Multi-Level Gazetteer-Free Geocoding

Sayali Kulkarni, Shailee Jain, Mohammad Javad Hosseini, Jason Baldridge, Eugene Ie, Li ZhangACL 2021

TL;DR

This paper introduces Multi-Level Geocoder (MLG), a novel gazetteer-free model that accurately geocodes location mentions in text by leveraging a hierarchical representation of the Earth's surface. MLG achieves state-of-the-art results by jointly optimizing predictions across multiple levels of granularity, demonstrating that it can learn spatial language representations without relying on external knowledge bases.

What Problem Does This Paper Address?

Geocoding, the task of converting textual location references (toponyms) into geographic coordinates, is crucial for many applications, especially in analyzing social media and historical texts. Traditional approaches often rely heavily on gazetteers (lists of known places) and population data, which can introduce biases towards well-known locations and struggle with less common or ambiguous references. This reliance on external, often curated, data limits the flexibility and generalizability of these models. The paper aims to address this by developing a geocoding model that can learn to associate toponyms with coordinates directly from text, without needing a gazetteer, and to improve accuracy by considering the hierarchical nature of geographic locations.

Methods

The core innovation of MLG lies in its multi-level approach to geocoding and its gazetteer-free nature.

  • Hierarchical Spatial Representation (S2 Geometry): The Earth's surface is discretized into a hierarchical grid using S2 geometry, which divides the sphere into a cube and recursively subdivides its faces. This provides a structured way to represent locations at various levels of granularity, from large regions to small areas (down to ~1cm² at the finest levels).
  • Multi-Level Prediction: Instead of predicting a single level of granularity, MLG jointly predicts cells at multiple S2 levels (e.g., L5, L6, L7). This allows the model to balance the accuracy of fine-grained predictions with the robustness of coarser-grained ones. A combined loss function is used to optimize predictions across these levels.
  • Gazetteer-Free Geocoding: MLG is designed to learn the connection between text and coordinates directly from training data, without requiring a gazetteer during inference. This makes it more flexible and less prone to biases introduced by external knowledge bases.
  • Convolutional Neural Network (CNN) Architecture: The model employs a CNN to process textual input (target toponym, toponyms in context, and surrounding words) and map it to the S2 grid cells.

Datasets

The paper uses three public datasets for evaluation:

  • Wikipedia Toponym Retrieval (WikToR): A large, programmatically created corpus designed for comprehensive toponym resolver evaluation, prioritizing ambiguous locations.
  • Local-Global Lexicon (LGL): A dataset derived from news articles, containing a mix of well-known and less common locations.
  • GeoVirus (GV): A dataset based on news articles about global epidemics, with manually tagged place mentions.

The authors also created unified evaluation sets by re-annotating coordinates from these datasets to ensure consistency, particularly addressing issues like latitude/longitude sign flips and differing coordinate representations across datasets.

Key Insights and Ideas

  • Gazetteer-Free is Feasible and Beneficial: The most striking insight is that MLG achieves strong performance, even outperforming models that use gazetteers, especially on datasets designed to counter population biases (like WikToR). This suggests that learning directly from text and spatial structure is highly effective.
  • Hierarchy Matters: Utilizing a multi-level hierarchical representation of space (S2 geometry) significantly improves geocoding performance. This allows the model to leverage the natural structure of geographic data, combining coarse and fine-grained information.
  • Balancing Granularity and Accuracy: The multi-level prediction strategy effectively balances the trade-off between the precision of fine-grained cells and the generalization capabilities of coarser cells. This is crucial for handling diverse location types and scales.
  • Generalization Across Domains: MLG demonstrates good generalization capabilities, performing well on out-of-domain datasets (LGL and GeoVirus) without dataset-specific tuning, highlighting the robustness of its gazetteer-free, multi-level approach.

Key Results

MLG achieves state-of-the-art results on the evaluated datasets.

  • Improved Accuracy: MLG shows significant gains over previous methods, particularly when not using a gazetteer filter. For example, it demonstrates a 6% improvement in AUC over CamCoder with a gazetteer filter and a 9% gain without it.
  • Robustness to Population Bias: MLG outperforms population baselines on WikToR, a dataset where population-based heuristics are counter-productive, indicating its ability to generalize beyond popular locations.
  • Effective Multi-Level Performance: Table 3 shows that using multi-level prediction (L5-L7) consistently outperforms using single levels at inference time, demonstrating the benefit of combining information across granularities.

Limitations and Future Work

  • Limitations:
    • While MLG is gazetteer-free during inference, it still relies on gazetteer-like information (population data) for gazetteer-constrained prediction, which is an optional post-processing step.
  • The current model predicts points, not polygons, which is a simplification for representing geographic regions.
  • Future Work:
    • Fine-tuning Large Pre-trained Language Models: Exploring the integration of large language models (like BERT) with MLG's multi-level approach.
  • Polygon-based Output: Extending the model to predict polygonal regions instead of points for more realistic geographic representation.
  • Earth Mover's Distance: Investigating the use of Earth Mover's Distance for smoothing the label space and capturing relationships between spatial cells.
  • Hierarchical Softmax: Further refining the multi-level loss by incorporating hierarchical softmax for conditional probabilities.

Summary updated

BibTeX
@inproceedings{kulkarni2021multi,
  author = {Sayali Kulkarni and Shailee Jain and Mohammad Javad Hosseini and Jason Baldridge and Eugene Ie and Li Zhang},
  title = {{Multi-Level Gazetteer-Free Geocoding}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}

Loading related content...

Figure