Scholarlink

Question-answering (QA) on handwritten documents is challenging but has valuable real-world applications. This paper presents a novel recognition-based QA approach that significantly improves accuracy over previous methods on handwritten datasets, including HW-SQuAD and BenthamQA. Our method integrates transformer-based retrieval and ensemble techniques, achieving Exact Match scores of 82.02% for HW-SQuAD and 69.1% for BenthamQA, with F1 Score improvements of 13.28% and 3.16%, respectively. It surpasses the previous best methods by 10.89% and 3.0%. Additionally, the document retrieval accuracy increased from 90.0% to 95.30% for HW-SQuAD and from 98.5% to 100% for BenthamQA. These results highlight the effectiveness of our approach in enhancing QA for handwritten documents. The data, model, and code are publicly available at  https://github.com/phc2017002/improved_hw_squad .

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

Advancing Question Answering on Handwritten Documents

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Aniket Pal,
  • Ajoy Mondal,
  • C. V. Jawahar

摘要

Question-answering (QA) on handwritten documents is challenging but has valuable real-world applications. This paper presents a novel recognition-based QA approach that significantly improves accuracy over previous methods on handwritten datasets, including HW-SQuAD and BenthamQA. Our method integrates transformer-based retrieval and ensemble techniques, achieving Exact Match scores of 82.02% for HW-SQuAD and 69.1% for BenthamQA, with F1 Score improvements of 13.28% and 3.16%, respectively. It surpasses the previous best methods by 10.89% and 3.0%. Additionally, the document retrieval accuracy increased from 90.0% to 95.30% for HW-SQuAD and from 98.5% to 100% for BenthamQA. These results highlight the effectiveness of our approach in enhancing QA for handwritten documents. The data, model, and code are publicly available at  https://github.com/phc2017002/improved_hw_squad .