| { | |
| "paper_id": "W09-0205", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T06:36:33.866585Z" | |
| }, | |
| "title": "BagPack: A general framework to represent semantic relations", | |
| "authors": [ | |
| { | |
| "first": "Ama\u00e7", | |
| "middle": [], | |
| "last": "Herdagdelen", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Trento Rovereto", | |
| "location": { | |
| "country": "Italy" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Marco", | |
| "middle": [], | |
| "last": "Baroni", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Trento", | |
| "location": { | |
| "settlement": "Rovereto", | |
| "country": "Italy" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "We introduce a way to represent word pairs instantiating arbitrary semantic relations that keeps track of the contexts in which the words in the pair occur both together and independently. The resulting features are of sufficient generality to allow us, with the help of a standard supervised machine learning algorithm, to tackle a variety of unrelated semantic tasks with good results and almost no task-specific tailoring.", | |
| "pdf_parse": { | |
| "paper_id": "W09-0205", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "We introduce a way to represent word pairs instantiating arbitrary semantic relations that keeps track of the contexts in which the words in the pair occur both together and independently. The resulting features are of sufficient generality to allow us, with the help of a standard supervised machine learning algorithm, to tackle a variety of unrelated semantic tasks with good results and almost no task-specific tailoring.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Co-occurrence statistics extracted from corpora lead to good performance on a wide range of tasks that involve the identification of the semantic relation between two words or concepts (Sahlgren, 2006; Turney, 2006) . However, the difficulty of such tasks and the fact that they are apparently unrelated has led to the development of largely ad-hoc solutions, tuned to specific challenges. For many practical applications, this is a drawback: Given the large number of semantic relations that might be relevant to one or the other task, we need a multi-purpose approach that, given an appropriate representation and training examples instantiating an arbitrary target relation, can automatically mine new pairs characterized by the same relation. Building on a recent proposal in this direction by Turney (2008) , we propose a generic method of this sort, and we test it on a set of unrelated tasks, reporting good performance across the board with very little task-specific tweaking.", | |
| "cite_spans": [ | |
| { | |
| "start": 185, | |
| "end": 201, | |
| "text": "(Sahlgren, 2006;", | |
| "ref_id": "BIBREF16" | |
| }, | |
| { | |
| "start": 202, | |
| "end": 215, | |
| "text": "Turney, 2006)", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 798, | |
| "end": 811, | |
| "text": "Turney (2008)", | |
| "ref_id": "BIBREF20" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "There has been much previous work on corpus-based models to extract broad classes of related words. The literature on word space models (Sahlgren, 2006) has focused on taxonomic similarity (synonyms, antonyms, co-hyponyms. . . ) and general association (e.g., finding topically related words), exploiting the idea that taxonomically or associated words will tend to occur in similar contexts, and thus share a vector of cooccurring words. The literature on relational similarity, on the other hand, has focused on pairs of words, devising various methods to compare how similar the contexts in which target pairs appear are to the contexts of other pairs that instantiate a relation of interest (Turney, 2006; Pantel and Pennacchiotti, 2006) . Beyond these domains, purely corpus-based methods play an increasingly important role in modeling constraints on composition of words, in particular verbal selectional preferences -finding out that, say, children are more likely to eat than apples, whereas the latter are more likely to be eaten (Erk, 2007; . Tasks of this sort differ from relation extraction in that we need to capture productive patterns: we want to find out that shabu shabu (a Japanese meat dish) is eaten whereas ink is not, even if in our corpus neither noun is attested in proximity to forms of the verb to eat. Turney (2008) is the first, to the best of our knowledge, to raise the issue of a unified approach. In particular, he treats synonymy and association as special cases of relational similarity: in the same way in which we might be able to tell that hands and arms are in a part-of relation by comparing the contexts in which they co-occur to the contexts of known part-of pairs, we can guess that cars and automobiles are synonyms by comparing the contexts in which they co-occur to the contexts linking known synonym pairs.", | |
| "cite_spans": [ | |
| { | |
| "start": 136, | |
| "end": 152, | |
| "text": "(Sahlgren, 2006)", | |
| "ref_id": "BIBREF16" | |
| }, | |
| { | |
| "start": 695, | |
| "end": 709, | |
| "text": "(Turney, 2006;", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 710, | |
| "end": 741, | |
| "text": "Pantel and Pennacchiotti, 2006)", | |
| "ref_id": "BIBREF14" | |
| }, | |
| { | |
| "start": 1040, | |
| "end": 1051, | |
| "text": "(Erk, 2007;", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 1331, | |
| "end": 1344, | |
| "text": "Turney (2008)", | |
| "ref_id": "BIBREF20" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Here, we build on Turney's work, adding two main methodological innovations that allow us further generalization. First, merging classic approaches to taxonomic and relational similarity, we represent concept pairs by a vector that concatenates information about the contexts in which the two words occur independently, and the contexts in which they co-occur (Mirkin et al. 2006 also integrate information from the lexical patterns in which two words co-occur and similarity of the contexts in which each word occurs on its own, to improve performance in lexical entailment acquisition). Second, we represent contexts as bag of words and bigrams, rather than strings of words (\"patterns\") of arbitrary length: we leave it to the machine learning algorithm to zero in on the most interesting words/bigrams. Thanks to the concatenated vector, we can tackle tasks in which the two words are not expected to co-occur even in very large corpora (such as selectional preference). Concatenation, together with unigram/bigram representation of context, allows us to scale down the approach to smaller training corpora (Turney used a corpus of more than 50 billion words), since we do not need to see the words directly cooccurring, and the unigram/bigram dimensions of the vectors are less sparse than dimensions based on longer strings of words. We show that our method produces reasonable results also on a corpus of 2 billion words, with many unseen pairs. Moreover, our bigram and unigram representation is general enough that we do not need to extract separate statistics nor perform adhoc feature selection for each task: we build the cooccurrence matrix once, and use the same matrix in all experiments. The bag-of-words assumption also makes for faster and more compact model building, since the number of features we extract from a context is linear in the number of words in the context, whereas it is exponential for Turney. On the other hand, our method is currently lagging behind Turney's in terms of performance, suggesting that at least some task-specific tuning will be necessary.", | |
