<![CDATA[This research investigates transfer learning efficacy for five-class emotion classification in Javanese Ngoko. A parallel Indonesian–Javanese Ngoko corpus was synthesized by translating 5,400 samples from the PRDECT-ID dataset using machine translation, with translation quality verified via a preliminary expert validation sample. Using IndoBERTweet as the backbone architecture, three paradigms were evaluated: zero-shot transfer (E1), fully supervised learning (E2), and cross-lingual transfer learning (E3) with identical hyperparameters. Empirical results indicate that the cross-lingual transfer (E3) setup achieved peak performance (67,5% accuracy; 0,67 weighted F1) under the evaluated dataset and experimental setting. Per-class analysis identified that positive affect (Happy) showed cross-lingual stability, whereas negative emotions (Sadness, Fear) suffered degradation due to lexical divergence between the two languages. Training dynamics revealed early-onset overfitting, suggesting model capacity exceeds current dataset density. This work establishes a baseline benchmark for Javanese emotion classification and provides a reproducible machine-translated parallel corpus, emphasizing the need for future validation with native-speaker data to mitigate translation bias.]]>
Copyrights © 2026
