OCR Cleaning of Scientific Texts with LLMs

Correcting Optical Character Recognition (OCR) errors is a major challenge in preprocessing datasets consisting of legacy PDF files. In this study, we develop Large Language Models specially finetuned to correct OCR errors. We experimented with the mT5 model (both the mT5-small and mT5-large configurations), a Text-to-Text Transfer Transformer-based machine translation model, for the post-correction of texts with OCR errors. We compiled a parallel corpus consisting of text corrupted with OCR errors as well as corresponding clean data. Our findings suggest that the mT5 model can be successfully applied to OCR error correction with improving accuracy. The results affirm the mT5 model as an effective tool for OCR post-correction, with prospects for achieving greater efficiency in future research.