Towards a Conversational Recommendation System with Item Representation Learning from Reviews

Abstract

none

Conversation-based recommendation systems are proposed to overcome the challenges of the static recommendation systems by taking real time user-system interactions into account for the user preference learning. However, less information of item is provided from the conversation. Our study proposed a conversation-based recommendation system named Review-Based Conversation Recommendation System(RBCRS). The main idea is to propose an item representation learning model to properly learn item representations from reviews of items. The pre-trained item representation is then used in the proposed review-based recommender model to better represent user preference according to their favorite items detected from the conversation. According to the results of experiments, the proposed recommender in RBCRS would recommend an item that reflect user’s favor except for the popular one. Besides, the RBCRS would provide more recommendations among dialogues and also obtain a higher ratio of making successful recommendations.