Augmenting Biomedical Named Entity Recognition with General-domain Resources.

Area of Science:

• Computational Biology
• Natural Language Processing

Background:

• Biomedical named entity recognition (BioNER) model training requires extensive, costly human annotations.
• Existing multi-task learning approaches with multiple BioNER datasets show inconsistent performance gains and potential label ambiguity.

Purpose of the Study:

• To develop a cost-effective BioNER training method using transfer learning from general-domain datasets.
• To improve BioNER model performance, particularly in low-resource scenarios.

Main Methods:

• Proposed GERBERA: a method utilizing general-domain NER datasets for training.
• Employed multi-task learning with a pre-trained biomedical language model, combining target BioNER and general-domain datasets.
• Fine-tuned models specifically on the target BioNER dataset.

Main Results:

• GERBERA models demonstrated superior performance compared to baseline models trained with additional BioNER datasets.
• Achieved an average improvement of 0.9% across eight entity types, outperforming baselines in six.
• Significantly improved performance on data-limited BioNER datasets, with a 4.7% F1 score increase on JNLPBA-RNA.

Conclusions:

• Leveraging cost-effective general-domain NER datasets can augment BioNER models effectively.
• The GERBERA method offers a valuable solution for scenarios with scarce or expensive biomedical annotation resources.
• This approach enhances BioNER model performance and reduces reliance on extensive manual curation.