Skip to content

Schema Generation#

Schema generation creates synthetic RDFS/OWL ontologies from statistical parameters. You define constraints like hierarchy depth and number of relations, and PyGraft-gen generates a valid ontology.

On this page:

About Schema Generation

Schema generation is the original PyGraft algorithm preserved.
The only changes are code modernization and the addition of seeded RNG for deterministic generation.

The Generation Pipeline#

flowchart LR
  SP[Schema Parameters]

  subgraph SG[ ]
    direction TB
    CG[Class Generator]
    RG[Relation Generator]
  end

  S((Schema))

  SP --> CG
  SP --> RG
  CG --> S
  RG --> S

  S -->|Consistency Checking| S

  style SP fill:#f8c9c9,stroke:#c55,stroke-width:2px
  style CG fill:#f8c9c9,stroke:#c55,stroke-width:2px
  style RG fill:#f8c9c9,stroke:#c55,stroke-width:2px
  style S fill:#eee,stroke:#666,stroke-width:2px

  style SG fill:transparent,stroke:#888,stroke-width:2px,stroke-dasharray:6 6

Schema generation happens in three phases: building the class hierarchy, generating relations between classes, and validating the result for logical consistency.

Class Hierarchy Generation#

The class hierarchy defines the taxonomy of entity types in your ontology.

Configuration parameters:

Parameter Description
num_classes Total number of classes
max_hierarchy_depth Maximum levels in the class tree
avg_class_depth Average depth of classes
avg_children_per_parent Branching factor
avg_disjointness Proportion of sibling classes marked disjoint

What gets generated:

  • Classes with rdfs:subClassOf relationships
  • owl:disjointWith declarations between siblings
  • Metadata tracking class depth and relationships

Example output: 

sc:C50 a owl:Class ;
    rdfs:subClassOf owl:Thing .

sc:C41 a owl:Class ;
    rdfs:subClassOf sc:C50 .

sc:C49 a owl:Class ;
    rdfs:subClassOf sc:C50 ;
    owl:disjointWith sc:C35 .

Relation Generation#

Once the class hierarchy exists, relations define how classes connect in the Knowledge Graph.

Configuration parameters:

Parameter Description
num_relations Total number of object properties
relation_specificity How precisely relations target specific classes (higher = more specific)
prop_profiled_relations Proportion with domain/range restrictions
profile_side Whether profiled relations have both domain and range ("both") or at least one ("partial")
prop_symmetric_relations Proportion of symmetric properties
prop_inverse_relations Proportion with inverse property pairs
prop_transitive_relations Proportion of transitive properties
prop_asymmetric_relations Proportion of asymmetric properties
prop_reflexive_relations Proportion of reflexive properties
prop_irreflexive_relations Proportion of irreflexive properties
prop_functional_relations Proportion of functional properties
prop_inverse_functional_relations Proportion of inverse functional properties
prop_subproperties Proportion in subproperty hierarchies

What gets generated:

  • Object properties with rdfs:domain and rdfs:range
  • OWL characteristics (owl:SymmetricProperty, owl:TransitiveProperty, etc.)
  • Inverse property pairs (owl:inverseOf)
  • Property hierarchies (rdfs:subPropertyOf)

Example output: 

sc:R1 a owl:ObjectProperty ;
    rdfs:domain sc:C32 ;
    rdfs:range sc:C18 .

sc:R15 a owl:ObjectProperty,
        owl:SymmetricProperty ;
    rdfs:domain sc:C6 ;
    rdfs:range sc:C6 .

sc:R10 a owl:ObjectProperty ;
    rdfs:domain sc:C41 ;
    rdfs:range sc:C4 ;
    owl:inverseOf sc:R14 .

sc:R6 a owl:ObjectProperty,
        owl:ReflexiveProperty ;
    rdfs:subPropertyOf sc:R11 .

Schema Building & Validation#

After generating classes and relations, PyGraft-gen combines everything into a complete ontology and validates it for logical consistency.

Serialization:

  1. Namespace prefixes are defined
  2. Classes written with hierarchy and disjointness axioms
  3. Relations written with constraints and characteristics
  4. Output serialized to your chosen RDF format

Consistency Checking:

The HermiT reasoner validates the schema for logical consistency. It checks for:

  • Conflicting disjointness declarations
  • Incompatible property characteristics (e.g., a property that's both symmetric and asymmetric)
  • Unsatisfiable classes

If inconsistencies are found, PyGraft-gen reports the schema as inconsistent. HermiT cannot provide details about which specific axioms are problematic, only whether the schema is Consistent or Inconsistent.

Output files:

output_pygraft/2026-01-16_14-30-22/
├── schema.ttl          # The complete ontology
├── class_info.json     # Class hierarchy metadata
└── relation_info.json  # Relation constraints metadata

What's Next#