(A) hierarchical aspect-sentiment model for predicting student’s performance on online programming education

Abstract

The number of students taking computer science (CS) courses is rapidly increasing as the demand for software engineers in the industry has exploded in recent years. However, for many students without much experience in programming, these exercises pose difficult challenges that may even prohibit the students from further studying CS even if they are intrigued by the theory and applications of CS. Such challenges can be overcome with dedicated teaching and mentoring by lecturers and, most importantly, TAs who give hands-up programming help. In many of those cases, the teaching staff would benefit from understanding the learning progress of each student, such that lecturers can design better lectures and exercises, and TAs can give individualized help to students who are having difficulties with the programming exercises. In this thesis, I first present a computer science education platform designed to lecture and collect the interaction data at the same time. From the collected data, I have identified the effect of interactive assistance and the six steps of the learning process in CS education. For each programming exercise, I map the students' actions to each of the steps in the CS-specific educational taxonomy, such that they can be analyzed to infer the students' learning progress in more detail than just completion of the exercise. Additionally, the analysis of Help Center posts, which is a set of conversational threads of asking a programming-related question, reveals that the question quality and topics can be used as a strong feature to predict the future academic performance of a student. To analyze the help center conversational data, I develop a hierarchical topic model that jointly discovers aspects and their related sentiments and applied the proposed model to the help center post conversation. Finally, I conduct case studies using the proposed model to show that students' question-asking behavior can predict their academic performance, and TA's assistance should be concentrated at the earliest stage of the course. I also report the correlation between sentiments in the student's question and the rate of problem-solving.