Henriques, Roberto André PereiraMaharramov, Arif2025-11-042025-11-042025-10-27http://hdl.handle.net/10362/190079Dissertation presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced Analytics, specialization in Data ScienceThis study investigates the automatic classification of exam questions according to Bloom’s Taxonomy, a hierarchical framework used to evaluate cognitive complexity in educational assessment. Addressing key challenges such as class imbalance and linguistic ambiguity, the research evaluates 26 model–feature combinations across classical machine learning, gradient boosting, and deep learning methods. A curated dataset of 1,800 questions—comprising both expert-labelled and AI-generated items—was used to ensure semantic diversity and balanced class representation across all six Bloom levels. The best-performing model, a Convolutional Neural Network (CNN) using fastText token-level embeddings, achieved a macro F1-score of 0.831 and a Cohen’s Kappa of 0.798, outperforming traditional models and other deep learning baselines. Analysis showed consistent high performance across all cognitive categories, eliminating the need for an ensemble. The results demonstrate that CNNs, when paired with subword-aware embeddings, offer an efficient and interpretable solution for cognitive-level classification, with potential for real-world integration into educational platforms and assessment design workflows.engBloom’s TaxonomyDeep LearningEducational NLPSynthetic DataCognitive Level ClassificationSDG 4 - Quality educationmaster thesis204073880