New Graph-Based AI Model Speeds Up Ontology Checking by 20x

A new approach to checking the logical consistency of knowledge bases could dramatically reduce the computational burden of semantic reasoning. Researchers have developed GLaMoR (Graph Language Model for Reasoning), a system that applies recent advances in graph-structured neural networks to validate OWL ontologies—machine-readable frameworks for organizing information that power knowledge graphs across domains like biomedicine.

The Problem with Current Methods#

Verifying that ontologies are logically consistent is a critical step in semantic reasoning, the process of inferring new facts from existing knowledge. However, state-of-the-art reasoners require substantial computational resources, and their performance degrades significantly as ontologies grow larger. While previous machine learning efforts have tackled the easier problem of checking smaller data elements (A-Box axioms), the more complex task of verifying core ontology rules (T-Boxes) has largely been neglected.

Large language models have shown strength in natural language reasoning but struggle with formal logical tasks, making them unsuitable for this kind of work.

How GLaMoR Works#

The solution leverages the emerging class of Graph Language Models (GLMs), which can process both graph-structured data and text simultaneously. The GLaMoR pipeline converts OWL ontologies into graph representations and feeds them through a GLM architecture specifically adapted for consistency checking. Researchers tested the approach on real-world ontologies from the NCBO BioPortal repository, translating them into triple formats compatible with the model.

Results#

The GLM-based approach substantially outperforms baseline methods, achieving 95% accuracy while running 20 times faster than classical reasoners. This speed improvement could make consistency checking practical for large, complex knowledge bases that are currently too expensive to verify regularly.