Moose: A New Neuro-Symbolic Method for OWL EL Ontologies

A new neuro-symbolic learning approach called Moose has been developed by researchers for the OWL 2 EL profile, which supports extensive production ontologies such as the Gene Ontology and SNOMED CT. In contrast to current neuro-symbolic techniques that utilize propositional theories or Datalog, Moose transforms an EL++ TBox and a finite ABox into a Sentential Decision Diagram (SDD), functioning as a differentiable weighted-model-counting layer. To enhance the limited expressivity of EL++ in partial supervision scenarios, closure clauses are incorporated beyond the EL++ profile on specified exhaustive families. The method ensures termination, soundness, completeness, and maintains polynomial intermediate sizes, with proofs confirmed in Lean. Moose also introduces the first formal partial-supervision latent-concept-learning task for an OWL EL ontology, facilitating individual classifiers for latent concepts. This research is accessible on arXiv (2608.12961), promoting the fusion of symbolic reasoning and machine learning in ontology contexts.