Base_JSON/prefixK/json/konvens/2021.konvens-1.5.json

{
    "paper_id": "2021",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:11:52.115970Z"
    },
    "title": "Emotion Recognition under Consideration of the Emotion Component Process Model",
    "authors": [
        {
            "first": "Felix",
            "middle": [],
            "last": "Casel",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Stuttgart",
                "location": {
                    "addrLine": "Pfaffenwaldring 5b",
                    "postCode": "70569",
                    "settlement": "Stuttgart",
                    "country": "Germany"
                }
            },
            "email": ""
        },
        {
            "first": "Amelie",
            "middle": [],
            "last": "Heindl",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Stuttgart",
                "location": {
                    "addrLine": "Pfaffenwaldring 5b",
                    "postCode": "70569",
                    "settlement": "Stuttgart",
                    "country": "Germany"
                }
            },
            "email": ""
        },
        {
            "first": "Roman",
            "middle": [],
            "last": "Klinger",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Stuttgart",
                "location": {
                    "addrLine": "Pfaffenwaldring 5b",
                    "postCode": "70569",
                    "settlement": "Stuttgart",
                    "country": "Germany"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Emotion classification in text is typically performed with neural network models which learn to associate linguistic units with emotions. While this often leads to good predictive performance, it does only help to a limited degree to understand how emotions are communicated in various domains. The emotion component process model (CPM) by Scherer (2005) is an interesting approach to explain emotion communication. It states that emotions are a coordinated process of various subcomponents, in reaction to an event, namely the subjective feeling, the cognitive appraisal, the expression, a physiological bodily reaction, and a motivational action tendency. We hypothesize that these components are associated with linguistic realizations: an emotion can be expressed by describing a physiological bodily reaction (\"he was trembling\"), or the expression (\"she smiled\"), etc. We annotate existing literature and Twitter emotion corpora with emotion component classes and find that emotions on Twitter are predominantly expressed by event descriptions or subjective reports of the feeling, while in literature, authors prefer to describe what characters do, and leave the interpretation to the reader. We further include the CPM in a multitask learning model and find that this supports the emotion categorization. The annotated corpora are available at https://www. ims.uni-stuttgart.de/data/emotion.",
    "pdf_parse": {
        "paper_id": "2021",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Emotion classification in text is typically performed with neural network models which learn to associate linguistic units with emotions. While this often leads to good predictive performance, it does only help to a limited degree to understand how emotions are communicated in various domains. The emotion component process model (CPM) by Scherer (2005) is an interesting approach to explain emotion communication. It states that emotions are a coordinated process of various subcomponents, in reaction to an event, namely the subjective feeling, the cognitive appraisal, the expression, a physiological bodily reaction, and a motivational action tendency. We hypothesize that these components are associated with linguistic realizations: an emotion can be expressed by describing a physiological bodily reaction (\"he was trembling\"), or the expression (\"she smiled\"), etc. We annotate existing literature and Twitter emotion corpora with emotion component classes and find that emotions on Twitter are predominantly expressed by event descriptions or subjective reports of the feeling, while in literature, authors prefer to describe what characters do, and leave the interpretation to the reader. We further include the CPM in a multitask learning model and find that this supports the emotion categorization. The annotated corpora are available at https://www. ims.uni-stuttgart.de/data/emotion.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "The task of emotion classification from written text is to map textual units, like documents, paragraphs, or sentences, to a predefined set of emotions. Common class inventories rely on psychological theories such as those proposed by Ekman (1992) (anger, disgust, fear, joy, sadness, surprise) or Plutchik (2001) . Often, emotion classification is tackled as an end-to-end learning task, potentially informed by lexical resources (see the SemEval Shared Task 1 on Affect in Tweets for an overview of recent approaches (Mohammad et al., 2018) ).",
                "cite_spans": [
                    {
                        "start": 235,
                        "end": 247,
                        "text": "Ekman (1992)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 248,
                        "end": 294,
                        "text": "(anger, disgust, fear, joy, sadness, surprise)",
                        "ref_id": null
                    },
                    {
                        "start": 298,
                        "end": 313,
                        "text": "Plutchik (2001)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 519,
                        "end": 542,
                        "text": "(Mohammad et al., 2018)",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "While end-to-end learning and fine-tuning of pretrained models for classification have shown great performance improvements in contrast to purely feature-based methods, such approaches typically neglect the existing knowledge about emotions in psychology (which might help in classification and to better understand how emotions are communicated). There are only very few approaches that aim at combining psychological theories (beyond basic emotion categories) with emotion classification models: We are only aware of the work by Hofmann et al. (2020) , who incorporate the cognitive appraisal of events, and Buechel et al. (2020) , who jointly learn affect (valence, arousal) and emotion classes; next to knowledge-base-oriented modelling of events by Balahur et al. (2012) and Cambria et al. (2014) .",
                "cite_spans": [
                    {
                        "start": 531,
                        "end": 552,
                        "text": "Hofmann et al. (2020)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 610,
                        "end": 631,
                        "text": "Buechel et al. (2020)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 754,
                        "end": 775,
                        "text": "Balahur et al. (2012)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 780,
                        "end": 801,
                        "text": "Cambria et al. (2014)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "An interesting and attractive theory for computational modelling of emotions that has not been used in natural language processing yet is the emotion component process model (Scherer, 2005, CPM) . This model states that emotions are a coordinated process in five subsystems, following an event that is relevant for the experiencer of the emotion, namely a motivational action tendency, the motor expression component, a neurophysiological, bodily symptom, the subjective feeling, and the cognitive appraisal. The cognitive appraisal has been explored in a fine-grained manner by Hofmann et al. (2020) , mentioned above. The subjective feeling component is related to the dimensions of affect. 1 We hypothesize (and subsequently analyze) that emotions in text are communicated in a variety of ways, and that these different stylistic means follow the emotion component process model. The communication of emotions can either be an explicit mention of the emotion name (\"I am angry\"), focus on the motivational aspect (\"He wanted to run away.\"), describe the expression (\"She smiled.\", \"He shouted.\") or a physiological bodily reaction (\"she was trembling\", \"a tear was running down his face\"), the subjective feeling (\"I felt so bad.\"), or, finally, describe a cognitive appraisal (\"I wasn't sure what was happening.\", \"I am not responsible.\").",
                "cite_spans": [
                    {
                        "start": 174,
                        "end": 194,
                        "text": "(Scherer, 2005, CPM)",
                        "ref_id": null
                    },
                    {
                        "start": 579,
                        "end": 600,
                        "text": "Hofmann et al. (2020)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 693,
                        "end": 694,
                        "text": "1",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "With this paper, we study how emotions are communicated (following the component model) in Tweets (based on the Twitter Emotion Corpus TEC, by Mohammad (2012) ) and literature (based on the REMAN corpus by Kim and Klinger (2018) ). We post-annotate a subset of 3041 instances with the use of emotion component-based emotion communication categories, analyze this corpus, and perform joint modelling/multi-task learning experiments. Our research goals are (1) to understand if emotion components are distributed similarly across emotion categories and domains, and (2) to evaluate if informing an emotion classifier about emotion components improves their performance (and to evaluate various classification approaches). We find that emotion component and emotion classification prediction interact and benefit from each other and that emotions are communicated by means of various components in literature and social media. The corpus is available at https: //www.ims.uni-stuttgart.de/data/emotion.",
                "cite_spans": [
                    {
                        "start": 143,
                        "end": 158,
                        "text": "Mohammad (2012)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 206,
                        "end": 228,
                        "text": "Kim and Klinger (2018)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Emotion models can be separated into those that consider a discrete set of categories or those that focus on underlying principles like affect. The model of basic emotions by Ekman (1992) considers anger, disgust, fear, joy, sadness, and surprise. According to his work, there are nine characteristics that a basic emotion fulfills: These are (1) distinctive universal signals, (2) presence in other primates, (3) distinctive physiology, (4) distinctive universals in antecedent events, (5) coherence among emotional response, (6) quick onset, (7) brief duration, (8) automatic appraisal, and (9) unbidden occurrence. His model of the six universal Rashkin et al., 2018) .",
                "cite_spans": [
                    {
