A Machine Learning Approach for Suggesting Feedback in Textual Exercises in Large Courses

Abstract

Open-ended textual exercises facilitate the comprehension of problem-solving skills. Students can learn from their mistakes when teachers provide individual feedback. However, courses with hundreds of students cause a heavy workload for teachers: providing individual feedback is mostly a manual, repetitive, and time-consuming activity. This paper presents CoFee, a machine learning approach designed to suggest computer-aided feedback in open-ended textual exercises. The approach uses topic modeling to split student answers into text segments and language embeddings to transform these segments. It then applies clustering to group the text segments by similarity so that the same feedback can be applied to all segments within the same cluster. We implemented this approach in a reference implementation called Athene and integrated it into Artemis. We used Athene to review 17 textual exercises in two large courses at the Technical University of Munich with 2,300 registered students and 53 teachers. On average, Athene suggested feedback for 26% of the submissions. Accordingly, 85% of these suggestions were accepted by the teachers, 5% were extended with a comment and then accepted, and 10% were changed.