Knowledge-Based Conversational Recommender Systems Enhanced by Dialogue Policy Learning

The conversational recommender system (CRS) provides personalized recommendations for users through dialogues. Knowledgebased CRS, which applies external knowledge graphs into the CRS, can provide knowledge-aware recommendations, and has proved successful in many fields. However, existing models suffer from two limitations. First, previous knowledge-based CRSs ignore the transfer of topics in the conversation and cannot handle multi-task recommendation dialogues. Second, it takes many inquiries for traditional models to obtain knowledge-based user profiles, which affect the user's interactive experience. In this work, we use the dialogue policy learning to tackle these issues, and propose a model called CRSDP, standing for knowledge-based ConversationalRecommender Systems enhanced byDialogue Policy learning. We leverage the actor-critic framework to learn a dialogue policy in the reinforcement learning paradigm. The optimized dialogue policy leads the conversation strategically in multi-task dialogue scenarios, determines the user preference in fewer turns, and makes effective recommendations. We evaluate the performance of the model in the recommendation task and the conversation task. Experimental results validate the effectiveness of our model.