                        "start": 175,
                        "end": 187,
                        "text": "Ekman (1992)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 649,
                        "end": 670,
                        "text": "Rashkin et al., 2018)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Models",
                "sec_num": "2.1"
            },
            {
                "text": "emotions constitutes one of the most popular emotion sets in natural language processing. Yet it might be doubted if this set is sufficient. Plutchik (2001) proposed a model with eight main emotions, visualized on a colored wheel. In this visualization, opposites and distance of emotion names are supposed to correspond to their respective relation.",
                "cite_spans": [
                    {
                        "start": 141,
                        "end": 156,
                        "text": "Plutchik (2001)",
                        "ref_id": "BIBREF22"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Models",
                "sec_num": "2.1"
            },
            {
                "text": "A complementary approach to categorizing emotions in discrete sets is advocated by Russell and Mehrabian (1977) . Their dimensional affect model corresponds to a 3-dimensional vector space with dimensions for pleasure-displeasure, the degree of arousal, and dominance-submissiveness (VAD). Emotion categories correspond to points in this vector space. A more expressive alternative to the VAD model of affect is motivated by the cognitive appraisal process that is part of emotions. The model of Smith and Ellsworth (1985) introduces a set of variables that they map to the principle components of pleasantness, responsibility/control, certainty, attention, effort, and situational control. They show that these dimensions are more powerful to distinguish emotion categories than VAD.",
                "cite_spans": [
                    {
                        "start": 83,
                        "end": 111,
                        "text": "Russell and Mehrabian (1977)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 496,
                        "end": 522,
                        "text": "Smith and Ellsworth (1985)",
                        "ref_id": "BIBREF29"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Models",
                "sec_num": "2.1"
            },
            {
                "text": "Appraisals are also part of the emotion component process model by Scherer (2005) , which is central to this paper. The five components are cognitive appraisal, neurophysiological bodily symptoms, motor expressions, motivational action tendencies, and subjective feelings. Cognitive appraisal is concerned with the evaluation of an event. The event is assessed regarding its relevance to the individual, the implications and consequences it might lead to, the possible ways to cope with it and control it, and its significance according to personal values and social norms. The component of neurophysiological symptoms regards automatically activated reactions and symptoms of the body, like changes in the heartbeat or breathing pattern. The motor expression component contains all movements, facial expressions, changes concerning the speech, and similar patterns. Actions like attention shifts and movement with respect to the position of the event are part of the motivational action tendencies component. Finally, the component of subjective feelings takes into account how strong, important, and persisting the felt sensations are. Scherer (2005) argues that it is possible to infer the emotion a person is experiencing by analyzing the set of changes in the five components. Scherer (2009) also points out that computational models must not ignore emotion components.",
                "cite_spans": [
                    {
                        "start": 67,
                        "end": 81,
                        "text": "Scherer (2005)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 1138,
                        "end": 1152,
                        "text": "Scherer (2005)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 1282,
                        "end": 1296,
                        "text": "Scherer (2009)",
                        "ref_id": "BIBREF26"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Models",
                "sec_num": "2.1"
            },
            {
                "text": "The majority of modelling approaches focuses on the analysis of fundamental emotions (see Alswaidan and Menai, 2020; Mohammad et al., 2018; Bostan and Klinger, 2018) or on the recognition of valence, arousal, and dominance (Buechel and Hahn, 2017) . Work with a focus on other aspects of emotions is scarce.",
                "cite_spans": [
                    {
                        "start": 90,
                        "end": 116,
                        "text": "Alswaidan and Menai, 2020;",
                        "ref_id": null
                    },
                    {
                        "start": 117,
                        "end": 139,
                        "text": "Mohammad et al., 2018;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 140,
                        "end": 165,
                        "text": "Bostan and Klinger, 2018)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 223,
                        "end": 247,
                        "text": "(Buechel and Hahn, 2017)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Analysis in Text",
                "sec_num": "2.2"
            },
            {
                "text": "Noteworthy, though this has not been a computational study, is the motivation of the ISEAR project (Scherer and Wallbott, 1994) , from which a textual corpus originated, which is frequently used in NLP. It consists of event descriptions and is therefore relevant for appraisal theories. Further, participants in that study have not only been asked to report on events they experienced, but they also report additional aspects, including the existence of bodily reactions. However, their work does not focus on the linguistic realization of emotion components, but on the existence in the described event.",
                "cite_spans": [
                    {
                        "start": 99,
                        "end": 127,
                        "text": "(Scherer and Wallbott, 1994)",
                        "ref_id": "BIBREF27"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Analysis in Text",
                "sec_num": "2.2"
            },
            {
                "text": "Similarly, Troiano et al. (2019) asked crowdworkers to report on events that caused an emotion. This resource has then been postannotated with appraisal dimensions (Hofmann et al., 2020) . This is the only recent work we are aware of that models appraisal as a component of the CPM to predict emotion categories, next to the rule-based classification approach by Shaikh et al. (2009) , who built on top of the work by Clore and Ortony (2013). Another noteworthy related work is SenticNet, which models event properties including people's goals, for sentiment analysis (Cambria et al., 2014) .",
                "cite_spans": [
                    {
                        "start": 11,
                        "end": 32,
                        "text": "Troiano et al. (2019)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 164,
                        "end": 186,
                        "text": "(Hofmann et al., 2020)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 363,
                        "end": 383,
                        "text": "Shaikh et al. (2009)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 568,
                        "end": 590,
                        "text": "(Cambria et al., 2014)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Analysis in Text",
                "sec_num": "2.2"
            },
            {
                "text": "The only work we are aware of that studies emotion components (though not following the CPM, and without computational modelling), is the corpus study by Kim and Klinger (2019) . They analyze if emotions in fan fiction are communicated via facial descriptions, body posture descriptions, the appearance, look, voice, gestures, subjective sensations, or spatial relations of characters. This set of variables is not the same as emotion components, however, it is related. They find that some emotions are preferred to be described with particular aspects by authors. Their work was motivated by the linguistic study of van Meel (1995) .",
                "cite_spans": [
                    {
                        "start": 154,
                        "end": 176,
                        "text": "Kim and Klinger (2019)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 622,
                        "end": 633,
                        "text": "Meel (1995)",
                        "ref_id": "BIBREF17"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Analysis in Text",
                "sec_num": "2.2"
            },
            {
                "text": "In contrast to their work, our study compares two different domains (Tweets and Literature), and follows the emotion component process model more strictly. Further, we show the use of that model for computational emotion classification through multi-task learning.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Analysis in Text",
                "sec_num": "2.2"
            },
            {
                "text": "3 Corpus Annotation",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Analysis in Text",
                "sec_num": "2.2"
            },
            {
                "text": "To study the relation between emotion components and emotions, we annotate subsets from two different existing emotion corpora from two different domains, namely literature and social media.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Selection",
                "sec_num": "3.1"
            },
            {
                "text": "For literature, we use the REMAN corpus (Kim and Klinger, 2018) , which consists of fiction written after the year 1800. It is manually annotated with text spans related to emotions, as well as their experiencers, causes, and targets. Emotion cue spans are annotated with the emotions of anger, fear, trust, disgust, joy, sadness, surprise, and anticipation, as well as 'other emotion'. From the 1720 instances, we randomly sample a subset of 1000. Each instance comprises a sentence triple and may contain any number of annotated spans. We map the emotions associated to spans to the text instances as the union of all labels, which leads to a multi-label classification task. Instances without emotion annotations are considered 'neutral'.",
                "cite_spans": [
                    {
                        "start": 40,
                        "end": 63,
                        "text": "(Kim and Klinger, 2018)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Selection",
                "sec_num": "3.1"
            },
            {
