Cross-Temporal Sinhala OCR: Page-Level Adaptation and Diachronic Analysis

2026-06-28 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, medium

Summary

A new dataset, "sinhala-ocr-lk-acts-1010", has been introduced to address the lack of real-world page-level Sinhala OCR data. This dataset comprises 1,010 annotated page images and their transcriptions from Sri Lankan Legislative Acts published between 1981-1989 and 2000-2019, split into 707 training, 101 validation, and 202 testing examples. Researchers fine-tuned three deep learning visual language processing models, DeepSeek-OCR V1, DeepSeek-OCR V2, and LightOnOCR-2-1B, using QLoRA. LightOnOCR-2-1B emerged as the top performer, achieving a Character Error Rate (CER) of 1.05% on all test examples. This significantly surpasses other models, including Surya-OCR (8.84%), Tesseract v5 (10.69%), and Google Document AI (2.06%), demonstrating robust performance on real-world Sinhala OCR tasks, even with severely degraded documents across various print periods.

Key takeaway

For Machine Learning Engineers developing OCR solutions for low-resource languages, you should consider adopting the "sinhala-ocr-lk-acts-1010" dataset for training and evaluation. Utilizing LightOnOCR-2-1B, fine-tuned with QLoRA, offers superior accuracy (1.05% CER) compared to commercial and open-source alternatives, even on degraded historical documents. This approach provides a robust framework for deploying high-performance OCR in challenging real-world scenarios, ensuring better data digitization outcomes.

Key insights

The "sinhala-ocr-lk-acts-1010" dataset and fine-tuned LightOnOCR-2-1B significantly advance real-world Sinhala OCR, outperforming commercial and open-source models.

Principles

• Real-world datasets are crucial for OCR model evaluation.
• Fine-tuning large visual language models improves OCR accuracy.
• Performance consistency across degraded documents is achievable.

Method

The method involved creating "sinhala-ocr-lk-1010" from legislative acts, then fine-tuning DeepSeek-OCR V1/V2 and LightOnOCR-2-1B using QLoRA on consumer and cloud GPUs.

In practice

• Use LightOnOCR-2-1B for high-accuracy Sinhala OCR.
• QLoRA fine-tuning is effective for OCR models.
• Prioritize real-world data for robust OCR development.

Topics

• Sinhala OCR
• Low-Resource Languages
• Deep Learning OCR
• LightOnOCR-2-1B
• QLoRA Fine-tuning
• Document Digitization

Code references

• JHLee0513/KLOCR

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian — QLoRA fine-tuning significantly boosts cross-lingual relation extraction performance for low-resource languages, outperforming prompt-only LLMs.
• Koshur Diacritizer: A Byte-Level Sequence-to-Sequence Model for Kashmiri Diacritic Restoration — Koshur Diacritizer is a ByT5-small model that restores Kashmiri diacritics using a new 23.7k sentence pair dataset.
• PorTEXTO: A European Portuguese Benchmark for Visual Text Extraction — Specialized multilingual data significantly improves European Portuguese OCR performance over model size or resolution.
• From Flutter On-Device to Gemini Vision: Building a Multi-Language Document Extraction Engine — Hybrid Vision LLM and OCR excels at structured data extraction from diverse, unstructured documents by combining visual and numeric accuracy.
• Adapting TrOCR for Printed Tigrinya Text Recognition: Word-Aware Loss Weighting for Cross-Script Transfer Learning — Adapting TrOCR for Tigrinya text requires tokenizer extension and Word-Aware Loss Weighting to overcome script-specific challenges.
• Building Singlish2Sinhala: A Machine Learning Approach to Sinhala Transliteration — Informal transliteration systems like Singlish present unique NLP challenges due to spelling inconsistencies.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.