Query Rewriting in Meituan Search: Iterative Directions and System Design

This article introduces query rewriting—a technique that expands user queries with better expressions to improve recall in text‑based Boolean retrieval systems—within Meituan's search scenario. It outlines the motivation, challenges, and the overall impact of query rewriting on search experience.

1. Introduction explains that mismatches between user queries and document texts cause severe recall loss. Query rewriting (also called Query Expansion) generates alternative terms that are semantically related to the original query, helping users retrieve more relevant merchants, products, and services.

2. Background and Challenges describes four types of semantic gaps in Meituan search: semantic expansion (synonyms, hyponyms, case variations), user‑merchant expression gaps, scenario expansion (e.g., “pick strawberries” → “strawberry garden”), and other recall issues such as missing characters or time‑sensitive concepts. It also highlights the unique difficulty of incorporating a third “regional” constraint, the diversity of local services, and the need for high‑precision, low‑latency rewriting across many business lines.

3. Technical Choices covers the end‑to‑end framework:

3.1 Raw Corpus Mining – Methods include search‑log co‑click mining, session‑based mining, word alignment, merchant SEO extraction, graph‑based mining (SimRank++, GNN), and semantic vector mining (word2vec, Doc2Vec, session‑based embeddings). FastText, LSH, DSSM, and XGBoost are used for candidate filtering.

3.2 Semantic Discrimination Models – BERT‑based sentence‑pair classifiers (MT‑BERT, NMT‑BERT co‑training) are fine‑tuned with semi‑supervised and hard‑negative data to achieve >94% accuracy and mitigate semantic drift.

3.3 Online Services – Four solutions: high‑precision dictionary rewriting, statistical machine translation (SMT) + XGBoost ranking, neural machine translation (NMT) with reinforcement learning, and vector‑based retrieval using dual‑tower models and ANN (Faiss/Antler). Each module is described with architecture diagrams, training pipelines, and deployment details.

4. Summary and Outlook reports that query rewriting now accounts for ~73% of Meituan App search traffic and ~67% of Dianping App search traffic, handling up to 60 k QPS. Future work includes deeper vector retrieval research, richer semantic discrimination (including multimodal signals), advanced generative rewriting (e.g., SeqGAN), and finer‑grained lexical relation graphs for the life‑service domain.

5. Authors : Yang Jian, Zong Yu, Xie Rui, Wu Wei (Meituan Search & NLP team).