LGCRS: LLM-Guided Representation-Enhancing for Conversational Recommender System

Conversational recommender systems primarily focus on acquiring user preferences through real-time interactions. However, their effectiveness in modeling user preferences is constrained by the amount of conversational information. Existing conversational recommendation methods typically leverage external knowledge graph to enhance recommendation performance, potentially overlooking valuable textual information related to users and items. With significant natural language understanding capabilities, large language models (LLMs) have the potential to improve the performance of conversational recommendation task by improving the quality of text representations. However, the incorporation of raw textual information may inadvertently introduce noise to recommendations. Additionally, the semantic gap between the textual information and the collaborative signals could potentially degrade the performance of the conversational recommender system. To tackle the aforementioned challenges, we propose a LLM-guided representation-enhancing method for conversational recommender system, which fuses collaborative signals and semantic information to improve recommendation performance and generate high-quality responses. Specifically, we leverage LLMs to refine item profiles while reducing noise in text information, construct fused item representations across multiple aspects, and align the LLMs-enhanced semantic representation with the CF-side rational representation through regularization terms in the training stage. Experiments on two publicly available conversational recommendation datasets show that our method exhibits better performance in both recommendation and conversation tasks.