DualGATs: Dual Graph Attention Networks for Emotion Recognition in Conversations

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7395–7408 July 9-14, 2023 ©2023 Association for Computational Linguistics DualGATs: Dual Graph Attention Networks for Emotion Recognition in Conversations Duzhen Zhang Baidu Inc, Beijing, China zhangduzhen@baidu.com Feilong Chen Huawei Inc, Beijing, China chenfeilong10@huawei.com Xiuyi Chen∗ Baidu Inc, Beijing, China chenxiuyi01@baidu.com Abstract Capturing complex contextual dependencies plays a vital role in Emotion Recognition in Conversations (ERC). Previous studies have predominantly focused on speaker-aware con- text modeling, overlooking the discourse struc- ture of the conversation. In this paper, we in- troduce Dual Graph ATtention networks (Dual- GATs) to concurrently consider the comple- mentary aspects of discourse structure and speaker-aware context, aiming for more pre- cise ERC. Specifically, we devise a Discourse- aware GAT (DisGAT) module to incorporate discourse structural information by analyz- ing the discourse dependencies between utter- ances. Additionally, we develop a Speaker- aware GAT (SpkGAT) module to incorporate speaker-aware contextual information by con- sidering the speaker dependencies between ut- terances. Furthermore, we design an interac- tion module that facilitates the integration of the DisGAT and SpkGAT modules, enabling the effective interchange of relevant informa- tion between the two modules. We extensively evaluate our method on four datasets, and exper- imental results demonstrate that our proposed DualGATs surpass state-of-the-art baselines on the majority of the datasets.1 1 Introduction With the increasing availability of conversational data on social media platforms (Poria et al., 2019a), Emotion Recognition in Conversations (ERC) has emerged as a popular research topic (Poria et al., 2019b). Its objective is to identify and track the emotional state of each utterance. ERC plays a cru- cial role in various applications, including opinion mining in social media (Chatterjee et al., 2019b) and the development of empathetic dialogue sys- tems that can analyze user emotional states and gen- ∗Corresponding author. 1Available at https://github.com/BladeDancer957/ DualGATs. erate emotion-aware responses (Zhou et al., 2018; Liu et al., 2021; Peng et al., 2022, 2023). However, analyzing emotions in conversations poses significant challenges. Unlike emotion recog- nition in isolated sentences (Seyeditabari et al., 2018), ERC requires careful consideration of con- textual dependencies. Previous ERC methods have primarily focused on capturing speaker or tempo- ral dependencies between utterances, making the modeling of speaker-aware context central to these approaches (Majumder et al., 2019). To incorporate speaker-aware contextual infor- mation, numerous methods have been proposed to model conversations as sequences (Poria et al., 2017; Hazarika et al., 2018a,b; Jiao et al., 2019; Hu et al., 2021; Ong et al., 2022) or graphs (Ghosal et al., 2019; Ishiwatari et al., 2020; Shen et al., 2021b; Li et al., 2022). Sequence-based methods capture sequential information by encoding utter- ances temporally using Recurrent Neural Networks (RNNs). Majumder et al. 2019 designed an inde- pendent Gated Recurrent Unit (GRU) (Cho et al., 2014) to track the emotional state of the speaker. However, these sequence-based methods often rely on limited information from nearby utterances to update the current utterance’s representation, mak- ing it challenging to capture distant contextual …

Content truncated to 500 words