<![CDATA[Emotion classification is an important task in Natural Language Processing (NLP) for identifying emotional information conveyed through Indonesian-language text. However, differences in model specialization may lead to different classification performance, while direct comparisons between language-oriented and emotion-oriented models under the same experimental setting remain limited. This study compares the performance of IndoRoBERTa and EmoRoBERTa for emotion classification using the Indonesian version of the GoEmotions dataset, with mBERT and TF-IDF with SVM included as baseline approaches. The dataset consists of 45,691 single-label samples categorized into seven emotions based on Ekman’s Taxonomy, namely anger, disgust, fear, joy, sadness, surprise, and neutral. The experimental process includes text preprocessing, class weighting to address class imbalance, and hyperparameter tuning to determine the optimal configuration for each model. Model performance is evaluated using accuracy, macro precision, macro recall, and macro F1-score, with macro F1-score as the primary metric. Under the dataset and experimental settings used in this study, IndoRoBERTa obtains an accuracy of 0.6603, macro precision of 0.5938, macro recall of 0.5975, and macro F1-score of 0.5887, while EmoRoBERTa obtains 0.6326, 0.5613, 0.5956, and 0.5742, respectively. mBERT and TF-IDF with SVM obtain macro F1-scores of 0.5495 and 0.5449, respectively. The results provide an empirical comparison of language-oriented and emotion-oriented Transformer models within the experimental setting and provide baseline context for interpreting their classification performance.]]>
