Scholarlink

Multiple-choice question (MCQ) difficulty estimation is a critical task in educational technology, with applications in personalized learning and adaptive testing. We propose CASSA, a cognitively motivated, context-aware attention mechanism that enhances traditional self-attention to better capture task-relevant cues in MCQs for difficulty estimation. Integrated with pretrained models such as BERT, mBERT, and IndicBERT, CASSA demonstrates consistent gains across four benchmark datasets of MCQ difficulty Estimation: RACE++, Ext-MCQ, TEEMIL-H, and TEEMIL-K, improving macro F1-scores by an average of 3% over state-of-the-art. Our experiments show that CASSA remains effective even under frozen pretrained settings, and ablation studies confirm the independent contributions of its core components. Code is available in https://github.com/manikandan-ravikiran/CASSA/

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

CASSA: Context-Aware Self-attention with Global Context Suppression and Relevance Modulation for MCQ Difficulty Estimation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Manikandan Ravikiran,
  • Tarun Sharma,
  • Arnav Bhavsar,
  • Rohit Saluja

摘要

Multiple-choice question (MCQ) difficulty estimation is a critical task in educational technology, with applications in personalized learning and adaptive testing. We propose CASSA, a cognitively motivated, context-aware attention mechanism that enhances traditional self-attention to better capture task-relevant cues in MCQs for difficulty estimation. Integrated with pretrained models such as BERT, mBERT, and IndicBERT, CASSA demonstrates consistent gains across four benchmark datasets of MCQ difficulty Estimation: RACE++, Ext-MCQ, TEEMIL-H, and TEEMIL-K, improving macro F1-scores by an average of 3% over state-of-the-art. Our experiments show that CASSA remains effective even under frozen pretrained settings, and ablation studies confirm the independent contributions of its core components. Code is available in https://github.com/manikandan-ravikiran/CASSA/