Georgia Tech's 847K-parameter system rivals larger models on relation extraction
Minimum character distance alone captures most of the relation signal—no pretrained LM needed.
Researchers from Georgia Tech (team DS@GT HIPE) tackled the HIPE-2026 shared task on person-place relation extraction from multilingual historical newspapers (English, French, German). Their key insight: can a lightweight, interpretable system compete without any pretrained language model at the classification stage? They built document-level dependency graphs, extracted proximity-based and part-of-speech features for each entity pair, and used small scikit-learn ensembles or compact Graph Attention Networks (GATs)—all under 847K parameters.
On the official newspaper test set, their best run achieved a macro recall of 0.5142, placing 3rd on the Efficiency profile and mid-table on Accuracy among 17 teams. Two critical findings emerged: first, minimum character distance alone captured most of the signal, with additional engineered features sometimes degrading performance. Second, document-grouped cross-validation was essential—splitting by pairs inflated scores by 25-37 percentage points due to recurring entity mentions across documents, a data-leakage effect that grouped splits eliminate.
The work demonstrates that lightweight, interpretable systems can be surprisingly effective for historical document analysis, offering a cost-efficient alternative to large language models. The open-source approach is practical for digital humanities archives scaling relation extraction across millions of pages.
- System uses only 847K parameters with no pretrained language models—just dependency graphs and proximity features.
- Minimum character distance between entity mentions outperforms all engineered features for classification.
- Document-grouped cross-validation prevents 25-37 point score inflation from data leakage across recurring entity pairs.
Why It Matters
Enables scalable, cost-effective historical newspaper analysis for digital humanities without relying on large, expensive models.