The paper introduces WeFair, a way to rewrite search queries on WeChat's short-video search so they match what a user actually wants. It trains a compact model, WeFair-2B, in three stages that use human feedback of decreasing directness.
Training stages: Demonstration uses supervised fine-tuning on search sessions and real user reformulations. Preference uses direct preference optimization on subsequent consumption feedback. Verification uses group relative policy optimization, rewarding exact agreement with reformulation or preservation targets to limit intent drift.
Results: Offline, WeFair-2B beats strong rewriting baselines on overall accuracy and intent preservation. In an online A/B test, effective video-view rate rose 2.593% relative and query reformulation rate fell 2.350% relative.
Deployment: The model is deployed in WeChat short-video search, which serves over 100 million queries per day.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper introduces WeFair, a way to rewrite search queries on WeChat's short-video search so they match what a user actually wants. It trains a compact model, WeFair-2B, in three stages that use human feedback of decreasing directness.