Applying Knowledge Graphs to Scene Understanding in Meituan's Hotel and Travel Search

The talk introduces the unique characteristics of Meituan's hotel and travel (酒旅) business, emphasizing the larger travel radius compared with local services and illustrating user scenarios such as skiing, beach trips, hotel searches by landmark or brand, and scenic‑spot queries.

It defines "scene cognition" as interpreting user search behavior as explicit expressions of demand scenarios, mapping these scenarios to tags, and retrieving matching merchants.

The knowledge‑graph construction covers four layers: category taxonomy, atomic concepts extracted from reviews and merchant data, demand concepts formed by semantic combination of atomic concepts, and POI layer linking merchants to these concepts. Extraction pipelines combine semi‑supervised pattern learning, NER, and BERT‑CRF models, while classification uses multi‑label approaches and multi‑task training.

Meituan's proprietary MT‑BERT model, enriched with massive user‑generated content, improves both extraction and classification performance.

A five‑layer service‑search architecture (L0–L4) is described: L0 builds knowledge indexes; L1 parses queries into structured intents; L2 applies deep‑learning ranking; L3 adjusts strategies per business scenario; L4 provides explainable tags and recommendation reasons.

Two main search intents are covered: precise merchant search (single‑point and multi‑point) and generic scene search (landmark+X, local+X, nationwide+X). Technical solutions include context‑aware entity linking, vector similarity via skip‑gram and GCN on heterogeneous graphs, and multi‑step query processing for generic scenes.

Recommendation explanations are generated either by extraction (using BERT‑MRC and pointer‑generator networks) or by generation (transformer‑based sentence compression and KG‑driven synthesis), followed by quality assessment and online traffic distribution.