MuSE: A Multi-Scale Emotional Flow Graph Model for Empathetic Dialogue Generation

Abstract

The purpose of empathetic dialogue generation is to fully understand the speakers’ emotional needs in dialogues and to generate appropriate empathetic responses. Existing works mainly focus on the overall coarse-grained emotion of the context while neglecting different utterances’ fine-grained emotions, which leads to the inability to detect the speakers’ fine-grained emotional changes during a conversation. However, in real-life dialogue scenarios, the speaker usually carries an initial emotional state that changes continuously during the conversation. Therefore, understanding a series of emotional states can help to better understand speakers’ emotions and generate empathetic responses. To address this issue, we propose a M u lti- S cale E motional flow model called MuSE , which simulates speakers’ emotional flow. First, we introduce a fine-grained expansion strategy to transform context into an emotional flow graph that combines multi-scale coarse and fine-grained information. This emotional flow graph captures speakers’ constant emotional changes at each turn of a conversation. And then, the emotion node and the situational node are introduced to the emotional flow graph respectively in order to extend the speakers’ initial emotion into the ensuing conversation. Finally, we conduct experiments on the public EMPATHETIC DIALOGUES dataset. The experimental results demonstrate that the MuSE model achieves superior performance under both automatic evaluation and human evaluation metrics compared with the existing baseline models. Our code is available at https://github.com/DericZhao/MuSE .