MuCDN: Mutual Conversational Detachment Network for Emotion Recognition in Multi-Party Conversations

Proceedings of the 29th International Conference on Computational Linguistics, pages 7020–7030 October 12–17, 2022. 7020 MuCDN: Mutual Conversational Detachment Network for Emotion Recognition in Multi-Party Conversations Weixiang Zhao, Yanyan Zhao∗, Bing Qin Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {wxzhao, yyzhao, qinb}@ir.hit.edu.cn Abstract As an emerging research topic in natural lan- guage processing community, emotion recog- nition in multi-party conversations has attained increasing interest. Previous approaches that focus either on dyadic or multi-party scenarios exert much effort to cope with the challenge of emotional dynamics and achieve appealing results. However, since emotional interactions among speakers are often more complicated within the entangled multi-party conversations, these works are limited in capturing effective emotional clues in conversational context. In this work, we propose Mutual Conversational Detachment Network (MuCDN) to clearly un- derstand the conversational context by separat- ing conversations into detached threads. Specif- ically, two detachment ways are devised to perform context and speaker-specific modeling within detached threads and they are bridged through a mutual module. Experimental results on two datasets show that our model achieves better performance over the baseline models. 1 Introduction Emotion recognition in conversations (ERC) is a task that predicts the emotion for each utterance in conversations. Since it plays a significant role in achieving empathetic systems and is of great val- ues to be applied in the fields of opinion mining in conversations, social media analysis, mental health care and other areas, ERC has received more and more attention in the natural language processing (NLP) community. We focus on emotion recogni- tion in multi-party conversations (ERMC) where three or more speakers are involved. The challenge of ERC, especially for that of ERMC, lies in a complicated emotional interac- tion referred to as emotional dynamics (Poria et al., 2019b) that utterances from one speaker would have impact on expressions of others. To cope ∗Corresponding author Figure 1: An example of a multi-party conversation from MELD dataset. with the challenge, many previous approaches ex- ploit recurrent neural network (RNN) (Majumder et al., 2019; Ghosal et al., 2020) and graph neural network (GNN) (Ghosal et al., 2019; Shen et al., 2021b; Zhang et al., 2022) to perform context and speaker-specific modeling. However, they are not effective enough to capture exact emotional clues for emotion recognition because the salient charac- teristic of multi-party conversations is ignored. Multiple conversation threads are intermingled in the same dialog history simultaneously in the multi-party scenario (Ma et al., 2021; Liu et al., 2021a), which makes it hard to capture emotional clues in such a complex and confusing interaction. As shown in Figure 1, there are two conversation threads in a multi-party conversation consisting of three speakers A, B and C. Targeting at the same utterance #2 from speaker B to reply, speaker A is consoled and becomes joyful to go shopping, while 7021 speaker C is unbearable towards the outfit of B and feels disgusted. Since speaker C and A originally focus on different aspects of the same dialog …

Content truncated to 500 words