Augmenting biomedical named entity recognition with general-domain resources.

Area of Science:

• Computational biology
• Natural Language Processing
• Machine Learning

Background:

• Biomedical Named Entity Recognition (BioNER) model training demands extensive, costly human annotations.
• Existing multi-task learning approaches for BioNER may not consistently improve performance and can introduce label ambiguity across datasets.

Purpose of the Study:

• To develop a novel training method for BioNER models that reduces reliance on costly human annotations.
• To improve BioNER model performance by leveraging transfer learning from general-domain datasets.

Main Methods:

• Proposed GERBERA, a method utilizing general-domain Named Entity Recognition (NER) datasets for training.
• Employed multi-task learning to train a pre-trained biomedical language model with both BioNER and general-domain datasets.
• Fine-tuned the models specifically on the target BioNER dataset.

Main Results:

• GERBERA models outperformed baseline models trained with additional BioNER datasets across five datasets and eight entity types.
• Achieved an average performance improvement of 0.9% over the best baseline across eight entities.
• Demonstrated significant performance gains on data-limited BioNER datasets, with a 4.7% F1 score improvement on JNLPBA-RNA.

Conclusions:

• Introduced a cost-effective training method augmenting BioNER models with general-domain NER datasets.
• Significantly enhanced BioNER model performance, offering a valuable solution for scenarios with scarce biomedical data.
• Made data, codes, and models publicly available to facilitate further research.