Generating Empathetic Responses By Injecting Anticipated Emotion

Showing empathy and reacting to users’ feeling are impor-tant social skills for current dialogue generation systems. Inprevious 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 intoresponse generation by minimizing the divergence betweendistribution of responses’ anticipated emotion and ground-truth emotion. The anticipated emotion is predicted by anauxiliary emotion predictor whose input is the previous ut-terances. Additionally, we treat the generation as delibera-tion process and design a two-round training method to refinethe response iteratively. Experimental results show that theproposed model outperforms the previous state-of-the-art foremphatic dialogue generation task.