Sentiment and Public Emotion Classification of Viral Content Using Transformer-Based Model

Abstract

The proliferation of social media platforms has generated an unprecedented volume of viral content, each drawing varied public responses expressed through sentiment and emotion. Mapping those responses — not merely counting them — is what separates surface-level monitoring from a genuine understanding of public perception. This study classified sentiment (positive, negative, neutral) and emotion (anger, joy, sadness, and fear) toward viral content using a fine-tuned Transformer-based model. Data were collected from social media via web scraping, then subjected to standard text preprocessing: case folding, tokenization, stopword removal, and stemming. The cleaned dataset was subsequently annotated with sentiment and emotion labels. BERT (Bidirectional Encoder Representations from Transformers) served as the base architecture, fine-tuned for multi-label classification. Evaluation relied on an 80:20 train-test split, with performance measured through accuracy, precision, recall, and F1-score. Across all sentiment and emotion categories, the model returned consistently high scores and handled ambiguous, context-dependent text more reliably than conventional machine learning baselines. The Transformer-based approach proved well-suited for sentiment and emotion analysis on social media data, with clear potential for deployment in public opinion monitoring systems.