                "text": "For the social media domain, we choose the Twitter Emotion Corpus (TEC) (Mohammad, 2012) . The emotion categories are anger, disgust, fear, joy, sadness, and surprise. TEC consists of approximately 21,000 posts from Twitter that have a hashtag at the end which states one of the six mentioned emotions. According to the authors, the validity of hashtags as classification labels is commensurable to the inter-annotator agreements of human annotators. We randomly sample 2041 instances with the emotion hashtags as labels for the creation of our corpus. Each instance equals one post and has exactly one emotion label.",
                "cite_spans": [
                    {
                        "start": 72,
                        "end": 88,
                        "text": "(Mohammad, 2012)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Selection",
                "sec_num": "3.1"
            },
            {
                "text": "Inter-Annotator Agreement",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Annotation Procedure and",
                "sec_num": "3.2"
            },
            {
                "text": "We annotate the emotion component dimensions independently: The existence of a CPM label means that this component is mentioned somewhere in the text, independent of its function to communicate one of the emotions. This is a simplification due to the fact that it turned out to be difficult to infer from the limited context of an instance if an emotion category and an emotion component mention are actually in relation. Further, this procedure also ensures that there is no information leak introduced in the annotation process (e.g., that components are only annotated if they indeed inform the emotion, and that a model could learn from its sheer presence).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Annotation Procedure and",
                "sec_num": "3.2"
            },
            {
                "text": "Cognitive appraisal evaluation of the pleasantness of an event. Thinks that @melbahughes had a great 50th birthday party Neurophysiol. symptoms change in someone's heartbeat.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explanation of Example Example",
                "sec_num": null
            },
            {
                "text": "Loves when a song makes your heart race [...] Motiv. Action tendencies urge to attack a person or object.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explanation of Example Example",
                "sec_num": null
            },
            {
                "text": "sometimes when i think bout you i want to beat the shit out of your face so everyone can see how ugly you are inside and out Motor expressions facial expression. @TheBodyShopUK when I walk in the room and my 9month old nephew recognises me and his face lights up with the biggest smile thats 100% Subjective feelings internal feeling state. Feelin a bit sad tonight We refined the annotation guidelines in an iterative process with two annotators. Annotator 1 is a 23 year-old female undergraduate computer science student, Annotator 2 is a 28 year-old male graduate student of computational linguistics. We first defined a list of guidelines for each emotion component, then let each annotator label 40 randomly sampled instances (20 each in two iterations) out of each corpus and measured the interannotator agreement. Based on instances with disagreement, we refined the guidelines. The achieved inter-annotator agreement scores are displayed in Table 2 . We observe that particularly the concepts of cognitive appraisal and motivational action tendencies have been clarified. During this process, for example, the discussion of the instance \"He did so, and to his surprise, found that all the bank stock had been sold, and transferred\" lead to the addition of a rule stating that the explicit mention of a feeling has to be annotated with subjective feeling. A rule for the annotation of tiredness as neurophysiological symptoms was created due to the instance \"Here he remained the whole night, feeling very tired and sorrowful.\". Concerning the annotation of verbal communication as motor expression, we decided to only annotate instances with verbal communications that address an emotional reaction or instances with interjections as for example 'oh' or 'wow'. With this clarification, the instance \"'Jolly rum thing about that boat,' said the spokesman of the party, as the boys continued their walk. 'I expect it got adrift somehow,' said another. 'I don't know,' said the first.\" should not be annotated, whereas \"'Sounds delightful.' 'Oh, it was actually pretty cool.'\" should (this aspect has particularly appeared in the second annotator training round, which lead to a slight decrease in agreement). We make the annotation guidelines available together with our corpus. Table 1 shows a short excerpt.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 949,
                        "end": 956,
                        "text": "Table 2",
                        "ref_id": "TABREF1"
                    },
                    {
                        "start": 2285,
                        "end": 2292,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Explanation of Example Example",
                "sec_num": null
            },
            {
                "text": "After the refinement process concluded, Annotator 1 annotated the subsample of TEC and Annotator 2 annotated the subsample of REMAN.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explanation of Example Example",
                "sec_num": null
            },
            {
                "text": "We show corpus statistics in Table 3 to develop an understanding how emotions are communicated in the two domains. For both corpora, we observe that cognitive appraisal is most frequent. In TEC, the second most dominant component is subjective feeling, while in REMAN it is the motor expression. The amount of subjective feeling descriptions is substantially lower for literature than for social media -which is in line with the show-don't-tell paradigm which is obviously not followed in social media as it is in literature.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 29,
                        "end": 36,
                        "text": "Table 3",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Corpus Statistics",
                "sec_num": "3.3"
            },
            {
                "text": "Components are not distributed equally across emotions. Particularly noteworthy is the cooccurrence of disgust with neurophysiological symptoms in social media, but not in literature where this component dominates the emotion of fear. We also observe a particularly high cooccurrence of the subjective feeling component with fear for social media, which is not the case for literature. In literature, the motivational action tendency component co-occurs with anger (and anticipation) more frequently than with all other emotions. This is not the case for the social media do- main. On the REMAN corpus, components occur least frequently when there is no emotion across all components. For both corpora, neurophysiological symptoms make up the smallest share of components, even more so in the case of TEC than REMAN.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Statistics",
                "sec_num": "3.3"
            },
            {
                "text": "In a comparison of social media and literature, we observe that emotions are distributed more uniformly in literature. The relative number of cooccurrences of CPM components with emotions varies more for REMAN than for the TEC corpus.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Statistics",
                "sec_num": "3.3"
            },
            {
                "text": "We will now turn to the computational modelling of emotion components and evaluate their usefulness for emotion classification. We evaluate a set of different feature-based and deep-learning based classification approaches to join the tasks of emotion classification and component classification.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods",
                "sec_num": "4"
            },
            {
                "text": "As baseline emotion classification models which are not particularly informed about components, we use two models: Emo-ME-Base is a maximum entropy (ME) classifier with TF-IDF-weighted bagof-words unigram and bigram features. As preprocessing, we convert all words to lowercase, and stem them with the PorterStemmer. On TEC, with its single-label annotation, Emo-ME-Base consists of one model, while on REMAN with multi-label annotation, we use 10 binary classifiers.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Classifier",
                "sec_num": "4.1"
            },
            {
                "text": "Our neural baseline Emo-NN-Base uses pretrained BERT sentence embeddings 2 (Devlin et al., 2019) as input features. Inspired by Chen and Wang (2018) ; Sosa (2017), the network architecture consists of a bidirectional LSTM layer (Hochreiter and Schmidhuber, 1997) , followed by a convolutional layer with kernel sizes 2, 3, 5, 7, 13, and 2 https://tfhub.dev/google/experts/bert/wiki books/sst2/1 25. The outputs of the convolutional layer are maxpooled over the dimension of the input sequence, inspired by Collobert et al. (2011) . Stacked on top of the pooling layer is a fully connected layer. Its outputs are finally fed into an output layer with a sigmoid activation function (see Figure 1a) . 3 We use dropout regularization after each layer. The network uses a weighted cross-entropy loss function, whereby the loss of false negatives is multiplied by 4 to increase recall. The model is trained using an Adam optimizer (Kingma and Ba, 2015). All network parameters of this model and subsequent neural models are determined using a subset of the training data as development set for the RE-MAN corpus and using 10-fold cross-validation for the TEC corpus. Details of the resulting hyperparameters are listed in the Appendix.",
                "cite_spans": [
                    {
                        "start": 75,
                        "end": 96,
                        "text": "(Devlin et al., 2019)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 128,
                        "end": 148,
                        "text": "Chen and Wang (2018)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 228,
                        "end": 262,
                        "text": "(Hochreiter and Schmidhuber, 1997)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 506,
                        "end": 529,
                        "text": "Collobert et al. (2011)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 698,
                        "end": 699,
                        "text": "3",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 685,
                        "end": 695,
                        "text": "Figure 1a)",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Emotion Classifier",
                "sec_num": "4.1"
            },
            {
                "text": "The emotion component classifiers predict which of the five CPM components occur in a text instance. Our Cpm-ME-Base baseline models (one for each component) only use bag-of-words features in the same configuration as Emo-ME-Base.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Component Classifier",
