Ontology Reuse in LLM-Generated Ontologies: A Comparative Evaluation Framework

Abstract

Ontologies are meant to be reused; however, reusing ontologies in practice is highly challenging. Recent advances in Large Language Models (LLMs) offer new possibilities for ontology engineering and competency question generation, but their potential for explicit ontology reuse is underexplored. In this paper, we introduce a framework for evaluating ontology reuse in ontologies generated by LLMs. The framework measures reuse across four dimensions: lexical reuse, structural reuse, logical consistency, and reuse depth. We systematically examine ontology reuse in LLM-generated ontologies across four models (GPT-4o-mini, GPT-4.1, Qwen2.5-7B, and Llama-3.1-8B) with various prompting strategies using established energy and IoT domain ontologies such as SAREF and the Fiesta-IoT Ontology. We observe that unstructured prompts result in minimal or no reuse, while reuse-orientated prompting increases the number of classes aligned with existing ontologies. Our results demonstrate that while LLMs achieve lexical reuse with reference ontologies in a controlled reuse-oriented prompting technique, deep structural and axiomatic reuse remains limited. This study highlights that effective ontology reuse with LLMs requires dedicated prompting, alignment mechanisms, LLM-ready ontology reuse methodologies and hybrid human-in-the loop workflows to ensure ontology reuse operations are considered and implemented.

This preprint corresponds to a paper accepted at the ELMKE 2026 Workshop, co-located with ESWC 2026. The manuscript is shared to facilitate early dissemination and discussion of the research results.