{ "paper_id": "2021", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T06:07:04.770625Z" }, "title": "Emotion Ratings: How Intensity, Annotation Confidence and Agreements are Entangled", "authors": [ { "first": "Enrica", "middle": [], "last": "Troiano", "suffix": "", "affiliation": { "laboratory": "", "institution": "Sprachverarbeitung University of Stuttgart", "location": { "country": "Germany" } }, "email": "" }, { "first": "Sebastian", "middle": [], "last": "Pad\u00f3", "suffix": "", "affiliation": { "laboratory": "", "institution": "Sprachverarbeitung University of Stuttgart", "location": { "country": "Germany" } }, "email": "" }, { "first": "Roman", "middle": [], "last": "Klinger", "suffix": "", "affiliation": { "laboratory": "", "institution": "Sprachverarbeitung University of Stuttgart", "location": { "country": "Germany" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "When humans judge the affective content of texts, they also implicitly assess the correctness of such judgment, that is, their confidence. We hypothesize that people's (in)confidence that they performed well in an annotation task leads to (dis)agreements among each other. If this is true, confidence may serve as a diagnostic tool for systematic differences in annotations. To probe our assumption, we conduct a study on a subset of the Corpus of Contemporary American English, in which we ask raters to distinguish neutral sentences from emotionbearing ones, while scoring the confidence of their answers. Confidence turns out to approximate inter-annotator disagreements. Further, we find that confidence is correlated to emotion intensity: perceiving stronger affect in text prompts annotators to more certain classification performances. This insight is relevant for modelling studies of intensity, as it opens the question wether automatic regressors or classifiers actually predict intensity, or rather human's self-perceived confidence.", "pdf_parse": { "paper_id": "2021", "_pdf_hash": "", "abstract": [ { "text": "When humans judge the affective content of texts, they also implicitly assess the correctness of such judgment, that is, their confidence. We hypothesize that people's (in)confidence that they performed well in an annotation task leads to (dis)agreements among each other. If this is true, confidence may serve as a diagnostic tool for systematic differences in annotations. To probe our assumption, we conduct a study on a subset of the Corpus of Contemporary American English, in which we ask raters to distinguish neutral sentences from emotionbearing ones, while scoring the confidence of their answers. Confidence turns out to approximate inter-annotator disagreements. Further, we find that confidence is correlated to emotion intensity: perceiving stronger affect in text prompts annotators to more certain classification performances. This insight is relevant for modelling studies of intensity, as it opens the question wether automatic regressors or classifiers actually predict intensity, or rather human's self-perceived confidence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "A plethora of theories exist on the matter of emotions: the intensity of affective states, their link to cognition, and their arrangement into categories are just a few of the angles from which psychology has tackled this complex phenomenon (Gendron and Feldman Barrett, 2009) . Correspondingly, in computational emotion analysis, texts have been associated to values of intensity (Strapparava and Mihalcea, 2007; Mohammad and Bravo-Marquez, 2017) , to cognitive components (Hofmann et al., 2020) , and discrete classes (Zhang et al., 2018; Zhong et al., 2019, i.a.) . In support of these tasks, substantial research effort has been directed to resource construction, which typically relies on the participation of human judges. Yet, emotions are a subjective experience. Their interpretation in text varies across individuals, and this poses a major challenge in emotion analysis: it is hard to reach acceptable levels of inter-annotator agreement (IAA) (Bobicev and Sokolova, 2017; Schuff et al., 2017; Troiano et al., 2019) .", "cite_spans": [ { "start": 241, "end": 276, "text": "(Gendron and Feldman Barrett, 2009)", "ref_id": "BIBREF11" }, { "start": 381, "end": 413, "text": "(Strapparava and Mihalcea, 2007;", "ref_id": "BIBREF28" }, { "start": 414, "end": 447, "text": "Mohammad and Bravo-Marquez, 2017)", "ref_id": "BIBREF19" }, { "start": 474, "end": 496, "text": "(Hofmann et al., 2020)", "ref_id": "BIBREF13" }, { "start": 520, "end": 540, "text": "(Zhang et al., 2018;", "ref_id": "BIBREF31" }, { "start": 541, "end": 566, "text": "Zhong et al., 2019, i.a.)", "ref_id": null }, { "start": 955, "end": 983, "text": "(Bobicev and Sokolova, 2017;", "ref_id": "BIBREF4" }, { "start": 984, "end": 1004, "text": "Schuff et al., 2017;", "ref_id": "BIBREF25" }, { "start": 1005, "end": 1026, "text": "Troiano et al., 2019)", "ref_id": "BIBREF30" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "On their side, humans (roughly) know how well they can \"read\" emotions (Realo et al., 2003) . They judge the affective content of texts, and at the same time, the correctness of such judgment: in other words, annotators can evaluate their own confidence with respect to their labelling decisions. This hints a possible relation between confidence and IAA. One could expect that annotators are more likely to incur inconsistencies when they feel uncertain about their answers. Hence, it would be interesting to verify if self-assessed confidence approximates inter-annotator disagreements in affectrelated annotation tasks. The collection of this type of judgments is not a common practice: past research has found that self-assigned scores of confidence are predictable based on some vocal attributes of emotion speech stimuli (Lausen and Hammerschmidt, 2020) , but this has not been done on text, to the best of our knowledge.", "cite_spans": [ { "start": 71, "end": 91, "text": "(Realo et al., 2003)", "ref_id": "BIBREF23" }, { "start": 827, "end": 859, "text": "(Lausen and Hammerschmidt, 2020)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Yet another aspect involved in emotion recognition and which, at first sight, relates to confidence, is emotion intensity. It would be intuitive to assume that emotions are recognized with higher confidence if they are expressed with stronger magnitude (e.g., \"The teacher exploded\" > \"He snapped his annoyed temper\", \"sadder\" > \"a bit sad\", \"ecstasy\" > \"joy\"). Still, in the sentence \"We had to cheer him up; later, he was off the ground\", readers have to choose what part of the text to attend to (the challenge that the speakers undertook -not intense, or the effect they had -intense) and be very confident about either choice. Similar counter examples reveal that the link between the perception of emotion intensity and the self-perceived confidence is opaque, and leaves room for exploration.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper, we experimentally investigate the relationship between three human judgments: about the presence of emotions, about their intensity, and about the confidence of the annotation decision. Leveraging such information, we aim at understanding in what cases annotators differ regarding the judgement that an emotion is expressed. Our first research question is: (RQ1) Are (dis)agreements with respect to the presence of emotions related to confidence? Second, (RQ2) Are judgments of intensity and confidence entangled? We bring these issues together in an annotation study based on emotion recognition, in which self-perceived confidence is a dimension to be rated. Given a subset of the Corpus of Contemporary American English (COCA) (Davies, 2015) , raters distinguish emotionbearing sentences from neutral ones, while quantifying both the intensity of the emotion and their confidence on a Likert scale (B\u00e8gue et al., 2018) .", "cite_spans": [ { "start": 745, "end": 759, "text": "(Davies, 2015)", "ref_id": "BIBREF8" }, { "start": 916, "end": 936, "text": "(B\u00e8gue et al., 2018)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We find that confidence can explain systematic differences in decisions of annotators; so does intensity; and impressions on intensity and confidence are correlated. Based on these results, we devise a strategy to leverage the two factors and smoothen out inter-annotator inconsistencies.