Fixing Poorly Written Questions and Classifying Their Difficulty with DistilBERT, ALBERT, CNN, and Explainable AI

Classifying question difficulty levels is critical in educational data mining, impacting adaptive learning systems and assessments. However, questions are sometimes poorly written, which can lead to misunderstandings. This study focuses on ensuring questions are clear and comprehensible before classification. By utilizing DistilBERT (Distilled Bidirectional Encoder Representations from Transformers) to clarify ambiguously written questions and ALBERT (A Lite BERT) integrated along with a convolutional neural network (CNN) to classify them, the proposed method enhances both the clarity and accuracy of the classification process. The methodology uses Random Over Sampling (ROS) for dataset balancing and Explainable AI (XAI) techniques, particularly SHAP; to understand model decisions. The models are evaluated on a labelled dataset of questions, with ALBERT achieves a validation accuracy of 94.76% and DistilBERT reaching 91.83%, demonstrating improved handling of class imbalance. Comparisons between ALBERT and RoBERTa (Robustly Optimized BERT Approach) indicate that ALBERT achieves higher accuracy, higher AUC, and lower loss, making it the preferred choice for classification. DistilBERT is used for efficient and effective text correction. This research underscores the importance of question clarification, robust evaluation strategies, and model interpretability in natural language processing tasks, offering valuable insights for the development of intelligent educational systems.