GLaMoR: Consistency Checking of OWL Ontologies using Graph Language Models

Details

OWL ontologies are widely used in knowledge management, biomedical informatics, and semantic web applications. Checking ontology consistency — verifying that axioms do not contradict each other — is a fundamental step, but classical reasoners become computationally prohibitive as ontology sizes grow. GLaMoR (Graph Language Model for Reasoning) reformulates consistency checking as a machine learning task. It transforms OWL ontologies into graph-structured triples and adapts the Graph Language Model (GLM) architecture, which can process both graph structure and text simultaneously, to classify ontology consistency. Evaluation on ontologies from the NCBO BioPortal repository shows 95% accuracy — outperforming all baselines — and 20x faster inference than classical reasoners. Importantly, prior ML approaches for consistency checking addressed only A-Box axioms (instance data); GLaMoR also handles T-Boxes (class definitions and property constraints), which had been unaddressed. Accepted at WI-IAT 2026. This has practical relevance for biomedical knowledge graph teams, enterprise data management, and anyone building systems on top of large ontologies where reasoning performance is a bottleneck. The 20x speed improvement makes incremental consistency checking during ontology development feasible for the first time.