DA-GCN: A Dependency-Aware Graph Convolutional Network for Emotion Recognition in Conversations

Yunhe Xie, Chengjie Sun, Bingquan Liu, Zhenzhou Ji

Abstract: Emotion Recognition in Conversations (ERC) has recently gained much attention from the NLP community. The contextual information and the dependency information are two key factors that contribute to the ERC task. Unfortunately, most of the existing approaches concentrate on mining contextual information while neglecting the dependency information. To address this problem, we propose a Dependency-Aware Graph Convolutional Network (DA-GCN) to jointly take advantage of these two kinds of information. The core module is a proposed dependency-aware graph interaction layer where a GCN is constructed and operates directly on the dependency tree of the utterance, achieving to consider the dependency information. In addition, the proposed layer can be stacked to further enhance the embeddings with multiple steps of propagation. Experimental results on three datasets show that our model achieves the state-of-the-art performance. Furthermore, comprehensive analysis empirically verifies the effectiveness of leveraging the dependency information and the multi-step propagation mechanism.