, given a movie review “The actors is good", by mirroring we get “doog si srotca ehT", or by random cropping we get “actors is", both of which are meaningless. Existing data augmentation methods for text are often loss of generality, which are developed with handcrafted rules or pipelines for specific domains. A general approach for text data augmentation is replacement-based method, which generates new sentences by replacing the words in the sentences with relevant words (e.g. synonyms). However, words with synonyms from a handcrafted lexical database likes WordNet BIBREF5 are very limited , and the replacement-based augmentation with synonyms can only produce limited diverse patterns from the original texts. To address the limitation of replacement-based methods, Kobayashi BIBREF6 proposed contextual augmentation for labeled sentences by offering a wide range of substitute words, which are predicted by a label-conditional bidirectional language model according to the context. But contextual augmentation suffers from two shortages: the bidirectional language model is simply shallow concatenation of a forward and backward model, and the usage of LSTM models restricts their prediction ability to a short range. BERT, which stands for Bidirectional Encoder Representations from Transformers, pre-trained deep bidirectional representations by jointly conditioning on both left and right context in all layers. BERT addressed the unidirectional constraint by proposing a “masked language model" (MLM) objective by masking some percentage of the input tokens at random, and predicting the masked words based on its context. This is very similar to how contextual augmentation predict the replacement words. But BERT was proposed to pre-train text representations, so MLM task is performed in an unsupervised way, taking no label variance into consideration. This paper focuses on the replacement-based methods, by proposing a novel data augmentation method called conditional BERT contextual augmentation. The method applies contextual augmentation by conditional BERT, which is fine-tuned on BERT. We adopt BERT as our pre-trained language model with two reasons. First, BERT is based on Transformer. Transformer provides us with a more structured memory for handling long-term dependencies in text. Second, BERT, as a deep bidirectional model, is strictly more powerful than the shallow concatenation of a left-to-right and right-to left model. So we apply BERT to contextual augmentation for labeled sentences, by offering a wider range of substitute words predicted by the masked language model task. However, the masked language model predicts the masked word based only on its context, so the predicted word maybe incompatible with the annotated labels of the original sentences. In order to address this issue, we introduce a new fine-tuning objective: the "conditional masked language model"(C-MLM). The conditional masked language model randomly masks some of the tokens from an input, and the objective is to predict a label-compatible word based on both its context and sentence label. Unlike Kobayashi's work, the C-MLM objective allows a deep bidirectional representations by jointly conditioning on both left and right context in all layers. In order to evaluate how well our augmentation method improves performance of deep neural network models, following Kobayashi BIBREF6 , we experiment it on two most common neural network structures, LSTM-RNN and CNN, on text classification tasks. Through the experiments on six various different text classification tasks, we demonstrate that the proposed conditional BERT model augments sentence better than baselines, and conditional BERT contextual augmentation method can be easily applied to both convolutional or recurrent neural networks classifier. We further explore our conditional MLM task’s connection with style transfer task and demonstrate that our conditional BERT can also be applied to style transfer too. Our contributions are concluded as follows: To our best knowledge, this is the first attempt to alter BERT to a conditional BERT or apply BERT on text generation tasks. Fine-tuning on Pre-trained Language Model Language model pre-training has attracted wide attention and fine-tuning on pre-trained language model has shown to be effective for improving many downstream natural language processing tasks. Dai BIBREF7 pre-trained unlabeled data to improve Sequence Learning with recurrent networks. Howard BIBREF8 proposed a general transfer learning method, Universal Language Model Fine-tuning (ULMFiT), with the key techniques for fine-tuning a language model. Radford BIBREF9 proposed that by generative pre-training of a language model on a diverse corpus of unlabeled text, large gains on a diverse range of tasks could be realized. Radford BIBREF9 achieved large improvements on many sentence-level tasks from the GLUE benchmark BIBREF10 . BERT BIBREF11 obtained new state-of-the-art results on a broad range of diverse tasks. BERT pre-trained deep bidirectional representations which jointly conditioned on both left and right context in all layers, following by discriminative fine-tuning on each specific task. Unlike previous works fine-tuning pre-trained language model to perform discriminative tasks, we aim to apply pre-trained BERT on generative tasks by perform the masked language model(MLM) task. To generate sentences that are compatible with given labels, we retrofit BERT to conditional BERT, by introducing a conditional masked language model task and fine-tuning BERT on the task. Text Data Augmentation Text data augmentation has been extensively studied in natural language processing. Sample-based methods includes downsampling from the majority classes and oversampling from the minority class, both of which perform weakly in practice. Generation-based methods employ deep generative models such as GANs BIBREF12 or VAEs BIBREF13 , BIBREF14 , trying to generate sentences from a continuous space with desired attributes of sentiment and tense. However, sentences generated in these methods are very hard to guarantee the quality both in label compatibility and sentence readability. In some specific areas BIBREF15 , BIBREF16 , BIBREF17 . word replacement augmentation was applied. Wang BIBREF18 proposed the use of neighboring words in continuous representations to create new instances for every word in a tweet to augment the training dataset. Zhang BIBREF19 extracted all replaceable words from the given text and randomly choose $r$ of them to be replaced, then substituted the replaceable words with synonyms from WordNet BIBREF5 . Kolomiyets BIBREF20 replaced only the headwords under a task-specific assumption that temporal trigger words usually occur as headwords. Kolomiyets BIBREF20 selected substitute words with top- $K$ scores given by the Latent Words LM BIBREF21 , which is a LM based on fixed length contexts. Fadaee BIBREF22 focused on the rare word problem in machine translation, replacing words in a source sentence with only rare words. A word in the translated sentence is also replaced using a word alignment method and a rightward LM. The work most similar to our research is Kobayashi BIBREF6 . Kobayashi used a fill-in-the-blank context for data augmentation by replacing every words in the sentence with language model. In order to prevent the generated words from reversing the information related to the labels of the sentences, Kobayashi BIBREF6 introduced a conditional constraint to control the replacement of words. Unlike previous works, we adopt a deep bidirectional language model to apply replacement, and the attention mechanism within our model allows a more structured memory for handling long-term dependencies in text, which resulting in more general and robust improvement on various downstream tasks. Preliminary: Masked Language Model Task In general, the language model(LM) models the probability of generating natural language sentences or documents. Given a sequence $\textbf {\textit {S}}$ of N tokens, $<t_1,t_2,...,t_N>$ , a forward language model allows us to predict the probability of the sequence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_1, t_2,..., t_{i-1}).$$ (Eq. 8) Similarly, a backward language model allows us to predict the probability of the sentence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_{i+1}, t_{i+2},..., t_N).$$ (Eq. 9) Traditionally, a bidirectional language model a shallow concatenation of independently trained forward and backward LMs. In order to train a deep bidirectional language model, BERT proposed Masked Language Model (MLM) task, which was also referred to Cloze Task BIBREF23 . MLM task randomly masks some percentage of the input tokens, and then predicts only those masked tokens according to their context. Given a masked token ${t_i}$ , the context is the tokens surrounding token ${t_i}$ in the sequence $\textbf {\textit {S}}$ , i.e. cloze sentence ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ . The final hidden vectors corresponding to the mask tokens are fed into an output softmax over the vocabulary to produce words with a probability distribution ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . MLM task only predicts the masked words rather than reconstructing the entire input, which suggests that more pre-training steps are required for the model to converge. Pre-trained BERT can augment sentences through MLM task, by predicting new words in masked positions according to their context. Conditional BERT As shown in Fig 1 , our conditional BERT shares the same model architecture with the original BERT. The differences are the input representation and training procedure. The input embeddings of BERT are the sum of the token embeddings, the segmentation embeddings and the position embeddings. For the segmentation embeddings in BERT, a learned sentence A embedding is added to every token of the first sentence, and if a second sentence exists, a sentence B embedding will be added to every token of the second sentence. However, the segmentation embeddings has no connection to the actual annotated labels of a sentence, like sense, sentiment or subjectivity, so predicted word is not always compatible with annotated labels. For example, given a positive movie remark “this actor is good", we have the word “good" masked. Through the Masked Language Model task by BERT, the predicted word in the masked position has potential to be negative word likes "bad" or "boring". Such new generated sentences by substituting masked words are implausible with respect to their original labels, which will be harmful if added to the corpus to apply augmentation. In order to address this issue, we propose a new task: “conditional masked language model". The conditional masked language model randomly masks some of the tokens from the labeled sentence, and the objective is to predict the original vocabulary index of the masked word based on both its context and its label. Given a masked token ${t_i}$ , the context ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ and label ${y}$ are both considered, aiming to calculate ${p(\cdot |y,\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ , instead of calculating ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . Unlike MLM pre-training, the conditional MLM objective allows the representation to fuse the context information and the label information, which allows us to further train a label-conditional deep bidirectional representations. To perform conditional MLM task, we fine-tune on pre-trained BERT. We alter the segmentation embeddings to label embeddings, which are learned corresponding to their annotated labels on labeled datasets. Note that the BERT are designed with segmentation embedding being embedding A or embedding B, so when a downstream task dataset with more than two labels, we have to adapt the size of embedding to label size compatible. We train conditional BERT using conditional MLM task on labeled dataset. After the model has converged, it is expected to be able to predict words in masked position both considering the context and the label. Conditional BERT Contextual Augmentation After the conditional BERT is well-trained, we utilize it to augment sentences. Given a labeled sentence from the corpus, we randomly mask a few words in the sentence. Through conditional BERT, various words compatibly with the label of the sentence are predicted by conditional BERT. After substituting the masked words with predicted words, a new sentences is generated, which shares similar context and same label with original sentence. Then new sentences are added to original corpus. We elaborate the entire process in algorithm "Conditional BERT Contextual Augmentation" . Conditional BERT contextual augmentation algorithm. Fine-tuning on the pre-trained BERT , we retrofit BERT to conditional BERT using conditional MLM task on labeled dataset. After the model converged, we utilize it to augment sentences. New sentences are added into dataset to augment the dataset. [1] Alter the segmentation embeddings to label embeddings Fine-tune the pre-trained BERT using conditional MLM task on labeled dataset D until convergence each iteration i=1,2,...,M Sample a sentence $s$ from D Randomly mask $k$ words Using fine-tuned conditional BERT to predict label-compatible words on masked positions to generate a new sentence $S^{\prime }$ Add new sentences into dataset $D$ to get augmented dataset $D^{\prime }$ Perform downstream task on augmented dataset $D^{\prime }$ Experiment In this section, we present conditional BERT parameter settings and, following Kobayashi BIBREF6 , we apply different augmentation methods on two types of neural models through six text classification tasks. The pre-trained BERT model we used in our experiment is BERT $_{BASE}$ , with number of layers (i.e., Transformer blocks) $L = 12$ , the hidden size $ H = 768$ , and the number of self-attention heads $A = 12$ , total parameters $= 110M$ . Detailed pre-train parameters setting can be found in original paper BIBREF11 . For each task, we perform the following steps independently. First, we evaluate the augmentation ability of original BERT model pre-trained on MLM task. We use pre-trained BERT to augment dataset, by predicted masked words only condition on context for each sentence. Second, we fine-tune the original BERT model to a conditional BERT. Well-trained conditional BERT augments each sentence in dataset by predicted masked words condition on both context and label. Third, we compare the performance of the two methods with Kobayashi's BIBREF6 contextual augmentation results. Note that the original BERT’s segmentation embeddings layer is compatible with two-label dataset. When the task-specific dataset is with more than two different labels, we should re-train a label size compatible label embeddings layer instead of directly fine-tuning the pre-trained one. Datasets Six benchmark classification datasets are listed in table 1 . Following Kim BIBREF24 , for a dataset without validation data, we use 10% of its training set for the validation set. Summary statistics of six classification datasets are shown in table 1. SST BIBREF25 SST (Stanford Sentiment Treebank) is a dataset for sentiment classification on movie reviews, which are annotated with five labels (SST5: very positive, positive, neutral, negative, or very negative) or two labels (SST2: positive or negative). Subj BIBREF26 Subj (Subjectivity dataset) is annotated with whether a sentence is subjective or objective. MPQA BIBREF27 MPQA Opinion Corpus is an opinion polarity detection dataset of short phrases rather than sentences, which contains news articles from a wide variety of news sources manually annotated for opinions and other private states (i.e., beliefs, emotions, sentiments, speculations, etc.). RT BIBREF28 RT is another movie review sentiment dataset contains a collection of short review excerpts from Rotten Tomatoes collected by Bo Pang and Lillian Lee. TREC BIBREF29 TREC is a dataset for classification of the six question types (whether the question is about person, location, numeric information, etc.). Text classification We evaluate the performance improvement brought by conditional BERT contextual augmentation on sentence classification tasks, so we need to prepare two common sentence classifiers beforehand. For comparison, following Kobayashi BIBREF6 , we adopt two typical classifier architectures: CNN or LSTM-RNN. The CNN-based classifier BIBREF24 has convolutional filters of size 3, 4, 5 and word embeddings. All outputs of each filter are concatenated before applied with a max-pooling over time, then fed into a two-layer feed-forward network with ReLU, followed by the softmax function. An RNN-based classifier has a single layer LSTM and word embeddings, whose output is fed into an output affine layer with the softmax function. For both the architectures, dropout BIBREF30 and Adam optimization BIBREF31 are applied during training. The train process is finish by early stopping with validation at each epoch. Sentence classifier hyper-parameters including learning rate, embedding dimension, unit or filter size, and dropout ratio, are selected using grid-search for each task-specific dataset. We refer to Kobayashi's implementation in the released code. For BERT, all hyper-parameters are kept the same as Devlin BIBREF11 , codes in Tensorflow and PyTorch are all available on github and pre-trained BERT model can also be downloaded. The number of conditional BERT training epochs ranges in [1-50] and number of masked words ranges in [1-2]. We compare the performance improvements obtained by our proposed method with the following baseline methods, “w/" means “with": w/synonym: Words are randomly replaced with synonyms from WordNet BIBREF5 . w/context: Proposed by Kobayashi BIBREF6 , which used a bidirectional language model to apply contextual augmentation, each word was replaced with a probability. w/context+label: Kobayashi’s contextual augmentation method BIBREF6 in a label-conditional LM architecture. Table 2 lists the accuracies of the all methods on two classifier architectures. The results show that, for various datasets on different classifier architectures, our conditional BERT contextual augmentation improves the model performances most. BERT can also augments sentences to some extent, but not as much as conditional BERT does. For we masked words randomly, the masked words may be label-sensitive or label-insensitive. If label-insensitive words are masked, words predicted through BERT may not be compatible with original labels. The improvement over all benchmark datasets also shows that conditional BERT is a general augmentation method for multi-labels sentence classification tasks. We also explore the effect of number of training steps to the performance of conditional BERT data augmentation. The fine-tuning epoch setting ranges in [1-50], we list the fine-tuning epoch of conditional BERT to outperform BERT for various benchmarks in table 3 . The results show that our conditional BERT contextual augmentation can achieve obvious performance improvement after only a few fine-tuning epochs, which is very convenient to apply to downstream tasks. Connection to Style Transfer In this section, we further deep into the connection to style transfer and apply our well trained conditional BERT to style transfer task. Style transfer is defined as the task of rephrasing the text to contain specific stylistic properties without changing the intent or affect within the context BIBREF32 . Our conditional MLM task changes words in the text condition on given label without changing the context. View from this point, the two tasks are very close. So in order to apply conditional BERT to style transfer task, given a specific stylistic sentence, we break it into two steps: first, we find the words relevant to the style; second, we mask the style-relevant words, then use conditional BERT to predict new substitutes with sentence context and target style property. In order to find style-relevant words in a sentence, we refer to Xu BIBREF33 , which proposed an attention-based method to extract the contribution of each word to the sentence sentimental label. For example, given a positive movie remark “This movie is funny and interesting", we filter out the words contributes largely to the label and mask them. Then through our conditional BERT contextual augmentation method, we fill in the masked position by predicting words conditioning on opposite label and sentence context, resulting in “This movie is boring and dull". The words “boring" and “dull" contribute to the new sentence being labeled as negative style. We sample some sentences from dataset SST2, transferring them to the opposite label, as listed in table 4 . Conclusions and Future Work In this paper, we fine-tune BERT to conditional BERT by introducing a novel conditional MLM task. After being well trained, the conditional BERT can be applied to data augmentation for sentence classification tasks. Experiment results show that our model outperforms several baseline methods obviously. Furthermore, we demonstrate that our conditional BERT can also be applied to style transfer task. In the future, (1)We will explore how to perform text data augmentation on imbalanced datasets with pre-trained language model, (2) we believe the idea of conditional BERT contextual augmentation is universal and will be applied to paragraph or document level data augmentation.

Yes

63bb39fd098786a510147f8ebc02408de350cb7c

63bb39fd098786a510147f8ebc02408de350cb7c_0

Q: Are other pretrained language models also evaluated for contextual augmentation? Text: Introduction Deep neural network-based models are easy to overfit and result in losing their generalization due to limited size of training data. In order to address the issue, data augmentation methods are often applied to generate more training samples. Recent years have witnessed great success in applying data augmentation in the field of speech area BIBREF0 , BIBREF1 and computer vision BIBREF2 , BIBREF3 , BIBREF4 . Data augmentation in these areas can be easily performed by transformations like resizing, mirroring, random cropping, and color shifting. However, applying these universal transformations to texts is largely randomized and uncontrollable, which makes it impossible to ensure the semantic invariance and label correctness. For example, given a movie review “The actors is good", by mirroring we get “doog si srotca ehT", or by random cropping we get “actors is", both of which are meaningless. Existing data augmentation methods for text are often loss of generality, which are developed with handcrafted rules or pipelines for specific domains. A general approach for text data augmentation is replacement-based method, which generates new sentences by replacing the words in the sentences with relevant words (e.g. synonyms). However, words with synonyms from a handcrafted lexical database likes WordNet BIBREF5 are very limited , and the replacement-based augmentation with synonyms can only produce limited diverse patterns from the original texts. To address the limitation of replacement-based methods, Kobayashi BIBREF6 proposed contextual augmentation for labeled sentences by offering a wide range of substitute words, which are predicted by a label-conditional bidirectional language model according to the context. But contextual augmentation suffers from two shortages: the bidirectional language model is simply shallow concatenation of a forward and backward model, and the usage of LSTM models restricts their prediction ability to a short range. BERT, which stands for Bidirectional Encoder Representations from Transformers, pre-trained deep bidirectional representations by jointly conditioning on both left and right context in all layers. BERT addressed the unidirectional constraint by proposing a “masked language model" (MLM) objective by masking some percentage of the input tokens at random, and predicting the masked words based on its context. This is very similar to how contextual augmentation predict the replacement words. But BERT was proposed to pre-train text representations, so MLM task is performed in an unsupervised way, taking no label variance into consideration. This paper focuses on the replacement-based methods, by proposing a novel data augmentation method called conditional BERT contextual augmentation. The method applies contextual augmentation by conditional BERT, which is fine-tuned on BERT. We adopt BERT as our pre-trained language model with two reasons. First, BERT is based on Transformer. Transformer provides us with a more structured memory for handling long-term dependencies in text. Second, BERT, as a deep bidirectional model, is strictly more powerful than the shallow concatenation of a left-to-right and right-to left model. So we apply BERT to contextual augmentation for labeled sentences, by offering a wider range of substitute words predicted by the masked language model task. However, the masked language model predicts the masked word based only on its context, so the predicted word maybe incompatible with the annotated labels of the original sentences. In order to address this issue, we introduce a new fine-tuning objective: the "conditional masked language model"(C-MLM). The conditional masked language model randomly masks some of the tokens from an input, and the objective is to predict a label-compatible word based on both its context and sentence label. Unlike Kobayashi's work, the C-MLM objective allows a deep bidirectional representations by jointly conditioning on both left and right context in all layers. In order to evaluate how well our augmentation method improves performance of deep neural network models, following Kobayashi BIBREF6 , we experiment it on two most common neural network structures, LSTM-RNN and CNN, on text classification tasks. Through the experiments on six various different text classification tasks, we demonstrate that the proposed conditional BERT model augments sentence better than baselines, and conditional BERT contextual augmentation method can be easily applied to both convolutional or recurrent neural networks classifier. We further explore our conditional MLM task’s connection with style transfer task and demonstrate that our conditional BERT can also be applied to style transfer too. Our contributions are concluded as follows: To our best knowledge, this is the first attempt to alter BERT to a conditional BERT or apply BERT on text generation tasks. Fine-tuning on Pre-trained Language Model Language model pre-training has attracted wide attention and fine-tuning on pre-trained language model has shown to be effective for improving many downstream natural language processing tasks. Dai BIBREF7 pre-trained unlabeled data to improve Sequence Learning with recurrent networks. Howard BIBREF8 proposed a general transfer learning method, Universal Language Model Fine-tuning (ULMFiT), with the key techniques for fine-tuning a language model. Radford BIBREF9 proposed that by generative pre-training of a language model on a diverse corpus of unlabeled text, large gains on a diverse range of tasks could be realized. Radford BIBREF9 achieved large improvements on many sentence-level tasks from the GLUE benchmark BIBREF10 . BERT BIBREF11 obtained new state-of-the-art results on a broad range of diverse tasks. BERT pre-trained deep bidirectional representations which jointly conditioned on both left and right context in all layers, following by discriminative fine-tuning on each specific task. Unlike previous works fine-tuning pre-trained language model to perform discriminative tasks, we aim to apply pre-trained BERT on generative tasks by perform the masked language model(MLM) task. To generate sentences that are compatible with given labels, we retrofit BERT to conditional BERT, by introducing a conditional masked language model task and fine-tuning BERT on the task. Text Data Augmentation Text data augmentation has been extensively studied in natural language processing. Sample-based methods includes downsampling from the majority classes and oversampling from the minority class, both of which perform weakly in practice. Generation-based methods employ deep generative models such as GANs BIBREF12 or VAEs BIBREF13 , BIBREF14 , trying to generate sentences from a continuous space with desired attributes of sentiment and tense. However, sentences generated in these methods are very hard to guarantee the quality both in label compatibility and sentence readability. In some specific areas BIBREF15 , BIBREF16 , BIBREF17 . word replacement augmentation was applied. Wang BIBREF18 proposed the use of neighboring words in continuous representations to create new instances for every word in a tweet to augment the training dataset. Zhang BIBREF19 extracted all replaceable words from the given text and randomly choose $r$ of them to be replaced, then substituted the replaceable words with synonyms from WordNet BIBREF5 . Kolomiyets BIBREF20 replaced only the headwords under a task-specific assumption that temporal trigger words usually occur as headwords. Kolomiyets BIBREF20 selected substitute words with top- $K$ scores given by the Latent Words LM BIBREF21 , which is a LM based on fixed length contexts. Fadaee BIBREF22 focused on the rare word problem in machine translation, replacing words in a source sentence with only rare words. A word in the translated sentence is also replaced using a word alignment method and a rightward LM. The work most similar to our research is Kobayashi BIBREF6 . Kobayashi used a fill-in-the-blank context for data augmentation by replacing every words in the sentence with language model. In order to prevent the generated words from reversing the information related to the labels of the sentences, Kobayashi BIBREF6 introduced a conditional constraint to control the replacement of words. Unlike previous works, we adopt a deep bidirectional language model to apply replacement, and the attention mechanism within our model allows a more structured memory for handling long-term dependencies in text, which resulting in more general and robust improvement on various downstream tasks. Preliminary: Masked Language Model Task In general, the language model(LM) models the probability of generating natural language sentences or documents. Given a sequence $\textbf {\textit {S}}$ of N tokens, $<t_1,t_2,...,t_N>$ , a forward language model allows us to predict the probability of the sequence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_1, t_2,..., t_{i-1}).$$ (Eq. 8) Similarly, a backward language model allows us to predict the probability of the sentence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_{i+1}, t_{i+2},..., t_N).$$ (Eq. 9) Traditionally, a bidirectional language model a shallow concatenation of independently trained forward and backward LMs. In order to train a deep bidirectional language model, BERT proposed Masked Language Model (MLM) task, which was also referred to Cloze Task BIBREF23 . MLM task randomly masks some percentage of the input tokens, and then predicts only those masked tokens according to their context. Given a masked token ${t_i}$ , the context is the tokens surrounding token ${t_i}$ in the sequence $\textbf {\textit {S}}$ , i.e. cloze sentence ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ . The final hidden vectors corresponding to the mask tokens are fed into an output softmax over the vocabulary to produce words with a probability distribution ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . MLM task only predicts the masked words rather than reconstructing the entire input, which suggests that more pre-training steps are required for the model to converge. Pre-trained BERT can augment sentences through MLM task, by predicting new words in masked positions according to their context. Conditional BERT As shown in Fig 1 , our conditional BERT shares the same model architecture with the original BERT. The differences are the input representation and training procedure. The input embeddings of BERT are the sum of the token embeddings, the segmentation embeddings and the position embeddings. For the segmentation embeddings in BERT, a learned sentence A embedding is added to every token of the first sentence, and if a second sentence exists, a sentence B embedding will be added to every token of the second sentence. However, the segmentation embeddings has no connection to the actual annotated labels of a sentence, like sense, sentiment or subjectivity, so predicted word is not always compatible with annotated labels. For example, given a positive movie remark “this actor is good", we have the word “good" masked. Through the Masked Language Model task by BERT, the predicted word in the masked position has potential to be negative word likes "bad" or "boring". Such new generated sentences by substituting masked words are implausible with respect to their original labels, which will be harmful if added to the corpus to apply augmentation. In order to address this issue, we propose a new task: “conditional masked language model". The conditional masked language model randomly masks some of the tokens from the labeled sentence, and the objective is to predict the original vocabulary index of the masked word based on both its context and its label. Given a masked token ${t_i}$ , the context ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ and label ${y}$ are both considered, aiming to calculate ${p(\cdot |y,\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ , instead of calculating ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . Unlike MLM pre-training, the conditional MLM objective allows the representation to fuse the context information and the label information, which allows us to further train a label-conditional deep bidirectional representations. To perform conditional MLM task, we fine-tune on pre-trained BERT. We alter the segmentation embeddings to label embeddings, which are learned corresponding to their annotated labels on labeled datasets. Note that the BERT are designed with segmentation embedding being embedding A or embedding B, so when a downstream task dataset with more than two labels, we have to adapt the size of embedding to label size compatible. We train conditional BERT using conditional MLM task on labeled dataset. After the model has converged, it is expected to be able to predict words in masked position both considering the context and the label. Conditional BERT Contextual Augmentation After the conditional BERT is well-trained, we utilize it to augment sentences. Given a labeled sentence from the corpus, we randomly mask a few words in the sentence. Through conditional BERT, various words compatibly with the label of the sentence are predicted by conditional BERT. After substituting the masked words with predicted words, a new sentences is generated, which shares similar context and same label with original sentence. Then new sentences are added to original corpus. We elaborate the entire process in algorithm "Conditional BERT Contextual Augmentation" . Conditional BERT contextual augmentation algorithm. Fine-tuning on the pre-trained BERT , we retrofit BERT to conditional BERT using conditional MLM task on labeled dataset. After the model converged, we utilize it to augment sentences. New sentences are added into dataset to augment the dataset. [1] Alter the segmentation embeddings to label embeddings Fine-tune the pre-trained BERT using conditional MLM task on labeled dataset D until convergence each iteration i=1,2,...,M Sample a sentence $s$ from D Randomly mask $k$ words Using fine-tuned conditional BERT to predict label-compatible words on masked positions to generate a new sentence $S^{\prime }$ Add new sentences into dataset $D$ to get augmented dataset $D^{\prime }$ Perform downstream task on augmented dataset $D^{\prime }$ Experiment In this section, we present conditional BERT parameter settings and, following Kobayashi BIBREF6 , we apply different augmentation methods on two types of neural models through six text classification tasks. The pre-trained BERT model we used in our experiment is BERT $_{BASE}$ , with number of layers (i.e., Transformer blocks) $L = 12$ , the hidden size $ H = 768$ , and the number of self-attention heads $A = 12$ , total parameters $= 110M$ . Detailed pre-train parameters setting can be found in original paper BIBREF11 . For each task, we perform the following steps independently. First, we evaluate the augmentation ability of original BERT model pre-trained on MLM task. We use pre-trained BERT to augment dataset, by predicted masked words only condition on context for each sentence. Second, we fine-tune the original BERT model to a conditional BERT. Well-trained conditional BERT augments each sentence in dataset by predicted masked words condition on both context and label. Third, we compare the performance of the two methods with Kobayashi's BIBREF6 contextual augmentation results. Note that the original BERT’s segmentation embeddings layer is compatible with two-label dataset. When the task-specific dataset is with more than two different labels, we should re-train a label size compatible label embeddings layer instead of directly fine-tuning the pre-trained one. Datasets Six benchmark classification datasets are listed in table 1 . Following Kim BIBREF24 , for a dataset without validation data, we use 10% of its training set for the validation set. Summary statistics of six classification datasets are shown in table 1. SST BIBREF25 SST (Stanford Sentiment Treebank) is a dataset for sentiment classification on movie reviews, which are annotated with five labels (SST5: very positive, positive, neutral, negative, or very negative) or two labels (SST2: positive or negative). Subj BIBREF26 Subj (Subjectivity dataset) is annotated with whether a sentence is subjective or objective. MPQA BIBREF27 MPQA Opinion Corpus is an opinion polarity detection dataset of short phrases rather than sentences, which contains news articles from a wide variety of news sources manually annotated for opinions and other private states (i.e., beliefs, emotions, sentiments, speculations, etc.). RT BIBREF28 RT is another movie review sentiment dataset contains a collection of short review excerpts from Rotten Tomatoes collected by Bo Pang and Lillian Lee. TREC BIBREF29 TREC is a dataset for classification of the six question types (whether the question is about person, location, numeric information, etc.). Text classification We evaluate the performance improvement brought by conditional BERT contextual augmentation on sentence classification tasks, so we need to prepare two common sentence classifiers beforehand. For comparison, following Kobayashi BIBREF6 , we adopt two typical classifier architectures: CNN or LSTM-RNN. The CNN-based classifier BIBREF24 has convolutional filters of size 3, 4, 5 and word embeddings. All outputs of each filter are concatenated before applied with a max-pooling over time, then fed into a two-layer feed-forward network with ReLU, followed by the softmax function. An RNN-based classifier has a single layer LSTM and word embeddings, whose output is fed into an output affine layer with the softmax function. For both the architectures, dropout BIBREF30 and Adam optimization BIBREF31 are applied during training. The train process is finish by early stopping with validation at each epoch. Sentence classifier hyper-parameters including learning rate, embedding dimension, unit or filter size, and dropout ratio, are selected using grid-search for each task-specific dataset. We refer to Kobayashi's implementation in the released code. For BERT, all hyper-parameters are kept the same as Devlin BIBREF11 , codes in Tensorflow and PyTorch are all available on github and pre-trained BERT model can also be downloaded. The number of conditional BERT training epochs ranges in [1-50] and number of masked words ranges in [1-2]. We compare the performance improvements obtained by our proposed method with the following baseline methods, “w/" means “with": w/synonym: Words are randomly replaced with synonyms from WordNet BIBREF5 . w/context: Proposed by Kobayashi BIBREF6 , which used a bidirectional language model to apply contextual augmentation, each word was replaced with a probability. w/context+label: Kobayashi’s contextual augmentation method BIBREF6 in a label-conditional LM architecture. Table 2 lists the accuracies of the all methods on two classifier architectures. The results show that, for various datasets on different classifier architectures, our conditional BERT contextual augmentation improves the model performances most. BERT can also augments sentences to some extent, but not as much as conditional BERT does. For we masked words randomly, the masked words may be label-sensitive or label-insensitive. If label-insensitive words are masked, words predicted through BERT may not be compatible with original labels. The improvement over all benchmark datasets also shows that conditional BERT is a general augmentation method for multi-labels sentence classification tasks. We also explore the effect of number of training steps to the performance of conditional BERT data augmentation. The fine-tuning epoch setting ranges in [1-50], we list the fine-tuning epoch of conditional BERT to outperform BERT for various benchmarks in table 3 . The results show that our conditional BERT contextual augmentation can achieve obvious performance improvement after only a few fine-tuning epochs, which is very convenient to apply to downstream tasks. Connection to Style Transfer In this section, we further deep into the connection to style transfer and apply our well trained conditional BERT to style transfer task. Style transfer is defined as the task of rephrasing the text to contain specific stylistic properties without changing the intent or affect within the context BIBREF32 . Our conditional MLM task changes words in the text condition on given label without changing the context. View from this point, the two tasks are very close. So in order to apply conditional BERT to style transfer task, given a specific stylistic sentence, we break it into two steps: first, we find the words relevant to the style; second, we mask the style-relevant words, then use conditional BERT to predict new substitutes with sentence context and target style property. In order to find style-relevant words in a sentence, we refer to Xu BIBREF33 , which proposed an attention-based method to extract the contribution of each word to the sentence sentimental label. For example, given a positive movie remark “This movie is funny and interesting", we filter out the words contributes largely to the label and mask them. Then through our conditional BERT contextual augmentation method, we fill in the masked position by predicting words conditioning on opposite label and sentence context, resulting in “This movie is boring and dull". The words “boring" and “dull" contribute to the new sentence being labeled as negative style. We sample some sentences from dataset SST2, transferring them to the opposite label, as listed in table 4 . Conclusions and Future Work In this paper, we fine-tune BERT to conditional BERT by introducing a novel conditional MLM task. After being well trained, the conditional BERT can be applied to data augmentation for sentence classification tasks. Experiment results show that our model outperforms several baseline methods obviously. Furthermore, we demonstrate that our conditional BERT can also be applied to style transfer task. In the future, (1)We will explore how to perform text data augmentation on imbalanced datasets with pre-trained language model, (2) we believe the idea of conditional BERT contextual augmentation is universal and will be applied to paragraph or document level data augmentation.