| "cite_spans": [ | |
| { | |
| "start": 360, | |
| "end": 379, | |
| "text": "(Mirkin et al. 2006", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Following Turney, we focus on devising a suitably general featural representation, and we see the specific machine learning algorithm employed to perform the various tasks as a parameter. Here, we use Support Vector Machines since they are a particularly effective general-purpose method. In terms of empirical evaluation of the model, besides experimenting with the \"classic\" SAT and TOEFL datasets, we show how our algorithm can tackle the selectional preference task proposed in Pad\u00f3 (2007) -a regression task -and we introduce to the corpus-based semantics community a challenge from the ConceptNet repository of commonsense knowledge (extending such repository by automated means is the original motivation of our project).", | |
| "cite_spans": [ | |
| { | |
| "start": 482, | |
| "end": 493, | |
| "text": "Pad\u00f3 (2007)", | |
| "ref_id": "BIBREF13" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In the next section, we will present our proposed method along with the corpora and model parameter choices used in the implementation. In Section 3, we describe the tasks that we use to evaluate the model. Results are reported in Section 4 and we conclude in Section 5, with a brief overview of the contributions of this paper.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The central idea in BagPack (Bag-of-words representation of Paired concept knowledge) is to construct a vector-based representation of a pair of words in such a way that the vector represents both the contexts where the two words co-occur and the contexts where the single words occur on their own. A straightforward approach is to construct three different sub-vectors, one for the first word, one for the second word, and one for the co-occurring pair. The concatenation of these three sub-vectors is the final vector that represents the pair.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "This approach provides us a graceful fall back mechanism in case of data scarcity. Even if the two words are not observed co-occurring in the corpus -no syntagmaic information about the pair -, the corresponding vector will still represent the individual contexts where the words are observed on their own. Our hypothesis (and hope) is that this information will be representative of the semantic relation between the pair, in the sense that, given pairs characterized by same relation, there should be paradigmatic similarity across the first, resp. second elements of the pairs (e.g., if the relation is between professionals and the typical tool of their trade, it is reasonable to expect that that both professionals and tools will tend to share similar contexts).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "Before going into further details, we need to describe what a \"co-occurrence\" precisely means, define the notion of context, and determine how to structure our vector. For a single word W , the following pseudo regular expression identifies an observation of occurrence:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\"C W D\"", | |
| "eq_num": "(1)" | |
| } | |
| ], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "where C and D can be empty strings or concatenations of up to 4 words separated by whitespace (i.e. C 1 , . . . , C i and D 1 , . . . , D j where i, j \u2264 4). Each observation of this pattern constitutes a single context of W . The pattern is matched with the longest possible substring without crossing sentence boundaries. Let (W 1 , W 2 ) denote an ordered pair of words W 1 and W 2 . We say the two words occur as a pair whenever one of the following pseudo regular expressions is observed in the corpus:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\"C W 1 D W 2 E\" (2) \"C W 2 D W 1 E\"", | |
| "eq_num": "(3)" | |
| } | |
| ], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "where C and E can be empty strings or concatenations of up to 2 words and similarly, D can be either an empty string or concatenation of up to 5 words (i.e. C 1 , . . . , C i , D 1 , . . . , D j , and E 1 , . . . , E k where i, j \u2264 2 and k \u2264 5). Together, patterns 2 and 3 constitute the pair context for W 1 and W 2 . The pattern is matched with the longest possible substring while making sure that D does not contain neither W 1 nor W 2 . The number of context words allowed before, after, and between the targets are actually model parameters but for the experiments reported in this study, we used the aforementioned values with no attempt at tuning.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The vector representing", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "(W 1 , W 2 ) is a concatenation v 1 v 2 v 1,2", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": ", where, the sub-vectors v 1 and v 2 are constructed by using the single contexts of W 1 and W 2 correspondingly (i.e. by pattern 1) and the sub-vector v 1,2 is built by using the pair contexts identified by the patterns 2 and 3. We refer to the components as single-occurrence vectors and pair-occurrence vector respectively.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The population of BagPack starts by identifying the b most frequent unigrams and the b most frequent bigrams as basis terms. Let T denote a basis term. For the construction of v 1 , we create two features for each term T : t pre corresponds to the number of observations of T in the single contexts of W 1 occurring before W 1 and t post corresponds to the number of observations of T in the single occurrence of W 1 where T occurs after W 1 (i.e. number of observations of the pattern 1 where T \u2208 C and T \u2208 D correspondingly). The construction of v 2 is identical except that this time the features correspond to the number of times the basis term is observed before and after the target word W 2 in single contexts. The construction of the pair-occurrence subvector v 1,2 proceeds in a similar fashion but in addition, we incorporate also the order of W 1 and W 2 as they co-occur in the pair context: The number of observations of the pair contexts where W 1 occurs before W 2 and T precedes (follows) the pair, are represented by feature t +pre (t +post ). The number of cases where the basis term is in between the target words is represented by t +betw . The number of cases where W 2 occurs before W 1 and T precedes the pair is represented by the feature t \u2212pre . Similarly the number of cases where T follows (is in between) the pair is represented by the feature t \u2212post (t \u2212betw ).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "Assume that the words \"only\" and \"that\" are our basis terms and consider the following context for the word pair (\"cat\", \"lion\"): \"Lion is the only cat that lives in large social groups.\" The observation of the basis terms should contribute to the pair-occurrence subvector v 1,2 and since the target words occur in reverse order, this context results in the incrementation of the features only \u2212betw and that \u2212post by one.