Automated Bloom's Taxonomy Classification of Teacher Questions Using Whisper and GPT-4

Abstract

Effective teacher questioning is central to fostering higher-order thinking, yet manually analyzing classroom discourse for cognitive demand is labor-intensive. This study proposes an automated system that classifies teacher questions into Bloom’s revised taxonomy levels by integrating Whisper speech-to-text transcription with GPT-4’s zero-shot classification capabilities. The end-to-end pipeline transcribes classroom audio, extracts teacher questions with surrounding context, and assigns one of six cognitive categories: Remember, Understand, Apply, Analyze, Evaluate, and Create. Using 350 questions annotated by three Bloom-trained experts (Fleiss’ Kappa = 0.85), the GPT-4–based classifier achieved 64.2% accuracy and a Cohen’s Kappa of 0.547 against human labels, outperforming a rule-based keyword baseline. Performance was strongest for higher-order categories such as Analyze and Create (precision ≥ 100%), while most misclassifications occurred between adjacent levels (e.g., Remember–Understand). Bloom-level distributions indicated a predominance of lower-order questions, highlighting the system’s potential for providing data-driven feedback to promote balanced cognitive engagement. This work demonstrates the feasibility of scalable, context-sensitive analysis of live classroom questioning, offering practical applications in teacher self-reflection, professional development, and AI-enhanced learning analytics.