Entity Linking System for Travel Knowledge Graph at Ctrip AI R&D

The Ctrip travel AI R&D team develops AI products for the travel division, focusing on a knowledge‑graph‑driven entity linking service that improves search, Q&A, and information extraction for tourism.

Background : Rapid growth of web data and linguistic ambiguity (polysemy, synonymy) make it hard to retrieve correct information. In tourism, points of interest (POI) – name, address, coordinates, category – are central, and linking textual mentions to POI entities is crucial.

Problem Analysis : Entity linking consists of mention detection, candidate generation, and candidate disambiguation. Traditional methods include dictionary‑based and statistical models (HMM, CRF), while modern approaches use neural networks (CNN, RNN, Transformer) and pretrained models such as BERT.

Travel Knowledge Graph : POI data are stored in Neo4j and Nebula graph databases, with a schema covering 18 entity types and 12 relation types, totaling about 10 million entities and 37 million triples. Automatic update and monitoring pipelines keep the graph fresh.

Technical Solution : The system follows a three‑stage pipeline. Mention detection combines a neural model (BERT‑based pointer network) with a prefix‑tree of all alias strings to achieve high‑recall coverage. Candidate generation retrieves POI candidates via alias‑to‑entity relationships, applying path‑based filtering. Disambiguation uses a BERT‑based interactive semantic‑matching model that concatenates the query and candidate description, extracts CLS, head, and tail token embeddings, and predicts a linking probability via a sigmoid‑activated linear layer.

Engineering optimizations include Redis caches for alias‑to‑entity‑id and entity‑attribute mappings, reducing latency from graph database queries.

Functional Modules :

5.1 Mention Detection – neural network + prefix‑tree.

5.2 Candidate Generation – alias‑driven graph lookup with rule‑based path filtering.

5.3 Candidate Disambiguation – BERT‑based interaction model trained with binary cross‑entropy loss.

Practical Scenarios :

Travel search – resolves ambiguous POI queries (e.g., "武汉东湖") to the correct entity.

Intelligent客服 – improves slot‑filling F1 by >12%.

POI key‑information updates – raises extraction accuracy by ~6%.

Duplicate POI and hierarchical POI detection – achieves >90% correctness.

Conclusion & Outlook : The system demonstrates the value of integrating entity linking with a tourism knowledge graph, and future work will explore tighter coupling with retrieval and ranking, lightweight models, and broader scenario coverage.