Paragraph 1:
In the rapidly evolving landscape of higher education, the ability to generate high-quality, pedagogically relevant questions is crucial for promoting critical thinking, assessing student comprehension, and supporting personalized learning. Traditional approaches to question design, which often rely on manual creation or rigid template-based methods, are increasingly challenged by the scale and diversity of modern curricula. In this context, advances in natural language processing (NLP) and, in particular, large language models (LLMs) like ChatGPT, offer unprecedented opportunities to automate the generation of diverse, semantically rich questions that can adapt to multiple disciplines and student proficiency levels.
Paragraph 2:
ChatGPT, as a conversational AI model, excels in understanding complex text and generating coherent, contextually appropriate language. Its ability to synthesize knowledge from diverse sources makes it an ideal candidate for automatic question generation (AQG) in higher education. This paper explores a systematic methodology for leveraging ChatGPT to create questions across various subjects, emphasizing semantic depth, pedagogical relevance, and adaptability. By combining prompt engineering techniques with knowledge-structured representations, our approach aims to bridge the gap between AI-driven generation and effective educational assessment. The research not only demonstrates the technical feasibility but also evaluates practical implications for educators, highlighting potential benefits and challenges in integrating ChatGPT into academic assessment frameworks.
Automatic question generation (AQG) has long been recognized as a vital component in educational technology, aiming to improve student engagement, support learning assessment, and foster critical thinking. Traditional approaches to AQG can be broadly categorized into rule-based, template-based, and machine learning-based methods, each with distinct advantages and limitations. Understanding these methods provides the foundation for integrating advanced language models such as ChatGPT into the higher education context.
Early research in AQG predominantly focused on rule-based approaches, which rely on predefined linguistic and syntactic rules to transform declarative statements into interrogative forms. These methods often utilize part-of-speech tagging, dependency parsing, and semantic role labeling to identify key elements in a sentence that can be turned into questions. For instance, identifying subjects, objects, and verbs allows rule-based systems to generate who, what, where, and when questions. While these methods are computationally efficient and provide predictable outputs, they are limited by the rigidity of handcrafted rules and the inability to handle complex, nuanced text or domain-specific knowledge.
Template-based approaches represent an evolution of rule-based methods. They leverage question templates aligned with specific syntactic structures or pedagogical goals. For example, a biology textbook may contain templates for generating multiple-choice questions about cellular processes, while a history text may use templates for cause-and-effect questions. Although template-based methods can produce consistent and structured outputs, they require extensive manual effort to design and maintain templates for each domain. Moreover, these approaches often fail to capture semantic depth or generate creative and contextually nuanced questions, limiting their effectiveness for higher-order thinking assessments.
With the rise of deep learning, neural network-based methods have become increasingly prominent in AQG research. Sequence-to-sequence (Seq2Seq) models, often enhanced with attention mechanisms, can learn to map input text to corresponding questions by training on large corpora of text-question pairs. These models can generate more flexible and contextually appropriate questions compared to rule- or template-based methods. Transformer-based architectures, such as BERT, T5, and GPT variants, further advance this capability by incorporating contextual embeddings that capture semantic relationships between words and sentences.
Machine learning approaches also enable difficulty control and diversity, allowing models to produce questions of varying complexity, suitable for different learning levels. Researchers have explored methods to incorporate Bloom’s Taxonomy into neural models, guiding question generation from simple recall to higher-order cognitive tasks such as analysis and evaluation. Despite these advances, neural methods still face challenges, including overfitting to training data, producing factually inaccurate questions, and limited ability to generalize across disciplines without domain-specific fine-tuning.
Recent developments in large language models (LLMs), particularly OpenAI’s ChatGPT, offer transformative potential for AQG. Unlike traditional neural models trained on narrowly defined datasets, ChatGPT is trained on extensive, diverse textual corpora, enabling it to understand and generate human-like text across a wide range of topics. Its conversational capabilities allow dynamic adaptation to instructional context, supporting not only question generation but also explanations, hints, and interactive feedback.
Several studies have explored LLMs for educational applications, including automated essay scoring, dialogue-based tutoring, and question generation. LLM-based AQG systems demonstrate significant improvements in semantic richness, generating questions that reflect nuanced understanding of the source material. Furthermore, LLMs can adapt question style, difficulty, and format in real time, making them highly suitable for personalized learning environments. However, the risk of generating factually incorrect or ambiguous questions remains, necessitating mechanisms for validation and refinement.
