Reliable Distillation of Recommendation Rationales From Large Language Models for Conversational Recommendation

Abstract

Conversational recommender systems (CRSs) aim to provide personalized recommendations through conversational interactions. Prior approaches that leverage knowledge graphs (KGs), referred to as KG-based CRSs, have demonstrated strong performance by incorporating domain-specific knowledge. However, they often struggle to comprehend users’ contextual conditions from conversations for context-sensitive recommendations. In contrast, large language models (LLMs) excel at understanding the context of a conversation, but their naive application to the CRS task reveals a lack of domain knowledge. To address these issues, we propose RADAR, a novel approach that combines the strength of KG-based CRSs with the contextual reasoning capability of LLMs. Accordingly, our approach exploits an LLM to distill recommendation rationales, composed of contextual conditions and preferred entities, from conversations. Moreover, we design filtering and scoring mechanisms to enhance the reliability of the distilled rationales. These high-quality rationales are injected into a smaller-sized reasoner, eliminating the need for further reliance on LLMs while equipping the reasoner with the ability to generate accurate and context-sensitive rationales. Using the specialized reasoner, we develop comprehensive user and item representations that enable recommending items whose content aligns with the inferred contextual conditions. In response generation, we improve recommendation explanations with contextual conditions and domain knowledge about a recommended item. Extensive experiments on two CRS benchmark datasets demonstrate that RADAR significantly outperforms existing baselines and delivers human-like recommendations and responses.