Scholarlink

This paper presents a novel multi-task learning (MTL) approach for medieval Hebrew manuscripts, simultaneously addressing script classification and manuscript dating. To facilitate this, we collected a VML-MSH - a comprehensive dataset of medieval manuscripts spanning a period from 850 to 1540 CE. Each manuscript is accompanied by annotations of script type, script mode, and manuscript production date. The MTL enhances generalization across diverse script styles and periods by leveraging shared feature representations, and employs advanced architectures, including Vision Transformer (ViT), MobileViT, Swin Transformer, FocalNet, BEiT, and ConvNeXT. Our training methodology combines classification and regression objectives with adaptive loss weighting, optimizing task performance. Experimental results demonstrate high accuracy in script type classification ranging from 84% to 100%, while highlighting challenges with a few script subtypes (60%–100%). For manuscript dating, the model achieves a mean error ranging from 1.37 to 8.97 decades. These findings underscore the effectiveness of the proposed MTL framework for advancing historical document image analysis and digital paleography. Code and dataset are available at https://github.com/atamnour/MTL-Hebrew-Paleography .

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

Multi-task Learning for Hebrew Paleography: Script Classification and Date Estimation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nour Atamni,
  • Boraq Madi,
  • Shoshana Bordman,
  • Daria Vasyutinsky Shapira,
  • Irina Rabaev,
  • Jihad El-Sana

摘要

This paper presents a novel multi-task learning (MTL) approach for medieval Hebrew manuscripts, simultaneously addressing script classification and manuscript dating. To facilitate this, we collected a VML-MSH - a comprehensive dataset of medieval manuscripts spanning a period from 850 to 1540 CE. Each manuscript is accompanied by annotations of script type, script mode, and manuscript production date. The MTL enhances generalization across diverse script styles and periods by leveraging shared feature representations, and employs advanced architectures, including Vision Transformer (ViT), MobileViT, Swin Transformer, FocalNet, BEiT, and ConvNeXT. Our training methodology combines classification and regression objectives with adaptive loss weighting, optimizing task performance. Experimental results demonstrate high accuracy in script type classification ranging from 84% to 100%, while highlighting challenges with a few script subtypes (60%–100%). For manuscript dating, the model achieves a mean error ranging from 1.37 to 8.97 decades. These findings underscore the effectiveness of the proposed MTL framework for advancing historical document image analysis and digital paleography. Code and dataset are available at https://github.com/atamnour/MTL-Hebrew-Paleography .