Scholarlink

We present the preliminary results of the work in building the DigitalMaktaba-LaPira-v1 (DM-LP-v1) dataset, a large-scale, openly available dataset aiming at advancing cataloguing and OCR research for multilingual Arabic script digital libraries. Derived from over 73,000 Arabic-script PDF volumes held by the FSCIRE “La Pira" Library, specialized in history and doctrines of Islam, the dataset includes frontispieces, indexes, and ISBN-bearing pages extracted as high-resolution images and structured OCR outputs. A reproducible pipeline combines Qwen-2VL-72B, a vision-language model for zero-shot page classification, with Google Vision AI for text extraction. Evaluation on a 100 books sample yields F1 scores above 94% across all tasks, confirming the pipeline’s suitability for enriching bibliographic metadata. The dataset, comprising around 5 TB of images, structured text, and metadata, is released under a permissive license, along with scripts for PDF preprocessing and initial layout tagging; validation and quality control pipelines are in preparation. DM-LP-v1 aims to offer a scalable foundation for research in multilingual cataloguing, document layout analysis, and OCR fine-tuning, addressing a critical gap in multilingual Arabic script heritage in the context of digital libraries and cataloguing while supporting inclusive digital library development.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DM-LP: A Large Arabic Script Languages Dataset for OCR and Cataloguing Research

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Luca Sala,
  • Riccardo Amerigo Vigliermo,
  • Giovanni Sullutrone,
  • Sonia Bergamaschi

摘要

We present the preliminary results of the work in building the DigitalMaktaba-LaPira-v1 (DM-LP-v1) dataset, a large-scale, openly available dataset aiming at advancing cataloguing and OCR research for multilingual Arabic script digital libraries. Derived from over 73,000 Arabic-script PDF volumes held by the FSCIRE “La Pira" Library, specialized in history and doctrines of Islam, the dataset includes frontispieces, indexes, and ISBN-bearing pages extracted as high-resolution images and structured OCR outputs. A reproducible pipeline combines Qwen-2VL-72B, a vision-language model for zero-shot page classification, with Google Vision AI for text extraction. Evaluation on a 100 books sample yields F1 scores above 94% across all tasks, confirming the pipeline’s suitability for enriching bibliographic metadata. The dataset, comprising around 5 TB of images, structured text, and metadata, is released under a permissive license, along with scripts for PDF preprocessing and initial layout tagging; validation and quality control pipelines are in preparation. DM-LP-v1 aims to offer a scalable foundation for research in multilingual cataloguing, document layout analysis, and OCR fine-tuning, addressing a critical gap in multilingual Arabic script heritage in the context of digital libraries and cataloguing while supporting inclusive digital library development.