Construction and Application of Meituan's Common‑Sense Concept Knowledge Graph

This article introduces the design, construction, and practical use of Meituan’s common‑sense concept knowledge graph, which models entities and their relationships to support various search, recommendation, and feed scenarios.

1. Introduction – Natural language understanding (NLU) benefits from external knowledge beyond raw text. Knowledge graphs provide explicit semantic information that helps machines reason like humans.

2. Concept Graph Overview – The graph focuses on three types of nodes (Taxonomy, Atomic concepts, Composite concepts) and four relation categories (Is‑what, What‑like, Provide‑what, POI/SPU‑concept). It is tailored to Meituan’s business domains (e.g., food, travel, medical‑beauty).

3. Construction Pipeline

3.1 Concept Mining – Atomic concepts are extracted from queries, UGC, and deals using popularity, meaningfulness, and completeness criteria, then filtered by an XGBoost classifier. Composite concepts are generated by combining atomic concepts and judged by a Wide&Deep model with graph‑embedding features.

3.2 Hierarchy Mining – Taxonomy‑concept links are treated as an entity‑typing task using BERT. Concept‑concept hierarchies are discovered via lexical rules and contextual sentence classification with BERT‑based models, enhanced by semi‑supervised UDA.

3.3 Attribute Mining – Public attributes are extracted through dependency parsing and fine‑grained NER; open attributes are mined via a bootstrapping template method and then modeled with a joint dependency‑NER network.

3.4 Bridging (Concept‑Concept) – A triplet‑based contrastive network discovers co‑occurrence signals, followed by a BERT‑based classifier that marks valid bridging relations. Missing links are completed using RGAT (Relational Graph Attention Network) combined with TransE.

3.5 POI/SPU‑Concept Linking – Synonym clustering, candidate sentence generation, and a discriminative model align real‑world items (POI/SPU) with graph concepts.

4. Applications

4.1 Category‑Word Graph – Enriches Meituan’s “to‑summary” category vocabularies, expanding nodes and synonyms dramatically.

4.2 Search Suggestion (SUG) – Generates refined queries (e.g., “strawberry cake”, “6‑inch cake”) from high‑level concepts to guide users and improve recall.

4.3 Medical‑Beauty Content Tagging – Improves label accuracy from 51% to 91% and recall from 77% to 91% by leveraging concept‑POI and concept‑UGC links.

5. Results & Outlook – The graph now contains over 2 million concepts and 3 million relations with ~90% accuracy. Ongoing work focuses on expanding relation types, refining models, and achieving higher precision while maintaining coverage.