harshildarji/JuraNER

German BERT for Legal NER

F1-Score: 99.762

This model is fine-tuned on the German LER dataset, introduced in this paper. The LER dataset provides annotations across 19 fine-grained legal entity classes, capturing the complexity of legal texts in German.

Class-wise Performance Metrics

The table below summarizes the class-wise performance metrics of our improved model:

Abbreviation	Class	Dataset %	F1-Score
PER	Person	3.26	94.47
RR	Judge	2.83	99.56
AN	Lawyer	0.21	92.31
LD	Country	2.66	96.30
ST	City	1.31	91.53
STR	Street	0.25	95.05
LDS	Landscape	0.37	88.24
ORG	Organization	2.17	93.72
UN	Company	1.97	98.16
INN	Institution	4.09	97.73
GRT	Court	5.99	98.32
MRK	Brand	0.53	98.65
GS	Law	34.53	99.46
VO	Ordinance	1.49	95.72
EUN	European legal norm	2.79	97.79
VS	Regulation	1.13	89.73
VT	Contract	5.34	99.22
RS	Court decision	23.46	99.76
LIT	Legal literature	5.60	98.09

Comparison of F1 Scores

Below is a comparison of F1 scores between our previous model, gbert-legal-ner, and JuraNER: