Reproduce “Beyond Human Annotation: Recent Advances in Data Generation Methods for Document Intelligence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dehao Ying, Fengchang Yu, Haihua Chen, Changjiang Jiang, Yurong Li, Wei Lu
This survey by Ying et al. provides the first comprehensive technical map and a novel, resource-centric taxonomy for data generation methods in Document Intelligence, redefining data generation as "supervisory signal production." It introduces a multi-level evaluation framework and analyzes the field's shift towards generative AI and self-supervised learning, with specific paradigm prevalence for perception and reasoning tasks across various DI benchmarks.
Key insights
- The proposed resource-centric taxonomy helps unify fragmented research and provides practical guidance for selecting data generation methods based on specific resource constraints.
- A "Perception-Reasoning Shift" is observed, with Data Generation from Scratch dominating perception tasks and Automated Data Annotation (especially LLM-driven) leading for high-level reasoning tasks.
- The future of data generation involves addressing fidelity and consistency challenges while exploring frontiers like agentic generation, knowledge injection, and co-evolutionary ecosystems with "Negentropy Filters."
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.