<![CDATA[Hate speech detection on Indonesian social media remains challenging due to the coexistence of formal and highly colloquial language, as well as the moderate class imbalance typical of real-world datasets. Models trained under these conditions often skew toward the majority class and generalize poorly across linguistic registers. This study investigates whether a simple, training-free model-level ensemble can improve Indonesian hate speech detection under such conditions without resampling the data. IndoBERTweet and IndoRoBERTa, pretrained respectively on informal Twitter text and broader formal corpora, serve as complementary base models, and their class probabilities are combined through equal-weight soft voting. On the Indonesian Hate Speech Superset (N = 14,306), evaluated across five random seeds with paired significance testing, the soft-voting ensemble attains a macro-averaged F1 of 0.898 ± 0.003 and a macro recall of 0.899 ± 0.003. It significantly outperforms a TF-IDF SVM baseline and the IndoRoBERTa base model, while showing no significant difference from the stronger IndoBERTweet base model and a trained logistic-regression stacking ensemble. Notably, the ensemble matches the stacking ensemble without any additional training stage or meta-learner, and a calibration analysis shows it improves probability calibration over both base models. These results indicate that equal-weight probability averaging is a simple, reproducible, and competitive strategy for Indonesian hate speech detection under moderate class imbalance.]]>
Copyrights © 2026
