Skip to content

Plan, apply and roll back changes to a production ontology, with a blast radius report and a proof an auditor can re-check.

3.6 KB
Raw Blame

Pizza Ontology: AI-Native vs Traditional

Comparison of an AI-generated Pizza ontology against the canonical Manchester University Pizza ontology — the most widely used OWL tutorial, with over 20 years of classroom use.

Method

Step Traditional (Protege Tutorial) AI-Native (Claude + Open Ontologies)
1 Open Protege, create blank ontology Describe the domain in natural language
2 Manually create each class via GUI Claude generates complete Turtle
3 Add properties via dialog boxes Properties emerge from domain description
4 Add restrictions one by one Restrictions declared inline with classes
5 Add disjoints via pairwise selection Key disjoints included, exhaustive set skipped
6 Run reasoner to check consistency onto_validate + onto_lint
Time ~4 hours (tutorial estimate) ~5 minutes

Results

Metric Reference AI-Generated Coverage
Total triples 2,332 1,168 50% of size
Classes 99 95 96%
Properties 8 8 100%
Disjoint pairs 398 101 24%
rdfs:labels 96 103 107%

Class Coverage by Category

Category Reference AI Coverage
Toppings 49 49 100%
Named Pizzas 24 24 100%
Pizza Types 13 9 69%
Bases 3 3 100%
Spiciness 4 4 100%
Other 6 6 100%

Missing Classes (4)

All four are teaching artifacts from the Protege tutorial, not domain concepts:

Class Purpose
UnclosedPizza Demonstrates the Open World Assumption
SpicyPizzaEquivalent Alternate syntax exercise for SpicyPizza
VegetarianPizzaEquivalent1 First alternate syntax for VegetarianPizza
VegetarianPizzaEquivalent2 Second alternate syntax for VegetarianPizza

None of these represent actual pizza domain knowledge. They exist only to teach OWL syntax variants.

Disjointness Analysis

The reference declares 398 pairwise disjointness axioms — every sibling class is marked disjoint from every other sibling. This is exhaustive and mechanistic: with N siblings, you get N×(N-1)/2 pairs.

The AI version declares 101 disjoint pairs covering the semantically important ones:

  • All 7 topping categories are mutually disjoint
  • Leaf toppings within each category are mutually disjoint
  • Pizza bases are mutually disjoint
  • Spiciness values are mutually disjoint

The missing disjoints are inferable — a reasoner can derive them from the existing hierarchy. The reference includes them explicitly because Protege generates them automatically via the "Make siblings disjoint" button.

What the AI Adds

  • rdfs:label on every class (103 vs 96 in reference) — better human readability
  • rdfs:comment on key classes — explains modeling decisions
  • Ontology-level metadata — title, description
  • Clean Turtle syntax — readable without tooling

Key Insight

The AI-native approach achieves 96% domain coverage in 50% of the triples. The "missing" 50% is almost entirely:

  • Exhaustive pairwise disjointness axioms (mechanical, not semantic)
  • Teaching-only duplicate classes
  • Verbose annotation triples from Protege

The domain modeling — classes, properties, restrictions, defined classes — is essentially identical.

Run It

cd benchmark
python3 pizza_compare.py

Requires rdflib:

pip install rdflib
python3 pizza_compare.py