A Multi-LLM Agent System for Modular Ontology Population: A Case Study on ADHD

Modular ontologies play a crucial role in structuring and interpreting complex knowledge, where they enable the representation of contextual information from diverse sources. To fully exploit these ontological models, a population phase is essential, involving the enrichment of the ontology with concrete instances. However, the automatic population of such ontologies from heterogeneous sources remains a significant challenge. Traditional rule-based or supervised learning approaches require extensive human supervision and often lack generalizability. The use of Large Language Models (LLMs) has recently emerged as a promising alternative, but their limitations become evident in complex modular architectures. This study proposes a multi-agent approach for the automatic population of modular ontologies. Each agent is an LLM guided by precise instructions and enhanced with appropriate tools, dedicated to one or more ontological modules. Our methodology combines advanced prompt engineering techniques with strategies for hallucination mitigation to ensure extractions that are both accurate and compliant with the ontology’s specifications. Our experimental setup involves extracting RDF triples from diagnostic PDF reports and activity schedules stored in a MongoDB database, aimed at the automatic population of an ADHD ontology. The results show that our multi-agent approach consistently outperforms mono-agent methods relying on a single LLM, regardless of the configuration used. We observe significant gains in precision, recall, and F1-score as well as a notable reduction in errors and hallucinations. This study demonstrates both the feasibility and effectiveness of a multi-agent architecture for the automatic population of a modular ontology.