Relation-aware Graph Attention Networks with Relational Position Encodings for Emotion Recognition in Conversations

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7360–7370, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 7360 Relation-aware Graph Attention Networks with Relational Position Encodings for Emotion Recognition in Conversations Taichi Ishiwatari Yuki Yasuda Taro Miyazaki NHK Science and Technology Research LaboratoriesJun Goto {ishiwatari.t-fa, yasuda.y-hk, miyazaki.t-jw, goto.j-fw} @nhk.or.jp Abstract Interest in emotion recognition in conversa- tions (ERC) has been increasing in various fields, because it can be used to analyze user behaviors and detect fake news. Many re- cent ERC methods use graph-based neural net- works to take the relationships between the ut- terances of the speakers into account. In par- ticular, the state-of-the-art method considers self- and inter-speaker dependencies in con- versations by using relational graph attention networks (RGAT). However, graph-based neu- ral networks do not take sequential informa- tion into account. In this paper, we propose re- lational position encodings that provide RGAT with sequential information reflecting the rela- tional graph structure. Accordingly, our RGAT model can capture both the speaker depen- dency and the sequential information. Exper- iments on four ERC datasets show that our model is beneficial to recognizing emotions expressed in conversations. In addition, our approach empirically outperforms the state-of- the-art on all of the benchmark datasets. 1 Introduction Interest in emotion recognition in conversations (ERC) has been increasing in various fields (Pi- card, 2010), because it can be used to analyze user behaviors (Lee and Hong, 2016) and detect fake news (Guo et al., 2019). With the recent prolifer- ation of social media platforms such as Facebook, Twitter, and YouTube, as well as conversational assistants such as Amazon Alexa, there is a need to study how emotions are expressed in natural con- versation. Recent research on ERC processes the utter- ances of dialogues in sequence by using recurrent neural network (RNN)-based methods (Hochreiter and Schmidhuber, 1997; Chung et al., 2014; Liu et al., 2016). However, these methods are not ♯ Speaker Utterance Emotion 1 A I’m just so tired all the time. Sad 2 B Well have you been trying to get a job, look for a job or...? Neutral 3 A I’ve been looking for like eight months. Frustrated 4 B I know., It- It’s really tough out there., It’s really hard to find a job. Frustrated 5 A I’m tired of the same excuses., No, no you’re not qualified enough, wish you had more education. Frustrated 6 B Well what are you looking for?, I mean– Neutral 7 B Well, okay. Well that’s– Neutral 8 A Cause I went to Harvard. Anger Table 1: Example for contextual emotion analysis on the IEMOCAP dataset (Busso et al., 2008), which con- tains emotion-labeled utterances in multi-party conver- sations. able to process long series of information (Brad- bury et al., 2016). DialogueRNN tries to make up for this problem by using an attention mecha- nism to focus on the relevant utterances in the en- tire conversation (Majumder et al., 2019). How- ever, these methods do not take self-dependency or …

Content truncated to 500 words