PitchayaS/xlm-roberta-estonian-pos

Estonian POS Tagging Model (XLM-RoBERTa-Base)

This model is a fine-tuned version of FacebookAI/xlm-roberta-base for Estonian Part-of-Speech (POS) tagging.
It is trained on the Universal Dependencies Treebank (UDT), specifically:

• Estonian Dependency Treebank (EDT)
• Estonian Web Treebank (EWT)

The model provides strong token-level linguistic annotation performance and can be used for downstream Estonian NLP tasks.

Evaluation Results

POS tagging accuracy on UDT test datasets (EDT + EWT): 0.9775

Label	Precision	Recall	F1-score	Support
ADJ	0.96	0.96	0.96	4902
ADP	0.95	0.97	0.96	1134
ADV	0.97	0.98	0.97	6511
AUX	0.98	0.99	0.98	3409
CCONJ	0.99	0.99	0.99	2510
DET	0.91	0.93	0.92	1142
INTJ	0.85	0.83	0.84	129
NOUN	0.98	0.98	0.98	15336
NUM	0.97	0.95	0.96	1104
PRON	0.98	0.98	0.98	3425
PROPN	0.96	0.95	0.96	3805
PUNCT	1.00	1.00	1.00	9939
SCONJ	0.98	0.98	0.98	1459
SYM	0.85	0.73	0.79	63
VERB	0.99	0.98	0.98	6746
X	0.78	0.53	0.63	75
Accuracy	—	—	0.98	61689
Macro avg	0.94	0.92	0.93	61689
Weighted avg	0.98	0.98	0.98	61689

---