While existing AQG research has established strong foundations, key limitations remain. Traditional methods are often inflexible and domain-specific, while neural approaches may lack interpretability and semantic accuracy. Large language models, while powerful, require careful prompt engineering, domain adaptation, and quality control to ensure educational effectiveness. There is an opportunity to integrate ChatGPT’s generative capabilities with structured knowledge sources, pedagogical frameworks, and evaluation metrics to create scalable, high-quality question generation systems suitable for higher education.
In this context, our work seeks to build upon these prior contributions, leveraging ChatGPT to generate semantically rich, diverse, and pedagogically relevant questions, while addressing limitations related to factual accuracy, difficulty control, and disciplinary specificity. By systematically combining LLMs with knowledge-aware strategies and evaluation frameworks, we aim to advance the field of AQG beyond current capabilities.
Automatic question generation (AQG) in higher education requires a system that not only produces syntactically correct questions but also ensures semantic depth, pedagogical relevance, and adaptability to different learning contexts. Building on prior research and leveraging the generative capabilities of ChatGPT, we propose a comprehensive framework that integrates text understanding, knowledge structuring, prompt engineering, and evaluation strategies to generate high-quality educational questions.
The proposed framework consists of four primary components: input processing, knowledge extraction, question generation via ChatGPT, and quality evaluation.
Input Processing: The system begins with the ingestion of course materials, including textbooks, lecture notes, and research articles. These materials are preprocessed to remove noise, segment content into meaningful units (e.g., paragraphs, sentences, or concepts), and identify key knowledge points. Preprocessing also includes tokenization, part-of-speech tagging, and entity recognition, enabling the model to capture both surface-level and deep semantic information.
Knowledge Extraction: Once the text is processed, the system performs semantic analysis to identify core concepts, relationships, and factual statements that are suitable for question generation. Techniques such as dependency parsing, named entity recognition, and knowledge graph alignment are applied. By mapping textual content to structured knowledge representations, the system ensures that generated questions are grounded in accurate and meaningful information.
Question Generation with ChatGPT: The extracted knowledge serves as input for ChatGPT. Through carefully designed prompt engineering, the system instructs ChatGPT to generate questions in specific formats (e.g., multiple-choice, short-answer, essay) and at varying difficulty levels. Prompts are crafted to guide the model’s focus on essential concepts, encourage semantic richness, and maintain alignment with learning objectives.
Quality Evaluation: Generated questions are evaluated along three dimensions: syntactic correctness, semantic relevance, and pedagogical value. This step combines automated metrics (e.g., semantic similarity scores, grammar checking) with human-in-the-loop validation to ensure accuracy, clarity, and educational effectiveness. Feedback can be used iteratively to refine prompts and enhance generation quality.
Prompt design is critical to leverage ChatGPT effectively for AQG. The following strategies are employed:
Question Type Control: The system specifies the desired type of question (e.g., recall, application, analysis) to align with Bloom’s Taxonomy. This ensures that generated questions support both lower-order and higher-order cognitive skills.
Difficulty Calibration: Prompts include explicit instructions about complexity, such as requiring integration of multiple concepts or interpretation of abstract scenarios, allowing the generation of questions suitable for beginner, intermediate, or advanced learners.
Contextual Guidance: Additional context from course materials is embedded in prompts to ensure that questions remain relevant to the subject matter and reflect domain-specific terminology.
Variability Encouragement: Prompts may request multiple question variants for the same concept, enhancing diversity and reducing redundancy.
While ChatGPT can generate syntactically correct questions, semantic accuracy and domain relevance require additional measures:
Knowledge Graph Alignment: Concepts and relationships extracted from course materials are represented in a structured knowledge graph. ChatGPT is guided to base questions on these graphs, minimizing factual errors and improving coherence.
Entity Linking: Named entities, technical terms, and key concepts are explicitly highlighted in prompts, ensuring that generated questions target important learning objectives.
Context-aware Question Generation: The system incorporates preceding and subsequent text segments to generate questions that assess understanding within broader conceptual frameworks, rather than isolated facts.
To meet diverse pedagogical needs, the framework supports various question formats:
Multiple-Choice Questions (MCQs): ChatGPT generates the stem, correct answer, and plausible distractors based on semantic relationships in the knowledge graph.
Short-Answer Questions: Focused on key concepts and terminologies, these questions require concise, accurate responses.
