Mapping Public Emotions Regarding the 2025 UTBK Announcement in Indonesia: A Multi-Label Approach with Targeted Calibration

This study maps public emotions in Indonesian-language tweets related to the 2025 UTBK announcement using multi-label emotion classification. The main challenges in multi-label emotion classification on social media include extreme label imbalance, distribution shift between training and application data, and weak lexical signals for specific emotions. This study aims to build a reliable emotion modeling framework for long-tail social media corpora while demonstrating generalizable post-training calibration practices. The novelty lies in the integration of four components: (1) per-label posterior calibration using Platt scaling, (2) precision-targeted per-label thresholding frozen from the development set, (3) score-quantile–based rate targeting to align predicted prevalence with domain-based rates, and (4) context-limited lexicon-aware boosting with a final clamp. The proposed pipeline is lightweight and model-agnostic. This research adopts a quantitative experimental approach by varying post-training calibration components to measure their impact on classification performance. An IndoBERTweet model is trained using BCEWithLogitsLoss on manually annotated data, then calibrated and evaluated on development and test sets. The results demonstrate balanced micro- and macro-level performance, improved detection of minority labels, and emotion mapping over 3,500 tweets with prevalence distributions consistent with Plutchik’s theory of emotions.