Augmenting LLM Generated Feedback with Data Mining

Augmenting LLM Generated Feedback with Data Mining

연구 분야: Databases

학회: International Conference on Artificial Intelligence in Education

Automated feedback systems are important in mathematics education for providing timely and scalable support to students. While pretrained Large Language Models (LLMs) such as GPT have shown promise in generating feedback, fine-tuning LLMs to improve their performance is costly and resource intensive. In this work, we explore cost-efficient alternatives, focusing on data mining to enhance LLMs for feedback generation for open-ended math answers. We evaluate the effectiveness and practicality of data mining for few-shot prompting in generating both descriptive feedback and numerical scores for middle school math open responses. Our results show that data mining significantly improves the result. Further, we explore why we believe the feedback is better by delving into teacher written reasonings for why they liked or disliked certain feedbacks.