Abstract:

With the recent increase in blended assessments, a debate has arisen regarding the appropriateness and utility of automated grading tools. The use of such tools provides learners with immediate feedback, while reducing instructor workload during grading. This work considers automatic short answer grading (ASAG), which involves the grading of short length, natural language student responses using machine learning and deep learning models. Various studies have shown that ASAG models exhibit limited performance due to the lack of training data. The aim of this study is to re-evaluate the robustness and generalizability of such systems, and to investigate how data/response diversity is addressed through the use of data augmentation techniques. Using a novel transformer based approach coupled with an enhanced paradigm for the augmentation of data, the proposed ASAG model demonstrates significant performance improvement using a representative open source dataset. The results suggest that it is feasible to use NLP support tools to reduce instructor workload.

Keywords:

Automated Assessment, Blended Learning, Automatic Short Answer Grading, Data Augmentation, Natural Language Processing.