DavidRivasPhD/covid-berts

BERT models pretrained on the CORD-19 Kaggle dataset

Covid-BERTs

This repository contains information on two BERT versions pretrained on a preprocessed version of the CORD-19 dataset, namely ClinicalCovidBERT and BioCovidBERT.

Contribution

This project was inspired by the covid_bert_base model from Deepset and discussions on potential improvements of this model on Kaggle. The contribution of this project is based on two pillars:

• better initialization: the training is initialized with existing BERT versions trained on scientific corpuses, namely ClinicalBERT and BioBERT.
• specialized vocabulary: the customized vocabulary provided on the BioBERT repository, used for training BioBERT-Large v1.1 (+ PubMed 1M), is used to train biocovid-bert-large-cased.

Download

Model	Downloads
clinicalcovid-bert-base-cased	config.json • tensorflow model • pytorch_model.bin • vocab.txt
biocovid-bert-large-cased	config.json • tensorflow model • pytorch_model.bin • vocab.txt

ClinicalCovidBERT

Model and training description

• BERT base default configuration
• Cased
• Initialized from Bio+Clinical BERT
• Using English bert_base_cased default vocabulary
• Using whole-word masking
• Pretrained on a preprocessed version of the CORD-19 dataset including titles, abstract and body text (approx. 1.5GB)
• Training parameters:
  • train_batch_size: 512
  • max_seq_length: 128
  • max_predictions_per_seq: 20
  • num_train_steps: 150000
  • num_warmup_steps: 10000
  • learning_rate: 2e-5

Usage

from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("manueltonneau/clinicalcovid-bert-base-cased")
model = AutoModel.from_pretrained("manueltonneau/clinicalcovid-bert-base-cased")

BioCovidBERT

Model and training description

• BERT large default configuration
• Cased
• Initialized from BioBERT-Large v1.1 (+ PubMed 1M) using their custom 30k vocabulary
• Using whole-word masking
• Pretrained on the same preprocessed version of the CORD-19 dataset including titles, abstract and body text (approx. 1.5GB)
• Training parameters:
  • train_batch_size: 512
  • max_seq_length: 128
  • max_predictions_per_seq: 20
  • num_train_steps: 200000
  • num_warmup_steps: 10000
  • learning_rate: 2e-5

Usage

from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("manueltonneau/biocovid-bert-large-cased")
model = AutoModel.from_pretrained("manueltonneau/biocovid-bert-large-cased")

References

Emily Alsentzer, John Murphy, William Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical BERT embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72-78, Minneapolis, Minnesota, USA. Association for Computational Linguistics.

Lee, Jinhyuk, et al. "BioBERT: a pre-trained biomedical language representation model for biomedical text mining." Bioinformatics 36.4 (2020): 1234-1240. BioBERT

Reimers, N., & Gurevych, I. (2019). Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084.

Acknowledgements

Research supported with Cloud TPUs from Google's TensorFlow Research Cloud (TFRC)