Lightweight Person-Place Relation Extraction from Historical Newspapers with Dependency Graphs and Proximity Features

· Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

A lightweight, interpretable system was developed for person-place relation extraction from multilingual historical newspapers, specifically for the HIPE-2026 shared task. This approach, submitted by team DS@GT HIPE, avoids pretrained language models at the relation classification stage, instead building a document-level graph from dependency parses and extracting proximity-based and part-of-speech features. These features are then classified using small scikit-learn ensembles or compact Graph Attention Networks, keeping all submitted runs under 847K parameters. On the official Test A evaluation, the best run achieved a macro recall of 0.5142, ranking 3rd on the Efficiency profile among 17 participating teams. Key findings indicate that minimum character distance alone captures most classification signal, and document-grouped cross-validation is essential to prevent score inflation by 25-37 percentage points due to recurring entity mentions.

Key takeaway

For NLP engineers building relation extraction systems, especially with historical or large-scale textual archives, your feature engineering efforts should prioritize simple proximity measures like minimum character distance. This approach offers high efficiency and interpretability. Crucially, ensure your evaluation methodology employs document-grouped cross-validation to avoid significantly inflated performance metrics caused by data leakage from recurring entity mentions across documents.

Key insights

Minimum character distance is a surprisingly strong signal for person-place relation extraction in historical texts.

Principles

Method

Build a document-level graph from dependency parses, extract proximity and part-of-speech features, then classify with small scikit-learn ensembles or compact Graph Attention Networks.

In practice

Topics

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.