No

6333845facb22f862ffc684293eccc03002a4830

6333845facb22f862ffc684293eccc03002a4830_0

Q: Do the authors report performance of conditional bert on tasks without data augmentation? Text: Introduction Deep neural network-based models are easy to overfit and result in losing their generalization due to limited size of training data. In order to address the issue, data augmentation methods are often applied to generate more training samples. Recent years have witnessed great success in applying data augmentation in the field of speech area BIBREF0 , BIBREF1 and computer vision BIBREF2 , BIBREF3 , BIBREF4 . Data augmentation in these areas can be easily performed by transformations like resizing, mirroring, random cropping, and color shifting. However, applying these universal transformations to texts is largely randomized and uncontrollable, which makes it impossible to ensure the semantic invariance and label correctness. For example, given a movie review “The actors is good", by mirroring we get “doog si srotca ehT", or by random cropping we get “actors is", both of which are meaningless. Existing data augmentation methods for text are often loss of generality, which are developed with handcrafted rules or pipelines for specific domains. A general approach for text data augmentation is replacement-based method, which generates new sentences by replacing the words in the sentences with relevant words (e.g. synonyms). However, words with synonyms from a handcrafted lexical database likes WordNet BIBREF5 are very limited , and the replacement-based augmentation with synonyms can only produce limited diverse patterns from the original texts. To address the limitation of replacement-based methods, Kobayashi BIBREF6 proposed contextual augmentation for labeled sentences by offering a wide range of substitute words, which are predicted by a label-conditional bidirectional language model according to the context. But contextual augmentation suffers from two shortages: the bidirectional language model is simply shallow concatenation of a forward and backward model, and the usage of LSTM models restricts their prediction ability to a short range. BERT, which stands for Bidirectional Encoder Representations from Transformers, pre-trained deep bidirectional representations by jointly conditioning on both left and right context in all layers. BERT addressed the unidirectional constraint by proposing a “masked language model" (MLM) objective by masking some percentage of the input tokens at random, and predicting the masked words based on its context. This is very similar to how contextual augmentation predict the replacement words. But BERT was proposed to pre-train text representations, so MLM task is performed in an unsupervised way, taking no label variance into consideration. This paper focuses on the replacement-based methods, by proposing a novel data augmentation method called conditional BERT contextual augmentation. The method applies contextual augmentation by conditional BERT, which is fine-tuned on BERT. We adopt BERT as our pre-trained language model with two reasons. First, BERT is based on Transformer. Transformer provides us with a more structured memory for handling long-term dependencies in text. Second, BERT, as a deep bidirectional model, is strictly more powerful than the shallow concatenation of a left-to-right and right-to left model. So we apply BERT to contextual augmentation for labeled sentences, by offering a wider range of substitute words predicted by the masked language model task. However, the masked language model predicts the masked word based only on its context, so the predicted word maybe incompatible with the annotated labels of the original sentences. In order to address this issue, we introduce a new fine-tuning objective: the "conditional masked language model"(C-MLM). The conditional masked language model randomly masks some of the tokens from an input, and the objective is to predict a label-compatible word based on both its context and sentence label. Unlike Kobayashi's work, the C-MLM objective allows a deep bidirectional representations by jointly conditioning on both left and right context in all layers. In order to evaluate how well our augmentation method improves performance of deep neural network models, following Kobayashi BIBREF6 , we experiment it on two most common neural network structures, LSTM-RNN and CNN, on text classification tasks. Through the experiments on six various different text classification tasks, we demonstrate that the proposed conditional BERT model augments sentence better than baselines, and conditional BERT contextual augmentation method can be easily applied to both convolutional or recurrent neural networks classifier. We further explore our conditional MLM task’s connection with style transfer task and demonstrate that our conditional BERT can also be applied to style transfer too. Our contributions are concluded as follows: To our best knowledge, this is the first attempt to alter BERT to a conditional BERT or apply BERT on text generation tasks. Fine-tuning on Pre-trained Language Model Language model pre-training has attracted wide attention and fine-tuning on pre-trained language model has shown to be effective for improving many downstream natural language processing tasks. Dai BIBREF7 pre-trained unlabeled data to improve Sequence Learning with recurrent networks. Howard BIBREF8 proposed a general transfer learning method, Universal Language Model Fine-tuning (ULMFiT), with the key techniques for fine-tuning a language model. Radford BIBREF9 proposed that by generative pre-training of a language model on a diverse corpus of unlabeled text, large gains on a diverse range of tasks could be realized. Radford BIBREF9 achieved large improvements on many sentence-level tasks from the GLUE benchmark BIBREF10 . BERT BIBREF11 obtained new state-of-the-art results on a broad range of diverse tasks. BERT pre-trained deep bidirectional representations which jointly conditioned on both left and right context in all layers, following by discriminative fine-tuning on each specific task. Unlike previous works fine-tuning pre-trained language model to perform discriminative tasks, we aim to apply pre-trained BERT on generative tasks by perform the masked language model(MLM) task. To generate sentences that are compatible with given labels, we retrofit BERT to conditional BERT, by introducing a conditional masked language model task and fine-tuning BERT on the task. Text Data Augmentation Text data augmentation has been extensively studied in natural language processing. Sample-based methods includes downsampling from the majority classes and oversampling from the minority class, both of which perform weakly in practice. Generation-based methods employ deep generative models such as GANs BIBREF12 or VAEs BIBREF13 , BIBREF14 , trying to generate sentences from a continuous space with desired attributes of sentiment and tense. However, sentences generated in these methods are very hard to guarantee the quality both in label compatibility and sentence readability. In some specific areas BIBREF15 , BIBREF16 , BIBREF17 . word replacement augmentation was applied. Wang BIBREF18 proposed the use of neighboring words in continuous representations to create new instances for every word in a tweet to augment the training dataset. Zhang BIBREF19 extracted all replaceable words from the given text and randomly choose $r$ of them to be replaced, then substituted the replaceable words with synonyms from WordNet BIBREF5 . Kolomiyets BIBREF20 replaced only the headwords under a task-specific assumption that temporal trigger words usually occur as headwords. Kolomiyets BIBREF20 selected substitute words with top- $K$ scores given by the Latent Words LM BIBREF21 , which is a LM based on fixed length contexts. Fadaee BIBREF22 focused on the rare word problem in machine translation, replacing words in a source sentence with only rare words. A word in the translated sentence is also replaced using a word alignment method and a rightward LM. The work most similar to our research is Kobayashi BIBREF6 . Kobayashi used a fill-in-the-blank context for data augmentation by replacing every words in the sentence with language model. In order to prevent the generated words from reversing the information related to the labels of the sentences, Kobayashi BIBREF6 introduced a conditional constraint to control the replacement of words. Unlike previous works, we adopt a deep bidirectional language model to apply replacement, and the attention mechanism within our model allows a more structured memory for handling long-term dependencies in text, which resulting in more general and robust improvement on various downstream tasks. Preliminary: Masked Language Model Task In general, the language model(LM) models the probability of generating natural language sentences or documents. Given a sequence $\textbf {\textit {S}}$ of N tokens, $<t_1,t_2,...,t_N>$ , a forward language model allows us to predict the probability of the sequence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_1, t_2,..., t_{i-1}).$$ (Eq. 8) Similarly, a backward language model allows us to predict the probability of the sentence as: $$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_{i+1}, t_{i+2},..., t_N).$$ (Eq. 9) Traditionally, a bidirectional language model a shallow concatenation of independently trained forward and backward LMs. In order to train a deep bidirectional language model, BERT proposed Masked Language Model (MLM) task, which was also referred to Cloze Task BIBREF23 . MLM task randomly masks some percentage of the input tokens, and then predicts only those masked tokens according to their context. Given a masked token ${t_i}$ , the context is the tokens surrounding token ${t_i}$ in the sequence $\textbf {\textit {S}}$ , i.e. cloze sentence ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ . The final hidden vectors corresponding to the mask tokens are fed into an output softmax over the vocabulary to produce words with a probability distribution ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . MLM task only predicts the masked words rather than reconstructing the entire input, which suggests that more pre-training steps are required for the model to converge. Pre-trained BERT can augment sentences through MLM task, by predicting new words in masked positions according to their context. Conditional BERT As shown in Fig 1 , our conditional BERT shares the same model architecture with the original BERT. The differences are the input representation and training procedure. The input embeddings of BERT are the sum of the token embeddings, the segmentation embeddings and the position embeddings. For the segmentation embeddings in BERT, a learned sentence A embedding is added to every token of the first sentence, and if a second sentence exists, a sentence B embedding will be added to every token of the second sentence. However, the segmentation embeddings has no connection to the actual annotated labels of a sentence, like sense, sentiment or subjectivity, so predicted word is not always compatible with annotated labels. For example, given a positive movie remark “this actor is good", we have the word “good" masked. Through the Masked Language Model task by BERT, the predicted word in the masked position has potential to be negative word likes "bad" or "boring". Such new generated sentences by substituting masked words are implausible with respect to their original labels, which will be harmful if added to the corpus to apply augmentation. In order to address this issue, we propose a new task: “conditional masked language model". The conditional masked language model randomly masks some of the tokens from the labeled sentence, and the objective is to predict the original vocabulary index of the masked word based on both its context and its label. Given a masked token ${t_i}$ , the context ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ and label ${y}$ are both considered, aiming to calculate ${p(\cdot |y,\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ , instead of calculating ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . Unlike MLM pre-training, the conditional MLM objective allows the representation to fuse the context information and the label information, which allows us to further train a label-conditional deep bidirectional representations. To perform conditional MLM task, we fine-tune on pre-trained BERT. We alter the segmentation embeddings to label embeddings, which are learned corresponding to their annotated labels on labeled datasets. Note that the BERT are designed with segmentation embedding being embedding A or embedding B, so when a downstream task dataset with more than two labels, we have to adapt the size of embedding to label size compatible. We train conditional BERT using conditional MLM task on labeled dataset. After the model has converged, it is expected to be able to predict words in masked position both considering the context and the label. Conditional BERT Contextual Augmentation After the conditional BERT is well-trained, we utilize it to augment sentences. Given a labeled sentence from the corpus, we randomly mask a few words in the sentence. Through conditional BERT, various words compatibly with the label of the sentence are predicted by conditional BERT. After substituting the masked words with predicted words, a new sentences is generated, which shares similar context and same label with original sentence. Then new sentences are added to original corpus. We elaborate the entire process in algorithm "Conditional BERT Contextual Augmentation" . Conditional BERT contextual augmentation algorithm. Fine-tuning on the pre-trained BERT , we retrofit BERT to conditional BERT using conditional MLM task on labeled dataset. After the model converged, we utilize it to augment sentences. New sentences are added into dataset to augment the dataset. [1] Alter the segmentation embeddings to label embeddings Fine-tune the pre-trained BERT using conditional MLM task on labeled dataset D until convergence each iteration i=1,2,...,M Sample a sentence $s$ from D Randomly mask $k$ words Using fine-tuned conditional BERT to predict label-compatible words on masked positions to generate a new sentence $S^{\prime }$ Add new sentences into dataset $D$ to get augmented dataset $D^{\prime }$ Perform downstream task on augmented dataset $D^{\prime }$ Experiment In this section, we present conditional BERT parameter settings and, following Kobayashi BIBREF6 , we apply different augmentation methods on two types of neural models through six text classification tasks. The pre-trained BERT model we used in our experiment is BERT $_{BASE}$ , with number of layers (i.e., Transformer blocks) $L = 12$ , the hidden size $ H = 768$ , and the number of self-attention heads $A = 12$ , total parameters $= 110M$ . Detailed pre-train parameters setting can be found in original paper BIBREF11 . For each task, we perform the following steps independently. First, we evaluate the augmentation ability of original BERT model pre-trained on MLM task. We use pre-trained BERT to augment dataset, by predicted masked words only condition on context for each sentence. Second, we fine-tune the original BERT model to a conditional BERT. Well-trained conditional BERT augments each sentence in dataset by predicted masked words condition on both context and label. Third, we compare the performance of the two methods with Kobayashi's BIBREF6 contextual augmentation results. Note that the original BERT’s segmentation embeddings layer is compatible with two-label dataset. When the task-specific dataset is with more than two different labels, we should re-train a label size compatible label embeddings layer instead of directly fine-tuning the pre-trained one. Datasets Six benchmark classification datasets are listed in table 1 . Following Kim BIBREF24 , for a dataset without validation data, we use 10% of its training set for the validation set. Summary statistics of six classification datasets are shown in table 1. SST BIBREF25 SST (Stanford Sentiment Treebank) is a dataset for sentiment classification on movie reviews, which are annotated with five labels (SST5: very positive, positive, neutral, negative, or very negative) or two labels (SST2: positive or negative). Subj BIBREF26 Subj (Subjectivity dataset) is annotated with whether a sentence is subjective or objective. MPQA BIBREF27 MPQA Opinion Corpus is an opinion polarity detection dataset of short phrases rather than sentences, which contains news articles from a wide variety of news sources manually annotated for opinions and other private states (i.e., beliefs, emotions, sentiments, speculations, etc.). RT BIBREF28 RT is another movie review sentiment dataset contains a collection of short review excerpts from Rotten Tomatoes collected by Bo Pang and Lillian Lee. TREC BIBREF29 TREC is a dataset for classification of the six question types (whether the question is about person, location, numeric information, etc.). Text classification We evaluate the performance improvement brought by conditional BERT contextual augmentation on sentence classification tasks, so we need to prepare two common sentence classifiers beforehand. For comparison, following Kobayashi BIBREF6 , we adopt two typical classifier architectures: CNN or LSTM-RNN. The CNN-based classifier BIBREF24 has convolutional filters of size 3, 4, 5 and word embeddings. All outputs of each filter are concatenated before applied with a max-pooling over time, then fed into a two-layer feed-forward network with ReLU, followed by the softmax function. An RNN-based classifier has a single layer LSTM and word embeddings, whose output is fed into an output affine layer with the softmax function. For both the architectures, dropout BIBREF30 and Adam optimization BIBREF31 are applied during training. The train process is finish by early stopping with validation at each epoch. Sentence classifier hyper-parameters including learning rate, embedding dimension, unit or filter size, and dropout ratio, are selected using grid-search for each task-specific dataset. We refer to Kobayashi's implementation in the released code. For BERT, all hyper-parameters are kept the same as Devlin BIBREF11 , codes in Tensorflow and PyTorch are all available on github and pre-trained BERT model can also be downloaded. The number of conditional BERT training epochs ranges in [1-50] and number of masked words ranges in [1-2]. We compare the performance improvements obtained by our proposed method with the following baseline methods, “w/" means “with": w/synonym: Words are randomly replaced with synonyms from WordNet BIBREF5 . w/context: Proposed by Kobayashi BIBREF6 , which used a bidirectional language model to apply contextual augmentation, each word was replaced with a probability. w/context+label: Kobayashi’s contextual augmentation method BIBREF6 in a label-conditional LM architecture. Table 2 lists the accuracies of the all methods on two classifier architectures. The results show that, for various datasets on different classifier architectures, our conditional BERT contextual augmentation improves the model performances most. BERT can also augments sentences to some extent, but not as much as conditional BERT does. For we masked words randomly, the masked words may be label-sensitive or label-insensitive. If label-insensitive words are masked, words predicted through BERT may not be compatible with original labels. The improvement over all benchmark datasets also shows that conditional BERT is a general augmentation method for multi-labels sentence classification tasks. We also explore the effect of number of training steps to the performance of conditional BERT data augmentation. The fine-tuning epoch setting ranges in [1-50], we list the fine-tuning epoch of conditional BERT to outperform BERT for various benchmarks in table 3 . The results show that our conditional BERT contextual augmentation can achieve obvious performance improvement after only a few fine-tuning epochs, which is very convenient to apply to downstream tasks. Connection to Style Transfer In this section, we further deep into the connection to style transfer and apply our well trained conditional BERT to style transfer task. Style transfer is defined as the task of rephrasing the text to contain specific stylistic properties without changing the intent or affect within the context BIBREF32 . Our conditional MLM task changes words in the text condition on given label without changing the context. View from this point, the two tasks are very close. So in order to apply conditional BERT to style transfer task, given a specific stylistic sentence, we break it into two steps: first, we find the words relevant to the style; second, we mask the style-relevant words, then use conditional BERT to predict new substitutes with sentence context and target style property. In order to find style-relevant words in a sentence, we refer to Xu BIBREF33 , which proposed an attention-based method to extract the contribution of each word to the sentence sentimental label. For example, given a positive movie remark “This movie is funny and interesting", we filter out the words contributes largely to the label and mask them. Then through our conditional BERT contextual augmentation method, we fill in the masked position by predicting words conditioning on opposite label and sentence context, resulting in “This movie is boring and dull". The words “boring" and “dull" contribute to the new sentence being labeled as negative style. We sample some sentences from dataset SST2, transferring them to the opposite label, as listed in table 4 . Conclusions and Future Work In this paper, we fine-tune BERT to conditional BERT by introducing a novel conditional MLM task. After being well trained, the conditional BERT can be applied to data augmentation for sentence classification tasks. Experiment results show that our model outperforms several baseline methods obviously. Furthermore, we demonstrate that our conditional BERT can also be applied to style transfer task. In the future, (1)We will explore how to perform text data augmentation on imbalanced datasets with pre-trained language model, (2) we believe the idea of conditional BERT contextual augmentation is universal and will be applied to paragraph or document level data augmentation.

Yes

a12a08099e8193ff2833f79ecf70acf132eda646

a12a08099e8193ff2833f79ecf70acf132eda646_0

Q: Do they cover data augmentation papers? Text: Introduction Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions. QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 . While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions. Fundamental Aspects of NQG For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research. Learning Paradigm QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility. Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates. In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates. However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" . Input Modality Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs. Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" . Cognitive Levels Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 . Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input. Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead. The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models. Corpora As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly. Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research. Evaluation Metrics Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks. As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet. To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking. Methodology Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ : $$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\ \vspace{-14.22636pt} & = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5) BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 . Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next. Encoding Answers The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 . The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD. To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder. We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training. Question Word Generation Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design. BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step. Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated. Paragraph-level Contexts Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information. To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations. Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors. Answer-unaware QG The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer. However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study. Technical Considerations Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics: 1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own. 2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder. 3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question. The State of the Art In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art. Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts. Emerging Trends We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation. Multi-task Learning As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve. Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer. Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration. Wider Input Modalities QG work now has incorporated input from knowledge bases (KBQG) and images (VQG). Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples. Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types. In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 . Generation of Deep Questions Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels. To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information. Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer. Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking. Conclusion – What's the Outlook? We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation. What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored. When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead. Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.

No

999b20dc14cb3d389d9e3ba5466bc3869d2d6190

999b20dc14cb3d389d9e3ba5466bc3869d2d6190_0

Q: What is the latest paper covered by this survey? Text: Introduction Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions. QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 . While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions. Fundamental Aspects of NQG For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research. Learning Paradigm QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility. Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates. In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates. However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" . Input Modality Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs. Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" . Cognitive Levels Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 . Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input. Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead. The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models. Corpora As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly. Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research. Evaluation Metrics Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks. As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet. To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking. Methodology Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ : $$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\ \vspace{-14.22636pt} & = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5) BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 . Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next. Encoding Answers The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 . The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD. To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder. We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training. Question Word Generation Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design. BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step. Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated. Paragraph-level Contexts Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information. To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations. Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors. Answer-unaware QG The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer. However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study. Technical Considerations Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics: 1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own. 2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder. 3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question. The State of the Art In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art. Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts. Emerging Trends We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation. Multi-task Learning As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve. Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer. Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration. Wider Input Modalities QG work now has incorporated input from knowledge bases (KBQG) and images (VQG). Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples. Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types. In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 . Generation of Deep Questions Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels. To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information. Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer. Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking. Conclusion – What's the Outlook? We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation. What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored. When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead. Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.

Kim et al. (2019)

ca4b66ffa4581f9491442dcec78ca556253c8146

ca4b66ffa4581f9491442dcec78ca556253c8146_0

Q: Do they survey visual question generation work? Text: Introduction Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions. QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 . While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions. Fundamental Aspects of NQG For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research. Learning Paradigm QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility. Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates. In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates. However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" . Input Modality Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs. Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" . Cognitive Levels Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 . Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input. Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead. The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models. Corpora As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly. Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research. Evaluation Metrics Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks. As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet. To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking. Methodology Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ : $$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\ \vspace{-14.22636pt} & = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5) BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 . Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next. Encoding Answers The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 . The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD. To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder. We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training. Question Word Generation Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design. BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step. Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated. Paragraph-level Contexts Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information. To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations. Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors. Answer-unaware QG The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer. However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study. Technical Considerations Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics: 1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own. 2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder. 3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question. The State of the Art In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art. Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts. Emerging Trends We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation. Multi-task Learning As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve. Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer. Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration. Wider Input Modalities QG work now has incorporated input from knowledge bases (KBQG) and images (VQG). Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples. Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types. In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 . Generation of Deep Questions Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels. To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information. Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer. Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking. Conclusion – What's the Outlook? We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation. What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored. When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead. Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.

Yes

b3ff166bd480048e099d09ba4a96e2e32b42422b

b3ff166bd480048e099d09ba4a96e2e32b42422b_0

Q: Do they survey multilingual aspects? Text: Introduction Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions. QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 . While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions. Fundamental Aspects of NQG For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research. Learning Paradigm QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility. Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates. In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates. However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" . Input Modality Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs. Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" . Cognitive Levels Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 . Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input. Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead. The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models. Corpora As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly. Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research. Evaluation Metrics Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks. As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet. To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking. Methodology Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ : $$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\ \vspace{-14.22636pt} & = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5) BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 . Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next. Encoding Answers The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 . The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD. To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder. We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training. Question Word Generation Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design. BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step. Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated. Paragraph-level Contexts Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information. To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations. Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors. Answer-unaware QG The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer. However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study. Technical Considerations Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics: 1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own. 2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder. 3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question. The State of the Art In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art. Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts. Emerging Trends We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation. Multi-task Learning As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve. Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer. Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration. Wider Input Modalities QG work now has incorporated input from knowledge bases (KBQG) and images (VQG). Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples. Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types. In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 . Generation of Deep Questions Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels. To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information. Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer. Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking. Conclusion – What's the Outlook? We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation. What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored. When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead. Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.

No

3703433d434f1913307ceb6a8cfb9a07842667dd

3703433d434f1913307ceb6a8cfb9a07842667dd_0

Q: What learning paradigms do they cover in this survey? Text: Introduction Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions. QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 . While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions. Fundamental Aspects of NQG For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research. Learning Paradigm QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility. Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates. In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates. However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" . Input Modality Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs. Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" . Cognitive Levels Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 . Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input. Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead. The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models. Corpora As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly. Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research. Evaluation Metrics Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks. As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet. To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking. Methodology Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ : $$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\ \vspace{-14.22636pt} & = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5) BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 . Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next. Encoding Answers The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 . The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD. To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder. We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training. Question Word Generation Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design. BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step. Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated. Paragraph-level Contexts Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information. To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations. Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors. Answer-unaware QG The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer. However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study. Technical Considerations Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics: 1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own. 2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder. 3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question. The State of the Art In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art. Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts. Emerging Trends We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation. Multi-task Learning As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve. Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer. Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration. Wider Input Modalities QG work now has incorporated input from knowledge bases (KBQG) and images (VQG). Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples. Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types. In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 . Generation of Deep Questions Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels. To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information. Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer. Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking. Conclusion – What's the Outlook? We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation. What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored. When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead. Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.

Considering "What" and "How" separately versus jointly optimizing for both.

f7c34b128f8919e658ba4d5f1f3fc604fb7ff793

f7c34b128f8919e658ba4d5f1f3fc604fb7ff793_0

Q: What are all the input modalities considered in prior work in question generation? Text: Introduction Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions. QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 . While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions. Fundamental Aspects of NQG For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research. Learning Paradigm QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility. Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates. In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates. However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" . Input Modality Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs. Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" . Cognitive Levels Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 . Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input. Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead. The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models. Corpora As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly. Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research. Evaluation Metrics Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks. As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet. To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking. Methodology Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ : $$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\ \vspace{-14.22636pt} & = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5) BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 . Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next. Encoding Answers The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 . The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD. To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder. We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training. Question Word Generation Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design. BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step. Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated. Paragraph-level Contexts Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information. To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations. Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors. Answer-unaware QG The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer. However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study. Technical Considerations Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics: 1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own. 2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder. 3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question. The State of the Art In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art. Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts. Emerging Trends We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation. Multi-task Learning As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve. Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer. Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration. Wider Input Modalities QG work now has incorporated input from knowledge bases (KBQG) and images (VQG). Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples. Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types. In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 . Generation of Deep Questions Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels. To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information. Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer. Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking. Conclusion – What's the Outlook? We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation. What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored. When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead. Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.

Textual inputs, knowledge bases, and images.

d42031893fd4ba5721c7d37e1acb1c8d229ffc21

d42031893fd4ba5721c7d37e1acb1c8d229ffc21_0

Q: Do they survey non-neural methods for question generation? Text: Introduction Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions. QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 . While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions. Fundamental Aspects of NQG For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research. Learning Paradigm QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility. Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates. In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates. However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" . Input Modality Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs. Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" . Cognitive Levels Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 . Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input. Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead. The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models. Corpora As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly. Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research. Evaluation Metrics Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks. As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet. To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking. Methodology Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ : $$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\ \vspace{-14.22636pt} & = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5) BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 . Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next. Encoding Answers The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 . The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD. To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder. We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training. Question Word Generation Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design. BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step. Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated. Paragraph-level Contexts Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information. To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations. Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors. Answer-unaware QG The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer. However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study. Technical Considerations Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics: 1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own. 2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder. 3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question. The State of the Art In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art. Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts. Emerging Trends We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation. Multi-task Learning As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve. Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer. Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration. Wider Input Modalities QG work now has incorporated input from knowledge bases (KBQG) and images (VQG). Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples. Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types. In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 . Generation of Deep Questions Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels. To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information. Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer. Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking. Conclusion – What's the Outlook? We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation. What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored. When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead. Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.

