Affiliations: [a] State Key Laboratory of Software Development Environment, Beihang University, Beijing, China
| [b] School of Computer Science and Engineering, Beihang University, Beijing, China
| [c]
Kuaishou Technology, Beijing, China
Abstract: The traditional end-to-end Neural Question Generation (NQG) models tend to generate generic and bland questions, as there are two obscure points: 1) the modifications of the answer in the context can be used as the clues to the answer mentioned in the question, while they are generally not unique and can be used independently for generating diverse questions; 2) the same question content can also be asked in diverse ways, which depends on personal preference in practice. The above-mentioned two points are indeed two variables to conduct question generation, but they are not annotated in the original dataset and are thus ignored by the traditional end-to-end models. In this paper we propose a framework that clarifies those two points through two sub-modules to better conduct question generation. We take experiments based on the GPT-2 model and the SQuAD dataset, and prove that our framework can improve the performance measured by similarity metrics, while it also provides appropriate alternatives for controllable diversity enhancement.
