DialogueGCN: A Graph Convolutional Neural Network for Emotion Recognition in Conversation

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 154–164, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 154 DialogueGCN: A Graph Convolutional Neural Network for Emotion Recognition in Conversation Deepanway Ghosal†, Navonil Majumder‡, Soujanya Poria†∗, Niyati Chhaya∇ and Alexander Gelbukh‡ † Singapore University of Technology and Design, Singapore ‡Instituto Polit´ecnico Nacional, CIC, Mexico ∇ Adobe Research, India {1004721@mymail.,sporia@}sutd.edu.sg, navo@nlp.cic.ipn.mx, nchhaya@adobe.com, gelbukh@gelbukh.com Abstract Emotion recognition in conversation (ERC) has received much attention, lately, from re- searchers due to its potential widespread ap- plications in diverse areas, such as health-care, education, and human resources. In this pa- per, we present Dialogue Graph Convolutional Network (DialogueGCN), a graph neural net- work based approach to ERC. We leverage self and inter-speaker dependency of the interlocu- tors to model conversational context for emo- tion recognition. Through the graph network, DialogueGCN addresses context propagation issues present in the current RNN-based meth- ods. We empirically show that this method al- leviates such issues, while outperforming the current state of the art on a number of bench- mark emotion classification datasets. 1 Introduction Emotion recognition has remained an active re- search topic for decades (K. D’Mello et al., 2006; Busso et al., 2008; Strapparava and Mihalcea, 2010). However, the recent proliferation of open conversational data on social media platforms, such as Facebook, Twitter, Youtube, and Red- dit, has warranted serious attention (Poria et al., 2019b; Majumder et al., 2019; Huang et al., 2019) from researchers towards emotion recognition in conversation (ERC). ERC is also undeniably im- portant in affective dialogue systems (as shown in Fig. 1) where bots understand users’ emotions and sentiment to generate emotionally coherent and empathetic responses. Recent works on ERC process the constituent utterances of a dialogue in sequence, with a re- current neural network (RNN). Such a scheme is illustrated in Fig. 2 (Poria et al., 2019b), that re- lies on propagating contextual and sequential in- formation to the utterances. Hence, we feed the ∗ Corresponding author My head is aching Frustrated I’m sorry to hear that! Sad Did you suffer a head injury earlier? Neutral Yes Neutral Please, immediately see a nearby doctor Excited No Neutral Pleased to hear that Happy Did you consume alcohol recently? Neutral User Health Assistant You might be concussed! Sad Figure 1: Illustration of an affective conversation where the emotion depends on the context. Health assistant understands affective state of the user in order to gen- erate affective and empathetic responses. conversation to a bidirectional gated recurrent unit (GRU) (Chung et al., 2014). However, like most of the current models, we also ignore intent mod- elling, topic, and personality due to lack of la- belling on those aspects in the benchmark datasets. In theory, RNNs like long short-term memory (LSTM) (Hochreiter and Schmidhuber, 1997) and GRU should propagate long-term contextual infor- mation. However, in practice it is not always the case (Bradbury et al., 2017). This affects the ef- …

Content truncated to 500 words