Attention-Enhanced Transformer Architecture for Automated Question Generation Based on Student Knowledge Level Estimation

Personalized question generation is critical for adaptive learning, enabling tailored assessments that align with individual student mastery. Existing approaches often rely solely on textual context without incorporating learner knowledge, resulting in generic questions that lack appropriate difficulty levels. To address these limitations, a novel framework integrates the Self-Attentive Knowledge Tracing (SAKT) model to estimate student knowledge states dynamically with an attention-enhanced T5 transformer for question generation. This dual-attention mechanism simultaneously attends to passage context and student mastery, allowing generation of contextually relevant, difficulty-adaptive questions. Fine-tuning is performed on the SQuAD dataset using Python and PyTorch tools. Experimental results show improved performance over baseline models, with BLEU-4 at 36.15, ROUGE-L at 44.78, and METEOR at 30.31, alongside a 96.3 accuracy in knowledge estimation by SAKT. The approach effectively personalizes learning, enhancing engagement and educational outcomes through adaptive question complexity.