Fine-Grained Emotion Classification in Student Tweets Using BERT and RoBERTa Models

Abstract

The most common themes mentioned in student tweets relate to examinations, assignment deadlines, and celebration after completing an examination. This dataset provides valuable insights into human emotions and therefore requires further investigation on how NLP models detect such sentiment signals from the dataset. Sentiment analysis techniques typically rely on polarity classifications of sentiments, for example, positive and negative sentiment types. Unfortunately, polarity-based sentiment classification techniques might be inefficient in capturing specific details regarding emotional expressions in social media content [16]. For this reason, six types of emotion, including anxiety, stress, sadness, happiness, motivation, and satisfaction, were studied in this paper. The dataset was generated from the Sentiment140 dataset and consisted of approximately 52,000 tweets related to students. As shown in the results, RoBERTa achieved better results in emotion detection with accuracy of 28.9% compared to the BERT model at 28.4% accuracy [2][3]. Though the performance of these NLP models remains average due to noisy labeling and overlapping emotions, class-wise results were helpful in comprehending the challenges of fine-grained emotion classification.