deprem-ml/adres_ner_v2_bert_128k

Address NER

• Language: Turkish
• PLM: dbmdz/bert-base-turkish-128k-cased
• Macro-F1 Score: 84%
• Dataset: NER v2 dataset
• Hyperparameters: per_device_train_batch_size = 16, per_device_eval_batch_size = 32, num_train_epochs = 5, weight_decay = 0.1, warmup_ratio = 0.1, learning_rate = 5e-5

Model Comparison

	Macro-F1
dbmdz/bert-base-turkish-128k-cased	0.84
dbmdz/bert-base-turkish-cased	0.83
bert-base-multilingual-cased	0.79
dbmdz/electra-base-turkish-mc4-cased-discriminator	0.76
xlm-roberta-base	0.75
dbmdz/convbert-base-turkish-cased	0.70

Class Performance

	support	precision	recall	f1
overall	957	0.84	0.88	0.86
bina	66	0.66	0.74	0.7
bulvar	13	0.92	0.92	0.92
cadde	57	0.77	0.84	0.81
diskapino	70	0.69	0.73	0.71
ilce	117	0.89	0.96	0.92
isim	113	0.86	0.9	0.88
mahalle	120	0.77	0.82	0.79
sehir	146	0.98	0.97	0.97
site	18	0.79	0.61	0.69
sokak	62	0.72	0.74	0.73
soyisim	98	0.94	0.95	0.94
telefonno	77	0.99	1	0.99