Construction and Application of Meituan's Life‑Service Knowledge Graph

Meituan, China’s largest online local life service platform, has built a large‑scale knowledge graph called “Meituan Brain” since 2018, covering billions of entities and triples across dining, delivery, hotel, and other domains.

The talk is organized into three parts: an overview of Meituan Brain, the construction and application of a tag knowledge graph, and the techniques for building a dish knowledge graph.

Tag knowledge graph construction involves four stages: knowledge extraction (using sequence labeling and span‑trans methods), relation mining (including synonym discovery via vector indexing and hashing), graph labeling (linking tags to merchants using BERT‑based models and multi‑task classification), and graph application (open QA, search recall/ranking, and embedding‑based recommendation). Various models such as word2vec, BERT, Sentence‑BERT, and contrastive learning are employed to improve synonym detection and hierarchical tag mining.

Dish knowledge graph construction focuses on dish name understanding, deep‑learning‑based inference, explicit reasoning, and multimodal pre‑training. Techniques include token‑level labeling, span‑trans, ResNet image encoding, BERT text encoding, and contrastive learning with dual‑tower models to align images, dish names, and categories. Multimodal and semi‑supervised learning are used for attribute extraction and cooking method prediction.

Applications of these knowledge graphs span open‑domain question answering, enhanced search recall and ranking, and recommendation systems where tag‑POI relationships and query‑POI interactions are modeled with GraphSAGE GNNs, leading to significant online performance gains.