Relevance Matters: A Multi-Task and Multi-Stage Large Language Model Approach for E-commerce Query Rewriting

A team of researchers from JD.com, one of China's largest e-commerce platforms, has published a novel AI framework that significantly improves online shopping search results. Their paper, "Relevance Matters: A Multi-Task and Multi-Stage Large Language Model Approach for E-commerce Query Rewriting," accepted at ICDE 2026, details a system that moves beyond simple query generation. The core innovation is injecting a dedicated relevance task directly into the query rewriting process, forcing the model to bridge the 'lexical gap' between how users phrase searches and how products are described in catalogs. This dual focus on relevance and user conversion (measured by click-through and purchase rates) addresses a fundamental challenge in e-commerce search.

The technical approach is a two-stage process leveraging a pre-trained LLM. First, the model undergoes multi-task supervised fine-tuning (SFT) on JD.com's user and product data, learning both to rewrite queries and to tag the relevance between original queries and their rewrites. Second, the model is aligned using a novel Group Relative Policy Optimization (GRPO) technique to explicitly optimize for the dual objectives of high relevance and stimulated user conversions. The results from offline evaluations and, critically, online A/B tests on JD.com's live platform show the framework delivers substantial improvements in search relevance and boosts key business metrics like User Conversion Value Rate (UCVR), which measures purchases per user. This represents a major step from academic research to real-world, revenue-impacting deployment for a massive online retailer.