Towards reliable named entity recognition in the biomedical domain

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Biomedical Named Entity Recognition (BioNER) is vital for biomedical information extraction.
• State-of-the-art BioNER models, including deep learning approaches like BiLSTM-CRF, often struggle with generalization to new datasets.
• Existing methods may not perform reliably on corpora beyond their training data.

Purpose of the Study:

• To evaluate strategies for improving the generalization of BioNER models.
• To address the limitations of current BioNER approaches in cross-corpus performance.
• To enhance the reliability and applicability of BioNER tools in diverse biomedical contexts.

Main Methods:

• Investigated three modifications to the BiLSTM-CRF architecture: variational dropout, transfer learning, and multi-task learning.
• Assessed model performance using both 'in-corpus' (training and testing on the same data) and 'out-of-corpus' (training on one corpus, testing on another) metrics.
• Developed and released Saber, an open-source tool implementing the optimized BioNER models.

Main Results:

• Variational dropout improved out-of-corpus performance by 4.62%.
• Transfer learning enhanced out-of-corpus performance by 6.48%.
• Multi-task learning boosted out-of-corpus performance by 8.42%, with the combination of multi-task learning and variational dropout achieving a 10.75% increase.

Conclusions:

• Multi-task learning and variational dropout are effective strategies for improving BioNER model generalization.
• The developed Saber tool offers a more reliable BioNER solution for diverse biomedical corpora.
• These findings contribute to more robust biomedical information extraction through enhanced BioNER capabilities.