A User Preference and Intent Extraction Framework for Explainable Conversational Recommender Systems

Conversational recommender systems (CRS) communicate with a user through natural language understanding to support the user finding necessary information. While the importance of user information extraction from a dialog is growing, previous systems rely on named-entity recognition to find out user preference based on deep learning methods. However, there is still scope for such recognition modules to perform better in terms of accuracy and richness of the elicited user preference information. Also, extracting user information solely depending on entities mentioned in user utterances might ignore contextual semantics. Besides, black-box recommender systems are widely used in previous CRSs whereas such methods undermine transparency and interpretability of recommended results. To alleviate these problems, we propose a novel framework to extract user preference and user intent and apply it to a recommender system. User preference is extracted from sets of an item feature entity detected by our item feature entity detection module and an estimated rating about each entity. Utilizing graph representation of user utterances, user intent is also elicited to consider the contextual semantic of each element word. Based on both outcomes, we implement recommendation by candidate selection and ranking, then provide explanation of the recommendation result to enhance interpretability and manipulability of the system. We illustrate how our framework works in practice by a sample conversation. Experiments present improvement and effectiveness of user information elicitation in recommendation.