Assessing the Efficacy of Transfer Learning for Emotion Classification in Low-Resource Javanese Text

Abstract

Despite recent developments in natural language processing (NLP), emotion recognition continues to face substantial computing problems when applied to under-resourced regional languages like Javanese. A fundamental, unresolved challenge in the current study is defining the appropriate adaptation strategy, whether to promote multilingual breadth, or rely on geographically proximal monolingual depth. To evaluate this trade-off, our work systematically compares three pre-trained Transformer architectures, namely IndoBERT, XLM-RoBERTa, and IndoRoBERTa, by fine-tuning them on the Nusa-Writes Javanese emotion dataset. We thoroughly assessed these models across multiple hyperparameter settings, specifically studying the influence of learning rate modifications and the number of training epochs. The ensuing performance analysis focuses on accuracy, weighted F1-score, and training efficiency measures. Our experiments show that IndoBERT achieved the highest weighted F1-score among the evaluated configurations. These results highlight the trade-off that exists in Javanese NLP between broad multilingual generalizability and the subtle benefits of domain-specific pre-training. In summary, we develop a reproducible benchmark pipeline for emotion classification in under-resourced languages.