EmoDialoGPT - Enhancing DialoGPT with Emotion.

Perceiving user emotions and generating responses with specific emotions are of great significance to a social dialogue system. Much previous work still utilizes seq2seq models as the backbone to generate emotional responses. With the popularity of the generative pre-training models, we propose an emotional response generation model EmoDialoGPT based on DialoGPT by introducing emotion embedding and emotion prediction loss. In order to obtain a large-scale dialogue dataset with emotion labels, we train an emotion classifier and automatically annotate the emotion labels for the dialogue data. We evaluate our models from three aspects, including emotion expression, response quality and human evaluation. The experimental results show that, the proposed models outperform seq2seq models and baseline DialoGPT models on most metrics.