NeurOWL: An LLM-Based Neural-symbolic Framework for Incomplete OWL Ontology Reasoning
06:00 · July 20, 2026 · arXiv cs.AI RSS
OWL ontologies provide a formal knowledge representation framework that enables semantic reasoning, and have been widely adopted across domains such as healthcare and bioinformatics. In practice, however, real-world ontologies are often incomplete, which pose challenges for reasoning. In this work, we focus on a fundamental subsumption reasoning problem: given an incomplete ontology and a candidate (non-entailed) subsumption, determine whether the subsumption is semantically plausible and, if so, providing a logically sound explanation containing potential missing axioms. This task unifies subsumption verification with ontology abduction, and generalizes the latter by removing the need for a predefined candidate set of missing axioms. To address this subsumption reasoning problem, we propose NeurOWL, an end-to-end neuro-symbolic framework that jointly performs verification and abduction, leveraging both formally defined semantics and textual semantics through Large Language Models and ontology embeddings. We evaluate NeurOWL on real-world ontologies across multiple domains, demonstrating strong and robust performance across different domains.
Summary
NeurOWL addresses the challenge of reasoning over incomplete OWL ontologies, which are common in domains such as biomedicine and healthcare because of evolving knowledge and curation costs. Standard description logic reasoners can verify subsumption and produce justifications only when an ontology is complete and consistent. When axioms are missing, valid subsumptions may not be entailed, and conventional abductive methods typically require a predefined candidate set of axioms together with an assumption that the target subsumption holds.
The framework integrates large language models with ontology embeddings to perform both verification of a candidate subsumption and generation of explanations in a single pipeline. It identifies intermediate concepts that could bridge the source and target classes, then proposes minimal sets of missing TBox axioms that would make the subsumption entailed. When no suitable intermediate concept is found, the target is treated directly as a candidate axiom and validated by the language model. This approach removes the need for a fixed hypothesis vocabulary and handles both positive and negative entailment cases.
Embeddings produced by the OnT model capture structural and semantic properties of EL ontologies and supply candidate intermediate concepts, while the language model checks semantic plausibility and constructs explanations. The system supports both training-free operation and fine-tuning. Evaluation on datasets derived from the real-world ontologies FoodOn and SNOMED CT shows F1 scores reaching 97 for plausibility checking and X-F1 scores reaching 89.3 for the correctness of predicted missing axioms, indicating robust performance across varying degrees of incompleteness.
Why it matters
Directly applicable by Dutch research teams working on semantic technologies, healthcare informatics, and ethical neuro-symbolic AI; aligns with EU strengths in transparent knowledge representation and supports reproducibility via open models and datasets.