Unanswerable

a999761aa976458bbc7b4f330764796446d030ff

a999761aa976458bbc7b4f330764796446d030ff_0

Q: What is their model? Text: Introduction Named Entity Recognition is a major natural language processing task that recognizes the proper labels such as LOC (Location), PER (Person), ORG (Organization), etc. Like words or phrase, being a sort of language constituent, named entities also benefit from better representation for better processing. Continuous word representations, known as word embeddings, well capture semantic and syntactic regularities of words BIBREF0 and perform well in monolingual NE recognition BIBREF1 , BIBREF2 . Word embeddings also exhibit isomorphism structure across languages BIBREF3 . On account of these characteristics above, we attempt to utilize word embeddings to improve NE recognition for resource-poor languages with the help of richer ones. The state-of-the-art cross-lingual NE recognition methods are mainly based on annotation projection methods according to parallel corpora, translations BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 and Wikipedia methods BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 . Most annotated corpus based NE recognition tasks can benefit a great deal from a known NE dictionary, as NEs are those words which carry common sense knowledge quite differ from the rest ones in any language vocabulary. This work will focus on the NE recognition from plain text instead of corpus based NE recognition. For a purpose of learning from limited annotated linguistic resources, our preliminary discovery shows that it is possible to build a geometric space projection between embedding spaces to help cross-lingual NE recognition. Our study contains two main steps: First, we explore the NE distribution in monolingual case. Next, we learn a hypersphere mapping between embedding spaces of languages with minimal supervision. Despite the simplicity of our model, we make the following contributions. First, for word embeddings generated by different dimensions and objective functions, all common NE types (PER, LOC, ORG) tend to be densely distributed in a hypersphere, which gives a better solution to characterize the general NE distribution rather than existing closed dictionary definition for NE. Second, with the help of the hypersphere mapping, it is possible to capture the NE distribution of resource-poor languages with only a small amount of annotated data. Third, our method is highly friendly to unregistered NEs, as the distance to each hypersphere center is the only factor needed to determine their NE categories. Finally, by adding hypersphere features we can significantly improve the performance of off-the-shelf named entity recognition (NER) systems. Word Embeddings Seok BIBREF2 proposed that similar words are more likely to occupy close spatial positions, since their word embeddings carries syntactical and semantical informative clues. For an intuitive understanding, they listed the nearest neighbors of words included in the PER and ORG tags under cosine similarity metric. To empirically verify this observation and explore the performance of this property in Euclidean space , we list Top-5 nearest neighbors under Euclidean distance metric in Table 1 and illustrate a standard t-SNE BIBREF12 2- $D$ projection of the embeddings of three entity types with a sample of 500 words for each type. Nearest neighbors are calculated by comparing the Euclidean distance between the embedding of each word (such as Fohnsdorf, Belgian, and Ltd.) and the embeddings of all other words in the vocabulary. We pre-train word embeddings using the continuous skip-gram model BIBREF13 with the tool, and obtain multi-word and single-word phrases with a maximum length of 8, and a minimum word frequency cutoff of 3. The examples in Table 1 and visualization in Figure 1 demonstrate that the above observation suits well under Euclidean distance metric for NE recognition either for monolingual or multilingual situations. Model Encouraged by the verification of nearest neighbors of NEs still being NEs, we attempt to build a model which can represent this property with least parameters. Namely, given an NE dictionary on a monolingual, we build a model to describe the distribution of the word embeddings of these entities, then we can easily use these parameters as a decoder for any word to directly determine whether it belongs to a certain type of entity. In this section, we first introduce the open modeling from embedding distribution in monolingual cases, and then put forward the mapping of the distribution model between languages, and then use the mapping to build named entity dataset for resource-poor languages. Finally, we use the proposed named entity model to improve the performance of state-of-the-art NE recognition systems. Open Monolingual NE Modeling As illustrated is Figure 1, the embedding distribution of NEs is aggregated, and there exists a certain boundary between different types of NEs. We construct an open representation for each type of NEs – hypersphere, the NE type of any entity can be easily judged by checking whether it is inside a hypersphere, which makes a difference from the defining way of any limited and insufficient NE dictionary. The hypersphere can be expressed as follows: $$E( X, O) \le r$$ (Eq. 9) where E represents the adopted Euclidean distance, X is referred to any point in the hypersphere, $ O $ and $ r $ are the center vector and radius. For each entity type, we attempt to construct a hypersphere which encompass as many congeneric NEs as possible, and as few as possible inhomogeneous NEs, we use $F_1$ score as a trade-off between these two concerns. We carefully tune the center and radius of the hypersphere to maximize its $F_1$ score: we first fix the center as the average of all NE embeddings from known NE dictionaries, and search the best radius in $[minDist, maxDist]$ , where $minDist/maxDist$ refers to the distance between the center and its nearest/farthest neighbors; Then, we kick NEs which are far from the center with the distance threshold $q$ (much larger than the radius) to generate a new center; Finally, we tune the threshold $q$ and repeat the above steps to find the most suitable center and radius. The mathematical intuition for using a hypersphere can be interpreted in a manner similar to support vector machine (SVM) BIBREF14 , which uses the kernel to obtain the optimal margin in very high dimensional spaces through linear hyperplane separation in Descartes coordination. We transfer the idea to the separation of NE distributions. The only difference is about boundary shape, what we need is a closed surface instead of an open hyperplane, and hypersphere is such a smooth, closed boundary (with least parameters as well) in polar coordinates as counterpart of hyperplane in Descartes coordinates. Using the least principle to model the mathematical objective also follows the Occam razor principle. Figure 1 also reveals that the distribution of PER NEs is compact, while ORG NE distribution is relatively sparse. Syntactically, PER NEs are more stable in terms of position and length in sentences compared to ORG NEs, so that they have a more accurate embedding representation with strong strong syntax and semantics, making the corresponding word embeddings closer to central region of the hypersphere. Embedding Distribution Mapping As the isomorphism characteristic exists between languages BIBREF3 , BIBREF15 , we can apply the distributional modeling for every languages in the same way. For a target language without an NE dictionary, its NE distribution can be obtained from a source language with known NE distributions by learning the transforming function between these two languages. We construct the transformation matrix $W$ via a set of parallel word pairs (the set will be referred to seed pairs hereafter) and their word embeddings $\lbrace X^{(i)}, Z^{(i)}\rbrace _{i=1}^m$ BIBREF3 , $\lbrace X^{(i)}\rbrace _{i=1}^m$ , $\lbrace Z^{(i)}\rbrace _{i=1}^m$ are the source and target word embeddings respectively. $W$ can be learned by solving the matrix equation $XW = Z$ . Then, given the source center vector ${ O_1}$ , the mapping center vector ${O_2}$ can be expressed as: $${ O_2} = W^T{O_1}$$ (Eq. 11) Actually, the isomorphism (mapping) between embedding spaces is the type of affine isomorphism by furthermore considering embedding in continuous space. The invariant characteristics of relative position BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 in affine transformation is applied to correct transformation matrix errors caused by limited amount of parallel word pairs (the set will be referred to seed pairs hereafter). As shown in Figure 2, the ratio of the line segments keep constant when the distance is linearly enlarged or shortened. Recall that point $Q$ is an affine combination of two other noncoincident points $Q_1$ and $Q_2$ on the line: $Q = (1-t)Q_1 + tQ_2 $ . We apply the affine mapping $f$ and get: $f(Q) = f((1-t)Q_1 + tQ_2) = (1-t)f(Q_1) + tf(Q_2)$ Obviously, the constant ratio $t$ is not affected by the affine transformation $f$ . That is, $Q$ has the same relative distances between it and $Q_1$ and $Q_2$ during the process of transformation. Based on the above characteristic, for any point $X^{(i)}$ in the source space and its mapping point $Z^{(i)}$ , $X^{(i)}$ and $f(Q) = f((1-t)Q_1 + tQ_2) = (1-t)f(Q_1) + tf(Q_2)$0 cut off radiuses with the same ratio, namely, the ratio of the distance of these two points to their centers and their radiuses remains unchanged. $$\frac{E( O_1, X^{(i)})}{r_1} = \frac{E( O_2, Z^{(i)})}{r_2}$$ (Eq. 15) where $E$ represents the adopted Euclidean distance, ${O_1, O_2, r_1, r_2}$ are the centers and radii of hyperspheres. We convert the equation and learn the optimized mapping center ${O_2}$ and ratio $K$ via the seed pairs: $${K = \frac{r_2}{r_1} = \frac{E( O_2, Z^{(i)})}{E( O_1, X^{(i)})}}$$ (Eq. 16) $$\begin{aligned} E( O_2, Z^{(i)}) &= K * E( O_1, X^{(i)}) \quad r_2 &= K * r_1 \\ \end{aligned}$$ (Eq. 17) Given the seed pairs $\lbrace X^{(i)}, Z^{(i)}\rbrace _{i=1}^m$ , the initialized center $O_2$ in Equation (3), the center $ O_1 $ and radius $ r_1 $ of the hypersphere in source language space, we may work out the optimized ratio $K$ , the mapping center $ O_2 $ and radius $ r_2 $ in target language space by solving the linear equation group (5). Hypersphere features for NE Recognition The Euclidean distance between word and hypersphere centers can be pre-computed as its NE likelihood, which may provide informative clues for NE recognition. We only consider three entity types in our experiment, and the Euclidean distance which is represented as a 3- $D$ vector and referred to HS vector hereafter) is added to four representative off-the-shelf NER systems to verify its effectiveness. We feed HS vector into different layers of the neural network: (1) input layer $[x_k; c_k; HS]$ ; (2) output layer of LSTM $[h_k; HS]$ , where $x_k$ , $w_k$ and $h_k$ represent word embeddings, char embeddings and the output of LSTM, respectively. All of these models are based on classical BiLSTM-CRF architecture BIBREF20 , except that BIBREF21 replaces CRF layer with softmax. These four baseline systems are introduced as follows. BIBREF22 concatenates ELMo with word embeddings as the input of LSTM to enhance word representations as it carries both syntactic and semantic information. BIBREF21 uses distant supervision for NER task and propose a new Tie or Break tagging scheme, where entity spans and entity types are encoded into two folds. They first build a binary classifier to distinguish Break from Tie, and then learn the entity types according to their occurrence and frequency in NE dictionary. The authors conduct their experiments on biomedical datasets rather than standard benchmark, so we extract the NEs in training data as the domain-specific dictionary. This work creates a promising prospect for using dictionary to replace the role of training data. BIBREF23 takes advantage of the power of the 120 entity types from annotated data in Wikipedia. Cosine similarity between the word embedding and the embedding of each entity type is concatenated as the 120- $D$ feature vector (which is called LS vector in their paper) and then fed into the input layer of LSTM. Lexical feature has been shown a key factor to NE recognition. BIBREF24 passes sentences as sequences of characters into a character-level language model to produce a novel type of word embedding, contextual string embeddings, where one word may have different embeddings as the embeddings are computed both on the characters of a word and its surrounding context. Such embeddings are then fed into the input layer of LSTM. Experiment In this section, we evaluate the hypersphere model based on the three models introduced above: open monolingual NE modeling, embedding distribution mapping and refinement NE recognition. Setup In this experiment, we adopt pre-trained word embeddings from Wikipedia corpus. Our preliminary experiments will be conducted on English and Chinese. For the former, we use NLTK toolkit and LANGID toolkit to perform the pre-processing. For the latter, we first use OpenCC to simplify characters, and then use THULAC to perform word segmentation. In order to make the experimental results more accurate and credible, we manually annotate two large enough Chinese and English NE dictionaries for training and test. Table 2 lists the statistics of Wikipedia corpus and the annotated data. Our dictionary contains many multi-word NEs in LOC and ORG types as accounted in the second column for each language in Table 2, while we only include single-word PER NEs in our dictionary, since the English first name and last name are separated, and Chinese word segmentation cuts most of the PER entities together. We pre-train quality multi-word and single-word embeddings and aim to maximize the coverage of the NEs in the dictionary. The pre-trained word embeddings cover 82.3% / 82.51% of LOC NEs and 70.2% / 63.61% of ORG NEs in English and Chinese, respectively. For other multi-word NEs, we simply calculate the average vector of each word embedding as their representations. Monolingual Embedding Distribution The NE distribution is closely correlated to the dimension of the embedding space, we train the word embeddings from 2- $D$ to 300- $D$ and search for the most suitable dimension for each NE type. For each dimension, we carefully tune the center and radius of the hypersphere using the method introduced in section 3.1 for maximize $F_1$ score, and select the dimension with maximize $F_1$ score. The most suitable dimension for ORG, PER, LOC are 16- ${D}$ /16- ${D}$ /24- ${D}$ (these dimensions will be used as parameters in the following experiments), respectively . We discover that in low-dimensional space, the distributions of NEs are better. In high dimensions, the curse of dimension could be the main reason to limit the performance. Table 3 lists the final maximum $F_1$ score of three NE types. The results of the three types of NE are almost 50%, and PER type performs best. The main factor may be that PER NEs are represented as single-word in our dictionary, and word embeddings can better represents their meanings. The result also states that better representations for multi-word NEs which are not covered by the dictionary instead of the average of each word may help bring better results. Besides, the incompleteness of NE dictionaries and noises during pre-processing may cause a decrease on the performance. Overall, hypersphere model has shown been effectively used as the open modeling for NEs. Hypersphere Mapping The following preparations were made for the mapping: $(i)$ A large enough NE dictionary in source (resource-rich) corpus; $(ii)$ A small amount of annotated seed pairs. We use $s$ to represent the number of seed pairs and $d$ to represent the number of unknown variables. With seed pair size $s < d$ , the matrix can be solved with much loose constraints, and $F_1$ score remarkably increases with more seed pairs. Once $s > d$ , the linear equation group will be always determined by strong enough constraints, which leads to a stable solution. Based on the characteristics, we only take two dozen of seed pairs on each type in following experiments. We combine human translation and online translation together for double verification for this small set of seed pairs. In this part, we utilize English and Chinese as the corpus of known NEs in turn, and predict the NE distribution of the other language. Evaluation In order to quantitatively represent the mapping effect, we present a new evaluation method to judge the hypersphere mapping between English and Chinese: $$\begin{aligned} P = \frac{V_i}{V_m} \quad R = \frac{V_i}{V_t} \quad F_1 = \frac{2 * P * R}{P + R} \end{aligned}$$ (Eq. 29) where ${V_t, V_m, V_i}$ represent the volumes of the target, mapping and intersection hyperspheres. Due to the difficulty of calculating the volume of hyperspheres in high dimensions, we adopt Monte Carlo methods to simulate the volume BIBREF25 . we generate a great quantity of points in the embedding spaces, and take the amount of the points falling in each hypersphere as its volume. Mapping between English and Chinese Table 4 shows the comparisons of cross-lingual named entity extraction performance. We use the unsupervised method proposed in BIBREF26 to generate cross-lingual embeddings. $k$ -NN and SVM are the same as monolingual cases in Table 3 except for the training set. $k$ -NN $_{150}$ and SVM $_{150}$ use 20% of the NEs in source language and 150 NEs (50 LOC, PER and ORG) in target language for training, while $k$ -NN $_{2500}$ and SVM $_{2500}$ use 20% of the NEs in source language and 2500 NEs (1000 LOC and PER, 500 ORG) in target language. $k$ -NN and SVM depend much on the annotated training set, requiring more than $1K$ training samples to provide a performance as our model offers. Due to the instability of ORG type in length, taking the average of each word embedding may disobey the syntactic and semantic regularities of ORG NEs, thereby undermines the multilingual isomorphism characteristics, which causes the inferior performance of our model on this type of NEs. This suggests that build better representations NEs for multi-word NEs may contribute to a better performance in our model. Mapping to truly Low-resource Language We build named entity dataset for a truly resource-poor language, Indonesian, and manually examine the nearest words to the hypersphere center for 'gold-standard' evaluation. We take English as the source language, the settings of the dimension $D$ and the number of seed pairs $s$ are the same as the above experiments between Chinese and English. From the results listed in Table 5, we can see that even the precision of the top-100 NEs are 0.350 $F_1$ /0.440 $F_1$ /0.310 $F_1$ , respectively, which proves the this distribution can indeed serves as a candidate NE dictionary for Indonesian. [9] The authors of BIBREF24 publish an updated results (92.98) on CoNLL-2003 dataset in https://github.com/zalandoresearch/flair/issues/206 on their 0.3.2 version, and this is the best result at our most try. [10] This is the reported state-of-the-art result in their github. [11]We use the same parameters as the authors release in https://github.com/zalandoresearch/flair/issues/173 and obtain the result of 89.45 on ONTONOTES 5.0 dataset. Off-the-shelf NE Recognition Systems To evaluate the influence of our hypersphere feature for off-the-shelf NER systems, we perform the NE recognition on two standard NER benchmark datasets, CoNLL2003 and ONTONOTES 5.0. Our results in Table 6 and Table 7 demonstrate the power of hypersphere features, which contribute to nearly all of the three types of entities as shown in Table 6, except for a slight drop in the PER type of BIBREF22 on a strong baseline. HS features stably enhance all strong state-of-the-art baselines, BIBREF22 , BIBREF21 and BIBREF23 by 0.33/0.72/0.23 $F_1$ point and 0.13/0.3/0.1 $F_1$ point on both benchmark datasets, CoNLL-2003 and ONTONOTES 5.0. We show that our HS feature is also comparable with previous much more complicated LS feature, and our model surpasses their baseline (without LS feature) by 0.58/0.78 $F_1$ point with only HS features. We establish a new state-of-the-art $F_1$ score of 89.75 on ONTONOTES 5.0, while matching state-of-the-art performance with a $F_1$ score of 92.95 on CoNLL-2003 dataset. Related Work In recent years, word embeddings have also been used as a feature to enhance the NE recognition, with the revealing of linguistic features in morphological, syntactic and semantic perspective. BIBREF1 clustered the word embeddings and combined multiple cluster granularities to improve the NE recognition performance. Our work likewise use word embeddings to help NE recognition, we make use of the characteristic that syntactically and semantically s are more likely to be neighbors in embedding spaces and construct a hypersphere model to encompass NEs. Cross-lingual knowledge transfer is a highly promising work for resource-poor languages, annotation projection and representation projection are widely used in NE recognition BIBREF27 , BIBREF5 , BIBREF4 , BIBREF28 , BIBREF29 , BIBREF30 . These works put forward inconvenient requirements for parallel or comparable corpora, a large amount of annotated or translation data or bilingual lexicon. Different from any existing work to the best of our knowledge, this is the first work that merely uses isomorphic mappings in low-dimensional embedding spaces to recognize NEs, and we introduce a mathematically simple model to describe NE embedding distribution from visualization results, showing it works for both monolingual and cross-lingual situations. Conclusion Named entities being an open set which keeps expanding are difficult to represent through a closed NE dictionary. This work mitigates significant defects in previous closed NE definitions and proposes a new open definition for NEs by modeling their embedding distributions with least parameters. We visualize NE distributions in monolingual case and perform an effective isomorphism spaces mapping in cross-lingual case. According to our work, we demonstrate that common named entity types (PER, LOC, ORG) tend to be densely distributed in a hypersphere and it is possible to build a mapping between the NE distributions in embedding spaces to help cross-lingual NE recognition. Experimental results show that the distribution of named entities via mapping can be used as a good enough replacement for the original distribution. Then the discovery is used to build an NE dictionary for Indonesian being a truly low-resource language, which also gives satisfactory precision. Finally, our simple hypersphere features being the representation of NE likelihood can be used for enhancing off-the-shelf NER systems by concatenating with word embeddings and the output of BiLSTM in the input layer and encode layer, respectively, and we achieve a new state-of-the-art $F_1$ score of 89.75 on ONTONOTES 5.0 benchmark. In this work, we also give a better solution for unregistered NEs. For any newly emerged NE together with its embedding, in case we obtain the hypersphere of each named entity, the corresponding named entity category can be determined by calculating the distance between its word embedding and the center of each hypersphere.

cross-lingual NE recognition

f229069bcb05c2e811e4786c89b0208af90d9a25

f229069bcb05c2e811e4786c89b0208af90d9a25_0

Q: Do they evaluate on NER data sets? Text: Introduction Named Entity Recognition is a major natural language processing task that recognizes the proper labels such as LOC (Location), PER (Person), ORG (Organization), etc. Like words or phrase, being a sort of language constituent, named entities also benefit from better representation for better processing. Continuous word representations, known as word embeddings, well capture semantic and syntactic regularities of words BIBREF0 and perform well in monolingual NE recognition BIBREF1 , BIBREF2 . Word embeddings also exhibit isomorphism structure across languages BIBREF3 . On account of these characteristics above, we attempt to utilize word embeddings to improve NE recognition for resource-poor languages with the help of richer ones. The state-of-the-art cross-lingual NE recognition methods are mainly based on annotation projection methods according to parallel corpora, translations BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 and Wikipedia methods BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 . Most annotated corpus based NE recognition tasks can benefit a great deal from a known NE dictionary, as NEs are those words which carry common sense knowledge quite differ from the rest ones in any language vocabulary. This work will focus on the NE recognition from plain text instead of corpus based NE recognition. For a purpose of learning from limited annotated linguistic resources, our preliminary discovery shows that it is possible to build a geometric space projection between embedding spaces to help cross-lingual NE recognition. Our study contains two main steps: First, we explore the NE distribution in monolingual case. Next, we learn a hypersphere mapping between embedding spaces of languages with minimal supervision. Despite the simplicity of our model, we make the following contributions. First, for word embeddings generated by different dimensions and objective functions, all common NE types (PER, LOC, ORG) tend to be densely distributed in a hypersphere, which gives a better solution to characterize the general NE distribution rather than existing closed dictionary definition for NE. Second, with the help of the hypersphere mapping, it is possible to capture the NE distribution of resource-poor languages with only a small amount of annotated data. Third, our method is highly friendly to unregistered NEs, as the distance to each hypersphere center is the only factor needed to determine their NE categories. Finally, by adding hypersphere features we can significantly improve the performance of off-the-shelf named entity recognition (NER) systems. Word Embeddings Seok BIBREF2 proposed that similar words are more likely to occupy close spatial positions, since their word embeddings carries syntactical and semantical informative clues. For an intuitive understanding, they listed the nearest neighbors of words included in the PER and ORG tags under cosine similarity metric. To empirically verify this observation and explore the performance of this property in Euclidean space , we list Top-5 nearest neighbors under Euclidean distance metric in Table 1 and illustrate a standard t-SNE BIBREF12 2- $D$ projection of the embeddings of three entity types with a sample of 500 words for each type. Nearest neighbors are calculated by comparing the Euclidean distance between the embedding of each word (such as Fohnsdorf, Belgian, and Ltd.) and the embeddings of all other words in the vocabulary. We pre-train word embeddings using the continuous skip-gram model BIBREF13 with the tool, and obtain multi-word and single-word phrases with a maximum length of 8, and a minimum word frequency cutoff of 3. The examples in Table 1 and visualization in Figure 1 demonstrate that the above observation suits well under Euclidean distance metric for NE recognition either for monolingual or multilingual situations. Model Encouraged by the verification of nearest neighbors of NEs still being NEs, we attempt to build a model which can represent this property with least parameters. Namely, given an NE dictionary on a monolingual, we build a model to describe the distribution of the word embeddings of these entities, then we can easily use these parameters as a decoder for any word to directly determine whether it belongs to a certain type of entity. In this section, we first introduce the open modeling from embedding distribution in monolingual cases, and then put forward the mapping of the distribution model between languages, and then use the mapping to build named entity dataset for resource-poor languages. Finally, we use the proposed named entity model to improve the performance of state-of-the-art NE recognition systems. Open Monolingual NE Modeling As illustrated is Figure 1, the embedding distribution of NEs is aggregated, and there exists a certain boundary between different types of NEs. We construct an open representation for each type of NEs – hypersphere, the NE type of any entity can be easily judged by checking whether it is inside a hypersphere, which makes a difference from the defining way of any limited and insufficient NE dictionary. The hypersphere can be expressed as follows: $$E( X, O) \le r$$ (Eq. 9) where E represents the adopted Euclidean distance, X is referred to any point in the hypersphere, $ O $ and $ r $ are the center vector and radius. For each entity type, we attempt to construct a hypersphere which encompass as many congeneric NEs as possible, and as few as possible inhomogeneous NEs, we use $F_1$ score as a trade-off between these two concerns. We carefully tune the center and radius of the hypersphere to maximize its $F_1$ score: we first fix the center as the average of all NE embeddings from known NE dictionaries, and search the best radius in $[minDist, maxDist]$ , where $minDist/maxDist$ refers to the distance between the center and its nearest/farthest neighbors; Then, we kick NEs which are far from the center with the distance threshold $q$ (much larger than the radius) to generate a new center; Finally, we tune the threshold $q$ and repeat the above steps to find the most suitable center and radius. The mathematical intuition for using a hypersphere can be interpreted in a manner similar to support vector machine (SVM) BIBREF14 , which uses the kernel to obtain the optimal margin in very high dimensional spaces through linear hyperplane separation in Descartes coordination. We transfer the idea to the separation of NE distributions. The only difference is about boundary shape, what we need is a closed surface instead of an open hyperplane, and hypersphere is such a smooth, closed boundary (with least parameters as well) in polar coordinates as counterpart of hyperplane in Descartes coordinates. Using the least principle to model the mathematical objective also follows the Occam razor principle. Figure 1 also reveals that the distribution of PER NEs is compact, while ORG NE distribution is relatively sparse. Syntactically, PER NEs are more stable in terms of position and length in sentences compared to ORG NEs, so that they have a more accurate embedding representation with strong strong syntax and semantics, making the corresponding word embeddings closer to central region of the hypersphere. Embedding Distribution Mapping As the isomorphism characteristic exists between languages BIBREF3 , BIBREF15 , we can apply the distributional modeling for every languages in the same way. For a target language without an NE dictionary, its NE distribution can be obtained from a source language with known NE distributions by learning the transforming function between these two languages. We construct the transformation matrix $W$ via a set of parallel word pairs (the set will be referred to seed pairs hereafter) and their word embeddings $\lbrace X^{(i)}, Z^{(i)}\rbrace _{i=1}^m$ BIBREF3 , $\lbrace X^{(i)}\rbrace _{i=1}^m$ , $\lbrace Z^{(i)}\rbrace _{i=1}^m$ are the source and target word embeddings respectively. $W$ can be learned by solving the matrix equation $XW = Z$ . Then, given the source center vector ${ O_1}$ , the mapping center vector ${O_2}$ can be expressed as: $${ O_2} = W^T{O_1}$$ (Eq. 11) Actually, the isomorphism (mapping) between embedding spaces is the type of affine isomorphism by furthermore considering embedding in continuous space. The invariant characteristics of relative position BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 in affine transformation is applied to correct transformation matrix errors caused by limited amount of parallel word pairs (the set will be referred to seed pairs hereafter). As shown in Figure 2, the ratio of the line segments keep constant when the distance is linearly enlarged or shortened. Recall that point $Q$ is an affine combination of two other noncoincident points $Q_1$ and $Q_2$ on the line: $Q = (1-t)Q_1 + tQ_2 $ . We apply the affine mapping $f$ and get: $f(Q) = f((1-t)Q_1 + tQ_2) = (1-t)f(Q_1) + tf(Q_2)$ Obviously, the constant ratio $t$ is not affected by the affine transformation $f$ . That is, $Q$ has the same relative distances between it and $Q_1$ and $Q_2$ during the process of transformation. Based on the above characteristic, for any point $X^{(i)}$ in the source space and its mapping point $Z^{(i)}$ , $X^{(i)}$ and $f(Q) = f((1-t)Q_1 + tQ_2) = (1-t)f(Q_1) + tf(Q_2)$0 cut off radiuses with the same ratio, namely, the ratio of the distance of these two points to their centers and their radiuses remains unchanged. $$\frac{E( O_1, X^{(i)})}{r_1} = \frac{E( O_2, Z^{(i)})}{r_2}$$ (Eq. 15) where $E$ represents the adopted Euclidean distance, ${O_1, O_2, r_1, r_2}$ are the centers and radii of hyperspheres. We convert the equation and learn the optimized mapping center ${O_2}$ and ratio $K$ via the seed pairs: $${K = \frac{r_2}{r_1} = \frac{E( O_2, Z^{(i)})}{E( O_1, X^{(i)})}}$$ (Eq. 16) $$\begin{aligned} E( O_2, Z^{(i)}) &= K * E( O_1, X^{(i)}) \quad r_2 &= K * r_1 \\ \end{aligned}$$ (Eq. 17) Given the seed pairs $\lbrace X^{(i)}, Z^{(i)}\rbrace _{i=1}^m$ , the initialized center $O_2$ in Equation (3), the center $ O_1 $ and radius $ r_1 $ of the hypersphere in source language space, we may work out the optimized ratio $K$ , the mapping center $ O_2 $ and radius $ r_2 $ in target language space by solving the linear equation group (5). Hypersphere features for NE Recognition The Euclidean distance between word and hypersphere centers can be pre-computed as its NE likelihood, which may provide informative clues for NE recognition. We only consider three entity types in our experiment, and the Euclidean distance which is represented as a 3- $D$ vector and referred to HS vector hereafter) is added to four representative off-the-shelf NER systems to verify its effectiveness. We feed HS vector into different layers of the neural network: (1) input layer $[x_k; c_k; HS]$ ; (2) output layer of LSTM $[h_k; HS]$ , where $x_k$ , $w_k$ and $h_k$ represent word embeddings, char embeddings and the output of LSTM, respectively. All of these models are based on classical BiLSTM-CRF architecture BIBREF20 , except that BIBREF21 replaces CRF layer with softmax. These four baseline systems are introduced as follows. BIBREF22 concatenates ELMo with word embeddings as the input of LSTM to enhance word representations as it carries both syntactic and semantic information. BIBREF21 uses distant supervision for NER task and propose a new Tie or Break tagging scheme, where entity spans and entity types are encoded into two folds. They first build a binary classifier to distinguish Break from Tie, and then learn the entity types according to their occurrence and frequency in NE dictionary. The authors conduct their experiments on biomedical datasets rather than standard benchmark, so we extract the NEs in training data as the domain-specific dictionary. This work creates a promising prospect for using dictionary to replace the role of training data. BIBREF23 takes advantage of the power of the 120 entity types from annotated data in Wikipedia. Cosine similarity between the word embedding and the embedding of each entity type is concatenated as the 120- $D$ feature vector (which is called LS vector in their paper) and then fed into the input layer of LSTM. Lexical feature has been shown a key factor to NE recognition. BIBREF24 passes sentences as sequences of characters into a character-level language model to produce a novel type of word embedding, contextual string embeddings, where one word may have different embeddings as the embeddings are computed both on the characters of a word and its surrounding context. Such embeddings are then fed into the input layer of LSTM. Experiment In this section, we evaluate the hypersphere model based on the three models introduced above: open monolingual NE modeling, embedding distribution mapping and refinement NE recognition. Setup In this experiment, we adopt pre-trained word embeddings from Wikipedia corpus. Our preliminary experiments will be conducted on English and Chinese. For the former, we use NLTK toolkit and LANGID toolkit to perform the pre-processing. For the latter, we first use OpenCC to simplify characters, and then use THULAC to perform word segmentation. In order to make the experimental results more accurate and credible, we manually annotate two large enough Chinese and English NE dictionaries for training and test. Table 2 lists the statistics of Wikipedia corpus and the annotated data. Our dictionary contains many multi-word NEs in LOC and ORG types as accounted in the second column for each language in Table 2, while we only include single-word PER NEs in our dictionary, since the English first name and last name are separated, and Chinese word segmentation cuts most of the PER entities together. We pre-train quality multi-word and single-word embeddings and aim to maximize the coverage of the NEs in the dictionary. The pre-trained word embeddings cover 82.3% / 82.51% of LOC NEs and 70.2% / 63.61% of ORG NEs in English and Chinese, respectively. For other multi-word NEs, we simply calculate the average vector of each word embedding as their representations. Monolingual Embedding Distribution The NE distribution is closely correlated to the dimension of the embedding space, we train the word embeddings from 2- $D$ to 300- $D$ and search for the most suitable dimension for each NE type. For each dimension, we carefully tune the center and radius of the hypersphere using the method introduced in section 3.1 for maximize $F_1$ score, and select the dimension with maximize $F_1$ score. The most suitable dimension for ORG, PER, LOC are 16- ${D}$ /16- ${D}$ /24- ${D}$ (these dimensions will be used as parameters in the following experiments), respectively . We discover that in low-dimensional space, the distributions of NEs are better. In high dimensions, the curse of dimension could be the main reason to limit the performance. Table 3 lists the final maximum $F_1$ score of three NE types. The results of the three types of NE are almost 50%, and PER type performs best. The main factor may be that PER NEs are represented as single-word in our dictionary, and word embeddings can better represents their meanings. The result also states that better representations for multi-word NEs which are not covered by the dictionary instead of the average of each word may help bring better results. Besides, the incompleteness of NE dictionaries and noises during pre-processing may cause a decrease on the performance. Overall, hypersphere model has shown been effectively used as the open modeling for NEs. Hypersphere Mapping The following preparations were made for the mapping: $(i)$ A large enough NE dictionary in source (resource-rich) corpus; $(ii)$ A small amount of annotated seed pairs. We use $s$ to represent the number of seed pairs and $d$ to represent the number of unknown variables. With seed pair size $s < d$ , the matrix can be solved with much loose constraints, and $F_1$ score remarkably increases with more seed pairs. Once $s > d$ , the linear equation group will be always determined by strong enough constraints, which leads to a stable solution. Based on the characteristics, we only take two dozen of seed pairs on each type in following experiments. We combine human translation and online translation together for double verification for this small set of seed pairs. In this part, we utilize English and Chinese as the corpus of known NEs in turn, and predict the NE distribution of the other language. Evaluation In order to quantitatively represent the mapping effect, we present a new evaluation method to judge the hypersphere mapping between English and Chinese: $$\begin{aligned} P = \frac{V_i}{V_m} \quad R = \frac{V_i}{V_t} \quad F_1 = \frac{2 * P * R}{P + R} \end{aligned}$$ (Eq. 29) where ${V_t, V_m, V_i}$ represent the volumes of the target, mapping and intersection hyperspheres. Due to the difficulty of calculating the volume of hyperspheres in high dimensions, we adopt Monte Carlo methods to simulate the volume BIBREF25 . we generate a great quantity of points in the embedding spaces, and take the amount of the points falling in each hypersphere as its volume. Mapping between English and Chinese Table 4 shows the comparisons of cross-lingual named entity extraction performance. We use the unsupervised method proposed in BIBREF26 to generate cross-lingual embeddings. $k$ -NN and SVM are the same as monolingual cases in Table 3 except for the training set. $k$ -NN $_{150}$ and SVM $_{150}$ use 20% of the NEs in source language and 150 NEs (50 LOC, PER and ORG) in target language for training, while $k$ -NN $_{2500}$ and SVM $_{2500}$ use 20% of the NEs in source language and 2500 NEs (1000 LOC and PER, 500 ORG) in target language. $k$ -NN and SVM depend much on the annotated training set, requiring more than $1K$ training samples to provide a performance as our model offers. Due to the instability of ORG type in length, taking the average of each word embedding may disobey the syntactic and semantic regularities of ORG NEs, thereby undermines the multilingual isomorphism characteristics, which causes the inferior performance of our model on this type of NEs. This suggests that build better representations NEs for multi-word NEs may contribute to a better performance in our model. Mapping to truly Low-resource Language We build named entity dataset for a truly resource-poor language, Indonesian, and manually examine the nearest words to the hypersphere center for 'gold-standard' evaluation. We take English as the source language, the settings of the dimension $D$ and the number of seed pairs $s$ are the same as the above experiments between Chinese and English. From the results listed in Table 5, we can see that even the precision of the top-100 NEs are 0.350 $F_1$ /0.440 $F_1$ /0.310 $F_1$ , respectively, which proves the this distribution can indeed serves as a candidate NE dictionary for Indonesian. [9] The authors of BIBREF24 publish an updated results (92.98) on CoNLL-2003 dataset in https://github.com/zalandoresearch/flair/issues/206 on their 0.3.2 version, and this is the best result at our most try. [10] This is the reported state-of-the-art result in their github. [11]We use the same parameters as the authors release in https://github.com/zalandoresearch/flair/issues/173 and obtain the result of 89.45 on ONTONOTES 5.0 dataset. Off-the-shelf NE Recognition Systems To evaluate the influence of our hypersphere feature for off-the-shelf NER systems, we perform the NE recognition on two standard NER benchmark datasets, CoNLL2003 and ONTONOTES 5.0. Our results in Table 6 and Table 7 demonstrate the power of hypersphere features, which contribute to nearly all of the three types of entities as shown in Table 6, except for a slight drop in the PER type of BIBREF22 on a strong baseline. HS features stably enhance all strong state-of-the-art baselines, BIBREF22 , BIBREF21 and BIBREF23 by 0.33/0.72/0.23 $F_1$ point and 0.13/0.3/0.1 $F_1$ point on both benchmark datasets, CoNLL-2003 and ONTONOTES 5.0. We show that our HS feature is also comparable with previous much more complicated LS feature, and our model surpasses their baseline (without LS feature) by 0.58/0.78 $F_1$ point with only HS features. We establish a new state-of-the-art $F_1$ score of 89.75 on ONTONOTES 5.0, while matching state-of-the-art performance with a $F_1$ score of 92.95 on CoNLL-2003 dataset. Related Work In recent years, word embeddings have also been used as a feature to enhance the NE recognition, with the revealing of linguistic features in morphological, syntactic and semantic perspective. BIBREF1 clustered the word embeddings and combined multiple cluster granularities to improve the NE recognition performance. Our work likewise use word embeddings to help NE recognition, we make use of the characteristic that syntactically and semantically s are more likely to be neighbors in embedding spaces and construct a hypersphere model to encompass NEs. Cross-lingual knowledge transfer is a highly promising work for resource-poor languages, annotation projection and representation projection are widely used in NE recognition BIBREF27 , BIBREF5 , BIBREF4 , BIBREF28 , BIBREF29 , BIBREF30 . These works put forward inconvenient requirements for parallel or comparable corpora, a large amount of annotated or translation data or bilingual lexicon. Different from any existing work to the best of our knowledge, this is the first work that merely uses isomorphic mappings in low-dimensional embedding spaces to recognize NEs, and we introduce a mathematically simple model to describe NE embedding distribution from visualization results, showing it works for both monolingual and cross-lingual situations. Conclusion Named entities being an open set which keeps expanding are difficult to represent through a closed NE dictionary. This work mitigates significant defects in previous closed NE definitions and proposes a new open definition for NEs by modeling their embedding distributions with least parameters. We visualize NE distributions in monolingual case and perform an effective isomorphism spaces mapping in cross-lingual case. According to our work, we demonstrate that common named entity types (PER, LOC, ORG) tend to be densely distributed in a hypersphere and it is possible to build a mapping between the NE distributions in embedding spaces to help cross-lingual NE recognition. Experimental results show that the distribution of named entities via mapping can be used as a good enough replacement for the original distribution. Then the discovery is used to build an NE dictionary for Indonesian being a truly low-resource language, which also gives satisfactory precision. Finally, our simple hypersphere features being the representation of NE likelihood can be used for enhancing off-the-shelf NER systems by concatenating with word embeddings and the output of BiLSTM in the input layer and encode layer, respectively, and we achieve a new state-of-the-art $F_1$ score of 89.75 on ONTONOTES 5.0 benchmark. In this work, we also give a better solution for unregistered NEs. For any newly emerged NE together with its embedding, in case we obtain the hypersphere of each named entity, the corresponding named entity category can be determined by calculating the distance between its word embedding and the center of each hypersphere.

