Directed Acyclic Graph Network for Conversational Emotion Recognition

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1551–1560 August 1–6, 2021. ©2021 Association for Computational Linguistics 1551 Directed Acyclic Graph Network for Conversational Emotion Recognition Weizhou Shen, Siyue Wu, Yunyi Yang, Xiaojun Quan∗ School of Computer Science and Engineering, Sun Yat-sen University, China {shenwzh3, wusy39, yangyy37}@mail2.sysu.edu.cn quanxj3@mail.sysu.edu.cn Abstract The modeling of conversational context plays a vital role in emotion recognition from con- versation (ERC). In this paper, we put for- ward a novel idea of encoding the utterances with a directed acyclic graph (DAG) to better model the intrinsic structure within a conversa- tion, and design a directed acyclic neural net- work, namely DAG-ERC1, to implement this idea. In an attempt to combine the strengths of conventional graph-based neural models and recurrence-based neural models, DAG-ERC provides a more intuitive way to model the in- formation flow between long-distance conver- sation background and nearby context. Exten- sive experiments are conducted on four ERC benchmarks with state-of-the-art models em- ployed as baselines for comparison. The empir- ical results demonstrate the superiority of this new model and confirm the motivation of the directed acyclic graph architecture for ERC. 1 Introduction Utterance-level emotion recognition in conversa- tion (ERC) is an emerging task that aims to identify the emotion of each utterance in a conversation. This task has been recently concerned by a con- siderable number of NLP researchers due to its potential applications in several areas, such as opin- ion mining in social media (Chatterjee et al., 2019) and building an emotional and empathetic dialog system (Majumder et al., 2020). The emotion of a query utterance is likely to be influenced by many factors such as the utterances spoken by the same speaker and the surrounding conversation context. Indeed, how to model the conversational context lies at the heart of this task (Poria et al., 2019a). Empirical evidence also shows ∗Corresponding author. 1The code is available at https://github.com/ shenwzh3/DAG-ERC Figure 1: Conversation as a directed acyclic graph, with brown directed edges representing the information propagation between speakers and blue ones represent- ing the information propagation inside a same speaker. that a good representation of conversation context significantly contributes to the model performance, especially when the content of query utterance is too short to be identified alone (Ghosal et al., 2019). Numerous efforts have been devoted to the mod- eling of conversation context. Basically, they can be divided into two categories: graph-based meth- ods (Zhang et al., 2019a; Ghosal et al., 2019; Zhong et al., 2019; Ishiwatari et al., 2020; Shen et al., 2020) and recurrence-based methods (Hazarika et al., 2018a; Hazarika et al., 2018b; Majumder et al., 2019; Ghosal et al., 2020). For the graph- based methods, they concurrently gather informa- tion of the surrounding utterances within a certain window, while neglecting the distant utterances and the sequential information. For the recurrence- based methods, they consider the distant utterances and sequential information by encoding the utter- ances temporally. …

Content truncated to 500 words