💻 computer science

Diacritic-Aware Post-OCR Correction for Vietnamese Scanned Documents: A Lightweight Baseline for Low-Quality Document Digitization

This paper proposes a lightweight, resource-efficient post-OCR correction pipeline specifically designed to improve the accuracy of Vietnamese scanned documents by addressing diacritic errors through a combination of image preprocessing, dictionary-based correction, and rule-based restoration, offering a practical baseline for low-resource digitization scenarios.

Original authors: Nguyễn Tuệ An, Trần Thị Lan

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nguyễn Tuệ An, Trần Thị Lan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Blurry Glasses" Effect

Imagine you are trying to read a handwritten note or a photocopy of an old document. If the paper is crumpled, the ink is faded, or the photo is taken in bad light, your eyes might struggle to see the small details.

In the world of computers, this is what happens with OCR (Optical Character Recognition). OCR is the technology that scans a picture of text and turns it into editable computer text.

For English, this is usually pretty easy. But for Vietnamese, it's like trying to read a note where the writer uses tiny, invisible ink dots to change the meaning of every word.

In Vietnamese, a word like "ban" (which could mean "committee") can become "bán" (to sell), "bàn" (table), "bản" (copy), or "bạn" (friend) just by adding or moving a tiny accent mark.
When a document is scanned poorly (blurry, low quality), the computer often sees the main letters ("ban") but misses the tiny dots and lines (the accents). It's like reading a sentence where the punctuation is missing, turning "Let's eat, grandma!" into "Let's eat grandma!"

The Solution: A "Smart Spellchecker"

The authors of this paper propose a lightweight fix that acts like a super-smart spellchecker specifically for Vietnamese.

Instead of trying to rebuild the entire camera or scanner (which is expensive and hard), they built a post-OCR correction pipeline. Think of it as a three-step assembly line:

Cleaning the Image (Preprocessing): Before the computer reads the text, they clean up the picture. It's like wiping fog off a car windshield or brightening a dim photo so the computer can see the letters better.
The First Read (OCR): The computer scans the text and gives a "raw" guess. It gets the main letters right but often misses the tiny accents.
The "Diacritic-Aware" Fix (Correction): This is the magic part. The system looks at the messy words and asks:
- "Is this word in our dictionary?"
- "If I remove the accents, does it match a real word?"
- "What words usually go together in this sentence?"

If the computer sees "hoa don," it knows that in Vietnamese, "hoa" and "don" often go together to mean "invoice" (hóa đơn). It uses a dictionary and rules to guess the missing dots and lines, turning the messy "hoa don" back into the correct "hóa đơn."

Why This Approach is Special

Most modern AI tries to learn everything from scratch by reading millions of books. This paper says, "Wait, we don't need a giant brain for this."

Lightweight: It's like using a simple, fast calculator instead of a supercomputer. It doesn't need massive amounts of data or expensive training.
Practical: It's designed for real-world messiness—old receipts, blurry student IDs, and faded historical papers.
Transparent: Because it uses rules and dictionaries, you can see why it made a change, unlike some "black box" AI models that just guess.

What They Found

The authors tested their system on a mix of clean documents and messy, low-quality scans.

The Result: The system significantly reduced errors. It didn't just fix random typos; it specifically fixed the "missing accent" errors that change the meaning of words.
The Analogy: If the raw OCR was like a student who forgot to put periods and commas in an essay, this correction tool is the teacher who goes through and adds them back in, making the essay readable again.

The Limits (What It Can't Do)

The authors are honest about what their tool can't do yet:

Handwriting: It works best on printed text. If the handwriting is messy, the computer might not recognize the base letters at all, and the spellchecker can't help.
Names: If a person has a very rare name that isn't in the dictionary, the system might accidentally "correct" it to a common word.
Tables: If a document has columns that the scanner reads in the wrong order (like reading a table sideways), this text-only fix can't rearrange the layout.

The Bottom Line

This paper offers a simple, affordable, and effective way to fix Vietnamese text after it has been scanned. It acknowledges that while we can't always make the original photo perfect, we can use smart rules to "heal" the text afterwards, making it useful for searching, archiving, and reading again.

Technical Summary: Diacritic-Aware Post-OCR Correction for Vietnamese Scanned Documents

Problem Statement
Optical Character Recognition (OCR) is critical for digitizing scanned documents, yet it faces significant challenges when applied to Vietnamese text, particularly under low-quality conditions (e.g., blur, low resolution, uneven illumination, and compression artifacts). Unlike non-tonal Latin-script languages, Vietnamese relies on a complex system of diacritics and tone marks to encode lexical meaning. A minor error in recognizing these marks—such as confusing ban (to give), bán (to sell), bàn (table), bản (copy), or bạn (friend)—can completely alter the semantic meaning of a word. In degraded scans, these small diacritical marks are often the first elements to be lost or misrecognized, rendering standard OCR outputs unusable for downstream tasks like search, text mining, and archival preservation. While deep learning models exist, they often require large annotated datasets and expensive training, which are not always feasible for practical, low-resource digitization scenarios.

