A primary study on diversity of question generation

Abstract

Question Generation aims to generate fluent and standard questions provided the context and the questions. The previous methods of question generation mainly generated top1 to get the target question. But there are some of the generated questions are less related to the reference questions. So we try to use diversity to improve the performance of question generation. First, we compared BLEU scores and consistency in Human Evaluation between generated 10 questions and generated 1 question; Second, we use a better pre-trained model to adjust the hyper-parameters to generate better questions. The experimental results show that our method outperforms top1 significantly and further demonstrates the advantage of diversity.