Beyond Human Annotation: Recent Advances in Data Generation Methods for Document Intelligence

TL;DR

This survey provides the first systematic, unified view of data generation for document intelligence, redefining data generation as supervisory-signal production and organizing methods into four resource-centric paradigms: Data Augmentation, Data Generation from Scratch, Automated Data Annotation, and Self-Supervised Signal Construction. It introduces a formal problem framing based on data and label availability, and a multi-level evaluation framework that combines intrinsic data quality with extrinsic task utility. Across four paradigms, the paper inventories techniques from template rendering and rule-based synthesis to LLM-driven labeling and self-supervised pretraining, highlighting how these approaches address data scarcity, privacy, and domain diversity. The work culminates in a vision of a co-evolving DI data ecosystem where generative AI and evaluation mechanisms jointly guide data production, aiming to close fidelity gaps and enable scalable, secure, and knowledge-rich document understanding.

Abstract

The advancement of Document Intelligence (DI) demands large-scale, high-quality training data, yet manual annotation remains a critical bottleneck. While data generation methods are evolving rapidly, existing surveys are constrained by fragmented focuses on single modalities or specific tasks, lacking a unified perspective aligned with real-world workflows. To fill this gap, this survey establishes the first comprehensive technical map for data generation in DI. Data generation is redefined as supervisory signal production, and a novel taxonomy is introduced based on the "availability of data and labels." This framework organizes methodologies into four resource-centric paradigms: Data Augmentation, Data Generation from Scratch, Automated Data Annotation, and Self-Supervised Signal Construction. Furthermore, a multi-level evaluation framework is established to integrate intrinsic quality and extrinsic utility, compiling performance gains across diverse DI benchmarks. Guided by this unified structure, the methodological landscape is dissected to reveal critical challenges such as fidelity gaps and frontiers including co-evolutionary ecosystems. Ultimately, by systematizing this fragmented field, data generation is positioned as the central engine for next-generation DI.

Figures (8)

• Figure 1: Overview of Representative Document Intelligence Tasks, Their Required Data Structures, and Benchmarks.
• Figure 2: The trend of publications on data generation for document intelligence over the past decade. Data is retrieved from Scopus using the query: TITLE-ABS-KEY(("document") AND ("data augmentation" OR "data synthesis" OR "data generation" OR "synthetic data" OR "artificial data")), filtered for publication years 2016--2025.
• Figure 3: The decision flowchart for the taxonomy, guiding paradigm selection based on resource constraints and learning objectives.
• Figure 4: A unified framework for Data Augmentation. The methods are organized hierarchically into four layers based on the augmentation objective: (1) Enhancing Visual Robustness via geometric and physical transformations; (2) Improving Structural Understanding by perturbing layout and reading order; (3) Strengthening Semantic Discrimination through content modification and instruction rewriting; and (4) Automated Strategies that dynamically optimizes augmentation policies.
• Figure 5: Comparison of Data Generation from Scratch paradigms: (1) Systematic Rendering: Populates templates with content to render documents with high annotation precision and controllability; (2) End-to-End Generative AI: Leverages diffusion models or LLMs to synthesize diverse and realistic document images, layouts, and handwritten text directly from noise or prompts, aiming to bridge the fidelity gap with real-world data.