S+PAGE: A Speaker and Position-Aware Graph Neural Network Model for Emotion Recognition in Conversation

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 148–157 November 20–23, 2022. ©2022 Association for Computational Linguistics 148 S+PAGE: A Speaker and Position-Aware Graph Neural Network Model for Emotion Recognition in Conversation Chen Liang, Jing Xu, Yangkun Lin, Chong Yang∗, Yongliang Wang Ant Group Hangzhou, China {liangchen.liangche, jill.xj, linyangkun.lyk, yangchong.yang, yongliang.wyl}@antgroup.com Abstract Emotion recognition in conversation (ERC) has attracted much attention in recent years for its necessity in widespread applications. With the development of graph neural network (GNN), recent state-of-the-art ERC models mostly use GNN to embed the intrinsic structure informa- tion of a conversation into the utterance fea- tures. In this paper, we propose a novel GNN- based model for ERC, namely S+PAGE, to bet- ter capture the speaker and position-aware con- versation structure information. Specifically, we add the relative positional encoding and speaker dependency encoding in the represen- tations of edge weights and edge types respec- tively to acquire a more reasonable aggregation algorithm for ERC. Besides, a two-stream con- versational Transformer is presented to extract both the self and inter-speaker contextual fea- tures for each utterance. Extensive experiments are conducted on four ERC benchmarks with state-of-the-art models employed as baselines for comparison, whose results demonstrate the superiority of our model. 1 Introduction Emotion recognition in conversation (ERC), which aims to identify the emotion of each utterance in a conversation, is a task arousing increasing inter- ests in many fields. With the prevalence of social media and intelligent assistants, ERC has great potential applications in several areas, such as emo- tional chatbots, sentiment analysis of comments in social media and healthcare intelligence, for understanding emotions in the conversation with emotion dynamics and generating emotionally co- herent responses. ERC problem still remains a challenge. Both lexicon-based (Wu et al., 2006; Shaheen et al., 2014) and deep learning-based (Col- neriˇc and Demšar, 2018) text emotion recogni- tion methods that treat each utterance individu- ∗ Corresponding author. A good friend of mine passed away the other day. [sad] Speaker A Speaker B Oh he had some pretty progressive cancer. [sad] Yeah, he was pretty awesome and supportive. Oh no, I'm so sorry. What happened? [sad] Did you know him well? [neutral] [happy] [sad] Figure 1: A dialogue from IEMOPCAP, in which the emotion of the last utterance by speaker A will be wrongly classified if the dialogue context is not taken into consideration. ally fail in this task as these works ignore some conversation-specific characteristics. In the past few years, recurrent neural network (RNN)-based solutions, such as CMN (Hazarika et al., 2018b), ICON (Hazarika et al., 2018a) and DialogueRNN (Majumder et al., 2019), have dom- inated this field due to the sequential nature of conversational context. Nonetheless, they share some inherent limitations: 1) RNN model performs poorly in grasping distant contextual information; 2) RNN-based methods are not capable of handling large-scale multiparty conversations. With the rise of graph neural …

Content truncated to 500 words