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Tasks. The first step in this study is to collect emotion assessments. 1 We are not interested in which emotion people interpret from text, but rather if they recognize any. Judges read sentences and answer the question: (EMO) Is it Emotional or Neutral? For the items deemed to express an emotion, we also ask (INT) How strong is it?, which enables us to obtain ratings about affective strengths on a Likert scale from 1 (not intense) to 3 (very). Lastly, since raters interpret emotions without an immediate first-hand experience, we have them self-evaluate their own judgments on a scale from 1 (unsure) to 3 (certain), in response to the question (CONF) How confident are you about your answer to EMO?", "cite_spans": [ { "start": 71, "end": 72, "text": "1", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Annotation Setup", "sec_num": "2" }, { "text": "As for EMO, we acknowledge that the emotional content of an utterance can be inferred from many perspectives. It is possible to assess one's own emotion after reading the text, to reconstruct the affective state of the writers who produced it, to guess the reaction that they intended to elicit in the readers, and so on. To avoid confusion, we instruct annotators to consider the presence of an emotion only with respect to their personal viewpoint.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Annotation Setup", "sec_num": "2" }, { "text": "We opt for an in-lab setting. Raters are three female master students aged between 24 and 27, who are proficient in English, and have some annotation experience and background in computational emotion analysis.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Annotation Setup", "sec_num": "2" }, { "text": "Data. Corpora that include emotion classes or gradations are tailored on specific domains, like self-reports (Scherer and Wallbott, 1997) , tweets (Mohammad, 2018) and newspapers (Strapparava and Mihalcea, 2008) . We broaden our focus to multiple genres, and annotate sentences from the 2020 version of COCA 2 , which includes unlabelled texts that occurred from 1990 to present in different domains, like blogs, magazines, newspapers, academic texts, spoken interactions, fictions, TV, and movie subtitles.", "cite_spans": [ { "start": 109, "end": 137, "text": "(Scherer and Wallbott, 1997)", "ref_id": "BIBREF24" }, { "start": 147, "end": 163, "text": "(Mohammad, 2018)", "ref_id": "BIBREF18" }, { "start": 179, "end": 211, "text": "(Strapparava and Mihalcea, 2008)", "ref_id": "BIBREF29" } ], "ref_spans": [], "eq_spans": [], "section": "Annotation Setup", "sec_num": "2" }, { "text": "With a corpus of this size (>1B words), considering all data points would be costly, and randomly selecting them may cause imbalance in the final annotation -i.e., a majority of neutral instances. Therefore, we draw a sample biased towards emotional sentences with a combination of rules and classifier-based information. To obtain such a classifier, we fine-tune the pre-trained BERT (Devlin et al., 2019 ) base-case model on a number of emotion analysis resources 3 , adding a classification layer that outputs the labels emotion or neutral. Having that, we filter academic texts out of COCA for their arguably impartial language, and from each of the other genres, we randomly pick 500 sentences; out of these, we sample 100 sentences balanced by class, i.e., 50 labelled as neutral by our classifier, 50 as bearing an emotion. Thus, the annotators are shown 700 items, 100 per domain, with a balanced class distribution, according to the classifier.", "cite_spans": [ { "start": 385, "end": 405, "text": "(Devlin et al., 2019", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Annotation Setup", "sec_num": "2" }, { "text": "To answer our research questions, we first observe annotators' agreements on EMO. The highest Cohen's \u03ba (1960) between pairs of human judges was .43 (Table 1a) ; Fleiss' \u03ba (1971) for the three annotators was .34.", "cite_spans": [ { "start": 94, "end": 110, "text": "Cohen's \u03ba (1960)", "ref_id": null }, { "start": 162, "end": 178, "text": "Fleiss' \u03ba (1971)", "ref_id": null } ], "ref_spans": [ { "start": 149, "end": 159, "text": "(Table 1a)", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Results", "sec_num": "3" }, { "text": "At first glance, these numbers appear unsatisfactory. On the one hand, they are due to the skewed class distribution in the annotators' choices. 4 the other, they can be traced back to the way in which the EMO task was formulated: asking if a text is emotional from the readers' own point of view (e.g., it \"describes an event [...] to which you would associate an emotion\", see Appendix) as opposed, for instance, to the writers', paves the way for more heterogeneous responses. However, a look at other IAA measures, like the absolute counts of items that were assigned to each label, leads to a more detailed picture. Table 1b breaks down the annotated categories by agreement: column 1 vs. 2 corresponds to the groups of items on which 1 annotator chose a label, while the majority opted for the other; column 3 vs. 0 shows how many times all three annotators agreed. We see that 138 sentences out of 700 were deemed emotionally charged by only 1 person (and hence, were associated to neutral by the two others). 2 annotators picked the emotion class for 170 sentences, i.e., those which were neutral according to just 1 rater. Overall, as the amount of considered judgments increases, so does their intersubjective validity about emotions. This tendency is clear in column 3 vs. 0, which shows that there were more emotional instances with 3 identical labels than those with conflicting ratings. Indeed, perfect agreement was reached for 392 items (304 emotion and 88 neutral), more than half the data, suggesting that people had a shared understanding.", "cite_spans": [ { "start": 145, "end": 146, "text": "4", "ref_id": null } ], "ref_spans": [ { "start": 621, "end": 629, "text": "Table 1b", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Results", "sec_num": "3" }, { "text": "(and so does Intensity)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Confidence Approximates Disagreements", "sec_num": "3.1" }, { "text": "We next focus on the items that received incoherent judgments. Annotators seem to diverge on the presence/absence of emotions in a systematic way. Specifically, their inconsistencies correspond to certain patterns in the ratings of confidence, as well as intensity. Taking pairs of annotators, we see that the one who picks the emotion class tends to do that with low confidence. 5 As an example representative of the general trend, on 11 sentences annotator A3 makes the neutral choice, while annotator A1 picks emotion, but rates such items with confidence 1 (5 sentences), or 2 (6 sentences) -never using the highest degree of confidence. The same holds for intensity: A1 rates 10 of the 11 sentences as having the lowest intensity, and 1 sentence as having intensity 2 -none with the highest intensity value.", "cite_spans": [ { "start": 380, "end": 381, "text": "5", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Confidence Approximates Disagreements", "sec_num": "3.1" }, { "text": "Hence, to answer RQ1, the evaluation of intensity and the self-evaluation of confidence underlie disagreements in discrete emotion annotations. Further, they show that different intuitions are not totally incompatible, since the annotator who takes the emotion decision does so without being extremely confident, and gauging intensity as rather weak.