Meituan Brain: Large‑Scale Knowledge Graph Construction and Applications

The article introduces Meituan’s large‑scale knowledge graph project, called Meituan Brain, which aims to enable cognitive intelligence by providing machines with structured, symbolic knowledge. It starts with a motivation that deep learning, while successful in perception tasks, suffers from lack of interpretability, common‑sense reasoning, semantic understanding, and heavy reliance on massive labeled data.

To overcome these limitations, knowledge graphs are presented as essential for organizing real‑world concepts, entities, and relations in a triple format. Meituan Brain extracts and integrates multi‑modal data from user reviews, merchant information, and external sources to build a graph covering billions of entities and triples across the food, entertainment, and travel domains.

The technical pipeline is described as a knowledge‑graph technology chain, including:

Knowledge acquisition from structured, semi‑structured, and unstructured data, using methods such as BiLSTM‑CRF for entity recognition, relation extraction (rule‑based, distant supervision, and neural joint models), and entity linking.

Knowledge validation through schema design, constraint checking, and confidence scoring of extracted triples.

Knowledge fusion that merges heterogeneous sources via schema merging, entity alignment, and linking.

Knowledge representation using both symbolic RDF triples and embedding‑based vector representations (Word2Vec, TransE, etc.), stored in a distributed graph database with CVT (Compound Value Type) support.

Knowledge reasoning that combines symbolic logical inference and statistical graph‑based inference to discover new facts and detect inconsistencies.

Knowledge empowerment for downstream AI services such as personalized recommendation, semantic search, and question answering.

Business applications of Meituan Brain are highlighted:

Intelligent Search: Enhances search relevance and diversity by leveraging entity relationships and user intent, providing explainable recommendations.

Merchant Empowerment (ToB): Supplies SaaS cash‑register systems with insights from aggregated user feedback, enabling data‑driven operational decisions.

Financial Risk Management & Anti‑Fraud: Utilizes graph‑based community detection and label propagation to identify risky users and fraudulent groups.

The article also discusses future challenges, including data ingestion and schema evolution, large‑scale graph storage and query, and the design of distributed graph algorithms for billions of edges.

Finally, acknowledgments are given to the many internal teams that contributed, followed by a list of academic references covering entity recognition, relation extraction, distant supervision, and knowledge‑graph embedding methods.