Yet another BERT NER implementation with 🤗 Transformers
Results on Conll2003 after 15 epochs on validation
| entity | precision | recall | f1-score | support |
|---|---|---|---|---|
| LOC | 0.97 | 0.97 | 0.97 | 2094 |
| MISC | 0.94 | 0.92 | 0.93 | 1268 |
| ORG | 0.95 | 0.95 | 0.95 | 2092 |
| PER | 0.98 | 0.99 | 0.98 | 3149 |
| O | 1.00 | 1.00 | 1.00 | 42759 |
| accuracy | - | - | 0.99 | 51362 |
| macro avg | 0.97 | 0.96 | 0.97 | 51362 |
| weighted avg | 0.99 | 0.99 | 0.99 | 51362 |
TODO:
- Save best model
- Add predict
- Add logging to tensorboard
- Try different datasets
- Make experiments with different hyperparams
- Try to tweak model architecture
pip install -r requirements.txt
python train.py
python predict.py