gysbert-citation-ner-v2b

Fine-tuned emanjavacas/GysBERT (historical Dutch, 1500-1950) for detecting Roman-law citation spans (label LEG, BIO-2 scheme) embedded in 17th/18th-century Dutch court-register prose. Trained entirely on synthetic Dutch-context sentences with injected citation spans and hard negatives — see notebooks/finetune_gysbert.ipynb in the roman-law-converter project.

Usage

from transformers import pipeline

pipe = pipeline(
    "token-classification",
    model="shashankmc/gysbert-citation-ner-v2b",
    aggregation_strategy="simple",
)
pipe("sulcx als te sien is in L 1 § 2 ff de minoribus conform L 4 D de instru:")

Evaluation

Held-out agreement against the project's regex-grammar reference method on 16585_Deel_2.txt (a manuscript excluded from synthetic-data generation), at a 512-subword serving window:

• F1 67.0%, precision 71.1%, recall 63.3%

Recommended inference threshold: min_score=0.5. This checkpoint's score distribution is less sharply separated than the prior v2 checkpoint, which was calibrated for a 0.9 threshold — using 0.9 here suppresses valid detections. See docs/experiments.md in the source repo for the full experiment log (v1/v2/v2b comparisons, window sweeps, error analysis).

Labels

BIO-2: O, B-LEG, I-LEG — LEG marks a full Roman-law citation span (e.g. L 1 § 2 ff de minoribus).