SyRACT: zero-shot biomedical document-level relation extraction with synergistic RAG and CoT.

Area of Science:

• Biomedical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) offer new possibilities for biomedical document-level relation extraction (BioDocRE).
• However, LLMs struggle with hallucination, limited reasoning, and lack of interpretability in BioDocRE tasks.

Purpose of the Study:

• To introduce the SyRACT framework for high-precision BioDocRE.
• To address challenges of hallucination, reasoning, and interpretability in LLM-based BioDocRE.

Main Methods:

• Reframed BioDocRE as a question-answering task for better LLM alignment.
• Utilized a PubMed-derived external database to supply reliable contextual information and reduce hallucinations.
• Developed a specific Chain of Thought (CoT) strategy for BioDocRE to improve reasoning and interpretability.

Main Results:

• SyRACT significantly improved F1 scores across three datasets (CDR, GDA, ADE).
• Improvements were 11.04% (CDR), 9.10% (GDA), and 41.00% (ADE) compared to standard LLM prompting (DocRE method).

Conclusions:

• The SyRACT framework offers a robust solution for high-precision BioDocRE.
• SyRACT effectively mitigates LLM limitations, enhancing both performance and interpretability in biomedical relation extraction.