Abstract
Automatic Question Generation (AQG) is a rapidly growing area within artificial intelligence (AI) and natural language processing (NLP), focused on creating questions automatically from various sources like raw text, databases, and semantic representations. This review explores a wide range of AQG approaches, from traditional rule-based methods to advanced neural network models, including sequence- to-sequence, transformer-based, and graph-based architectures, as well as hybrid methods that combine linguistic rules with machine learning techniques. While rule-based systems offer clarity and control, they often struggle with complex language structures, whereas neural models, especially those using transformers like T5 and BART, have transformed AQG by enabling end-to-end learning and generating more contextually relevant questions. Hybrid models aim to balance the strengths of both approaches, enhancing flexibility and adaptability. The review also discusses evaluation methods, including automated metrics like BLEU, ROUGE, and METEOR, along with human assessments. Despite notable progress, challenges remain in achieving natural question fluency, semantic accuracy, and the generation of high- quality distractors for multiple-choice questions. Looking ahead, promising research directions include lifelong learning models, multimodal question generation that integrates text with images or code, and more robust evaluation frameworks. This review offers insights for researchers and practitioners, emphasizing AQG’s potential to improve educational tools, conversational agents, and information retrieval systems.