Abstract
Assessment is a pivotal part of the educational process. Current trends in education are focused towards balanced and effective assessment of student knowledge. The six levels of Bloom’s Taxonomy Framework for knowledge level assessment are related to the Cognitive Process Domain. Multiple Choice Question based assessments are one of the most common assessment methods for academic assessment. Developing these assessments while accurately depicting levels of knowledge is a major focus area in the educational arena. Using Natural Language Processing based methods of text, sentiment & question classification, this task can be further advanced. The focus of this study is effective classification of Multiple Choice Question Stems into the six cognitive levels of Bloom’s Taxonomy using Deep Learning methods. To assess this, three datasets were utilized: A custom built single-domain Multiple Choice Question stem based dataset, second a public domain general question dataset and the third from a previous relevant study. Deep Learning Model ‘bert-base-uncased’ was chosen as the base model, and trained on the datasets for classification. Results indicated that action verbs, used as the main criterion for Bloom level classification, were insufficient for accurate classification. Through the comparison of different trained models, it was deduced that even highly context aware attention based DL models were dependent on more than only verb based keywords pattern recognition for classification. When the best performing trained model was tested on a specialized domain specific test dataset, the above hypothesis was confirmed. Therefore, it was concluded that additional parameters and large domain specific datasets were necessary for more accurate content-aware question classification.
Replies