This repository provides fine-tuning codes of BioBERT, a language representation model for biomedical domain, especially designed for biomedical text mining tasks such as biomedical named entity recognition, relation extraction, question answering, etc. Please refer to our paper BioBERT: a pre-trained biomedical language representation model for biomedical text mining for more details.
- (12 Sep 2019) Our paper has been accepted to Bioinformatics (link). Please update your BibTex.
- (19 June 2019) Check out our BioBERT based multi-type NER and normalization model (BERN) with its paper!
- (16 May 2019) A new version of BioBERT (v1.1) is now available.
- (15 May 2019) Updated code and dataset for BioASQ
7ed3d7a.
To use BioBERT, we need pre-trained weights of BioBERT, which you can download from NAVER GitHub repository for BioBERT pre-trained weights. Make sure to specify the versions of pre-trained weights used in your works. Also, note that this repository is based on the BERT repository by Google.
All the fine-tuning experiments were conducted on a single TITAN Xp GPU machine which has 12GB of RAM. The code was tested with Python2 and Python3 (We used Python2 for experiments). You might want to install java to use official evaluation script of BioASQ. See requirements.txt for other details.
We provide pre-processed version of benchmark datasets for each task as follows:
Named Entity Recognition: (17.3 MB), 8 datasets on biomedical named entity recognitionRelation Extraction: (2.5 MB), 2 datasets on biomedical relation extractionQuestion Answering: (5.23 MB), 3 datasets on biomedical question answering task.
For details on NER datasets, please see A Neural Network Multi-Task Learning Approach to Biomedical Named Entity Recognition (Crichton et al. 2017). The source of pre-processed datasets are from https://github.com/cambridgeltl/MTL-Bioinformatics-2016 and https://github.com/spyysalo/s800.
For details on QA datasets, please see An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition (Tsatsaronis et al. 2015).
Due to the copyright issue, we provide links of some datasets as follows:
After downloading one of the pre-trained models from NAVER GitHub repository for BioBERT pre-trained weights, unpack it to any directory you want, which we will denote as $BIOBERT_DIR.
Download and unpack the NER datasets provided above (Named Entity Recognition). From now on, $NER_DIR indicates a folder for a single dataset which should include train_dev.tsv, train.tsv, devel.tsv and test.tsv. For example, export NER_DIR=~/bioBERT/biodatasets/NERdata/NCBI-disease. Following command runs fine-tuining code on NER with default arguments.
mkdir /tmp/bioner/
python run_ner.py \
--do_train=true \
--do_eval=true \
--vocab_file=$BIOBERT_DIR/vocab.txt \
--bert_config_file=$BIOBERT_DIR/bert_config.json \
--init_checkpoint=$BIOBERT_DIR/biobert_model.ckpt \
--num_train_epochs=10.0 \
--data_dir=$NER_DIR/ \
--output_dir=/tmp/bioner/
You can change the arguments as you want. Once you have trained your model, you can use it in inference mode by using --do_train=false --do_predict=true for evaluating test.tsv.
The token-level evaluation result will be printed as stdout format. For example, the result for NCBI-disease dataset will be like this:
INFO:tensorflow:***** token-level evaluation results *****
INFO:tensorflow: eval_f = 0.9028707
INFO:tensorflow: eval_precision = 0.8839457
INFO:tensorflow: eval_recall = 0.92273223
INFO:tensorflow: global_step = 2571
INFO:tensorflow: loss = 25.894125
(tips : You should go up a few lines to find the result. It comes before INFO:tensorflow:**** Trainable Variables **** )
Note that this result is the token-level evaluation measure while the official evaluation should use the entity-level evaluation measure.
The results of python run_ner.py will be recorded as two files: token_test.txt and label_test.txt in output_dir.
Use ner_detokenize.py in ./biocodes/ to obtain word level prediction file.
python biocodes/ner_detokenize.py \
--token_test_path=/tmp/bioner/token_test.txt \
--label_test_path=/tmp/bioner/label_test.txt \
--answer_path=$NER_DIR/test.tsv \
--output_dir=/tmp/bioner
This will generate NER_result_conll.txt in output_dir.
Use conlleval.pl in ./biocodes/ for entity-level exact match evaluation results.
perl biocodes/conlleval.pl < /tmp/bioner/NER_result_conll.txt
The entity-level results for NCBI-disease dataset will be like :
processed 24497 tokens with 960 phrases; found: 993 phrases; correct: 866.
accuracy: 98.57%; precision: 87.21%; recall: 90.21%; FB1: 88.68
MISC: precision: 87.21%; recall: 90.21%; FB1: 88.68 993
Note that this is a sample run of an NER model. Performance of NER models usually converges at more than 50 epochs (learning rate = 1e-5 is recommended).