                "sec_num": "4.2"
            },
            {
                "text": "In the model Cpm-ME-Adv, we add taskspecific features, namely features derived from manually crafted small dictionaries with words associated with the different components. Those dictionaries were developed without considering the corpora and with inspiration from Scherer (2005) and contain on average 26 items. Further, we add part-of-speech tags (calculated with spaCy 4 , Honnibal et al. (2020)) and glove-twitter-100 embeddings 5 (Pennington et al., 2014) . Additionally, only for the cognitive appraisal component, we run the appraisal classifier developed by Hofmann et al. (2020) and use the predictions as features. 6 For each component individually, the best-performing combination of these features is chosen.",
                "cite_spans": [
                    {
                        "start": 265,
                        "end": 279,
                        "text": "Scherer (2005)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 435,
                        "end": 460,
                        "text": "(Pennington et al., 2014)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 566,
                        "end": 587,
                        "text": "Hofmann et al. (2020)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 625,
                        "end": 626,
                        "text": "6",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Component Classifier",
                "sec_num": "4.2"
            },
            {
                "text": "The Cpm-NN-Base is configured analogously to Emo-NN-Base. The primary reason for using an equivalent setup is to facilitate a multi-head architecture as joint model for both tasks in the next step. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Component Classifier",
                "sec_num": "4.2"
            },
            {
                "text": "Cognitive Phys. Motiv. Action Motor Exp. Subject. Total TEC",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion",
                "sec_num": null
            },
            {
                "text": "To analyze if emotion classification benefits from the component prediction (and partially also vice versa), we set up several model configurations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint Modelling and Multi-Task Learning of Emotions and Components",
                "sec_num": "4.3"
            },
            {
                "text": "In Emo-Cpm-ME-Pred, we predict the emotion with Cpm-ME-Adv and use these predictions as features. Other than that, Emo-Cpm-ME-Pred corresponds to Emo-ME-Base. In Emo-Cpm-ME-Gold, we replace the predictions by gold component annotations to analyze error propagation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint Modelling and Multi-Task Learning of Emotions and Components",
                "sec_num": "4.3"
            },
            {
                "text": "Emo-Cpm-NN-Pred and Emo-Cpm-NN-Gold are configured analogously and follow the same architecture as Emo-NN-Base with the following differences: A binary vector with the CPM annotations is introduced as additional input feature, feeding into a fully connected layer. Its outputs are concatenated with the outputs of the penultimate layer and passed to another fully connected layer, followed by the output layer.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint Modelling and Multi-Task Learning of Emotions and Components",
                "sec_num": "4.3"
            },
            {
                "text": "Emo-Cpm-NN-Pred uses Cpm-NN-Base to obtain component predictions, but the weights of Cpm-NN-Base are frozen. The basic network architecture resembles that of the Emo-Cpm-NN-Gold model, replacing the additional CPM input vector with the Cpm-NN-Base model (see Figure 1b) . Its outputs are, again, fed into a fully connected layer which is connected to the output layer.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 259,
                        "end": 269,
                        "text": "Figure 1b)",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Joint Modelling and Multi-Task Learning of Emotions and Components",
                "sec_num": "4.3"
            },
            {
                "text": "Next to the models that make use of the output of the CPM classifiers for prediction, we use two multi-task learning models which predict emotions and components based on shared latent variables. For a multi-head variant (MTL-MH), the basic architectures of the individual models for both tasks remain the same. Outputs of the CNN layer are fed to two separate, task-specific, fully connected layers. This model has two output layers, one for emotion classification and one for CPM component classification. Both tasks use the weighted cross entropy loss function to increase recall. Based on the model proposed by Misra et al. (2016), we use cross-stitch units in our model MTL-XS. This model employs two separate parallel instances of the Cpm-NN-Base architecture introduced above, one for the CPM classification task and one for emotion classification. The model additionally employs one cross-stitch unit after the respective CNN layers. This sharing unit learns a linear combination of the pooled task-specific CNN activation maps which is then passed to the taskspecific fully connected layers. The cross-stitch unit learns during training which information to share across tasks (see Figure 1c) .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 1191,
                        "end": 1201,
                        "text": "Figure 1c)",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Joint Modelling and Multi-Task Learning of Emotions and Components",
                "sec_num": "4.3"
            },
            {
                "text": "For our experiments, we use our reannotated subsample of TEC and REMAN (not all instances available in TEC and REMAN). We split the corpora into 90% for training and 10% to test. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "We start the discussion of the results with the component classification, a classification task that has not been addressed before and for which our data set is the first that becomes available to the research community. Table 4 shows the results. The model performances are acceptable. Macroaverage F 1 scores on REMAN range from .42 of MTL-MH to .59 for Cpm-ME-Adv, and from .53 (Cpm-ME-Base) to .57 (Cpm-NN-Base) on TEC. There are, however, differences for the components: On TEC, there are difficulties in predicting neurophysiological symptoms. The addition of taskspecific features in Cpm-ME-Adv shows a clear improvement across all components.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 221,
                        "end": 228,
                        "text": "Table 4",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Component Prediction",
                "sec_num": "5.1"
            },
            {
                "text": "The neural baseline Cpm-NN-Base outperforms Cpm-ME-Adv on TEC, and does so without feature engineering. On REMAN, the feature-based model is superior which might be due to the engineered features being more commonly represented in the literature domain than in social media. This is partially leveraged in the MTL-XS model on REMAN.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Component Prediction",
                "sec_num": "5.1"
            },
            {
                "text": "The components are not equally difficult to predict; the relations between the components are comparable across models. The lowest performance scores are observed for neurophysiological symptoms. This holds across models and corpora. For the neurophysiological component on the literature domain, however, the engineered features in Cpm-ME-Adv show substantial improvement, yielding an F 1 score of 0.44. Cognitive appraisal shows best prediction performances, with F 1 between .73 and .86. For TEC, we observe a correlation between performance and class size for all components.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Component Prediction",
                "sec_num": "5.1"
            },
            {
                "text": "For REMAN, Cpm-ME-Adv is the bestperforming model. Cpm-ME-Adv's macro average F 1 of 0.59 is 9pp higher than the second best F 1score. For TEC, the best results are achieved by Cpm-NN-Base with a macro F 1 of 0.57.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Component Prediction",
                "sec_num": "5.1"
            },
            {
                "text": "In this section, we discuss the performance of our emotion classification models across different configurations. One question is how providing component information to them helps most. Table 5 shows the results for all experiments.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 186,
                        "end": 193,
                        "text": "Table 5",
                        "ref_id": "TABREF8"
                    }
                ],
                "eq_spans": [],
                "section": "Emotion Classification",
                "sec_num": "5.2"
            },
            {
                "text": "The comparison of Emo-ME-Base and Emo-NN-Base reveals that a pure word-based model is not able to categorize emotions in REMAN, due to the imbalancedness in this multilabel classification setup. This observation is in line with previous results (Kim and Klinger, 2018) . The use of BERT's contextualized sentence embeddings leads to a strong improvement of 43pp (against a 0 F 1 for Emo-ME-Base). The performance of the ME models is comparably limited also on TEC, though this is less obvious on the micro-averaged F 1 due to the imbalancedness of the resource (.35 macro, .54 micro F 1 ).",
                "cite_spans": [
                    {
                        "start": 245,
                        "end": 268,
                        "text": "(Kim and Klinger, 2018)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Classification",
                "sec_num": "5.2"
            },
            {
                "text": "Our main research question is if emotion components help emotion classification. In our first attempt to include this information as features, we see some improvement. On REMAN, Emo-Cpm-ME-Pred \"boosts\" from 0 to 6 F 1 , on TEC we 0  0  0  0  0  0  0  0  0  0  0  0  Emo-Cpm-ME-Gold  18  0  0  25  16  62  0  0  0  0  12  14  Emo-Cpm-ME-Pred  0  0  0  12  15  0  0  0  0  14  4  6  Emo-NN-Base  36  18  29  41  59  46  14  36  71  50  40 To answer the question if this limited improvement is only due to a limited performance of the component classification model, we compare these results to a setting, in which the predicted values are replaced by gold labels from the annotation. This setup does show an improvement with Emo-Cpm-ME-Gold to .14 F 1 on REMAN, which is obviously still very low; and no improvement on TEC. However, with our neural model Emo-Cpm-NN-Gold, we see the potential of gold information increasing the score for emotion classification to .45 F 1 on REMAN and .62 F 1 on TEC.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 231,
                        "end": 437,
