Generating empathetic responses by injecting anticipated emotion

Showing empathy and reacting to users’ feeling are important social skills for current dialogue generation systems. In previous research, empathetic responses are generated by 1) only modeling the emotion of dialogue history or 2) indirectly leveraging the predicted emotion label of responses. In this paper, we propose a novel empathetic response generation method that incorporates the anticipated emotion into response generation by minimizing the divergence between distribution of responses’ anticipated emotion and ground-truth emotion. The anticipated emotion is predicted by an auxiliary emotion predictor whose input is the previous utterances. Additionally, we treat the generation as deliberation process and design a two-round training method to refine the response iteratively. Experimental results show that the proposed model outperforms the previous state-of-the-art for emphatic dialogue generation task.