Semantic Understanding of Merchant Signboards for Automatic POI Name Generation at Amap

Introduction: Amap maintains tens of millions of POI (Point‑of‑Interest) entries such as buildings, shops, and schools. Keeping POI names fresh requires massive manual effort, which is costly. Automatic generation of POI names therefore relies on accurate semantic understanding of merchant signboards.

Background: Merchant signboards contain heterogeneous text lines that can be categorized into four groups: main name, business nature (scope), branch name, and noise (e.g., address, phone). Only the first three participate in the final POI name. Correctly interpreting the semantics of each line, especially when the same phrase (e.g., “China Telecom”) appears with different meanings, demands contextual and visual analysis.

Technical Solution – Two‑Stages Cascade Model:

1. Stage One – Single‑line Feature Extraction gathers token‑level and sentence‑level features together with positional information (PV). Token‑level features use three part‑of‑speech tags defined for POI names: Core (C), Universal (U), and Tail (T). An LSTM learns the sequence of these tags. Sentence‑level features are derived from a domain‑adapted BERT model (BERT‑POI) that is further pre‑trained on POI text pairs using a multi‑task objective (character‑mask completion and same‑POI detection). Experiments show BERT‑POI improves missing‑character and same‑POI prediction accuracy by about 20 % over the original Google BERT.

2. Stage Two – Contextual Relationship Modeling employs a Bi‑LSTM to fuse the Stage‑One outputs, allowing the model to capture interactions among multiple text lines on the same signboard and resolve ambiguities.

Business Impact: The semantic understanding module achieves >95 % accuracy. In downstream POI name generation, BERT‑POI raises recall for main name, branch name, and business scope by 3 %–6 % compared with the baseline. The system can automatically apply different naming strategies based on the detected composition (e.g., “main + noise”, “main + branch + business nature”).

Conclusion: The signboard semantic module significantly improves POI name automation and will be extended to other tasks such as noise filtering, building attachment detection, sensitive category handling, missing‑text recovery, name generation, and correction. Future work will deepen multimodal (image‑text) exploration to further enhance data quality for real‑world scenarios.