What a DL Reasoner Actually Computes

What a DL Reasoner Actually Computes

Consistency, classification, realisation, entailment — four jobs, one tool.

Why it matters

Most teams treat reasoners as a black box. Knowing which of the four jobs is being run lets you pick the right profile, debug performance, and explain results to stakeholders.

Going deeper

Practical mapping of job → cost → mitigation:

Job	Typical cost driver	Mitigation
Consistency	A single bad axiom (e.g. unintended disjointness)	Run consistency first and isolate culprit axioms with explanations
Classification	Number of classes; complex class expressions	Pick the smallest OWL profile that works (EL for taxonomy-heavy ontologies)
Realisation	Number of individuals	ABox reasoning is the most expensive; consider materialising once and querying after
Entailment	A single complex query	Targeted reasoning; many reasoners support per-query SPARQL entailment

Almost every production performance story is 'we were doing realisation when we only needed classification' — or 'we were running on full OWL DL when EL would have answered the question 1000× faster'.

A reasoner is a four-job clerk in the records office, doing distinct work under one job title:

1. Consistency — 'do any of the records contradict each other?' (Is there an Alice-is-a-Person and Alice-is-an-Organisation somewhere?) The clerk's first sanity check; if it fails, every other job is meaningless.
2. Classification — 'given all the rules, what's the true taxonomy of classes?' The clerk redraws the org chart so every implicit sub-class relationship is visible.
3. Realisation — 'for each individual, which classes do they belong to?' The clerk re-tags Alice as Adult + EligibleVoter + RegisteredCustomer based on the rules and her known properties.
4. Entailment — 'is this specific statement true given everything we know?' The clerk's yes/no service for downstream queries.

Same clerk, four desks. When 'OWL is slow' you almost always need to ask: which desk is the queue forming at?

Tools & resources

• Protégé 'Explain inference' workflow — click any inferred axiom → see the proof tree: https://protegewiki.stanford.edu/wiki/Explanation_Workbench
• OWL API — the Java library most reasoners are built against: https://github.com/owlcs/owlapi
• OWL 2 Reasoner survey (Matentzoglu et al.): https://www.cs.ox.ac.uk/files/4859/RR-13-04.pdf
• « What is a reasoner doing? » tutorial (Stanford BMIR): https://stanford.edu/class/cs520/2020/notes/Reasoning_Over_Knowledge_Graphs.html