Yes

6b55b558ed581759425ede5d3a6fcdf44b8082ac

6b55b558ed581759425ede5d3a6fcdf44b8082ac_0

Q: What previously proposed methods is this method compared against? Text: Introduction A lot of work has been done in the field of Twitter sentiment analysis till date. Sentiment analysis has been handled as a Natural Language Processing task at many levels of granularity. Most of these techniques use Machine Learning algorithms with features such as unigrams, n-grams, Part-Of-Speech (POS) tags. However, the training datasets are often very large, and hence with such a large number of features, this process requires a lot of computation power and time. The following question arises: What to do if we do not have resources that provide such a great amount of computation power? The existing solution to this problem is to use a smaller sample of the dataset. For sentiment analysis, if we train the model using a smaller randomly chosen sample, then we get low accuracy [16, 17]. In this paper, we propose a novel technique to sample tweets for building a sentiment classification model so that we get higher accuracy than the state-of-the-art baseline method, namely Distant Supervision, using a smaller set of tweets. Our model has lower computation time and higher accuracy compared to baseline model. Users often express sentiment using subjective expression. Although objective expressions can also have sentiment, it is much rare. Determining subjectivity is quite efficient compared to determining sentiment. Subjectivity can be determined for individual tweets. But to do sentiment classification, we need to build a classification model with positive and negative sentiment tweets. The time to train a sentiment classification model increases with the increase in the number of training tweets. In this paper, we use tweet subjectivity to select the best training tweets. This not only lowers the computation time but also increases the accuracy because we have training data with less noise. Even the created features will be more relevant to the classification task. The computation cost will reduce due to small training data size and better set of features. Thus if users do not have enough computational resources, they can filter the training dataset using a high value of subjectivf ity threshold. This ensures reliable prediction on a smaller training dataset, and eventually requires less computational time. The above approach, and some of the intricacies that invariably seep in, need to be considered, and are described in the later sections of the paper. In this paper we also integrate a lot of meticulous preprocessing steps. This makes our model more robust, and hence leads to higher accuracy. Along with the machine learning algorithms being used, we use a heuristic-based classification of tweets. This is based on the EFWS of a tweet, which is described in later sections. This heuristic basically takes into account the polarity scores of frequently used words in tweets, and is able to achieve around 85% accuracy on our dataset, hence boosting the overall accuracy by a considerable amount. Our training data consists of generic (not topic-specific) Twitter messages with emoticons, which are used as noisy labels. We show that the accuracy obtained on a training dataset comprising 100K tweets, and a test dataset of 5000 tweets gives an accuracy of around 80% on the following classifiers: Naive Bayes, RBF-kernel Support Vector Machine, and Logistic Regression. Our model takes roughly half the time to train and achieves higher accuracy (than the baseline model) on all the classifiers. Because the amount of training time is expected to increase exponentially as the training data increases, we expect our model to outperform (in terms of higher accuracy) the baseline model at a speed which is at least twofold the speed of the baseline model on larger datasets. Related Work There has been a large amount of prior research in sentiment analysis of tweets. Read [10] shows that using emoticons as labels for positive and sentiment is effective for reducing dependencies in machine learning techniques. Alec Go [1] used Naive Bayes, SVM, and MaxEnt classifiers to train their model. This, as mentioned earlier, is our baseline model. Our model builds on this and achieves higher accuracy on a much smaller training dataset. Ayushi Dalmia [6] proposed a model with a more involved preprocessing stage, and used features like scores from Bing Liu’s Opinion Lexicon, and number of positive, negative POS tags. This model achieved considerably high accuracies considering the fact that their features were the not the conventional bag-of-words, or any n-grams. The thought of using the polarity scores of frequently used tweet words (as described in our EFWS heuristic) was inspired from this work. [14] created prior probabilities using the datasets for the average sentiment of tweets in different spatial, temporal and authorial contexts. They then used a Bayesian approach to combine these priors with standard bigram language models. Another significant effort in sentiment analysis on Twitter data is by Barbosa [16]. They use polarity predictions from three websites as noisy labels to train a model and use 1000 manually labelled tweets for tuning and another 1000 for testing. They propose the use of syntax features of tweets like punctuation, retweet, hashtags, link, and exclamation marks in addition with features like prior polarity of words and POS of words. Some works leveraged the use of existing hashtags in the Twitter data for building the training data. (Davidov, Tsur, and Rappoport 2010) also use hashtags for creating training data, but they limit their experiments to sentiment/non-sentiment classification, rather than 3-way polarity classification, as [15] does. Our model integrates some of the preprocessing techniques this work used. Hassan Saif [9] introduced a novel approach of adding semantics as additional features into the training set for sentiment analysis. This approach works well for topic specific data. Hence, we thought of taking a different approach for a generic tweet dataset like ours. Subjectivity Subjectivity refers to how someone's judgment is shaped by personal opinions and feelings instead of outside influences. An objective perspective is one that is not influenced by emotions, opinions, or personal feelings - it is a perspective based in fact, in things quantifiable and measurable. A subjective perspective is one open to greater interpretation based on personal feeling, emotion, aesthetics, etc. Subjectivity classification is another topic in the domain of text classification which is garnering more and more interest in the field of sentiment analysis. Since a single sentence may contain multiple opinions and subjective and factual clauses, this problem is not as straightforward as it seems. Below are some examples of subjective and objective sentences. Objective sentence with no sentiment: So, the Earth revolves around the Sun. Objective sentence with sentiment: The drug relieved my pain. Subjective sentence with no sentiment: I believe he went home yesterday. Subjective sentence with sentiment: I am so happy you got the scholarship. Classifying a sentence as subjective or objective provides certain conclusions. Purely objective sentences do not usually convey any sentiment, while most of the purely subjective sentences have a clear inclination towards either the positive or negative sentiment. Sentences which are not completely subjective or objective may or may not convey a sentiment. Libraries like TextBlob, and tools like Opinion Finder can be used to find the extent to which a sentence can be considered subjective. Since tweets are usually person-specific, or subjective, we use this intuition to reduce the size of the training set by filtering the sentences with a subjectivity level below a certain threshold (fairly objective tweets). Implementation In this section, we explain the various preprocessing techniques used for feature reduction, and also the additional step of filtering the training dataset using the subjectivity score of tweets. We further describe our approach of using different machine learning classifiers and feature extractors. We also propose an additional heuristic for sentiment classification which can be used as a tag-along with the learning heuristics. Corpus Our training dataset has 1.6 million tweets, and 5000 tweets in the test dataset. Since the test dataset provided comprised only 500 tweets, we have taken part of the training data (exactly 5000 tweets, distinct from the training dataset) as the test dataset. We remove emoticons from our training and test data. The table below shows some sample tweets. Subjectivity Filtering This is a new step we propose to achieve higher accuracy on a smaller training dataset. We use TextBlob to classify each tweet as subjective or objective. We then remove all tweets which have a subjectivity level/score (score lies between 0 and 1) below a specified threshold. The remaining tweets are used for training purposes. We observe that a considerable number of tweets are removed as the subjectivity threshold increases. We show the effect of doing this procedure on the overall accuracy in the evaluation section of the paper. Preprocessing The Twitter language model has many unique properties. We take advantage of the following properties to reduce the feature space. Most of the preprocessing steps are common to most of the previous works in the field. However, we have added some more steps to this stage of our model. We first strip off the emoticons from the data. Users often include twitter usernames in their tweets in order to direct their messages. We also strip off usernames (e.g. @Chinmay) and URLs present in tweets because they do not help us in sentiment classification. Apart from full stops, which are dealt in the next point, other punctuations and special symbols are also removed. Repeated whitespaces are replaced with a single space. We also perform stemming to reduce the size of the feature space. In the previous works, full stops are just usually replaced by a space. However, we have observed that casual language in tweets is often seen in form of repeated punctuations. For example, “this is so cool...wow". We take into consideration this format, and replace two or more occurrences of “." and “-" with a space. Also, full stops are also quite different in usage. Sometimes, there isn't any space in between sentences. For example, “It’s raining.Feeling awesome". We replace a single occurrence of a full stop with a space to ensure correct feature incorporation. In the case of hashtags, most of the previous works just consider the case of hashtags followed by a single word; they just remove the hashtag and add the word to the feature vector. However, sometimes, there are multiple words after a hashtag, and more often than not, these words form an important, conclusive part of the Tweet. For example, #ThisSucks, or #BestMomentEver. These hashtags need to be dealt with in a proper fashion. We split the text after hashtags after before each capital letter, and add these as tokens to the feature vector. For hashtags followed by a single word, we just replace the pattern #word with the word, as conventional models do. The intuition behind this step is that quite often, the sentiment of a tweet is expressed in form of a hashtag. For example, #happy or #disappointed are frequently used hashtags, and we don’t want to lose this information during sentiment classification. Tweets contain very casual language as mentioned earlier. For example, if we search “wow" with an arbitrary number of o's in the middle (e.g. wooow, woooow) on Twitter, there will most likely be a non-empty result set. We use preprocessing so that any letter occurring more than two times in a row is replaced with two occurrences. In the samples above, these words would be converted into the token “woow". After all the above modifications, tweets are converted into lowercase to avoid confusion between features having same content, but are different in capitalization. We gather a list of 400 stopwords. These words, if present in the tweets, are not considered in the feature vector. We store an acronym dictionary which has over 5000, frequently-used acronyms and their abbreviations. We replace such acronyms in tweets with their abbreviation, since these can be of great use while sentiment classification. All negative words like 'cannot', 'can't', 'won't', 'don't' are replaced by 'not', which effectively keeps the sentiment stable. It is observed that doing this makes the training faster, since the model has to deal with a smaller feature vector. Baseline model The baseline model for our experiments is explained in the paper by Alec Go [1]. The model uses the Naive Bayes, SVM, and the Maximum Entropy classifiers for their experiment. Their feature vector is either composed of Unigrams, Bigrams, Unigrams + Bigrams, or Unigrams + POS tags. This work achieved the following maximum accuracies: a) 82.2 for the Unigram feature vector, using the SVM classifier, b) 83.0 for the Unigram + Bigram feature vector, using the MaxEnt classifier, and 82.7 using the Naive Bayes classifier. c) 81.9 for the Unigram + POS feature vector, using the SVM classifier. These baseline accuracies were on a training dataset of 1.6 million tweets, and a test dataset of 500 tweets. We are using the same training dataset for our experiments. We later present the baseline accuracies on a training set of 200K tweets, and a test dataset of 5000 tweets; we compare our model's accuracy with these baseline accuracy values on the same test data of 5000 tweets. Effective Word Score (EFWS) Heuristic We have described our baseline model above. So the feature vectors we collate results for, are Unigram, Unigram + Bigram, and Unigram + POS. We have already made two major changes before the training starts on our dataset as compared to our baseline model. Firstly, our training dataset will be filtered according to the subjectivity threshold. And secondly, our preprocessing is much more robust as compared to their work. Now let us look at an additional heuristic we use to obtain labels for our test data. Along with dictionaries for stop words and acronyms, we also maintain a dictionary of a list of frequently used words and their polarity scores. This dictionary has around 2500 words and their polarity score ranging from -5 to 5. At runtime, we also use all synonyms of a word (from WordNet) present in a tweet and also the dictionary, and assign them the same score as the dictionary word. There is a reasonable assumption here, that the synonyms aren't very extremal in nature, that is, a word with a polarity score of 2 cannot have a synonym which has a polarity score of 5. Now, we calculate the Effective Word Scores of a tweet. We define the Effective Word Score of score x as EFWS(x) = N(+x) - N(-x), where N(x) is the number of words in the tweet with polarity score x. For example, if a tweet has one word with score 5, three words with score 4, two with score 2, three with with score -2, one with score -3, and finally two with score -4, then the effective word scores are: EFWS(5) = N(5) - N(-5) = 1 - 0 = 1 EFWS(4) = N(4) - N(-4) = 3 - 2 = 1 EFWS(3) = N(3) - N(-3) = 0 - 1 = -1 EFWS(2) = N(2) - N(-2) = 2 - 3 = -1 EFWS(1) = N(1) - N(-1) = 2 - 0 = 2 We now define the heuristic for obtaining the label of a Tweet. (EFWS(5) INLINEFORM0 1 or EFWS(4) INLINEFORM1 1) and (EFWS(2) INLINEFORM2 1) Label = positive Similarly, (EFWS(5) INLINEFORM0 -1 or EFWS(4) INLINEFORM1 -1) and (EFWS(2) INLINEFORM2 -1) Label = negative The basic intuition behind such a heuristic is that we found tweets having one strongly positive and one moderately positive word more than the number of strongly negative and the moderately negative words respectively, usually conveyed a positive sentiment. Similar was the case for negative sentiments. The tweets getting a label from this heuristic are not sent into the training phase. After considerable amount of experimenting, and analyzing the nature of our dataset, which is not domain specific, we have reached the conclusion that the heuristic mentioned above is optimal for obtaining labels. We found that the heuristic accuracy was around 85% for a training dataset of 100K and a test dataset of 5K, where the total number of test tweets labelled by the heuristic were around 500. This means that around 425 out of the 500 tweets received a correct prediction of sentiment using this heuristic. Thus, using this heuristic improves the overall accuracy, as well as saves time by reducing the number of tweets to be tested by the ML algorithms. Training Model We use the following classifiers for our model. Naive Bayes is a simple model which works well on text categorization. We use a Naive Bayes model. Class c* is assigned to tweet d, where c* = argmax P(c INLINEFORM0 d). INLINEFORM1 And INLINEFORM0 is calculated using Bayes Rule. In this formula, f represents a feature and INLINEFORM1 represents the count of feature INLINEFORM2 found in tweet d. There are a total of m features. Parameters P(c) and INLINEFORM3 are obtained through maximum likelihood estimates. Support vector machines are based on the Structural Risk Minimization principle from computational learning theory. SVM classification algorithms for binary classification is based on finding a separation between hyperplanes defined by classes of data. One remarkable property of SVMs is that their ability to learn can be independent of the dimensionality of the feature space. SVMs can generalize even in the presence of many features as in the case of text data classification. We use a non-linear Support Vector Machine with an RBF kernel. Maximum Entropy Model belongs to the family of discriminative classifiers also known as the exponential or log-linear classifiers.. In the naive Bayes classifier, Bayes rule is used to estimate this best y indirectly from the likelihood INLINEFORM0 (and the prior INLINEFORM1 ) but a discriminative model takes this direct approach, computing INLINEFORM2 by discriminating among the different possible values of the class y rather than first computing a likelihood. INLINEFORM3 Logistic regression estimates INLINEFORM0 by combining the feature set linearly (multiplying each feature by a weight and adding them up), and then applying a function to this combination. Evaluation In this section, we present the collated results of our experiments. To show that our model achieves higher accuracy than the baseline model and on a smaller training dataset, we first fix the test dataset. Our test dataset, as mentioned before, consists of 5000 tweets. We conducted our experiments on an Intel Core i5 machine (4 cores), with 8 GB RAM. The following are the accuracies of the baseline model on a training set of 200K tweets: We filtered the training set with a subjectivity threshold of 0.5. By doing this, we saw that the number of tweets reduced to approximately 0.6 million tweets from an earlier total of 1.6 million. We then trained our model described in earlier sections on a 100K tweets randomly picked from this filtered training dataset, and observed the following accuracies: Note that all the accuracies in the tables above have been recorded as the average of 3 iterations of our experiment. We achieve higher accuracy for all feature vectors, on all classifiers, and that too from a training dataset half the size of the baseline one. We now see the intricacies of the subjectivity threshold parameter. It is clear that more and more tweets get filtered as the subjectivity threshold parameter increases. This can be seen in the Figure 1 shown below. We have plotted the number of tweets that remain after filtering from two sources: TextBlob, Opinion Finder Tool. TextBlob has an inbuilt function that provides us the subjectivity level of a tweet. On the other hand, Opinion Finder only provides the information of which parts of the text are subjective, and which are objective. From that, we define the subjectivity level of that text as: Subjectivity level = INLINEFORM0 [ xlabel=Subjectivity Threshold, ylabel=Tweets (in millions), xmin=0, xmax=1, ymin=0, ymax=2000000, xtick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, ytick=0,200000,400000,600000,800000,1000000,1200000,1400000,1600000,1800000, legend pos=north east, ] [color=red] coordinates (0, 1600000) (0.1, 939785) (0.2, 873054) (0.3, 804820) (0.4, 712485) (0.5, 571864) (0.6, 449286) (0.7, 304874) (0.8, 211217) (0.9, 135788) ; [color=blue] coordinates (0, 1600000) (0.1, 602313) (0.2, 499173) (0.3, 392223) (0.4, 262109) (0.5, 169477) (0.6, 154667) (0.7, 139613) (0.8, 126148) (0.9, 116842) ; Textblob, Opinion Finder Figure 1: Number of tweets with subjectivity greater than the subjectivity threshold [ xlabel=Subjectivity Threshold, ylabel=Accuracy (from 0 to 1), xmin=0, xmax=1, ymin=0.7, ymax=1, xtick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, ytick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, legend pos=north east, ] [color=red] coordinates (0.1, 0.753871866) (0.2, 0.779442897) (0.3, 0.763421155) (0.4, 0.783231198) (0.5,0.805132645) (0.6,0.807373259) (0.7,0.808587744) (0.8,0.817799443) (0.9,0.823872989) ; Figure 2: Variation of accuracy (*Training data of 100K, Test data of 5K) with subjectivity threshold. *TextBlob is used to filter the tweets to form the training dataset. We now focus on the issue of choosing the optimum threshold value. As the subjectivity threshold parameter increases, our model trains on tweets with a higher subjectivity level, and the overall accuracy increases. We observed the following accuracies on subjectivity level 0.8 (Unigrams as features): Naive Bayes: 80.32% Non-linear SVM: 80.15 % Logistic Regression: 81.77% We should consider the fact that a lot of useful tweets are also lost in the process of gradually increasing the parameter, and this could cause a problem in cases when the test data is very large, because the model will not train on a generic dataset. Researchers may use a higher subjectivity threshold for their experiments if they are confident that most of the important information would be retained. This is most likely to happen in case of topic-specific or domain-specific data. [ ybar, enlargelimits=0.15, legend style=anchor=north, legend pos= north east, ylabel=Training time (in minutes), symbolic x coords=baseline,subjectivity=0.5,subjectivity=0.8, xtick=data, ] coordinates (baseline,17.4) (subjectivity=0.5,12.55) (subjectivity=0.8,10.68); coordinates (baseline,16.23) (subjectivity=0.5,12.31) (subjectivity=0.8,10.34); coordinates (baseline,31.9) (subjectivity=0.5,18.24) (subjectivity=0.8,16.3); Logistic Regression,Naive Bayes,SVM Figure 3: Comparison of training times for Unigrams [ ybar, enlargelimits=0.15, legend style=anchor=north, legend pos= north east, ylabel=Training time (in minutes), symbolic x coords=baseline,subjectivity=0.5,subjectivity=0.8, xtick=data, ] coordinates (baseline,28.41) (subjectivity=0.5,14.09) (subjectivity=0.8,11.3); coordinates (baseline,16.6) (subjectivity=0.5,13.51) (subjectivity=0.8,12.66); coordinates (baseline,35.2) (subjectivity=0.5,20.6) (subjectivity=0.8,19.2); Logistic Regression,Naive Bayes,SVM Figure 4: Comparison of training times for Unigrams + Bigrams We use Logistic regression for classification and unigrams as the feature vector with K-fold cross validation for determining the accuracy. We choose an optimal threshold value of 0.5 for our experiment, considering the fact that the model should train on a more generic dataset. Figure 2 shows the variation of accuracy with the subjectivity threshold. The training size is fixed at 100K and the test dataset (5K tweets) is also same for all the experiments. We also measure the time taken to train our model, and compare it to the baseline model. Our observation was that our model took roughly half the amount of time in some cases and yet obtained a higher accuracy. Figures 3 and 4 show the difference in training time of the baseline model, our model on a 0.5 subjectivity-filtered dataset, and our model on a 0.8 subjectivity-filtered dataset on unigrams and unigrams + bigrams respectively. The times recorded are on a training dataset of 100K for our model and 200K for the baseline model, and a test dataset of 5K was fixed in all the recordings. The winning point, which can be seen from the plots, is that our model is considerably faster, and even has twofold speed in some cases. And alongside saving computation time, it achieves higher accuracy. This can be attributed to the fact that as the subjectivity threshold increases, only the tweets with highly polar words are retained in the training set and this makes the whole process faster. Conclusion We show that a higher accuracy can be obtained in sentiment classification of Twitter messages training on a smaller dataset and with a much faster computation time, and hence the issue of constraint on computation power is resolved to a certain extent. This can be achieved using a subjectivity threshold to selectively filter the training data, incorporating a more complex preprocessing stage, and using an additional heuristic for sentiment classification, along with the conventional machine learning techniques. As Twitter data is abundant, our subjectivity filtering process can achieve a better generalised model for sentiment classification.