                        "text": "0  0  0  0  0  0  0  0  0  0  0  0  Emo-Cpm-ME-Gold  18  0  0  25  16  62  0  0  0  0  12  14  Emo-Cpm-ME-Pred  0  0  0  12  15  0  0  0  0  14  4  6  Emo-NN-Base  36  18  29  41  59  46  14  36  71  50  40",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Emotion Classification",
                "sec_num": "5.2"
            },
            {
                "text": "This is an unrealistic setting -the classifier does not have access to annotated labels in real world applications. However, in the (realistic) crossstitch multi-task learning setting of MTL-XS, we observe further improvements: On REMAN, we achieve .47 F 1 (which is even slightly higher than with gold component labels), which constitutes an achieved improvement by 4pp to the emotion classifier which is not informed about components. On TEC, we achieve .61 F 1 , which is close to the model that has access to gold components (.62). This is an improvement of 4pp as well in comparison to the model that has no access to components but follows the same architecture.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Classification",
                "sec_num": "5.2"
            },
            {
                "text": "Particularly, we observe that models with component information perform better across all emotions, with the exception of surprise on the REMAN corpus and anger on the TEC corpus. We can therefore conclude that emotion component information does contribute to emotion classification; the bestperforming combination is via a cross-stitch model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Classification",
                "sec_num": "5.2"
            },
            {
                "text": "A detailed discussion based on example predictions of the various models is available in the Appendix.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emotion Classification",
                "sec_num": "5.2"
            },
            {
                "text": "We presented the first data sets (based on existing emotion corpora) with emotion component annotation. While Hofmann et al. (2020) has proposed to use the cognitive appraisal for emotion classification, they did not succeed to present models that actually benefit in emotion classification performance. That might be due to the fact that cognitive appraisal classification itself is challenging, and that they did not compare multiple multi-task learning approaches.",
                "cite_spans": [
                    {
                        "start": 104,
                        "end": 131,
                        "text": "While Hofmann et al. (2020)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "6"
            },
            {
                "text": "With this paper we moved to another psychological theory, namely the emotion component process model, and make the first annotations available that closely follow this theory. Based on this resource, we have shown that, even with a comparably limited data set size, emotion components contribute to emotion classification. We expect that with a larger corpus the improvement would be more substantial than it is already now. A manual introspection of the data instances also shows that the components indeed help. Further, we have seen that emotions are communicated quite differently in the two domains, which is an explanation why emotion classification systems (up-to-today) need to be developed particularly for domains of interest. We propose that future work analyzes further which information is relevant and should be shared across these tasks in multi-task learning models.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "6"
            },
            {
                "text": "Further, we propose that larger corpora should be created across more domains, and also that multitask learning is not only performed individually, but also across corpora. Presumably, the component information in different domains is not the same, but might be helpful across them.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "6"
            },
            {
                "text": "We did not collect a new data set from individuals, but did reannotate existing and publicly available resources. Therefore, this paper does not pose ethical questions regarding data collection.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ethical Considerations",
                "sec_num": null
            },
            {
                "text": "However, emotion analysis has the principled potential to be misused, and researchers need to be aware that their findings (though they are not in themselves harmful) might lead to software that can do harm. We assume that sentiment and emotion analysis are sufficiently well-known that users of social media might be aware that their data could be automatically analyzed. However, we propose that no automatic system ever does report back analyses of individuals and instead does aggregate data of anonymized posts. We do not assume that analyzing literature data poses any risk.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ethical Considerations",
                "sec_num": null
            },
            {
                "text": "One aspect of our work we would like to point out is that, in contrast to other and previous emotion analysis research, we focus and enable particularly the analysis of implicit (and perhaps even unconcious) communication of emotions. That might further mean that authors of posts in social media are not aware that their emotional state could be computationally analyzed, potentially, they are not even fully aware of their own affective state. We would like to point out that automatically analyzing social media data without the explicit consent of the users is unethical at least when the user can be identified or identify themselves, particularly if they might not be aware of the details of an analysis system.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ethical Considerations",
                "sec_num": null
            },
            {
                "text": "A Ablation Study for Feature Based Maximum Entropy Classification Model of Emotion Components Table 6 shows the performance scores if just one additional feature is enabled (while bag-of-words always remains available). It can be seen, that the most advantageous feature are word embeddings. On REMAN, Cpm-ME-Adv achieves a macro F1-score of 0.59 and a micro F 1 -score of 0.67. On TEC, we have respective values of 0.56 and 0.71, with the high micro score resulting from cognitive appraisal being the best performing class while also being more than twice as frequent as any other component. Table 6 : Overview over the single feature's impact in classification with Cpm-ME-Adv. Each column displays the classification results if only this column's feature is additionally to bag-of-words features, enabled. In the last column, the additional feature is only used for the prediction of cognitive appraisal, due to the classification assumption that the components can appear individually of each other in text.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 94,
                        "end": 101,
                        "text": "Table 6",
                        "ref_id": null
                    },
                    {
                        "start": 593,
                        "end": 600,
                        "text": "Table 6",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Ethical Considerations",
                "sec_num": null
            },
            {
                "text": "The results table in the main paper did, for space reasons, only show F 1 scores. Table 8 shows the network parameters that were determined during the development process of the neural models.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 82,
                        "end": 89,
                        "text": "Table 8",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "B Detailed Emotion Results for Emotion Classification",
                "sec_num": null
            },
            {
                "text": "Parameter Cpm-NN-Base Emo-NN-Base Emo-Cpm-NN-Gold Emo-Cpm-NN-Pred",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C Neural Network Parameters",
                "sec_num": null
            },
            {
                "text": "There exists other work that has been motivated by appraisal theories, but that is either rule-based(Shaikh et al., 2009;Udochukwu and He, 2015) or does not explicitly model appraisal or component dimensions(Balahur et al., 2012;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "We selected this architecture based on preliminary experiments on the validation data. We evaluated it against LSTM-Dense Layer and CNN-LSTM architectures.4 https://spacy.io/usage/linguistic-features#pos-tagging 5 https://nlp.stanford.edu/projects/glove/ 6 http://www.ims.uni-stuttgart.de/data/appraisalemotion",
                "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).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            },
            {
                "text": " Minibatch size  40  80  80  80  80  80   Table 8 : Neural network parameters. In cases where multiple values are displayed, the first value refers to the emotion detection part of the network, while the second value refers to CPM detection.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 1,
                        "end": 49,
                        "text": "Minibatch size  40  80  80  80  80  80   Table 8",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "annex",
                "sec_num": null
            },
            {
                "text": "We show examples in Table 9 where component information is helpful for emotion classification. Regarding the neural classifiers, MTL-XS generally tends to predict fewer false positives when there are no strong correlations among the potential emotions to the predicted CPM, like in (1). Similarly, in (2) the model predicts only 'fear', which is more likely to occur together with the 'subjective feeling' component than 'anger' or 'disgust', according to Table 3 in the paper. Additionally, CPM information helps to solve ambiguities: In (3), the model predicts 'anticipation' rather than 'sadness', presumably because of the stronger correlation to the predicted CPM component 'action tendency'. In the two TEC examples (4-5), the baseline detects 'joy', while MTL-XS correctly detects 'sadness'. The cross-stitch model predicts a 'subjective feeling' component in both instances and a 'cognitive appraisal' component in one instance. Both components are more strongly correlated with 'sadness' than with 'joy' (see Table 3 in main paper).We also show some examples that exemplify differences in prediction of the ME-based models (6-8). Generally, the CPM information leads to little improvement in emotion detection on TEC. Nevertheless, there are some cases in which the correct emotion was predicted by at least one of Emo-Cpm-ME-Gold and Emo-Cpm-ME-Pred, whereas it was not detected by Emo-ME-Base. In both examples (6-7), the correct emotions 'surprise' and 'sadness' have not been found by Emo-ME-Base (predicting 'joy' and 'surprise' respectively). Emo-Cpm-ME-Gold and Emo-Cpm-ME-Pred both correctly predicted 'surprise' for (6) and 'sadness' for (7). There are indications of 'subjective feeling' in the second and of 'motor expression' and 'cognitive appraisal' in both examples, that were also predicted by Cpm-ME-Adv, which might have helped assigning the correct emotion class. On REMAN, the ME models were able to classify a small fraction of the instances correctly, which is still an improvement compared to the miserably failing baseline. An example with improved prediction for REMAN is (8), where the emotion 'joy' was correctly identified by Emo-Cpm-ME-Gold and Emo-Cpm-ME-Pred, while not being detected by Emo-ME-Base.(1) As for the hero of this story, 'His One Fault' was absent-mindedness. He forgot to lock his uncle's stable door, and the horse was stolen. In seeking to recover the stolen horse, he unintentionally stole another. (REMAN would be the very last to complain of it. We went to bed again, and the forsaken child of some half-animal mother, now perhaps asleep in some filthy lodging for tramps, lay in my Ethelwyn's bosom. I loved her the more for it; though, I confess, it would have been very painful to me had she shown it possible for her to treat the baby otherwise, especially after what we had been talking about that same evening. (REMAN) CPM Cpm-ME-Adv cognitive appraisal, action tendency, subjective feeling Emotion Emo-ME-Base / Emotion Emo-Cpm-ME-Pred joy Emotion Emo-Cpm-ME-Gold joy CPM Gold cognitive appraisal, subjective feeling Emotion Gold disgust, joy, sadness, trust Table 9 : Examples in which components support emotion classification.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 20,
