Abstract
Conversational recommender systems (CRSs) aim to mine user preferences and recommend appropriate items through natural language dialogue. A complete CRS typically consists of a recommendation module and a conversation module, which generate high-quality recommended items and fluent natural language responses, respectively. Existing research usually constructs and trains the two modules separately, leading to inconsistencies in input information and construction methods for different subtasks. To address the limitations of the above methods, we propose a joint training framework based on knowledge-aware prompt tuning to build an integrated conversational recommender system (KPICRS). First, we adopt contrastive learning to align the semantic space of the embeddings of the conversation context and the knowledge graph entity, and then generate a prompt template as the unified input of the joint training framework through the prompt encoder. Specifically, we incorporate augmented similar user representations into the prompt template, which helps to alleviate the data sparsity and cold-start problems. Next, we jointly train the pretrained language model (PLM) and the prompt encoder so that the PLM can simultaneously generate predictions for both conversation subtask and recommendation subtask. Extensive experiments on two public English and Chinese CRS datasets demonstrate that our model achieves highly competitive performance.