Dimensional Emotion Detection from Categorical Emotion

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4367–4380 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 4367 Dimensional Emotion Detection from Categorical Emotion Sungjoon Park1 Jiseon Kim1 Seonghyeon Ye1 Jaeyeol Jeon2 Hee Young Park3 Alice Oh1 1 School of Computing, KAIST, Republic of Korea 2 Upstage AI Research, Upstage, Republic of Korea 3 Department of Psychology, Seoul National University, Republic of Korea {sungjoon.park, jiseon_kim, vano1205}@kaist.ac.kr, jaeyeol.jeon@upstage.ai, heeyoungpark@snu.ac.kr, alice.oh@kaist.edu Abstract We present a model to predict fine-grained emotions along the continuous dimensions of valence, arousal, and dominance (VAD) with a corpus with categorical emotion annotations. Our model is trained by minimizing the EMD (Earth Mover’s Distance) loss between the pre- dicted VAD score distribution and the categor- ical emotion distributions sorted along VAD, and it can simultaneously classify the emo- tion categories and predict the VAD scores for a given sentence. We use pre-trained RoBERTa-Large and fine-tune on three differ- ent corpora with categorical labels and evalu- ate on EmoBank corpus with VAD scores. We show that our approach reaches comparable performance to that of the state-of-the-art clas- sifiers in categorical emotion classification and shows significant positive correlations with the ground truth VAD scores. Also, further train- ing with supervision of VAD labels leads to im- proved performance especially when dataset is small. We also present examples of predic- tions of appropriate emotion words that are not part of the original annotations. 1 Introduction In psychology literature, basic emotions are catego- rized as happy, sad, angry and so on (Ekman, 1992; Plutchik, 2001), however, we can feel and express more subtle and complex emotions beyond them. They can be systematically represented with the Valence-Arousal-Dominance (VAD) model which maps emotional states to 3-dimensional continu- ous VAD space. This space allows various emo- tions to be projected into the space with mea- surable distances from one another (Russell and Mehrabian, 1977), covering a wider range of sub- tle emotions compared to the categorical mod- els with a finite set of basic emotions. Captur- ing such fine-grained emotions with dimensional VAD models could benefit clinical natural language processing (NLP) (Desmet and Hoste, 2013; Sa- hana and Girish, 2015), emotion regulation such as psychotherapy (Torre and Lieberman, 2018). For example, analyzing the client’s utterance and ac- knowledging the negative emotion as ‘neglected’ rather than ‘sad’, which is known as ‘affect label- ing’, would reduce negative physiological, behav- ioral, and psychological responses resulting from that emotional state. Thus developing a dimensional emotion detec- tion model would be very useful, but one problem is a lack of required annotated resources. There is a relatively small sentence-level corpus with full VAD annotations (Buechel and Hahn, 2017), and a corpus annotated with V and A dimensions (Preo¸tiuc-Pietro et al., 2016a; Yu et al., 2016a), and only with V (Lykousas et al., 2019). We could build additional resources by labeling VAD scores by Best-Worst Scaling (Kiritchenko and Moham- mad, 2017). Instead, we approach this problem with a novel and more efficient method to predict …

Content truncated to 500 words