                        "end": 27,
                        "text": "Table 9",
                        "ref_id": null
                    },
                    {
                        "start": 456,
                        "end": 463,
                        "text": "Table 3",
                        "ref_id": null
                    },
                    {
                        "start": 1018,
                        "end": 1025,
                        "text": "Table 3",
                        "ref_id": null
                    },
                    {
                        "start": 2457,
                        "end": 2463,
                        "text": "(REMAN",
                        "ref_id": null
                    },
                    {
                        "start": 3126,
                        "end": 3133,
                        "text": "Table 9",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "D Discussion of Instances",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "A survey of state-of-the-art approaches for emotion recognition in text",
                "authors": [],
                "year": null,
                "venue": "Knowledge and Information Systems",
                "volume": "62",
                "issue": "",
                "pages": "2937--2987",
                "other_ids": {
                    "DOI": [
                        "10.1007/s10115-020-01449-0"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Nourah Alswaidan and Mohamed Menai. 2020. A sur- vey of state-of-the-art approaches for emotion recog- nition in text. Knowledge and Information Systems, 62:2937-2987.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Building and exploiting emotinet, a knowledge base for emotion detection based on the appraisal theory model",
                "authors": [
                    {
                        "first": "Alexandra",
                        "middle": [],
                        "last": "Balahur",
                        "suffix": ""
                    },
                    {
                        "first": "Jesus",
                        "middle": [
                            "M"
                        ],
                        "last": "Hermida",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Montoyo",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "IEEE Transactions on Affective Computing",
                "volume": "3",
                "issue": "1",
                "pages": "88--101",
                "other_ids": {
                    "DOI": [
                        "10.1109/T-AFFC.2011.33"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Alexandra Balahur, Jesus M. Hermida, and Andrew Montoyo. 2012. Building and exploiting emotinet, a knowledge base for emotion detection based on the appraisal theory model. IEEE Transactions on Affective Computing, 3(1):88-101.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "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
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "EmoBank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis",
                "authors": [
                    {
                        "first": "Sven",
                        "middle": [],
                        "last": "Buechel",
                        "suffix": ""
                    },
                    {
                        "first": "Udo",
                        "middle": [],
                        "last": "Hahn",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 15th Conference of the European Chapter",
                "volume": "2",
                "issue": "",
                "pages": "578--585",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sven Buechel and Udo Hahn. 2017. EmoBank: Study- ing the impact of annotation perspective and repre- sentation format on dimensional emotion analysis. In Proceedings of the 15th Conference of the Euro- pean Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 578-585, Valencia, Spain. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Towards a unified framework for emotion analysis",
                "authors": [
                    {
                        "first": "Sven",
                        "middle": [],
                        "last": "Buechel",
                        "suffix": ""
                    },
                    {
                        "first": "Luise",
                        "middle": [],
                        "last": "Modersohn",
                        "suffix": ""
                    },
                    {
                        "first": "Udo",
                        "middle": [],
                        "last": "Hahn",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sven Buechel, Luise Modersohn, and Udo Hahn. 2020. Towards a unified framework for emotion analysis.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Senticnet 3: A common and common-sense knowledge base for cognition-driven sentiment analysis",
                "authors": [
                    {
                        "first": "Erik",
                        "middle": [],
                        "last": "Cambria",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Olsher",
                        "suffix": ""
                    },
                    {
                        "first": "Dheeraj",
                        "middle": [],
                        "last": "Rajagopal",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the AAAI",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Erik Cambria, Daniel Olsher, and Dheeraj Rajagopal. 2014. Senticnet 3: A common and common-sense knowledge base for cognition-driven sentiment anal- ysis. In Proceedings of the AAAI.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Advanced combined LSTM-CNN model for Twitter sentiment analysis",
                "authors": [
                    {
                        "first": "Nan",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Peikang",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "5th IEEE International Conference on Cloud Computing and Intelligence Systems (CCIS)",
                "volume": "",
                "issue": "",
                "pages": "684--687",
                "other_ids": {
                    "DOI": [
                        "10.1109/CCIS.2018.8691381"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Nan Chen and Peikang Wang. 2018. Advanced com- bined LSTM-CNN model for Twitter sentiment anal- ysis. In 2018 5th IEEE International Conference on Cloud Computing and Intelligence Systems (CCIS), pages 684-687.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Psychological construction in the OCC model of emotion",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Gerald",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Clore",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ortony",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Emotion Review",
                "volume": "5",
                "issue": "4",
                "pages": "335--343",
                "other_ids": {
                    "DOI": [
                        "10.1177/1754073913489751"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Gerald L. Clore and Andrew Ortony. 2013. Psycho- logical construction in the OCC model of emotion. Emotion Review, 5(4):335-343.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Natural language processing (almost) from scratch",
                "authors": [
                    {
                        "first": "Ronan",
                        "middle": [],
                        "last": "Collobert",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Weston",
                        "suffix": ""
                    },
                    {
                        "first": "L\u00e9on",
                        "middle": [],
                        "last": "Bottou",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Karlen",
                        "suffix": ""
                    },
                    {
                        "first": "Koray",
                        "middle": [],
                        "last": "Kavukcuoglu",
                        "suffix": ""
                    },
                    {
                        "first": "Pavel",
                        "middle": [],
                        "last": "Kuksa",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Journal of machine learning research",
                "volume": "12",
                "issue": "",
                "pages": "2493--2537",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ronan Collobert, Jason Weston, L\u00e9on Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. 2011. Natural language processing (almost) from scratch. Journal of machine learning research, 12:2493- 2537.",
                "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 Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "An argument for basic emotions",
                "authors": [
                    {
                        "first": "Paul",
                        "middle": [],
                        "last": "Ekman",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Cognition & emotion",
                "volume": "6",
                "issue": "3-4",
                "pages": "169--200",
                "other_ids": {
                    "DOI": [
                        "10.1080/02699939208411068"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Paul Ekman. 1992. An argument for basic emotions. Cognition & emotion, 6(3-4):169-200.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Long short-term memory",
                "authors": [
                    {
                        "first": "Sepp",
                        "middle": [],
                        "last": "Hochreiter",
                        "suffix": ""