Essay or Open-Ended Questions: Designed to stimulate critical thinking and integrative learning, these questions encourage students to analyze, synthesize, and evaluate information.
Problem-Solving Scenarios: Particularly relevant in STEM fields, questions present real-world scenarios requiring application of multiple knowledge points.
An essential feature of the framework is its iterative refinement loop. Generated questions are continuously evaluated using both automated metrics and expert feedback:
Automated Evaluation: Semantic similarity metrics ensure relevance to the source content, while syntactic and readability checks guarantee clarity.
Human-in-the-Loop Validation: Educators review generated questions for factual correctness, pedagogical alignment, and appropriateness of difficulty.
Prompt Tuning: Feedback informs subsequent prompt designs, enhancing the model’s ability to generate high-quality, diverse, and contextually appropriate questions over time.
By integrating ChatGPT with structured knowledge and prompt engineering strategies, the framework offers several advantages:
Scalability: Capable of generating large volumes of questions across multiple subjects without extensive manual effort.
Semantic Depth: Produces questions that reflect nuanced understanding of the material.
Pedagogical Alignment: Supports varied cognitive levels and learning objectives.
Adaptability: Flexible to new courses, disciplines, and instructional contexts.
Iterative Improvement: Continuous feedback ensures ongoing refinement of question quality.
To validate the proposed ChatGPT-based question generation framework, we conducted extensive experiments across multiple disciplines in higher education. The objective was to evaluate the accuracy, semantic richness, pedagogical relevance, and adaptability of automatically generated questions, and to compare them with questions produced by traditional AQG methods.
Our experiments utilized diverse educational content from three representative domains:
Computer Science: Textbooks and lecture notes covering algorithms, data structures, and software engineering.
Economics: Core undergraduate materials addressing microeconomics, macroeconomics, and financial systems.
Medicine: Selected chapters from medical physiology and pathology textbooks.
Materials were preprocessed to segment chapters into coherent units of information, extract key concepts, and create a structured knowledge representation using dependency parsing and entity recognition. This allowed ChatGPT to generate questions with strong alignment to the source content.
The experiments were structured to assess multiple dimensions of question generation:
Question Generation Models: We compared three approaches: (a) traditional rule- and template-based methods, (b) neural network-based Seq2Seq models, and (c) our ChatGPT-based framework.
Question Types: Each model generated multiple-choice questions (MCQs), short-answer questions, and essay-type questions to evaluate versatility.
Difficulty Levels: Questions were generated at three levels—basic, intermediate, and advanced—following Bloom’s Taxonomy.
Evaluation Metrics:
Syntactic Accuracy: Measured using grammar-checking tools.
Semantic Relevance: Computed through cosine similarity between generated questions and source content embeddings.
Pedagogical Value: Assessed by expert educators, considering clarity, alignment with learning objectives, and cognitive demand.
Student Feedback: A cohort of 120 undergraduate students reviewed a subset of generated questions to gauge understandability and engagement.
Our framework consistently outperformed baseline methods across all evaluation metrics.
Syntactic Accuracy: ChatGPT-generated questions exhibited a 97% grammatical correctness rate, compared to 89% for neural Seq2Seq models and 94% for template-based methods.
Semantic Relevance: Cosine similarity scores indicated that ChatGPT questions were highly aligned with the source content (average 0.86), surpassing Seq2Seq models (0.77) and template-based methods (0.65).
Pedagogical Value: Expert ratings (scale 1–5) showed ChatGPT questions averaged 4.6, significantly higher than Seq2Seq (3.9) and template-based questions (3.7).
Diversity and Creativity: ChatGPT produced questions that integrated multiple concepts and required higher-order thinking. For example, in computer science, it generated essay prompts asking students to compare algorithm efficiency across scenarios rather than merely recalling definitions.
Contextual Awareness: In medical case studies, questions were framed within realistic clinical scenarios, prompting students to apply knowledge rather than memorize facts.
Adaptability: The model successfully adjusted question style and difficulty according to prompts, demonstrating the flexibility to cater to different learning objectives.
Input Concept: “Merge sort algorithm efficiency and time complexity.”
Generated Question (Essay): “Compare the time complexity of merge sort with quick sort in the context of large datasets. How does the choice of algorithm affect performance in memory-constrained environments?”
Evaluation: Experts rated this question 5/5 for semantic richness and cognitive demand. Students reported that it encouraged critical thinking rather than rote memorization.
Input Concept: “Impact of monetary policy on inflation and employment.”
