Enhancing Response Relevance and Emotional Consistency for Dialogue Response Generation

VAE (Variational Autoencoder) and CVAE (Conditional V AE) encode the sentence with the latent variable to generate response in Dialogue. However, studies have shown that the latent variables obtained are more inclined to remember the first words and the length of the sentence, and only represents limited local features. In order to alleviate this problem, we propose to involve contrastive learning to generate positive and negative samples for training process, which enriches the latent variables representation with the global information of sentence and generates more relevant response. On the other hand, those generative models do not consider emotional information of dialogue, a sentiment discrimination module is introduced in our model to maintain the emotional consistency. Experiments on two public datasets - DailyDialog and PERSONA-CHAT demonstrate the effectiveness of our method, the evaluation results of BLEU and Rouge are both improved. The sentiment discrimination network also forces the model to generating emotional consistency response with share embedding.