Modeling The Dynamic Note-Taking Behavior of MOOC Video Learners via Representation Learning

Keeping notes is an important learning method for MOOC learners that can bring fun and new ideas to themselves and others. We identified the opportunity lied in the mass data of dynamic MOOC video notes, which can help us better understand the behavior and evaluate effectiveness of online learning. In this paper, we propose the task of modeling MOOC learners' dynamic video note-taking behavior and solving some subsequent down-stream tasks, which to the best of our knowledge is the first in literature. To model the video-commenting behavior, most existing methods consider singular factors of either video frames or surrounding texts. Moreover, they only consider the information `visible' to a learner, while in the scenario of MOOC what the lecturer speaks is sometimes more informative than what the camara records. To overcome these shortcomings, an extended neural model to jointly encode note of not only visual/text content but also the speech content is proposed. Evaluation on a large real-world dataset shows that our model has better performance than previous methods.