Abstract
Neural Question Generation (NQG) has emerged as a foundation for educational technologies and automated assessments. In this study, we perform a survey of NQG over its emerging era from 2020 to 2026. To assess a pool of 109 highly relevant research articles, we systematically follow a five-phase identification and filtering approach. Using this analysis, the study proposes an exhaustive taxonomy that classifies the recent advancements in NQG based on question generation methodology, generation type, input modality, datasets, application domain, control objectives, and evaluation metrics. The findings indicate that there is a paradigm shift in NQG methodologies from rule-based and syntactic heuristics to highly advanced and popular Transformer-based methodologies, accounting for 58% of the methodologies. The results also indicate that although multiple-choice questions are the most prevalent question type in NQG methodologies, there is an increasing trend towards the generation of cognitively complex questions. Furthermore, we conclude that NQG approaches suffer from a lack of appropriate evaluation measures as well as an inability to generate questions at the higher cognitive order of Bloom’s Taxonomy.