LDComKG: un marco de doble mejora impulsado por LLM para la finalización de grafos de conocimiento conscientes de la...

The intricate semantics, diverse terminology, and lack of standardization in ancient Chinese medical corpora present formidable obstacles to constructing knowledge graphs (KGs). Traditional rule-based methods exhibit limited efficacy. Leveraging the unprecedented capabilities of Large Language Models (LLMs), this study explores novel solutions to unlock the knowledge locked within these complex historical texts.

To enhance KG construction efficiency and completion effectiveness, we propose LDComKG, an LLM-powered dual-enhanced framework. At its core, our graph data auto-generation module employs advanced LLMs to process hierarchical text chunks (e.g., architectures adept at understanding classical Chinese), intelligently and efficiently extract entities and relationships from unstructured raw texts of ancient Chinese medical corpora. This auto-generation significantly reduces manual effort and improves initial graph quality. Subsequently, we integrate the Leiden algorithm to enhance community structure recognition and GraphSAGE for effective node representation learning, refining the graph for link prediction and KG completion tasks. This synergistic combination leverages the semantic comprehension power of LLMs and the structural analysis power of graph algorithms.

Extensive experiments on annotated ancient medical text datasets demonstrate that our LLM-enhanced LDComKG framework surpasses baseline models. With GPT-4o-mini, it achieves Accuracy: 93.33%, F1-Score: 93.69%, and AUC: 95.87%, while with GPT-4o, performance further improves to Accuracy: 95.94%, F1-Score: 95.07%, and AUC: 97.72%. These substantial improvements validate the framework's effectiveness, particularly highlighting the critical contribution of LLMs in enabling accurate initial graph construction from challenging corpora. Additionally, we conducted comprehensive comparisons with state-of-the-art LLM-KG fusion methods including KEPLER and CoLAKE, as well as systematic evaluations of different LLM backbones across GraphRAG stages, further confirming the robustness and scalability of our framework.

The LDComKG framework demonstrates the transformative potential of integrating advanced large language models with graph representation learning and community detection for processing semantically complex ancient medical corpora. By harnessing LLMs' ability to navigate intricate semantics and terminology, it achieves high accuracy in link prediction and KG completion, providing a scalable, LLM-empowered solution for building high-quality TCM knowledge graphs. This work exemplifies a practical application of LLMs in healthcare knowledge discovery, specifically addressing the critical challenge of digitizing and utilizing traditional medical knowledge, offering significant potential for intelligent TCM information systems.