Learning-Related Emotion Detection from Serbian Text

This paper tackles the problem of learning-related emotion detection in Serbian text. Previous studies that detected learning-related emotions predominantly focused on feedback provided after learning sessions (e.g., at the end of the semester). Though this approach provides insights into students’ emotional states, it cannot support dynamic adaptability to their needs. To address this problem, we asked students to express their emotions as they arise during learning sessions. Previous studies focused on large-resource languages, such as English, and, to the best of our knowledge, none considered Serbian. We performed a case study to build Machine Learning (ML) models that can detect learning-related emotions. We asked the students to explicitly express their learning-related emotions every 20 min during their interactions with the Intelligent Tutoring System that supported pre-class activities for a blended flipped classroom software engineering course. Five annotators annotated the collected 955 instances for five emotions. We trained ML models using different text representations – BERTić, ELMO, and TF-IDF. The highest F-measure (0.75) was achieved using TF-IDF and BERTić representations.