Improving Biomedical ReQA With Consistent NLI-Transfer and Post-Whitening.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Retrieval Question Answering (ReQA) is crucial for information access.
• Dual-Encoder models are effective for general ReQA but under-researched in biomedicine.
• Biomedical ReQA faces challenges due to scarce annotated data, leading to overfitting.

Purpose of the Study:

• To develop a robust Dual-Encoder framework for biomedical ReQA.
• To address the overfitting problem in training biomedical ReQA models.
• To create new ReQA BioASQ datasets for advancing biomedical research.

Main Methods:

• Pre-training Dual-Encoder models on a modified Natural Language Inference (NLI) task to improve transferability.
• Implementing consistent post-whitening to decorrelate sentence embeddings and reduce feature redundancy.
• Training and evaluating the proposed framework on newly built ReQA BioASQ datasets.

Main Results:

• The proposed framework significantly improves the robustness and performance of biomedical ReQA.
• Achieved promising results, outperforming various competitive methods in experiments.
• Demonstrated enhanced transferability from NLI pre-training to biomedical ReQA tasks.

Conclusions:

• The developed framework effectively mitigates overfitting in biomedical ReQA.
• The modified NLI pre-training and post-whitening techniques enhance Dual-Encoder model performance.
• This work provides a valuable resource and methodology for biomedical information retrieval research.