Automated classification of assessment questions according to Bloom’s taxonomy is increasingly used to support curriculum design and educational analytics. Many existing approaches rely heavily on instructional verbs as proxies for cognitive demand, despite longstanding concerns about their interpretive reliability. This paper adopts a data-centric perspective to examine why verb-centric Bloom-level classification remains fragile when applied to authentic multiple-choice question (MCQ) stems. The study is based on a custom, single-domain dataset of MCQ stems annotated according to the revised Bloom’s taxonomy, with intentional class imbalance preserved to reflect realistic assessment pra... More >