Efficient Tuning Framework for Resource- Constrained Biomedical Question Answering.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Natural Language Processing

Background:

• Automatic question-answering (QA) systems are crucial for enhancing clinical decision-making.
• Large-scale language models (LLMs) show promise but face challenges in specialized domains like biomedicine due to data privacy and scarcity.
• Efficient fine-tuning methods are needed for LLMs in resource-constrained biomedical settings.

Purpose of the Study:

• To develop an efficient fine-tuning method for pre-trained language models (PLMs) in biomedical QA tasks.
• To address challenges of data privacy and scarcity in the biomedical domain.
• To improve the accuracy and performance of biomedical QA systems.

Main Methods:

• A multi-stage fine-tuning approach is proposed for biomedical QA.
• Incorporation of a multi-prompt-based contrastive learning strategy.
• Integration of a multi-prompt self-consistency voting module to enhance accuracy.

Main Results:

• The proposed method significantly improves the performance of PLMs on biomedical QA tasks.
• Experiments on the PubMedQA dataset demonstrate superior performance compared to domain-specific pre-training models.
• The approach achieves performance comparable to GPT-4 with substantially fewer fine-tuned parameters.

Conclusions:

• The multi-stage fine-tuning approach offers an effective solution for resource-constrained biomedical QA.
• This method enhances the utility of LLMs in the biomedical domain, improving precision and efficiency.
• The strategy provides a scalable and efficient way to adapt powerful language models for specialized scientific applications.