Automated Detection of Fundamental Emotions from Literary Texts of Middle Grade Students

Abstract

Early adolescence, which largely coincides with the middle school years, is characterized by heightened emotional complexity, during which children may experience intense negative emotions such as anxiety, loneliness, guilt, depression, and anger. Accurately identifying these emotional states is critical for supporting emotion regulation, facilitating healthy coping during developmental transitions, and reducing the risk of long-term psychological distress. From both educational and psychological guidance perspectives, early detection of students' emotional conditions through their written expression can enable timely intervention and preventive support within school settings. Emotions are central to literary expression and are often more authentically conveyed through informal and creative writing. Prior research suggests that such texts provide a rich medium for emotion analysis. In this study, we argue that literary texts written freely by children without imposed topics allow for spontaneous and uninhibited emotional expression, making them a valuable data source for educational data mining and student well-being analytics. To this end, we designed and implemented a web-based platform that enables middle-grade students to upload literary works such as poems, fairy tales, and short stories. Using these texts, we applied sentiment analysis and machine learning techniques to identify five fundamental emotions: anger, fear, disgust, sadness, and joy. Multiple text representation methods (Bag-of-Words, TF-IDF, Word2Vec, and Integer Tokenization) and classification models (Logistic Regression (LR), Support Vector Classifiers (SVCs), Multi-layer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and Transformers) were evaluated. Among these, MLP achieved the highest average F1-scores. Across models, joy was consistently detected with the highest accuracy, whereas disgust proved the most challenging emotion to identify, reflecting differences in linguistic expression and emotional salience. Our findings further indicate that machine learning-based emotion classification yields comparable performance on translated and original-language texts, highlighting the feasibility of multilingual or cross-linguistic applications in educational contexts. Model performance is expected to improve with expanded data collection, which can be facilitated by increasing the accessibility and adoption of the platform. With further validation on larger datasets, this system has the potential to be integrated into school guidance and psychological counseling services, enabling systematic monitoring of students' emotional trajectories and supporting early intervention strategies. Such an approach may enhance affective learning environments by bridging educational practice with psychological support mechanisms.