TRANSFORMER-BASED EMOTION CLASSIFICATION FOR INDONESIAN POLITICAL SPEECH

Abstract

This study aims to classify emotions in Indonesian political speeches using the RoBERTa Transformer model. The dataset comprises 2,952 paragraphs derived from transcripts of closed-meeting political speeches delivered between 2014 and 2024. The research methodology encompasses data collection, augmentation, automatic emotion labeling, text pre-processing, and the fine-tuning of the RoBERTa model. The model was trained with a batch size of 16, a learning rate of 1e-5, and over three epochs. Evaluation was performed using a multi-class confusion matrix based on accuracy, precision, recall, and F1-score. Experimental results demonstrate that RoBERTa performs effectively, achieving 90% accuracy with a macro-average precision of 91%, recall of 90%, and an F1-score of 90%. Notably, the "happy" label achieved perfect scores across all metrics, while other labels exhibited consistent results. These findings validate the feasibility of Transformer-based approaches for recognizing implicit emotions in formal Indonesian political discourse. This research contributes to expanding Indonesian NLP resources within the political domain and provides empirical evidence for computational political discourse analysis. Future work may focus on expanding emotion categories, enhancing expert validation, and testing alternative Transformer architectures.