A Review of Automatic Question Generation based on Deep Neural Networks

Abstract

Item writing is one of the expensive and time-consuming processes in test development. Recently, automatic question generation using artificial intelligence has gained attention as a way to reduce the burden of item writing. Traditional automatic question generation methods have heavily relied on manually designed templates and rules, while question generation methods based on neural networks, including large language models, have recently become popular and have demonstrated the capability to generate high-quality items. Furthermore, advances in multimodal data processing technologies enable the generation of test items involving images, knowledge graphs, equations, programming code, and more. Additionally, controllability in automatic question generation over various aspects, such as difficulty and item types, has become another focus of recent research. Consequently, this paper provides an overview of such latest neural automatic question generation methods and discusses their limitations and future challenges.