One of the student's abilities that must be measured in the learning process is cognitive ability. This measurement is essential to establish the level of students' understanding of the material that has been taught. Giving questions with the appropriate difficulty level is an attempt to measure these cognitive abilities. According to Bloom's taxonomy, there are six levels of cognitive abilities, remembering, understanding, applying, analyzing, evaluating, and creating. An automatic question classifier is helpful for teachers to develop a high-quality question and comply with different cognitive levels. This paper proposed automation question classification using the classical classification method (Support Vector Machine and Multinomial Naïve Bayes) and the pre-trained language model, IndoBERT. We conducted an experiment based on 670 mathematics and Indonesian language questions in primary and high school. The evaluation results show that the best classification accuracy was achieved by the IndoBERT method, which is 82% for classifying mathematics questions and 63% for classifying Indonesian Language questions. We analyzed that the small number of datasets and unbalanced at each taxonomy level affected the classification results.