Abstract:Search relevance modeling is a core task in e-commerce search systems, assessing how well a user query matches candidate products. Rather than relying on a single holistic matching signal, relevance judgment often requires structured reasoning over query understanding, product understanding, and facet-level matching. With large language models (LLMs), this process is increasingly formulated as chain-of-thought (CoT) reasoning and optimized with reinforcement learning (RL). However, existing RL methods mainly rely on outcome-level rewards and treat the entire reasoning chain as a single optimization unit. This makes it difficult to distinguish faulty reasoning steps from correct intermediate ones, leading to misaligned credit assignment. Although process-reward methods provide denser supervision, they often treat reasoning steps independently and ignore dependency-driven error propagation, making responsibility attribution difficult and limiting the optimization of structured relevance reasoning. We propose Graph-GRPO, a graph-structured extension of GRPO for multi-component relevance reasoning. Graph-GRPO constructs a relevance reasoning dependency graph, where CoT steps are modeled as nodes and their logical dependencies as edges. It propagates outcome-level rewards over the graph to derive step-level credit signals, enabling more accurate fine-grained credit assignment. We further introduce a main-loss-driven controller that adaptively adjusts edge-wise credit-propagation coefficients. Together with CoT random masking for supervised policy initialization and graph-node-based multi-head distillation, we build a trainable and deployable framework for generative relevance modeling. Extensive offline evaluations and online A/B tests on a leading e-commerce platform demonstrate that the Graph-GRPO-based framework improves relevance classification metrics and key engagement metrics.
Abstract:For e-commerce search, user experience is measured by users' behavioral responses to returned products, like click-through rate and conversion rate, as well as the relevance between returned products and search queries. Consequently, relevance and user conversion constitute the two primary objectives in query rewriting, a strategy to bridge the lexical gap between user expressions and product descriptions. This research proposes a multi-task and multi-stage query rewriting framework grounded in large language models (LLMs). Critically, in contrast to previous works that primarily emphasized rewritten query generation, we inject the relevance task into query rewriting. Specifically, leveraging a pretrained model on user data and product information from JD.com, the approach initiates with multi-task supervised fine-tuning (SFT) comprising of the rewritten query generation task and the relevance tagging task between queries and rewrites. Subsequently, we employ Group Relative Policy Optimization (GRPO) for the model's objective alignment oriented toward enhancing the relevance and stimulating user conversions. Through offline evaluation and online A/B test, our framework illustrates substantial improvements in the effectiveness of e-commerce query rewriting, resulting in elevating the search results' relevance and boosting the number of purchases made per user (UCVR). Since August 2025, our approach has been implemented on JD.com, one of China's leading online shopping platforms.