Automating Characterization of Peer Review Comments in Chemistry Courses

Abstract

While writing-to-learn (WTL) pedagogies are a promising way for students to construct knowledge, one limiting factor to implementation is time the instructor spends grading. We conducted two WTL assignments in two undergraduate general chemistry courses combined with collaborative peer review. We used a previously developed scheme to classify peer review comments generated by 1,732 students enrolled in two undergraduate chemistry courses as praise, problem/solution, and verification/summary. Problem/solution comments were further separated into greater-level, mid-level, and word-sentence descriptors. Using the SciBERT language model we then developed a classifier which accurately identifies comments where human coding was considered the ground truth. In the future, this model can provide an efficient way for instructors to monitor peer review collaborations and help instructors use peer-generated insights to guide their instruction.