LEAP: LLM instruction-example adaptive prompting framework for biomedical relation extraction.

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Large language models (LLMs) show promise for biomedical relation extraction.
• Current demonstration methods in LLMs require optimization for complex biomedical tasks.
• Adaptive tuning strategies are needed to enhance LLM performance in specialized domains.

Purpose of the Study:

• To investigate adaptive tuning methods for large language models (LLMs) in biomedical relation extraction.
• To introduce and evaluate the LLM instruction-example adaptive prompting (LEAP) framework.
• To assess the impact of adaptive task descriptions and examples on LLM performance.

Main Methods:

• Analyzed demonstration components (task descriptions, examples) for LLMs in biomedical data tasks.
• Developed and implemented the LEAP framework with three adaptive tuning methods: instruction, example, and instruction-example adaptive tuning.
• Evaluated LEAP on DDI, ChemProt, and BioRED datasets using LLMs like Llama2 and MedLLaMA.

Main Results:

• Instruction + Options + Example prompting significantly improved F1 scores compared to standard methods for zero-shot LLMs.
• The LEAP framework, particularly example adaptive prompting, outperformed conventional instruction tuning across all tested models.
• MedLLAMA_13B achieved a 95.13 F1 score on ChemProt, demonstrating robust performance on DDI and BioRED datasets.

Conclusions:

• The LEAP framework provides an effective strategy for enhancing LLM training in biomedical relation extraction.
• This approach favors dynamic, contextually enriched prompting over extensive fine-tuning.
• LEAP offers a promising direction for improving LLM capabilities in sophisticated data extraction scenarios.