                    },
                    {
                        "first": "J\u00fcrgen",
                        "middle": [],
                        "last": "Schmidhuber",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Neural Computation",
                "volume": "9",
                "issue": "8",
                "pages": "1735--1780",
                "other_ids": {
                    "DOI": [
                        "10.1162/neco.1997.9.8.1735"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9(8):1735-1780.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "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 emotion classification in text. In Proceedings of the 28th In- ternational Conference on Computational Linguis- tics, pages 125-138, Barcelona, Spain (Online). In- ternational Committee on Computational Linguis- tics.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "spaCy: Industrial-strength Natural Language Processing in Python",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.5281/zenodo.1212303"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Matthew Honnibal, Ines Montani, Sofie Van Lan- deghem, and Adriane Boyd. 2020. spaCy: Industrial-strength Natural Language Processing in Python.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Who feels what and why? annotation of a literature corpus with semantic roles of emotions",
                "authors": [
                    {
                        "first": "Evgeny",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Roman",
                        "middle": [],
                        "last": "Klinger",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 27th International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "1345--1359",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Evgeny Kim and Roman Klinger. 2018. Who feels what and why? annotation of a literature corpus with semantic roles of emotions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1345-1359. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "An analysis of emotion communication channels in fan-fiction: Towards emotional storytelling",
                "authors": [
                    {
                        "first": "Evgeny",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Roman",
                        "middle": [],
                        "last": "Klinger",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the Second Workshop on Storytelling",
                "volume": "",
                "issue": "",
                "pages": "56--64",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-3406"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Evgeny Kim and Roman Klinger. 2019. An analysis of emotion communication channels in fan-fiction: Towards emotional storytelling. In Proceedings of the Second Workshop on Storytelling, pages 56-64, Florence, Italy. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Adam: A method for stochastic optimization",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Diederik",
                        "suffix": ""
                    },
                    {
                        "first": "Jimmy",
                        "middle": [],
                        "last": "Kingma",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "International Conference on Learning Representations (ICLR Poster)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Inter- national Conference on Learning Representations (ICLR Poster).",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Representing emotions in literature and paintings: A comparative analysis",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Jacques",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Van Meel",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Emotions and Cultural Products",
                "volume": "23",
                "issue": "",
                "pages": "159--176",
                "other_ids": {
                    "DOI": [
                        "10.1016/0304-422X(94)00013-V"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jacques M. van Meel. 1995. Representing emotions in literature and paintings: A comparative analysis. Po- etics, 23(1):159 -176. Emotions and Cultural Prod- ucts.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Cross-stitch networks for multi-task learning",
                "authors": [
                    {
                        "first": "Ishan",
                        "middle": [],
                        "last": "Misra",
                        "suffix": ""
                    },
                    {
                        "first": "Abhinav",
                        "middle": [],
                        "last": "Shrivastava",
                        "suffix": ""
                    },
                    {
                        "first": "Abhinav",
                        "middle": [],
                        "last": "Gupta",
                        "suffix": ""
                    },
                    {
                        "first": "Martial",
                        "middle": [],
                        "last": "Hebert",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
                "volume": "",
                "issue": "",
                "pages": "3994--4003",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. 2016. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pages 3994-4003.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "#emotional tweets",
                "authors": [
                    {
                        "first": "Saif",
                        "middle": [],
                        "last": "Mohammad",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "*SEM 2012: The First Joint Conference on Lexical and Computational Semantics",
                "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 Vol- ume 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
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "SemEval-2018 task 1: Affect in tweets",
                "authors": [
                    {
                        "first": "Saif",
                        "middle": [],
                        "last": "Mohammad",
                        "suffix": ""
                    },
                    {
                        "first": "Felipe",
                        "middle": [],
                        "last": "Bravo-Marquez",
                        "suffix": ""
                    },
                    {
                        "first": "Mohammad",
                        "middle": [],
                        "last": "Salameh",
                        "suffix": ""
                    },
                    {
                        "first": "Svetlana",
                        "middle": [],
                        "last": "Kiritchenko",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of The 12th International Workshop on Semantic Evaluation",
                "volume": "",
                "issue": "",
                "pages": "1--17",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/S18-1001"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Saif Mohammad, Felipe Bravo-Marquez, Mohammad Salameh, and Svetlana Kiritchenko. 2018. SemEval- 2018 task 1: Affect in tweets. In Proceedings of The 12th International Workshop on Semantic Eval- uation, pages 1-17, New Orleans, Louisiana. Asso- ciation for Computational Linguistics.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "GloVe: Global vectors for word representation",
                "authors": [
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Pennington",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "1532--1543",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/D14-1162"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "The nature of emotions: Human emotions have deep evolutionary roots, a fact that may explain their complexity and provide tools for clinical practice",
                "authors": [
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Plutchik",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "American scientist",
                "volume": "89",
                "issue": "4",
                "pages": "344--350",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Robert Plutchik. 2001. The nature of emotions: Hu- man emotions have deep evolutionary roots, a fact that may explain their complexity and provide tools for clinical practice. American scientist, 89(4):344- 350.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Event2Mind: Commonsense inference on events, intents, and reactions",
                "authors": [
                    {
                        "first": "Maarten",
                        "middle": [],
                        "last": "Hannah Rashkin",
                        "suffix": ""
                    },
                    {
                        "first": "Emily",
                        "middle": [],
                        "last": "Sap",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Allaway",
                        "suffix": ""
                    },
                    {
                        "first": "Yejin",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Choi",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "463--473",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P18-1043"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hannah Rashkin, Maarten Sap, Emily Allaway, Noah A. Smith, and Yejin Choi. 2018. Event2Mind: Commonsense inference on events, intents, and reac- tions. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 463-473, Melbourne, Australia. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Evidence for a three-factor theory of emotions",
                "authors": [
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Russell",
                        "suffix": ""
                    },
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Mehrabian",
                        "suffix": ""
                    }
                ],
                "year": 1977,
                "venue": "Journal of Research in Personality",
                "volume": "11",
                "issue": "3",
                "pages": "273--294",
                "other_ids": {
                    "DOI": [
                        "10.1016/0092-6566(77)90037-X"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "James Russell and Albert Mehrabian. 1977. Evidence for a three-factor theory of emotions. Journal of Re- search in Personality, 11(3):273-294.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "What are emotions? and how can they be measured?",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Klaus",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Scherer",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Social Science Information",
                "volume": "44",
                "issue": "4",
                "pages": "695--729",
                "other_ids": {
                    "DOI": [
                        "10.1177/0539018405058216"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Klaus R. Scherer. 2005. What are emotions? and how can they be measured? Social Science Information, 44(4):695-729.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Emotions are emergent processes: they require a dynamic computational architecture",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Klaus",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Scherer",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Philosophical Transactions of the Royal Society B",
