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

· Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A lightweight system for person-place relation extraction from multilingual historical newspapers was developed for the HIPE-2026 shared task. This system, avoiding large pretrained language models, utilizes dependency parses, 15 engineered proximity-based and part-of-speech features, and small scikit-learn ensembles or compact Graph Attention Networks (GATs). The best submission, Run 1, achieved a macro recall of 0.5142 on Test A, securing 3rd place on the Efficiency profile among 17 participating teams, with parameter counts under 847K. Key findings indicate that minimum character distance alone captures most of the classification signal, and document-grouped cross-validation is essential to prevent significant score inflation (25-37 percentage points) caused by data leakage from recurring entity mentions. Parse quality also influences GAT performance, with GATs proving more beneficial for noisier English parses than for cleaner French parses.

Key takeaway

For NLP Engineers developing relation extraction systems for historical or noisy text, you should prioritize simple proximity features like minimum character distance, as they often provide the strongest signal. Implement document-grouped cross-validation to avoid inflated performance estimates from data leakage. Consider lightweight, feature-engineered approaches over large language models when efficiency is critical, especially for large-scale archival processing.

Key insights

Minimum character distance is the dominant signal for person-place relation extraction in historical texts, often outperforming complex features.

Principles

Method

A three-stage pipeline: document-level graph construction from dependency parses, extraction of 15 proximity-based and POS-based features, and classification using scikit-learn ensembles or small GATs.

In practice

Topics

Code references

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.