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "To sum up, we have 2b basis terms (b unigrams and b bigrams). Each of the single-occurrence sub-vectors v 1 and v 2 consists of 4b features: Each basis term gives rise to 2 features incorporating the relative position of basis term with respect to the single word. The pair-occurrence sub-vector, v 1,2 , consists of 12b features: Each basis term gives rise to 6 new features; \u00d73 for possible relative positions of the basis term with respect to the pair and \u00d72 for the order of the words. Importantly, the 2b basis terms are picked only once, and the overall co-occurrence matrix is built once and for all for all the tasks: unlike Turney, we do not need to go back to the corpus to pick basis terms and collect separate statistics for different tasks.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The specifics of the adaptation to each task will be detailed in Section 3. For the moment, it should suffice to note that the vectors v 1 and v 2 represent the contexts in which the two words occur on their own, thus encode paradigmatic information. However, v 1,2 represents the contexts in which the two words co-occur, thus encode sytagmatic information.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The model training and evaluation is done in a 10fold cross-validation setting whenever applicable. The reported performance measures are the averages over all folds and the confidence intervals are calculated by using the distribution of fold-specific results. The only exception to this setting is the SAT analogy questions task simply because we consider each question as a separate mini dataset as described in Section 3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "We carried out our tests on two different corpora: ukWaC, a Web-derived, POS-tagged and lemmatized collection of about 2 billion tokens, 1 and the Yahoo! database queried via the BOSS service. 2 We will refer to these corpora as ukWaC and Yahoo from now on.", | |
| "cite_spans": [ | |
| { | |
| "start": 193, | |
| "end": 194, | |
| "text": "2", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Source Corpora", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "In ukWaC, we limited the number of occurrence and co-occurrence queries to the first 5000 observations for computational efficiency. Since we collect corpus statistics at the lemma level, we construct Yahoo! queries using disjunctions of inflected forms that were automatically generated with the NodeBox Linguistics library. 3 For example, the query to look for \"lion\" and \"cat\" with 4 words in the middle is: \"(lion OR lions) * * * * (cat OR cats OR catting OR catted)\". Each pair requires 14 Yahoo! queries (one for W 1 , one for W 2 , 6 for (W 1 , W 2 ), in that order, with 0-to-5 intervening words, 6 analogous queries for (W 2 , W 1 )). Yahoo! returns maximally 1,000 snippets per query, and the latter are lemmatized with the TreeTagger 4 before feature extraction.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Source Corpora", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "We did not carry out a search for \"good\" parameter values. Instead, the model parameters are generally picked at convenience to ease memory requirements and computational efficiency. For instance, in all experiments, b is set to 1500 unless noted otherwise in order to fit the vectors of all pairs at our hand into the computer memory.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model implementation", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "Once we construct the vectors for a set of word pairs, we get a co-occurrence matrix with pairs on the rows and the features on the columns. In all of our experiments, the same normalization method and classification algorithm is used with the default parameters: First, a TF-IDF feature weighting is applied to the cooccurrence matrix (Salton and Buckley, 1988) . Then following the suggestion of Hsu and Chang (2003) , each feature t's [\u03bc t \u2212 2\u03c3 t ,\u03bc t + 2\u03c3 t ] interval is scaled to [0, 1], trimming the exceeding values from upper and lower bounds (the symbols\u03bc t and\u03c3 t denote the average and standard deviation of the feature values respectively). For the classification algorithm, we use the C-SVM classifier and for regression the -SVM regressor, both implemented in the Matlab toolbox of Canu et al. (2005) . We employed a linear kernel. The cost parameter C is set to 1 for all experiments; for the regressor, = 0.2. For other pattern recognition related coding (e.g., cross validation, scaling, etc.) we made use of the Matlab PRTools (Duin, 2001) .", | |
| "cite_spans": [ | |
| { | |
| "start": 336, | |
| "end": 362, | |
| "text": "(Salton and Buckley, 1988)", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 398, | |
| "end": 418, | |
| "text": "Hsu and Chang (2003)", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 797, | |
| "end": 815, | |
| "text": "Canu et al. (2005)", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 1046, | |
| "end": 1058, | |
| "text": "(Duin, 2001)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model implementation", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "For each task that will be defined in the next section, we evaluated our algorithm on the following representations: 1) Single-occurrence vectors (", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model implementation", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "v 1 v 2 condition) 2) Pair-occurrence vectors (v 1,2 condition) 3) Entire co-occurrence matrix (v 1 v 2 v 1,2 condition).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model implementation", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "The first task we evaluated our algorithm on is the SAT analogy questions task introduced by Turney et al. (2003) . In this task, there are 374 multiple choice questions with a pair of related words like (lion,cat) as the stem and 5 other pairs as the choices. The correct answer is the choice pair which has the relationship most similar to that in the stem pair.", | |
| "cite_spans": [ | |
| { | |
| "start": 93, | |
| "end": 113, | |
| "text": "Turney et al. (2003)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SAT Analogy Questions", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "We adopt a similar approach to the one used in Turney (2008) and consider each question as a separate binary classification problem with one positive training instance and 5 unknown pairs. For a question, we pick a pair at random from the stems of other questions as a pseudo negative instance and train our classifier on this two-instance training set. Then the trained classifier is evaluated on the choice pairs and the pair with the highest posterior probability for the positive class is called the winner. The procedure is repeated 10 times picking a different pseudo-negative instance each time and the choice pair which is selected as the winner most often is taken as the answer to that question. The performance measure on this task is defined as the percentage of correctly answered questions. The mean score and confidence intervals are calculated over the performance scores obtained for all folds.