Towards retrieval-based conversational recommendation

Conversational recommender systems have attracted immense attention recently. The most recent approaches rely on neural models trained recorded dialogs between humans, implementing an end-to-end learning process. These are commonly designed to generate responses given the user's utterances in natural language. One main challenge is that these generated both be appropriate for dialog context and must grammatically semantically correct. An alternative such generation-based retrieve from pre-recorded data adapt them if needed. Such retrieval-based were successfully explored of general conversational systems, but received limited years CRS. In this work, we re-assess potential design evaluate a novel technique response retrieval ranking. A user study (N=90) revealed by our system average higher quality than those two systems. We furthermore found ranking not aligned with results literature, which points open methodological questions. Overall, research underlines should considered or complement language generation approaches.