Analisis deteksi emosi pada teks Bahasa Indonesia dan Bahasa Jawa dengan metode cross-lingual menggunakan modern multilingual bert

Emotion detection in text is a major challenge in Natural Language Processing (NLP), especially for low-resource languages such as Indonesian, Sundanese, and Javanese. This study aims to evaluate the performance of Transformer-based models in solving multi-label emotion classification tasks in cross-language scenarios. The tested models include ModernBERT and mmBERT trained on English and Sundanese datasets. Data analysis revealed challenges in the form of class imbalance, so this study applied a class weighting configuration to the BCEWithLogitsLoss loss function to increase the model's sensitivity to minority classes. Performance evaluation was conducted using the macro F1-score metric. Experimental results show that the ModernBERT model achieves optimal performance in monolingual English scenarios with a macro F1-score of 0.685. Meanwhile, the mmBERT model shows consistent superiority across all cross-language tasks. Comparison of the effects of training languages reveals that the volume of English data provides an advantage in training compared to the proximity of the language family.