AllenNLP NER: AllenNLP Models for Named Entity Recognition

Installation

AllenNLP NER uses poetry for library management. If you don't have poetry, please run pip3 install poetry.

git clone [email protected]:himkt/allennlp_ner && cd allennlp_ner
poetry install

Run model training

poetry run allennlp train config/bert.jsonnet --include-pacakge allennlp_ner -s model/bert
poetry run allennlp train config/bert-crf.jsonnet -s model/bert-crf
poetry run allennlp train config/bilstm-crf.jsonnet -s model/bilstm-crf
poetry run allennlp train config/bilstm-cnn-crf.jsonnet -s model/bilstm-cnn-crf

[New!] Hyperparameter Optimization

You can try hyperparameter optimization for BiLSTM-CNN-CRF! It uses Optuna, please read AllenNLP guide if you want to learn more.

poetry run python hpo.py

Or, you can also use allennlp-optuna.

poetry run allennlp tune \
  config/bilstm-cnn-crf-hpo.jsonnet  \
  config/bilstm-cnn-crf.hparams.json \
  --serialization-dir result \
  --metrics best_validation_f1-measure-overall \
  --direction maximize

References

• BiLSTM-CRF [Lample+, NAACL2016]: https://www.aclweb.org/anthology/N16-1030/
• BiLSTM-CNN-CRF [Ma+, ACL2016]: https://www.aclweb.org/anthology/P16-1101/
• BERT [Devlin+, NAACL2019]: https://www.aclweb.org/anthology/N19-1423/
• BERT-CRF [Devlin+, NAACL2019]: https://www.aclweb.org/anthology/N19-1423/ + CRF