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Confidence Approximates Disagreements", "sec_num": "3.1" }, { "text": "We corroborate this finding by looking at a more standard measure of IAA, namely Fleiss' \u03ba, which turns out to be affected by both CONF and INT. We compute it once more for the answers of EMO, but here we consider to be emotional only the items on which all annotators choose a certain level of confidence or intensity. Table 1c displays \u03ba separately for different levels (rows) of CONF and INT. Low values aside, these results inform us that the lower CONF/INT, the more prominent are disagreements. The highest IAA (.39), instead, is achieved for the most confident answers.", "cite_spans": [], "ref_spans": [ { "start": 320, "end": 328, "text": "Table 1c", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Confidence Approximates Disagreements", "sec_num": "3.1" }, { "text": "Post-Processing Disagreements. If systematic differences among annotators can be diagnosed with the help of confidence and intensity, can they also be resolved to some extent? We use the CONF and INT scores as acceptance thresholds for the label emotion, so to post-process the EMO decisions of each judge: they turn into neutral if the corresponding CONF or INT answer does not reach a certain threshold t. For instance, using INT as thres- hold, with t<2 all items labeled emotion in EMO are kept as such only in case the INT is 2 or more, all the others are mapped to neutral. Agreement counts on the post-processed annotation of EMO are in Table 2 . We see, again, that the number of agreed upon items increases by increasing the sets of equal ratings. For instance, 283 sentences received 2 unanimous judgments (column 1 vs. 2, under EMO INT<2), and 417 received 3. In comparison to the original annotation in Table 1b , we can observe a considerable change in the number of items with perfect agreement. While in the raw judgments they were 392, with t<3 they increase to 626. We find a similar pattern when leveraging confidence: with t<2 (low confidence), it is obtained for 359 items, and with t<3 (moderate), perfect agreement increases to 486 items.", "cite_spans": [], "ref_spans": [ { "start": 644, "end": 651, "text": "Table 2", "ref_id": "TABREF3" }, { "start": 915, "end": 923, "text": "Table 1b", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Confidence Approximates Disagreements", "sec_num": "3.1" }, { "text": "This comes at the cost of agreeing on fewer emotion sentences (304 before filtering, 7 and 56 after applying the highest threshold to INT and CONF), but it indicates that the better raters agree on intensity or confidence, the more they agree regarding the presence or absence of emotions.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Confidence Approximates Disagreements", "sec_num": "3.1" }, { "text": "Having found that confidence and intensity have a similar relationship to disagreements, we move to analyzing how they link to one another. To address RQ2, we focus on the ratings of the 304 sentences with the unanimous emotion judgment. For them, we compute the intra-annotator correlation between the answers to INT and the corresponding ratings of CONF. A Spearman's \u03c1 (Spearman, 1904) of .5 for annotator A1, .58 for A2 and .64 for A3 (p-value <.05 for all) reveals a significant positive correlati-", "cite_spans": [ { "start": 372, "end": 388, "text": "(Spearman, 1904)", "ref_id": "BIBREF27" } ], "ref_spans": [], "eq_spans": [], "section": "Stronger Intensity, Higher Confidence", "sec_num": "3.2" }, { "text": "Ann2 Ann3", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ann1", "sec_num": null }, { "text": "Conf 1 Conf 2 Conf 3 Int 2 Int 3 Int 1 Int 2 Int 3 Int 1 Int 2 Int 3 Int 1 Figure 1 : Cross-tabulation of INT and CONF by annotator, for the items that each of them deems emotional.", "cite_spans": [], "ref_spans": [ { "start": 5, "end": 63, "text": "1 Conf 2 Conf 3 Int 2 Int 3 Int 1 Int 2 Int 3 Int 1", "ref_id": "TABREF1" }, { "start": 82, "end": 90, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Ann1", "sec_num": null }, { "text": "on between intensity and confidence. This suggests that people believe they correctly classified a text if they also perceived high emotion intensity. Figure 1 gives an in-depth account of the CONF-INT relation. It plots the counts of items that were labeled with a certain emotion intensity together with a certain confidence level, separately for each annotator. The columns INT3 tell us that rarely annotators perceive intensity as strong without being extremely confident that the text expresses an emotion. In fact, no instance was rated with the highest intensity and the lowest confidence (INT3-CONF1) at the same time. Conversely, for cases of low intensity, annotators tend to stay low also on the scale of confidence.", "cite_spans": [], "ref_spans": [ { "start": 151, "end": 159, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Ann1", "sec_num": null }, { "text": "On what do People Agree? On the 304 sentences considered emotional by all, if the annotators gave the same score to intensity (i.e., perfect agreement is both on EMO and INT), they did not have a total disagreement on confidence (i.e., there are always at least 2 people with the same CONF) and vice versa. This may be a sign of the correlation between the two variables.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ann1", "sec_num": null }, { "text": "Moreover, some items were scored with perfect agreement on all questions. It is the case of sentences like \"I can't believe that you saved my life\", considered, with the highest level of confidence, to convey an emotion of intensity 3, \"Get off my back!\" deemed to have a mild intensity, and \",,You have such an interesting life,\" she said, after a little small talk\" with intensity 1. While these examples are rated as highly certain (CONF3), there are also sentences on which people agree across all confidence-intensity combinations. 6 None of the instances has CONF1 and INT3, while a number of examples have CONF3 and INT1, indicating that it is harder to be uncertain of a strong emotion than to be sure of a weak one.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ann1", "sec_num": null }, { "text": "This annotation investigated if the perceived emotion (class), a perceived feature of emotion (intensity), and self-perception (confidence) are tied together -and can help understand inconsistent annotations. We found that (RQ1) both intensity and confidence account for inter-annotator inconsistencies relative to binary decisions. Adopting confidence as an acceptance threshold showed that higher scores lead to more uniform assessments of emotions; though not a surprising effect of confidence, this also applies to intensity. Moreover, (RQ2) the two variables are correlated, that is, people feel more certain about their emotion recognition performance on items with high intensity.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Conclusion", "sec_num": "4" }, { "text": "We acknowledge that our design of EMO naturally incurs the risk of inconsistent answers. However, precisely for the subjective nature of the task, the finding that disagreements decrease with high CONF/INT is interesting in itself: some judgments which are seemingly unsolvable can be explained by certain perceived properties of emotions (intensity) or self-perceived features (confidence).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Conclusion", "sec_num": "4" }, { "text": "From these results we can draw some lessons. First, the correlation between confidence and intensity brings relevant implications to all those studies that focus on emotional strength. When asked to evaluate intensity, do people confound that with confidence? Even more, is there a causal relation between the two? As a best practice to put safeguards in their guidelines, experimenters may ask people to tease the two variables apart. Potentially, this issue concerns modelling studies as well: do classifiers for emotion intensity predict such feature, or rather the confidence with which people judge emotions? We provided reasons to look into this further.