HIPE-2026: How to Extract Person-Place Relations from Historical Newspapers Without Neural Networks

Team DS@GT HIPE showed at the HIPE-2026 shared task that person-place relations can be extracted from historical newspapers without pretrained language models. Their system, under 847 thousand parameters in size, took 3rd place in the efficiency profile among 17 teams, reaching a macro recall of 0.5142. The preprint was published on arXiv (July 2026).

What the study showed

HIPE-2026 for the first time made person-place relation extraction from multilingual historical newspapers a separate evaluation track, covering English, French and German. The track's task is to classify at and isAt relations between pre-annotated mentions of people and locations. Team DS@GT HIPE (team 2 in the official results) tested how far a lightweight, interpretable system could go when it uses no large language model at the relation classification stage.

The DS@GT HIPE system builds a document-level graph from dependency parses, extracts proximity and part-of-speech features for each entity pair, and classifies the pairs with small scikit-learn ensembles or compact Graph Attention Networks. Every run submitted for evaluation stayed under 847 thousand parameters.

• HIPE-2026 is a new track: person-place relations in English, French and German
• The task is to classify at and isAt relations between annotated mentions of people and places
• Team DS@GT HIPE is team 2 in the official results
• Best run: macro recall of 0.5142 on test A (the newspaper set)
• Result: 3rd place in efficiency, mid-table in accuracy among 17 teams

Which feature solves the task?

The minimum character distance between entities captures most of the classification signal on its own — the first of the team's two main findings. Adding other engineered features yields uneven gains and sometimes makes results worse. This echoes earlier findings that the distance between arguments dominates relation extraction.

"Minimum character distance alone captures most of the classification

signal," the DS@GT HIPE team writes in its arXiv paper.

The practical upshot is simple: expensive feature engineering and heavy models do not pay off here. How close two mentions sit in the text is almost enough to predict the relation between a person and a place.

Why does ordinary cross-validation inflate results?

Splitting data at the pair level inflates scores by 25-37 percentage points, because the same entity mentions recur across different documents and cause data leakage. Team DS@GT HIPE showed that on this corpus it is critical to use document-grouped cross-validation, which removes the leakage and restores an honest estimate of quality.

This finding matters more than the team's exact place in the standings. It warns other researchers working with historical archives: standard random cross-validation systematically paints an overly optimistic picture when mentions of people and places migrate from document to document.

What this means

When digitizing historical archives, the scale and cost of processing matter as much as accuracy. The DS@GT HIPE work shows that lightweight, interpretable methods with under a million parameters remain competitive on efficiency, and that an honest evaluation scheme is worth more than one more complex feature.

Frequently asked questions

What are the at and isAt relations in HIPE-2026?

These are the person-place relation types that must be classified between pre-annotated mentions of people and locations in historical newspapers in English, French and German. The track was added to HIPE-2026 as a new evaluation task.

Why did the system get by without large language models?

Team DS@GT HIPE tested how far a lightweight, interpretable approach could go: a graph built from dependency parses plus proximity features, with classification by scikit-learn ensembles or compact Graph Attention Networks. All runs stayed under 847 thousand parameters, which earned the 3rd place in efficiency.