Advancing Learning Systems: Enhancing Explanation Generation Through AI Innovations

AI-enhanced adaptive learning systems, such as Intelligent Tutoring Systems (ITS), improve teaching methods and enhance learning experiences across various domains, especially in technical subjects like mathematics and sciences. They provide personalized and adaptable environments, fostering engagement and academic achievement through tailored educational approaches. Despite the significant advancements in mathematical language processing driven by the advent of Large Language Models (LLMs), fine-tuning these expansive models for specialized mathematical tasks poses notable challenges. This is often a consequence of the scarcity of labeled datasets in this domain. To bridge this gap, this study curates a unique dataset, comprising around 4,000 formulas from mathematics, physics, and engineering, which features computationally rearranged equations paired with expert-reviewed, student-friendly explanations. Developed using the SymPy Python library, it serves as a comprehensive resource for evaluating language models in mathematical education. Furthermore, utilizing both the mathematical LaTeX and the string format of equations, this research fine-tunes several LLMs to highlight the differences between the models’ explanation generation capabilities and performance in generating accurate and informative mathematical explanations. Additionally, the comparison between outputs generated using LaTeX and string representations offers insights into the effectiveness of each format in conveying mathematical semantics. This research study contributes to the field of automated mathematics education, providing a deeper understanding of the potential of language models in creating explanations for intelligent tutoring systems. Our key achievements include the tbs17/MathBERT-custom model's enhanced Rouge-1 score, which marginally increased from 0.9678 in LaTeX format to 0.9679 in string format. More notably, the AnReu_math_pretrained_bert model's Rouge-1 score significantly improved from 0.9116 when using LaTeX to 0.9479 with string representations. This underscores the superior handling of string formats by these models, leading to improved explanation generation in the field of mathematics.