📜 Archival Transcription with LLMs: Can LLMs facilitate the OCR/HTR tasks for historical records? 📚

This repository accompanies our study on using Large Language Models (LLMs) like GPT-4o and Claude Sonnet 3.5 to transcribe historical handwritten documents (French) in a tabular format. We compare their performance against traditional OCR/HTR systems: EasyOCR, Keras OCR, Pytesseract, and TrOCR.

📜 Paper

https://arxiv.org/abs/2501.11623

🔍 Key Highlights

• Models Tested: GPT-4o, Claude Sonnet 3.5, EasyOCR, Keras OCR, Pytesseract, TrOCR.
• Metrics: Character Error Rate (CER) and Bilingual Evaluation Understudy (BLEU).
• Findings:
  • GPT-4o is best for line-by-line transcription.
  • Claude Sonnet 3.5 excels at whole-scan transcription.
  • In both line-by-line and whole-scan experiments, the two-shot approach yielded the best output.
  • BLEU better captures transcription quality compared to CER.

🚀 Quick Start

1. Clone the repo:

   git clone https://github.com/jbaudru/llm-ocr-htr-historical-text-recognition
   cd llm-ocr-htr-historical-text-recognition
   

2. Install dependencies:

   cd main/
   pip install -r requirements.txt

3. Run

• For the whole-scan experiments: run the scripts.
• For the line-by-line experiments, consult the notebook.

📁 Data

The repository includes:

• Historical document scans (images) and transcriptions.
• Ground truth transcriptions for evaluation.
• Sample model predictions.

📜 License

This project is licensed under the MIT License. (https://opensource.org/license/mit)