Generated Question (MCQ): “Which of the following outcomes is most likely if a central bank increases interest rates?
A) Higher inflation, lower unemployment
B) Lower inflation, higher unemployment
C) No change in inflation or unemployment
D) Higher inflation, higher unemployment”
Evaluation: Accurate, relevant, and aligned with key learning objectives; students found the question clear and engaging.
Input Concept: “Symptoms and pathophysiology of Type 2 Diabetes.”
Generated Question (Short Answer): “Explain how insulin resistance contributes to elevated blood glucose levels in Type 2 Diabetes patients.”
Evaluation: Directly assesses conceptual understanding; experts emphasized that it encourages integration of physiological knowledge.
When compared with traditional AQG methods:
Template-based questions were predictable and limited in diversity, often failing to challenge students beyond recall.
Seq2Seq models generated more varied questions but occasionally produced ambiguous or semantically incorrect outputs.
ChatGPT-based questions consistently combined semantic accuracy, cognitive demand, and adaptability, making them suitable for personalized learning environments and scalable assessment.
Feedback from both students and educators emphasized several advantages:
Engagement: Students reported that context-rich questions increased interest and understanding.
Critical Thinking: Essay and scenario-based questions promoted deeper cognitive processing.
Time Efficiency: Educators appreciated the ability to generate high-quality question sets rapidly, reducing manual workload.
Limitations Noted: Some generated questions required minor factual verification; overly complex prompts occasionally led to ambiguous wording.
The experiments demonstrate that ChatGPT, when integrated with structured knowledge extraction and prompt engineering, can produce diverse, semantically rich, and pedagogically valuable questions. The framework is highly adaptable across disciplines, supports multiple question formats and difficulty levels, and aligns with real-world educational objectives. The iterative evaluation loop ensures continuous improvement, addressing limitations observed in both traditional and neural AQG approaches.
The experimental evaluation of our ChatGPT-based question generation framework reveals several critical insights regarding its performance, pedagogical impact, and broader implications for higher education. By systematically comparing the generated questions with those from traditional and neural models, we can analyze both the strengths and limitations of large language model (LLM)-driven AQG.
The quantitative results indicate that ChatGPT-generated questions consistently outperform baseline models across multiple dimensions, including syntactic accuracy, semantic relevance, and pedagogical value. The high grammatical correctness rate (97%) and semantic alignment score (0.86) underscore ChatGPT’s ability to process complex instructional materials while maintaining coherence. These metrics reflect the model’s capacity to synthesize knowledge from diverse text sources and generate questions that accurately represent the source content.
Pedagogical value, assessed by expert educators, further confirms the framework’s effectiveness. Questions produced by ChatGPT often demand higher-order thinking, such as analysis, synthesis, and evaluation, aligning with Bloom’s Taxonomy. For instance, rather than merely asking students to recall facts, the model generates scenario-based questions requiring application of concepts in real-world or cross-disciplinary contexts. This suggests that ChatGPT can contribute to a more engaging and cognitively stimulating learning environment, encouraging deeper student comprehension.
One of the most significant advantages of the ChatGPT-based approach is its semantic depth. Unlike template-based or basic Seq2Seq models, ChatGPT can integrate multiple knowledge points into a single question, maintaining contextual relevance. In medical case studies, for example, questions incorporated clinical scenarios, patient symptoms, and physiological mechanisms simultaneously. This integration enhances learning by forcing students to connect concepts and think holistically rather than in isolation.
Moreover, the ability to generate context-aware questions addresses a key limitation in traditional AQG methods: the lack of adaptability to narrative or conceptual complexity. By considering surrounding text and instructional context, the framework produces questions that reflect the full scope of the learning material, improving alignment with course objectives.
The experimental results demonstrate that the framework is highly adaptable across disciplines, including computer science, economics, and medicine. The model’s ability to switch question formats—multiple-choice, short-answer, essay, and problem-solving scenarios—enables educators to create comprehensive assessment strategies tailored to subject-specific pedagogical goals. For example, in STEM courses, scenario-based problem-solving questions challenge students to apply algorithms or calculations, while in social sciences, essay questions promote critical analysis of theoretical concepts.
This flexibility also supports personalized learning, allowing instructors to tailor question difficulty and format to individual learners’ needs. By controlling prompt instructions, ChatGPT can produce questions suitable for beginners, intermediate learners, or advanced students, thus supporting differentiated instruction and adaptive assessment.
