Filtering-based automatic cloze test generation

We propose a method to generate high-quality cloze test questions using a computational approach. Previous methods for automatic cloze test generation have contained some problems; specifically, there can be multiple correct answers. We found that approximately 50% of the generated answers have such errors with previous methods, which requires human post-editing was necessary in previous research. We propose an N-gram filtering method that can detect the answer to a given question. We compare the errors of the generated questions before and after applying the filtering methods. We found that our filtering method can select quality distractors by reducing errors in the generated questions. Moreover, when we generate cloze tests using semantic similarity, non-native speakers are very hard to answer the questions.