Leveraging Generative AI for Enhanced Question Generation in Reading Comprehension

Automatic Question Generation (QG) is a cornerstone of digital learning systems, enabling scalable creation of engaging and pedagogically relevant questions. However, traditional approaches often suffer from issues such as contextual irrelevance, repetition, and lack of diversity in question formats. To overcome these limitations, we propose a novel QG framework built on the LLaMA 3.1 language model, designed to generate three key types of questions: plain text, multiple-choice (MCQs), and fill-in-the-blank, each aligned with specific cognitive learning goals. Our methodology integrates optimized token handling and long-context management with advanced keyword extraction techniques, such as Sense2Vec and KeyBERT, and Maximal Marginal Relevance (MMR) to ensure semantic diversity and relevance, particularly in distractor generation for MCQs. Our system dynamically adapts question types to the input content, promoting active recall, comprehension, and analytical thinking. We evaluate our approach using automated metrics on the SQuAD 1.1 dataset. The model achieves an overall F1 score of 0.8, reflecting high alignment with reference questions in both lexical and semantic dimensions. These results demonstrate that our LLaMA 3.1-based framework generates high-quality educational questions that can be deployed in digital and blended learning environments.