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SAT Analogy Questions", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "This task, introduced by Landauer and Dumais (1997), consists of 80 multiple choice questions in which a word is given as the stem and the correct choice is the word which has the closest meaning to that of the stem, among 4 candidates. To fit the task into our framework, we pair each choice with the stem word and obtain 4 word pairs for each question. The word pair constructed with the stem and the correct choice is labeled as positive and the other pairs are labeled as negative. We consider all 320 pairs constructed for all 80 questions as our dataset. Thus, the problem is turned into a binary classification problem where the task is to discriminate the synonymous word pairs (i.e. positive class) from the other pairs (i.e. negative class). We made sure that the pairs constructed for the same question were never split between training and test set, so that no question-specific learning is performed. The reason for this precaution is that the evaluation is done on a per-question basis. The estimated posterior class probabilities of the pairs constructed for the same question are compared to each other and the pair with the highest probability for the positive class is selected as the answer for the question. By keeping the pairs of a question in the same set we make sure their posteriors are calculated by the same trained classifier. The performance measure is the percentage of correctly answered questions and we report the mean performance over all 10 folds.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "TOEFL Synonym Questions", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "Linguists have long been interested in the semantic constraints that verbs impose on their arguments, a broad area that has also attracted computational modeling, with increasing interest in purely corpus-based methods (Erk, 2007; . This task is of particular interest to us as an example of a broader class of linguistic problems that involve productive constraints on composition. As has been stressed at least since Chomsky's early work (Chomsky, 1957) , no matter how large a corpus is, if a phenomenon is productive there will always be new well-formed instances that are not in the corpus. In the domain of selectional restrictions this is particularly obvious: we would not say that an algorithm learned the constraints on the possible objects/patients of eating simply by producing the list of all the attested objects of this verb in a very large corpus; the interesting issue is whether the algorithm can detect if an unseen object is or is not a plausible \"eatee\", like humans do without problems. Specifically, we test selectional preferences on the dataset constructed by Pad\u00f3 (2007) , that collects average plausibility judgments (from 20 speakers) for nouns as either subjects or objects of verbs (211 noun-verb pairs).", | |
| "cite_spans": [ | |
| { | |
| "start": 219, | |
| "end": 230, | |
| "text": "(Erk, 2007;", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 440, | |
| "end": 455, | |
| "text": "(Chomsky, 1957)", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 1085, | |
| "end": 1096, | |
| "text": "Pad\u00f3 (2007)", | |
| "ref_id": "BIBREF13" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selectional Preference Judgments", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "We formulate this task as a regression problem. We train the -SVM regressor with 18-fold cross validation: Since the pair instances are not independent but grouped according to the verbs, one fold is constructed for each of the 18 verbs used in the dataset. In each fold, all instances sharing the corresponding verb are left out as the test set. The performance measure for this task is the Spearman correlation between the human judgments and our algorithm's estimates. There are two possible ways to calculate this measure. One is to get the overall correlation between the human judgments and our estimates obtained by concatenating the output of each cross-validation fold. That measure allows us to compare our method with the previously reported results. However, it cannot control for a possible verb-effect on the human judgment values: If the average judgment values of the pairs associated with a specific verb is significantly higher (or lower) than the average of the pairs associated with another verb, then any regressor which simply learns to assign the average value to all pairs associated with that verb (regardless of whether there is a patient or agent relation between the pairs) will still get a reasonably high correlation because of the variation of judgment scores across the verbs. To control for this effect, we also calculated the correlation between the human judgments and our estimates for each verb's plausibility values separately, and we report averages across these separate correlations (the \"mean\" results reported below).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selectional Preference Judgments", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "Open Mind Common Sense 5 is an ongoing project of acquisition of common-sense knowledge from ordinary people by letting them carry out simple semantic and linguistics tasks. An end result of the project is Con-ceptNet 3, a large scale semantic network consisting of relations between concept pairs (Havasi et al., 2007) . It is possible to view this network as a collection of semantic assertions, each of which can be represented by a triple involving two concepts and a relation between them, e.g. UsedFor(piccolo, make music). One motivation for this project is the fact that common-sense knowledge is assumed to be known by both parties in a communication setting and usually is not expressed explicitly. Thus, corpus-based approaches may have serious difficulties in capturing these relations (Havasi et al., 2007) , but there are reasons to believe that they could still be useful: Eslick (2006) uses the assertions of ConceptNet as seeds to parse Web search results and augment ConceptNet by new candidate relations. We use the ConceptNet snapshot released in June 2008, containing more than 200.000 assertions with around 20 semantic relations like UsedFor, Desirious-EffectOf, or SubEventOf. Each assertion has a confidence rating based on the number of people who expressed or confirmed that assertion. For simplicity we limited ourselves to single word concepts and the relations between them. Furthermore, we eliminated the assertions with a confidence score lower than 3 in an attempt to increase the \"quality\" of the assertions and focused on the most populated 5 relations of the remaining set, as given in Table 3 .4. There may be more than one relation between a pair of concepts, so the total number is less than the sum of the size of the individual relation sets.", | |
| "cite_spans": [ | |
| { | |
| "start": 298, | |
| "end": 319, | |
| "text": "(Havasi et al., 2007)", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 798, | |
| "end": 819, | |
| "text": "(Havasi et al., 2007)", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 888, | |
| "end": 901, | |
| "text": "Eslick (2006)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 1622, | |
| "end": 1629, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Common-sense Relations from ConceptNet", | |
| "sec_num": "3.4" | |
