Research map / Large language models / Language, safety and interpretability
Reference-Based Post-OCR Processing with LLM for Precise Diacritic Text in Historical Document Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Uses ebook text as a reference with GPT-4o mini to produce pseudo page-to-page OCR labels for classical Vietnamese books, releasing the VieBookRead dataset and fine-tuned Tesseract/Donut baselines.
Problem
Obtaining precise page-level OCR ground truth for diacritic-heavy historical Vietnamese books, where standalone post-OCR spell correction struggles with many error combinations and modern-vs-classical distribution shift.
Approach
Treat content-focused ebooks (no page alignment) as reference sources: fuzzy-match OCR paragraphs to local ebook windows and have an LLM correct them with the reference in context, plus hallucination trimming and strict constraints on minor context-based corrections, yielding pseudo page2page labels without annotation.
Key results
- Pseudo-labels scored 8.72/10 by human graders versus 7.03/10 for the transformer-based Vietnamese spell correction model on a 450-sample subset.
- GPT-4o judge preferred the proposed pseudo-labels in 86.66% of comparisons versus 6.81% for the baseline spell corrector.
- Fine-tuned Tesseract reduced CER from 0.12 to 0.08 and WER from 0.27 to 0.18 compared with the off-the-shelf Tesseract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ebook reference text with LLM-based correction instead of transformer-based Vietnamese spell correction | Post-OCR processing on VieBookRead · VieBookRead sampled subset | mean grading score (out of 10): 7.03 -> 8.72 | Helped |
| Use ebook reference text with LLM-based correction instead of transformer-based Vietnamese spell correction | Post-OCR processing on VieBookRead · VieBookRead | semantic preference rate: 6.81% -> 86.66% | Helped |
| Fine-tune Tesseract on VieBookRead | Tesseract · VieBookRead test | Norm Edit Dist: 0.12 -> 0.08 | Helped |
| Fine-tune Tesseract on VieBookRead | Tesseract · VieBookRead test | WER: 0.27 -> 0.18 | Helped |
Surprising findings
- Fine-tuned Donut performed worse than the two-stage fine-tuned Tesseract on dense book pages (CER 0.14 vs 0.08).
Limitations
- Human evaluation covered only 450 of 1025 sampled instances due to availability.
- Pipeline depends on OCR quality and on the availability of matching ebook references.
- Donut full-page recognition underperforms two-stage approaches on dense text.
Experimental setup
Datasets: VieBookRead ViTextVQA NomNaOCR Vi-BCI
Benchmarks: VieBookRead
Metrics: Normalized edit distance BLEU Precision Recall F1 CER WER human grading score (1-10)
Baselines: GPT-4o Azure Document Intelligence Google Document AI Tesseract EasyOCR Vietnamese transformer spell correction model
Builds on: Donut Tesseract YOLO-World LayoutParser
Closest papers by meaning
- Reading between the Lines: Image-Based Order Detection in OCR for Chinese Historical Documents (AAAI 2024 · similarity 0.83)
- Knowledge-Enhanced Historical Document Segmentation and Recognition (AAAI 2024 · similarity 0.83)
- Predicting the Original Appearance of Damaged Historical Documents (AAAI 2025 · similarity 0.83)
- PP-OCRv5: A Specialized 5M-Parameter Model Rivaling Billion-Parameter Vision-Language Models on OCR Tasks (CVPR 2026 · similarity 0.83)
- Seeing is Believing? Mitigating OCR Hallucinations in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.82)
- DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.