Automated Question Generation Using T5 Transformer in Conversational Educational Assistants

Conversational educational assistants increasingly generate comprehension and review questions from instructional text on demand, a capability whose value depends less on surface fluency than on whether the generated questions are factually grounded, pedagogically meaningful, and pitched at an appropriate cognitive level—criteria that standard text-generation metrics only partially capture. T5 (Text-to-Text Transfer Transformer), which reframes every NLP task, including question generation, as a unified text-to-text mapping, is fine-tuned in this study on 84,000 instructional passages spanning six subjects (Science: 21,000; Mathematics: 18,480; History: 14,700; Literature: 12,600; Language Arts: 10,500; Social Studies: 6,720), producing three questions per passage (252,000 total generated questions) at a target answer-type distribution spanning factual, inferential, and critical-thinking categories. Against a held-out test set, T5 achieves BLEU-4=0.324 and ROUGE$-\mathrm{L}=0.424$, exceeding a BERT2BERT benchmark (BLEU-4 0.284; +14.1%), a sequence-to-sequence LSTM with attention (0.248; +30.6%), and rule-based template generation (0.184; +76.1%); on human evaluation, 81.2% of T5-generated questions were rated pedagogically acceptable by trained reviewers, against 72.4% for BERT2BERT and 54.8% for rule-based generation. Generated questions skew toward lower Bloom's Taxonomy cognitive levels (61.6% Remember/Understand combined, 7.6% Evaluate/Create combined), a distributional finding with direct implications for how the system should be used alongside, rather than in place of, teacher-authored higher-order questions. Deployed within a conversational tutoring assistant across one semester, students using assistant-generated review questions showed a 41.3% increase in voluntary review-session engagement and an 48.4% increase in mean session length relative to a static-question-bank control group. A manual audit of 1,200 sampled question-answer pairs found that 84 (7.0%) contained an answer not actually supported by the source passage (a hallucination failure mode requiring human review before deployment to summative assessment contexts), mathematics-specific generation underperformed the cross-subject average by 23.5% on BLEU-4, and low-resource subjects with under 5,000 training passages showed a 33.3% BLEU-4 degradation relative to well-represented subjects-each bounding the system's current scope to formative, teacher-reviewed use rather than unsupervised question authoring.

Hali tarjima qilinmagan