Difficulty-Controllable Neural Multiple Question Generation for Reading Comprehension Using Item Response Theory

Abstract

In recent years, there has been growing interest in the automatic generation of reading comprehension questions with controllable difficulty levels in educational settings. We have developed a technology that generates reading comprehension questions of a difficulty level suitable for the learner's ability using Item Response Theory. However, this method targets only extractive question formats, where the answer exists within the reading text, and does not support the multiple-choice question format that is widely used in educational settings. Therefore, in this study, we develop an automatic generation method for multiple-choice questions with controllable difficulty levels. Furthermore, we evaluate the performance of difficulty controllability based on the correctness of responses and the characteristics of the options chosen, by answering the questions generated by the proposed method using a Question Answering system and analyzing the response data through Item Response Theory. The results confirm that the proposed method is capable of generating multiple-choice questions reflecting the difficulty information.