Unveiling Patterns in Student Problem-Solving: Semantic Clustering and Topic Modeling with T5 and BART Embeddings

For students, it is very important to know our strengths and weaknesses in mathematics so that we can improve overall performance. Traditional grading systems often provide only a summary of marks without offering insights into where we are losing them. This study proposes a transformative approach to assessments by leveraging transformer-based models, T5((Text-to-Text Transfer Transformer) and BART(Bidirectional and Auto-Regressive Transformer), to analyze and cluster student responses based on mathematical subdomains like trigonometry, geometry, and algebra. It will identify patterns in our problem-solving techniques by extracting semantic embeddings from our answers and applying clustering algorithms such as K-Means and Mean Shift. This insight will enable us to know the specific areas where mistakes are happening, thus pinpointing the subdomains that need improvement. A personalized learning experience will be achieved this way. Furthermore, an auto-recommendation system is foreseen to recommend targeted resources and strategies to enhance our skills in those areas. This work envisions assessments as something serving growth and self-awareness, fostering not just better academic performance but a deeper grasp of mathematical concepts through innovative machine learning applications.