Dialogue based recommender system that flexibly mixes utterances and recommendations

Much of the prior research in the recommendations through dialogue separate dialogue and recommendations. However, since the accuracy of the recommendations themselves is not necessarily high, recommendation results rarely meet user needs. However, as human we can find the solutions that satisfy users by appropriately repeating the cycle of checking mismatched reasons and making another recommendations in our conversations. In this paper, we propose a system for leveraging a dialogue strategy for reinforcement learning using recommendation results based on user utterances. We constructed a dialogue system to perform adaptive behavior that naturally incorporates recommendations into conversation with users.