Cross-Course and Multi-course Sentiment Classification of Student Posts

Abstract

Affective Computing is one of the most active research topics in education. Increased interest in emotion recognition through text channels makes sentiment analysis (i.e., the Natural Language Processing task of determining the valence in texts) a state-of-the-practice tool. Considering the domain-dependent nature of sentiment analysis as well as the heterogeneity of the educational domain, development of robust sentiment classifiers requires an in-depth understanding of the effect of the teaching-learning context on model performance. This work investigates machine learning-based sentiment classification on datasets comprised of student posts in forums, pertaining to two different academic courses. Different dataset configurations were tested, aiming to compare performance: i) between single-course and multi-course classifiers, ii) between in-course and cross-course classification. A sentiment classifier was built for each course, exhibiting a fair performance. However, classification performance dramatically decreased, when the two models were transferred between courses. Additionally, classifiers trained on a mixture of courses underperformed single-course classifiers. Findings suggested that sentiment analysis is a course-dependent task and, as a rule of thumb, less but course-specific information results in more effective models than more but non-specialized information.