AI Scoring of Scientific Explanations: Data Augmentation for Class Imbalance

A study investigates data augmentation strategies to address class imbalance in transformer-based automated scoring of students' scientific explanations. Using a dataset of 1,466 high school responses to a physical science assessment aligned with an NGSS learning progression, the rubric includes 11 binary-coded analytic categories covering six complete explanation components and five common incomplete or inaccurate ideas. The baseline model is SciBERT, fine-tuned and tested with three augmentation methods: GPT-4-generated synthetic responses, EASE (a word-level extraction and filtering approach), and ALP (Augmentation using Lexicalized Probabilistic context-free grammar). The research aims to improve accuracy in scoring advanced reasoning, which is often underrepresented in rubric categories.