Transfer learning for Bloom’s taxonomy-based question classification

Bloom’s taxonomy (BT) is commonly employed to categorize assessment questions posed to students based on different degrees of complexity. The existing datasets based on BT are small in size. Due to the substantial data requirements of deep learning models, the existing datasets are inadequate for adequately training these models and attaining favorable outcomes. This paper presents a new dataset referred to as “NCERT question classification” which is significantly larger than the existing BT-based question classification datasets. Experimental results demonstrate that an ensemble of pre-trained language models achieves state-of-the-art accuracy of 92.77%, 88.50%, and 78.44% on the existing Yahya et al., Jain et al., and CLO datasets, respectively, and an accuracy of 94.10% on our proposed dataset. Furthermore, we show that the results on existing datasets can be improved by up to 3% through fine-tuning a combined dataset that includes training instances from both the proposed and existing datasets.