Mathemyths: Leveraging Large Language Models to Teach Mathematical Language through Child-AI Co-Creative Storytelling

Title of the Paper

Mathemyths: Leveraging Large Language Models to Teach Mathematical Language through Child-AI Co-Creative Storytelling

Paper Information

• Subject Areas: Educational Technology, Artificial Intelligence, and Children's Learning
• Keywords: Storytelling, Mathematical Language, Conversational Interface, Large Language Models, Child-AI Collaboration, Co-Creation, Children

Research Background and Problem

• Problem or Challenge: A crucial component of early childhood mathematical development is mastering mathematical language, such as terms like "equal to" and "half." However, acquiring such language typically relies on daily conversations with parents and teachers, the quantity and quality of which vary significantly due to socioeconomic differences, potentially leading to gaps in early mathematical abilities. Integrating abstract mathematical language into education through storytelling is a common approach, but such methods often require one-on-one tutoring, which is challenging due to limited educational resources.
• Importance: Understanding mathematical language is a predictor of children's future academic achievement. Developing educational interventions to improve children's mathematical language comprehension is practically significant for bridging educational gaps.
• Motivation and Related Work: With advancements in large language models (LLMs), these models have demonstrated potential for spontaneous and creative conversations with children. However, systematic research on the feasibility and effectiveness of conversational AI in education remains underexplored.

Solution

• Proposed Method or Solution: This study developed a co-creative storytelling system, Mathemyths, leveraging large language models (such as GPT-4) to teach mathematical language to children aged 4-8 through collaborative storytelling.
• Innovative Features:
  • Utilized prompt engineering techniques to optimize AI-generated language, tailoring it to child users and aligning it with educational goals.
  • Incorporated support strategies (e.g., follow-up questions and rephrased prompts) to help children engage in storytelling, with specially designed contextual explanations of mathematical language.
• Implementation Steps and Techniques:
  • Developed a system to alternate storytelling with children, encouraging them to use key mathematical vocabulary.
  • The model employed a three-part conversational structure through question generation, story continuation, and interactive feedback mechanisms (including feedback and scaffolding strategies).
  • Conducted user research and experimental evaluations to validate the system's performance in educational contexts.

Research Outcomes

• Specific Results:
  • Children interacting with Mathemyths demonstrated mathematical language learning outcomes comparable to those guided by experienced human partners.
  • The system effectively supported children in completing relevant language tasks, including keyword recall, definitions, and understanding in transfer scenarios.
  • Subtle differences in learning and interaction patterns were observed across age groups: older children exhibited more complex responses during AI interactions, while younger children benefited more from the scaffolding functions provided by the AI.
• Advantages Compared to Existing Methods:
  • The system enabled more open and creative dialogues with children, surpassing the limitations of traditional AI structured interactions.
  • It introduced a novel approach to learning mathematical language through contextual storytelling, addressing constraints caused by limited parental or educational resources.
• Experimental or Evaluation Results:
  • User research was conducted with 35 children. Results showed an increase in mathematical language learning scores from a pre-experiment mean of 16.69 to a post-experiment mean of 18.97. While the outcomes of AI interaction were not significantly superior to human interaction, their comparability established a foundation for promoting this new learning model.
  • Younger children exhibited a slightly higher rate of "uncertain" responses to AI-generated questions, attributed to the more prominent imaginative elements in some questions, such as "cloud-accelerated travel."
• Limitations and Future Directions:
  • While the system demonstrated high mathematical language generation capabilities, there is room for improvement in creative extension and specific logical reasoning.
  • Long-term user studies have yet to be fully conducted. Future research could evaluate the effects of AI learning tools in long-term applications and explore multimodal interactions combining images and text.

Output Format

• The content is structured clearly.
• All experimental and research data mentioned in the paper are included.
• AI systems designed for children's learning, such as Mathemyths, provide valuable design frameworks and empirical evidence for reference.