Naive Bayes, SVM, Maximum Entropy classifiers

3e3f5254b729beb657310a5561950085fa690e83

3e3f5254b729beb657310a5561950085fa690e83_0

Q: How is effective word score calculated? Text: Introduction A lot of work has been done in the field of Twitter sentiment analysis till date. Sentiment analysis has been handled as a Natural Language Processing task at many levels of granularity. Most of these techniques use Machine Learning algorithms with features such as unigrams, n-grams, Part-Of-Speech (POS) tags. However, the training datasets are often very large, and hence with such a large number of features, this process requires a lot of computation power and time. The following question arises: What to do if we do not have resources that provide such a great amount of computation power? The existing solution to this problem is to use a smaller sample of the dataset. For sentiment analysis, if we train the model using a smaller randomly chosen sample, then we get low accuracy [16, 17]. In this paper, we propose a novel technique to sample tweets for building a sentiment classification model so that we get higher accuracy than the state-of-the-art baseline method, namely Distant Supervision, using a smaller set of tweets. Our model has lower computation time and higher accuracy compared to baseline model. Users often express sentiment using subjective expression. Although objective expressions can also have sentiment, it is much rare. Determining subjectivity is quite efficient compared to determining sentiment. Subjectivity can be determined for individual tweets. But to do sentiment classification, we need to build a classification model with positive and negative sentiment tweets. The time to train a sentiment classification model increases with the increase in the number of training tweets. In this paper, we use tweet subjectivity to select the best training tweets. This not only lowers the computation time but also increases the accuracy because we have training data with less noise. Even the created features will be more relevant to the classification task. The computation cost will reduce due to small training data size and better set of features. Thus if users do not have enough computational resources, they can filter the training dataset using a high value of subjectivf ity threshold. This ensures reliable prediction on a smaller training dataset, and eventually requires less computational time. The above approach, and some of the intricacies that invariably seep in, need to be considered, and are described in the later sections of the paper. In this paper we also integrate a lot of meticulous preprocessing steps. This makes our model more robust, and hence leads to higher accuracy. Along with the machine learning algorithms being used, we use a heuristic-based classification of tweets. This is based on the EFWS of a tweet, which is described in later sections. This heuristic basically takes into account the polarity scores of frequently used words in tweets, and is able to achieve around 85% accuracy on our dataset, hence boosting the overall accuracy by a considerable amount. Our training data consists of generic (not topic-specific) Twitter messages with emoticons, which are used as noisy labels. We show that the accuracy obtained on a training dataset comprising 100K tweets, and a test dataset of 5000 tweets gives an accuracy of around 80% on the following classifiers: Naive Bayes, RBF-kernel Support Vector Machine, and Logistic Regression. Our model takes roughly half the time to train and achieves higher accuracy (than the baseline model) on all the classifiers. Because the amount of training time is expected to increase exponentially as the training data increases, we expect our model to outperform (in terms of higher accuracy) the baseline model at a speed which is at least twofold the speed of the baseline model on larger datasets. Related Work There has been a large amount of prior research in sentiment analysis of tweets. Read [10] shows that using emoticons as labels for positive and sentiment is effective for reducing dependencies in machine learning techniques. Alec Go [1] used Naive Bayes, SVM, and MaxEnt classifiers to train their model. This, as mentioned earlier, is our baseline model. Our model builds on this and achieves higher accuracy on a much smaller training dataset. Ayushi Dalmia [6] proposed a model with a more involved preprocessing stage, and used features like scores from Bing Liu’s Opinion Lexicon, and number of positive, negative POS tags. This model achieved considerably high accuracies considering the fact that their features were the not the conventional bag-of-words, or any n-grams. The thought of using the polarity scores of frequently used tweet words (as described in our EFWS heuristic) was inspired from this work. [14] created prior probabilities using the datasets for the average sentiment of tweets in different spatial, temporal and authorial contexts. They then used a Bayesian approach to combine these priors with standard bigram language models. Another significant effort in sentiment analysis on Twitter data is by Barbosa [16]. They use polarity predictions from three websites as noisy labels to train a model and use 1000 manually labelled tweets for tuning and another 1000 for testing. They propose the use of syntax features of tweets like punctuation, retweet, hashtags, link, and exclamation marks in addition with features like prior polarity of words and POS of words. Some works leveraged the use of existing hashtags in the Twitter data for building the training data. (Davidov, Tsur, and Rappoport 2010) also use hashtags for creating training data, but they limit their experiments to sentiment/non-sentiment classification, rather than 3-way polarity classification, as [15] does. Our model integrates some of the preprocessing techniques this work used. Hassan Saif [9] introduced a novel approach of adding semantics as additional features into the training set for sentiment analysis. This approach works well for topic specific data. Hence, we thought of taking a different approach for a generic tweet dataset like ours. Subjectivity Subjectivity refers to how someone's judgment is shaped by personal opinions and feelings instead of outside influences. An objective perspective is one that is not influenced by emotions, opinions, or personal feelings - it is a perspective based in fact, in things quantifiable and measurable. A subjective perspective is one open to greater interpretation based on personal feeling, emotion, aesthetics, etc. Subjectivity classification is another topic in the domain of text classification which is garnering more and more interest in the field of sentiment analysis. Since a single sentence may contain multiple opinions and subjective and factual clauses, this problem is not as straightforward as it seems. Below are some examples of subjective and objective sentences. Objective sentence with no sentiment: So, the Earth revolves around the Sun. Objective sentence with sentiment: The drug relieved my pain. Subjective sentence with no sentiment: I believe he went home yesterday. Subjective sentence with sentiment: I am so happy you got the scholarship. Classifying a sentence as subjective or objective provides certain conclusions. Purely objective sentences do not usually convey any sentiment, while most of the purely subjective sentences have a clear inclination towards either the positive or negative sentiment. Sentences which are not completely subjective or objective may or may not convey a sentiment. Libraries like TextBlob, and tools like Opinion Finder can be used to find the extent to which a sentence can be considered subjective. Since tweets are usually person-specific, or subjective, we use this intuition to reduce the size of the training set by filtering the sentences with a subjectivity level below a certain threshold (fairly objective tweets). Implementation In this section, we explain the various preprocessing techniques used for feature reduction, and also the additional step of filtering the training dataset using the subjectivity score of tweets. We further describe our approach of using different machine learning classifiers and feature extractors. We also propose an additional heuristic for sentiment classification which can be used as a tag-along with the learning heuristics. Corpus Our training dataset has 1.6 million tweets, and 5000 tweets in the test dataset. Since the test dataset provided comprised only 500 tweets, we have taken part of the training data (exactly 5000 tweets, distinct from the training dataset) as the test dataset. We remove emoticons from our training and test data. The table below shows some sample tweets. Subjectivity Filtering This is a new step we propose to achieve higher accuracy on a smaller training dataset. We use TextBlob to classify each tweet as subjective or objective. We then remove all tweets which have a subjectivity level/score (score lies between 0 and 1) below a specified threshold. The remaining tweets are used for training purposes. We observe that a considerable number of tweets are removed as the subjectivity threshold increases. We show the effect of doing this procedure on the overall accuracy in the evaluation section of the paper. Preprocessing The Twitter language model has many unique properties. We take advantage of the following properties to reduce the feature space. Most of the preprocessing steps are common to most of the previous works in the field. However, we have added some more steps to this stage of our model. We first strip off the emoticons from the data. Users often include twitter usernames in their tweets in order to direct their messages. We also strip off usernames (e.g. @Chinmay) and URLs present in tweets because they do not help us in sentiment classification. Apart from full stops, which are dealt in the next point, other punctuations and special symbols are also removed. Repeated whitespaces are replaced with a single space. We also perform stemming to reduce the size of the feature space. In the previous works, full stops are just usually replaced by a space. However, we have observed that casual language in tweets is often seen in form of repeated punctuations. For example, “this is so cool...wow". We take into consideration this format, and replace two or more occurrences of “." and “-" with a space. Also, full stops are also quite different in usage. Sometimes, there isn't any space in between sentences. For example, “It’s raining.Feeling awesome". We replace a single occurrence of a full stop with a space to ensure correct feature incorporation. In the case of hashtags, most of the previous works just consider the case of hashtags followed by a single word; they just remove the hashtag and add the word to the feature vector. However, sometimes, there are multiple words after a hashtag, and more often than not, these words form an important, conclusive part of the Tweet. For example, #ThisSucks, or #BestMomentEver. These hashtags need to be dealt with in a proper fashion. We split the text after hashtags after before each capital letter, and add these as tokens to the feature vector. For hashtags followed by a single word, we just replace the pattern #word with the word, as conventional models do. The intuition behind this step is that quite often, the sentiment of a tweet is expressed in form of a hashtag. For example, #happy or #disappointed are frequently used hashtags, and we don’t want to lose this information during sentiment classification. Tweets contain very casual language as mentioned earlier. For example, if we search “wow" with an arbitrary number of o's in the middle (e.g. wooow, woooow) on Twitter, there will most likely be a non-empty result set. We use preprocessing so that any letter occurring more than two times in a row is replaced with two occurrences. In the samples above, these words would be converted into the token “woow". After all the above modifications, tweets are converted into lowercase to avoid confusion between features having same content, but are different in capitalization. We gather a list of 400 stopwords. These words, if present in the tweets, are not considered in the feature vector. We store an acronym dictionary which has over 5000, frequently-used acronyms and their abbreviations. We replace such acronyms in tweets with their abbreviation, since these can be of great use while sentiment classification. All negative words like 'cannot', 'can't', 'won't', 'don't' are replaced by 'not', which effectively keeps the sentiment stable. It is observed that doing this makes the training faster, since the model has to deal with a smaller feature vector. Baseline model The baseline model for our experiments is explained in the paper by Alec Go [1]. The model uses the Naive Bayes, SVM, and the Maximum Entropy classifiers for their experiment. Their feature vector is either composed of Unigrams, Bigrams, Unigrams + Bigrams, or Unigrams + POS tags. This work achieved the following maximum accuracies: a) 82.2 for the Unigram feature vector, using the SVM classifier, b) 83.0 for the Unigram + Bigram feature vector, using the MaxEnt classifier, and 82.7 using the Naive Bayes classifier. c) 81.9 for the Unigram + POS feature vector, using the SVM classifier. These baseline accuracies were on a training dataset of 1.6 million tweets, and a test dataset of 500 tweets. We are using the same training dataset for our experiments. We later present the baseline accuracies on a training set of 200K tweets, and a test dataset of 5000 tweets; we compare our model's accuracy with these baseline accuracy values on the same test data of 5000 tweets. Effective Word Score (EFWS) Heuristic We have described our baseline model above. So the feature vectors we collate results for, are Unigram, Unigram + Bigram, and Unigram + POS. We have already made two major changes before the training starts on our dataset as compared to our baseline model. Firstly, our training dataset will be filtered according to the subjectivity threshold. And secondly, our preprocessing is much more robust as compared to their work. Now let us look at an additional heuristic we use to obtain labels for our test data. Along with dictionaries for stop words and acronyms, we also maintain a dictionary of a list of frequently used words and their polarity scores. This dictionary has around 2500 words and their polarity score ranging from -5 to 5. At runtime, we also use all synonyms of a word (from WordNet) present in a tweet and also the dictionary, and assign them the same score as the dictionary word. There is a reasonable assumption here, that the synonyms aren't very extremal in nature, that is, a word with a polarity score of 2 cannot have a synonym which has a polarity score of 5. Now, we calculate the Effective Word Scores of a tweet. We define the Effective Word Score of score x as EFWS(x) = N(+x) - N(-x), where N(x) is the number of words in the tweet with polarity score x. For example, if a tweet has one word with score 5, three words with score 4, two with score 2, three with with score -2, one with score -3, and finally two with score -4, then the effective word scores are: EFWS(5) = N(5) - N(-5) = 1 - 0 = 1 EFWS(4) = N(4) - N(-4) = 3 - 2 = 1 EFWS(3) = N(3) - N(-3) = 0 - 1 = -1 EFWS(2) = N(2) - N(-2) = 2 - 3 = -1 EFWS(1) = N(1) - N(-1) = 2 - 0 = 2 We now define the heuristic for obtaining the label of a Tweet. (EFWS(5) INLINEFORM0 1 or EFWS(4) INLINEFORM1 1) and (EFWS(2) INLINEFORM2 1) Label = positive Similarly, (EFWS(5) INLINEFORM0 -1 or EFWS(4) INLINEFORM1 -1) and (EFWS(2) INLINEFORM2 -1) Label = negative The basic intuition behind such a heuristic is that we found tweets having one strongly positive and one moderately positive word more than the number of strongly negative and the moderately negative words respectively, usually conveyed a positive sentiment. Similar was the case for negative sentiments. The tweets getting a label from this heuristic are not sent into the training phase. After considerable amount of experimenting, and analyzing the nature of our dataset, which is not domain specific, we have reached the conclusion that the heuristic mentioned above is optimal for obtaining labels. We found that the heuristic accuracy was around 85% for a training dataset of 100K and a test dataset of 5K, where the total number of test tweets labelled by the heuristic were around 500. This means that around 425 out of the 500 tweets received a correct prediction of sentiment using this heuristic. Thus, using this heuristic improves the overall accuracy, as well as saves time by reducing the number of tweets to be tested by the ML algorithms. Training Model We use the following classifiers for our model. Naive Bayes is a simple model which works well on text categorization. We use a Naive Bayes model. Class c* is assigned to tweet d, where c* = argmax P(c INLINEFORM0 d). INLINEFORM1 And INLINEFORM0 is calculated using Bayes Rule. In this formula, f represents a feature and INLINEFORM1 represents the count of feature INLINEFORM2 found in tweet d. There are a total of m features. Parameters P(c) and INLINEFORM3 are obtained through maximum likelihood estimates. Support vector machines are based on the Structural Risk Minimization principle from computational learning theory. SVM classification algorithms for binary classification is based on finding a separation between hyperplanes defined by classes of data. One remarkable property of SVMs is that their ability to learn can be independent of the dimensionality of the feature space. SVMs can generalize even in the presence of many features as in the case of text data classification. We use a non-linear Support Vector Machine with an RBF kernel. Maximum Entropy Model belongs to the family of discriminative classifiers also known as the exponential or log-linear classifiers.. In the naive Bayes classifier, Bayes rule is used to estimate this best y indirectly from the likelihood INLINEFORM0 (and the prior INLINEFORM1 ) but a discriminative model takes this direct approach, computing INLINEFORM2 by discriminating among the different possible values of the class y rather than first computing a likelihood. INLINEFORM3 Logistic regression estimates INLINEFORM0 by combining the feature set linearly (multiplying each feature by a weight and adding them up), and then applying a function to this combination. Evaluation In this section, we present the collated results of our experiments. To show that our model achieves higher accuracy than the baseline model and on a smaller training dataset, we first fix the test dataset. Our test dataset, as mentioned before, consists of 5000 tweets. We conducted our experiments on an Intel Core i5 machine (4 cores), with 8 GB RAM. The following are the accuracies of the baseline model on a training set of 200K tweets: We filtered the training set with a subjectivity threshold of 0.5. By doing this, we saw that the number of tweets reduced to approximately 0.6 million tweets from an earlier total of 1.6 million. We then trained our model described in earlier sections on a 100K tweets randomly picked from this filtered training dataset, and observed the following accuracies: Note that all the accuracies in the tables above have been recorded as the average of 3 iterations of our experiment. We achieve higher accuracy for all feature vectors, on all classifiers, and that too from a training dataset half the size of the baseline one. We now see the intricacies of the subjectivity threshold parameter. It is clear that more and more tweets get filtered as the subjectivity threshold parameter increases. This can be seen in the Figure 1 shown below. We have plotted the number of tweets that remain after filtering from two sources: TextBlob, Opinion Finder Tool. TextBlob has an inbuilt function that provides us the subjectivity level of a tweet. On the other hand, Opinion Finder only provides the information of which parts of the text are subjective, and which are objective. From that, we define the subjectivity level of that text as: Subjectivity level = INLINEFORM0 [ xlabel=Subjectivity Threshold, ylabel=Tweets (in millions), xmin=0, xmax=1, ymin=0, ymax=2000000, xtick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, ytick=0,200000,400000,600000,800000,1000000,1200000,1400000,1600000,1800000, legend pos=north east, ] [color=red] coordinates (0, 1600000) (0.1, 939785) (0.2, 873054) (0.3, 804820) (0.4, 712485) (0.5, 571864) (0.6, 449286) (0.7, 304874) (0.8, 211217) (0.9, 135788) ; [color=blue] coordinates (0, 1600000) (0.1, 602313) (0.2, 499173) (0.3, 392223) (0.4, 262109) (0.5, 169477) (0.6, 154667) (0.7, 139613) (0.8, 126148) (0.9, 116842) ; Textblob, Opinion Finder Figure 1: Number of tweets with subjectivity greater than the subjectivity threshold [ xlabel=Subjectivity Threshold, ylabel=Accuracy (from 0 to 1), xmin=0, xmax=1, ymin=0.7, ymax=1, xtick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, ytick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, legend pos=north east, ] [color=red] coordinates (0.1, 0.753871866) (0.2, 0.779442897) (0.3, 0.763421155) (0.4, 0.783231198) (0.5,0.805132645) (0.6,0.807373259) (0.7,0.808587744) (0.8,0.817799443) (0.9,0.823872989) ; Figure 2: Variation of accuracy (*Training data of 100K, Test data of 5K) with subjectivity threshold. *TextBlob is used to filter the tweets to form the training dataset. We now focus on the issue of choosing the optimum threshold value. As the subjectivity threshold parameter increases, our model trains on tweets with a higher subjectivity level, and the overall accuracy increases. We observed the following accuracies on subjectivity level 0.8 (Unigrams as features): Naive Bayes: 80.32% Non-linear SVM: 80.15 % Logistic Regression: 81.77% We should consider the fact that a lot of useful tweets are also lost in the process of gradually increasing the parameter, and this could cause a problem in cases when the test data is very large, because the model will not train on a generic dataset. Researchers may use a higher subjectivity threshold for their experiments if they are confident that most of the important information would be retained. This is most likely to happen in case of topic-specific or domain-specific data. [ ybar, enlargelimits=0.15, legend style=anchor=north, legend pos= north east, ylabel=Training time (in minutes), symbolic x coords=baseline,subjectivity=0.5,subjectivity=0.8, xtick=data, ] coordinates (baseline,17.4) (subjectivity=0.5,12.55) (subjectivity=0.8,10.68); coordinates (baseline,16.23) (subjectivity=0.5,12.31) (subjectivity=0.8,10.34); coordinates (baseline,31.9) (subjectivity=0.5,18.24) (subjectivity=0.8,16.3); Logistic Regression,Naive Bayes,SVM Figure 3: Comparison of training times for Unigrams [ ybar, enlargelimits=0.15, legend style=anchor=north, legend pos= north east, ylabel=Training time (in minutes), symbolic x coords=baseline,subjectivity=0.5,subjectivity=0.8, xtick=data, ] coordinates (baseline,28.41) (subjectivity=0.5,14.09) (subjectivity=0.8,11.3); coordinates (baseline,16.6) (subjectivity=0.5,13.51) (subjectivity=0.8,12.66); coordinates (baseline,35.2) (subjectivity=0.5,20.6) (subjectivity=0.8,19.2); Logistic Regression,Naive Bayes,SVM Figure 4: Comparison of training times for Unigrams + Bigrams We use Logistic regression for classification and unigrams as the feature vector with K-fold cross validation for determining the accuracy. We choose an optimal threshold value of 0.5 for our experiment, considering the fact that the model should train on a more generic dataset. Figure 2 shows the variation of accuracy with the subjectivity threshold. The training size is fixed at 100K and the test dataset (5K tweets) is also same for all the experiments. We also measure the time taken to train our model, and compare it to the baseline model. Our observation was that our model took roughly half the amount of time in some cases and yet obtained a higher accuracy. Figures 3 and 4 show the difference in training time of the baseline model, our model on a 0.5 subjectivity-filtered dataset, and our model on a 0.8 subjectivity-filtered dataset on unigrams and unigrams + bigrams respectively. The times recorded are on a training dataset of 100K for our model and 200K for the baseline model, and a test dataset of 5K was fixed in all the recordings. The winning point, which can be seen from the plots, is that our model is considerably faster, and even has twofold speed in some cases. And alongside saving computation time, it achieves higher accuracy. This can be attributed to the fact that as the subjectivity threshold increases, only the tweets with highly polar words are retained in the training set and this makes the whole process faster. Conclusion We show that a higher accuracy can be obtained in sentiment classification of Twitter messages training on a smaller dataset and with a much faster computation time, and hence the issue of constraint on computation power is resolved to a certain extent. This can be achieved using a subjectivity threshold to selectively filter the training data, incorporating a more complex preprocessing stage, and using an additional heuristic for sentiment classification, along with the conventional machine learning techniques. As Twitter data is abundant, our subjectivity filtering process can achieve a better generalised model for sentiment classification.

We define the Effective Word Score of score x as EFWS(x) = N(+x) - N(-x), where N(x) is the number of words in the tweet with polarity score x.

5bb96b255dab3e47a8a68b1ffd7142d0e21ebe2a

5bb96b255dab3e47a8a68b1ffd7142d0e21ebe2a_0

Q: How is tweet subjectivity measured? Text: Introduction A lot of work has been done in the field of Twitter sentiment analysis till date. Sentiment analysis has been handled as a Natural Language Processing task at many levels of granularity. Most of these techniques use Machine Learning algorithms with features such as unigrams, n-grams, Part-Of-Speech (POS) tags. However, the training datasets are often very large, and hence with such a large number of features, this process requires a lot of computation power and time. The following question arises: What to do if we do not have resources that provide such a great amount of computation power? The existing solution to this problem is to use a smaller sample of the dataset. For sentiment analysis, if we train the model using a smaller randomly chosen sample, then we get low accuracy [16, 17]. In this paper, we propose a novel technique to sample tweets for building a sentiment classification model so that we get higher accuracy than the state-of-the-art baseline method, namely Distant Supervision, using a smaller set of tweets. Our model has lower computation time and higher accuracy compared to baseline model. Users often express sentiment using subjective expression. Although objective expressions can also have sentiment, it is much rare. Determining subjectivity is quite efficient compared to determining sentiment. Subjectivity can be determined for individual tweets. But to do sentiment classification, we need to build a classification model with positive and negative sentiment tweets. The time to train a sentiment classification model increases with the increase in the number of training tweets. In this paper, we use tweet subjectivity to select the best training tweets. This not only lowers the computation time but also increases the accuracy because we have training data with less noise. Even the created features will be more relevant to the classification task. The computation cost will reduce due to small training data size and better set of features. Thus if users do not have enough computational resources, they can filter the training dataset using a high value of subjectivf ity threshold. This ensures reliable prediction on a smaller training dataset, and eventually requires less computational time. The above approach, and some of the intricacies that invariably seep in, need to be considered, and are described in the later sections of the paper. In this paper we also integrate a lot of meticulous preprocessing steps. This makes our model more robust, and hence leads to higher accuracy. Along with the machine learning algorithms being used, we use a heuristic-based classification of tweets. This is based on the EFWS of a tweet, which is described in later sections. This heuristic basically takes into account the polarity scores of frequently used words in tweets, and is able to achieve around 85% accuracy on our dataset, hence boosting the overall accuracy by a considerable amount. Our training data consists of generic (not topic-specific) Twitter messages with emoticons, which are used as noisy labels. We show that the accuracy obtained on a training dataset comprising 100K tweets, and a test dataset of 5000 tweets gives an accuracy of around 80% on the following classifiers: Naive Bayes, RBF-kernel Support Vector Machine, and Logistic Regression. Our model takes roughly half the time to train and achieves higher accuracy (than the baseline model) on all the classifiers. Because the amount of training time is expected to increase exponentially as the training data increases, we expect our model to outperform (in terms of higher accuracy) the baseline model at a speed which is at least twofold the speed of the baseline model on larger datasets. Related Work There has been a large amount of prior research in sentiment analysis of tweets. Read [10] shows that using emoticons as labels for positive and sentiment is effective for reducing dependencies in machine learning techniques. Alec Go [1] used Naive Bayes, SVM, and MaxEnt classifiers to train their model. This, as mentioned earlier, is our baseline model. Our model builds on this and achieves higher accuracy on a much smaller training dataset. Ayushi Dalmia [6] proposed a model with a more involved preprocessing stage, and used features like scores from Bing Liu’s Opinion Lexicon, and number of positive, negative POS tags. This model achieved considerably high accuracies considering the fact that their features were the not the conventional bag-of-words, or any n-grams. The thought of using the polarity scores of frequently used tweet words (as described in our EFWS heuristic) was inspired from this work. [14] created prior probabilities using the datasets for the average sentiment of tweets in different spatial, temporal and authorial contexts. They then used a Bayesian approach to combine these priors with standard bigram language models. Another significant effort in sentiment analysis on Twitter data is by Barbosa [16]. They use polarity predictions from three websites as noisy labels to train a model and use 1000 manually labelled tweets for tuning and another 1000 for testing. They propose the use of syntax features of tweets like punctuation, retweet, hashtags, link, and exclamation marks in addition with features like prior polarity of words and POS of words. Some works leveraged the use of existing hashtags in the Twitter data for building the training data. (Davidov, Tsur, and Rappoport 2010) also use hashtags for creating training data, but they limit their experiments to sentiment/non-sentiment classification, rather than 3-way polarity classification, as [15] does. Our model integrates some of the preprocessing techniques this work used. Hassan Saif [9] introduced a novel approach of adding semantics as additional features into the training set for sentiment analysis. This approach works well for topic specific data. Hence, we thought of taking a different approach for a generic tweet dataset like ours. Subjectivity Subjectivity refers to how someone's judgment is shaped by personal opinions and feelings instead of outside influences. An objective perspective is one that is not influenced by emotions, opinions, or personal feelings - it is a perspective based in fact, in things quantifiable and measurable. A subjective perspective is one open to greater interpretation based on personal feeling, emotion, aesthetics, etc. Subjectivity classification is another topic in the domain of text classification which is garnering more and more interest in the field of sentiment analysis. Since a single sentence may contain multiple opinions and subjective and factual clauses, this problem is not as straightforward as it seems. Below are some examples of subjective and objective sentences. Objective sentence with no sentiment: So, the Earth revolves around the Sun. Objective sentence with sentiment: The drug relieved my pain. Subjective sentence with no sentiment: I believe he went home yesterday. Subjective sentence with sentiment: I am so happy you got the scholarship. Classifying a sentence as subjective or objective provides certain conclusions. Purely objective sentences do not usually convey any sentiment, while most of the purely subjective sentences have a clear inclination towards either the positive or negative sentiment. Sentences which are not completely subjective or objective may or may not convey a sentiment. Libraries like TextBlob, and tools like Opinion Finder can be used to find the extent to which a sentence can be considered subjective. Since tweets are usually person-specific, or subjective, we use this intuition to reduce the size of the training set by filtering the sentences with a subjectivity level below a certain threshold (fairly objective tweets). Implementation In this section, we explain the various preprocessing techniques used for feature reduction, and also the additional step of filtering the training dataset using the subjectivity score of tweets. We further describe our approach of using different machine learning classifiers and feature extractors. We also propose an additional heuristic for sentiment classification which can be used as a tag-along with the learning heuristics. Corpus Our training dataset has 1.6 million tweets, and 5000 tweets in the test dataset. Since the test dataset provided comprised only 500 tweets, we have taken part of the training data (exactly 5000 tweets, distinct from the training dataset) as the test dataset. We remove emoticons from our training and test data. The table below shows some sample tweets. Subjectivity Filtering This is a new step we propose to achieve higher accuracy on a smaller training dataset. We use TextBlob to classify each tweet as subjective or objective. We then remove all tweets which have a subjectivity level/score (score lies between 0 and 1) below a specified threshold. The remaining tweets are used for training purposes. We observe that a considerable number of tweets are removed as the subjectivity threshold increases. We show the effect of doing this procedure on the overall accuracy in the evaluation section of the paper. Preprocessing The Twitter language model has many unique properties. We take advantage of the following properties to reduce the feature space. Most of the preprocessing steps are common to most of the previous works in the field. However, we have added some more steps to this stage of our model. We first strip off the emoticons from the data. Users often include twitter usernames in their tweets in order to direct their messages. We also strip off usernames (e.g. @Chinmay) and URLs present in tweets because they do not help us in sentiment classification. Apart from full stops, which are dealt in the next point, other punctuations and special symbols are also removed. Repeated whitespaces are replaced with a single space. We also perform stemming to reduce the size of the feature space. In the previous works, full stops are just usually replaced by a space. However, we have observed that casual language in tweets is often seen in form of repeated punctuations. For example, “this is so cool...wow". We take into consideration this format, and replace two or more occurrences of “." and “-" with a space. Also, full stops are also quite different in usage. Sometimes, there isn't any space in between sentences. For example, “It’s raining.Feeling awesome". We replace a single occurrence of a full stop with a space to ensure correct feature incorporation. In the case of hashtags, most of the previous works just consider the case of hashtags followed by a single word; they just remove the hashtag and add the word to the feature vector. However, sometimes, there are multiple words after a hashtag, and more often than not, these words form an important, conclusive part of the Tweet. For example, #ThisSucks, or #BestMomentEver. These hashtags need to be dealt with in a proper fashion. We split the text after hashtags after before each capital letter, and add these as tokens to the feature vector. For hashtags followed by a single word, we just replace the pattern #word with the word, as conventional models do. The intuition behind this step is that quite often, the sentiment of a tweet is expressed in form of a hashtag. For example, #happy or #disappointed are frequently used hashtags, and we don’t want to lose this information during sentiment classification. Tweets contain very casual language as mentioned earlier. For example, if we search “wow" with an arbitrary number of o's in the middle (e.g. wooow, woooow) on Twitter, there will most likely be a non-empty result set. We use preprocessing so that any letter occurring more than two times in a row is replaced with two occurrences. In the samples above, these words would be converted into the token “woow". After all the above modifications, tweets are converted into lowercase to avoid confusion between features having same content, but are different in capitalization. We gather a list of 400 stopwords. These words, if present in the tweets, are not considered in the feature vector. We store an acronym dictionary which has over 5000, frequently-used acronyms and their abbreviations. We replace such acronyms in tweets with their abbreviation, since these can be of great use while sentiment classification. All negative words like 'cannot', 'can't', 'won't', 'don't' are replaced by 'not', which effectively keeps the sentiment stable. It is observed that doing this makes the training faster, since the model has to deal with a smaller feature vector. Baseline model The baseline model for our experiments is explained in the paper by Alec Go [1]. The model uses the Naive Bayes, SVM, and the Maximum Entropy classifiers for their experiment. Their feature vector is either composed of Unigrams, Bigrams, Unigrams + Bigrams, or Unigrams + POS tags. This work achieved the following maximum accuracies: a) 82.2 for the Unigram feature vector, using the SVM classifier, b) 83.0 for the Unigram + Bigram feature vector, using the MaxEnt classifier, and 82.7 using the Naive Bayes classifier. c) 81.9 for the Unigram + POS feature vector, using the SVM classifier. These baseline accuracies were on a training dataset of 1.6 million tweets, and a test dataset of 500 tweets. We are using the same training dataset for our experiments. We later present the baseline accuracies on a training set of 200K tweets, and a test dataset of 5000 tweets; we compare our model's accuracy with these baseline accuracy values on the same test data of 5000 tweets. Effective Word Score (EFWS) Heuristic We have described our baseline model above. So the feature vectors we collate results for, are Unigram, Unigram + Bigram, and Unigram + POS. We have already made two major changes before the training starts on our dataset as compared to our baseline model. Firstly, our training dataset will be filtered according to the subjectivity threshold. And secondly, our preprocessing is much more robust as compared to their work. Now let us look at an additional heuristic we use to obtain labels for our test data. Along with dictionaries for stop words and acronyms, we also maintain a dictionary of a list of frequently used words and their polarity scores. This dictionary has around 2500 words and their polarity score ranging from -5 to 5. At runtime, we also use all synonyms of a word (from WordNet) present in a tweet and also the dictionary, and assign them the same score as the dictionary word. There is a reasonable assumption here, that the synonyms aren't very extremal in nature, that is, a word with a polarity score of 2 cannot have a synonym which has a polarity score of 5. Now, we calculate the Effective Word Scores of a tweet. We define the Effective Word Score of score x as EFWS(x) = N(+x) - N(-x), where N(x) is the number of words in the tweet with polarity score x. For example, if a tweet has one word with score 5, three words with score 4, two with score 2, three with with score -2, one with score -3, and finally two with score -4, then the effective word scores are: EFWS(5) = N(5) - N(-5) = 1 - 0 = 1 EFWS(4) = N(4) - N(-4) = 3 - 2 = 1 EFWS(3) = N(3) - N(-3) = 0 - 1 = -1 EFWS(2) = N(2) - N(-2) = 2 - 3 = -1 EFWS(1) = N(1) - N(-1) = 2 - 0 = 2 We now define the heuristic for obtaining the label of a Tweet. (EFWS(5) INLINEFORM0 1 or EFWS(4) INLINEFORM1 1) and (EFWS(2) INLINEFORM2 1) Label = positive Similarly, (EFWS(5) INLINEFORM0 -1 or EFWS(4) INLINEFORM1 -1) and (EFWS(2) INLINEFORM2 -1) Label = negative The basic intuition behind such a heuristic is that we found tweets having one strongly positive and one moderately positive word more than the number of strongly negative and the moderately negative words respectively, usually conveyed a positive sentiment. Similar was the case for negative sentiments. The tweets getting a label from this heuristic are not sent into the training phase. After considerable amount of experimenting, and analyzing the nature of our dataset, which is not domain specific, we have reached the conclusion that the heuristic mentioned above is optimal for obtaining labels. We found that the heuristic accuracy was around 85% for a training dataset of 100K and a test dataset of 5K, where the total number of test tweets labelled by the heuristic were around 500. This means that around 425 out of the 500 tweets received a correct prediction of sentiment using this heuristic. Thus, using this heuristic improves the overall accuracy, as well as saves time by reducing the number of tweets to be tested by the ML algorithms. Training Model We use the following classifiers for our model. Naive Bayes is a simple model which works well on text categorization. We use a Naive Bayes model. Class c* is assigned to tweet d, where c* = argmax P(c INLINEFORM0 d). INLINEFORM1 And INLINEFORM0 is calculated using Bayes Rule. In this formula, f represents a feature and INLINEFORM1 represents the count of feature INLINEFORM2 found in tweet d. There are a total of m features. Parameters P(c) and INLINEFORM3 are obtained through maximum likelihood estimates. Support vector machines are based on the Structural Risk Minimization principle from computational learning theory. SVM classification algorithms for binary classification is based on finding a separation between hyperplanes defined by classes of data. One remarkable property of SVMs is that their ability to learn can be independent of the dimensionality of the feature space. SVMs can generalize even in the presence of many features as in the case of text data classification. We use a non-linear Support Vector Machine with an RBF kernel. Maximum Entropy Model belongs to the family of discriminative classifiers also known as the exponential or log-linear classifiers.. In the naive Bayes classifier, Bayes rule is used to estimate this best y indirectly from the likelihood INLINEFORM0 (and the prior INLINEFORM1 ) but a discriminative model takes this direct approach, computing INLINEFORM2 by discriminating among the different possible values of the class y rather than first computing a likelihood. INLINEFORM3 Logistic regression estimates INLINEFORM0 by combining the feature set linearly (multiplying each feature by a weight and adding them up), and then applying a function to this combination. Evaluation In this section, we present the collated results of our experiments. To show that our model achieves higher accuracy than the baseline model and on a smaller training dataset, we first fix the test dataset. Our test dataset, as mentioned before, consists of 5000 tweets. We conducted our experiments on an Intel Core i5 machine (4 cores), with 8 GB RAM. The following are the accuracies of the baseline model on a training set of 200K tweets: We filtered the training set with a subjectivity threshold of 0.5. By doing this, we saw that the number of tweets reduced to approximately 0.6 million tweets from an earlier total of 1.6 million. We then trained our model described in earlier sections on a 100K tweets randomly picked from this filtered training dataset, and observed the following accuracies: Note that all the accuracies in the tables above have been recorded as the average of 3 iterations of our experiment. We achieve higher accuracy for all feature vectors, on all classifiers, and that too from a training dataset half the size of the baseline one. We now see the intricacies of the subjectivity threshold parameter. It is clear that more and more tweets get filtered as the subjectivity threshold parameter increases. This can be seen in the Figure 1 shown below. We have plotted the number of tweets that remain after filtering from two sources: TextBlob, Opinion Finder Tool. TextBlob has an inbuilt function that provides us the subjectivity level of a tweet. On the other hand, Opinion Finder only provides the information of which parts of the text are subjective, and which are objective. From that, we define the subjectivity level of that text as: Subjectivity level = INLINEFORM0 [ xlabel=Subjectivity Threshold, ylabel=Tweets (in millions), xmin=0, xmax=1, ymin=0, ymax=2000000, xtick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, ytick=0,200000,400000,600000,800000,1000000,1200000,1400000,1600000,1800000, legend pos=north east, ] [color=red] coordinates (0, 1600000) (0.1, 939785) (0.2, 873054) (0.3, 804820) (0.4, 712485) (0.5, 571864) (0.6, 449286) (0.7, 304874) (0.8, 211217) (0.9, 135788) ; [color=blue] coordinates (0, 1600000) (0.1, 602313) (0.2, 499173) (0.3, 392223) (0.4, 262109) (0.5, 169477) (0.6, 154667) (0.7, 139613) (0.8, 126148) (0.9, 116842) ; Textblob, Opinion Finder Figure 1: Number of tweets with subjectivity greater than the subjectivity threshold [ xlabel=Subjectivity Threshold, ylabel=Accuracy (from 0 to 1), xmin=0, xmax=1, ymin=0.7, ymax=1, xtick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, ytick=0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1, legend pos=north east, ] [color=red] coordinates (0.1, 0.753871866) (0.2, 0.779442897) (0.3, 0.763421155) (0.4, 0.783231198) (0.5,0.805132645) (0.6,0.807373259) (0.7,0.808587744) (0.8,0.817799443) (0.9,0.823872989) ; Figure 2: Variation of accuracy (*Training data of 100K, Test data of 5K) with subjectivity threshold. *TextBlob is used to filter the tweets to form the training dataset. We now focus on the issue of choosing the optimum threshold value. As the subjectivity threshold parameter increases, our model trains on tweets with a higher subjectivity level, and the overall accuracy increases. We observed the following accuracies on subjectivity level 0.8 (Unigrams as features): Naive Bayes: 80.32% Non-linear SVM: 80.15 % Logistic Regression: 81.77% We should consider the fact that a lot of useful tweets are also lost in the process of gradually increasing the parameter, and this could cause a problem in cases when the test data is very large, because the model will not train on a generic dataset. Researchers may use a higher subjectivity threshold for their experiments if they are confident that most of the important information would be retained. This is most likely to happen in case of topic-specific or domain-specific data. [ ybar, enlargelimits=0.15, legend style=anchor=north, legend pos= north east, ylabel=Training time (in minutes), symbolic x coords=baseline,subjectivity=0.5,subjectivity=0.8, xtick=data, ] coordinates (baseline,17.4) (subjectivity=0.5,12.55) (subjectivity=0.8,10.68); coordinates (baseline,16.23) (subjectivity=0.5,12.31) (subjectivity=0.8,10.34); coordinates (baseline,31.9) (subjectivity=0.5,18.24) (subjectivity=0.8,16.3); Logistic Regression,Naive Bayes,SVM Figure 3: Comparison of training times for Unigrams [ ybar, enlargelimits=0.15, legend style=anchor=north, legend pos= north east, ylabel=Training time (in minutes), symbolic x coords=baseline,subjectivity=0.5,subjectivity=0.8, xtick=data, ] coordinates (baseline,28.41) (subjectivity=0.5,14.09) (subjectivity=0.8,11.3); coordinates (baseline,16.6) (subjectivity=0.5,13.51) (subjectivity=0.8,12.66); coordinates (baseline,35.2) (subjectivity=0.5,20.6) (subjectivity=0.8,19.2); Logistic Regression,Naive Bayes,SVM Figure 4: Comparison of training times for Unigrams + Bigrams We use Logistic regression for classification and unigrams as the feature vector with K-fold cross validation for determining the accuracy. We choose an optimal threshold value of 0.5 for our experiment, considering the fact that the model should train on a more generic dataset. Figure 2 shows the variation of accuracy with the subjectivity threshold. The training size is fixed at 100K and the test dataset (5K tweets) is also same for all the experiments. We also measure the time taken to train our model, and compare it to the baseline model. Our observation was that our model took roughly half the amount of time in some cases and yet obtained a higher accuracy. Figures 3 and 4 show the difference in training time of the baseline model, our model on a 0.5 subjectivity-filtered dataset, and our model on a 0.8 subjectivity-filtered dataset on unigrams and unigrams + bigrams respectively. The times recorded are on a training dataset of 100K for our model and 200K for the baseline model, and a test dataset of 5K was fixed in all the recordings. The winning point, which can be seen from the plots, is that our model is considerably faster, and even has twofold speed in some cases. And alongside saving computation time, it achieves higher accuracy. This can be attributed to the fact that as the subjectivity threshold increases, only the tweets with highly polar words are retained in the training set and this makes the whole process faster. Conclusion We show that a higher accuracy can be obtained in sentiment classification of Twitter messages training on a smaller dataset and with a much faster computation time, and hence the issue of constraint on computation power is resolved to a certain extent. This can be achieved using a subjectivity threshold to selectively filter the training data, incorporating a more complex preprocessing stage, and using an additional heuristic for sentiment classification, along with the conventional machine learning techniques. As Twitter data is abundant, our subjectivity filtering process can achieve a better generalised model for sentiment classification.

