Knowledge Enrichment for Conversational Recommendations

Recommender systems utilize past interaction data to infer user preferences and have been widely adopted across various industry applications. However, traditional recommendation systems frequently become obsolete because they struggle to adjust to evolving user preferences and item availability, resulting in imprecise recommendations. Conversational recommendation systems (CRSs) overcome these limitations by facilitating dynamic, engaging conversations with users. However, existing CRSs face challenges, such as: 1) providing recommendations based on inadequate context, resulting in repetitive suggestions and subsequent user dissatisfaction and 2) struggling to recognize contextual connections between entities, thereby constraining the range of recommendations. To tackle these challenges, we introduce a Knowledge-Enriched CRS framework that utilizes knowledge graphs to improve the comprehension of user preferences. Our model bridges the gap between different attribute representations by integrating structured knowledge from knowledge graphs and incorporating data from external sources. This approach enhances recommendation relevance and user experience by providing more diverse and contextually enriched suggestions. Additionally, by leveraging a pre-trained language model, our model offers detailed explanations for recommendations, boosting user satisfaction. Experiments conducted on the ReDial dataset validate the effectiveness of our method.