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Conclusion", "sec_num": "4" }, { "text": "Second, confidence turned out to be an important dimension of rating, because it can inform us when the annotators expect to disagree. When judgments diverge, annotators do not deem their intuition credible. Hence, our finding that confidence approximates disagreements means precisely that people themselves predict their performance to differ from that of the others.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Conclusion", "sec_num": "4" }, { "text": "Concretely, all this knowledge can come in support of annotation studies. Including confidence as a rating dimension may give an additional source of information about annotators' reliability. This can help experimenters to refine the guidelines in a pre-testing phase: one might want to disentan-gle cases in which annotators' disagreements are random -signaling a lack of annotators' reliability, and when, instead, they are due to consistently different ways of perceiving and reporting on confidence. In this second case, disagreements may be normalized to an extent, by post-processing the annotation results. For instance, as people seem to agree on the class emotion only if they also agree on certain degrees of CONF, there might be some levels of confidence (or intensity) that one filters in/out of the final annotation labels.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Conclusion", "sec_num": "4" }, { "text": "A similar strategy may be somewhat restrictive, since it accepts as emotional only those items on which humans' intuition fall above a pre-defined threshold. While we observed agreement on such items, it is possible to adopt more nuanced evaluation approaches and integrate information about intensity or confidence into IAA measures. As an example, disagreements between two raters can be penalized more when the one choosing the emotion label does so by perceiving extreme confidence or intensity -even though we provided evidence that these cases are rare. Future work could explore this direction.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Conclusion", "sec_num": "4" }, { "text": "In summary, we uphold that disagreements are not necessarily symptomatic of unreliability. This claim has so far not found much attention in emotion annotations, but is in line with a more general body of research dedicated to the reasons and the patterns underlying annotators' disagreements and to the ways in which their intuitions should be aggregated and evaluated (Bayerl and Paul, 2011; Bhardwaj et al., 2010; Qing et al., 2014; Peldszus and Stede, 2013; Plank et al., 2014; Sommerauer et al., 2020, i.a.) . The applicability of these ideas to emotions should not come as a surprise -their assessment can derive from perceptive and metaperceptual phenomena (intensity and confidence, for instance). Therefore, if emotion judgments alone might not be sufficient to measure the quality of annotations, they can be enriched and, eventually, explained by the knowledge of such phenomena.", "cite_spans": [ { "start": 370, "end": 393, "text": "(Bayerl and Paul, 2011;", "ref_id": null }, { "start": 394, "end": 416, "text": "Bhardwaj et al., 2010;", "ref_id": "BIBREF3" }, { "start": 417, "end": 435, "text": "Qing et al., 2014;", "ref_id": "BIBREF22" }, { "start": 436, "end": 461, "text": "Peldszus and Stede, 2013;", "ref_id": "BIBREF20" }, { "start": 462, "end": 481, "text": "Plank et al., 2014;", "ref_id": "BIBREF21" }, { "start": 482, "end": 512, "text": "Sommerauer et al., 2020, i.a.)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Discussion and Conclusion", "sec_num": "4" }, { "text": "In what follows are the detailed guidelines of the annotation. Q1 corresponds to EMO in the paper, Q2 corresponds to CONF, and Q3 to INT.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Annotation Guidelines", "sec_num": null }, { "text": "In Q1, annotators were required to disregard the emotion that the author of the utterances intended to express or to elicit in others. Their task was to give their immediate, personal impression with respect to the presence of an emotion.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Annotation Guidelines", "sec_num": null }, { "text": "Before annotating the 700 sentences selected from COCA, we observe inter-annotator agreement on a pre-annotation trial. With 70 sentences, Cohen's \u03ba for pairs of annotators was found to be satisfactory (.52, .6 and .43) and motivated us to complete the job on the rest of the sentences. The job was completed upon a compensation of 60e.", "cite_spans": [ { "start": 202, "end": 219, "text": "(.52, .6 and .43)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "A Annotation Guidelines", "sec_num": null }, { "text": "In this annotation trial, you will assess if texts are emotional or neutral.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Task Description Approx Duration: 3 hours", "sec_num": null }, { "text": "Neutral sentences are those which 1. bear no affective connotation. A war started in Westeros. The king found an old sausage under his bed.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Task Description Approx Duration: 3 hours", "sec_num": null }, { "text": "I am so happy to see you. She was bursting with arrogance. And there she was, desperate for her family. I couldn't stand the catering food, bleark!", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Examples of 3. are:", "sec_num": null }, { "text": "You will be shown individual sentences and, for each of them, you will answer 3 questions (Q1, Q2, Q3). Go for your immediate reaction to the text -avoid over-assessments.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "Q1 Given a sentence, please ask yourself: is it emotional (E) or neutral (N)? Type:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 N, if the text does not convey any emotion, like in Examples of 1.;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 E, if an emotion could be inferred from the text, like in Examples of 2., or an emotion is a central part of the text, like in Examples of 3.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "Q2 Ask yourself: how confident am I about my answer to Q1? Give yourself a rating on a scale from 1 to 3. Indicate if you are", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 1, not confident at all;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 2, confident;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 3, sure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "Q3 This question only applies if you answered 2 or 3 in Q1: in case the sentence expresses an emotion, how strong is such emotion? Assess the degree of its intensity on a scale from 1 to 3, where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 1 is mild;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 2 is intense;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "\u2022 3 is very intense.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.2 Guidelines", "sec_num": null }, { "text": "After data pre-processing step (i.e. exclusion of academic texts and sentences containing words that are masked for copyright reasons), we use a binary classifier as a guide for the selection of sentences from COCA. Using HugginFace 7 , we fine-tune the pre-trained BERT base-case model on data for emotion recognition, adding a classification layer that outputs the labels emotion or neutral. Data are the resources by Liu et al. (2007) , Troiano et al. (2019) , Scherer and Wallbott (1997), Alm et al. (2005) , Li et al. (2017) , Ghazi et al. (2015) , Mohammad and Bravo-Marquez (2017) , (Mohammad, 2012) and Schuff et al. (2017) . We make their format homogeneous with the tool made available by Bostan and Klinger (2018) ; next, as the labels in the resulting unified emotion data (UED) are not binary, we map neutral and no emotion instances into neutral, and the rest into emotion. The total 136891 sentences are then split into train (70%), validation (10%) and test (20%) sets. Classifier's performance is in Table 3 .", "cite_spans": [ { "start": 420, "end": 437, "text": "Liu et al. (2007)", "ref_id": "BIBREF16" }, { "start": 440, "end": 461, "text": "Troiano et al. (2019)", "ref_id": "BIBREF30" }, { "start": 464, "end": 475, "text": "Scherer