Learning by Analogy: Diverse Questions Generation in Math Word Problem
Project Overview
The document elaborates on the implementation of a Diverse Questions Generation Framework (DQGF) designed to enhance the capabilities of AI systems in solving Math Word Problems (MWPs) through the generation of diverse and coherent questions. This innovative framework comprises three main components: a Diverse Equations Generator that creates a variety of equations, an Equation-aware Question Generator that formulates relevant questions, and a Data Filter that ensures the quality of the generated content. By leveraging this structured approach, the DQGF seeks to improve the effectiveness of MWP solvers, facilitating learning through analogy and overcoming the limitations of existing methods that often depend on superficial heuristics. The findings suggest that such a framework not only enhances the diversity of questions posed to students but also contributes to deeper learning experiences, indicating significant potential for generative AI in educational settings. Overall, the DQGF serves as a pivotal development in the application of AI in education, particularly in fostering improved problem-solving skills in mathematics through more engaging and varied instructional materials.
Key Applications
Diverse Questions Generation Framework (DQGF)
Context: Educational resource for students solving Math Word Problems (MWPs)
Implementation: DQGF is implemented using three components: Diverse Equations Generator, Equation-aware Question Generator, and Data Filter to produce varied MWP datasets.
Outcomes: Improved accuracy and group-accuracy of MWP solvers, demonstrating enhanced learning by analogy.
Challenges: Generated data may still contain noise, affecting performance on original single-question MWPs.
Implementation Barriers
Data Quality
Generated data may not always meet the quality standards required for effective learning.
Proposed Solutions: Utilization of an expert model for filtering and enhancing the quality of generated MWPs.
Diversity of Questions
The diversity of questions generated is limited by the equations produced, which are based on heuristic rules.
Proposed Solutions: Future work will explore model-based equation generation for greater diversity.
Operator Limitations
Current models can only recognize basic arithmetic operators due to limitations in the training dataset.
Proposed Solutions: Expand the set of recognized operators in future iterations of the question generator.
Project Team
Zihao Zhou
Researcher
Maizhen Ning
Researcher
Qiufeng Wang
Researcher
Jie Yao
Researcher
Wei Wang
Researcher
Xiaowei Huang
Researcher
Kaizhu Huang
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Zihao Zhou, Maizhen Ning, Qiufeng Wang, Jie Yao, Wei Wang, Xiaowei Huang, Kaizhu Huang
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