The paper introduces OntoEKG, a large-language-model-driven pipeline intended to reduce the manual effort involved in building domain-specific ontologies for enterprise knowledge graphs. It addresses the problem of turning unstructured enterprise data into structured semantic models, a process the authors describe as resource-intensive and dependent on domain expertise. OntoEKG divides the task into two stages: an extraction module identifies core classes and properties, while an entailment module organizes those elements into a logical hierarchy. The resulting ontology is serialized in the standard RDF format. Because comprehensive end-to-end benchmarks for ontology construction are lacking, the authors create an evaluation dataset from documents in the Data, Finance, and Logistics sectors. In experiments, the system achieves a fuzzy-match F1 score of 0.724 in the Data domain. The paper presents this result as evidence of potential for LLM-assisted ontology construction, while also identifying unresolved limitations in defining the scope of an ontology and performing hierarchical reasoning. The work is scheduled for presentation at the 20th International Conference on Semantic Computing.
