Generative Ontology Induction Discovers Schemas from Documents.

Ontology engineering is a significant bottleneck in developing knowledge-intensive AI systems. Existing automated methods often require predefined schemas, are limited to narrow domains, or produce unstructured outputs that are difficult to integrate into downstream pipelines. This research addresses these limitations with Generative Ontology Induction (GOI). GOI is a novel, domain-agnostic framework that leverages large language models to induce a "generative blueprint" from a corpus of example documents. This blueprint includes entities, dimensions, properties, relationships, and constraints, which are then exported as a typed graph in formats like YAML or JSON. The framework also introduces the Node Coverage Score, a new metric to evaluate the fraction of structural ontology nodes present in the generated outputs. Controlled validation across four diverse ontologies—including software invoices, job descriptions, and clinical records—demonstrates GOI's effectiveness. It consistently achieves 95-100% structural backbone coverage, significantly outperforming generic templates which often drop below 60% on less familiar domains. This high coverage holds regardless of the document type's familiarity to the model, indicating robust schema discovery capabilities.