Exploring Data Augmentation and Resampling Strategies for Transformer-Based Models to Address Class Imbalance in AI Scoring of Scientific Explanations in NGSS Classrooms

A team of researchers from the University of California, Berkeley, explored the effectiveness of data augmentation and resampling strategies in addressing class imbalance in AI scoring of scientific explanations in Next Generation Science Standards (NGSS) classrooms. The study focused on transformer-based models, which are commonly used in natural language processing tasks. The researchers found that data augmentation techniques, such as oversampling the minority class and generating new samples using a generative model, significantly improved the performance of the models in scoring advanced reasoning in scientific explanations. However, the study also highlighted the importance of careful selection of resampling strategies, as some methods can lead to overfitting and decreased performance. The findings have significant implications for the development of AI-powered educational tools that can provide immediate and accurate feedback to students. As the use of AI in education continues to grow, the ability to address class imbalance in AI scoring will be crucial in ensuring that AI-powered tools can provide fair and accurate assessments of student performance.