FlexDoc: Parameterized Sampling for Diverse Multilingual Synthetic Documents for Training Document Understanding Models

TL;DR

FlexDoc tackles the data bottleneck in document understanding by introducing a scalable synthetic data framework built on Stochastic Schemas and Parameterized Sampling. A Dynamic Virtual Grid algorithm organizes elements to preserve layout fidelity while enabling hundreds of thousands of document permutations across languages. Empirical results on KIE tasks show up to $11\%$ absolute improvements in $F1$ with synthetic augmentation and over $90\%$ annotation-effort reductions compared to hard-template methods, with deployment in enterprise settings. The approach offers a practical path to multilingual, layout-aware model training, while acknowledging limitations for fully structured, richly styled documents and outlining directions for semantic value generation and broader cultural adaptation.

Abstract

Developing document understanding models at enterprise scale requires large, diverse, and well-annotated datasets spanning a wide range of document types. However, collecting such data is prohibitively expensive due to privacy constraints, legal restrictions, and the sheer volume of manual annotation needed - costs that can scale into millions of dollars. We introduce FlexDoc, a scalable synthetic data generation framework that combines Stochastic Schemas and Parameterized Sampling to produce realistic, multilingual semi-structured documents with rich annotations. By probabilistically modeling layout patterns, visual structure, and content variability, FlexDoc enables the controlled generation of diverse document variants at scale. Experiments on Key Information Extraction (KIE) tasks demonstrate that FlexDoc-generated data improves the absolute F1 Score by up to 11% when used to augment real datasets, while reducing annotation effort by over 90% compared to traditional hard-template methods. The solution is in active deployment, where it has accelerated the development of enterprise-grade document understanding models while significantly reducing data acquisition and annotation costs.

Figures (10)

• Figure 1: Typical patterns in an invoice
• Figure 2: High-Level description of FlexDoc for generating Synthetic Annotated Documents
• Figure 3: Dynamic Virtual Grid Algorithm
• Figure 7: Overall Algorithm
• Figure 8: Hard Template Based Synthetic Document Generation