Moose Improves Latent Concept Learning in OWL EL Ontologies

Large-scale production ontologies, such as the Gene Ontology and SNOMED CT, frequently utilize the OWL 2 EL profile. However, existing neuro-symbolic (NeSy) learning methods typically work with propositional theories or Datalog, and the concept of "reasoning-shortcut awareness" has not been explored in ontology-based learning. Researchers have developed Moose, a novel method designed to address these gaps. Moose compiles an OWL EL TBox and ABox into a Sentential Decision Diagram (SDD), which functions as a differentiable weighted-model-counting layer. To overcome the expressivity limitations of OWL EL under partial supervision, Moose incorporates closure clauses outside the standard profile for declared exhaustive families. The method demonstrates termination, soundness, completeness, and polynomial intermediate sizes, with proofs validated in Lean. Moose also defines the first formal task for partial-supervision latent-concept learning over an OWL EL ontology. Evaluations on datasets like MNIST-with-ontology and Pizzaïolo show Moose significantly outperforms propositional-NeSy, fuzzy-logic, and ontology embedding baselines, while also providing the first analysis of reasoning shortcuts in an OWL EL context.