Beyond Single Labels: Improving Conversational Recommendation through LLM-Powered Data Augmentation

Haozhe Xu, Xiaohua Wang, Changze Lv, Xiaoqing Zheng · 2025

Read Beyond Single Labels: Improving Conversational Recommendation through LLM-Powered Data Augmentation by Haozhe Xu, Xiaohua Wang, Changze Lv, Xiaoqing Zheng (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-05657.txt.

Description

A shopping assistant that chats can learn what you like, but its training data has a hidden flaw. Items you would have loved are often marked as things you rejected, simply because you never saw them. Haozhe Xu and co-authors fix these false negatives by expanding the label set with help from a large language model. Their framework first retrieves extra items that are semantically relevant to the conversation, then filters the list with a scorer to drop noise.

A two-stage training plan balances the new semantic labels against the collaborative patterns already in the data, who liked what alongside whom. Across two benchmark datasets and simulated users, the augmented training improved several recommender models consistently. The authors present the method as a general remedy for label scarcity in conversational recommendation.

Catalog record

CREATORHaozhe Xu, Xiaohua Wang, Changze Lv, Xiaoqing Zheng
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-05657

Browse the full library · About the Tower of Babel