Scholarlink

<p>This paper introduces an AI-based automation tool, the “Automated Answer Sheet Corrector”, designed to enhance the traditional manual grading of handwritten answer sheets. The system addresses the time-consuming and error-prone nature of manual evaluation by employing advanced technologies. It uniquely integrates a custom heuristic method for text line extraction, Tesseract OCR for converting handwritten text to digital format, Sentence-BERT (SBERT) for semantic similarity assessment of textual answers, and Contrastive Language-Image Pre-training (CLIP) for analyzing diagrammatic content. The proposed framework processes PDF answer sheets by first converting them to images. It then applies histogram projection analysis to differentiate and extract text and diagrams. SBERT evaluates the semantic closeness of student textual responses to an answer key, moving beyond simple keyword matching. Simultaneously, CLIP compares extracted diagrams with reference diagrams. The similarity scores from both SBERT and CLIP are then combined to generate a final score. Experimental results on a custom college dataset demonstrate that the proposed system achieves an average grading accuracy of 93.3% compared to evaluations by human educators, significantly enhancing both efficiency and consistency in the evaluation process. This multi-modal approach aims to provide accurate, efficient, and consistent grading for both textual and visual elements in answer sheets, offering a significant improvement over existing systems that often struggle with non-textual components or rely on less sophisticated semantic understanding. The system demonstrates improved OCR accuracy, particularly for handwritten text on ruled paper, and introduces a comprehensive evaluation by incorporating diagram analysis. The code and dataset supporting the findings of this study are openly available on GitHub: <a href="https://github.com/RatnakirtiKamble/AutoChecker">https://github.com/RatnakirtiKamble/AutoChecker</a>.</p>

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-Modal Automated Evaluation of Handwritten Answer Sheets: A Framework Integrating SBERT and CLIP Encoding

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Ratnakirti Kamble,
  • Hriddhisha Halder,
  • Niharika Mishra,
  • Vikas Chaurasia,
  • Sangita Chaudhari,
  • Gajanan Birajdar

摘要

This paper introduces an AI-based automation tool, the “Automated Answer Sheet Corrector”, designed to enhance the traditional manual grading of handwritten answer sheets. The system addresses the time-consuming and error-prone nature of manual evaluation by employing advanced technologies. It uniquely integrates a custom heuristic method for text line extraction, Tesseract OCR for converting handwritten text to digital format, Sentence-BERT (SBERT) for semantic similarity assessment of textual answers, and Contrastive Language-Image Pre-training (CLIP) for analyzing diagrammatic content. The proposed framework processes PDF answer sheets by first converting them to images. It then applies histogram projection analysis to differentiate and extract text and diagrams. SBERT evaluates the semantic closeness of student textual responses to an answer key, moving beyond simple keyword matching. Simultaneously, CLIP compares extracted diagrams with reference diagrams. The similarity scores from both SBERT and CLIP are then combined to generate a final score. Experimental results on a custom college dataset demonstrate that the proposed system achieves an average grading accuracy of 93.3% compared to evaluations by human educators, significantly enhancing both efficiency and consistency in the evaluation process. This multi-modal approach aims to provide accurate, efficient, and consistent grading for both textual and visual elements in answer sheets, offering a significant improvement over existing systems that often struggle with non-textual components or rely on less sophisticated semantic understanding. The system demonstrates improved OCR accuracy, particularly for handwritten text on ruled paper, and introduces a comprehensive evaluation by incorporating diagram analysis. The code and dataset supporting the findings of this study are openly available on GitHub: https://github.com/RatnakirtiKamble/AutoChecker.