Predicting Speech Acts in MOOC Forum Posts Using Conditional Random Fields
Kyle Shaffer · Carolina Digital Repository (University of North Carolina at Chapel Hill) · 2019
Massive Open Online Courses (MOOCs) have emerged as a way to reach large numbers of students by providing course materials as free online resources. The popularity of these courses has been reflected in high enrollment numbers, however it is unclear how successful MOOCs are at educating their students given their high attrition rates. One cause for this may be due to instructors' inability to manage the large number of students that enroll. While discussion forums are available for students to seek help, instructors are unable to monitor the large number of posts written in these forums. This study investigates the effectiveness of using machine learning models to classify posts into speech acts as a way to help instructors monitor these discussion forums. Speech acts describe the purpose of a post and may be indicative of common functions such as asking questions or raising issues. A linear classifier is compared against a conditional random field (CRF) classifier, which is able to leverage contextual information about the forum in order to make predictions. The results of this study find that CRFs outperform a simpler linear classifier, and this suggests that casting this prediction problem as a sequence labeling task is fruitful for predicting these speech acts, and automatically identifying posts of interest.