Leveraging explainability for discussion forum classification: Using confusion detection as an example

Abstract

Abstract Online discussion forums are highly valued by instructors due to their affordance for understanding class activities and learning. However, a discussion forum with a great number of posts requires a large amount of time to view, and help requests are easily overlooked. Various machine-learning–based tools have been developed to help instructors monitor or identify posts that require immediate responses. However, the black-box nature of deep learning cannot explain why and how decisions are achieved, raising trust and reliability issues. To address the gap, this work developed an explainable text classifier framework based on a model originally designed for legal services. We used the Stanford MOOCPost dataset to identify posts of confusion. Our results showed that the framework can not only identify discussion forum posts with confusion of different levels, but also provide explanation in terms of words from the identified posts.