This repo contains code to train a POS tagger on Universal Dependencies' Ro RRT dataset.
It is used to eval transformer language models on Romanian Transformers . Maybe it will develop into a usable model sometime in the future. In the mean time, if you want POS tagging, use NLP-Cube, Stanza or Spacy ;)
Here are the evals on the following models, with default params (except large models with batch_size 1 and grad_accumulation 8)
| Model | UPOS F1 | XPOS F1 |
|---|---|---|
| dumitrescustefan/bert-base-romanian-cased-v1 | 0.9821 | 0.9741 |
| dumitrescustefan/bert-base-romanian-uncased-v1 | 0.9826 | 0.9728 |
| racai/distilbert-base-romanian-cased | 0.9637 | 0.9255 |
| readerbench/RoGPT2-base | 0.8982 | 0.8015 |
| readerbench/RoGPT2-medium | 0.9092 | 0.8201 |
| xlm-roberta-base | 0.9706 | 0.9581 |