RoBERTa for Multilabel Language Segmentation
Training
RoBERTa fine-tuned on small parts of Open Subtitles, Oscar and Tatoeba datasets (~9k samples per language).
Implemented heuristic algorithm for multilingual training data creation with generation of target masks- https://github.com/n1kstep/lang-classifier
| data source | language |
|---|---|
| open_subtitles | ka, he, en, de |
| oscar | be, kk, az, hu |
| tatoeba | ru, uk |
Validation
The metrics obtained from validation on the another part of dataset (~1k samples per language).
| Validation Loss | Precision | Recall | F1-Score | Accuracy |
|---|---|---|---|---|
| 0.029172 | 0.919623 | 0.933586 | 0.926552 | 0.991883 |