                "volume": "364",
                "issue": "",
                "pages": "3459--3474",
                "other_ids": {
                    "DOI": [
                        "10.1098/rstb.2009.0141"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Klaus R. Scherer. 2009. Emotions are emergent pro- cesses: they require a dynamic computational archi- tecture. Philosophical Transactions of the Royal So- ciety B, 364(1535):3459-3474.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Evidence for universality and cultural variation of differential emotion response patterning",
                "authors": [
                    {
                        "first": "Klaus",
                        "middle": [
                            "R"
                        ],
                        "last": "Scherer",
                        "suffix": ""
                    },
                    {
                        "first": "Harald",
                        "middle": [
                            "G"
                        ],
                        "last": "Wallbott",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "Journal of personality and social psychology",
                "volume": "66",
                "issue": "2",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.1037//0022-3514.66.2.310"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Klaus R. Scherer and Harald G. Wallbott. 1994. Evi- dence for universality and cultural variation of differ- ential emotion response patterning. Journal of per- sonality and social psychology, 66(2):310.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "A linguistic interpretation of the OCC emotion model for affect sensing from text",
                "authors": [
                    {
                        "first": "Mostafa Al Masum",
                        "middle": [],
                        "last": "Shaikh",
                        "suffix": ""
                    },
                    {
                        "first": "Helmut",
                        "middle": [],
                        "last": "Prendinger",
                        "suffix": ""
                    },
                    {
                        "first": "Mitsuru",
                        "middle": [],
                        "last": "Ishizuka",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Affective Information Processing",
                "volume": "",
                "issue": "",
                "pages": "45--73",
                "other_ids": {
                    "DOI": [
                        "10.1007/978-1-84800-306-4_4"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Mostafa Al Masum Shaikh, Helmut Prendinger, and Mitsuru Ishizuka. 2009. A linguistic interpretation of the OCC emotion model for affect sensing from text. Affective Information Processing, pages 45-73.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Patterns of cognitive appraisal in emotion",
                "authors": [
                    {
                        "first": "Craig",
                        "middle": [
                            "A"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "Phoebe",
                        "middle": [
                            "C"
                        ],
                        "last": "Ellsworth",
                        "suffix": ""
                    }
                ],
                "year": 1985,
                "venue": "Journal of personality and social psychology",
                "volume": "48",
                "issue": "4",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Craig A. Smith and Phoebe C. Ellsworth. 1985. Pat- terns of cognitive appraisal in emotion. Journal of personality and social psychology, 48(4):813.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Twitter sentiment analysis using combined LSTM-CNN models",
                "authors": [
                    {
                        "first": "Pedro",
                        "middle": [
                            "M"
                        ],
                        "last": "Sosa",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "1--9",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Pedro M. Sosa. 2017. Twitter sentiment analysis using combined LSTM-CNN models. Eprint Arxiv, pages 1-9.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "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 Pro- ceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 4005- 4011, Florence, Italy. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "A rule-based approach to implicit emotion detection in text",
                "authors": [
                    {
                        "first": "Orizu",
                        "middle": [],
                        "last": "Udochukwu",
                        "suffix": ""
                    },
                    {
                        "first": "Yulan",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Natural Language Processing and Information Systems",
                "volume": "",
                "issue": "",
                "pages": "197--203",
                "other_ids": {
                    "DOI": [
                        "10.1007/978-3-319-19581-0_17"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Orizu Udochukwu and Yulan He. 2015. A rule-based approach to implicit emotion detection in text. In Natural Language Processing and Information Sys- tems, pages 197-203, Cham. Springer International Publishing.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "text": "Neural Model Architectures (subset)",
                "uris": null,
                "num": null,
                "type_str": "figure"
            },
            "TABREF0": {
                "text": "Excerpt of the final annotation guidelines including examples from TEC.",
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table><tr><td>Component</td><td colspan=\"2\">round 1 round 2</td></tr><tr><td>Cognitive appraisal</td><td>0.288</td><td>0.777</td></tr><tr><td>Neurophysiological symptoms</td><td>0.459</td><td>-</td></tr><tr><td>Motiv. Action tendencies</td><td>0.444</td><td>0.732</td></tr><tr><td>Motor expressions</td><td>0.643</td><td>0.617</td></tr><tr><td>Subjective feelings</td><td>0.733</td><td>0.793</td></tr></table>"
            },
            "TABREF1": {
                "text": "Inter-annotator agreement after the different annotation rounds during the guideline creation process measured with Cohen's \u03ba. In the second round, no annotator detected the neurophysiological component in the sample instances.",
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table/>"
            },
            "TABREF3": {
                "text": "",
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table/>"
            },
            "TABREF5": {
                "text": "",
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table/>"
            },
            "TABREF8": {
                "text": "F 1 (/100) results across models and emotion categories. (empty cells denote that this category is not available in the respective corpus. The best scores (except the gold setting) are printed bold face.",
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table><tr><td>observe an improvement by 1pp, to .55 F 1 . The</td></tr><tr><td>inclusion of predicted component information as</td></tr><tr><td>features in the neural network model shows no im-</td></tr><tr><td>provement on REMAN or on TEC.</td></tr></table>"
            },
            "TABREF10": {
                "text": "present the complete results for the neural network, including precision and recall values.",
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table><tr><td/><td/><td colspan=\"9\">Emo-NN-Base Emo-Cpm-NN-Gold Emo-Cpm-NN-Pred</td><td/><td>MTL-MH</td><td/><td>MTL-XS</td></tr><tr><td/><td>Emotion</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td/><td>Anger</td><td colspan=\"3\">28 50 36</td><td>47</td><td>70</td><td>56</td><td colspan=\"2\">33 30</td><td>32</td><td colspan=\"3\">31 40 35 31</td><td>50</td><td>38</td></tr><tr><td/><td colspan=\"4\">Anticipation 18 18 18</td><td>19</td><td>27</td><td>22</td><td>0</td><td>0</td><td>0</td><td colspan=\"3\">12 27 16 17</td><td>36</td><td>24</td></tr><tr><td/><td>Disgust</td><td colspan=\"3\">20 56 29</td><td>20</td><td>44</td><td>28</td><td colspan=\"2\">24 56</td><td>33</td><td colspan=\"3\">16 56 24 18</td><td>44</td><td>26</td></tr><tr><td>REMAN</td><td>Fear Joy Neutral Other</td><td colspan=\"5\">35 50 41 47 77 59 40 55 46 100 55 25 71 74 64 33 9 14 50 9</td><td>37 68 71 15</td><td colspan=\"2\">33 36 70 73 29 64 17 18</td><td>34 71 40 17</td><td colspan=\"3\">28 64 39 40 65 59 62 57 35 82 49 38 15 45 22 29</td><td>57 73 91 55</td><td>47 64 54 37</td></tr><tr><td/><td>Sadness</td><td colspan=\"3\">27 53 36</td><td>31</td><td>53</td><td>39</td><td colspan=\"2\">50 53</td><td>52</td><td colspan=\"3\">37 67 48 44</td><td>53</td><td>48</td></tr><tr><td/><td>Surprise</td><td colspan=\"3\">65 79 71</td><td>41</td><td>64</td><td>50</td><td colspan=\"2\">53 64</td><td>58</td><td colspan=\"5\">55 86 67 47 100 64</td></tr><tr><td/><td>Trust</td><td colspan=\"3\">39 69 50</td><td>86</td><td>46</td><td>60</td><td colspan=\"2\">67 31</td><td>42</td><td colspan=\"3\">43 77 56 50</td><td>62</td><td>55</td></tr><tr><td/><td>Macro avg.</td><td colspan=\"3\">35 52 40</td><td>49</td><td>50</td><td>45</td><td colspan=\"2\">38 42</td><td>38</td><td colspan=\"3\">34 60 42 37</td><td>62</td><td>46</td></tr><tr><td/><td>Micro avg.</td><td/><td/><td>43</td><td/><td/><td>45</td><td/><td/><td>43</td><td/><td>42</td><td/><td/><td>47</td></tr><tr><td/><td>Anger</td><td colspan=\"3\">50 35 41</td><td>57</td><td>47</td><td>52</td><td colspan=\"2\">30 35</td><td>32</td><td colspan=\"3\">29 12 17 42</td><td>29</td><td>34</td></tr><tr><td/><td>Disgust</td><td colspan=\"3\">40 50 44</td><td>50</td><td>25</td><td>33</td><td>0</td><td>0</td><td>0</td><td colspan=\"3\">67 50 57 50</td><td>50</td><td>50</td></tr><tr><td/><td>Fear</td><td colspan=\"3\">65 50 56</td><td>86</td><td>55</td><td>67</td><td colspan=\"2\">73 50</td><td>59</td><td colspan=\"3\">48 59 53 54</td><td>68</td><td>60</td></tr><tr><td>TEC</td><td>Joy Sadness Surprise</td><td colspan=\"3\">60 82 69 57 47 51 48 32 39</td><td>68 61 45</td><td>78 58 50</td><td>72 60 47</td><td colspan=\"2\">67 72 61 47 40 50</td><td>70 53 44</td><td colspan=\"3\">79 74 76 66 66 44 53 61 36 62 45 60</td><td>82 53 35</td><td>73 57 44</td></tr><tr><td/><td>Macro avg.</td><td colspan=\"3\">53 49 50</td><td>61</td><td>52</td><td>55</td><td colspan=\"2\">45 42</td><td>43</td><td colspan=\"3\">54 50 50 55</td><td>53</td><td>53</td></tr><tr><td/><td>Micro avg.</td><td/><td/><td>57</td><td/><td/><td>62</td><td/><td/><td>56</td><td/><td>58</td><td/><td/><td>61</td></tr></table>"
            },
            "TABREF11": {
                "text": "Performance of the neural network emotion classifiers. The highest F 1 scores are printed bold face.",
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table/>"
            }
        }
    }
}