| }, | |
| { | |
| "text": "For the multiple choice question tasks (i.e. SAT and TOEFL), we say a question is complete when all of the related pairs (stem and choice) are represented by vectors with at least one non-zero component. If a question has at least one pair represented by a zero-vector (missing pairs), then we say that the question is partial. For these tasks, we report the worst-case performance scores where we assume that a random guessing performance is obtained on the partial questions. This is a strict lower bound because it discards all information we have about a partial question even if it has only one missing pair. We define coverage as the percentage of complete questions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Results", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In Yahoo, the coverage is quite high. In the v 1,2 only condition, 4 questions had at least some choice/stem pairs with all zero components. In all other cases, all of the pairs were represented by vectors with at least one non-zero component. The highest score is obtained for the v 1 v 2 v 1,2 condition with a 44.1% of correct questions, that is not significantly above the 42.5% performance of v 1,2 (paired t-test, \u03b1 = 0.05). The v 1 v 2 only condition results in a poorer performance of 33.9% correct questions, statistically lower than the former two conditions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "For ukWaC, the v 1,2 only condition provides a relatively low coverage. Only 238 questions out of 374 were complete. For the other conditions, we get a complete coverage. The performances are statistically indistinguishable from each other and are 38.0%, 38.2%, and 39.6% Table 2 : Percentage of correctly answered questions in SAT analogy task, worst-case scenario.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 272, | |
| "end": 279, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "for v 1,2 , v 1 v 2 , and v 1 v 2 v 1,2 respectively. Condition Yahoo ukWaC v 1,2 42.5% 38.0% v 1 v 2 33.9% 38.2% v 1 v 2 v 1,2 44.1% 39.6%", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "In Fig. 1 , the best performances we get for Yahoo and ukWaC are compared to previous studies with 95% binomial confidence intervals plotted. The reported values are taken from the ACL wiki page on the state of the art for SAT analogy questions 6 . The algorithm proposed by Turney (2008) is labeled as Turney-PairClass. M a n g a la th e t a l.", | |
| "cite_spans": [ | |
| { | |
| "start": 275, | |
| "end": 288, | |
| "text": "Turney (2008)", | |
| "ref_id": "BIBREF20" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 3, | |
| "end": 9, | |
| "text": "Fig. 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "V e a le \u2212 K N O W \u2212 B E S T B ic ic i a n d Y u re t B a g P a ck \u2212 Y a h o o T u rn e y a n d L itt m a n T u rn e y\u2212 P a ir C la ss T u rn e y\u2212 P E R T T u rn e y\u2212 L R A Figure 1 : Comparison with previous algorithms on SAT analogy questions.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 173, | |
| "end": 181, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Overall, the performance of BagPack is not at the level of the state of the art but still provides a reasonable level even in the v 1 v 2 only condition for which we do not utilize the contexts where the two words co-occur. This aspect is most striking for ukWaC where the coverage is low and by only utilizing the single-occurrence sub-vectors we obtain a performance of 38.2% correct answers (the comparable \"attributional\" models re-ported in Turney, 2006 , have an average performance of 31%).", | |
| "cite_spans": [ | |
| { | |
| "start": 446, | |
| "end": 458, | |
| "text": "Turney, 2006", | |
| "ref_id": "BIBREF19" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "For the v 1,2 sub-vector calculated for Yahoo, we have two partial questions out of 80 and the system answers 80.0% of the questions correctly. The single occurrence case v 1 v 2 instead provides a correct percentage of 41.2% which is significantly above the random performance of 25% but still very poor. The combined case v 1 v 2 v 1,2 provides a score of 75.0% with no statistically significant difference from the v 1,2 case. The reason of the low performance for", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "TOEFL", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "v 1 v 2 is an open question.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "TOEFL", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "For ukWaC, the coverage for the v 1 v 2 case is pretty low. Out of 320 pairs, 70 were represented by zerovectors, resulting in 34 partial questions out of 80. The performance is at 33.8%. The v 1 v 2 case on its own does not lead to a performance better than random guessing (27.5%) but the combined case", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "TOEFL", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "v 1 v 2 v 1,2", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "TOEFL", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "provides the highest ukWaC score of 42.5%. Table 3 : Percentage of correctly answered questions in TOEFL synonym task, worst-case scenario.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 43, | |
| "end": 50, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "TOEFL", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "v 1,2 80.0% 33.8% v 1 v 2 41.2% 27.5% v 1 v 2 v 1,2 75.0% 42.5%", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Condition Yahoo ukWaC", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "To our knowledge, the best performance with a purely corpus-based approach is that of Rapp (2003) who obtained a score of 92.5% with SVD. Fig. 2 reports our results and a list of other corpus-based systems which achieve scores higher than 70%, along with 95% confidence interval values. The results are taken from the ACL wiki page on the state of the art for TOEFL synonym questions. T u rn e y\u2212 P M I\u2212 IR T u rn e y\u2212 P a ir C la ss B a g P a ck \u2212 Y a h o o T e rr a a n d C la rk e B u lli n a ri a a n d L e vy M a tv e e va e t a l. R a p p Figure 2 : Comparison with previous algorithms on TOEFL synonym questions with 95% confidence intervals.", | |
| "cite_spans": [ | |
| { | |
| "start": 86, | |
| "end": 97, | |
| "text": "Rapp (2003)", | |
| "ref_id": "BIBREF15" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 138, | |
| "end": 144, | |
| "text": "Fig. 2", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 545, | |
