ChatGPT + Knowledge Graphs = Smart Math Tutor that knows exactly where you're stuck
This paper introduces a novel approach combining knowledge graphs with LLMs to provide personalized educational feedback in e-learning environments. The system evaluates student interactions and prerequisites to deliver tailored guidance through ChatGPT, addressing individual learning challenges while maintaining accuracy and relevance.
-----
https://arxiv.org/abs/2412.03856
🎯 Original Problem:
Traditional e-learning systems and Intelligent Tutoring Systems (ITSs) struggle to provide truly personalized feedback, often relying on predefined pathways that don't address specific student impasses[2].
-----
🔧 Solution in this Paper:
→ The system integrates dynamic knowledge graphs with ChatGPT to map student prerequisites and knowledge states.
→ It categorizes students into three profiles: S1 (lacking foundational knowledge), S2 (middling understanding), and S3 (advanced with occasional challenges).
→ Knowledge graphs determine topic hierarchies and relationships, using "GO for Help" indicators from textbooks to establish prerequisites.
→ The system generates personalized feedback using a specialized prompt template that includes the question, standard solution, and student's impasse.
-----
💡 Key Insights:
→ LLM feedback accuracy varies with question difficulty - more consistent for basic questions, more diverse for advanced topics
→ Human oversight remains crucial to prevent potential misinformation
→ Knowledge graphs effectively map prerequisite relationships and student knowledge states
-----
📊 Results:
→ High correctness scores (mean 4.67-5.00) across all difficulty levels
→ Low hallucination rates (mean 4.67-5.00)
→ Moderate to high precision in feedback (mean 3.67-4.67)
→ Cohen's Kappa values: 0.47 (easy), 0.42 (moderate), 0.30 (hard) questions
Share this post