Knowledge-based Question Answering (KBQA) System at Meituan: Design, Challenges, and Solutions

Knowledge-based Question Answering (KBQA) refers to answering natural‑language questions by understanding the query, retrieving and reasoning over a knowledge base. Meituan applies KBQA across pre‑sale, in‑sale, and post‑sale scenarios to automatically or recommend replies, improving merchant response efficiency.

The article presents Meituan’s practical experience, based on an EMNLP 2021 paper, covering overall system design, key challenges, and end‑to‑end QA exploration.

Background & Challenges : The platform serves diverse domains (hotels, tourism, food, etc.) with varied user intents, constraint‑rich queries, and multi‑hop questions. Challenges include handling numerous business scenarios, constraint extraction, and multi‑hop reasoning.

Solution Overview : The KBQA pipeline consists of four stages – Query Understanding, Relation Recognition, Sub‑graph Retrieval, and Answer Ranking. An architecture diagram (Fig. 1) shows the application layer, query understanding modules, and knowledge‑graph storage.

Query Understanding : Performs fine‑grained semantic parsing, including entity recognition & linking and dependency analysis. Optimizations such as knowledge injection for OOV words, nested entity handling, and context‑aware linking are applied.

Relation Recognition : Models the task as text matching between the query and candidate predicates, using both representation‑based (dual‑tower) and interaction‑based architectures. BERT is pruned (layer‑wise skipping), fine‑tuned on domain data, and enhanced with domain and syntactic features, achieving higher accuracy with lower latency.

Complex Question Handling : Addresses constraint questions by extracting constraint tokens via dependency parsing and performing key‑value matching; multi‑hop questions are solved by identifying hop count and traversing the graph accordingly. Opinion QA extracts sentiment‑type answers from user reviews, providing aggregated support counts and evidence snippets.

End‑to‑End Exploration : Introduces three pre‑training tasks – Relation Extraction, Relation Matching, and Relation Reasoning – to bridge the gap between textual pre‑training and graph‑structured reasoning, yielding notable performance gains (Fig. 17).

Applications : Deployed in hotel “Ask‑Me‑Anything”, ticket promotion, and merchant reply recommendation, the system has improved response speed and conversion rates. It achieved top rankings in the 2020 CCKS KBQA benchmark.

Summary & Outlook : While the current system handles head‑type queries well, long‑tail and more complex question types remain challenging. Future directions include unsupervised domain transfer, business‑knowledge injection, handling comparative and multi‑relation queries, and fully end‑to‑end KBQA solutions.