| "end": 553, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Condition Yahoo ukWaC", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We note that our results obtained for Yahoo are comparable to the results of Turney but even the best results obtained for ukWaC and the Yahoo's results for v 1 v 2 only condition are very poor. Whether this is because of the inability of the sub-vectors to capture synonymity or because the default parameter values of SVM are not adequate is an open question. Notice that our concatenated v 1 v 2 vector does not exploit information about the similarity of v 1 to v 2 , that, presumably, should be of great help in solving the synonym task.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Condition Yahoo ukWaC", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The coverage for this dataset is quite high. All pairs were represented by non-zero vectors for Yahoo while only two pairs had zero-vectors for ukWaC. The two pairs are discarded in our experiments. For Yahoo, the best results are obtained for the v 1,2 case. The singleoccurrence case, v 1 v 2 , provides an overall correlation of 0.36 and mean correlation of 0.26. However low, in case of rarely co-occurring word pairs this data could be the only data we have in our hands and it is important that it provides reasonable judgment estimates.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selectional Preference", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "For the ukWaC corpus, the best results we get are an overall correlation of 0.60 and a mean correlation of 0.52 for the combined case", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selectional Preference", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "v 1 v 2 v 1,2 . The results for v 1,2 and v 1 v 2 v 1,2 are statistically indistinguishable. Yahoo ukWaC Condition Overall Mean Overall Mean v 1,2 0.60 0.45 0.58 0.48 v 1 v 2 0.36 0.26 0.33 0.22 v 1 v 2 v 1,2", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selectional Preference", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "0.55 0.42 0.60 0.52 Table 4 : Spearman correlations between the targets and estimations for selectional preference task.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 20, | |
| "end": 27, | |
| "text": "Table 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Selectional Preference", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "In Fig. 3 , we present a comparison of our results with some previous studies reported in . The best result reported so far is a correlation of 0.52. Our results for Yahoo and ukWaC are currently the highest correlation values reported. Even the verbeffect-controlled correlations achieve competitive performance. ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 3, | |
| "end": 9, | |
| "text": "Fig. 3", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Selectional Preference", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "Only for this task, (because of practical memory limitations) we reduced the model parameter b to 500, which means we used the 500 most frequent unigrams and 500 most frequent bigrams as our basis terms. For each of the 5 relations at our hand, we trained a different SVM classifier by labeling the pairs with the corresponding relation as positive and the rest as negative.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "ConceptNet", | |
| "sec_num": "4.4" | |
| }, | |
| { | |
| "text": "To eliminate the issue of unbalanced number of negative and positive instances we randomly down-sampled the positive or negative instances set (whichever is larger). For the IsA, UsedFor, CapableOf, and PartOf relations, the down-sampling procedure means keeping some of the negative instances out of the training and test sets while for the LocationOf relation it means keeping a subset of the positive instances out. We performed 5 iterations of the down-sampling procedure and for each iteration we carried out a 10-fold crossvalidation to train and test our classifier. The results are test set averages over all iterations and folds. The performance measure we use is the area under the receiver operating characteristic (AUC in short for area under the curve). The AUC of a classifier is the area under the curve defined by the corresponding true positive rate and false positive rate values obtained for varying the threshold of the classifier to accept an instance as positive. Intuitively, AUC is the probability that a randomly picked positive instance's estimated posterior probability is higher than a randomly picked negative instance's estimated posterior probability (Fawcett, 2006) . The coverage is quite high for both corpora: Out of 1943 pairs,only 3 were represented by a zero-vector in Yahoo while in ukWaC this number is 68. For simplicity, we discarded missing pairs from our analysis. We report only the results obtained for the entire cooccurrence matrix. The results are virtually identical for the other conditions too: Both for Yahoo and ukWaC, almost all of the AUC values obtained for all relations and for all conditions are above 95%. Only the PartOf relation has AUC values above 90% (which is still a very good result).", | |
| "cite_spans": [ | |
| { | |
| "start": 1182, | |
| "end": 1197, | |
| "text": "(Fawcett, 2006)", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "ConceptNet", | |
| "sec_num": "4.4" | |
| }, | |
| { | |
| "text": "Yahoo ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Relation", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "v 1 v 2 v 1,2 condition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Relation", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The very high performance we observe for the Con-ceptNet task is surprising when compared to the moderate performance we observe for other tasks. Our extensive filtering of the assertions could have resulted in a biased dataset which might have made the job of the classifier easy while reducing its generalization ca-pacity. To investigate this, we decided to use the pairs coming from the SAT task as a validation set.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Relation", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Again, we trained an SVM classifier on the Concept-Net data for each of the 5 relations like we did previously, but this time without cross-validation (i.e. after the down-sampling, we used the entire set as the training dataset in each iteration). Then we evaluated the classifiers on the 2224 pairs of the SAT analogy task (removing pairs that were in the training data) and averaged the posterior probability reported by each SVM over each down-sampling iteration. The 5 pairs which are assigned the highest posterior probability for each relation are reported in Table 6 . We have not yet quantified the performance of BagPack in this task but the preliminary results in this table are, qualitatively, exceptionally good.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 567, | |
| "end": 574, | |
| "text": "Table 6", | |
| "ref_id": "TABREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Relation", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We presented a general way to build a vector-based space to represent the semantic relations between word pairs and showed how that representation can be used to solve various tasks involving semantic similarity. For SAT and TOEFL, we obtained reasonable performances comparable to the state of the art. For the estimation of selective preference judgments about verbnoun pairs, we achieved state of the art performance. Perhaps more importantly, our representation format allows us to provide meaningful estimates even when the verb and noun are not observed co-occurring in the corpus -which is an obvious advantage over the models which rely on sytagmatic contexts alone and cannot provide estimates for word pairs that are not seen directly co-occurring. We also obtained very promising results for the automated augmentation of ConceptNet.