language:
- en
license: apache-2.0
NOTE: bioformer-cased-v1.0 has been renamed to bioformer-8L. All links to bioformer-cased-v1.0 will automatically redirect to bioformer-8L, including git operations. However, to avoid confusion, we recommend updating any existing local clones to point to the new repository URL.
Bioformer-8L is a lightweight BERT model for biomedical text mining. Bioformer-8L uses a biomedical vocabulary and is pre-trained from scratch only on biomedical domain corpora. Our experiments show that Bioformer-8L is 3x as fast as BERT-base, and achieves comparable or even better performance than BioBERT/PubMedBERT on downstream NLP tasks.
Bioformer-8L has 8 layers (transformer blocks) with a hidden embedding size of 512, and the number of self-attention heads is 8. Its total number of parameters is 42,820,610.
The usage of Bioformer-8L is the same as a standard BERT model. The documentation of BERT can be found here.
Vocabulary of Bioformer-8L
Bioformer-8L uses a cased WordPiece vocabulary trained from a biomedical corpus, which included all PubMed abstracts (33 million, as of Feb 1, 2021) and 1 million PMC full-text articles. PMC has 3.6 million articles but we down-sampled them to 1 million such that the total size of PubMed abstracts and PMC full-text articles are approximately equal. To mitigate the out-of-vocabulary issue and include special symbols (e.g. male and female symbols) in biomedical literature, we trained Bioformer’s vocabulary from the Unicode text of the two resources. The vocabulary size of Bioformer-8L is 32768 (2^15), which is similar to that of the original BERT.
Pre-training of Bioformer-8L
Bioformer-8L was pre-trained from scratch on the same corpus as the vocabulary (33 million PubMed abstracts + 1 million PMC full-text articles). For the masked language modeling (MLM) objective, we used whole-word masking with a masking rate of 15%. There are debates on whether the next sentence prediction (NSP) objective could improve the performance on downstream tasks. We include it in our pre-training experiment in case the prediction of the next sentence is needed by end-users. Sentence segmentation of all training text was performed using SciSpacy.
Pre-training of Bioformer-8L was performed on a single Cloud TPU device (TPUv2, 8 cores, 8GB memory per core). The maximum input sequence length was fixed to 512, and the batch size was set to 256. We pre-trained Bioformer-8L for 2 million steps, which took about 8.3 days.
Awards
Bioformer-8L achieved top performance (highest micro-F1 score) in the BioCreative VII COVID-19 multi-label topic classification challenge (https://doi.org/10.1093/database/baac069)
Links
Acknowledgment
Training and evaluation of Bioformer-8L is supported by the Google TPU Research Cloud (TRC) program, the Intramural Research Program of the National Library of Medicine (NLM), National Institutes of Health (NIH), and NIH/NLM grants LM012895 and 1K99LM014024-01.
Questions
If you have any questions, please submit an issue here: https://github.com/WGLab/bioformer/issues
You can also send an email to Li Fang ([email protected], https://fangli80.github.io/).