Large language models for intelligent RDF knowledge graph construction: results from medical ontology mapping.

Abstract

The exponential growth of digital data, particularly in specialized domains like healthcare, necessitates advanced knowledge representation and integration techniques. RDF knowledge graphs offer a powerful solution, yet their creation and maintenance, especially for complex medical ontologies like Systematized Nomenclature of Medicine - Clinical Terms (SNOMED CT), remain challenging. Traditional methods often struggle with the scale, heterogeneity, and semantic complexity of medical data. This paper introduces a methodology leveraging the contextual understanding and reasoning capabilities of Large Language Models (LLMs) to automate and enhance medical ontology mapping for Resource Description Framework (RDF) knowledge graph construction. We conduct a comprehensive comparative analysis of six systems-GPT-4o, Claude 3.5 Sonnet v2, Gemini 1.5 Pro, Llama 3.3 70B, DeepSeek R1, and BERTMap-using a novel evaluation framework that combines quantitative metrics (precision, recall, and F1-score) with qualitative assessments of semantic accuracy. Our approach integrates a data preprocessing pipeline with an LLM-powered semantic mapping engine, utilizing BioBERT embeddings and ChromaDB vector database for efficient concept retrieval. Experimental results on a dataset of 108 medical terms demonstrate the superior performance of modern LLMs, particularly GPT-4o, achieving a precision of 93.75% and an F1-score of 96.26%. These findings highlight the potential of LLMs in bridging the gap between structured medical data and semantic knowledge representation, toward more accurate and interoperable medical knowledge graphs.