Unanswerable

129c03acb0963ede3915415953317556a55f34ee

129c03acb0963ede3915415953317556a55f34ee_0

Q: Why is supporting fact supervision necessary for DMN? Text: Introduction Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 . The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging. We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set. In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 . Dynamic Memory Networks We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 . The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows: Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words. As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by $$u_i &=& \sigma \left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \right)\\ r_i &=& \sigma \left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \right)\\ \tilde{h}_i &=& \tanh \left(Wx_{i} + r_i \circ U h_{i-1} + b^{(h)}\right)\\ h_i &=& u_i\circ \tilde{h}_i + (1-u_i) \circ h_{i-1}$$ (Eq. 2) where $\sigma $ is the sigmoid activation function, $\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \in \mathbb {R}^{n_H \times n_I}$ , $U^{(z)}, U^{(r)}, U \in \mathbb {R}^{n_H \times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size. Question Module: This module computes a vector representation $q$ of the question, where $q \in \mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question. Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \in \mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ . The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \in \mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \in \mathbb {R}^{n_H}$0 . Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network. Improved Dynamic Memory Networks: DMN+ We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset. Input Module for Text QA In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. "Model Analysis" ). We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU. Input Fusion Layer For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction. Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence. The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. For the positional encoding scheme, the sentence representation is produced by $f_i = \sum ^{j=1}_M l_j \circ w^i_j$ , where $\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding. The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU. $$\overrightarrow{f_i} = GRU_{fwd}(f_i, \overrightarrow{f_{i-1}}) \\ \overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \overleftarrow{f_{i+1}}) \\ \overleftrightarrow{f_i} = \overleftarrow{f_i} + \overrightarrow{f_i}$$ (Eq. 5) where $f_i$ is the input fact at timestep $i$ , $ \overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\overleftrightarrow{f_i}$ . We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. Input Module for VQA To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. "Input Module for Text QA" . Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \times 14 \times 14$ . The pooling layer divides the image into a grid of $14 \times 14$ , resulting in 196 local regional vectors of $d = 512$ . Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ . Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems. To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. "Input Module for Text QA" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information. The Episodic Memory Module The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\overleftrightarrow{F} = [\overleftrightarrow{f_1}, \hdots , \overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state. $$z^t_i &=& [\overleftrightarrow{f_i} \circ q; \overleftrightarrow{f_i} \circ m^{t-1}; \vert \overleftrightarrow{f_i} - q \vert ; \vert \overleftrightarrow{f_i} - m^{t-1} \vert ] \\ Z^t_i &=& W^{(2)} \tanh \left(W^{(1)}z^t_i + b^{(1)} \right)+ b^{(2)} \\ g^t_i &=& \frac{\exp (Z^t_i)}{\sum _{k=1}^{M_i} \exp (Z^t_k)} $$ (Eq. 10) where $\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\circ $ is the element-wise product, $|\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors. The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required. Attention Mechanism Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+. Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \sum _{i=1}^N g^t_i \overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient. Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation . We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory. By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 . $$h_i &=& g^t_i \circ \tilde{h}_i + (1-g^t_i) \circ h_{i-1}$$ (Eq. 12) An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \in \mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \in \mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU. Episode Memory Updates After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by $$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13) The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the “Mem Weights” row in Table 1 . Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by $$m^t = ReLU\left(W^t [m^{t-1} ; c^t ; q] + b\right)$$ (Eq. 14) where $;$ is the concatenation operator, $W^t \in \mathbb {R}^{n_H \times n_H}$ , $b \in \mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN. Related Work The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities. Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps. Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 . Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 . Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset. Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 . Datasets To analyze our proposed model changes and compare our performance with other architectures, we use three datasets. bAbI-10k For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k. DAQUAR-ALL visual dataset The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. "Model Analysis" ). Visual Question Answering The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 . Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task. Model Analysis To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets. The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. "Input Module for Text QA" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. "The Episodic Memory Module" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 . A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required. The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases. From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets. Comparison to state of the art using bAbI-10k We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\sqrt{3}, \sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 . On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss. Text QA Results We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training. The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component. The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time. In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset. One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks. The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting. Comparison to state of the art using VQA For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \frac{1}{N}\sum _{i=1}^Nmin(\frac{\sum _{t\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\%$ accurate if at least 3 people provide that exact answer. Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ . Results and Analysis The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities. The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach. Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image. However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries. Conclusion We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains.

First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU.

58b3b630a31fcb9bffb510390e1ec30efe87bfbf

58b3b630a31fcb9bffb510390e1ec30efe87bfbf_0

Q: What does supporting fact supervision mean? Text: Introduction Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 . The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging. We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set. In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 . Dynamic Memory Networks We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 . The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows: Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words. As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by $$u_i &=& \sigma \left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \right)\\ r_i &=& \sigma \left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \right)\\ \tilde{h}_i &=& \tanh \left(Wx_{i} + r_i \circ U h_{i-1} + b^{(h)}\right)\\ h_i &=& u_i\circ \tilde{h}_i + (1-u_i) \circ h_{i-1}$$ (Eq. 2) where $\sigma $ is the sigmoid activation function, $\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \in \mathbb {R}^{n_H \times n_I}$ , $U^{(z)}, U^{(r)}, U \in \mathbb {R}^{n_H \times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size. Question Module: This module computes a vector representation $q$ of the question, where $q \in \mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question. Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \in \mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ . The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \in \mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \in \mathbb {R}^{n_H}$0 . Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network. Improved Dynamic Memory Networks: DMN+ We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset. Input Module for Text QA In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. "Model Analysis" ). We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU. Input Fusion Layer For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction. Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence. The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. For the positional encoding scheme, the sentence representation is produced by $f_i = \sum ^{j=1}_M l_j \circ w^i_j$ , where $\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding. The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU. $$\overrightarrow{f_i} = GRU_{fwd}(f_i, \overrightarrow{f_{i-1}}) \\ \overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \overleftarrow{f_{i+1}}) \\ \overleftrightarrow{f_i} = \overleftarrow{f_i} + \overrightarrow{f_i}$$ (Eq. 5) where $f_i$ is the input fact at timestep $i$ , $ \overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\overleftrightarrow{f_i}$ . We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. Input Module for VQA To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. "Input Module for Text QA" . Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \times 14 \times 14$ . The pooling layer divides the image into a grid of $14 \times 14$ , resulting in 196 local regional vectors of $d = 512$ . Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ . Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems. To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. "Input Module for Text QA" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information. The Episodic Memory Module The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\overleftrightarrow{F} = [\overleftrightarrow{f_1}, \hdots , \overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state. $$z^t_i &=& [\overleftrightarrow{f_i} \circ q; \overleftrightarrow{f_i} \circ m^{t-1}; \vert \overleftrightarrow{f_i} - q \vert ; \vert \overleftrightarrow{f_i} - m^{t-1} \vert ] \\ Z^t_i &=& W^{(2)} \tanh \left(W^{(1)}z^t_i + b^{(1)} \right)+ b^{(2)} \\ g^t_i &=& \frac{\exp (Z^t_i)}{\sum _{k=1}^{M_i} \exp (Z^t_k)} $$ (Eq. 10) where $\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\circ $ is the element-wise product, $|\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors. The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required. Attention Mechanism Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+. Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \sum _{i=1}^N g^t_i \overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient. Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation . We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory. By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 . $$h_i &=& g^t_i \circ \tilde{h}_i + (1-g^t_i) \circ h_{i-1}$$ (Eq. 12) An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \in \mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \in \mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU. Episode Memory Updates After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by $$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13) The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the “Mem Weights” row in Table 1 . Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by $$m^t = ReLU\left(W^t [m^{t-1} ; c^t ; q] + b\right)$$ (Eq. 14) where $;$ is the concatenation operator, $W^t \in \mathbb {R}^{n_H \times n_H}$ , $b \in \mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN. Related Work The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities. Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps. Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 . Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 . Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset. Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 . Datasets To analyze our proposed model changes and compare our performance with other architectures, we use three datasets. bAbI-10k For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k. DAQUAR-ALL visual dataset The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. "Model Analysis" ). Visual Question Answering The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 . Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task. Model Analysis To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets. The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. "Input Module for Text QA" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. "The Episodic Memory Module" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 . A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required. The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases. From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets. Comparison to state of the art using bAbI-10k We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\sqrt{3}, \sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 . On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss. Text QA Results We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training. The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component. The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time. In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset. One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks. The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting. Comparison to state of the art using VQA For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \frac{1}{N}\sum _{i=1}^Nmin(\frac{\sum _{t\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\%$ accurate if at least 3 people provide that exact answer. Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ . Results and Analysis The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities. The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach. Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image. However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries. Conclusion We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains.

the facts that are relevant for answering a particular question) are labeled during training.

141dab98d19a070f1ce7e7dc384001d49125d545

141dab98d19a070f1ce7e7dc384001d49125d545_0

Q: What changes they did on input module? Text: Introduction Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 . The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging. We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set. In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 . Dynamic Memory Networks We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 . The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows: Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words. As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by $$u_i &=& \sigma \left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \right)\\ r_i &=& \sigma \left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \right)\\ \tilde{h}_i &=& \tanh \left(Wx_{i} + r_i \circ U h_{i-1} + b^{(h)}\right)\\ h_i &=& u_i\circ \tilde{h}_i + (1-u_i) \circ h_{i-1}$$ (Eq. 2) where $\sigma $ is the sigmoid activation function, $\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \in \mathbb {R}^{n_H \times n_I}$ , $U^{(z)}, U^{(r)}, U \in \mathbb {R}^{n_H \times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size. Question Module: This module computes a vector representation $q$ of the question, where $q \in \mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question. Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \in \mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ . The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \in \mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \in \mathbb {R}^{n_H}$0 . Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network. Improved Dynamic Memory Networks: DMN+ We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset. Input Module for Text QA In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. "Model Analysis" ). We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU. Input Fusion Layer For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction. Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence. The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. For the positional encoding scheme, the sentence representation is produced by $f_i = \sum ^{j=1}_M l_j \circ w^i_j$ , where $\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding. The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU. $$\overrightarrow{f_i} = GRU_{fwd}(f_i, \overrightarrow{f_{i-1}}) \\ \overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \overleftarrow{f_{i+1}}) \\ \overleftrightarrow{f_i} = \overleftarrow{f_i} + \overrightarrow{f_i}$$ (Eq. 5) where $f_i$ is the input fact at timestep $i$ , $ \overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\overleftrightarrow{f_i}$ . We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. Input Module for VQA To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. "Input Module for Text QA" . Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \times 14 \times 14$ . The pooling layer divides the image into a grid of $14 \times 14$ , resulting in 196 local regional vectors of $d = 512$ . Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ . Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems. To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. "Input Module for Text QA" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information. The Episodic Memory Module The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\overleftrightarrow{F} = [\overleftrightarrow{f_1}, \hdots , \overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state. $$z^t_i &=& [\overleftrightarrow{f_i} \circ q; \overleftrightarrow{f_i} \circ m^{t-1}; \vert \overleftrightarrow{f_i} - q \vert ; \vert \overleftrightarrow{f_i} - m^{t-1} \vert ] \\ Z^t_i &=& W^{(2)} \tanh \left(W^{(1)}z^t_i + b^{(1)} \right)+ b^{(2)} \\ g^t_i &=& \frac{\exp (Z^t_i)}{\sum _{k=1}^{M_i} \exp (Z^t_k)} $$ (Eq. 10) where $\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\circ $ is the element-wise product, $|\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors. The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required. Attention Mechanism Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+. Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \sum _{i=1}^N g^t_i \overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient. Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation . We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory. By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 . $$h_i &=& g^t_i \circ \tilde{h}_i + (1-g^t_i) \circ h_{i-1}$$ (Eq. 12) An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \in \mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \in \mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU. Episode Memory Updates After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by $$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13) The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the “Mem Weights” row in Table 1 . Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by $$m^t = ReLU\left(W^t [m^{t-1} ; c^t ; q] + b\right)$$ (Eq. 14) where $;$ is the concatenation operator, $W^t \in \mathbb {R}^{n_H \times n_H}$ , $b \in \mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN. Related Work The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities. Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps. Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 . Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 . Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset. Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 . Datasets To analyze our proposed model changes and compare our performance with other architectures, we use three datasets. bAbI-10k For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k. DAQUAR-ALL visual dataset The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. "Model Analysis" ). Visual Question Answering The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 . Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task. Model Analysis To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets. The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. "Input Module for Text QA" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. "The Episodic Memory Module" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 . A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required. The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases. From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets. Comparison to state of the art using bAbI-10k We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\sqrt{3}, \sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 . On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss. Text QA Results We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training. The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component. The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time. In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset. One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks. The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting. Comparison to state of the art using VQA For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \frac{1}{N}\sum _{i=1}^Nmin(\frac{\sum _{t\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\%$ accurate if at least 3 people provide that exact answer. Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ . Results and Analysis The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities. The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach. Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image. However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries. Conclusion We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains.

For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, The second component is the input fusion layer

afdad4c9bdebf88630262f1a9a86ac494f06c4c1

afdad4c9bdebf88630262f1a9a86ac494f06c4c1_0

Q: What improvements they did for DMN? Text: Introduction Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 . The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging. We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set. In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 . Dynamic Memory Networks We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 . The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows: Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words. As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by $$u_i &=& \sigma \left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \right)\\ r_i &=& \sigma \left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \right)\\ \tilde{h}_i &=& \tanh \left(Wx_{i} + r_i \circ U h_{i-1} + b^{(h)}\right)\\ h_i &=& u_i\circ \tilde{h}_i + (1-u_i) \circ h_{i-1}$$ (Eq. 2) where $\sigma $ is the sigmoid activation function, $\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \in \mathbb {R}^{n_H \times n_I}$ , $U^{(z)}, U^{(r)}, U \in \mathbb {R}^{n_H \times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size. Question Module: This module computes a vector representation $q$ of the question, where $q \in \mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question. Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \in \mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ . The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \in \mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \in \mathbb {R}^{n_H}$0 . Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network. Improved Dynamic Memory Networks: DMN+ We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset. Input Module for Text QA In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. "Model Analysis" ). We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU. Input Fusion Layer For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction. Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence. The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. For the positional encoding scheme, the sentence representation is produced by $f_i = \sum ^{j=1}_M l_j \circ w^i_j$ , where $\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding. The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU. $$\overrightarrow{f_i} = GRU_{fwd}(f_i, \overrightarrow{f_{i-1}}) \\ \overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \overleftarrow{f_{i+1}}) \\ \overleftrightarrow{f_i} = \overleftarrow{f_i} + \overrightarrow{f_i}$$ (Eq. 5) where $f_i$ is the input fact at timestep $i$ , $ \overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\overleftrightarrow{f_i}$ . We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. Input Module for VQA To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. "Input Module for Text QA" . Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \times 14 \times 14$ . The pooling layer divides the image into a grid of $14 \times 14$ , resulting in 196 local regional vectors of $d = 512$ . Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ . Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems. To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. "Input Module for Text QA" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information. The Episodic Memory Module The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\overleftrightarrow{F} = [\overleftrightarrow{f_1}, \hdots , \overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state. $$z^t_i &=& [\overleftrightarrow{f_i} \circ q; \overleftrightarrow{f_i} \circ m^{t-1}; \vert \overleftrightarrow{f_i} - q \vert ; \vert \overleftrightarrow{f_i} - m^{t-1} \vert ] \\ Z^t_i &=& W^{(2)} \tanh \left(W^{(1)}z^t_i + b^{(1)} \right)+ b^{(2)} \\ g^t_i &=& \frac{\exp (Z^t_i)}{\sum _{k=1}^{M_i} \exp (Z^t_k)} $$ (Eq. 10) where $\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\circ $ is the element-wise product, $|\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors. The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required. Attention Mechanism Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+. Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \sum _{i=1}^N g^t_i \overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient. Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation . We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory. By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 . $$h_i &=& g^t_i \circ \tilde{h}_i + (1-g^t_i) \circ h_{i-1}$$ (Eq. 12) An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \in \mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \in \mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU. Episode Memory Updates After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by $$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13) The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the “Mem Weights” row in Table 1 . Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by $$m^t = ReLU\left(W^t [m^{t-1} ; c^t ; q] + b\right)$$ (Eq. 14) where $;$ is the concatenation operator, $W^t \in \mathbb {R}^{n_H \times n_H}$ , $b \in \mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN. Related Work The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities. Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps. Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 . Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 . Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset. Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 . Datasets To analyze our proposed model changes and compare our performance with other architectures, we use three datasets. bAbI-10k For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k. DAQUAR-ALL visual dataset The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. "Model Analysis" ). Visual Question Answering The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 . Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task. Model Analysis To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets. The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. "Input Module for Text QA" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. "The Episodic Memory Module" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 . A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required. The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases. From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets. Comparison to state of the art using bAbI-10k We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\sqrt{3}, \sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 . On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss. Text QA Results We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training. The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component. The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time. In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset. One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks. The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting. Comparison to state of the art using VQA For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \frac{1}{N}\sum _{i=1}^Nmin(\frac{\sum _{t\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\%$ accurate if at least 3 people provide that exact answer. Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ . Results and Analysis The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities. The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach. Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image. However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries. Conclusion We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains.

the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training., In addition, we introduce a new input module to represent images.

bfd4fc82ffdc5b2b32c37f4222e878106421ce2a

bfd4fc82ffdc5b2b32c37f4222e878106421ce2a_0

Q: How does the model circumvent the lack of supporting facts during training? Text: Introduction Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 . The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging. We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set. In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 . Dynamic Memory Networks We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 . The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows: Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words. As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by $$u_i &=& \sigma \left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \right)\\ r_i &=& \sigma \left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \right)\\ \tilde{h}_i &=& \tanh \left(Wx_{i} + r_i \circ U h_{i-1} + b^{(h)}\right)\\ h_i &=& u_i\circ \tilde{h}_i + (1-u_i) \circ h_{i-1}$$ (Eq. 2) where $\sigma $ is the sigmoid activation function, $\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \in \mathbb {R}^{n_H \times n_I}$ , $U^{(z)}, U^{(r)}, U \in \mathbb {R}^{n_H \times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size. Question Module: This module computes a vector representation $q$ of the question, where $q \in \mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question. Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \in \mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ . The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \in \mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \in \mathbb {R}^{n_H}$0 . Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network. Improved Dynamic Memory Networks: DMN+ We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset. Input Module for Text QA In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. "Model Analysis" ). We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU. Input Fusion Layer For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction. Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence. The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. For the positional encoding scheme, the sentence representation is produced by $f_i = \sum ^{j=1}_M l_j \circ w^i_j$ , where $\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding. The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU. $$\overrightarrow{f_i} = GRU_{fwd}(f_i, \overrightarrow{f_{i-1}}) \\ \overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \overleftarrow{f_{i+1}}) \\ \overleftrightarrow{f_i} = \overleftarrow{f_i} + \overrightarrow{f_i}$$ (Eq. 5) where $f_i$ is the input fact at timestep $i$ , $ \overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\overleftrightarrow{f_i}$ . We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. Input Module for VQA To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. "Input Module for Text QA" . Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \times 14 \times 14$ . The pooling layer divides the image into a grid of $14 \times 14$ , resulting in 196 local regional vectors of $d = 512$ . Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ . Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems. To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. "Input Module for Text QA" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information. The Episodic Memory Module The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\overleftrightarrow{F} = [\overleftrightarrow{f_1}, \hdots , \overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state. $$z^t_i &=& [\overleftrightarrow{f_i} \circ q; \overleftrightarrow{f_i} \circ m^{t-1}; \vert \overleftrightarrow{f_i} - q \vert ; \vert \overleftrightarrow{f_i} - m^{t-1} \vert ] \\ Z^t_i &=& W^{(2)} \tanh \left(W^{(1)}z^t_i + b^{(1)} \right)+ b^{(2)} \\ g^t_i &=& \frac{\exp (Z^t_i)}{\sum _{k=1}^{M_i} \exp (Z^t_k)} $$ (Eq. 10) where $\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\circ $ is the element-wise product, $|\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors. The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required. Attention Mechanism Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+. Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \sum _{i=1}^N g^t_i \overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient. Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation . We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory. By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 . $$h_i &=& g^t_i \circ \tilde{h}_i + (1-g^t_i) \circ h_{i-1}$$ (Eq. 12) An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \in \mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \in \mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU. Episode Memory Updates After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by $$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13) The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the “Mem Weights” row in Table 1 . Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by $$m^t = ReLU\left(W^t [m^{t-1} ; c^t ; q] + b\right)$$ (Eq. 14) where $;$ is the concatenation operator, $W^t \in \mathbb {R}^{n_H \times n_H}$ , $b \in \mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN. Related Work The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities. Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps. Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 . Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 . Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset. Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 . Datasets To analyze our proposed model changes and compare our performance with other architectures, we use three datasets. bAbI-10k For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k. DAQUAR-ALL visual dataset The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. "Model Analysis" ). Visual Question Answering The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 . Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task. Model Analysis To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets. The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. "Input Module for Text QA" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. "The Episodic Memory Module" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 . A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required. The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases. From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets. Comparison to state of the art using bAbI-10k We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\sqrt{3}, \sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 . On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss. Text QA Results We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training. The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component. The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time. In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset. One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks. The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting. Comparison to state of the art using VQA For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \frac{1}{N}\sum _{i=1}^Nmin(\frac{\sum _{t\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\%$ accurate if at least 3 people provide that exact answer. Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ . Results and Analysis The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities. The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach. Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image. However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries. Conclusion We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains.

the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs.

1ce26783f0ff38925bfc07bbbb65d206e52c2d21

1ce26783f0ff38925bfc07bbbb65d206e52c2d21_0

Q: Does the DMN+ model establish state-of-the-art ? Text: Introduction Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 . The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging. We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set. In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 . Dynamic Memory Networks We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 . The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows: Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words. As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by $$u_i &=& \sigma \left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \right)\\ r_i &=& \sigma \left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \right)\\ \tilde{h}_i &=& \tanh \left(Wx_{i} + r_i \circ U h_{i-1} + b^{(h)}\right)\\ h_i &=& u_i\circ \tilde{h}_i + (1-u_i) \circ h_{i-1}$$ (Eq. 2) where $\sigma $ is the sigmoid activation function, $\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \in \mathbb {R}^{n_H \times n_I}$ , $U^{(z)}, U^{(r)}, U \in \mathbb {R}^{n_H \times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size. Question Module: This module computes a vector representation $q$ of the question, where $q \in \mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question. Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \in \mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ . The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \in \mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \in \mathbb {R}^{n_H}$0 . Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network. Improved Dynamic Memory Networks: DMN+ We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset. Input Module for Text QA In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. "Model Analysis" ). We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU. Input Fusion Layer For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction. Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence. The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. For the positional encoding scheme, the sentence representation is produced by $f_i = \sum ^{j=1}_M l_j \circ w^i_j$ , where $\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding. The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU. $$\overrightarrow{f_i} = GRU_{fwd}(f_i, \overrightarrow{f_{i-1}}) \\ \overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \overleftarrow{f_{i+1}}) \\ \overleftrightarrow{f_i} = \overleftarrow{f_i} + \overrightarrow{f_i}$$ (Eq. 5) where $f_i$ is the input fact at timestep $i$ , $ \overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\overleftrightarrow{f_i}$ . We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used. Input Module for VQA To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. "Input Module for Text QA" . Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \times 14 \times 14$ . The pooling layer divides the image into a grid of $14 \times 14$ , resulting in 196 local regional vectors of $d = 512$ . Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ . Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems. To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. "Input Module for Text QA" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information. The Episodic Memory Module The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\overleftrightarrow{F} = [\overleftrightarrow{f_1}, \hdots , \overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state. $$z^t_i &=& [\overleftrightarrow{f_i} \circ q; \overleftrightarrow{f_i} \circ m^{t-1}; \vert \overleftrightarrow{f_i} - q \vert ; \vert \overleftrightarrow{f_i} - m^{t-1} \vert ] \\ Z^t_i &=& W^{(2)} \tanh \left(W^{(1)}z^t_i + b^{(1)} \right)+ b^{(2)} \\ g^t_i &=& \frac{\exp (Z^t_i)}{\sum _{k=1}^{M_i} \exp (Z^t_k)} $$ (Eq. 10) where $\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\circ $ is the element-wise product, $|\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors. The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required. Attention Mechanism Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+. Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \sum _{i=1}^N g^t_i \overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient. Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation . We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory. By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 . $$h_i &=& g^t_i \circ \tilde{h}_i + (1-g^t_i) \circ h_{i-1}$$ (Eq. 12) An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \in \mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \in \mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU. Episode Memory Updates After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by $$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13) The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the “Mem Weights” row in Table 1 . Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by $$m^t = ReLU\left(W^t [m^{t-1} ; c^t ; q] + b\right)$$ (Eq. 14) where $;$ is the concatenation operator, $W^t \in \mathbb {R}^{n_H \times n_H}$ , $b \in \mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN. Related Work The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities. Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps. Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 . Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 . Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset. Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 . Datasets To analyze our proposed model changes and compare our performance with other architectures, we use three datasets. bAbI-10k For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k. DAQUAR-ALL visual dataset The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. "Model Analysis" ). Visual Question Answering The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 . Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task. Model Analysis To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets. The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. "Input Module for Text QA" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. "The Episodic Memory Module" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 . A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required. The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases. From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets. Comparison to state of the art using bAbI-10k We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\sqrt{3}, \sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 . On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss. Text QA Results We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training. The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component. The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time. In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset. One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks. The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting. Comparison to state of the art using VQA For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \frac{1}{N}\sum _{i=1}^Nmin(\frac{\sum _{t\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\%$ accurate if at least 3 people provide that exact answer. Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ . Results and Analysis The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities. The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach. Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image. However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries. Conclusion We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains.

