Reproduce “PEaCE: A Chemistry-Oriented Dataset for Optical Character Recognition on Scientific Documents” 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
Nan Zhang, Connor Heaton, Sean Timothy Okonsky, Prasenjit Mitra, Hilal Ezgi Toraman
PEaCE (Printed English and Chemical Equations), a chemistry-oriented dataset, was developed for optical character recognition on scientific documents, providing both synthetic and real-world records with LaTeX ground-truth. Experiments with transformer-based models show that multi-domain training, smaller patch sizes, and specific image transformations enhance performance on real-world scientific texts by up to 102.25% compared to models without these enhancements.
Key insights
- OCR model performance experiences a sharp decline when transitioning from synthetic data to real-world scientific documents, validating the importance of realistic evaluation datasets like PEaCE.
- Smaller patch sizes (e.g., 10x10) are crucial for Vision Transformer models to achieve fine-grained character recognition accuracy necessary for complex scientific text.
- Multi-domain training and synthetic data augmentation via targeted image transformations significantly improve OCR model generalization and robustness to imperfections in real-world scientific documents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.