pprokopidis/elNER18-bert-base-greek-uncased-v1-bs8-e150-lr5e-06

Greek Named Entity Model finetuned on the elNER Dataset

This Greek NER model was fine-tuned by researchers at the Institute for Language and Speech Processing/Athena RC. The model was finetuned on the elNER-18 dataset using the nlpaueb/bert-base-greek-uncased-v1 as backbone LM.

Dataset

The elNER-18 dataset consists of 21K sentences, 623K tokens and 94K annotated named entities for 18 NE classes.

The following 18 named entities are annotated in the train partition:

Class	#
ORG	10944
PERSON	8774
CARDINAL	7343
GPE	6781
DATE	6338
ORDINAL	1438
PERCENT	1437
LOC	1404
NORP	1396
MONEY	1012
TIME	1011
EVENT	962
PRODUCT	668
WORK_OF_ART	608
FAC	567
QUANTITY	565
LAW	235
LANGUAGE	55

Fine-Tuning

Flair version 0.14 was used for fine-tuning.

The model was trained with the following hyper-parameters:

• Batch Size: [8]
• Learning Rate: [5e-05]

Results

• F-score (micro) 0.9173
• F-score (macro) 0.8778
• Accuracy 0.8651

Class	precision	recall	f1-score	support
ORG	0.8931	0.8847	0.8889	1388
PERSON	0.9516	0.9724	0.9619	1051
CARDINAL	0.9330	0.9627	0.9476	911
DATE	0.9403	0.9403	0.9403	838
GPE	0.9282	0.9552	0.9415	826
PERCENT	0.9807	0.9854	0.9831	206
LOC	0.8011	0.7921	0.7966	178
ORDINAL	0.9477	0.9477	0.9477	172
NORP	0.8690	0.8936	0.8811	141
TIME	0.8951	0.9343	0.9143	137
EVENT	0.6395	0.7231	0.6787	130
MONEY	0.9818	0.9730	0.9774	111
PRODUCT	0.7882	0.8072	0.7976	83
WORK_OF_ART	0.8313	0.8214	0.8263	84
FAC	0.6933	0.6753	0.6842	77
QUANTITY	0.8636	0.8769	0.8702	65
LAW	0.8214	0.8214	0.8214	28
LANGUAGE	1.0000	0.8889	0.9412	9
micro avg	0.9112	0.9235	0.9173	6435
macro avg	0.8755	0.8809	0.8778	6435
weighted avg	0.9116	0.9235	0.9174	6435

Files

The Flair training log has also been uploaded to the model hub.