Reading Emotions Behind TikTok Text: Fine-Tuning IndoBERT for Nine-Class Emotion Classification in the Indonesian Language

Abstract

This study focuses on fine-tuning IndoBERT for nine-class emotion classification of Indonesian TikTok text. A computational BAHAGIA pipeline is developed covering text preprocessing, label construction, class-imbalance handling, transformer fine-tuning, evaluation, and model export. The primary dataset contains 8,834 TikTok records, of which 2,775 contain emotion labels and form the labeled classification basis. Fine-tuned IndoBERT is compared with a TF-IDF–Logistic Regression baseline using accuracy, macro-F1, and weighted-F1. The documented repository results report 78.0% accuracy, 0.80 macro-F1, and 0.78 weighted-F1 for IndoBERT, compared with 73.7%, 0.73, and 0.74 for the baseline. The 0.07 macro-F1 improvement indicates better average performance across classes in the available evaluation. The main contribution is a contextual IndoBERT-based computational approach for fine-grained emotion classification in Indonesian social-media text. Heuristic labels and limited human validation constrain interpretation; the model is intended for text classification and social listening rather than psychological diagnosis.