ncbi-nlp/bluebert ncbi-nlp/bluebert

  • Stars
    star
    545
  • Rank 81,520 (Top 2 %)
  • Language
    Python
  • License
    Other
  • Created over 5 years ago
  • Updated over 1 year ago
Twitter logo Share LinkedIn logo Share Facebook logo Share
ncbi-nlp/bluebert
github

ncbi-nlp

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

BlueBERT, pre-trained on PubMed abstracts and clinical notes (MIMIC-III).

BlueBERT

***** New Nov 1st, 2020: BlueBERT can be found at huggingface *****

***** New Dec 5th, 2019: NCBI_BERT is renamed to BlueBERT *****

***** New July 11th, 2019: preprocessed PubMed texts *****

We uploaded the preprocessed PubMed texts that were used to pre-train the BlueBERT models.

This repository provides codes and models of BlueBERT, pre-trained on PubMed abstracts and clinical notes (MIMIC-III). Please refer to our paper Transfer Learning in Biomedical Natural Language Processing: An Evaluation of BERT and ELMo on Ten Benchmarking Datasets for more details.

Pre-trained models and benchmark datasets

The pre-trained BlueBERT weights, vocab, and config files can be downloaded from:

The pre-trained weights can also be found at Huggingface:

The benchmark datasets can be downloaded from https://github.com/ncbi-nlp/BLUE_Benchmark

Fine-tuning BlueBERT

We assume the BlueBERT model has been downloaded at $BlueBERT_DIR, and the dataset has been downloaded at $DATASET_DIR.

Add local directory to $PYTHONPATH if needed.

export PYTHONPATH=.;$PYTHONPATH

Sentence similarity

python bluebert/run_bluebert_sts.py \
  --task_name='sts' \
  --do_train=true \
  --do_eval=false \
  --do_test=true \
  --vocab_file=$BlueBERT_DIR/vocab.txt \
  --bert_config_file=$BlueBERT_DIR/bert_config.json \
  --init_checkpoint=$BlueBERT_DIR/bert_model.ckpt \
  --max_seq_length=128 \
  --num_train_epochs=30.0 \
  --do_lower_case=true \
  --data_dir=$DATASET_DIR \
  --output_dir=$OUTPUT_DIR

Named Entity Recognition

python bluebert/run_bluebert_ner.py \
  --do_prepare=true \
  --do_train=true \
  --do_eval=true \
  --do_predict=true \
  --task_name="bc5cdr" \
  --vocab_file=$BlueBERT_DIR/vocab.txt \
  --bert_config_file=$BlueBERT_DIR/bert_config.json \
  --init_checkpoint=$BlueBERT_DIR/bert_model.ckpt \
  --num_train_epochs=30.0 \
  --do_lower_case=true \
  --data_dir=$DATASET_DIR \
  --output_dir=$OUTPUT_DIR

The task name can be

  • bc5cdr: BC5CDR chemical or disease task
  • clefe: ShARe/CLEFE task

Relation Extraction

python bluebert/run_bluebert.py \
  --do_train=true \
  --do_eval=false \
  --do_predict=true \
  --task_name="chemprot" \
  --vocab_file=$BlueBERT_DIR/vocab.txt \
  --bert_config_file=$BlueBERT_DIR/bert_config.json \
  --init_checkpoint=$BlueBERT_DIR/bert_model.ckpt \
  --num_train_epochs=10.0 \
  --data_dir=$DATASET_DIR \
  --output_dir=$OUTPUT_DIR \
  --do_lower_case=true

The task name can be

  • chemprot: BC6 ChemProt task
  • ddi: DDI 2013 task
  • i2b2_2010: I2B2 2010 task

Document multilabel classification

python bluebert/run_bluebert_multi_labels.py \
  --task_name="hoc" \
  --do_train=true \
  --do_eval=true \
  --do_predict=true \
  --vocab_file=$BlueBERT_DIR/vocab.txt \
  --bert_config_file=$BlueBERT_DIR/bert_config.json \
  --init_checkpoint=$BlueBERT_DIR/bert_model.ckpt \
  --max_seq_length=128 \
  --train_batch_size=4 \
  --learning_rate=2e-5 \
  --num_train_epochs=3 \
  --num_classes=20 \
  --num_aspects=10 \
  --aspect_value_list="0,1" \
  --data_dir=$DATASET_DIR \
  --output_dir=$OUTPUT_DIR

Inference task

python bluebert/run_bluebert.py \
  --do_train=true \
  --do_eval=false \
  --do_predict=true \
  --task_name="mednli" \
  --vocab_file=$BlueBERT_DIR/vocab.txt \
  --bert_config_file=$BlueBERT_DIR/bert_config.json \
  --init_checkpoint=$BlueBERT_DIR/bert_model.ckpt \
  --num_train_epochs=10.0 \
  --data_dir=$DATASET_DIR \
  --output_dir=$OUTPUT_DIR \
  --do_lower_case=true

Preprocessed PubMed texts

We provide preprocessed PubMed texts that were used to pre-train the BlueBERT models. The corpus contains ~4000M words extracted from the PubMed ASCII code version. Other operations include

Below is a code snippet for more details.

