Optical Character Recognition for Low-Resource Languages: Benchmarking Data Generation Strategies and Their Impact on Accuracy

Abstract

Even though there have been groundbreaking achievements in Vision-Language Models leading Optical Character Recognition to human-like performance on major languages, billions of documents written in under-resourced scripts are still invisible to computers yet. This paper aims to resolve this issue by focusing on synthetic data generation and creating a data synthesis pipeline that combines both text-to-image and image-to-image translation algorithms to generate big labeled datasets. Targeting both Kazakh and Russian languages—where Kazakh remains significantly underrepresented in current OCR systems and Russian provides a complementary benchmark due to its substantial overlap in Cyrillic script—we develop a data synthesis pipeline that integrates text-to-image rendering and image-to-image translation techniques to generate labeled training data. For evaluation, we construct benchmark test sets in both languages to assess model performance across shared and language-specific Cyrillic characteristics. Carefully selected real-world documents, such as handwritten notes, transcripts, and domain-specific contracts, serve as further foundation for these models. Fine-tuning the state-of-the-art Dots.OCR model, a 1.7-billion-parameter unified Vision-Language Model, via Low-Rank Adaptation in nine experiments, outperforming both Google Vision API and the baseline by more than 25 percentage points on recognizing words in Kazakh; more importantly, however, we show that while synthetic data amount is important, it is not enough to ensure real-life generalization without grounding – a mere 37,621 real-life images are sufficient to prevent catastrophic forgetting.