Combine Pre-Trained Language Model and Classical Recommendation Network for Conversational Recommender Systems

Conversational recommender systems (CRSs) typically require natural language interaction with users and complete recommendations within the conversation. Existing approaches often struggle with the integration of generated responses and recommended items. To address this, we propose a novel CRS model named KGPLM, incorporating the Knowledge Graph based recommender module and the dialogue module based on the fine-tuned Pre-trained Language Model (PLM). We decompose the dialogue generation task into two subtasks: dialogue goal prediction and response generation. The former enables the system to make recommendations appropriately during the conversation, while the latter, through the introduction of item tokens, achieves the seamless integration between the recommended items and the generated responses. To fine-tune PLM for the CRS, we propose a two-stage fine-tuning approach: (1) multi-task learning explores the connections between the two sub-tasks; (2) prompt learning optimizes the model's performance on specific sub-tasks. Experimental results demonstrate the effectiveness of our approach on both item recommendation and response generation tasks in the final responses.