Yes

9213159f874b3bdd9b4de956a88c703aac988411

9213159f874b3bdd9b4de956a88c703aac988411_0

Q: Is this style generator compared to some baseline? Text: Introduction All text has style, whether it be formal or informal, polite or aggressive, colloquial, persuasive, or even robotic. Despite the success of style transfer in image processing BIBREF0, BIBREF1, there has been limited progress in the text domain, where disentangling style from content is particularly difficult. To date, most work in style transfer relies on the availability of meta-data, such as sentiment, authorship, or formality. While meta-data can provide insight into the style of a text, it often conflates style with content, limiting the ability to perform style transfer while preserving content. Generalizing style transfer requires separating style from the meaning of the text itself. The study of literary style can guide us. For example, in the digital humanities and its subfield of stylometry, content doesn't figure prominently in practical methods of discriminating authorship and genres, which can be thought of as style at the level of the individual and population, respectively. Rather, syntactic and functional constructions are the most salient features. In this work, we turn to literary style as a test-bed for style transfer, and build on work from literature scholars using computational techniques for analysis. In particular we draw on stylometry: the use of surface level features, often counts of function words, to discriminate between literary styles. Stylometry first saw success in attributing authorship to the disputed Federalist Papers BIBREF2, but is recently used by scholars to study things such as the birth of genres BIBREF3 and the change of author styles over time BIBREF4. The use of function words is likely not the way writers intend to express style, but they appear to be downstream realizations of higher-level stylistic decisions. We hypothesize that surface-level linguistic features, such as counts of personal pronouns, prepositions, and punctuation, are an excellent definition of literary style, as borne out by their use in the digital humanities, and our own style classification experiments. We propose a controllable neural encoder-decoder model in which these features are modelled explicitly as decoder feature embeddings. In training, the model learns to reconstruct a text using only the content words and the linguistic feature embeddings. We can then transfer arbitrary content words to a new style without parallel data by setting the low-level style feature embeddings to be indicative of the target style. This paper makes the following contributions: A formal model of style as a suite of controllable, low-level linguistic features that are independent of content. An automatic evaluation showing that our model fools a style classifier 84% of the time. A human evaluation with English literature experts, including recommendations for dealing with the entanglement of content with style. Related Work ::: Style Transfer with Parallel Data Following in the footsteps of machine translation, style transfer in text has seen success by using parallel data. BIBREF5 use modern translations of Shakespeare plays to build a modern-to-Shakespearan model. BIBREF6 compile parallel data for formal and informal sentences, allowing them to successfully use various machine translation techniques. While parallel data may work for very specific styles, the difficulty of finding parallel texts dramatically limits this approach. Related Work ::: Style Transfer without Parallel Data There has been a decent amount of work on this approach in the past few years BIBREF7, BIBREF8, mostly focusing on variations of an encoder-decoder framework in which style is modeled as a monolithic style embedding. The main obstacle is often to disentangle style and content. However, it remains a challenging problem. Perhaps the most successful is BIBREF9, who use a de-noising auto encoder and back translation to learn style without parallel data. BIBREF10 outline the benefits of automatically extracting style, and suggest there is a formal weakness of using linguistic heuristics. In contrast, we believe that monolithic style embeddings don't capture the existing knowledge we have about style, and will struggle to disentangle content. Related Work ::: Controlling Linguistic Features Several papers have worked on controlling style when generating sentences from restaurant meaning representations BIBREF11, BIBREF12. In each of these cases, the diversity in outputs is quite small given the constraints of the meaning representation, style is often constrained to interjections (like “yeah”), and there is no original style from which to transfer. BIBREF13 investigate using stylistic parameters and content parameters to control text generation using a movie review dataset. Their stylistic parameters are created using word-level heuristics and they are successful in controlling these parameters in the outputs. Their success bodes well for our related approach in a style transfer setting, in which the content (not merely content parameters) is held fixed. Related Work ::: Stylometry and the Digital Humanities Style, in literary research, is anything but a stable concept, but it nonetheless has a long tradition of study in the digital humanities. In a remarkably early quantitative study of literature, BIBREF14 charts sentence-level stylistic attributes specific to a number of novelists. Half a century later, BIBREF15 builds on earlier work in information theory by BIBREF16, and defines a literary text as consisting of two “materials": “the vocabulary, and some structural properties, the style, of its author." Beginning with BIBREF2, statistical approaches to style, or stylometry, join the already-heated debates over the authorship of literary works. A noteable example of this is the “Delta" measure, which uses z-scores of function word frequencies BIBREF17. BIBREF18 find that Shakespeare added some material to a later edition of Thomas Kyd's The Spanish Tragedy, and that Christopher Marlowe collaborated with Shakespeare on Henry VI. Models ::: Preliminary Classification Experiments The stylometric research cited above suggests that the most frequently used words, e.g. function words, are most discriminating of authorship and literary style. We investigate these claims using three corpora that have distinctive styles in the literary community: gothic novels, philosophy books, and pulp science fiction, hereafter sci-fi. We retrieve gothic novels and philosophy books from Project Gutenberg and pulp sci-fi from Internet Archive's Pulp Magazine Archive. We partition this corpus into train, validation, and test sets the sizes of which can be found in Table TABREF12. In order to validate the above claims, we train five different classifiers to predict the literary style of sentences from our corpus. Each classifier has gradually more content words replaced with part-of-speech (POS) tag placeholder tokens. The All model is trained on sentences with all proper nouns replaced by `PROPN'. The models Ablated N, Ablated NV, and Ablated NVA replace nouns, nouns & verbs, and nouns, verbs, & adjectives with the corresponding POS tag respectively. Finally, Content-only is trained on sentences with all words that are not tagged as NOUN, VERB, ADJ removed; the remaining words are not ablated. We train the classifiers on the training set, balancing the class distribution to make sure there are the same number of sentences from each style. Classifiers are trained using fastText BIBREF19, using tri-gram features with all other settings as default. table:classifiers shows the accuracies of the classifiers. The styles are highly distinctive: the All classifier has an accuracy of 86%. Additionally, even the Ablated NVA is quite successful, with 75% accuracy, even without access to any content words. The Content only classifier is also quite successful, at 80% accuracy. This indicates that these stylistic genres are distinctive at both the content level and at the syntactic level. Models ::: Formal Model of Style Given that non-content words are distinctive enough for a classifier to determine style, we propose a suite of low-level linguistic feature counts (henceforth, controls) as our formal, content-blind definition of style. The style of a sentence is represented as a vector of counts of closed word classes (like personal pronouns) as well as counts of syntactic features like the number of SBAR non-terminals in its constituency parse, since clause structure has been shown to be indicative of style BIBREF20. Controls are extracted heuristically, and almost all rely on counts of pre-defined word lists. For constituency parses we use the Stanford Parser BIBREF21. table:controlexamples lists all the controls along with examples. Models ::: Formal Model of Style ::: Reconstruction Task Models are trained with a reconstruction task, in which a distorted version of a reference sentence is input and the goal is to output the original reference. fig:sentenceinput illustrates the process. Controls are calculated heuristically. All words found in the control word lists are then removed from the reference sentence. The remaining words, which represent the content, are used as input into the model, along with their POS tags and lemmas. In this way we encourage models to construct a sentence using content and style independently. This will allow us to vary the stylistic controls while keeping the content constant, and successfully perform style transfer. When generating a new sentence, the controls correspond to the counts of the corresponding syntactic features that we expect to be realized in the output. Models ::: Neural Architecture We implement our feature controlled language model using a neural encoder-decoder with attention BIBREF22, using 2-layer uni-directional gated recurrent units (GRUs) for the encoder and decoder BIBREF23. The input to the encoder is a sequence of $M$ content words, along with their lemmas, and fine and coarse grained part-of-speech (POS) tags, i.e. $X_{.,j} = (x_{1,j},\ldots ,x_{M,j})$ for $j \in \mathcal {T} = \lbrace \textrm {word, lemma, fine-pos, coarse-pos}\rbrace $. We embed each token (and its lemma and POS) before concatenating, and feeding into the encoder GRU to obtain encoder hidden states, $ c_i = \operatorname{gru}(c_{i-1}, \left[E_j(X_{i,j}), \; j\in \mathcal {T} \right]; \omega _{enc}) $ for $i \in {1,\ldots ,M},$ where initial state $c_0$, encoder GRU parameters $\omega _{enc}$ and embedding matrices $E_j$ are learned parameters. The decoder sequentially generates the outputs, i.e. a sequence of $N$ tokens $y =(y_1,\ldots ,y_N)$, where all tokens $y_i$ are drawn from a finite output vocabulary $\mathcal {V}$. To generate the each token we first embed the previously generated token $y_{i-1}$ and a vector of $K$ control features $z = ( z_1,\ldots , z_K)$ (using embedding matrices $E_{dec}$ and $E_{\textrm {ctrl-1}}, \ldots , E_{\textrm {ctrl-K}}$ respectively), before concatenating them into a vector $\rho _i,$ and feeding them into the decoder side GRU along with the previous decoder state $h_{i-1}$: where $\omega _{dec}$ are the decoder side GRU parameters. Using the decoder hidden state $h_i$ we then attend to the encoder context vectors $c_j$, computing attention scores $\alpha _{i,j}$, where before passing $h_i$ and the attention weighted context $\bar{c}_i=\sum _{j=1}^M \alpha _{i,j} c_j$ into a single hidden-layer perceptron with softmax output to compute the next token prediction probability, where $W,U,V$ and $u,v, \nu $ are parameter matrices and vectors respectively. Crucially, the controls $z$ remain fixed for all input decoder steps. Each $z_k$ represents the frequency of one of the low-level features described in sec:formalstyle. During training on the reconstruction task, we can observe the full output sequence $y,$ and so we can obtain counts for each control feature directly. Controls receive a different embedding depending on their frequency, where counts of 0-20 each get a unique embedding, and counts greater than 20 are assigned to the same embedding. At test time, we set the values of the controls according to procedure described in Section SECREF25. We use embedding sizes of 128, 128, 64, and 32 for token, lemma, fine, and coarse grained POS embedding matrices respectively. Output token embeddings $E_{dec}$ have size 512, and 50 for the control feature embeddings. We set 512 for all GRU and perceptron output sizes. We refer to this model as the StyleEQ model. See fig:model for a visual depiction of the model. Models ::: Neural Architecture ::: Baseline Genre Model We compare the above model to a similar model, where rather than explicitly represent $K$ features as input, we have $K$ features in the form of a genre embedding, i.e. we learn a genre specific embedding for each of the gothic, scifi, and philosophy genres, as studied in BIBREF8 and BIBREF7. To generate in a specific style, we simply set the appropriate embedding. We use genre embeddings of size 850 which is equivalent to the total size of the $K$ feature embeddings in the StyleEQ model. Models ::: Neural Architecture ::: Training We train both models with minibatch stochastic gradient descent with a learning rate of 0.25, weight decay penalty of 0.0001, and batch size of 64. We also apply dropout with a drop rate of 0.25 to all embedding layers, the GRUs, and preceptron hidden layer. We train for a maximum of 200 epochs, using validation set BLEU score BIBREF26 to select the final model iteration for evaluation. Models ::: Neural Architecture ::: Selecting Controls for Style Transfer In the Baseline model, style transfer is straightforward: given an input sentence in one style, fix the encoder content features while selecting a different genre embedding. In contrast, the StyleEQ model requires selecting the counts for each control. Although there are a variety of ways to do this, we use a method that encourages a diversity of outputs. In order to ensure the controls match the reference sentence in magnitude, we first find all sentences in the target style with the same number of words as the reference sentence. Then, we add the following constraints: the same number of proper nouns, the same number of nouns, the same number of verbs, and the same number of adjectives. We randomly sample $n$ of the remaining sentences, and for each of these `sibling' sentences, we compute the controls. For each of the new controls, we generate a sentence using the original input sentence content features. The generated sentences are then reranked using the length normalized log-likelihood under the model. We can then select the highest scoring sentence as our style-transferred output, or take the top-$k$ when we need a diverse set of outputs. The reason for this process is that although there are group-level distinctive controls for each style, e.g. the high use of punctuation in philosophy books or of first person pronouns in gothic novels, at the sentence level it can understandably be quite varied. This method matches sentences between styles, capturing the natural distribution of the corpora. Automatic Evaluations ::: BLEU Scores & Perplexity In tab:blueperpl we report BLEU scores for the reconstruction of test set sentences from their content and feature representations, as well as the model perplexities of the reconstruction. For both models, we use beam decoding with a beam size of eight. Beam candidates are ranked according to their length normalized log-likelihood. On these automatic measures we see that StyleEQ is better able to reconstruct the original sentences. In some sense this evaluation is mostly a sanity check, as the feature controls contain more locally specific information than the genre embeddings, which say very little about how many specific function words one should expect to see in the output. Automatic Evaluations ::: Feature Control Designing controllable language models is often difficult because of the various dependencies between tokens; when changing one control value it may effect other aspects of the surface realization. For example, increasing the number of conjunctions may effect how the generator places prepositions to compensate for structural changes in the sentence. Since our features are deterministically recoverable, we can perturb an individual control value and check to see that the desired change was realized in the output. Moreover, we can check the amount of change in the other non-perturbed features to measure the independence of the controls. We sample 50 sentences from each genre from the test set. For each sample, we create a perturbed control setting for each control by adding $\delta $ to the original control value. This is done for $\delta \in \lbrace -3, -2, -1, 0, 1, 2, 3\rbrace $, skipping any settings where the new control value would be negative. table:autoeval:ctrl shows the results of this experiment. The Exact column displays the percentage of generated texts that realize the exact number of control features specified by the perturbed control. High percentages in the Exact column indicate greater one-to-one correspondence between the control and surface realization. For example, if the input was “Dracula and Frankenstein and the mummy,” and we change the conjunction feature by $\delta =-1$, an output of “Dracula, Frankenstein and the mummy,” would count towards the Exact category, while “Dracula, Frankenstein, the mummy,” would not. The Direction column specifies the percentage of cases where the generated text produces a changed number of the control features that, while not exactly matching the specified value of the perturbed control, does change from the original in the correct direction. For example, if the input again was “Dracula and Frankenstein and the mummy,” and we change the conjunction feature by $\delta =-1$, both outputs of “Dracula, Frankenstein and the mummy,” and “Dracula, Frankenstein, the mummy,” would count towards Direction. High percentages in Direction mean that we could roughly ensure desired surface realizations by modifying the control by a larger $\delta $. Finally, the Atomic column specifies the percentage of cases where the generated text with the perturbed control only realizes changes to that specific control, while other features remain constant. For example, if the input was “Dracula and Frankenstein in the castle,” and we set the conjunction feature to $\delta =-1$, an output of “Dracula near Frankenstein in the castle,” would not count as Atomic because, while the number of conjunctions did decrease by one, the number of simple preposition changed. An output of “Dracula, Frankenstein in the castle,” would count as Atomic. High percentages in the Atomic column indicate this feature is only loosely coupled to the other features and can be changed without modifying other aspects of the sentence. Controls such as conjunction, determiner, and punctuation are highly controllable, with Exact rates above 80%. But with the exception of the constituency parse features, all controls have high Direction rates, many in the 90s. These results indicate our model successfully controls these features. The fact that the Atomic rates are relatively low is to be expected, as controls are highly coupled – e.g. to increase 1stPer, it is likely another pronoun control will have to decrease. Automatic Evaluations ::: Automatic Classification For each model we look at the classifier prediction accuracy of reconstructed and transferred sentences. In particular we use the Ablated NVA classifier, as this is the most content-blind one. We produce 16 outputs from both the Baseline and StyleEq models. For the Baseline, we use a beam search of size 16. For the StyleEQ model, we use the method described in Section SECREF25 to select 16 `sibling' sentences in the target style, and generated a transferred sentence for each. We look at three different methods for selection: all, which uses all output sentences; top, which selects the top ranked sentence based on the score from the model; and oracle, which selects the sentence with the highest classifier likelihood for the intended style. The reason for the third method, which indeed acts as an oracle, is that using the score from the model didn't always surface a transferred sentence that best reflected the desired style. Partially this was because the model score was mostly a function of how well a transferred sentence reflected the distribution of the training data. But additionally, some control settings are more indicative of a target style than others. The use of the classifier allows us to identify the most suitable control setting for a target style that was roughly compatible with the number of content words. In table:fasttext-results we see the results. Note that for both models, the all and top classification accuracy tends to be quite similar, though for the Baseline they are often almost exactly the same when the Baseline has little to no diversity in the outputs. However, the oracle introduces a huge jump in accuracy for the StyleEQ model, especially compared to the Baseline, partially because the diversity of outputs from StyleEQ is much higher; often the Baseline model produces no diversity – the 16 output sentences may be nearly identical, save a single word or two. It's important to note that neither model uses the classifier in any way except to select the sentence from 16 candidate outputs. What this implies is that lurking within the StyleEQ model outputs are great sentences, even if they are hard to find. In many cases, the StyleEQ model has a classification accuracy above the base rate from the test data, which is 75% (see table:classifiers). Human Evaluation table:cherrypicking shows example outputs for the StyleEQ and Baseline models. Through inspection we see that the StyleEQ model successfully changes syntactic constructions in stylistically distinctive ways, such as increasing syntactic complexity when transferring to philosophy, or changing relevant pronouns when transferring to sci-fi. In contrast, the Baseline model doesn't create outputs that move far from the reference sentence, making only minor modifications such changing the type of a single pronoun. To determine how readers would classify our transferred sentences, we recruited three English Literature PhD candidates, all of whom had passed qualifying exams that included determining both genre and era of various literary texts. Human Evaluation ::: Fluency Evaluation To evaluate the fluency of our outputs, we had the annotators score reference sentences, reconstructed sentences, and transferred sentences on a 0-5 scale, where 0 was incoherent and 5 was a well-written human sentence. table:fluency shows the average fluency of various conditions from all three annotators. Both models have fluency scores around 3. Upon inspection of the outputs, it is clear that many have fluency errors, resulting in ungrammatical sentences. Notably the Baseline often has slightly higher fluency scores than the StyleEQ model. This is likely because the Baseline model is far less constrained in how to construct the output sentence, and upon inspection often reconstructs the reference sentence even when performing style transfer. In contrast, the StyleEQ is encouraged to follow the controls, but can struggle to incorporate these controls into a fluent sentence. The fluency of all outputs is lower than desired. We expect that incorporating pre-trained language models would increase the fluency of all outputs without requiring larger datasets. Human Evaluation ::: Human Classification Each annotator annotated 90 reference sentences (i.e. from the training corpus) with which style they thought the sentence was from. The accuracy on this baseline task for annotators A1, A2, and A3 was 80%, 88%, and 80% respectively, giving us an upper expected bound on the human evaluation. In discussing this task with the annotators, they noted that content is a heavy predictor of genre, and that would certainly confound their annotations. To attempt to mitigate this, we gave them two annotation tasks: which-of-3 where they simply marked which style they thought a sentence was from, and which-of-2 where they were given the original style and marked which style they thought the sentence was transferred into. For each task, each annotator marked 180 sentences: 90 from each model, with an even split across the three genres. Annotators were presented the sentences in a random order, without information about the models. In total, each marked 270 sentences. (Note there were no reconstructions in this annotation task.) table:humanclassifiers shows the results. In both tasks, accuracy of annotators classifying the sentence as its intended style was low. In which-of-3, scores were around 20%, below the chance rate of 33%. In which-of-2, scores were in the 50s, slightly above the chance rate of 50%. This was the case for both models. There was a slight increase in accuracy for the StyleEQ model over the Baseline for which-of-3, but the opposite trend for which-of-2, suggesting these differences are not significant. It's clear that it's hard to fool the annotators. Introspecting on their approach, the annotators expressed having immediate responses based on key words – for instance any references of `space' implied `sci-fi'. We call this the `vampires in space' problem, because no matter how well a gothic sentence is rewritten as a sci-fi one, it's impossible to ignore the fact that there is a vampire in space. The transferred sentences, in the eyes of the Ablated NVA classifier (with no access to content words), did quite well transferring into their intended style. But people are not blind to content. Human Evaluation ::: The `Vampires in Space' Problem Working with the annotators, we regularly came up against the 'vampires in space' problem: while syntactic constructions account for much of the distinction of literary styles, these constructions often co-occur with distinctive content. Stylometrics finds syntactic constructions are great at fingerprinting, but suggests that these constructions are surface realizations of higher-level stylistic decisions. The number and type of personal pronouns is a reflection of how characters feature in a text. A large number of positional prepositions may be the result of a writer focusing on physical descriptions of scenes. In our attempt to decouple these, we create Frankenstein sentences, which piece together features of different styles – we are putting vampires in space. Another way to validate our approach would be to select data that is stylistically distinctive but with similar content: perhaps genres in which content is static but language use changes over time, stylistically distinct authors within a single genre, or parodies of a distinctive genre. Conclusion and Future Work We present a formal, extendable model of style that can add control to any neural text generation system. We model style as a suite of low-level linguistic controls, and train a neural encoder-decoder model to reconstruct reference sentences given only content words and the setting of the controls. In automatic evaluations, we show that our model can fool a style classifier 84% of the time and outperforms a baseline genre-embedding model. In human evaluations, we encounter the `vampires in space' problem in which content and style are equally discriminative but people focus more on the content. In future work we would like to model higher-level syntactic controls. BIBREF20 show that differences in clausal constructions, for instance having a dependent clause before an independent clause or vice versa, is a marker of style appreciated by the reader. Such features would likely interact with our lower-level controls in an interesting way, and provide further insight into style transfer in text. Acknowledgements Katy Gero is supported by an NSF GRF (DGE - 1644869). We would also like to thank Elsbeth Turcan for her helpful comments.

Yes

5f4e6ce4a811c4b3ab07335d89db2fd2a8d8d8b2

5f4e6ce4a811c4b3ab07335d89db2fd2a8d8d8b2_0

Q: How they perform manual evaluation, what is criteria? Text: Introduction All text has style, whether it be formal or informal, polite or aggressive, colloquial, persuasive, or even robotic. Despite the success of style transfer in image processing BIBREF0, BIBREF1, there has been limited progress in the text domain, where disentangling style from content is particularly difficult. To date, most work in style transfer relies on the availability of meta-data, such as sentiment, authorship, or formality. While meta-data can provide insight into the style of a text, it often conflates style with content, limiting the ability to perform style transfer while preserving content. Generalizing style transfer requires separating style from the meaning of the text itself. The study of literary style can guide us. For example, in the digital humanities and its subfield of stylometry, content doesn't figure prominently in practical methods of discriminating authorship and genres, which can be thought of as style at the level of the individual and population, respectively. Rather, syntactic and functional constructions are the most salient features. In this work, we turn to literary style as a test-bed for style transfer, and build on work from literature scholars using computational techniques for analysis. In particular we draw on stylometry: the use of surface level features, often counts of function words, to discriminate between literary styles. Stylometry first saw success in attributing authorship to the disputed Federalist Papers BIBREF2, but is recently used by scholars to study things such as the birth of genres BIBREF3 and the change of author styles over time BIBREF4. The use of function words is likely not the way writers intend to express style, but they appear to be downstream realizations of higher-level stylistic decisions. We hypothesize that surface-level linguistic features, such as counts of personal pronouns, prepositions, and punctuation, are an excellent definition of literary style, as borne out by their use in the digital humanities, and our own style classification experiments. We propose a controllable neural encoder-decoder model in which these features are modelled explicitly as decoder feature embeddings. In training, the model learns to reconstruct a text using only the content words and the linguistic feature embeddings. We can then transfer arbitrary content words to a new style without parallel data by setting the low-level style feature embeddings to be indicative of the target style. This paper makes the following contributions: A formal model of style as a suite of controllable, low-level linguistic features that are independent of content. An automatic evaluation showing that our model fools a style classifier 84% of the time. A human evaluation with English literature experts, including recommendations for dealing with the entanglement of content with style. Related Work ::: Style Transfer with Parallel Data Following in the footsteps of machine translation, style transfer in text has seen success by using parallel data. BIBREF5 use modern translations of Shakespeare plays to build a modern-to-Shakespearan model. BIBREF6 compile parallel data for formal and informal sentences, allowing them to successfully use various machine translation techniques. While parallel data may work for very specific styles, the difficulty of finding parallel texts dramatically limits this approach. Related Work ::: Style Transfer without Parallel Data There has been a decent amount of work on this approach in the past few years BIBREF7, BIBREF8, mostly focusing on variations of an encoder-decoder framework in which style is modeled as a monolithic style embedding. The main obstacle is often to disentangle style and content. However, it remains a challenging problem. Perhaps the most successful is BIBREF9, who use a de-noising auto encoder and back translation to learn style without parallel data. BIBREF10 outline the benefits of automatically extracting style, and suggest there is a formal weakness of using linguistic heuristics. In contrast, we believe that monolithic style embeddings don't capture the existing knowledge we have about style, and will struggle to disentangle content. Related Work ::: Controlling Linguistic Features Several papers have worked on controlling style when generating sentences from restaurant meaning representations BIBREF11, BIBREF12. In each of these cases, the diversity in outputs is quite small given the constraints of the meaning representation, style is often constrained to interjections (like “yeah”), and there is no original style from which to transfer. BIBREF13 investigate using stylistic parameters and content parameters to control text generation using a movie review dataset. Their stylistic parameters are created using word-level heuristics and they are successful in controlling these parameters in the outputs. Their success bodes well for our related approach in a style transfer setting, in which the content (not merely content parameters) is held fixed. Related Work ::: Stylometry and the Digital Humanities Style, in literary research, is anything but a stable concept, but it nonetheless has a long tradition of study in the digital humanities. In a remarkably early quantitative study of literature, BIBREF14 charts sentence-level stylistic attributes specific to a number of novelists. Half a century later, BIBREF15 builds on earlier work in information theory by BIBREF16, and defines a literary text as consisting of two “materials": “the vocabulary, and some structural properties, the style, of its author." Beginning with BIBREF2, statistical approaches to style, or stylometry, join the already-heated debates over the authorship of literary works. A noteable example of this is the “Delta" measure, which uses z-scores of function word frequencies BIBREF17. BIBREF18 find that Shakespeare added some material to a later edition of Thomas Kyd's The Spanish Tragedy, and that Christopher Marlowe collaborated with Shakespeare on Henry VI. Models ::: Preliminary Classification Experiments The stylometric research cited above suggests that the most frequently used words, e.g. function words, are most discriminating of authorship and literary style. We investigate these claims using three corpora that have distinctive styles in the literary community: gothic novels, philosophy books, and pulp science fiction, hereafter sci-fi. We retrieve gothic novels and philosophy books from Project Gutenberg and pulp sci-fi from Internet Archive's Pulp Magazine Archive. We partition this corpus into train, validation, and test sets the sizes of which can be found in Table TABREF12. In order to validate the above claims, we train five different classifiers to predict the literary style of sentences from our corpus. Each classifier has gradually more content words replaced with part-of-speech (POS) tag placeholder tokens. The All model is trained on sentences with all proper nouns replaced by `PROPN'. The models Ablated N, Ablated NV, and Ablated NVA replace nouns, nouns & verbs, and nouns, verbs, & adjectives with the corresponding POS tag respectively. Finally, Content-only is trained on sentences with all words that are not tagged as NOUN, VERB, ADJ removed; the remaining words are not ablated. We train the classifiers on the training set, balancing the class distribution to make sure there are the same number of sentences from each style. Classifiers are trained using fastText BIBREF19, using tri-gram features with all other settings as default. table:classifiers shows the accuracies of the classifiers. The styles are highly distinctive: the All classifier has an accuracy of 86%. Additionally, even the Ablated NVA is quite successful, with 75% accuracy, even without access to any content words. The Content only classifier is also quite successful, at 80% accuracy. This indicates that these stylistic genres are distinctive at both the content level and at the syntactic level. Models ::: Formal Model of Style Given that non-content words are distinctive enough for a classifier to determine style, we propose a suite of low-level linguistic feature counts (henceforth, controls) as our formal, content-blind definition of style. The style of a sentence is represented as a vector of counts of closed word classes (like personal pronouns) as well as counts of syntactic features like the number of SBAR non-terminals in its constituency parse, since clause structure has been shown to be indicative of style BIBREF20. Controls are extracted heuristically, and almost all rely on counts of pre-defined word lists. For constituency parses we use the Stanford Parser BIBREF21. table:controlexamples lists all the controls along with examples. Models ::: Formal Model of Style ::: Reconstruction Task Models are trained with a reconstruction task, in which a distorted version of a reference sentence is input and the goal is to output the original reference. fig:sentenceinput illustrates the process. Controls are calculated heuristically. All words found in the control word lists are then removed from the reference sentence. The remaining words, which represent the content, are used as input into the model, along with their POS tags and lemmas. In this way we encourage models to construct a sentence using content and style independently. This will allow us to vary the stylistic controls while keeping the content constant, and successfully perform style transfer. When generating a new sentence, the controls correspond to the counts of the corresponding syntactic features that we expect to be realized in the output. Models ::: Neural Architecture We implement our feature controlled language model using a neural encoder-decoder with attention BIBREF22, using 2-layer uni-directional gated recurrent units (GRUs) for the encoder and decoder BIBREF23. The input to the encoder is a sequence of $M$ content words, along with their lemmas, and fine and coarse grained part-of-speech (POS) tags, i.e. $X_{.,j} = (x_{1,j},\ldots ,x_{M,j})$ for $j \in \mathcal {T} = \lbrace \textrm {word, lemma, fine-pos, coarse-pos}\rbrace $. We embed each token (and its lemma and POS) before concatenating, and feeding into the encoder GRU to obtain encoder hidden states, $ c_i = \operatorname{gru}(c_{i-1}, \left[E_j(X_{i,j}), \; j\in \mathcal {T} \right]; \omega _{enc}) $ for $i \in {1,\ldots ,M},$ where initial state $c_0$, encoder GRU parameters $\omega _{enc}$ and embedding matrices $E_j$ are learned parameters. The decoder sequentially generates the outputs, i.e. a sequence of $N$ tokens $y =(y_1,\ldots ,y_N)$, where all tokens $y_i$ are drawn from a finite output vocabulary $\mathcal {V}$. To generate the each token we first embed the previously generated token $y_{i-1}$ and a vector of $K$ control features $z = ( z_1,\ldots , z_K)$ (using embedding matrices $E_{dec}$ and $E_{\textrm {ctrl-1}}, \ldots , E_{\textrm {ctrl-K}}$ respectively), before concatenating them into a vector $\rho _i,$ and feeding them into the decoder side GRU along with the previous decoder state $h_{i-1}$: where $\omega _{dec}$ are the decoder side GRU parameters. Using the decoder hidden state $h_i$ we then attend to the encoder context vectors $c_j$, computing attention scores $\alpha _{i,j}$, where before passing $h_i$ and the attention weighted context $\bar{c}_i=\sum _{j=1}^M \alpha _{i,j} c_j$ into a single hidden-layer perceptron with softmax output to compute the next token prediction probability, where $W,U,V$ and $u,v, \nu $ are parameter matrices and vectors respectively. Crucially, the controls $z$ remain fixed for all input decoder steps. Each $z_k$ represents the frequency of one of the low-level features described in sec:formalstyle. During training on the reconstruction task, we can observe the full output sequence $y,$ and so we can obtain counts for each control feature directly. Controls receive a different embedding depending on their frequency, where counts of 0-20 each get a unique embedding, and counts greater than 20 are assigned to the same embedding. At test time, we set the values of the controls according to procedure described in Section SECREF25. We use embedding sizes of 128, 128, 64, and 32 for token, lemma, fine, and coarse grained POS embedding matrices respectively. Output token embeddings $E_{dec}$ have size 512, and 50 for the control feature embeddings. We set 512 for all GRU and perceptron output sizes. We refer to this model as the StyleEQ model. See fig:model for a visual depiction of the model. Models ::: Neural Architecture ::: Baseline Genre Model We compare the above model to a similar model, where rather than explicitly represent $K$ features as input, we have $K$ features in the form of a genre embedding, i.e. we learn a genre specific embedding for each of the gothic, scifi, and philosophy genres, as studied in BIBREF8 and BIBREF7. To generate in a specific style, we simply set the appropriate embedding. We use genre embeddings of size 850 which is equivalent to the total size of the $K$ feature embeddings in the StyleEQ model. Models ::: Neural Architecture ::: Training We train both models with minibatch stochastic gradient descent with a learning rate of 0.25, weight decay penalty of 0.0001, and batch size of 64. We also apply dropout with a drop rate of 0.25 to all embedding layers, the GRUs, and preceptron hidden layer. We train for a maximum of 200 epochs, using validation set BLEU score BIBREF26 to select the final model iteration for evaluation. Models ::: Neural Architecture ::: Selecting Controls for Style Transfer In the Baseline model, style transfer is straightforward: given an input sentence in one style, fix the encoder content features while selecting a different genre embedding. In contrast, the StyleEQ model requires selecting the counts for each control. Although there are a variety of ways to do this, we use a method that encourages a diversity of outputs. In order to ensure the controls match the reference sentence in magnitude, we first find all sentences in the target style with the same number of words as the reference sentence. Then, we add the following constraints: the same number of proper nouns, the same number of nouns, the same number of verbs, and the same number of adjectives. We randomly sample $n$ of the remaining sentences, and for each of these `sibling' sentences, we compute the controls. For each of the new controls, we generate a sentence using the original input sentence content features. The generated sentences are then reranked using the length normalized log-likelihood under the model. We can then select the highest scoring sentence as our style-transferred output, or take the top-$k$ when we need a diverse set of outputs. The reason for this process is that although there are group-level distinctive controls for each style, e.g. the high use of punctuation in philosophy books or of first person pronouns in gothic novels, at the sentence level it can understandably be quite varied. This method matches sentences between styles, capturing the natural distribution of the corpora. Automatic Evaluations ::: BLEU Scores & Perplexity In tab:blueperpl we report BLEU scores for the reconstruction of test set sentences from their content and feature representations, as well as the model perplexities of the reconstruction. For both models, we use beam decoding with a beam size of eight. Beam candidates are ranked according to their length normalized log-likelihood. On these automatic measures we see that StyleEQ is better able to reconstruct the original sentences. In some sense this evaluation is mostly a sanity check, as the feature controls contain more locally specific information than the genre embeddings, which say very little about how many specific function words one should expect to see in the output. Automatic Evaluations ::: Feature Control Designing controllable language models is often difficult because of the various dependencies between tokens; when changing one control value it may effect other aspects of the surface realization. For example, increasing the number of conjunctions may effect how the generator places prepositions to compensate for structural changes in the sentence. Since our features are deterministically recoverable, we can perturb an individual control value and check to see that the desired change was realized in the output. Moreover, we can check the amount of change in the other non-perturbed features to measure the independence of the controls. We sample 50 sentences from each genre from the test set. For each sample, we create a perturbed control setting for each control by adding $\delta $ to the original control value. This is done for $\delta \in \lbrace -3, -2, -1, 0, 1, 2, 3\rbrace $, skipping any settings where the new control value would be negative. table:autoeval:ctrl shows the results of this experiment. The Exact column displays the percentage of generated texts that realize the exact number of control features specified by the perturbed control. High percentages in the Exact column indicate greater one-to-one correspondence between the control and surface realization. For example, if the input was “Dracula and Frankenstein and the mummy,” and we change the conjunction feature by $\delta =-1$, an output of “Dracula, Frankenstein and the mummy,” would count towards the Exact category, while “Dracula, Frankenstein, the mummy,” would not. The Direction column specifies the percentage of cases where the generated text produces a changed number of the control features that, while not exactly matching the specified value of the perturbed control, does change from the original in the correct direction. For example, if the input again was “Dracula and Frankenstein and the mummy,” and we change the conjunction feature by $\delta =-1$, both outputs of “Dracula, Frankenstein and the mummy,” and “Dracula, Frankenstein, the mummy,” would count towards Direction. High percentages in Direction mean that we could roughly ensure desired surface realizations by modifying the control by a larger $\delta $. Finally, the Atomic column specifies the percentage of cases where the generated text with the perturbed control only realizes changes to that specific control, while other features remain constant. For example, if the input was “Dracula and Frankenstein in the castle,” and we set the conjunction feature to $\delta =-1$, an output of “Dracula near Frankenstein in the castle,” would not count as Atomic because, while the number of conjunctions did decrease by one, the number of simple preposition changed. An output of “Dracula, Frankenstein in the castle,” would count as Atomic. High percentages in the Atomic column indicate this feature is only loosely coupled to the other features and can be changed without modifying other aspects of the sentence. Controls such as conjunction, determiner, and punctuation are highly controllable, with Exact rates above 80%. But with the exception of the constituency parse features, all controls have high Direction rates, many in the 90s. These results indicate our model successfully controls these features. The fact that the Atomic rates are relatively low is to be expected, as controls are highly coupled – e.g. to increase 1stPer, it is likely another pronoun control will have to decrease. Automatic Evaluations ::: Automatic Classification For each model we look at the classifier prediction accuracy of reconstructed and transferred sentences. In particular we use the Ablated NVA classifier, as this is the most content-blind one. We produce 16 outputs from both the Baseline and StyleEq models. For the Baseline, we use a beam search of size 16. For the StyleEQ model, we use the method described in Section SECREF25 to select 16 `sibling' sentences in the target style, and generated a transferred sentence for each. We look at three different methods for selection: all, which uses all output sentences; top, which selects the top ranked sentence based on the score from the model; and oracle, which selects the sentence with the highest classifier likelihood for the intended style. The reason for the third method, which indeed acts as an oracle, is that using the score from the model didn't always surface a transferred sentence that best reflected the desired style. Partially this was because the model score was mostly a function of how well a transferred sentence reflected the distribution of the training data. But additionally, some control settings are more indicative of a target style than others. The use of the classifier allows us to identify the most suitable control setting for a target style that was roughly compatible with the number of content words. In table:fasttext-results we see the results. Note that for both models, the all and top classification accuracy tends to be quite similar, though for the Baseline they are often almost exactly the same when the Baseline has little to no diversity in the outputs. However, the oracle introduces a huge jump in accuracy for the StyleEQ model, especially compared to the Baseline, partially because the diversity of outputs from StyleEQ is much higher; often the Baseline model produces no diversity – the 16 output sentences may be nearly identical, save a single word or two. It's important to note that neither model uses the classifier in any way except to select the sentence from 16 candidate outputs. What this implies is that lurking within the StyleEQ model outputs are great sentences, even if they are hard to find. In many cases, the StyleEQ model has a classification accuracy above the base rate from the test data, which is 75% (see table:classifiers). Human Evaluation table:cherrypicking shows example outputs for the StyleEQ and Baseline models. Through inspection we see that the StyleEQ model successfully changes syntactic constructions in stylistically distinctive ways, such as increasing syntactic complexity when transferring to philosophy, or changing relevant pronouns when transferring to sci-fi. In contrast, the Baseline model doesn't create outputs that move far from the reference sentence, making only minor modifications such changing the type of a single pronoun. To determine how readers would classify our transferred sentences, we recruited three English Literature PhD candidates, all of whom had passed qualifying exams that included determining both genre and era of various literary texts. Human Evaluation ::: Fluency Evaluation To evaluate the fluency of our outputs, we had the annotators score reference sentences, reconstructed sentences, and transferred sentences on a 0-5 scale, where 0 was incoherent and 5 was a well-written human sentence. table:fluency shows the average fluency of various conditions from all three annotators. Both models have fluency scores around 3. Upon inspection of the outputs, it is clear that many have fluency errors, resulting in ungrammatical sentences. Notably the Baseline often has slightly higher fluency scores than the StyleEQ model. This is likely because the Baseline model is far less constrained in how to construct the output sentence, and upon inspection often reconstructs the reference sentence even when performing style transfer. In contrast, the StyleEQ is encouraged to follow the controls, but can struggle to incorporate these controls into a fluent sentence. The fluency of all outputs is lower than desired. We expect that incorporating pre-trained language models would increase the fluency of all outputs without requiring larger datasets. Human Evaluation ::: Human Classification Each annotator annotated 90 reference sentences (i.e. from the training corpus) with which style they thought the sentence was from. The accuracy on this baseline task for annotators A1, A2, and A3 was 80%, 88%, and 80% respectively, giving us an upper expected bound on the human evaluation. In discussing this task with the annotators, they noted that content is a heavy predictor of genre, and that would certainly confound their annotations. To attempt to mitigate this, we gave them two annotation tasks: which-of-3 where they simply marked which style they thought a sentence was from, and which-of-2 where they were given the original style and marked which style they thought the sentence was transferred into. For each task, each annotator marked 180 sentences: 90 from each model, with an even split across the three genres. Annotators were presented the sentences in a random order, without information about the models. In total, each marked 270 sentences. (Note there were no reconstructions in this annotation task.) table:humanclassifiers shows the results. In both tasks, accuracy of annotators classifying the sentence as its intended style was low. In which-of-3, scores were around 20%, below the chance rate of 33%. In which-of-2, scores were in the 50s, slightly above the chance rate of 50%. This was the case for both models. There was a slight increase in accuracy for the StyleEQ model over the Baseline for which-of-3, but the opposite trend for which-of-2, suggesting these differences are not significant. It's clear that it's hard to fool the annotators. Introspecting on their approach, the annotators expressed having immediate responses based on key words – for instance any references of `space' implied `sci-fi'. We call this the `vampires in space' problem, because no matter how well a gothic sentence is rewritten as a sci-fi one, it's impossible to ignore the fact that there is a vampire in space. The transferred sentences, in the eyes of the Ablated NVA classifier (with no access to content words), did quite well transferring into their intended style. But people are not blind to content. Human Evaluation ::: The `Vampires in Space' Problem Working with the annotators, we regularly came up against the 'vampires in space' problem: while syntactic constructions account for much of the distinction of literary styles, these constructions often co-occur with distinctive content. Stylometrics finds syntactic constructions are great at fingerprinting, but suggests that these constructions are surface realizations of higher-level stylistic decisions. The number and type of personal pronouns is a reflection of how characters feature in a text. A large number of positional prepositions may be the result of a writer focusing on physical descriptions of scenes. In our attempt to decouple these, we create Frankenstein sentences, which piece together features of different styles – we are putting vampires in space. Another way to validate our approach would be to select data that is stylistically distinctive but with similar content: perhaps genres in which content is static but language use changes over time, stylistically distinct authors within a single genre, or parodies of a distinctive genre. Conclusion and Future Work We present a formal, extendable model of style that can add control to any neural text generation system. We model style as a suite of low-level linguistic controls, and train a neural encoder-decoder model to reconstruct reference sentences given only content words and the setting of the controls. In automatic evaluations, we show that our model can fool a style classifier 84% of the time and outperforms a baseline genre-embedding model. In human evaluations, we encounter the `vampires in space' problem in which content and style are equally discriminative but people focus more on the content. In future work we would like to model higher-level syntactic controls. BIBREF20 show that differences in clausal constructions, for instance having a dependent clause before an independent clause or vice versa, is a marker of style appreciated by the reader. Such features would likely interact with our lower-level controls in an interesting way, and provide further insight into style transfer in text. Acknowledgements Katy Gero is supported by an NSF GRF (DGE - 1644869). We would also like to thank Elsbeth Turcan for her helpful comments.

