MC-CRS: enhanced conversational recommender system based on multi-contrastive learning

Accurately learning dynamic user preferences from limited conversations and generating responses with interpretations is crucial for conversational recommender systems (CRS). Existing research has mainly focused on enhancing the understanding of dialogue context with the help of specific types of external knowledge bases (especially knowledge graphs). This process often neglects the learning and modeling of the original dialogue and lacks the utilization and integration of multi-type data. To this end, we propose a new multi-contrastive learning approach for conversational recommender systems, called (MC-CRS), which first obtains two representations of a contextual information through a text encoder and a ‘perturb’ encoder, and utilizes contrastive learning to mine the deep semantic information hidden in the contextual information. Second, we use structured knowledge graphs and personalized multi-reviews to pre-train the recommendation module, which uses contrastive learning to bridge the semantic gap between multi-types of data to achieve diverse recommendations. We conduct a large number of experiments on two public CRS datasets, and the final results demonstrate the effectiveness of our approach in recommendation and conversation generation tasks.