This improved version adds native German training data (municipal decisions from Freiburg and Bamberg) on top of the original Dutch (Ghent) and English (translated) data, improving detection quality on German text and on text translated from German.

Base model: xlm-roberta-base
Languages: English (en), Dutch (nl), German (de)
Labels (BIO): DATE, LOCATION, LEGAL_GROUND, ADMINISTRATIVE_BODY, MANDATARY

Quickstart

python

from transformers import pipeline

ner = pipeline("token-classification", model="PedroDKE/multilingual-ner-abb-improved", aggregation_strategy="simple")

# German
print(ner("Der Stadtrat der Stadt Bamberg hat am 12. März 2024 beschlossen."))

# Dutch
print(ner("De gemeenteraad van Gent heeft op 12 maart 2024 besloten."))

# English
print(ner("The city council of London decided on March 12, 2024."))

Evaluation

strict (seqeval, entity-level) micro on held-out validation set.

Language	Samples	Precision	Recall	F1
ALL	920	0.766	0.823	0.794
EN	57	0.675	0.705	0.690
NL	69	0.751	0.783	0.767
DE	794	0.785	0.855	0.818

Per-label (strict) — ALL

Label	Precision	Recall	F1	Support
ADMINISTRATIVE_BODY	0.76	0.86	0.81	459
DATE	0.84	0.88	0.86	773
LEGAL_GROUND	0.82	0.85	0.84	362
LOCATION	0.61	0.66	0.63	678
MANDATARY	0.82	0.88	0.85	580

Per-label (strict) — EN

Label	Precision	Recall	F1	Support
ADMINISTRATIVE_BODY	0.64	0.72	0.68	64
DATE	0.67	0.66	0.67	59
LEGAL_GROUND	0.63	0.75	0.69	32
LOCATION	0.38	0.62	0.47	32
MANDATARY	0.85	0.72	0.78	138

Per-label (strict) — NL

Label	Precision	Recall	F1	Support
ADMINISTRATIVE_BODY	0.60	0.67	0.63	58
DATE	0.77	0.91	0.84	104
LEGAL_GROUND	0.67	0.74	0.71	47
LOCATION	0.70	0.61	0.65	174
MANDATARY	0.85	0.91	0.88	188

Per-label (strict) — DE

Label	Precision	Recall	F1	Support
ADMINISTRATIVE_BODY	0.81	0.92	0.86	337
DATE	0.87	0.90	0.89	610
LEGAL_GROUND	0.88	0.88	0.88	283
LOCATION	0.61	0.68	0.64	472
MANDATARY	0.80	0.95	0.87	254

lblod/multilingual-ner-abb-improved

PedroDKE/multilingual-ner-abb-improved