Conversational Emotion-Cause Pair Extraction with Guided Mixture of Experts

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3288–3298 May 2-6, 2023 ©2023 Association for Computational Linguistics Conversational Emotion-Cause Pair Extraction with Guided Mixture of Experts DongJin Jeong Sungkyunkwan University Department of Artificial Intelligence South Korea jdjin3000@g.skku.edu JinYeong Bak Sungkyunkwan University Department of Artificial Intelligence South Korea jy.bak@skku.edu Abstract Emotion-Cause Pair Extraction (ECPE) task aims to pair all emotions and corresponding causes in documents. ECPE is an impor- tant task for developing human-like responses. However, previous ECPE research is conducted based on news articles, which has different characteristics compared to dialogues. To ad- dress this issue, we propose a Pair-Relationship Guided Mixture-of-Experts (PRG-MoE) model, which considers dialogue features (e.g., speaker information). PRG-MoE automatically learns relationship between utterances and advises a gating network to incorporate dialogue fea- tures in the evaluation, yielding substantial per- formance improvement. We employ a new ECPE dataset, which is an English dialogue dataset, with more emotion-cause pairs in documents than news articles. We also pro- pose Cause Type Classification that classifies emotion-cause pairs according to the types of the cause of a detected emotion. For reproduc- ing the results, we make available all our code and data1. 1 Introduction With increased interest in developing human-like responses, it is crucial to determine the cause of a given emotion. As part of such interest, there is a surge of research activities that analyze the cause of emotions (Yan et al., 2021; Turcan et al., 2021; Li et al., 2022a). Recently, Poria et al. (2021) presents RECCON, a new dataset for Emotion Cause Ex- traction (ECE) task in dialogue. ECE is a task to find a clause that contains the cause of an annotated emotion in a clause of a given document. However, ECE is limited in that the model requires manually annotated emotions. To overcome the limitation of ECE (Lee et al., 2010), Xia and Ding (2019) suggest Emotion Cause Pair Extraction (ECPE) task, which automatically predicts emotion clauses in a given document and 1https://github.com/jdjin3000/PRG-MoE # of Emotion-Cause Pairs ECPE-news (Xia and Ding, 2019) ECPE-D 1 1,746 (89.77%) 9 (0.80%) 2 177 (9.10%) 19 (1.69%) ≥ 3 22 (1.13%) 1,094 (97.50%) Table 1: The amount of emotion-cause pairs in a docu- ment compared with the ECPE-news corpus (Xia and Ding, 2019) and ECPE-D corpus. ECPE-D is a dialogue dataset reconstructed based on RECCON (Poria et al., 2021). On average, ECPE-D corpus has more emotion- cause pairs than ECPE-news corpus. identifies their corresponding causes. They also build a new ECPE corpus from Chinese news arti- cles. Since ECPE and the dataset were proposed, it has attracted the interest of numerous researchers Ding et al. (2020a,b); Wei et al. (2020); Fan et al. (2020); Cheng et al. (2020); Chen et al. (2020a,b). However, ECPE in dialogues is different from ECPE in news articles. A dialogue is an interaction between two or more people, while a news article describes a fact. So, dialogues contains meta in- formation such as the speakers, which is …

Content truncated to 500 words