Reproduce “Combining Synthetic and Real Data for Low-Resource Historical OCR: A Manchu Case Study” 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
Yan Hon Michael Chung, Hanlin Wang
A systematic Manchu OCR study compares synthetic-only, real-only, joint, and sequential training across three vision-language models and a compact CRNN. Adding annotated historical images raises archival word accuracy from a synthetic-only ceiling of 87.92% to as high as 96.28%, while combining complementary recognizers reaches 98.27% on a 753-image Qing-era benchmark.
Key insights
- Moderate amounts of annotated real data substantially reduce the synthetic-to-real gap, and model size is not a reliable proxy for archival OCR performance: the CRNN performs comparably to or better than larger VLMs under real-data training.
- Sequential training has higher point estimates than joint training across models, but the differences are statistically inconclusive; its effect depends on the recognizer and can reduce synthetic-domain accuracy without improving archival results.
- Recognition errors are complementary across models: only seven benchmark images are misrecognized by all three selected systems, enabling ensemble correction, although dictionary lookup does not improve over majority voting on the final test set.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.