An Empirical Study on Zero-Data Bootstrapping for Conversational Recommender Systems
Synthetic dialogue training from reviews, metadata, and user-item interactions beat zero-shot prompting for conversational recommenders.
The paper tests “zero-data” bootstrapping, where a recommender is trained without an in-domain dialogue corpus. Its synthetic, domain-grounded data outperformed naive synthetic baselines, and active selection methods improved efficiency over random sampling. The authors also report that metadata and collaborative-filtering signals helped choose better training examples. In low-resource settings, the synthetic data could outperform scarce real dialogues while still complementing them. HF Daily Papers' note
The paper tests “zero-data” bootstrapping, where a recommender is trained without an in-domain dialogue corpus. Its synthetic, domain-grounded data outperformed naive synthetic baselines, and active selection methods improved efficiency over random sampling. The authors also report that metadata and collaborative-filtering signals helped choose better training examples. In low-resource settings, the synthetic data could outperform scarce real dialogues while still complementing them. HF Daily Papers' note
score 4