The paper discusses Conversational Recommender Systems (CRS) which recommend items to users through natural language conversation. Traditional methods utilize external knowledge graphs, language models, and a recommendation module. This often leads to a cumbersome training process and issues with semantic misalignment. The authors propose a new approach, PECRS (Parameter-Efficient CRS), which represents items in natural language and formulates CRS as a natural language processing task. This model uses pre-trained language models to encode items, understand user intent, perform item recommendation, and generate dialogues. The model can be optimized in a single stage without relying on non-textual metadata like a knowledge graph. Experiments on two benchmark CRS datasets, ReDial and INSPIRED, show the effectiveness of PECRS on recommendation and conversation.
Publication date: 25 Jan 2024
Project Page: https://github.com/Ravoxsg/efficient_unified_crs
Paper: https://arxiv.org/pdf/2401.14194
