OntoForm: Ontology-Grounded Semantic Data Generation for Enterprise Document Understanding

Abstract

Enterprise forms encode related entities across fields, recurring sections, and document packets, but authentic training data is difficult to obtain because of privacy, annotation cost, and organisation-specific layouts. OntoForm is a semantic-first framework that generates enterprise form packets from a lightweight task ontology, hierarchical schemas, dependency-aware entity sampling, deterministic validation, and template bindings.

Framework

OntoForm treats the business entity, rather than an individual page, as the primary object of generation. An ontology map defines concepts, value types, cardinalities, dependencies, and constraints. A compiler binds those layout-independent concepts to template fields, while the sampler creates a coherent latent entity record before the renderer projects it into one or more forms.

Supervision signals

Every rendered value remains traceable to its ontology path, sampling rule, and template location. The pipeline exports filled forms, pixel-level field boxes, ontology-keyed extraction targets, grounded question-answer pairs, and validation metadata. The controlled insurance instantiation contains 706 forms, 63,318 field annotations, and 7,395 grounded question-answer pairs.

Results

Fine-tuning vision-language models on OntoForm data improves ontology-keyed extraction F1 by 4.99 points and grounded-QA ROUGE-L by 44.1 points over zero-shot baselines. Ontology guidance improves held-out-form extraction F1 by 36.05 points overall and by 42.67 points for concepts absent from training. It also increases exact cross-form reasoning accuracy by 34.17 points.

Why it matters

The framework provides scalable, traceable supervision without using private production documents. It also preserves cross-field and cross-document consistency, supporting document systems that must reason over complete enterprise submissions instead of treating pages as isolated visual artefacts.