Ontology Learning from Text: an Analysis on LLM Performance

Ontologies provide a structured framework to represent and integrate domain knowledge. Developing them is a complex and time-consuming task, requiring domain expertise to ensure accuracy and consistency. Ontology learning aims to automate this process by learning full ontologies, or parts of them, from sources such as textual data. In this paper, we research the potential of Large Language Models (LLMs), specifically GPT-4o, in ontology learning, using a real-world use case. We introduce a manually constructed ontology based on knowledge in a news article, and compare it to ontologies extracted using three different prompting approaches over multiple runs. The resulting ontologies are evaluated both quantitatively and qualitatively, to ensure that differences in performance due to modelling choices are also considered. The results show that, while the LLM effectively identifies important classes and individuals, it often does not include properties between classes, and adds inconsistent and incorrect properties between individuals. Prompting on a sentence level leads to more correct individuals and properties, however, quantitative evaluation shows more hallucinations and incorrect triples. Despite these issues, LLMs advance previous ontology learning methods by considering classes, individuals, and properties as a whole, creating a more complete ontology rather than isolated elements. This provides a new perspective on ontology learning and highlights the potential of LLMs to offer a first version of an ontology or an extension to an existing one based on new information.