Supervised Prototypical Contrastive Learning for Emotion Recognition in Conversation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5197 - 5206 December 7-11, 2022 ©2022 Association for Computational Linguistics Supervised Prototypical Contrastive Learning for Emotion Recognition in Conversation Xiaohui Song1,2⋆, Longtao Huang3, Hui Xue3 and Songlin Hu1,2† 1Institute of Information Engineering, Chinese Academy of Sciences 2School of Cyber Security, University of Chinese Academy of Sciences 3Alibaba Group {songxiaohui,husonglin}@iie.ac.cn, {kaiyang.hlt, hui.xueh}@alibaba-inc.com Abstract Capturing emotions within a conversation plays an essential role in modern dialogue systems. However, the weak correlation between emo- tions and semantics brings many challenges to emotion recognition in conversation (ERC). Even semantically similar utterances, the emo- tion may vary drastically depending on con- texts or speakers. In this paper, we propose a Supervised Prototypical Contrastive Learning (SPCL) loss for the ERC task. Leveraging the Prototypical Network, the SPCL targets at solving the imbalanced classification problem through contrastive learning and does not re- quire a large batch size. Meanwhile, we design a difficulty measure function based on the dis- tance between classes and introduce curricu- lum learning to alleviate the impact of extreme samples. We achieve state-of-the-art results on three widely used benchmarks. Further, we conduct analytical experiments to demonstrate the effectiveness of our proposed SPCL and curriculum learning strategy. We release the code at https://github.com/caskcsg/SPCL. 1 Introduction With the development of online social networks, capturing emotions during conversations has gained increasing attention in both academia and industry(Li et al., 2020; Shen et al., 2021; Wang et al., 2020; Ghosal et al., 2020; Song et al., 2022; Zhu et al., 2021). Emotion recognition in conver- sation (ERC) is critical in many scenarios, such as chatbots, healthcare applications, mining opin- ions on social media, and so on(Poria et al., 2019b). The ERC task aims to identify different emotions at each turn within a conversation based on the transcript. A conversation often contains several speakers and runs several turns; thus, emotions can vary drastically during the conversation. Compared to traditional text classification tasks, figuring out ⋆Work done during internship at Alibaba Group. † Corresponding Author. Absolutely, oh I promise. Tell her what? I don't care. I am trying to get a person out of my body here, and you're not making it any easier. Now go! Thanks a lot. Listen, can you promise me that you won’t tell her though? Carol Ross Pete Phoebe Thanks a lot. Pete Figure 1: Examples of emotion recognition in conver- sation. The same utterance "Thanks a lot" can convey different emotions in different contexts. emotions needs not only one turn of textual utter- ance but also contextual information. An example of ERC is illustrated in Figure 1. Contrastive learning applied to self-supervised representation learning has seen a resurgence in recent years. Khosla et al. (2020) extended the self-supervised batch contrastive approach to the fully-supervised setting and show outperformance over cross-entropy loss in several benchmarks. Al- though CoG-BART(Li et al., 2021) has demon- strated the effectiveness of supervised contrastive learning (SCL) in the ERC task, there are still two issues worth …

Content truncated to 500 words