LLMs build domain ontologies but need refinement
GPT-4 generates maritime ontologies but human experts still refine them
Researchers from the University of São Paulo demonstrated that large language models (LLMs) can assist in constructing domain-specific ontologies—structured representations of knowledge that bridge human understanding and machine processing. In their paper titled 'Specific Domain Ontology Construction Using Large Language Models,' Vivian Magri Alcaldi Soares and Renata Wassermann evaluated how well GPT-3.5 and GPT-4 could generate conceptual hierarchies for Brazil's maritime territory, known as the Blue Amazon.
The team generated 20 ontologies automatically, which were then reviewed by human experts. While the models produced overall coherent conceptualizations of the domain, none of the outputs was deemed fully satisfactory without subsequent refinement. The findings suggest that LLMs can significantly expedite the initial stages of ontology development, but human oversight remains critical for accuracy and contextual relevance. The research was presented at NeLaMKRR@KR 2025 and published on arXiv (arXiv:2606.20691).
- LLMs (GPT-3.5 and GPT-4) were tested for automating ontology construction for Brazil's maritime domain (Blue Amazon)
- 20 ontologies were generated and evaluated by human experts, who found them coherent but requiring refinement
- Study highlights LLMs' potential in ontology development while emphasizing the need for human oversight
Why It Matters
LLMs can accelerate domain ontology creation but still need human refinement for real-world applications.