Multi-granularity contrastive learning for conversational recommender system

Conversational recommender systems (CRS) use natural language dialogues to capture user preferences and provide personalized recommendations. CRS typically includes two modules: recommendation and dialogue, which serve different functions but are both essential. Early approaches used external knowledge graphs and pre-trained language models (PLMs) to capture semantic information and generate recommendations. However, these methods often focus on entity relationships while neglecting the deep semantics of the text, and the distinct construction of the recommendation and dialogue modules makes it challenging to ensure semantic consistency. To address these issues, we propose a Multi-granularity Contrastive Learning CRS (MGCCRS), which unifies the recommendation and dialogue modules using prompt learning and optimizes them with a multi-granularity contrastive learning strategy. This strategy uses both coarse-grained and fine-grained contrastive learning to enhance the model's understanding of overall and local semantics. Our approach improves the performance of both modules, ensures semantic consistency, and outperforms previous methods, as demonstrated by experiments on the ReDial and INSPIRED datasets.