Improving Controllability of Educational Question Generation by Keyword Provision
Project Overview
The document explores the role of generative AI, particularly in Question Generation (QG), within the educational sector, emphasizing advancements aimed at improving the quality and relevance of generated questions. It addresses the current limitations of QG models, which often struggle to produce complex questions suitable for higher-level assessments. To mitigate these challenges, the authors introduce the Keyword Provision Question Generation (KPQG) model, which empowers users to influence the question generation process through the provision of specific keywords. This new approach enhances the controllability of the output, allowing for the creation of tailored questions that align more closely with educational objectives. The findings suggest that the KPQG model significantly improves the overall effectiveness of question generation, making it a valuable tool for educators in developing assessments that meet diverse learning needs. Overall, the integration of generative AI in education, particularly through models like KPQG, demonstrates promising outcomes in facilitating personalized learning experiences and improving student engagement.
Key Applications
Keyword Provision Question Generation (KPQG)
Context: Educational reading practice and assessments for advanced learners.
Implementation: KPQG uses pre-trained language models to generate questions based on a context, answer, and user-provided keywords.
Outcomes: Improved question diversity and controllability, with user feedback indicating better alignment with expected results.
Challenges: Initial models lacked controllability and produced overly simple questions for advanced assessments.
Implementation Barriers
Technical Barrier
Current QG models suffer from controllability issues, making it difficult to generate expected questions.
Proposed Solutions: Implementing a keyword provision system to guide question generation.
Quality Barrier
Generated questions are often too simple for advanced educational assessments.
Proposed Solutions: Training QG models with exam-like datasets and using pre-trained language models for better performance.
Project Team
Ying-Hong Chan
Researcher
Ho-Lam Chung
Researcher
Yao-Chung Fan
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Ying-Hong Chan, Ho-Lam Chung, Yao-Chung Fan
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai