Generative Ontology Induction: Domain-Agnostic Schema Discovery from Document Corpora Using Large Language Models
06:00 · July 21, 2026 · arXiv cs.AI RSS
Ontology engineering remains a critical bottleneck in knowledge-intensive AI systems. Existing automated approaches either depend on predefined schemas, operate within narrow domains, or produce unstructured outputs unsuitable for downstream pipelines. We introduce Generative Ontology Induction (GOI), a domain-agnostic framework that induces a generative blueprint - entities, dimensions, properties, relationships, and constraints - from a corpus of examples and exports it as a typed graph (six node types, seven edge types) in YAML/JSON. We introduce the Node Coverage Score, a novel evaluation metric that measures the fraction of structural ontology nodes (classes, properties, and dimensions) appearing in generated outputs. A controlled generative validation on four contrasting ontologies - a familiar Software Services Invoice schema, a custom Job Description Ontology, a confidential Pain-Management Clinical Visit Record Ontology, and a Professional Services Contract & Statement of Work Ontology - shows that GOI-prompted generation covers 95-100% of the structural backbone in every case; a generic three-field template holds at 97.8% on the invoice schema but drops to 52.2% on the Job Description Ontology, 62.2% on the Pain-Management ontology, and 78.3% on the Professional Services Contract ontology. The structural coverage holds regardless of how familiar the document type is to the model.
Summary
Ontology engineering has long been a manual bottleneck for knowledge-intensive AI systems, as most automated methods either presuppose fixed schemas, remain confined to narrow domains, or emit unstructured results that cannot feed downstream pipelines. Generative Ontology Induction (GOI) tackles this limitation by inducing a reusable generative blueprint directly from a corpus of example documents. The framework extracts recurring entities, dimensions, properties, relationships, and constraints, then represents them as a typed graph comprising six node types and seven edge types.
The induced schema is exported in both YAML and JSON, allowing immediate injection into retrieval-augmented generation, agent memory structures, or knowledge-graph construction workflows without additional graph traversal. A central innovation is the Node Coverage Score, which quantifies structural completeness by measuring the fraction of ontology nodes—classes, properties, and dimensions—that appear in outputs generated from the induced schema. This metric focuses on structural fidelity rather than token-level overlap.
Controlled experiments across four contrasting document classes—a standard Software Services Invoice schema, a Job Description Ontology, a clinical Pain-Management Visit Record Ontology, and a Professional Services Contract & Statement of Work Ontology—demonstrate that GOI-prompted generation consistently recovers 95–100 % of the structural backbone. In contrast, a generic three-field template achieves comparable coverage only on the most familiar invoice schema and falls to 52–78 % on the remaining cases, indicating that coverage is independent of a model’s prior exposure to the document type. The approach also supplies a shared, human-readable artifact intended to align researchers, engineers, and domain experts within cross-functional teams.
The paper supplies full code, an interactive demo, and supplementary materials to support replication and extension.
Why it matters
Offers actionable, reproducible ontology engineering methods directly usable by Dutch AI teams for RAG, knowledge graphs, and ethical structured AI systems; aligns with NL strengths in SME adoption and transparent AI.
