D-Man: a Distance-based Multi-channel Attention Network for ERC
Fefei Xu, Guangzhen Li, Zheng Zhong, Yingchen Zhou, Zhou Wang · 2024
Emotion recognition in conversation (ERC) has gained significant attention in the field of natural language processing (NLP) due to its diverse applications. The task of ERC is to recognize the emotion of each utterance within a conversation, and the modeling of conversational information plays a crucial role in achieving accurate results. Previous studies have not taken into account the effect of different distances between utterances, which leads to the problems of incomplete global information and the same importance for each utterance. In this paper, we propose a Distance-based Multi-channel Attention network (D-Man) for emotion recognition. We first design a distance-based multi-head attention mechanism that aims to efficiently capture conversational information with distance. By dynamically adjusting the attention between utterances according to the distance, we are not only able to deeply understand conversation on a global scale, but also assign corresponding attention weights to utterances with different distances, which makes the model focused on more important utterances. In addition, we decouple the conversation into temporal information, speaker-aware information, and discourse dependency information for fully mining the conversation semantics. Specifically, we design three distance matrices using three types of information: temporal sequence, speaker identity, and discourse dependency, and then feed them into distance-based multi-head attention to capture the three types of distance-based information in the time sequence channel, speaker channel, and dependency channel, respectively. We conduct extensive experiments on four benchmark datasets. The results show that our proposed model outperforms all baseline models. Ablation experiments further validate the effectiveness of each module.