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "The generality of the proposed method is also reflected in the fact that we built a single feature space based on frequent basis terms and used the same features for all pairs coming from different tasks. The use of the same feature set for all pairs makes it possible to build a single database of word-pair vectors. For example, we were able to re-use the vectors constructed for SAT pairs as a validation set in the Con-ceptNet task. Furthermore, the results reported here are obtained for the same machine learning model (SVM) without any parameter tweaking, which renders them very strict lower bounds.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Another contribution is that the proposed method provides a way to represent the relations between words even if they are not observed co-occurring in the corpus. Employing a larger corpus can be an alternative solution for some cases but this is not always possible and some tasks, like estimating selectional preference judgments, inherently call for a method that does not exclusively depends on paired co-occurrence observations.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Finally, we introduced ConceptNet, a common-sense semantic network, to the corpus-based semantics community, both as a new challenge and as a repository we", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "UsedFor PartOf CapableOf LocationOf 1 watch,timepiece pencil,draw vehicle,wheel motorist,drive spectator,arena 2 emerald,gem blueprint,build spider,leg volatile,vaporize water,riverbed 3 cherry,fruit detergent,clean keyboard,finger concrete,harden bovine,pasture 4 dinosaur,reptile guard,protect train,caboose parasite,contribute benediction,church 5 ostrich,bird buttress,support hub,wheel immature,develop byline,newspaper ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 8, | |
| "end": 375, | |
| "text": "PartOf CapableOf LocationOf 1 watch,timepiece pencil,draw vehicle,wheel motorist,drive spectator,arena 2 emerald,gem blueprint,build spider,leg volatile,vaporize water,riverbed 3 cherry,fruit detergent,clean keyboard,finger concrete,harden bovine,pasture 4 dinosaur,reptile guard,protect train,caboose parasite,contribute benediction,church 5", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "IsA", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "v 1 v 2 v 1,2 condition.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "IsA", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "can benefit from.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "IsA", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In future work, one of the most pressing issue we want to explore is how to better exploit the information in the single occurrence vectors: currently, we do not make any use of the overlap between v 1 and v 2 . In this way, we are missing the classic intuition that taxonomically similar words tend to occur in similar contexts, and it is thus not surprising that v 1 v 2 flunks the TOEFL. We are currently looking at ways to augment our concatenated vector with \"meta-information\" about vector overlap.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "IsA", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://wacky.sslmit.unibo.it", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://developer.yahoo.com/search/ boss/ 3 http://nodebox.net/code/index.php/ Linguistics 4 http://www.ims.uni-stuttgart.de/ projekte/corplex/TreeTagger/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://commons.media.mit.edu/en/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "See http://aclweb.org/aclwiki/ for further information and references", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "SVM and Kernel Methods Matlab Toolbox, Perception Syst\u00e8mes et Information, INSA de Rouen", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Canu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Grandvalet", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Guigue", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Rakotomamonjy", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Canu, Y. Grandvalet, V. Guigue and A. Rakotoma- monjy. 2005. SVM and Kernel Methods Matlab Toolbox, Perception Syst\u00e8mes et Information, INSA de Rouen, Rouen, France", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Syntactic structures", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Chomsky", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1957, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Chomsky. 1957. Syntactic structures. Mouton, The Hague.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "PRTOOLD (Version 3.1.7), A Matlab toolbox for pattern recognition", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "P W" | |
| ], | |
| "last": "Duin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. P. W. Duin. 2001. PRTOOLD (Version 3.1.7), A Matlab toolbox for pattern recognition. Pattern Recognition Group. Delft University of Technology.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "A simple, similarity-based model for selectional preferences", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Erk", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings of ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Erk. 2007. A simple, similarity-based model for selectional preferences. Proceedings of ACL 2007.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "A structured vector space model for word meaning in context", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Erk", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Pad\u00f3", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proceedings of EMNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Erk and S. Pad\u00f3. 2008. A structured vector space model for word meaning in context. Proceedings of EMNLP 2008.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Searching for commonsense", | |
| "authors": [ | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Eslick", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "I. Eslick. 2006. Searching for commonsense. Master's thesis, Massachusetts Institute of Technology.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "An introduction to roc analysis", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Fawcett", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Pattern Recogn. Lett", | |
| "volume": "27", | |
| "issue": "8", | |
| "pages": "861--874", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Fawcett. 2006. An introduction to roc analysis. Pattern Recogn. Lett., 27(8):861-874.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Conceptnet 3: a flexible, multilingual semantic network for common sense knowledge", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Havasi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Speer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Alonso", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Recent Advances in Natural Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "C. Havasi, R. Speer and J. Alonso. 2007. Concept- net 3: a flexible, multilingual semantic network for common sense knowledge. In Recent Advances in Natural Language Processing, Borovets, Bulgaria, September.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "A practical guide to support vector classification", | |
| "authors": [ | |
| { | |
| "first": "C.-W", | |
| "middle": [], | |
| "last": "Hsu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C.-C", | |
| "middle": [], | |
