Goal-oriented conditional variational autoencoders for proactive and knowledge-aware conversational recommender system

Conversational recommender system is designed to proactively elicit the user preferences in a dialogue manner, which could effectively improve the user experience as well as the accuracy of recommendation compared with the traditional static recommender systems. As a powerful technique to flexibly produce context-dependent responses, generative dialogue systems have been widely studied. To further improve the meaningfulness and diversity of responses, a number of remarkable researches have been proposed to enrich the inputs of generative model and utilize them effectively. In this work, we focus on capturing the discourse-level features of responses to improve the quality of generation, and propose a novel goal-oriented conditional variational autoencoders model. Our model uses the latent variable guided by dialogue goal to learn the distribution over potential responses and generates informative and diverse results. Moreover, a response-aware knowledge discernment mechanism is proposed which employs the discourse-level response features to accurately discern related knowledge facts and further facilitate the response generation. Extensive experimental studies are conducted to prove the effectiveness of the proposed approaches, the results of the automatic evaluation, the human evaluation as well as the ablation studies demonstrate the potential of this work.¹