UniMRE: a unified framework for zero-shot medicial relation extraction with large language models

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence

Background:

• Medical relation extraction is crucial for understanding unstructured clinical text but suffers from limited labeled data.
• Existing zero-shot methods lack robust domain representation, hindering accurate relation extraction.
• Large Language Models (LLMs) offer advanced contextual understanding, promising for zero-shot tasks.

Purpose of the Study:

• To introduce UniMRE, a Unified framework for zero-shot Medical Relation Extraction utilizing LLMs.
• To address the scarcity of labeled data in medical relation extraction through a novel approach.
• To enhance the capability of LLMs for extracting complex medical relationships in a zero-shot setting.

Main Methods:

• UniMRE employs a knowledge injection strategy to integrate medical expertise into LLMs.
• It generates silver labels which are used to retrieve relevant samples and relation rules.
• A relation extraction agent processes retrieved information, refining labels based on confidence scores.

Main Results:

• UniMRE demonstrated superior performance compared to baseline models on medical datasets.
• The framework successfully extracts relation triplets in a zero-shot setting.
• Knowledge injection and label refinement strategies proved effective in improving accuracy.

Conclusions:

• UniMRE offers an effective solution for zero-shot medical relation extraction by leveraging LLMs.
• The proposed method enhances the extraction of structured medical knowledge from unstructured text.
• This framework has the potential to significantly advance medical informatics and clinical data analysis.