Can AI co-design ontologies?
We’re excited to share that HACID researchers presented groundbreaking results at the Extended Semantic Web Conference (ESWC) 2025 on how Large Language Models (LLMs) can support ontology engineering, which is one of the most cognitively demanding tasks in artificial intelligence and knowledge graph development.
What’s the research about?
The paper, titled “Ontology Generation Using Large Language Models”, explores how state-of-the-art LLMs like GPT-4 and OpenAI’s o1-preview can generate OWL ontologies directly from textual requirements such as user stories and competency questions. This process, usually requiring extensive knowledge engineering expertise, is notoriously slow and error-prone. The authors assess two novel prompting strategies (i.e. Memoryless CQbyCQ and Ontogenia) that significantly improve the quality of generated ontologies. In fact, results show that these LLM-based methods outperform novice ontology engineers and previous work presented at ESWC 2024, thus approaching expert-level outputs under certain conditions.
Why does this matter for HACID?
HACID aims to build hybrid human-AI collective intelligence in open-ended domains like work, governance, and science. Ontologies are essential for structuring this intelligence, by defining the concepts, relations, and rules that underpin semantic interoperability. This research showcases how AI can act as a co-pilot in this critical process, accelerating knowledge modelling and reducing barriers to entry for non-experts.
What did we learn?
- LLMs can generate useful OWL ontologies, particularly when guided by high-quality prompts and structural evaluation criteria.
- The Ontogenia technique, in combination with OpenAI’s o1-preview model, produced the best results in the study achieving high coverage of modelling requirements with fewer critical flaws.
- However, some limitations remain especially in terms of superfluous or overlapping ontology elements, and the need for human oversight in validating semantic correctness.
Tools and Data
To support further research, the team released:
- A benchmark dataset of 100 competency questions and 29 user stories
- Open-source prompting pipelines and code
Read the full paper
Lippolis, A.S. et al. (2025). Ontology Generation Using Large Language Models. In ESWC 2025. https://doi.org/10.1007/978-3-031-94575-5_18
News
HACID is an official sponsor of the 2025 edition of the International Semantic Web Research Summer School…
We’re excited to share that HACID researchers presented groundbreaking results at the Extended Semantic…
This webinar explores how Artificial Intelligence can support the growing domain of climate services…
The webinar explored how LLMs can be employed to support collective intelligence, and how they influence…
We are organising a series of webinars to dig into the opportunities offered by hybrid collective intelligence,…
On Tuesday, February the 25th 2025, we started the HACID Webinar series with the first event dedicated…
The first webinar introduced the series with a broad discussion on hybrid collective intelligence. We…
We’re excited to share our latest video on participatory AI, a cornerstone of HACID’s approach to…
The HACID project will be present at the Maker Faire Rome with an interactive game to understand…
We are excited to share our latest findings in the preprint "Human-AI collectives produce the most accurate…
People ➞