Automatic short answer grading with SBERT on out-of-sample questions
Aubrey Condor, Max Litster, Zachary A. Pardos
Abstract
We explore how different components of an Automatic Short Answer Grading (ASAG) model affect the model’s ability to generalize to questions outside of those used for training. For supervised automatic grading models, human ratings are primarily used as ground truth labels. Producing such ratings can be resource heavy, as subject matter experts spend vast amounts of time carefully rating a sample of responses. Further, it is often the case that multiple raters must come to a census before a final ground-truth rating is established. If ASAG models were developed that could generalize to out-of-sample questions, educators may be able to quickly add new questions to an auto-graded assessment without a continued manual rating process. For this project we explore various methods for producing vector representations of student responses including state-of-the-art representation methods such as Sentence-BERT as well as more traditional approaches including Word2Vec and Bag-of-words. We experiment with including previously untapped question-related information within the model input, such as the question text, question context text, scoring rubric information and a question-bundle identifier. The out-of-sample generalizability of the model is examined with both a leave-one-question-out and leave-one-bundle-out evaluation method and compared against a typical student-level cross validation.