A Coarse-to-Fine Model for Geolocating Chinese Addresses

Chunyao Qian, Chao Yi, Chengqi Cheng, Guoliang Pu, Jiashu LiuISPRS Int. J. Geo Inf 2020 · DOI

TL;DR

This paper introduces a novel "coarse-to-fine" deep learning model for geolocating unstructured Chinese addresses. By framing the problem as a sequence-to-sequence task and leveraging BERT for contextual understanding and LSTM for hierarchical prediction, the model achieves state-of-the-art accuracy in pinpointing locations, overcoming challenges like polysemy in Chinese addresses.

What Problem Does This Paper Address?

Accurately determining the geographic location of an address from its textual description, known as address geolocation, is crucial for many location-based services (LBS) like ride-sharing and delivery. In China, addresses are often unstructured and can be ambiguous due to polysemy (words having multiple meanings). Traditional methods that directly predict coordinates (latitude/longitude) often ignore the inherent hierarchical structure of addresses (e.g., province, city, district), which can lead to inaccuracies. Furthermore, predicting fine-grained locations directly is significantly harder than predicting broader regions. This paper addresses the challenge of accurately geolocating Chinese addresses by proposing a method that leverages the hierarchical nature of addresses and addresses the ambiguity of Chinese text.

Methods

The core of this paper is a novel coarse-to-fine geolocation model that treats address geolocation as a Sequence-to-Sequence (Seq2Seq) learning task.

  • Input: A textual Chinese address.
  • Output: A GeoSOT grid code. GeoSOT is a hierarchical subdivision scheme for the Earth's surface, providing a unique, multi-level code that represents geographic regions. The model predicts this code digit by digit.

The model employs an encoder-decoder architecture with an attention mechanism:

  • Encoder: A BERT (Bidirectional Encoder Representations from Transformers) model is used to capture the complex contextual relationships between Chinese characters in the address. BERT's ability to generate context-aware embeddings is key to handling the polysemy inherent in Chinese addresses.
  • Decoder: A Long Short-Term Memory (LSTM) network is used to predict the GeoSOT grid code sequentially. The LSTM decoder, combined with an attention mechanism, allows the model to focus on relevant parts of the encoded address representation at each step of predicting the GeoSOT code. This hierarchical prediction, from coarse to fine, is the "coarse-to-fine" aspect of the model.

Datasets

The authors utilized a dataset collected from Amap, a leading digital map provider in China.

  • Size: Approximately 400,000 addresses initially, preprocessed to 385,793 addresses.
  • Attributes: Includes POI name, category, hierarchical address description, latitude, and longitude.
  • Preprocessing: Duplicate or non-conforming records were removed.
  • GeoSOT Code Generation: GeoSOT codes were calculated for each address based on its latitude and longitude, specifically at the 17th level of granularity.

Key Insights and Ideas

  • Seq2Seq for Geolocation: Framing geolocation as a Seq2Seq task, where the output is a structured code (GeoSOT), is a novel approach that allows for hierarchical prediction.
  • BERT for Chinese Polysemy: Leveraging BERT's contextual embeddings is crucial for disambiguating Chinese characters that have different meanings in different address contexts (e.g., "市" as a city vs. a market).
  • Coarse-to-Fine Prediction: Predicting GeoSOT codes digit by digit, effectively moving from broader regions to finer ones, significantly simplifies the prediction task and improves accuracy compared to direct coordinate regression or flat classification.
  • GeoSOT as a Unified Benchmark: The GeoSOT grid system provides a standardized, hierarchical representation of geographic space, enabling a consistent benchmark for evaluating geolocation models without relying on external metadata.

Key Results

  • Accuracy: The proposed coarse-to-fine model achieved the highest geolocation accuracy among the baselines, outperforming Vanilla-RNN and Bi-LSTM models. For example, at an average address length of 10 and predicting GeoSOT codes up to level 17, the model achieved 96.27% accuracy.
  • Robustness: The model's performance remained stable across different input and output lengths, attributed to the self-attention mechanism in the BERT encoder, which is less sensitive to sequence length.
  • Polysemy Distinction: Visualizations of character embeddings (e.g., for "市" and "区") demonstrated the model's ability to distinguish between different semantic meanings of characters, clustering them into distinct groups.
  • Distance-based Metrics: Compared to XGBoost-regression and XGBoost-classification, the proposed model showed significantly lower mean and median error distances, indicating more precise location predictions. For instance, with a hidden size of 2048, the mean error was 552.6 meters.

Limitations and Future Work

  • Generalization to Other Languages: The current model is optimized for Chinese addresses. Future work could explore generalizing it to other languages and geographical text types (e.g., Weibo posts, travel notes).
  • Alternative Geocodes: While GeoSOT is used, the model could be adapted to support other geocoding systems like GeoHash or Google S2.
  • Automatic Granularity Prediction: Currently, the granularity of the predicted GeoSOT code is explicitly set. Future research could enable the model to predict the appropriate granularity automatically based on the input address and application needs.

Summary updated

BibTeX
@article{qian2020coarse,
  author = {Chunyao Qian and Chao Yi and Chengqi Cheng and Guoliang Pu and Jiashu Liu},
  title = {{A Coarse-to-Fine Model for Geolocating Chinese Addresses}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2020},
  doi = {10.3390/ijgi9120698},
}
Figure