A knowledge-enhanced dual-perspective reasoning encoding framework that integrates large language models

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence

Background:

• Joint extraction models excel in general domains but struggle with complex medical texts.
• Large Language Models (LLMs) show promise for few-shot medical knowledge extraction but have architectural limitations for sequence-labeling tasks.

Purpose of the Study:

• To propose MedDPE-KE, a dual-perspective reasoning encoding framework enhanced with LLM-based knowledge for Chinese medical entity-relation extraction.
• To address challenges in medical text processing, including dense terminology and intricate logical structures.

Main Methods:

• A dual-perspective reasoning encoder leveraging Graph Transformer and Attention-based Graph Convolutional Neural Network to capture local and global relationships.
• Integration of LLM-analyzed prior medical knowledge with text features into a triple-decoding layer.
• Utilizing Chinese medical public datasets CMeIE and CMedCasual for evaluation.

Main Results:

• The MedDPE-KE framework significantly improves entity-relation extraction accuracy in Chinese medical texts.
• The method demonstrates particular strength in handling complex medical scenarios.
• Ablation studies confirm the contribution of individual modules to the overall performance.

Conclusions:

• The proposed MedDPE-KE framework effectively enhances Chinese medical entity-relation extraction by integrating dual-perspective reasoning and LLM knowledge.
• This approach offers a promising solution for overcoming limitations of existing models in specialized medical domains.