ZeroTuneBio NER: A three-stage framework for zero-shot and zero-tuning biomedical entity extraction using large...

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Biomedical entity extraction is crucial for knowledge discovery.
• Large Language Models (LLMs) show promise but often require extensive fine-tuning.
• Zero-shot capabilities in LLMs for specialized domains like biomedicine are underexplored.

Purpose of the Study:

• To enhance LLM performance in biomedical entity extraction.
• To investigate zero-shot named entity recognition (NER) without LLM fine-tuning.
• To compare the proposed framework against existing models and human annotation.

Main Methods:

• A three-stage NER framework, ZeroTuneBio NER, was developed.
• The framework integrates chain-of-thought reasoning and prompt engineering.
• Evaluation was conducted on disease, chemistry, and gene datasets without task-specific examples or LLM fine-tuning.

Main Results:

• ZeroTuneBio NER achieved an average F1-score improvement of 0.28 over direct LLM queries.
• The framework demonstrated a partial-matching F1-score of approximately 88%.
• Performance rivaled fine-tuned models and surpassed others when excluding strict-matching errors, while optimizing manual annotation speed and cost.

Conclusions:

• LLMs can achieve high-quality NER without fine-tuning, reducing manual annotation needs.
• The ZeroTuneBio NER framework expands LLM applications in biomedical NER.
• The study highlights scalability and suggests future research directions.