and", "ref_id": "BIBREF24" }, { "start": 476, "end": 510, "text": "Wallbott (1997), Alm et al. (2005)", "ref_id": null }, { "start": 513, "end": 529, "text": "Li et al. (2017)", "ref_id": "BIBREF15" }, { "start": 532, "end": 551, "text": "Ghazi et al. (2015)", "ref_id": "BIBREF12" }, { "start": 554, "end": 587, "text": "Mohammad and Bravo-Marquez (2017)", "ref_id": "BIBREF19" }, { "start": 590, "end": 606, "text": "(Mohammad, 2012)", "ref_id": "BIBREF17" }, { "start": 611, "end": 631, "text": "Schuff et al. (2017)", "ref_id": "BIBREF25" }, { "start": 699, "end": 724, "text": "Bostan and Klinger (2018)", "ref_id": "BIBREF5" } ], "ref_spans": [ { "start": 1017, "end": 1024, "text": "Table 3", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "B Binary Classification", "sec_num": null }, { "text": "C Further Analysis C.1 Disagreements Table 4 reports the distribution of the scores of confidence and intensity for the items where the annotators disagree. This is observed on annotator pairs. A row considers all those items on which either annotators (on the columns) chooses the emotion label and the other selects the neutral one. For instance, A1-A2 disagree in total 201 times: on 24 sentences, A1 makes the emotion choice, and on 177 sentences it is A2 who picks the class emotion. Out of the 24 items, A1 rated 23 as having low intensity and 1 as medium intensity; out of the 177 sentences, 149 are considered of low intensity by A2, 27 as mild, and only 1 as highly intense.", "cite_spans": [], "ref_spans": [ { "start": 37, "end": 44, "text": "Table 4", "ref_id": null } ], "eq_spans": [], "section": "B Binary Classification", "sec_num": null }, { "text": "What emerges overall is that people rarely disagree when the emotion-leaning annotator has extreme confidence, or perceives very high intensity.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "B Binary Classification", "sec_num": null }, { "text": "A manual analysis of the annotations reveals that perfect agreement occurs in the presence of certain patterns. Items unanimously considered emotional often report personal impressions about state of affairs or the speakers' interlocutors (e.g., \"Paris is so sexy\", \"Your expression changed from excited puppy to crestfallen\"), and mostly involve first-hand experiences of the speakers themselves (e.g., \"We'll miss you, but we'll be watching\", \"I'm afraid I don't see anything very beautiful right now\", \"Others helped me and it made a huge difference\"). Instead, sentences that received 3 neutral labels seem to be centered on factual statements, like \"Furthermore, the types of materials of manufacture are different\", \"They continue walking\".", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "C.2 Agreements", "sec_num": null }, { "text": "One difference between the emotion and neutral labels is the frequency of agreement, as we found that people concur more on the former -and this invalidates our expectation that, not being given a varied set of affective categories, and not identifying what emotion they are judging, people would tend to resort to the neutral choice. Moreover, annotators converge more on one or more on the other label depending on the genre of a text: looking at the distribution of the 304 unanimous emotions (magazine: 28 sentences, blogs: 44, news: 27, tv: 67, fiction: 54, spoken: 39, web: 45) and the 88 neutrals (magazine: 18 sentences, blogs: 12, news: 22, tv: 4, fiction: 5, spoken: 12, web: 15), we see that people recognize that affect often manifests itself in fictions, for instance, but is rarer in newsthe opposite holds for the neutral expressions. An obvious strategy to recognize emotions would be to find an emotion name in text. But this is not the case. Sentences that contain emotion words considered less emotionally intense than others: the majority of sentences with CONF3-INT2 contain emotion words (e.g.,\"I was sad to leave.\", while those with CONF3-INT3 are related to extremely negative states of mind (e.g.,\" [...] if I could die and bring her back , I would , but I can't , and I have to deal with that now\").", "cite_spans": [ { "start": 1224, "end": 1229, "text": "[...]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "C.2 Agreements", "sec_num": null }, { "text": "In Table 5 , we report some example sentences on which annotators reached perfect agreement across all confidence-intensity combinations, having choosen the label emotion. The sentences are extracted from a number of genres in COCA, and are associated to different scores of intensity (INT) and confidence (CONF). Note that there is no instance that elicited a high intensity evaluation (3) and a low confidence (either 2 or 3) in the annotators. Instances of CONF3 and INT1 show that the correlation between confidence and intensity, though intuitive, has counterexamples.", "cite_spans": [], "ref_spans": [ { "start": 3, "end": 10, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "C.2 Agreements", "sec_num": null }, { "text": "The guidelines are in the Appendix. Our data is accessible at https://www.ims.uni-stuttgart.de/ data/emotion-confidence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "https://www.english-corpora.org/coca/ 3 Details and classifier's performance in Appendix.4 With skewed class distribution, chance agreement increases, penalizing the resulting \u03ba (Cicchetti and Feinstein, 1990).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Ratings on disagreements are in Appendix,Table 4.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "A detailed analysis is inTable 5, Appendix.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This work was supported by Deutsche Forschungsgemeinschaft (project CEAT, KL 2869/1-2) and the Leibniz WissenschaftsCampus T\u00fcbingen \"Cognitive Interfaces\". We thank Laura Oberl\u00e4nder, Agnieszka Fale\u0144ska and the anonymous reviewers for their constructive feedback.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": null }, { "text": "A2 A3 CONF INT CONF INT CONF INT 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 A1-A2 15 8 1 23 1 0 54 95 28 149 27 1 ------A2-A3 ------13 25 6 37 7 0 56 61 3 94 26 0 A3-A1 5 6 0 10 1 0 ------95 128 17 169 70 1 Magazine 1 2 Nature always solves her own problems ; and we can go far toward solving our own if we will listen to her teachings and consort with those who love her. Web 3 3 I will completely destroy them and make them an object of horror and scorn , and an everlasting ruin. Spoken 3 3 \"We 're very worried .\"", "cite_spans": [], "ref_spans": [ { "start": 3, "end": 80, "text": "A3 CONF INT CONF INT CONF INT 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3", "ref_id": null } ], "eq_spans": [], "section": "A1", "sec_num": null }, { "text": "If -if I could die and bring her back , I would , but I can't , and I have to deal with that now. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Spoken 3 3", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Emotions from text: Machine learning for text-based emotion prediction", "authors": [ { "first": "Cecilia", "middle": [], "last": "Ovesdotter Alm", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Roth", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Sproat", "suffix": "" } ], "year": 2005, "venue": "Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "579--586", "other_ids": {}, "num": null, "urls": [], "raw_text": "Cecilia Ovesdotter Alm, Dan Roth, and Richard Sproat. 2005. Emotions from text: Machine learning for text-based emotion prediction. In Proceedings of Human Language Technology Conference and Con- ference on Empirical Methods in Natural Language Processing, pages 579-586, Vancouver, British Co- lumbia, Canada. Association for Computational Lin- guistics.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "What determines inter-coder agreement in manual annotations? a meta-analytic investigation", "authors": [], "year": 2011, "venue": "Computational Linguistics", "volume": "37", "issue": "4", "pages": "699--725", "other_ids": { "DOI": [ "10.1162/COLI_a_00074" ] }, "num": null, "urls": [], "raw_text": "Petra Saskia Bayerl and Karsten Ingmar Paul. 2011. What determines inter-coder agreement in manual annotations? a meta-analytic investigation. Compu- tational Linguistics, 37(4):699-725.