LetGrade: An Automated Grading System for Programming Assignments

Abstract

Manually grading programming assignments is time consuming and tedious, especially if they are incorrect and incomplete. Most existing automated grading systems use testing or program analysis. These systems rely on a single reference solution and award no marks to submissions that differ from the reference solution. In this research, we introduce an automated grading model LetGrade. LetGrade is a supervised machine learning-based mechanism for automatically identifying the approach of solving and grading a student’s submission. The method looks for a score of the similarities between the submitted solution and multiple correct solutions available to determine the solution’s approach. The calculated similarity score is then entered into a pretrained supervised machine learning model that grades the submission. Our models were evaluated against datasets containing Python and C programming problems. The average variance in the grade predicted by the supervised machine learning model is consistently close to 0.5. This indicates that the models can accurately predict the grade within a $10\%$ margin of error.