Download and unpack the RE datasets provided above (Relation Extraction). From now on, $RE_DIR indicates a folder for a single dataset. {TASKNAME} means the name of task such as gad or euadr. For example, export RE_DIR=~/bioBERT/biodatasets/REdata/GAD/1 and --task_name=gad. Following command runs fine-tuining code on RE with default arguments.
python run_re.py \
--task_name={TASKNAME} \
--do_train=true \
--do_eval=true \
--do_predict=true \
--vocab_file=$BIOBERT_DIR/vocab.txt \
--bert_config_file=$BIOBERT_DIR/bert_config.json \
--init_checkpoint=$BIOBERT_DIR/biobert_model.ckpt \
--max_seq_length=128 \
--train_batch_size=32 \
--learning_rate=2e-5 \
--num_train_epochs=3.0 \
--do_lower_case=false \
--data_dir=$RE_DIR/ \
--output_dir=/tmp/RE_output/
The predictions will be saved into a file called test_results.tsv in the output_dir. Once you have trained your model, you can use it in inference mode by using --do_train=false --do_predict=true for evaluating test.tsv. Use ./biocodes/re_eval.py in ./biocodes/ folder for evaluation. Also, note that CHEMPROT dataset is a multi-class classification dataset. To evaluate CHEMPROT result, run re_eval.py with additional --task=chemprot flag.
python ./biocodes/re_eval.py --output_path={output_dir}/test_results.tsv --answer_path=$RE_DIR/test.tsv
The result for GAD dataset will be like this:
.tsv
recall : 92.88%
specificity : 67.19%
f1 score : 83.52%
precision : 75.87%
Please be aware that you have to move output_dir to make new model. As some RE datasets are 10-fold divided, you have to make different output directories to train a model with different datasets.
To download QA datasets, you should register in BioASQ website. After the registration, download BioASQ Task B data, and unpack it to some directory $BIOASQ_DIR. Finally, download Question Answering, our pre-processed version of BioASQ-4/5/6b datasets, and unpack it to $BIOASQ_DIR.
Please use BioASQ-*.json for training and testing the model. This is necessary as the input data format of BioBERT is different from BioASQ dataset format. Also, please be informed that the do_lower_case flag should be set as --do_lower_case=False. Following command runs fine-tuining code on QA with default arguments.
python run_qa.py \
--do_train=True \
--do_predict=True \
--vocab_file=$BIOBERT_DIR/vocab.txt \
--bert_config_file=$BIOBERT_DIR/bert_config.json \
--init_checkpoint=$BIOBERT_DIR/biobert_model.ckpt \
--max_seq_length=384 \
--train_batch_size=12 \
--learning_rate=5e-6 \
--doc_stride=128 \
--num_train_epochs=5.0 \
--do_lower_case=False \
--train_file=$BIOASQ_DIR/BioASQ-train-4b.json \
--predict_file=$BIOASQ_DIR/BioASQ-test-4b-1.json \
--output_dir=/tmp/QA_output/
The predictions will be saved into a file called predictions.json and nbest_predictions.json in the output_dir.
Run transform_nbset2bioasqform.py in ./biocodes/ folder to convert nbest_predictions.json to BioASQ JSON format, which will be used for the official evaluation.
python ./biocodes/transform_nbset2bioasqform.py --nbest_path={QA_output_dir}/nbest_predictions.json --output_path={output_dir}
This will generate BioASQform_BioASQ-answer.json in {output_dir}.
Clone evaluation code from BioASQ github and run evaluation code on Evaluation-Measures directory. Please note that you should always put 5 as parameter for -e.
cd Evaluation-Measures
java -Xmx10G -cp $CLASSPATH:./flat/BioASQEvaluation/dist/BioASQEvaluation.jar evaluation.EvaluatorTask1b -phaseB -e 5 \
$BIOASQ_DIR/4B1_golden.json \
RESULTS_PATH/BioASQform_BioASQ-answer.json
As our model is only on factoid questions, the result will be like
0.0 0.4358974358974359 0.6153846153846154 0.5072649572649572 0.0 0.0 0.0 0.0 0.0 0.0
where the second, third and fourth numbers will be SAcc, LAcc and MRR of factoid questions respectively. Note that we pre-trained our model on SQuAD dataset to get the state-of-the-art performance. Please check our paper for details.
- How can I use BioBERT with PyTorch?
- Can I get word/sentence embeddings using BioBERT?
- How can I pre-train QA models on SQuAD?
- What vocabulary does BioBERT use?
Please see LICENSE file for details. Downloading data indicates your acceptance of our disclaimer.
@article{10.1093/bioinformatics/btz682,
author = {Lee, Jinhyuk and Yoon, Wonjin and Kim, Sungdong and Kim, Donghyeon and Kim, Sunkyu and So, Chan Ho and Kang, Jaewoo},
title = "{BioBERT: a pre-trained biomedical language representation model for biomedical text mining}",
journal = {Bioinformatics},
year = {2019},
month = {09},
issn = {1367-4803},
doi = {10.1093/bioinformatics/btz682},
url = {https://doi.org/10.1093/bioinformatics/btz682},
}
For help or issues using BioBERT, please submit a GitHub issue. Please contact Jinhyuk Lee
(lee.jnhk (at) gmail.com), or Wonjin Yoon (wonjin.info (at) gmail.com) for communication related to BioBERT.