BioM-Transformers: Building Large Biomedical Language Models with BERT, ALBERT and ELECTRA

Abstract

The impact of design choices on the performance of biomedical language models recently has been a subject for investigation. In this paper, we empirically study biomedical domain adaptation with large transformer models using different design choices. We evaluate the performance of our pretrained models against other existing biomedical language models in the literature. Our results show that we achieve state-of-the-art results on several biomedical domain tasks despite using similar or less computational cost compared to other models in the literature. Our findings highlight the significant effect of design choices on improving the performance of biomedical language models.

Model Description

This model was pre-trained on PMC full article for further 64k steps with a batch size of 8192, where we initiate our weights from our model BioM-ALBERT-xxlarge. Thus, the total training steps for this model is 264k+64K=328K steps. The model is very large due to the number of hidden layer size (4096). In order to help researchers with limited resources to fine-tune larger models, we created an example with PyTorch XLA. PyTorch XLA (https://github.com/pytorch/xla) is a library that allows you to use PyTorch on TPU units, which is provided for free by Google Colab and Kaggle. Follow this example to work with PyTorch/XLA Link. In this example we achieve 80.74 micro F1 score on ChemProt task with BioM-ALBERTxxlarge . Fine-tuning takes 43 minutes for 5 epochs .

Check our GitHub repo at https://github.com/salrowili/BioM-Transformers for TensorFlow and GluonNLP checkpoints. We also updated this repo with a couple of examples on how to fine-tune LMs on text classification and questions answering tasks such as ChemProt, SQuAD, and BioASQ.

Colab Notebook Examples

BioM-ELECTRA-LARGE on NER and ChemProt Task

BioM-ELECTRA-Large on SQuAD2.0 and BioASQ7B Factoid tasks

BioM-ALBERT-xxlarge on SQuAD2.0 and BioASQ7B Factoid tasks

Text Classification Task With HuggingFace Transformers and PyTorchXLA on Free TPU

Reproducing our BLURB results with JAX

Finetunning BioM-Transformers with Jax/Flax on TPUv3-8 with free Kaggle resource

Acknowledgment

We would like to acknowledge the support we have from Tensorflow Research Cloud (TFRC) team to grant us access to TPUv3 units.

Citation

@inproceedings{alrowili-shanker-2021-biom,
title = "{B}io{M}-Transformers: Building Large Biomedical Language Models with {BERT}, {ALBERT} and {ELECTRA}",
author = "Alrowili, Sultan and
Shanker, Vijay",
booktitle = "Proceedings of the 20th Workshop on Biomedical Language Processing",
month = jun,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2021.bionlp-1.24",
pages = "221--227",
abstract = "The impact of design choices on the performance of biomedical language models recently has been a subject for investigation. In this paper, we empirically study biomedical domain adaptation with large transformer models using different design choices. We evaluate the performance of our pretrained models against other existing biomedical language models in the literature. Our results show that we achieve state-of-the-art results on several biomedical domain tasks despite using similar or less computational cost compared to other models in the literature. Our findings highlight the significant effect of design choices on improving the performance of biomedical language models.",
}