User Linguistic Style Awareness and Interest-Driven Conversational Recommender Systems

Generation-based conversational recommender systems (CRSs) are tailored to individual user needs, employing cutting-edge text generation techniques for generating contextually relevant and fluent responses. Nevertheless, current CRSs, including advanced general language models like ChatGPT, often overlook the subtle differences in users’ linguistic styles and the evolving nature of their interests, which leads to a tendency for uniform responses at the onset of dialogues and unnecessarily prolonged interaction cycles. To overcome these challenges, our approach leverages user historical data and information from similar user profiles to more accurately capture the user linguistic feature within dialogue subtasks. We introduce an innovative multi-factor cross-attention mechanism, specifically designed for the efficient integration of diverse quantitative relationships. In the recommendation subtask, our strategy involves the intuitive representation of user interest migration trajectories and the stereotype. This not only allows for a more granular understanding of the trajectory of user interests but also alleviates data sparsity for new users. Our proposed framework, User Linguistic Style Awareness and Interest-Driven CRSs (LSAID), has undergone extensive testing on well-established CRS datasets such as ReDial and INSPIRED. The empirical results underscore LSAID’s effectiveness, showcasing its state-of-the-art performance. Our code will be released at http://github.com/zhizhaixingchen/LSAID.