/BioMedicalBertNer

Named Entity Recognition using BERT

Primary LanguagePython

Bio Medical Bert Named Entity Recognition

- Named Entity Recognition for biomedical data using [Scibert](https://arxiv.org/abs/1903.10676) and [Biobert](https://arxiv.org/abs/1901.08746) .
- Different classification architectures like Bert-Crf, Bert-LSTM-Crf on top of Bert Layer.
- Finetuning script is compatible with the datasets provided by [Scibert git repo](https://github.com/allenai/scibert/tree/master/data/ner) data folder.

Finetuning Steps

- add labels file to the data_dir; a labels file should have all the labels from the data, one label per line.
- Edit the config file of your chosen architecture and save.
- Execute the below command

python finettune_with_logger.py
    --model_config_path=/path/to/the/edited/config/file
    --data_dir=/path/to/the/data_dir/
    --logger_file_dir=/path/to/the/dir/where/logs/should/be/saved/
    --labels_file=/path/to/the/labels_file/of/the/dataset/file

Results

Data Architecture Best Test F1
JNLPBA Bert-CRF -
Bert-TokenClassification -
Bert-LSTM-CRF -
NCBI-Disease Bert-CRF
Bert-TokenClassification -
Bert-LSTM-CRF(no finetuning) 0.88
BC5CDR Bert-CRF -
Bert-TokenClassification -
Bert-LSTM-CRF -
SCIIE Bert-CRF -
Bert-TokenClassification -
Bert-LSTM-CRF -