Automated Feedback Generation for Programming Assignments Through Diversification

Immediate and personalized feedback on students' programming assignments is important for improving their programming skills. However, it is challenging for instructors to give personalized feedback to every student since each program is written differently. To address this problem, the Automated Feedback Generation (AFG) technique has been proposed, which identifies faults from the wrong program, generates patches, and provides feedback if validation is passed. AFG relies on students' correct programs to find faults and generate patches. Therefore, having diverse correct programs is important for the performance of AFG. However, in small-scale programming courses or new online judge problems, there might be a lack of diversity in correct programs. In this paper, we propose Mentored, a new AFG for students' programming assignments through diversification. Mentored generates new structures of programs through various combinations of programs to gener-ate modifications optimized for wrong programs while solving the problem of dependency on correct programs. Additionally, Mentored provides transparent feedback on the process of repairing wrong programs. We evaluate Mentored on real student programming assignments and compare it with state-of-the-art AFG approaches. Our dataset includes real university in-troductory programming assignments and online judge problems. Experimental results show that Mentored generates higher repair rates and more diverse program structures than other AFG approaches. Moreover, by providing a transparent sequence of repair processes, Mentored is expected to improve students' programming skills and reduce instructors' manual effort in feedback generation. These results indicate that Mentored can be a useful tool in proaramming education.