Improving the Responses Diversity of Persona-based Neural Conversation System: A Conditional Variational Autoencoders-based Approach

Given a conversational context based on persona information to a chatbot, how to generate diverse and relevant responses? Many existing works could effectively capture useful information from external knowledge, but they rarely consider the diversity of responses. In this paper, we consider both the persona information and the responses' diversity, upon which we develop a Seq2Seq-based model with hierarchical RNN context encoder, Conditional Variational Autoencoder (CVAE) and pointer generator. Hierarchical RNN allows those context utterances to be used more effectively. CVAE is used to capture the relationship between question and several appropriate answers. With pointer generator, an output token in a response could either be generated or copied from persona facts. The model is trained and tested on Persona-Chat dataset. Finally, automatic evaluation shows that compared with the baseline model, this developed model is able to better integrate the persona information and generate more diverse responses.