value = value.lower()
value = re.sub(r'[\r\n]+', ' ', value)
value = re.sub(r'[^\x00-\x7F]+', ' ', value)

tokenized = TreebankWordTokenizer().tokenize(value)
sentence = ' '.join(tokenized)
sentence = re.sub(r"\s's\b", "'s", sentence)

Pre-training with BERT

Afterwards, we used the following code to generate pre-training data. Please see https://github.com/google-research/bert for more details.

python bert/create_pretraining_data.py \
  --input_file=pubmed_uncased_sentence_nltk.txt \
  --output_file=pubmed_uncased_sentence_nltk.tfrecord \
  --vocab_file=bert_uncased_L-12_H-768_A-12_vocab.txt \
  --do_lower_case=True \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --masked_lm_prob=0.15 \
  --random_seed=12345 \
  --dupe_factor=5

We used the following code to train the BERT model. Please do not include init_checkpoint if you are pre-training from scratch. Please see https://github.com/google-research/bert for more details.

python bert/run_pretraining.py \
  --input_file=pubmed_uncased_sentence_nltk.tfrecord \
  --output_dir=$BlueBERT_DIR \
  --do_train=True \
  --do_eval=True \
  --bert_config_file=$BlueBERT_DIR/bert_config.json \
  --init_checkpoint=$BlueBERT_DIR/bert_model.ckpt \
  --train_batch_size=32 \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --num_train_steps=20000 \
  --num_warmup_steps=10 \
  --learning_rate=2e-5

Citing BlueBERT

@InProceedings{peng2019transfer,
  author    = {Yifan Peng and Shankai Yan and Zhiyong Lu},
  title     = {Transfer Learning in Biomedical Natural Language Processing: An Evaluation of BERT and ELMo on Ten Benchmarking Datasets},
  booktitle = {Proceedings of the 2019 Workshop on Biomedical Natural Language Processing (BioNLP 2019)},
  year      = {2019},
  pages     = {58--65},
}

Acknowledgments

This work was supported by the Intramural Research Programs of the National Institutes of Health, National Library of Medicine and Clinical Center. This work was supported by the National Library of Medicine of the National Institutes of Health under award number K99LM013001-01.

We are also grateful to the authors of BERT and ELMo to make the data and codes publicly available.

We would like to thank Dr Sun Kim for processing the PubMed texts.

Disclaimer

This tool shows the results of research conducted in the Computational Biology Branch, NCBI. The information produced on this website is not intended for direct diagnostic use or medical decision-making without review and oversight by a clinical professional. Individuals should not change their health behavior solely on the basis of information produced on this website. NIH does not independently verify the validity or utility of the information produced by this tool. If you have questions about the information produced on this website, please see a health care professional. More information about NCBI's disclaimer policy is available.

More Repositories

1

BioSentVec

BioWordVec & BioSentVec: pre-trained embeddings for biomedical words and sentences
Jupyter Notebook
567
star
2

BLUE_Benchmark

BLUE benchmark consists of five different biomedicine text-mining tasks with ten corpora.
Python
283
star
3

NegBio

πŸ“° High-performance tool for negation and uncertainty detection in radiology reports
Python
155
star
4

BioWordVec

Python
142
star
5

DeepSeeNet

πŸ‘€ DeepSeeNet is a deep learning framework for classifying patient-based age-related macular degeneration severity in retinal color fundus photographs.
Python
57
star
6

PhenoTagger

PhenoTagger
GAP
50
star
7

Ab3P

Abbreviation definition dection library trained on PubMed abstracts
C
45
star
8

PubMed-Best-Match

Machine-learning based pipeline relying on LambdaMART currently used in PubMed for relevance (Best Match) searches
Python
39
star
9

TeamTat

Text annotation tool for team collaboration
JavaScript
35
star
10

ML_Net

ML-Net is a novel end-to-end deep learning framework for multi-label classification of biomedical tasks. ML-Net combines the label prediction network with a label count prediction network, which can determine the output labels based on both label confidence scores and document context in an end-to-end manner.
Python
33
star
11

TrialGPT

Code and data for TrialGPT.
Python
23
star
12

COVID-19-CT-CXR

COVID-19-CT-CXR, a public database of COVID-19 CXR and CT images
Python
20
star
13

MedCalc-Bench

Benchmarking the medical calculation capabilities of large language models.
Python
17
star
14

ezTag

Web interface that allows users to perform computer-assisted text annotation
JavaScript
15
star
15

BioC-JSON

Tool that converts between BioC XML files and BioC JSON files
Python
15
star
16

NQAC

Python
15
star
17

pubtator-gpt

13
star
18

DeepRel

A convolutional neural network model for relation extraction.
Python
12
star
19

NCBITextLib

Software library for building a large-scale data infrastructure for text mining
C++
7
star
20

BC6PM

Evaluation scripts for BioCreative VI Precision Medicine Track
Python
6
star
21

VarTriage

Python
5
star
22

PubTator-Covid19

4
star
23

CovidTermVar

Python
2
star
24

TaggerOne

General-purpose tagger for joint named entity recognition and normalization. Includes models for both diseases and chemicals (drugs) in biomedical publications.
2
star
25

Medical-Imaging-Causal-Fairness

Python
2
star
26

Unmasking-GPT-Bias

Jupyter Notebook
1
star
27

PhenoRerank

Python
1
star