Intelligent Question Answering Technology Framework and Practices at Meituan

The article introduces Meituan's intelligent question answering (QA) technology, which aims to provide accurate answers to open‑ended user queries in life‑service scenarios such as travel, hotels, and shopping. Traditional QA systems rely on pre‑built knowledge bases that handle frequent questions well but struggle with dynamic, open‑domain queries.

To address this, Meituan builds a three‑part QA capability: Document QA (extracting answers from unstructured documents using machine reading comprehension), Community QA (leveraging user‑generated Q&A from the "Ask Everyone" module), and Knowledge‑Graph QA (answering queries over structured knowledge graphs).

Document QA uses deep neural networks for machine reading comprehension (MRC). The pipeline consists of a retriever that selects relevant documents, a reader (based on BERT) that predicts answer spans, and a ranker that scores candidate answers. The system also incorporates a no‑answer detection head and a Yes/No classification task to improve answer completeness and correctness.

Community QA processes user‑generated question‑answer pairs. Low‑quality answers are filtered using keyword and pattern matching, while answer quality is further improved by a pairwise RoBERTa ranking model that considers both question and answer semantics. Online matching combines coarse retrieval with fine‑grained semantic re‑ranking.

Knowledge‑Graph QA adopts a semantic‑parser approach to map natural language questions to logical forms (e.g., SPARQL) and retrieve answers from a graph. Relation detection combines pattern matching and phrase similarity, and constraint understanding uses a deep biaffine model enriched with NER features. Answer ranking leverages graph‑based constraints and a scoring model.

Since each QA module may return multiple candidate answers for the same user query, a multi‑answer fusion ranking stage is introduced. The fusion model considers semantic relevance, answer completeness, and factual correctness, using features such as answer length, information entropy, and cross‑answer attention, as well as intent information.

The integrated QA system is deployed in several real‑world applications: a conversational QA assistant for travel and hotel queries, QA‑enhanced search, an intelligent ticket‑booking chatbot at scenic spots, and the "Ask Everyone" feature that supplements user‑generated answers with automated responses.

Overall, the paper demonstrates how advanced NLP techniques—MRC, multi‑task learning, ranking models, and knowledge‑graph parsing—are combined to build a scalable, high‑quality intelligent QA service within Meituan's ecosystem.