ZennRefining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization
Published in Proceedings of EMNLP 2025 (Main Conference), 2025
Peer-reviewed, accepted at the EMNLP 2025 Main Conference (Suzhou, China).
We extend SumRec, a summary-based CRS, by introducing a two-stage training procedure that first trains a score predictor, then uses DPO to optimize the generators with the predictor-induced preferences — improving realism of recommendation dialogue.
Recommended citation: Manato Tajiri, Michimasa Inaba. (2025). "Refining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization." Proceedings of EMNLP 2025, pp. 28628–28649. Suzhou, China.
Download Paper | Download Bibtex