[Paper Review] Leveraging Text Repetitions and Denoising Autoencoders in OCR Post-correction

Kai Hakala, Aleksi Vesanto+3 more|arXiv (Cornell University)|Jun 26, 2019

Handwritten Text Recognition Techniques17 references5 citations

TL;DR

This paper proposes a novel method for OCR post-correction that leverages text repetitions in large OCR corpora to estimate realistic error distributions, enabling synthetic training data generation without manual annotations. By training a character-level seq2seq model on this realistic noise, the approach achieves significant improvements over uniform noise baselines and approaches gold-standard performance, reducing WER by 12.38pp over original OCR.

ABSTRACT

A common approach for improving OCR quality is a post-processing step based on models correcting misdetected characters and tokens. These models are typically trained on aligned pairs of OCR read text and their manually corrected counterparts. In this paper we show that the requirement of manually corrected training data can be alleviated by estimating the OCR errors from repeating text spans found in large OCR read text corpora and generating synthetic training examples following this error distribution. We use the generated data for training a character-level neural seq2seq model and evaluate the performance of the suggested model on a manually corrected corpus of Finnish newspapers mostly from the 19th century. The results show that a clear improvement over the underlying OCR system as well as previously suggested models utilizing uniformly generated noise can be achieved.

Motivation & Objective

• To reduce reliance on expensive manually corrected OCR training data by estimating real OCR error patterns from large-scale text repetitions.
• To develop a synthetic data generation method that mimics actual OCR errors using character-level replacement distributions derived from aligned repeating text spans.
• To train a neural seq2seq model on this synthetic data and evaluate its performance against uniform noise and gold-standard baselines.
• To investigate whether training data from corpora with similar linguistic style and time period to the target data is critical for model performance.
• To explore the feasibility of extending the method to handle complex OCR errors such as one-to-many character replacements.

Proposed method

• Text repetitions are identified in a large OCR-processed newspaper corpus, and aligned at the character level to estimate error patterns.
• A character replacement distribution is computed from the differences between repeated spans, capturing systematic OCR errors like 's' → 'f' or 'b' → 'd'.
• Synthetic training data is generated by applying these empirically derived error distributions to clean text from parallel corpora (Gutenberg and Kotus).
• A character-level sequence-to-sequence neural model is trained to denoise the artificially corrupted text, learning to correct OCR errors.
• The model is evaluated using both character error rate (CER) and word error rate (WER) on a held-out test set of manually corrected Finnish newspaper texts.
• The approach is compared against models trained with uniformly generated noise and a gold-standard model trained on manually corrected data.

Experimental results

Research questions

• RQ1 Can text repetitions in large OCR corpora be used to reliably estimate the true error distribution of an OCR system?
• RQ2 Does training a denoising autoencoder on synthetic data generated from empirically derived error distributions outperform training with uniformly generated noise?
• RQ3 How does the choice of clean pre-training corpus (e.g., Gutenberg vs. Kotus) affect model performance when the target data is from the same time period but different genre?
• RQ4 Can a model trained on synthetic data from repetitions achieve performance close to that of a model trained on gold-standard manual corrections?
• RQ5 What are the limitations of current synthetic data generation in capturing complex OCR errors such as one-to-many character substitutions?

Key findings

• The model trained on synthetic data using realistic error distributions (GUT-REAL) achieved a 23.24% word error rate (WER), significantly outperforming the uniform noise baseline (GUT-UNI, 25.35% WER).
• The model trained on synthetic data from the Kotus corpus (KOT-REAL) achieved a WER of 23.45%, demonstrating that corpus genre similarity is less critical than temporal and linguistic consistency.
• The gold-standard model reduced WER to 16.76%, representing a 12.38 percentage point improvement over the original OCR system (29.14% WER), highlighting the gap between synthetic and manual data.
• The model trained on synthetic data from Gutenberg achieved a 6.17pp improvement over original OCR in WER when applied to full sentences, indicating potential for context-aware correction.
• The model trained on gold-standard data performed poorly on full sentences (23.40pp WER decrease), suggesting that larger context requires significantly more data or architectural adaptation.
• The study concludes that text repetition-based error estimation enables high-quality synthetic data generation, making manual annotation less essential for effective OCR post-correction.