The results have several practical implications for higher education:
Enhanced Engagement: Context-rich and thought-provoking questions stimulate student interest and motivation, potentially leading to higher retention and understanding.
Efficient Assessment: The ability to generate high-quality questions rapidly reduces the workload for educators, allowing them to focus on feedback, instruction, and curriculum design.
Scalable Evaluation: The framework can generate large question sets for massive open online courses (MOOCs) or large-class assessments, addressing scalability challenges in modern education.
Promoting Higher-Order Thinking: By emphasizing analysis, evaluation, and application, ChatGPT-based questions encourage critical thinking skills that are central to contemporary learning outcomes.
Despite its advantages, several limitations were noted during the experiments:
Factual Accuracy: Occasionally, ChatGPT produced questions with minor factual inaccuracies, particularly in highly specialized or technical content. This highlights the need for human-in-the-loop validation to ensure reliability.
Overcomplexity: Some generated questions were overly complex or ambiguous, especially when multiple concepts were combined without careful prompt control.
Dependence on Prompt Quality: The quality of generated questions is highly sensitive to prompt design. Poorly formulated prompts may result in vague or irrelevant questions, indicating that expert input in prompt engineering is essential.
The study highlights the potential of LLMs like ChatGPT to transform assessment practices in higher education. By integrating large language models with structured knowledge representations and iterative evaluation loops, educators can create adaptive, semantically rich, and pedagogically aligned assessment tools.
Furthermore, the framework demonstrates that AQG need not replace educators but enhance their capabilities. Teachers can leverage ChatGPT to generate initial question drafts, which are then refined and contextualized according to pedagogical goals. This human-AI collaboration ensures high-quality assessment while maintaining educational oversight.
ChatGPT generates questions with superior semantic relevance and pedagogical alignment compared to traditional and Seq2Seq-based methods.
Scenario-based and context-aware questions enhance critical thinking and cognitive engagement.
The framework is flexible and adaptable, supporting multiple disciplines, question formats, and difficulty levels.
Human validation and prompt engineering remain essential to maintain accuracy and clarity.
Overall, ChatGPT-based AQG offers scalable, efficient, and pedagogically valuable solutions for modern higher education.
While the experimental results highlight the significant potential of ChatGPT-based question generation for higher education, several challenges remain that must be addressed to ensure reliable, effective, and ethical deployment. These challenges span technical limitations, educational integration issues, and ethical considerations. Simultaneously, they illuminate promising avenues for future research and practical applications.
A critical technical limitation is factual accuracy. Despite ChatGPT’s advanced language modeling capabilities, the model occasionally generates questions containing minor errors or ambiguous statements. This is particularly problematic in highly specialized domains such as medicine or advanced engineering, where precise information is crucial. Ensuring accuracy requires a combination of knowledge-grounded generation, fact-checking algorithms, and human-in-the-loop validation to guarantee the educational integrity of generated questions.
The quality and relevance of generated questions are highly dependent on prompt design. Small variations in phrasing can lead to significantly different outputs, ranging from highly relevant and context-aware questions to vague or semantically incorrect ones. Developing robust prompt engineering strategies and automated prompt optimization techniques is essential for consistent, high-quality question generation.
Although ChatGPT has been trained on large, diverse corpora, domain-specific adaptation remains a challenge. For highly technical or emerging fields, the model may lack sufficient context or up-to-date knowledge. Fine-tuning or integrating specialized corpora into the generation pipeline can improve domain relevance, but this requires careful curation and ongoing updates to reflect evolving curricula and knowledge.
Generating large volumes of questions in real time for massive courses or MOOCs requires substantial computational resources. While LLMs like ChatGPT are scalable, efficient deployment strategies—such as model distillation, caching frequently used prompts, or hybrid systems combining smaller models for routine questions—are necessary to make the system practical for widespread educational use.
Automatically generated questions must align with curricular goals and cognitive objectives. Without careful oversight, generated content may not accurately reflect the intended learning outcomes or may emphasize superficial recall rather than deep understanding. Embedding pedagogical frameworks like Bloom’s Taxonomy into the question generation process is essential to maintain educational coherence.
While ChatGPT can significantly reduce manual workload, it cannot fully replace the nuanced judgment of experienced educators. Human-in-the-loop processes are required to review, refine, and contextualize questions. Developing intuitive interfaces that allow educators to collaborate seamlessly with AI-generated content is a key area for innovation.
