Unaligned Message-Passing and Contextualized-Pretraining for Robust Geo-Entity Resolution

Yuwen Ji, Wenbo Xie, Jiaqi Zhang, Chao Wang, Ning Guo, Lei Shi, Yue ZhangAAAI 2025 · OpenReview

Introduction

Geo-Entity Resolution (GER) is the critical task of determining if two records in different databases—such as Yelp and OpenStreetMap—refer to the same physical location. This process is the backbone of modern location-based services. However, real-world geo-data is notoriously "noisy," often containing spelling errors, abbreviations (e.g., "A.W." vs. "Art Wall"), or missing attributes like postcodes. Traditional Large Language Models (LLMs) like BERT excel at semantics but struggle when text is corrupted or incomplete.

Illustration of Geo-Entity Resolution challenges and the proposed neighbor-based matching approach.

The Challenge of Noisy Geo-Data

The core challenge in GER is effectively integrating geographical context—the "neighbors" of a business—to resolve ambiguities. As illustrated in the figure above, when a record for "Art Wall" has missing or abbreviated data, its spatial neighbors (like "KFC" or "Beer") can serve as a source of truth for error correction. Previous attempts to use this context often failed because they tried to force unordered spatial data into linear sequences or ignored fine-grained details. The authors propose that by incorporating neighbors, we can increase the mutual information between the observed record and the true underlying entity, effectively filtering out noise.

Proposed Methodology: UMP and CP

The authors introduce a two-pronged framework designed for robustness:

Unaligned Message-Passing (UMP)

Standard Graph Neural Networks (GNNs) typically require fixed-length input vectors. However, geo-entities are represented by variable-length sequences of tokens. The UMP module utilizes a cross-attention mechanism that allows a "pivot" entity to query specific, fine-grained tokens from its spatial neighbors. This allows for "selective token-level denoising"; for example, if a pivot entity lacks a postcode, UMP allows it to "attend" to the postcode tokens of its neighbors to refine its own representation.

The Unaligned Message-Passing (UMP) architecture utilizing BERT and cross-attention over neighbor tokens.

Contextualized-Pretraining (CP)

To help the model understand spatial relationships before the matching task, the authors introduce two pre-training objectives:

  1. Adjacent Entity Prediction (AEP): A binary classification task where the model predicts if two entities are close neighbors, teaching it the underlying "geography."
  2. Masked Language Modeling (MLM): The model must recover masked tokens using both the entity’s own text and the attributes of its neighbors.

Experimental Results

The framework was evaluated using real-world databases (OpenStreetMap, Yelp, and Foursquare) across four major cities: Singapore, Edinburgh, Toronto, and Pittsburgh.

  • Superior Performance: The proposed method consistently ranked first across all eight evaluation datasets, achieving up to a 7.9% F1-score improvement over the previous state-of-the-art, Geo-ER.
  • Resilience to Noise: In perturbation-based evaluations (simulating 20% to 60% data corruption), the model's performance degraded much more slowly than baselines. While other models' accuracy plummeted as data quality worsened, this framework remained stable.
  • Ablation Success: Removing either the UMP or the CP components led to significant performance drops, proving that both the architectural innovation and the spatial pre-training are vital for robustness.

Conclusion and Future Work

By treating spatial neighbors as active denoisers rather than just extra features, this framework sets a new benchmark for robust Geo-Entity Resolution. While highly effective, future research is needed to optimize the computational overhead of integrating GNN layers with Transformer encoders and to explore how these models handle dynamic environments where businesses open and close over time.

Summary updated

BibTeX
@inproceedings{ji2025unaligned,
  author = {Yuwen Ji and Wenbo Xie and Jiaqi Zhang and Chao Wang and Ning Guo and Lei Shi and Yue Zhang},
  title = {{Unaligned Message-Passing and Contextualized-Pretraining for Robust Geo-Entity Resolution}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=fmyzm8hZDT},
}
Figure