In the realm of e-commerce search, the significance of semantic matching cannot be overstated, as it directly impacts both user experience and company revenue. Query rewriting serves as an important technique to bridge semantic gaps inherent in the semantic matching process. However, existing query rewriting methods often struggle to effectively optimize long-tail queries and alleviate the phenomenon of \textit{``\nothing''} caused by semantic gap. In this paper, we present \textbf{\method}, a comprehensive framework that \textbf{B}ridges the s\textbf{E}mantic gap for long-tail \textbf{QUE}ries. \method comprises three stages: multi-instruction supervised fine tuning (SFT), offline feedback, and objective alignment. Specifically, we first construct a rewriting dataset based on rejection sampling, and mix it with multiple auxiliary tasks data to fine tune our large language model (LLM) in a supervised fashion during the first stage. Subsequently, with the well-trained LLM, we employ beam search to generate multiple candidate rewrites, which would be fed into Taobao offline system to simulate the retrieval process and obtain the partial order. Leveraging the partial order of candidate rewrites, we introduce a contrastive learning method to highlight the distinctions between rewrites and align the model with the Taobao online objectives. Offline experiments prove the effectiveness of our method in enhancing retrieval performance. Online A/B tests reveal that our method can significantly boost gross merchandise volume (GMV), number of transaction (\#Trans) and unique visitor (UV) for long-tail queries. \method has been deployed on Taobao, one of most popular online shopping platforms in China, since October 2023.

翻译：在电商搜索领域，语义匹配的重要性不言而喻，它直接影响用户体验和公司收入。查询重写作为弥合语义匹配过程中固有语义鸿沟的重要技术，其现有方法往往难以有效优化长尾查询并缓解语义鸿沟导致的"无结果"现象。本文提出**BESQUE**框架，这是一个弥合长尾查询语义鸿沟的综合体系，包含多指令监督微调（SFT）、离线反馈和目标对齐三个阶段。具体而言，我们首先基于拒绝采样构建重写数据集，并将其与多项辅助任务数据混合，在第一阶段对大型语言模型（LLM）进行监督微调。随后利用训练好的LLM，通过束搜索生成多个候选重写结果，输入淘宝离线系统模拟检索过程并获取偏序关系。基于候选重写的偏序关系，我们引入对比学习方法凸显重写差异，使模型与淘宝在线目标对齐。离线实验证明该方法能有效提升检索性能。在线A/B测试表明，我们的方法显著提升了长尾查询的商品交易总额（GMV）、交易笔数（#Trans）和独立访客数（UV）。自2023年10月起，**BESQUE**方法已部署于中国最受欢迎的在线购物平台之一——淘宝。