OCR Error Correction Framework

• OCR Error Correction Framework is a system that combines synthetic handwriting generation and neural sequence-to-sequence models to reduce transcription errors.
• The framework employs Bézier curve-based data synthesis with augmentation techniques alongside a fine-tuned T5 model processing 836,000 error–correction pairs.
• Performance improvements are demonstrated via metrics like WAR and CAR, with practical applications in education, archival preservation, and reproducible research.

Optical Character Recognition (OCR) Error Correction Frameworks are specialized post-processing systems designed to reduce transcription errors in text produced by OCR engines. These frameworks address inaccuracies such as character substitutions, word segmentation mistakes, and context-inappropriate substitutions—critical for XML, digital humanities, archival preservation, and downstream NLP applications. Modern frameworks span approaches from character-level statistical models and lexicon-driven correction algorithms to advanced neural sequence-to-sequence architectures and LLM–based infilling, reflecting the increasing complexity and diversity of both source materials and error distributions.

1. Synthetic Handwriting Data Generation Using Bézier Curves

A central challenge for post-OCR correction in Cyrillic and other scripts is the lack of large, annotated corpora pairing OCR errors with their correct forms. The framework described in (Davydkin et al., 2023) addresses this by generating high-fidelity synthetic handwriting data using Bézier curves. Each handwritten character or word is composed from concatenated Bézier segments, defined by four control points:

$$P = (1-t)^3 P_1 + 3(1-t)^2 t P_2 + 3(1-t) t^2 P_3 + t^3 P_4,\quad t \in [0,1]$$

Vector parameters $V_1$, $V_4$ and the careful alignment of segment terminal and start points ($P_2 = P_1 + V_1$, $P_3 = P_4 - V_4$) ensure realistic stroke transitions. Complex augmentation routines—e.g., noise injection, stroke skew, random scaling, and inter-point jitter—simulate the variability of real handwriting observed in educational, historical, and in-the-wild contexts.

This process uses large-scale Russian text corpora scraped from the Internet, producing synthetic “handwritten” images which are then processed by a Handwritten Text Recognition (HTR) model. The resulting OCR errors are paired with their known, ground truth textual input, creating a vast set of error–correction pairs for supervised model training.

2. Sequence-to-Sequence Post-OCR Correction with Pretrained Transformer Models

The backbone of the correction framework is a sequence-to-sequence (seq2seq) model fine-tuned for error correction. Specifically, the architecture leverages the T5 (‘cointegrated/rut5-base-multitask’) transformer, pre-trained on a bilingual Russian–English corpus and supplemented with various multitask learning objectives. The correction model is trained to map HTR outputs (which can include a range of OCR errors) to their error-free reference texts. Inputs are fixed to a context window of 90 symbols—balancing the need for local syntactic and semantic context against model memory and efficiency constraints.

Training utilizes a CrossEntropy loss and the Adam optimizer (learning rate $1 \times 10^{-5}$), with further regularization via dropout ($0.1$). The corpus includes approximately 836,000 error–correction sentence pairs, with batch size $32$ over $12$ epochs, typically converging within a day on a consumer-grade NVIDIA 2080 Ti GPU.

3. Evaluation Metrics, Results, and Error Analysis

Performance evaluation employs both Word Accuracy Rate (WAR) and Character Accuracy Rate (CAR), providing complementary views of correction effectiveness:

• WAR: Proportion of words exactly matched to the ground truth after correction.
• CAR: Fraction of correct characters (including insertions, deletions, substitutions) relative to the reference.

On the HWR200 dataset, the framework improves WAR from 43.2% to 56.0% (scans, raw text scenario), with CAR improving marginally from 78.7% to 78.8%. These metrics are complemented by results on School_notebooks_RU and internal student essay datasets, all indicating robust error reduction in challenging (e.g., poor lighting, non-ideal imaging) conditions.

The incremental improvement in CAR (compared to WAR) suggests that the seq2seq model is particularly effective at addressing errors at the word level—often entire incorrect recognitions—while already-high character-level accuracy means that further gains are more difficult.

4. Applications in Education and Reproducibility

A notable application domain is educational assessment. By comparing raw and post-correction text versions, teachers can efficiently identify which errors originated from handwriting as opposed to OCR misrecognition, directly aiding student feedback. This comparative error highlighting is realized by a simple text diff between pre- and post-correction outputs, allowing for transparent and actionable error reporting.

The methodology is explicitly designed for reproducibility; the paper details the augmentation parameters, corpora utilized, and model hyperparameters. The code and segments for data generation, HTR simulation, and model fine-tuning are available at the authors’ repository, enabling further experimentation and extension to new scripts and languages.

5. Technical and Mathematical Underpinnings

Key technical advances in this framework include:

• Bézier Curve Handwriting Generation: Mathematical modeling of handwriting with explicit parametric (Bézier) representations, ensuring the generation of diverse, smooth, and realistic synthetic text.
• Augmentation Strategies: Randomization of stroke properties and structural perturbations generate variability while preserving script legibility.
• Seq2Seq Model Training: Application of transformer-based T5 models, facilitating the “translation” of erroneous OCR output into canonical text.
• Windowed Context Correction: Bounded context windows cooperate with the model’s memory and attention mechanisms to emphasize local consistency and prevent overfitting to global document structure.

6. Broader Implications and Future Directions

This OCR error correction framework exemplifies the synergy between synthetic data generation and modern neural error correction. Its reproducible pipeline can be adapted to other handwriting scripts and OCR error models, facilitating advancements across languages where annotated real-world corpora are lacking. Potential future directions include:

• Extending the approach to multi-script or code-switched documents.
• Integrating richer linguistic context into training (matic analysis, named entity features).
• Optimizing augmentation parameters to maximize coverage of real-world handwriting variability.

Its approach positions it as a robust baseline for educational, archival, and large-scale document digitization, where correction accuracy must keep pace with the demands of heterogeneous handwriting and degraded input quality.