[Paper Review] Upcycle Your OCR: Reusing OCRs for Post-OCR Text Correction in Romanised Sanskrit

Amrith Krishna, Bodhisattwa Prasad Majumder+2 more|arXiv (Cornell University)|Sep 6, 2018

Natural Language Processing Techniques32 references4 citations

TL;DR

This paper proposes a post-OCR text correction model for Romanised Sanskrit that reuses pre-trained OCR systems by applying a sequence-to-sequence model with a copying mechanism (CopyNet) to correct OCR errors. It achieves a 7.69% higher Character Recognition Rate (CRR) than prior state-of-the-art models and improves OCR output from 35.76% CRR to 87.01% CRR, with human evaluations showing faster and more accurate corrections from the proposed model.

ABSTRACT

We propose a post-OCR text correction approach for digitising texts in Romanised Sanskrit. Owing to the lack of resources our approach uses OCR models trained for other languages written in Roman. Currently, there exists no dataset available for Romanised Sanskrit OCR. So, we bootstrap a dataset of 430 images, scanned in two different settings and their corresponding ground truth. For training, we synthetically generate training images for both the settings. We find that the use of copying mechanism (Gu et al., 2016) yields a percentage increase of 7.69 in Character Recognition Rate (CRR) than the current state of the art model in solving monotone sequence-to-sequence tasks (Schnober et al., 2016). We find that our system is robust in combating OCR-prone errors, as it obtains a CRR of 87.01% from an OCR output with CRR of 35.76% for one of the dataset settings. A human judgment survey performed on the models shows that our proposed model results in predictions which are faster to comprehend and faster to improve for a human than the other systems.

Motivation & Objective

• To address the lack of available datasets and OCR systems for Romanised Sanskrit by proposing a post-OCR correction framework that reuses existing OCR models trained on other Roman-alphabet languages.
• To improve the robustness and accuracy of OCR output for Romanised Sanskrit texts, which often suffer from high error rates due to poor scan quality and script-specific challenges.
• To evaluate whether sequence-to-sequence models with copying mechanisms can outperform traditional sequence labelling models in monotone sequence correction tasks for low-resource languages like Sanskrit.
• To demonstrate that synthetic data generation with realistic distortions can effectively train models for real-world OCR error correction without requiring large-scale annotated datasets.
• To validate the model’s practical utility through human evaluation, measuring correction speed and accuracy of human annotators on model predictions.

Proposed method

• The proposed model uses a sequence-to-sequence architecture with an attention mechanism and a copying mechanism (CopyNet) to correct OCR errors in unsegmented Romanised Sanskrit text.
• Synthetic training images are generated using digitised Sanskrit texts, with controlled distortions (e.g., noise, blur, erosion) to emulate real-world scan variations from two distinct settings.
• The model is trained on synthetic data that mimics the visual characteristics of scanned manuscripts, including variations in DPI, contrast, and textline quality.
• The copying mechanism allows the model to retain correct characters from the input OCR output, reducing errors from misrecognition of similar-looking graphemes.
• The model is evaluated on two real test sets: one from high-quality scans and another from low-quality scans (60 DPI), simulating OCR robustness under poor conditions.
• A human evaluation study compares correction speed and accuracy across OCR output, CRF-based model (PCRF), and the proposed CopyNet model, using anonymised predictions from 45 test strings.

Experimental results

Research questions

• RQ1 Can a sequence-to-sequence model with a copying mechanism outperform traditional sequence labelling models in correcting OCR errors for Romanised Sanskrit?
• RQ2 To what extent can synthetic data with realistic distortions improve the robustness of a post-OCR correction model for low-resource languages?
• RQ3 How does the performance of the proposed model compare to commercial OCR systems (e.g., Google Cloud Vision) on low-quality scans of Romanised Sanskrit texts?
• RQ4 Does the proposed model produce outputs that are more comprehensible and easier to correct for humans than baseline systems?
• RQ5 Can the model generalize across different manuscript styles and OCR quality levels without retraining?

Key findings

• The proposed CopyNet model achieves a 7.69% higher Character Recognition Rate (CRR) than the prior state-of-the-art model (Schnober et al., 2016) on the post-OCR correction task for Romanised Sanskrit.
• The model improves OCR output with an initial CRR of 35.76% to 87.01% on a low-quality scan setting (60 DPI), demonstrating strong robustness to OCR errors.
• In human evaluation, participants corrected 4.44 out of 5 sentences on average from the CopyNet predictions, compared to 3.56 for the PCRF model and 3.11 for raw OCR output.
• Correction time was significantly faster for CopyNet (81.4 seconds per string) than for PCRF (106.6 seconds) and OCR (127.6 seconds), indicating higher usability.
• The model achieved a CRR of 89.02% on a generalized setting using 10 random noise configurations, closely matching the 89.65% CRR reported on a curated test set.
• The model shows strong performance even on OCR outputs with very low CRR (as low as 36%), indicating resilience to severe OCR degradation.