Abstract: Conversational Recommender Systems are known to benefit from explanations of why an entity is recommended. In this work, we implement an attribute-based approach to generating such explanations, which we call AtCRS. We show that this approach is preferred by humans -- making them feel more confident in the recommended item. We also show that attribute-first models provide benefits for automatic generation: AtCRS generates fewer hallucinations and is more consistent with the previous conversation than current state-of-the-art end-to-end systems.The newly annotated StrAtData and the code used in this paper will be made available upon acceptance.
Paper Type: short
Research Area: Dialogue and Interactive Systems
0 Replies
Loading