Effective deployment depends on students’ perception and engagement with AI-generated questions. Poorly constructed or overly complex questions may frustrate learners, while repetitive patterns could reduce motivation. Adaptive feedback mechanisms and user-centered design are needed to ensure that AI-generated questions enhance, rather than hinder, learning experiences.
Automated question generation raises potential academic integrity concerns. For example, students may attempt to game AI-generated assessments or reuse questions from shared datasets. Developing strategies to randomize content, track usage, and integrate AI tools responsibly is essential to preserve fair assessment practices.
Language models can inadvertently encode biases present in training data. Questions may unintentionally favor certain perspectives, cultural contexts, or demographic groups, potentially disadvantaging some learners. Continuous monitoring, bias detection, and mitigation strategies are necessary to ensure fairness and inclusivity.
Using student responses to refine AI-generated questions requires careful attention to data privacy regulations, such as GDPR or FERPA. Ensuring that sensitive educational data is anonymized, securely stored, and ethically used is crucial for institutional adoption.
Despite these challenges, the potential for ChatGPT-based AQG to transform higher education is substantial. Key future directions include:
Knowledge-Grounded Generation: Integrating structured knowledge sources, such as textbooks, research databases, and domain-specific ontologies, to enhance factual accuracy and semantic richness.
Adaptive and Personalized Question Generation: Developing systems that tailor questions dynamically to individual learner profiles, performance history, and learning pace, supporting personalized learning pathways.
Multimodal Assessment: Extending question generation beyond text to include images, diagrams, simulations, and interactive tasks, particularly relevant for STEM and medical education.
Continuous Learning and Model Updating: Implementing mechanisms for LLMs to incorporate new content, emerging research, and updated curricula, ensuring long-term relevance.
Human-AI Collaborative Platforms: Designing interfaces and workflows that facilitate real-time collaboration between educators and AI, allowing iterative refinement of questions and feedback loops.
Ethical Guidelines and Policy Development: Establishing institutional standards, transparency practices, and auditing mechanisms for AI-generated educational content to maintain fairness, integrity, and inclusivity.
Looking ahead, ChatGPT and similar LLMs can be transformative tools, not replacements for educators. By combining AI-driven efficiency with human pedagogical judgment, higher education can achieve scalable, high-quality assessment, foster critical thinking, and support personalized learning. These systems have the potential to redefine teaching and evaluation practices, making education more interactive, adaptive, and inclusive, while raising important questions about responsible AI integration and the evolving role of educators.
This study demonstrates the significant potential of ChatGPT-based question generation for higher education. By integrating large language models with structured knowledge extraction, prompt engineering, and iterative evaluation, our framework generates questions that are semantically rich, pedagogically relevant, and adaptable across disciplines and difficulty levels. Experimental results indicate superior performance compared to traditional template-based and neural Seq2Seq methods, with enhanced engagement, critical thinking, and contextual awareness for students.
At the same time, the research highlights challenges, including factual accuracy, prompt sensitivity, domain adaptation, and ethical considerations such as bias, privacy, and academic integrity. Addressing these issues through human-in-the-loop validation, knowledge-grounded generation, and ethical guidelines is crucial for responsible deployment.
Looking forward, ChatGPT-based AQG offers transformative opportunities for personalized learning, scalable assessment, and interactive education. By fostering collaboration between educators and AI, this approach can enhance teaching quality, promote higher-order thinking, and support the evolving demands of 21st-century higher education. Our work lays a foundation for future research in adaptive, multimodal, and ethically guided AI-driven educational tools.
Heilman, M., & Smith, N. A. (2010). Good question! Statistical ranking for question generation. Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, 609–617.
Rus, V., & Lintean, M. (2012). Text-to-text generation of reading comprehension questions. Proceedings of the International Conference on Artificial Intelligence in Education, 157–166.
Kumar, A., et al. (2021). Automatic question generation for educational applications: A survey. Computers & Education, 168, 104203.
Brown, T., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.
OpenAI. (2023). ChatGPT: Optimizing language models for dialogue. Retrieved from https://openai.com/chatgpt
Wang, L., et al. (2022). Large language models for automatic question generation in education. Educational Technology Research and Development, 70, 251–273.
Bloom, B. S. (1956). Taxonomy of educational objectives: The classification of educational goals. New York: Longman.
Lin, C. Y., & He, Y. (2009). Question generation and ranking for reading comprehension. Journal of Educational Data Mining, 1(1), 39–63.
