Validate Devanagari post-OCR correction on real scans and with lightweight fine-tuning

Validate CharBM25-based Devanagari post-OCR correction on real scanned documents and determine whether lightweight LoRA fine-tuning with CharBM25-retrieved examples improves correction performance.

Background

The study evaluates Hindi and Marathi post-OCR correction on a synthetic benchmark generated by rendering clean sentences, applying image degradations, and processing them with Tesseract 5. Although CharBM25 retrieval yields substantial gains, the benchmark does not establish how well the method transfers to authentic scanned documents, which may contain different noise, layout, typography, and degradation patterns.

The experiments use open-weight LLMs without task-specific fine-tuning. The conclusion therefore identifies lightweight LoRA fine-tuning, using examples retrieved by CharBM25, as an unresolved direction for determining whether retrieval-guided adaptation can further improve correction quality while retaining the method's computational advantages.