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Confidence of emotion expression recognition recruits brain regions outside the face perception network", "authors": [ { "first": "Indrit", "middle": [], "last": "B\u00e8gue", "suffix": "" }, { "first": "Maarten", "middle": [], "last": "Vaessen", "suffix": "" }, { "first": "Jeremy", "middle": [], "last": "Hofmeister", "suffix": "" }, { "first": "Marice", "middle": [], "last": "Pereira", "suffix": "" } ], "year": 2018, "venue": "Social Cognitive and Affective Neuroscience", "volume": "14", "issue": "1", "pages": "81--95", "other_ids": { "DOI": [ "10.1093/scan/nsy102" ] }, "num": null, "urls": [], "raw_text": "Indrit B\u00e8gue, Maarten Vaessen, Jeremy Hofmeister, Marice Pereira, Sophie Schwartz, and Patrik Vuil- leumier. 2018. Confidence of emotion expression recognition recruits brain regions outside the face perception network. Social Cognitive and Affective Neuroscience, 14(1):81-95.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Anveshan: A framework for analysis of multiple annotators' labeling behavior", "authors": [ { "first": "Vikas", "middle": [], "last": "Bhardwaj", "suffix": "" }, { "first": "Rebecca", "middle": [], "last": "Passonneau", "suffix": "" }, { "first": "Ansaf", "middle": [], "last": "Salleb-Aouissi", "suffix": "" }, { "first": "Nancy", "middle": [], "last": "Ide", "suffix": "" } ], "year": 2010, "venue": "Proceedings of the Fourth Linguistic Annotation Workshop", "volume": "", "issue": "", "pages": "47--55", "other_ids": {}, "num": null, "urls": [], "raw_text": "Vikas Bhardwaj, Rebecca Passonneau, Ansaf Salleb- Aouissi, and Nancy Ide. 2010. Anveshan: A frame- work for analysis of multiple annotators' labeling be- havior. In Proceedings of the Fourth Linguistic An- notation Workshop, pages 47-55, Uppsala, Sweden. Association for Computational Linguistics.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Interannotator agreement in sentiment analysis: Machine learning perspective", "authors": [ { "first": "Victoria", "middle": [], "last": "Bobicev", "suffix": "" }, { "first": "Marina", "middle": [], "last": "Sokolova", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the International Conference Recent Advances in Natural Language Processing", "volume": "", "issue": "", "pages": "97--102", "other_ids": { "DOI": [ "10.26615/978-954-452-049-6_015" ] }, "num": null, "urls": [], "raw_text": "Victoria Bobicev and Marina Sokolova. 2017. Inter- annotator agreement in sentiment analysis: Machine learning perspective. In Proceedings of the Interna- tional Conference Recent Advances in Natural Lan- guage Processing, RANLP 2017, pages 97-102, Var- na, Bulgaria. INCOMA Ltd.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "An analysis of annotated corpora for emotion classification in text", "authors": [ { "first": "Laura-Ana-Maria", "middle": [], "last": "Bostan", "suffix": "" }, { "first": "Roman", "middle": [], "last": "Klinger", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 27th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "2104--2119", "other_ids": {}, "num": null, "urls": [], "raw_text": "Laura-Ana-Maria Bostan and Roman Klinger. 2018. An analysis of annotated corpora for emotion clas- sification in text. In Proceedings of the 27th Inter- national Conference on Computational Linguistics, pages 2104-2119, Santa Fe, New Mexico, USA. As- sociation for Computational Linguistics.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "High agreement but low kappa: Ii. resolving the paradoxes", "authors": [ { "first": "V", "middle": [], "last": "Domenic", "suffix": "" }, { "first": "", "middle": [], "last": "Cicchetti", "suffix": "" }, { "first": "", "middle": [], "last": "Alvan R Feinstein", "suffix": "" } ], "year": 1990, "venue": "Journal of clinical epidemiology", "volume": "43", "issue": "6", "pages": "551--558", "other_ids": {}, "num": null, "urls": [], "raw_text": "Domenic V Cicchetti and Alvan R Feinstein. 1990. High agreement but low kappa: Ii. resolving the para- doxes. Journal of clinical epidemiology, 43(6):551- 558.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "A coefficient of agreement for nominal scales. Educational and Psychological Measurement", "authors": [ { "first": "Jacob", "middle": [], "last": "Cohen", "suffix": "" } ], "year": 1960, "venue": "", "volume": "20", "issue": "", "pages": "37--46", "other_ids": { "DOI": [ "10.1177/001316446002000104" ] }, "num": null, "urls": [], "raw_text": "Jacob Cohen. 1960. A coefficient of agreement for no- minal scales. Educational and Psychological Mea- surement, 20(1):37-46.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Corpus of Contemporary American English (COCA)", "authors": [ { "first": "Mark", "middle": [], "last": "Davies", "suffix": "" } ], "year": 2015, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "DOI": [ "10.7910/DVN/AMUDUW" ] }, "num": null, "urls": [], "raw_text": "Mark Davies. 2015. Corpus of Contemporary Ameri- can English (COCA).", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "BERT: Pre-training of deep bidirectional transformers for language understanding", "authors": [ { "first": "Jacob", "middle": [], "last": "Devlin", "suffix": "" }, { "first": "Ming-Wei", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Kenton", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Kristina", "middle": [], "last": "Toutanova", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "4171--4186", "other_ids": { "DOI": [ "10.18653/v1/N19-1423" ] }, "num": null, "urls": [], "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kri- stina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understan- ding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers), pa- ges 4171-4186, Minneapolis, Minnesota. Associati- on for Computational Linguistics.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Measuring nominal scale agreement among many raters", "authors": [ { "first": "Joseph", "middle": [ "L" ], "last": "Fleiss", "suffix": "" } ], "year": 1971, "venue": "Psychological bulletin", "volume": "76", "issue": "5", "pages": "378--382", "other_ids": { "DOI": [ "10.1037/h0031619" ] }, "num": null, "urls": [], "raw_text": "Joseph L. Fleiss. 1971. Measuring nominal scale agree- ment among many raters. Psychological bulletin, 76(5):378-382.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Reconstructing the past: A century of ideas about emotion in psychology", "authors": [ { "first": "Maria", "middle": [], "last": "Gendron", "suffix": "" }, { "first": "Lisa", "middle": [ "Feldman" ], "last": "Barrett", "suffix": "" } ], "year": 2009, "venue": "Emotion Review", "volume": "1", "issue": "4", "pages": "316--339", "other_ids": { "DOI": [ "10.1177/1754073909338877" ] }, "num": null, "urls": [], "raw_text": "Maria Gendron and Lisa Feldman Barrett. 2009. Re- constructing the past: A century of ideas about emo- tion in psychology. Emotion Review, 1(4):316-339.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Detecting emotion stimuli in emotion-bearing sentences", "authors": [ { "first": "Diman", "middle": [], "last": "Ghazi", "suffix": "" }, { "first": "Diana", "middle": [], "last": "Inkpen", "suffix": "" }, { "first": "Stan", "middle": [], "last": "Szpakowicz", "suffix": "" } ], "year": 2015, "venue": "Computational Linguistics and Intelligent Text Processing", "volume": "", "issue": "", "pages": "152--165", "other_ids": { "DOI": [ "https://link.springer.com/chapter/10.1007/978-3-319-18117-2_12" ] }, "num": null, "urls": [], "raw_text": "Diman Ghazi, Diana Inkpen, and Stan Szpakowicz. 2015. Detecting emotion stimuli in emotion-bearing sentences. In Computational Linguistics and Intelli- gent Text Processing, pages 152-165. Springer Inter- national Publishing.