Automatic Grading of Student Code with Similarity Measurement

Nowadays, online judges are extensively used for automatically grading student code. However, they grade code by only counting the number of passed test cases, which is not fair for assessing the overall quality of a code snippet. On the other hand, existing studies have used machine learning techniques for code grading. However, they usually require large amounts of labeled code to enable supervised learning and heavily rely on feature engineering. In this work, we design SimGrader, a code grading system that grades student code based on the measurement of similarity to the "good" code, and thus save the effort for code labeling. We extract three types of features to capture the overall quality of a code snippet, and design specific methods to enhance the feature discrimination, which facilitates the similarity measurement. We conduct extensive experiments to show the superiority of SimGrader over existing methods and justify the effect of the its system components. We deploy SimGrader to grade the student code submitted in an introductory programming course.