BioADAPT-MRC: adversarial learning-based domain adaptation improves biomedical machine reading comprehension task

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Machine Learning

Background:

• Biomedical machine reading comprehension (biomedical-MRC) systems require large, human-annotated datasets for high performance.
• Domain knowledge requirement for biomedical data leads to scarcity of labeled data.
• Direct transfer learning from general domains to biomedical domains is hindered by distribution discrepancies.

Purpose of the Study:

• To develop a domain adaptation framework for biomedical-MRC.
• To address the marginal distribution discrepancies between general and biomedical datasets.
• To improve the performance of biomedical-MRC models without relying on biomedical-specific labeled data.

Main Methods:

• An adversarial learning-based domain adaptation framework, BioADAPT-MRC, was developed.
• The framework utilizes neural networks to bridge the distribution gap between domains.
• It relaxes the need for pseudo-label generation for training.

Main Results:

• BioADAPT-MRC achieved state-of-the-art performance on benchmark biomedical-MRC datasets (BioASQ-7b, 8b, 9b).
• The framework demonstrated effectiveness without using any synthetic or human-annotated biomedical data.
• Performance was evaluated against existing state-of-the-art methods.

Conclusions:

• BioADAPT-MRC effectively adapts general domain models to the biomedical domain for MRC tasks.
• The framework offers a viable solution for overcoming data scarcity in biomedical NLP.
• Open-source availability facilitates further research and application.