Methodology
The paper proposes a lightweight, diacritic-aware post-OCR correction pipeline designed to operate as a modular stage after any standard OCR engine (e.g., Tesseract, VietOCR, PaddleOCR). The methodology consists of five distinct stages:

Image Preprocessing: The input image undergoes grayscale conversion, denoising, contrast enhancement, adaptive thresholding (or Otsu binarization), and deskewing. This stage aims to maximize the visibility of small diacritics without over-aggressive binarization that might erase them.
OCR Recognition: The preprocessed image is passed to a "black-box" OCR engine to generate raw text tokens ( $T_{raw}$ ).
Error Candidate Detection: The system identifies "suspicious" tokens that require correction. A token is flagged if it:
- Is absent from the Vietnamese dictionary.
- Has a base form (stripped of diacritics) matching multiple dictionary words.
- Contains unusual character combinations or missing tone marks in high-probability contexts.
- Exhibits visually similar character confusion (e.g., o/ô/ơ, a/ă/â, u/ư, e/ê).
Diacritic-Aware Correction: For each candidate, the system strips diacritics to find the base form and retrieves all valid Vietnamese dictionary words sharing that base.
- If the candidate set is empty, an edit-distance search finds visually similar words.
- If multiple candidates exist, they are ranked using a scoring function based on character similarity, diacritic similarity, word frequency, local context, domain dictionary scores, and n-gram probabilities.
- The highest-scoring candidate is selected as the corrected token ( $\hat{w}_i$ ).
Text Reconstruction: The corrected tokens are reassembled into the final text ( $T_{corr}$ ).

The approach explicitly avoids retraining large end-to-end models, relying instead on dictionary lookups, rule-based substitutions, and statistical ranking.

Key Contributions
The paper outlines four primary contributions:

A Lightweight Pipeline: A practical, reproducible post-OCR correction framework specifically targeting diacritic and tone-mark errors in Vietnamese, designed for scenarios with limited data and computational resources.
Error Taxonomy: A defined taxonomy of Vietnamese OCR errors, categorizing issues such as missing diacritics, wrong tone marks, vowel-shape confusion, character substitution, word segmentation errors, and layout-related noise.
Correction Strategy: A dictionary-based and rule-based correction mechanism that functions without large-scale model training, utilizing candidate generation and ranking to resolve ambiguity.
Evaluation Framework: A comprehensive set of metrics and qualitative analyses, including Character Error Rate (CER), Word Error Rate (WER), Diacritic Error Rate (DER), and Exact Word Accuracy, alongside sample document analyses to support future research.

Experimental Results
The authors evaluated the pipeline on a collection of Vietnamese scanned documents featuring various layouts, font sizes, and noise levels. While the specific numerical values in the paper are noted as illustrative for a preprint draft, the reported trends indicate:

Performance Gains: The combination of preprocessing and post-correction yielded the best results. Compared to raw OCR, the full pipeline reduced the Diacritic Error Rate (DER) from 18.95% to 6.91% and improved Exact Word Accuracy from 78.33% to 87.15%.
Quality Sensitivity: The correction module showed significant improvement across all image quality subsets. For instance, on "Low resolution" images, the Word Error Rate (WER) improved by 6.75%, and on "Blurred images," it improved by 8.88%.
Error Distribution: The analysis confirmed that "Missing tone" (38.6%) and "Missing vowel mark" (25.1%) are the dominant error types, validating the focus of the proposed method.

Significance and Claims
The paper positions this work as a simple and reproducible baseline for improving Vietnamese OCR, particularly for low-quality document digitization, historical document analysis, and low-resource language technology.

The authors claim that significant OCR improvement does not necessarily require retraining large, data-hungry models. Instead, a lightweight post-processing module can effectively repair systematic diacritic errors, which are the primary source of semantic degradation in Vietnamese OCR. The method is highlighted as particularly useful for administrative and educational documents where repetitive, domain-specific phrases allow for effective dictionary-based correction.

However, the paper maintains a modest stance regarding limitations. It acknowledges that the method depends heavily on dictionary coverage (struggling with rare names or non-standard spellings), is primarily designed for printed text rather than handwriting, and does not solve complex layout understanding issues (e.g., table column ordering). The authors emphasize that while the current approach is a strong baseline, future work should integrate neural language models (like PhoBERT) and layout-aware understanding to address these remaining challenges.