accuracy

a234bcbf2e41429422adda37d9e926b49ef66150

a234bcbf2e41429422adda37d9e926b49ef66150_0

Q: What metrics are used for automatic evaluation? Text: Introduction All text has style, whether it be formal or informal, polite or aggressive, colloquial, persuasive, or even robotic. Despite the success of style transfer in image processing BIBREF0, BIBREF1, there has been limited progress in the text domain, where disentangling style from content is particularly difficult. To date, most work in style transfer relies on the availability of meta-data, such as sentiment, authorship, or formality. While meta-data can provide insight into the style of a text, it often conflates style with content, limiting the ability to perform style transfer while preserving content. Generalizing style transfer requires separating style from the meaning of the text itself. The study of literary style can guide us. For example, in the digital humanities and its subfield of stylometry, content doesn't figure prominently in practical methods of discriminating authorship and genres, which can be thought of as style at the level of the individual and population, respectively. Rather, syntactic and functional constructions are the most salient features. In this work, we turn to literary style as a test-bed for style transfer, and build on work from literature scholars using computational techniques for analysis. In particular we draw on stylometry: the use of surface level features, often counts of function words, to discriminate between literary styles. Stylometry first saw success in attributing authorship to the disputed Federalist Papers BIBREF2, but is recently used by scholars to study things such as the birth of genres BIBREF3 and the change of author styles over time BIBREF4. The use of function words is likely not the way writers intend to express style, but they appear to be downstream realizations of higher-level stylistic decisions. We hypothesize that surface-level linguistic features, such as counts of personal pronouns, prepositions, and punctuation, are an excellent definition of literary style, as borne out by their use in the digital humanities, and our own style classification experiments. We propose a controllable neural encoder-decoder model in which these features are modelled explicitly as decoder feature embeddings. In training, the model learns to reconstruct a text using only the content words and the linguistic feature embeddings. We can then transfer arbitrary content words to a new style without parallel data by setting the low-level style feature embeddings to be indicative of the target style. This paper makes the following contributions: A formal model of style as a suite of controllable, low-level linguistic features that are independent of content. An automatic evaluation showing that our model fools a style classifier 84% of the time. A human evaluation with English literature experts, including recommendations for dealing with the entanglement of content with style. Related Work ::: Style Transfer with Parallel Data Following in the footsteps of machine translation, style transfer in text has seen success by using parallel data. BIBREF5 use modern translations of Shakespeare plays to build a modern-to-Shakespearan model. BIBREF6 compile parallel data for formal and informal sentences, allowing them to successfully use various machine translation techniques. While parallel data may work for very specific styles, the difficulty of finding parallel texts dramatically limits this approach. Related Work ::: Style Transfer without Parallel Data There has been a decent amount of work on this approach in the past few years BIBREF7, BIBREF8, mostly focusing on variations of an encoder-decoder framework in which style is modeled as a monolithic style embedding. The main obstacle is often to disentangle style and content. However, it remains a challenging problem. Perhaps the most successful is BIBREF9, who use a de-noising auto encoder and back translation to learn style without parallel data. BIBREF10 outline the benefits of automatically extracting style, and suggest there is a formal weakness of using linguistic heuristics. In contrast, we believe that monolithic style embeddings don't capture the existing knowledge we have about style, and will struggle to disentangle content. Related Work ::: Controlling Linguistic Features Several papers have worked on controlling style when generating sentences from restaurant meaning representations BIBREF11, BIBREF12. In each of these cases, the diversity in outputs is quite small given the constraints of the meaning representation, style is often constrained to interjections (like “yeah”), and there is no original style from which to transfer. BIBREF13 investigate using stylistic parameters and content parameters to control text generation using a movie review dataset. Their stylistic parameters are created using word-level heuristics and they are successful in controlling these parameters in the outputs. Their success bodes well for our related approach in a style transfer setting, in which the content (not merely content parameters) is held fixed. Related Work ::: Stylometry and the Digital Humanities Style, in literary research, is anything but a stable concept, but it nonetheless has a long tradition of study in the digital humanities. In a remarkably early quantitative study of literature, BIBREF14 charts sentence-level stylistic attributes specific to a number of novelists. Half a century later, BIBREF15 builds on earlier work in information theory by BIBREF16, and defines a literary text as consisting of two “materials": “the vocabulary, and some structural properties, the style, of its author." Beginning with BIBREF2, statistical approaches to style, or stylometry, join the already-heated debates over the authorship of literary works. A noteable example of this is the “Delta" measure, which uses z-scores of function word frequencies BIBREF17. BIBREF18 find that Shakespeare added some material to a later edition of Thomas Kyd's The Spanish Tragedy, and that Christopher Marlowe collaborated with Shakespeare on Henry VI. Models ::: Preliminary Classification Experiments The stylometric research cited above suggests that the most frequently used words, e.g. function words, are most discriminating of authorship and literary style. We investigate these claims using three corpora that have distinctive styles in the literary community: gothic novels, philosophy books, and pulp science fiction, hereafter sci-fi. We retrieve gothic novels and philosophy books from Project Gutenberg and pulp sci-fi from Internet Archive's Pulp Magazine Archive. We partition this corpus into train, validation, and test sets the sizes of which can be found in Table TABREF12. In order to validate the above claims, we train five different classifiers to predict the literary style of sentences from our corpus. Each classifier has gradually more content words replaced with part-of-speech (POS) tag placeholder tokens. The All model is trained on sentences with all proper nouns replaced by `PROPN'. The models Ablated N, Ablated NV, and Ablated NVA replace nouns, nouns & verbs, and nouns, verbs, & adjectives with the corresponding POS tag respectively. Finally, Content-only is trained on sentences with all words that are not tagged as NOUN, VERB, ADJ removed; the remaining words are not ablated. We train the classifiers on the training set, balancing the class distribution to make sure there are the same number of sentences from each style. Classifiers are trained using fastText BIBREF19, using tri-gram features with all other settings as default. table:classifiers shows the accuracies of the classifiers. The styles are highly distinctive: the All classifier has an accuracy of 86%. Additionally, even the Ablated NVA is quite successful, with 75% accuracy, even without access to any content words. The Content only classifier is also quite successful, at 80% accuracy. This indicates that these stylistic genres are distinctive at both the content level and at the syntactic level. Models ::: Formal Model of Style Given that non-content words are distinctive enough for a classifier to determine style, we propose a suite of low-level linguistic feature counts (henceforth, controls) as our formal, content-blind definition of style. The style of a sentence is represented as a vector of counts of closed word classes (like personal pronouns) as well as counts of syntactic features like the number of SBAR non-terminals in its constituency parse, since clause structure has been shown to be indicative of style BIBREF20. Controls are extracted heuristically, and almost all rely on counts of pre-defined word lists. For constituency parses we use the Stanford Parser BIBREF21. table:controlexamples lists all the controls along with examples. Models ::: Formal Model of Style ::: Reconstruction Task Models are trained with a reconstruction task, in which a distorted version of a reference sentence is input and the goal is to output the original reference. fig:sentenceinput illustrates the process. Controls are calculated heuristically. All words found in the control word lists are then removed from the reference sentence. The remaining words, which represent the content, are used as input into the model, along with their POS tags and lemmas. In this way we encourage models to construct a sentence using content and style independently. This will allow us to vary the stylistic controls while keeping the content constant, and successfully perform style transfer. When generating a new sentence, the controls correspond to the counts of the corresponding syntactic features that we expect to be realized in the output. Models ::: Neural Architecture We implement our feature controlled language model using a neural encoder-decoder with attention BIBREF22, using 2-layer uni-directional gated recurrent units (GRUs) for the encoder and decoder BIBREF23. The input to the encoder is a sequence of $M$ content words, along with their lemmas, and fine and coarse grained part-of-speech (POS) tags, i.e. $X_{.,j} = (x_{1,j},\ldots ,x_{M,j})$ for $j \in \mathcal {T} = \lbrace \textrm {word, lemma, fine-pos, coarse-pos}\rbrace $. We embed each token (and its lemma and POS) before concatenating, and feeding into the encoder GRU to obtain encoder hidden states, $ c_i = \operatorname{gru}(c_{i-1}, \left[E_j(X_{i,j}), \; j\in \mathcal {T} \right]; \omega _{enc}) $ for $i \in {1,\ldots ,M},$ where initial state $c_0$, encoder GRU parameters $\omega _{enc}$ and embedding matrices $E_j$ are learned parameters. The decoder sequentially generates the outputs, i.e. a sequence of $N$ tokens $y =(y_1,\ldots ,y_N)$, where all tokens $y_i$ are drawn from a finite output vocabulary $\mathcal {V}$. To generate the each token we first embed the previously generated token $y_{i-1}$ and a vector of $K$ control features $z = ( z_1,\ldots , z_K)$ (using embedding matrices $E_{dec}$ and $E_{\textrm {ctrl-1}}, \ldots , E_{\textrm {ctrl-K}}$ respectively), before concatenating them into a vector $\rho _i,$ and feeding them into the decoder side GRU along with the previous decoder state $h_{i-1}$: where $\omega _{dec}$ are the decoder side GRU parameters. Using the decoder hidden state $h_i$ we then attend to the encoder context vectors $c_j$, computing attention scores $\alpha _{i,j}$, where before passing $h_i$ and the attention weighted context $\bar{c}_i=\sum _{j=1}^M \alpha _{i,j} c_j$ into a single hidden-layer perceptron with softmax output to compute the next token prediction probability, where $W,U,V$ and $u,v, \nu $ are parameter matrices and vectors respectively. Crucially, the controls $z$ remain fixed for all input decoder steps. Each $z_k$ represents the frequency of one of the low-level features described in sec:formalstyle. During training on the reconstruction task, we can observe the full output sequence $y,$ and so we can obtain counts for each control feature directly. Controls receive a different embedding depending on their frequency, where counts of 0-20 each get a unique embedding, and counts greater than 20 are assigned to the same embedding. At test time, we set the values of the controls according to procedure described in Section SECREF25. We use embedding sizes of 128, 128, 64, and 32 for token, lemma, fine, and coarse grained POS embedding matrices respectively. Output token embeddings $E_{dec}$ have size 512, and 50 for the control feature embeddings. We set 512 for all GRU and perceptron output sizes. We refer to this model as the StyleEQ model. See fig:model for a visual depiction of the model. Models ::: Neural Architecture ::: Baseline Genre Model We compare the above model to a similar model, where rather than explicitly represent $K$ features as input, we have $K$ features in the form of a genre embedding, i.e. we learn a genre specific embedding for each of the gothic, scifi, and philosophy genres, as studied in BIBREF8 and BIBREF7. To generate in a specific style, we simply set the appropriate embedding. We use genre embeddings of size 850 which is equivalent to the total size of the $K$ feature embeddings in the StyleEQ model. Models ::: Neural Architecture ::: Training We train both models with minibatch stochastic gradient descent with a learning rate of 0.25, weight decay penalty of 0.0001, and batch size of 64. We also apply dropout with a drop rate of 0.25 to all embedding layers, the GRUs, and preceptron hidden layer. We train for a maximum of 200 epochs, using validation set BLEU score BIBREF26 to select the final model iteration for evaluation. Models ::: Neural Architecture ::: Selecting Controls for Style Transfer In the Baseline model, style transfer is straightforward: given an input sentence in one style, fix the encoder content features while selecting a different genre embedding. In contrast, the StyleEQ model requires selecting the counts for each control. Although there are a variety of ways to do this, we use a method that encourages a diversity of outputs. In order to ensure the controls match the reference sentence in magnitude, we first find all sentences in the target style with the same number of words as the reference sentence. Then, we add the following constraints: the same number of proper nouns, the same number of nouns, the same number of verbs, and the same number of adjectives. We randomly sample $n$ of the remaining sentences, and for each of these `sibling' sentences, we compute the controls. For each of the new controls, we generate a sentence using the original input sentence content features. The generated sentences are then reranked using the length normalized log-likelihood under the model. We can then select the highest scoring sentence as our style-transferred output, or take the top-$k$ when we need a diverse set of outputs. The reason for this process is that although there are group-level distinctive controls for each style, e.g. the high use of punctuation in philosophy books or of first person pronouns in gothic novels, at the sentence level it can understandably be quite varied. This method matches sentences between styles, capturing the natural distribution of the corpora. Automatic Evaluations ::: BLEU Scores & Perplexity In tab:blueperpl we report BLEU scores for the reconstruction of test set sentences from their content and feature representations, as well as the model perplexities of the reconstruction. For both models, we use beam decoding with a beam size of eight. Beam candidates are ranked according to their length normalized log-likelihood. On these automatic measures we see that StyleEQ is better able to reconstruct the original sentences. In some sense this evaluation is mostly a sanity check, as the feature controls contain more locally specific information than the genre embeddings, which say very little about how many specific function words one should expect to see in the output. Automatic Evaluations ::: Feature Control Designing controllable language models is often difficult because of the various dependencies between tokens; when changing one control value it may effect other aspects of the surface realization. For example, increasing the number of conjunctions may effect how the generator places prepositions to compensate for structural changes in the sentence. Since our features are deterministically recoverable, we can perturb an individual control value and check to see that the desired change was realized in the output. Moreover, we can check the amount of change in the other non-perturbed features to measure the independence of the controls. We sample 50 sentences from each genre from the test set. For each sample, we create a perturbed control setting for each control by adding $\delta $ to the original control value. This is done for $\delta \in \lbrace -3, -2, -1, 0, 1, 2, 3\rbrace $, skipping any settings where the new control value would be negative. table:autoeval:ctrl shows the results of this experiment. The Exact column displays the percentage of generated texts that realize the exact number of control features specified by the perturbed control. High percentages in the Exact column indicate greater one-to-one correspondence between the control and surface realization. For example, if the input was “Dracula and Frankenstein and the mummy,” and we change the conjunction feature by $\delta =-1$, an output of “Dracula, Frankenstein and the mummy,” would count towards the Exact category, while “Dracula, Frankenstein, the mummy,” would not. The Direction column specifies the percentage of cases where the generated text produces a changed number of the control features that, while not exactly matching the specified value of the perturbed control, does change from the original in the correct direction. For example, if the input again was “Dracula and Frankenstein and the mummy,” and we change the conjunction feature by $\delta =-1$, both outputs of “Dracula, Frankenstein and the mummy,” and “Dracula, Frankenstein, the mummy,” would count towards Direction. High percentages in Direction mean that we could roughly ensure desired surface realizations by modifying the control by a larger $\delta $. Finally, the Atomic column specifies the percentage of cases where the generated text with the perturbed control only realizes changes to that specific control, while other features remain constant. For example, if the input was “Dracula and Frankenstein in the castle,” and we set the conjunction feature to $\delta =-1$, an output of “Dracula near Frankenstein in the castle,” would not count as Atomic because, while the number of conjunctions did decrease by one, the number of simple preposition changed. An output of “Dracula, Frankenstein in the castle,” would count as Atomic. High percentages in the Atomic column indicate this feature is only loosely coupled to the other features and can be changed without modifying other aspects of the sentence. Controls such as conjunction, determiner, and punctuation are highly controllable, with Exact rates above 80%. But with the exception of the constituency parse features, all controls have high Direction rates, many in the 90s. These results indicate our model successfully controls these features. The fact that the Atomic rates are relatively low is to be expected, as controls are highly coupled – e.g. to increase 1stPer, it is likely another pronoun control will have to decrease. Automatic Evaluations ::: Automatic Classification For each model we look at the classifier prediction accuracy of reconstructed and transferred sentences. In particular we use the Ablated NVA classifier, as this is the most content-blind one. We produce 16 outputs from both the Baseline and StyleEq models. For the Baseline, we use a beam search of size 16. For the StyleEQ model, we use the method described in Section SECREF25 to select 16 `sibling' sentences in the target style, and generated a transferred sentence for each. We look at three different methods for selection: all, which uses all output sentences; top, which selects the top ranked sentence based on the score from the model; and oracle, which selects the sentence with the highest classifier likelihood for the intended style. The reason for the third method, which indeed acts as an oracle, is that using the score from the model didn't always surface a transferred sentence that best reflected the desired style. Partially this was because the model score was mostly a function of how well a transferred sentence reflected the distribution of the training data. But additionally, some control settings are more indicative of a target style than others. The use of the classifier allows us to identify the most suitable control setting for a target style that was roughly compatible with the number of content words. In table:fasttext-results we see the results. Note that for both models, the all and top classification accuracy tends to be quite similar, though for the Baseline they are often almost exactly the same when the Baseline has little to no diversity in the outputs. However, the oracle introduces a huge jump in accuracy for the StyleEQ model, especially compared to the Baseline, partially because the diversity of outputs from StyleEQ is much higher; often the Baseline model produces no diversity – the 16 output sentences may be nearly identical, save a single word or two. It's important to note that neither model uses the classifier in any way except to select the sentence from 16 candidate outputs. What this implies is that lurking within the StyleEQ model outputs are great sentences, even if they are hard to find. In many cases, the StyleEQ model has a classification accuracy above the base rate from the test data, which is 75% (see table:classifiers). Human Evaluation table:cherrypicking shows example outputs for the StyleEQ and Baseline models. Through inspection we see that the StyleEQ model successfully changes syntactic constructions in stylistically distinctive ways, such as increasing syntactic complexity when transferring to philosophy, or changing relevant pronouns when transferring to sci-fi. In contrast, the Baseline model doesn't create outputs that move far from the reference sentence, making only minor modifications such changing the type of a single pronoun. To determine how readers would classify our transferred sentences, we recruited three English Literature PhD candidates, all of whom had passed qualifying exams that included determining both genre and era of various literary texts. Human Evaluation ::: Fluency Evaluation To evaluate the fluency of our outputs, we had the annotators score reference sentences, reconstructed sentences, and transferred sentences on a 0-5 scale, where 0 was incoherent and 5 was a well-written human sentence. table:fluency shows the average fluency of various conditions from all three annotators. Both models have fluency scores around 3. Upon inspection of the outputs, it is clear that many have fluency errors, resulting in ungrammatical sentences. Notably the Baseline often has slightly higher fluency scores than the StyleEQ model. This is likely because the Baseline model is far less constrained in how to construct the output sentence, and upon inspection often reconstructs the reference sentence even when performing style transfer. In contrast, the StyleEQ is encouraged to follow the controls, but can struggle to incorporate these controls into a fluent sentence. The fluency of all outputs is lower than desired. We expect that incorporating pre-trained language models would increase the fluency of all outputs without requiring larger datasets. Human Evaluation ::: Human Classification Each annotator annotated 90 reference sentences (i.e. from the training corpus) with which style they thought the sentence was from. The accuracy on this baseline task for annotators A1, A2, and A3 was 80%, 88%, and 80% respectively, giving us an upper expected bound on the human evaluation. In discussing this task with the annotators, they noted that content is a heavy predictor of genre, and that would certainly confound their annotations. To attempt to mitigate this, we gave them two annotation tasks: which-of-3 where they simply marked which style they thought a sentence was from, and which-of-2 where they were given the original style and marked which style they thought the sentence was transferred into. For each task, each annotator marked 180 sentences: 90 from each model, with an even split across the three genres. Annotators were presented the sentences in a random order, without information about the models. In total, each marked 270 sentences. (Note there were no reconstructions in this annotation task.) table:humanclassifiers shows the results. In both tasks, accuracy of annotators classifying the sentence as its intended style was low. In which-of-3, scores were around 20%, below the chance rate of 33%. In which-of-2, scores were in the 50s, slightly above the chance rate of 50%. This was the case for both models. There was a slight increase in accuracy for the StyleEQ model over the Baseline for which-of-3, but the opposite trend for which-of-2, suggesting these differences are not significant. It's clear that it's hard to fool the annotators. Introspecting on their approach, the annotators expressed having immediate responses based on key words – for instance any references of `space' implied `sci-fi'. We call this the `vampires in space' problem, because no matter how well a gothic sentence is rewritten as a sci-fi one, it's impossible to ignore the fact that there is a vampire in space. The transferred sentences, in the eyes of the Ablated NVA classifier (with no access to content words), did quite well transferring into their intended style. But people are not blind to content. Human Evaluation ::: The `Vampires in Space' Problem Working with the annotators, we regularly came up against the 'vampires in space' problem: while syntactic constructions account for much of the distinction of literary styles, these constructions often co-occur with distinctive content. Stylometrics finds syntactic constructions are great at fingerprinting, but suggests that these constructions are surface realizations of higher-level stylistic decisions. The number and type of personal pronouns is a reflection of how characters feature in a text. A large number of positional prepositions may be the result of a writer focusing on physical descriptions of scenes. In our attempt to decouple these, we create Frankenstein sentences, which piece together features of different styles – we are putting vampires in space. Another way to validate our approach would be to select data that is stylistically distinctive but with similar content: perhaps genres in which content is static but language use changes over time, stylistically distinct authors within a single genre, or parodies of a distinctive genre. Conclusion and Future Work We present a formal, extendable model of style that can add control to any neural text generation system. We model style as a suite of low-level linguistic controls, and train a neural encoder-decoder model to reconstruct reference sentences given only content words and the setting of the controls. In automatic evaluations, we show that our model can fool a style classifier 84% of the time and outperforms a baseline genre-embedding model. In human evaluations, we encounter the `vampires in space' problem in which content and style are equally discriminative but people focus more on the content. In future work we would like to model higher-level syntactic controls. BIBREF20 show that differences in clausal constructions, for instance having a dependent clause before an independent clause or vice versa, is a marker of style appreciated by the reader. Such features would likely interact with our lower-level controls in an interesting way, and provide further insight into style transfer in text. Acknowledgements Katy Gero is supported by an NSF GRF (DGE - 1644869). We would also like to thank Elsbeth Turcan for her helpful comments.

classification accuracy, BLEU scores, model perplexities of the reconstruction