Generating Educational Questions with Similar Difficulty Level

Question Framing is an essential task of preparing an assessment for astudent’s skill evaluation. In the COVID-19 phase, the education is shifted to an on-line phase, and the challenge that needs to be solved is the question framing tech- niques for providing different sets of the same question papers to a class of students. Our work provides insight into the paraphrasing of questions by fine-tuning a pretrained Text-to-Text Transfer Transformer (T5) to paraphrase questions according to the difficulty level of an exam. Later, the paraphrased dataset is back-translated to enhance the quality of the generated question. Experimental results on the Mohler dataset show that our system generates semantic equality of all the paraphrased questions and provides all the possible paraphrased versions of a given question. This approach is a cost-effective method for the preparation of the assessment sheet. The questions generated by a fine-tuned model are evaluated by their BLEU and METEOR scores. Our experiments demonstrate that questions generated with our model are of high quality, diverse.