Answers are grounded in this paper and cite their sources.
Managing data across large organizations is challenging due to the heterogeneity and volume of information systems. Ontologies ensure semantic coherence and interoperability, but their manual development is labor-intensive, motivating the need for automated solutions. Large Language Models offer promising support for automating ontology engineering, but there is still no sufficient knowledge about their performance in this task. Consequently, this study presents a systematic evaluation of these models for ontology generation from structured data in the e-commerce domain. Beyond comparing models, a central contribution is a structured, reproducible, and scalable evaluation workflow enabling consistent assessment of generated ontologies across different approaches. Across 144 executions on six datasets, we compare proprietary and open-source models under base and fine-tuned conditions. Our findings show that GPT models consistently generate the most syntactically valid ontologies with the highest vocabulary reuse, while quantized LLaMA-3-70B obtains the highest qualitative ranking among base models, narrowing the gap with proprietary models. Fine-tuned models score substantially higher than base models in readability and class coverage (+90.32% and +45.50% in mean score), and also in logical consistency and structure (+58.16% and +83.72%), although these remain the lowest-scoring criteria. The evaluation workflow integrates external tools such as OOPS! and OQuaRE, which provide useful quality-related information to identify pitfalls in ontology modeling across models and to analyze their performance along different quality metrics. Overall, our findings highlight both the current advantages of proprietary models and the growing potential of open-source alternatives for scalable, automated ontology engineering in enterprise settings.
No takes yet. Share an insight, caveat, or question.
Riquelme-García et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: