MEGE: A mixed emotion graph model for empathetic dialogue generation

In daily life, emotions tend to exhibit amalgamated forms. For instance, when someone is involved in an interview, excitement and nervousness consistently coexist. In previous studies on empathetic dialogue, the mixed emotions in the conversation have not been taken into consideration. The neglect of the fact that human emotions in dialogue are intertwined with different intensities of various emotions, affects the accuracy of understanding the speaker's emotions. To address this issue, we model the mixed emotions and propose a Mixed Emotion Graph model for Empathetic dialogue generation, called MEGE. The MEGE model constructs dialogue graph structures to capture the dynamics of information flow in the conversation. We decompose the speaker's mixed emotions into multiple single-emotion channels across multiple parallel timelines. And then we propose a multi-channel convolutional fusion mechanism to fuse information from different channels. Each channel can be driven by customized external knowledge to achieve the specific purpose, including dialogue actions and fine-grained emotions of each utterance. Therefore, the dialogue graph structure proposed in this paper is an extensible graph structure that can easily introduce customized external knowledge to help generate reliable responses. Experimental results show that our proposed MEGE method achieves state-of-the-art performance on the public dataset.