Research map / Graph learning / Knowledge and temporal graphs

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

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Graph-based dialog systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Unaligned Message-Passing (UMP), a cross-attention graph module over token-level neighbor embeddings, plus a contextualized pretraining (AEP and MLM) framework, to make geo-entity resolution robust to noisy attributes.

Problem

Geo-entity resolution across spatial datasets degrades when attribute quality is poor (spelling errors, missing values), because pairwise semantic comparison is unreliable and existing neighbor-aware methods pool or serialize neighbor tokens.

Approach

Use cross-attention-based message passing that lets each entity's token embeddings query unaligned token embeddings of spatial neighbors (avoiding pooling), and pretrain on unlabeled multi-source geo data with adjacent-entity prediction and masked language modeling so the module learns geo-context.

Key results

• At 20% perturbation, Geo-ER+UMP&CP reaches F1 .802 on OSM-FSQ Singapore versus .725 for Geo-ER, with improvement up to 7.9% over the best alternative.
• At 40% perturbation, the method improves over the best alternative by 6.0% to 7.5% on OSM-FSQ datasets.
• Removing AEP or MLM from CP lowers results more than removing CP's contribution in some cases, indicating both tasks matter; BERT+UMP&CP shows about 9% average gain over Ditto-BERT.

Measured effects reported in the paper

Surprising findings

• Geo-ER's neighbor-based performance on Singapore drops steeply at 40% perturbation, falling behind several general ER methods, since it does not denoise neighbor tokens.
• With trivial neighbors, the proposed method still stays at top-tier performance while Geo-ER declines.

Limitations

• Robustness gains diminish at very high perturbation ratios (>=80%), where neighbors provide little useful information.
• Evaluation is restricted to OSM, Yelp, and Foursquare data in four cities.

Experimental setup

Datasets: OpenStreetMap Yelp Foursquare GER dataset from Balsebre et al. 2022

Benchmarks: OSM-FSQ GER dataset (Singapore, Edinburgh, Toronto, Pittsburgh) OSM-Yelp GER dataset (Singapore, Edinburgh, Toronto, Pittsburgh)

Metrics: F1 F1 improvement over best alternative

Baselines: DeepMatcher Ditto-BERT GeoVector SpaBERT GeoLM Geo-ER

Builds on: Geo-ER Ditto Graph Attention Network BERT SpaBERT

Closest papers by meaning

• GeoLLM: Extracting Geospatial Knowledge from Large Language Models (ICLR 2024 · similarity 0.85)
• HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models (AAAI 2025 · similarity 0.83)
• Unifying Knowledge from Diverse Datasets to Enhance Spatial-Temporal Modeling: A Granularity-Adaptive Geographical Embedding Approach (ICML 2025 · similarity 0.83)
• Textual Supervision Enhances Geospatial Representations in Vision-Language Models (ICML 2026 · similarity 0.82)
• Entity Alignment with Noisy Annotations from Large Language Models (NeurIPS 2024 · similarity 0.82)
• Geolocation Representation from Large Language Models Are Generic Enhancers for Spatio-Temporal Learning (AAAI 2025 · similarity 0.82)