Learner affect through the looking glass: Characterization and detection of confusion in online courses

Characterizing the nature of students’ affective and emotional states and detecting them is of fundamental importance in online course platforms. In this paper, we study this problem by using discussion forum posts derived from large open online courses. We find that posts identified as encoding confusion are actually manifestations of different learner affects pertaining to their informational needs–primarily seeking factual answers. We quantitatively demonstrate that the use of content-related linguistic features and community-related features derived from a post serve as reliable detectors of confusion while widely outperforming currently available algorithms of confusion detection. We also point out that several prediction tasks in this domain (e.g., confusion and urgency detection) can be correlated, and that a model trained for one task can effectively be used for making predictions on the other task without requiring labeled examples. Finally, we highlight a very significant problem of adapting the classifier to unseen courses.