CoRSAL-OCR: Evaluating Zero-Shot OCR for Language Archive Materials

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, quick

Summary

CoRSAL-OCR is a new OCR evaluation dataset comprising over 200 document pages with gold-standard transcriptions from Bodo (Devanagari script) and Garo (Latin script) languages, designed to assess OCR quality for language archive materials. This dataset, alongside the 8-language AILLA-OCR benchmark, was used to evaluate four OCR systems: Tesseract, Google Cloud Vision, Gemini 3 Flash, and the open-weight Qwen3.5-27B. The evaluation revealed that vision language models (VLMs) achieve the lowest error rates on these datasets when provided with appropriate prompts. A critical finding was the significant impact of prompt design on VLM performance, with a detailed generic prompt reducing the character error rate (CER) by up to six-fold compared to a minimal prompt. The CoRSAL-OCR dataset is publicly released to foster further research in this area.

Key takeaway

For NLP Engineers or archivists digitizing diverse language materials, you should prioritize evaluating vision language models (VLMs) for OCR tasks. Your choice of prompt design is paramount; a detailed generic prompt can drastically improve VLM accuracy, potentially reducing character error rates by up to six-fold. Consider leveraging the CoRSAL-OCR dataset to benchmark VLM performance on specific low-resource or multi-script archival content, ensuring optimal accessibility for undigitized collections.

Key insights

VLMs excel at zero-shot OCR for language archives, but prompt engineering is crucial for optimal performance.

Principles

• VLMs can achieve low OCR error rates.
• Prompt design critically impacts VLM OCR.
• Zero-shot OCR is viable for diverse scripts.

Method

Evaluate OCR systems using a multi-language, multi-script dataset with gold-standard transcriptions, comparing traditional OCR with VLM performance under varied prompt conditions.

In practice

• Use VLMs for digitizing language archives.
• Experiment with detailed VLM prompts.
• Leverage CoRSAL-OCR for VLM benchmarking.

Topics

• OCR
• Vision Language Models
• Language Archives
• Prompt Engineering
• Zero-Shot OCR
• CoRSAL-OCR Dataset

Code references

• larc-iu/corsal-ocr

Best for: Research Scientist, AI Scientist, NLP Engineer, Domain Expert

Related on AIssential

• Multilingual OCR-Aware Fine-Tuning and Prompt-Guided Chain-of-Thought Reasoning for Multimodal Large Language Models — OCR-aware data curation and fine-tuning significantly improve MLLM robustness for multilingual text in degraded images.
• The easiest way to improve prompts — Zero-shot and few-shot prompting are core techniques for controlling language model output based on guidance provided.
• Vision LLM Output Control for Better OCR with Prompt Hints — Prompt hints in Sparrow enable direct VLM output control for improved OCR data extraction and formatting.
• Wall-OSS-0.5: 4B VLA with open training code and zero-shot real-robot evaluation[D] — Wall-OSS-0.5 demonstrates strong zero-shot real-robot performance and significant fine-tuned gains using a novel VLA architecture.
• Vision-Language Models as Zero-Annotation Oracles in Histopathology — General-purpose Vision-Language Models (VLMs) offer robust, zero-annotation foreground segmentation in histopathology, generalizing across diverse stains.
• GLM-OCR (9B) - Local OCR Test | OCR, Document Extraction, Table Recognition — GM OCR is a small, MIT-licensed, two-stage open-source model excelling in complex document and custom data extraction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.