Emotion Recognition under Consideration of the Emotion Component Process Model

Emotion Recognition under Consideration of the Emotion Component Process Model Felix Casel∗, Amelie Heindl∗, and Roman Klinger Institut f¨ur Maschinelle Sprachverarbeitung, University of Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {firstname.lastname}@ims.uni-stuttgart.de Abstract Emotion classification in text is typically per- formed with neural network models which learn to associate linguistic units with emo- tions. While this often leads to good predic- tive performance, it does only help to a limited degree to understand how emotions are com- municated in various domains. The emotion component process model (CPM) by Scherer (2005) is an interesting approach to explain emotion communication. It states that emo- tions are a coordinated process of various sub- components, in reaction to an event, namely the subjective feeling, the cognitive appraisal, the expression, a physiological bodily reac- tion, and a motivational action tendency. We hypothesize that these components are associ- ated with linguistic realizations: an emotion can be expressed by describing a physiologi- cal bodily reaction (“he was trembling”), or the expression (“she smiled”), etc. We an- notate existing literature and Twitter emotion corpora with emotion component classes and find that emotions on Twitter are predomi- nantly expressed by event descriptions or sub- jective reports of the feeling, while in liter- ature, authors prefer to describe what char- acters do, and leave the interpretation to the reader. We further include the CPM in a mul- titask learning model and find that this sup- ports the emotion categorization. The anno- tated corpora are available at https://www. ims.uni-stuttgart.de/data/emotion. 1 Introduction The task of emotion classification from written text is to map textual units, like documents, paragraphs, or sentences, to a predefined set of emotions. Com- mon class inventories rely on psychological the- ories such as those proposed by Ekman (1992) (anger, disgust, fear, joy, sadness, surprise) or ∗The first two authors contributed equally to this work. 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)). While end-to-end learning and fine-tuning of pre- trained 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 communi- cated). There are only very few approaches that aim at combining psychological theories (beyond basic emotion categories) with emotion classifica- tion models: We are only aware of the work by Hofmann et al. (2020), who incorporate the cogni- tive appraisal of events, and Buechel et al. (2020), who jointly learn affect (valence, arousal) and emo- tion classes; next to knowledge-base-oriented mod- elling of events by Balahur et al. (2012) and Cam- bria et al. (2014). An interesting and attractive theory for computa- tional 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 …

Content truncated to 500 words