Talk the Walk: Synthetic Data Generation for Conversational Music Recommendation

Recommendation systems are ubiquitous yet often difficult for users to control and adjust when recommendation quality is poor. This has motivated the development of conversational recommendation systems (CRSs), with control over recommendations provided through natural language feedback. However, building conversational recommendation systems requires conversational training data involving user utterances paired with items that cover a diverse range of preferences. Such data has proved challenging to collect scalably using conventional methods like crowdsourcing. We address it in the context of item-set recommendation, noting the increasing attention to this task motivated by use cases like music, news and recipe recommendation. We present a new technique, TalkTheWalk, that synthesizes realistic high-quality conversational data by leveraging domain expertise encoded in widely available curated item collections, showing how these can be transformed into corresponding item set curation conversations. Specifically, TalkTheWalk generates a sequence of hypothetical yet plausible item sets returned by a system, then uses a language model to produce corresponding user utterances. Applying TalkTheWalk to music recommendation, we generate over one million diverse playlist curation conversations. A human evaluation shows that the conversations contain consistent utterances with relevant item sets, nearly matching the quality of small human-collected conversational data for this task. At the same time, when the synthetic corpus is used to train a CRS, it improves Hits@100 by 10.5 points on a benchmark dataset over standard baselines and is preferred over the top-performing baseline in an online evaluation.