Abstract
Classifying exam questions using Bloom's Taxonomy is a critical task in education. However, performing this classification manually is often inefficient, prone to subjective errors and demands significant time from educators. To address this challenge, this study proposes and evaluates a robust system using state-of-the-art Deep Learning (DL) techniques. This system uses the pre-trained language model RoBERTa to understand the meaning of text and turn it into useful features. These features are then used by three different types of models to classify the text: a Gated Recurrent Unit (GRU), a Long Short-Term Memory (LSTM) and a Convolutional Neural Network (CNN). To further enhance performance, a Weighted Ensemble model is implemented, which intelligently combines the predictions of the three models. The models were trained and evaluated using a publically available dataset. Experimental results show that all models demonstrate an outstanding efficacy and the weighted ensemble model outperforms all individual models. The ensemble model established a new standard of performance at 92.37% accuracy, macro F1-Score of 0.923 and an Area Under the Curve (AUC) of 0.992. Crucially, the marginal gain over the best single model was confirmed to be statistically significant via McNemar's Test (P = 0.048). This paper presents a scalable system that will enable the automatic assessment of the objects in real-time with high accuracy that could ensure the objects in the assessment are not biased, would significantly reduce efforts put in by teachers and enhance the overall quality of educational practices.