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Appraisal theories for emotion classification in text", "authors": [ { "first": "Jan", "middle": [], "last": "Hofmann", "suffix": "" }, { "first": "Enrica", "middle": [], "last": "Troiano", "suffix": "" }, { "first": "Kai", "middle": [], "last": "Sassenberg", "suffix": "" }, { "first": "Roman", "middle": [], "last": "Klinger", "suffix": "" } ], "year": 2020, "venue": "Proceedings of the 28th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "125--138", "other_ids": { "DOI": [ "10.18653/v1/2020.coling-main.11" ] }, "num": null, "urls": [], "raw_text": "Jan Hofmann, Enrica Troiano, Kai Sassenberg, and Ro- man Klinger. 2020. Appraisal theories for emoti- on classification in text. In Proceedings of the 28th International Conference on Computational Lingui- stics, pages 125-138, Barcelona, Spain (Online). International Committee on Computational Lingui- stics.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Emotion recognition and confidence ratings predicted by vocal stimulus type and prosodic parameters", "authors": [ { "first": "Adi", "middle": [], "last": "Lausen", "suffix": "" }, { "first": "Kurt", "middle": [], "last": "Hammerschmidt", "suffix": "" } ], "year": 2020, "venue": "Humanities and Social Sciences Communications", "volume": "7", "issue": "1", "pages": "1--17", "other_ids": {}, "num": null, "urls": [], "raw_text": "Adi Lausen and Kurt Hammerschmidt. 2020. Emotion recognition and confidence ratings predicted by vo- cal stimulus type and prosodic parameters. Huma- nities and Social Sciences Communications, 7(1):1- 17.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "DailyDialog: A manually labelled multi-turn dialogue dataset", "authors": [ { "first": "Yanran", "middle": [], "last": "Li", "suffix": "" }, { "first": "Hui", "middle": [], "last": "Su", "suffix": "" }, { "first": "Xiaoyu", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Wenjie", "middle": [], "last": "Li", "suffix": "" }, { "first": "Ziqiang", "middle": [], "last": "Cao", "suffix": "" }, { "first": "Shuzi", "middle": [], "last": "Niu", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the Eighth International Joint Conference on Natural Language Processing", "volume": "1", "issue": "", "pages": "986--995", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. DailyDialog: A manu- ally labelled multi-turn dialogue dataset. In Procee- dings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986-995, Taipei, Taiwan. Asian Fe- deration of Natural Language Processing.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Grounded emotions", "authors": [ { "first": "Vicki", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Carmen", "middle": [], "last": "Banea", "suffix": "" }, { "first": "Rada", "middle": [], "last": "Mihalcea", "suffix": "" } ], "year": 2007, "venue": "Seventh International Conference on Affective Computing and Intelligent Interaction (ACII 2017)", "volume": "", "issue": "", "pages": "477--483", "other_ids": { "DOI": [ "https://doi.ieeecomputersociety.org/10.1109/ACII.2017.8273642" ] }, "num": null, "urls": [], "raw_text": "Vicki Liu, Carmen Banea, and Rada Mihalcea. 2007. Grounded emotions. In Seventh International Con- ference on Affective Computing and Intelligent Inter- action (ACII 2017), pages 477-483. IEEE Computer Society.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "#emotional tweets", "authors": [ { "first": "Saif", "middle": [], "last": "Mohammad", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the main conference and the shared task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation", "volume": "1", "issue": "", "pages": "246--255", "other_ids": {}, "num": null, "urls": [], "raw_text": "Saif Mohammad. 2012. #emotional tweets. In *SEM 2012: The First Joint Conference on Lexical and Computational Semantics -Volume 1: Proceedings of the main conference and the shared task, and Vo- lume 2: Proceedings of the Sixth International Work- shop on Semantic Evaluation (SemEval 2012), pages 246-255, Montr\u00e9al, Canada. Association for Com- putational Linguistics.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Word affect intensities", "authors": [ { "first": "Saif", "middle": [], "last": "Mohammad", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Saif Mohammad. 2018. Word affect intensities. In Pro- ceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Re- sources Association (ELRA).", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "WASSA-2017 shared task on emotion intensity", "authors": [ { "first": "Saif", "middle": [], "last": "Mohammad", "suffix": "" }, { "first": "Felipe", "middle": [], "last": "Bravo-Marquez", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis", "volume": "", "issue": "", "pages": "34--49", "other_ids": { "DOI": [ "10.18653/v1/W17-5205" ] }, "num": null, "urls": [], "raw_text": "Saif Mohammad and Felipe Bravo-Marquez. 2017. WASSA-2017 shared task on emotion intensity. In Proceedings of the 8th Workshop on Computatio- nal Approaches to Subjectivity, Sentiment and Soci- al Media Analysis, pages 34-49, Copenhagen, Den- mark. Association for Computational Linguistics.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Ranking the annotators: An agreement study on argumentation structure", "authors": [ { "first": "Andreas", "middle": [], "last": "Peldszus", "suffix": "" }, { "first": "Manfred", "middle": [], "last": "Stede", "suffix": "" } ], "year": 2013, "venue": "Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse", "volume": "", "issue": "", "pages": "196--204", "other_ids": {}, "num": null, "urls": [], "raw_text": "Andreas Peldszus and Manfred Stede. 2013. Ranking the annotators: An agreement study on argumenta- tion structure. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Dis- course, pages 196-204, Sofia, Bulgaria. Association for Computational Linguistics.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Linguistically debatable or just plain wrong?", "authors": [ { "first": "Barbara", "middle": [], "last": "Plank", "suffix": "" }, { "first": "Dirk", "middle": [], "last": "Hovy", "suffix": "" }, { "first": "Anders", "middle": [], "last": "S\u00f8gaard", "suffix": "" } ], "year": 2014, "venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics", "volume": "2", "issue": "", "pages": "507--511", "other_ids": { "DOI": [ "10.3115/v1/P14-2083" ] }, "num": null, "urls": [], "raw_text": "Barbara Plank, Dirk Hovy, and Anders S\u00f8gaard. 2014. Linguistically debatable or just plain wrong? In Pro- ceedings of the 52nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Papers), pages 507-511, Baltimore, Maryland. As- sociation for Computational Linguistics.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Empirical analysis of aggregation methods for collective annotation", "authors": [ { "first": "Ciyang", "middle": [], "last": "Qing", "suffix": "" }, { "first": "Ulle", "middle": [], "last": "Endriss", "suffix": "" }, { "first": "Raquel", "middle": [], "last": "Fern\u00e1ndez", "suffix": "" }, { "first": "Justin", "middle": [], "last": "Kruger", "suffix": "" } ], "year": 2014, "venue": "Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1533-1542", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ciyang Qing, Ulle Endriss, Raquel Fern\u00e1ndez, and Ju- stin Kruger. 2014. Empirical analysis of aggregation methods for collective annotation. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pa- ges 1533-1542, Dublin, Ireland. Dublin City Univer- sity and Association for Computational Linguistics.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Mind-reading ability: Beliefs and performance", "authors": [ { "first": "Anu", "middle": [], "last": "Realo", "suffix": "" }, { "first": "J\u00fcri", "middle": [], "last": "Allik", "suffix": "" }, { "first": "Aire", "middle": [], "last": "N\u00f5lvak", "suffix": "" }, { "first": "Raivo", "middle": [], "last": "Valk", "suffix": "" }, { "first": "Tuuli", "middle": [], "last": "Ruus", "suffix": "" }, { "first": "Monika", "middle": [], "last": "Schmidt", "suffix": "" }, { "first": "Tiina", "middle": [], "last": "Eilola", "suffix": "" } ], "year": 2003, "venue": "Journal of Research in Personality", "volume": "37", "issue": "5", "pages": "420--445", "other_ids": {}, "num": null, "urls": [], "raw_text": "Anu Realo, J\u00fcri Allik, Aire N\u00f5lvak, Raivo Valk, Tuuli Ruus, Monika Schmidt, and Tiina Eilola. 