HIPE-2026: как извлечь связи «человек-место» из исторических газет без нейросетей

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

What the study found

HIPE-2026 for the first time introduced the extraction of person-place relations from multilingual historical newspapers as a separate evaluation track — in English, French, and German. The track's task is to classify the at and isAt relations between pre-annotated mentions of people and locations. The DS@GT HIPE team (team2 in the official results) tested how far a lightweight, interpretable system could go without using a single 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 them with small scikit-learn ensembles or compact Graph Attention Networks. Every run submitted for evaluation stayed under 847 thousand parameters.

*HIPE-2026 — a new track: person-place relations, in English, French, and German

*The at and isAt relations are classified between annotated mentions of people and places

*The DS@GT HIPE team is team2 in the official results

*Best run: macro recall of 0.5142 on Test A (the newspaper set)

*Result: 3rd place for efficiency, mid-table for accuracy among 17 teams

Which feature solves the task?

Minimal character distance between entities alone captures most of the classification signal — this is the first of the team's two main findings. Adding other engineered features gives uneven gains, and sometimes worsens the result. This echoes prior evidence that the distance between arguments dominates the relation-extraction task.

"Minimal character distance alone captures most of the classification

signal," the DS@GT HIPE team's paper on arXiv states.

The practical implication is simple: expensive feature engineering and heavy models don't pay off here. The proximity of two mentions in the text is an almost sufficient signal to predict a relation between a person and a place.

Why does standard cross-validation inflate the result?

Splitting the data at the pair level inflates the score by 25-37 percentage points — because the same entity mentions recur across different documents and create data leakage. The DS@GT HIPE team showed that for this corpus it is critical to apply document-grouped cross-validation: it removes the leakage and returns a fair quality estimate.

This finding matters more than the specific position on the leaderboard. It warns other researchers of historical archives: standard random cross-validation systematically paints too optimistic a picture when mentions of people and places migrate from document to document.

What it means

When digitizing historical archives, scale and processing cost matter as much as accuracy. The DS@GT HIPE team's work shows that lightweight, interpretable methods under a million parameters remain competitive on efficiency, and that a fair 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 types of person-place relations 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 do without large language models?

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