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.

Pre-Trained Models ( PyTorch )

BioM-ELECTRA-Base-Discriminator Link
BioM-ELECTRA-Base-Generator Link
BioM-ELECTRA-Large-Discriminator Link
BioM-ELECTRA-Large-Generator Link
BioM-BERT-PubMed-PMC-Large Link
BioM-ALBERT-xxlarge Link
BioM-ALBERT-xxlarge-PMC Link ( +PMC for 64k steps)
BioM-ELECTRA-Large-SQuAD2 Link
BioM-ALBERT-xxlarge-SQuAD2 Link

Pre-Trained LM Models ( TensorFlow )

Model	Corpus	Vocab	Batch Size	Training Steps	Link
BioM-ELECTRA-Base	PubMed Abstracts	29K PubMed	1024	500K	link
BioM-ELECTRA-Large	PubMed Abstracts	29K PubMed	4096	434K	link
BioM-BERT-Large	PubMed Abstracts + PMC	30K EN Wiki + Books Corpus	4096	690K	link
BioM-ALBERT-xxlarge	PubMed Abstracts	30K PubMed	8192	264k	link
BioM-ALBERT-xxlarge-PMC	PubMed Abstracts + PMC	30K PubMed	8192	+64k	link

SQuAD Fine-Tuned Checkpoints ( TensorFlow )

Model	Exact Match (EM)	F1 Score	Link
BioM-ELECTRA-Base-SQuAD2	81.35	84.20	Link
BioM-ELECTRA-Large-SQuAD2	85.48	88.27	Link
BioM-ELECTRA-Large-MNLI-SQuAD2	85.24	88.01	Link
BioM-ALBERT-xxlarge-SQuAD2	83.86	86.99	Link
BioM-ALBERT-xxlarge-MNLI-SQuAD2	84.35	87.31	Link

We implement transferability between MNLI and SQuAD, which was explained in details by (Jeong, et al., 2020). We detailed our particpiation in BioASQ9B in this Paper. To check the performance of our systems (UDEL-LAB) from the official BioASQ leaderboard visit http://participants-area.bioasq.org/results/9b/phaseB/ .

GluonNLP (MXNet) Checkpoints

More information about GlounNLP https://github.com/dmlc/gluon-nlp

Model	Link
BioM-ELECTRA-Base	Link
BioM-ELECTRA-Large	Link

Model	Exact Match (EM)	F1 Score	Link
BioM-ELECTRA-Base-SQuAD2	80.93	83.86	Link
BioM-ELECTRA-Large-SQuAD2	85.34	88.09	Link

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

Fine-Tunning BioM-Transformers on Question Answering dataset with TPU and Torch XLA

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

BioM-Transfomers Model

@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.",
}

Question Answering with BioM-Transformers

@article{alrowili2021large,
  title={Large biomedical question answering models with ALBERT and ELECTRA},
  author={Alrowili, Sultan and Shanker, K},
  url = "http://ceur-ws.org/Vol-2936/paper-14.pdf",
  journal={CLEF (Working Notes)},
  year={2021}
}

@inproceedings{alrowili2022exploring,
  title={Exploring Biomedical Question Answering with BioM-Transformers At BioASQ10B challenge: Findings and Techniques},
  author={Alrowili, Sultan and Vijay-Shanker, K},
  year={2022},
  organization={CEUR Workshop Bologna, Italy}
}

salrowili/BioM-Transformers