Towards Policy-Guided Conversational Recommendation with Dialogue Acts

Abstract

Conversation Recommender System (CRS) aims to recommend items through nature conversation. Existing works in open-ended CRS mainly focus on recommendation and generation, but lacks of control over dialogue policy. In addition, the system is unable to adapt user profile to the user's feedback. Thus, we present a new dataset named DA-ReDial (Recommendation through Dialogue guided by Dialogue Act). We summarize 10 representative Dialog Acts and label dialogue with the DAs schema. To solve the problems above, we also propose a novel CRS called PGCR which stands for Policy-Guided Conversational Recommendation. It is able to formulate a DA-aware user profile, leverage Dialogue Acts to explicitly model the discourse structure of conversation and better guide the response generation. Extensive experiments on the new dataset show that our proposed model outperforms most baselines in dialog generation and recommendation. Also, the Policy Network fine-tuned by self-play can better control the dialogue policy and contribute a lot to recommendation strategy and user engagement in conversation.

Tasks

Reproductions