| "last": "Chang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "C.-W. Hsu, C.-C Chang. 2003. A practical guide to support vector classification. Technical report, Department of Computer Science, National Taiwan University.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "A solution to Plato's problem: The Latent Semantic Analysis theory of acquisition, induction and representation of knowledge", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Landauer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Dumais", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "Psychological Review", | |
| "volume": "104", | |
| "issue": "2", | |
| "pages": "211--240", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T.K. Landauer and S.T. Dumais. 1997. A solution to Plato's problem: The Latent Semantic Analysis theory of acquisition, induction and representation of knowledge. Psychological Review, 104(2): 211- 240.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "ConceptNet -A practical commonsense reasoning tool-kit", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Singh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "BT Technology Journal", | |
| "volume": "22", | |
| "issue": "4", | |
| "pages": "211--226", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. Liu and P. Singh. 2004. ConceptNet -A practi- cal commonsense reasoning tool-kit. BT Technology Journal, 22(4) 211-226.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Integrating pattern-based and distributional similarity methods for lexical entailment acquisition", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Mirkin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Dagan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Geffet", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of COLING/ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "579--586", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Mirkin, I. Dagan and M. Geffet. 2006. Integrat- ing pattern-based and distributional similarity meth- ods for lexical entailment acquisition. Proceedings of COLING/ACL 2006, 579-586.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Flexible, corpusbased modelling of human plausibility judgements", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Pad\u00f3", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Pad\u00f3", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Erk", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings EMNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "400--409", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Pad\u00f3, S. Pad\u00f3 and K. Erk. 2007. Flexible, corpus- based modelling of human plausibility judgements. Proceedings EMNLP 2007, 400-409.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "The Integration of Syntax and Semantic Plausibility in a Wide-Coverage Model of Sentence Processing", | |
| "authors": [ | |
| { | |
| "first": "U", | |
| "middle": [], | |
| "last": "Pad\u00f3", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "U. Pad\u00f3. 2007. The Integration of Syntax and Semantic Plausibility in a Wide-Coverage Model of Sentence Processing. Ph.D. thesis, Saarland University.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Espresso: Leveraging generic patterns for automatically harvesting semantic relations", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Pantel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Pennacchiotti", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of COL-ING/ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "113--120", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P. Pantel and M. Pennacchiotti. 2006. Espresso: Leveraging generic patterns for automatically har- vesting semantic relations. Proceedings of COL- ING/ACL 2006, 113-120.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Word sense discovery based on sense descriptor dissimilarity", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Rapp", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proceedings of MT", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "315--322", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. Rapp. 2003. Word sense discovery based on sense descriptor dissimilarity. Proceedings of MT Summit IX: 315-322.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "The Word-space model", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Sahlgren", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "M. Sahlgren. 2006. The Word-space model. Ph.D. dis- sertation, Stockholm University, Stockholm.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Term-weighting approaches in automatic text retrieval", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Salton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Buckley", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1988, | |
| "venue": "Information Processing and Management", | |
| "volume": "24", | |
| "issue": "5", | |
| "pages": "513--523", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Salton and C. Buckley. 1988. Term-weighting approaches in automatic text retrieval. Information Processing and Management, 24(5): 513-523.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Analogyspace: Reducing the dimensionality of common sense knowledge", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Speer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Havasi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Lieberman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "AAAI", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "548--553", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "R. Speer, C. Havasi and H. Lieberman. 2008. Anal- ogyspace: Reducing the dimensionality of common sense knowledge. In Dieter Fox and Carla P. Gomes, editors, AAAI, pages 548-553. AAAI Press.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Similarity of semantic relations", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Turney", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Computational Linguistics", | |
| "volume": "32", | |
| "issue": "3", | |
| "pages": "379--416", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P. Turney. 2006. Similarity of semantic relations. Computational Linguistics, 32(3): 379-416.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "A uniform approach to analogies, synonyms, antonyms and associations", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Turney", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proceedings of COLING 2008", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "905--912", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P. Turney. 2008. A uniform approach to analogies, synonyms, antonyms and associations. Proceedings of COLING 2008, 905-912.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF2": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Comparison of algorithms on selectional preference task.", | |
| "num": null | |
| }, | |
| "TABREF2": { | |
| "num": null, | |
| "type_str": "table", | |
| "text": "AUC scores for 5 relations of ConceptNet, classifier trained for", | |
| "content": "<table/>", | |
| "html": null | |
| }, | |
| "TABREF3": { | |
| "num": null, | |
| "type_str": "table", | |
| "text": "Top 5 SAT pairs classified as positive for ConceptNet relations, classifier trained for", | |
| "content": "<table/>", | |
| "html": null | |
| } | |
| } | |
| } | |
| } |