2003. Mind-reading ability: Beliefs and performan- ce. Journal of Research in Personality, 37(5):420- 445.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "The ISEAR questionnaire and codebook", "authors": [ { "first": "Klaus", "middle": [ "R" ], "last": "Scherer", "suffix": "" }, { "first": "Harald", "middle": [ "G" ], "last": "Wallbott", "suffix": "" } ], "year": 1997, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Klaus R. Scherer and Harald G. Wallbott. 1997. The ISEAR questionnaire and co- debook. Geneva Emotion Research Group.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Annotation, modelling and analysis of fine-grained emotions on a stance and sentiment detection corpus", "authors": [ { "first": "Hendrik", "middle": [], "last": "Schuff", "suffix": "" }, { "first": "Jeremy", "middle": [], "last": "Barnes", "suffix": "" }, { "first": "Julian", "middle": [], "last": "Mohme", "suffix": "" }, { "first": "Sebastian", "middle": [], "last": "Pad\u00f3", "suffix": "" }, { "first": "Roman", "middle": [], "last": "Klinger", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis", "volume": "", "issue": "", "pages": "13--23", "other_ids": { "DOI": [ "10.18653/v1/W17-5203" ] }, "num": null, "urls": [], "raw_text": "Hendrik Schuff, Jeremy Barnes, Julian Mohme, Seba- stian Pad\u00f3, and Roman Klinger. 2017. Annotation, modelling and analysis of fine-grained emotions on a stance and sentiment detection corpus. In Pro- ceedings of the 8th Workshop on Computational Ap- proaches to Subjectivity, Sentiment and Social Me- dia Analysis, pages 13-23, Copenhagen, Denmark. Association for Computational Linguistics.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "Would you describe a leopard as yellow? evaluating crowd-annotations with justified and informative disagreement", "authors": [ { "first": "Pia", "middle": [], "last": "Sommerauer", "suffix": "" }, { "first": "Antske", "middle": [], "last": "Fokkens", "suffix": "" }, { "first": "Piek", "middle": [], "last": "Vossen", "suffix": "" } ], "year": 2020, "venue": "Proceedings of the 28th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "4798--4809", "other_ids": { "DOI": [ "10.18653/v1/2020.coling-main.422" ] }, "num": null, "urls": [], "raw_text": "Pia Sommerauer, Antske Fokkens, and Piek Vossen. 2020. Would you describe a leopard as yellow? evaluating crowd-annotations with justified and in- formative disagreement. In Proceedings of the 28th International Conference on Computational Lingui- stics, pages 4798-4809, Barcelona, Spain (Online). International Committee on Computational Lingui- stics.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "The proof and measurement of association between two things", "authors": [ { "first": "Charles", "middle": [], "last": "Spearman", "suffix": "" } ], "year": 1904, "venue": "The American Journal of Psychology", "volume": "15", "issue": "1", "pages": "71--101", "other_ids": { "DOI": [ "10.2307/1412159" ] }, "num": null, "urls": [], "raw_text": "Charles Spearman. 1904. The proof and measurement of association between two things. The American Journal of Psychology, 15(1):71-101.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "SemEval-2007 task 14: Affective text", "authors": [ { "first": "Carlo", "middle": [], "last": "Strapparava", "suffix": "" }, { "first": "Rada", "middle": [], "last": "Mihalcea", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007)", "volume": "", "issue": "", "pages": "70--74", "other_ids": {}, "num": null, "urls": [], "raw_text": "Carlo Strapparava and Rada Mihalcea. 2007. SemEval- 2007 task 14: Affective text. In Proceedings of the Fourth International Workshop on Semantic Evalua- tions (SemEval-2007), pages 70-74, Prague, Czech Republic. Association for Computational Lingui- stics.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "Learning to identify emotions in text", "authors": [ { "first": "Carlo", "middle": [], "last": "Strapparava", "suffix": "" }, { "first": "Rada", "middle": [], "last": "Mihalcea", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 2008 ACM symposium on Applied computing", "volume": "", "issue": "", "pages": "1556--1560", "other_ids": { "DOI": [ "https://dl.acm.org/doi/pdf/10.1145/1363686.1364052" ] }, "num": null, "urls": [], "raw_text": "Carlo Strapparava and Rada Mihalcea. 2008. Learning to identify emotions in text. In Proceedings of the 2008 ACM symposium on Applied computing, pages 1556-1560.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "Crowdsourcing and validating event-focused emotion corpora for German and English", "authors": [ { "first": "Enrica", "middle": [], "last": "Troiano", "suffix": "" }, { "first": "Sebastian", "middle": [], "last": "Pad\u00f3", "suffix": "" }, { "first": "Roman", "middle": [], "last": "Klinger", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "4005--4011", "other_ids": { "DOI": [ "10.18653/v1/P19-1391" ] }, "num": null, "urls": [], "raw_text": "Enrica Troiano, Sebastian Pad\u00f3, and Roman Klinger. 2019. Crowdsourcing and validating event-focused emotion corpora for German and English. In Procee- dings of the 57th Annual Meeting of the Associati- on for Computational Linguistics, pages 4005-4011, Florence, Italy. Association for Computational Lin- guistics.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "Text emotion distribution learning via multi-task convolutional neural network", "authors": [ { "first": "Yuxiang", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Jiamei", "middle": [], "last": "Fu", "suffix": "" }, { "first": "Dongyu", "middle": [], "last": "She", "suffix": "" }, { "first": "Ying", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Senzhang", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Jufeng", "middle": [], "last": "Yang", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18", "volume": "", "issue": "", "pages": "4595--4601", "other_ids": { "DOI": [ "10.24963/ijcai.2018/639" ] }, "num": null, "urls": [], "raw_text": "Yuxiang Zhang, Jiamei Fu, Dongyu She, Ying Zhang, Senzhang Wang, and Jufeng Yang. 2018. Text emo- tion distribution learning via multi-task convolutio- nal neural network. In Proceedings of the Twenty- Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 4595-4601. Internatio- nal Joint Conferences on Artificial Intelligence Orga- nization.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "Knowledge-enriched transformer for emotion detection in textual conversations", "authors": [ { "first": "Peixiang", "middle": [], "last": "Zhong", "suffix": "" }, { "first": "Di", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Chunyan", "middle": [], "last": "Miao", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)", "volume": "", "issue": "", "pages": "165--176", "other_ids": { "DOI": [ "10.18653/v1/D19-1016" ] }, "num": null, "urls": [], "raw_text": "Peixiang Zhong, Di Wang, and Chunyan Miao. 2019. Knowledge-enriched transformer for emotion detec- tion in textual conversations. In Proceedings of the 2019 Conference on Empirical Methods in Na- tural Language Processing and the 9th Internatio- nal Joint Conference on Natural Language Pro- cessing (EMNLP-IJCNLP), pages 165-176, Hong Kong, China. Association for Computational Lingui- stics.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "text": "sentences are those which either 2. describe an event, a concept or state of affairs to which you would associate an emotion; 3. have an emotion as a central component of their meaning. Examples of 1. are: I am wearing my mask. She answered her phone. A new deal was established between the parties. The elections are over. Examples of 2. are: I saw my bestfriend. She was being pretty arrogant to me.", "num": null, "uris": null }, "TABREF1": { "content": "", "type_str": "table", "text": "", "html": null, "num": null }, "TABREF3": { "content": "
", "type_str": "table", "text": "Counts of labels for subsets of ratings on EMO, post-processed with acceptance thresholds <2 and <3, for both CONF (top) and INT (bottom).", "html": null, "num": null }, "TABREF5": { "content": "
", "type_str": "table", "text": "Binary classification on UED test set.", "html": null, "num": null } } } }