Topic-Driven and Knowledge-Aware Transformer for Dialogue Emotion Detection

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1571–1582 August 1–6, 2021. ©2021 Association for Computational Linguistics 1571 Topic-Driven and Knowledge-Aware Transformer for Dialogue Emotion Detection Lixing Zhu†, Gabriele Pergola†, Lin Gui†, Deyu Zhou§, Yulan He† †Department of Computer Science, University of Warwick, UK §School of Computer Science and Engineering, Key Laboratory of Computer Network and Information Integration, Ministry of Education, Southeast University, China {lixing.zhu,gabriele.pergola,lin.gui,yulan.he}@warwick.ac.uk d.zhou@seu.edu.cn Abstract Emotion detection in dialogues is challenging as it often requires the identification of the- matic topics underlying a conversation, the rel- evant commonsense knowledge, and the intri- cate transition patterns between the affective states. In this paper, we propose a Topic- Driven Knowledge-Aware Transformer to han- dle the challenges above. We firstly design a topic-augmented language model (LM) with an additional layer specialized for topic detec- tion. The topic-augmented LM is then com- bined with commonsense statements derived from a knowledge base based on the dialogue contextual information. Finally, a transformer- based encoder-decoder architecture fuses the topical and commonsense information, and performs the emotion label sequence predic- tion. The model has been experimented on four datasets in dialogue emotion detection, demonstrating its superiority empirically over the existing state-of-the-art approaches. Quan- titative and qualitative results show that the model can discover topics which help in dis- tinguishing emotion categories. 1 Introduction The abundance in dialogues extracted from on- line conversations and TV series provides unprece- dented opportunity to train models for automatic emotion detection, which are important for the de- velopment of empathetic conversational agents or chat bots for psychotherapy (Hsu and Ku, 2018; Jiao et al., 2019; Zhang et al., 2019; Cao et al., 2019). However, it is challenging to capture the contextual semantics of personal experience de- scribed in one’s utterance. For example, the emo- tion of the sentence “I just passed the exam” can be either happy or sad depending on the expectation of the subject. There are strands of works utilizing the dialogue context to enhance the utterance rep- resentation (Jiao et al., 2019; Zhang et al., 2019; (a): Food and Restaurant (b): Marriage and Death A: Could I have some fish? B: Certainly. And what vegetables would you like? A: Oh , spinach , I think. A: I like drinking tea at teahouses. B: Oh, so do I. ☺ B: Great. We can chat while enjoying a cup there. A: Why don't we go for one now? ☺ A: Let's go! ☺ ☺ A: Johnny died yesterday, we knew that it was coming, but... B: Like just last week, he was doing so well. ☹ A: Then all of a sudden they gave him a microphone, he asked me to marry him, like, onstage. B: He was doing so well. ☹ ☺ ☺ Figure 1: Utterances around particular topics carry spe- cific emotions. Utterances carrying positive (smiling face) or negative (crying face) emotions are highlighted in colour. Other utterances are labeled as ‘Neutral’. In …

Content truncated to 500 words