<![CDATA[Sarcasm detection in Indonesian YouTube comments remains challenging due to contextual ambiguity, limited labeled data, and class imbalance. This study proposes a fine-tuned RoBERTa-based auto-labeling pipeline to generate high-confidence pseudo-labels for unlabeled comments. The scientific contribution lies in integrating Back-Translation augmentation, confidence-threshold filtering at P ≥ 0.85, and comparative evaluation against baseline models on a YouTube sarcasm corpus. The data were collected from 10 public Indonesian YouTube videos covering public service, political, social, and entertainment topics during January-March 2025. From 1,000 raw comments, 493 clean comments, 200 manually labeled instances, and 536 final instances were obtained after augmentation and pseudo-label filtering. The 5-fold cross-validation results show that RoBERTa achieved an accuracy of 0.89 and a macro F1-Score of 0.87, with class-wise precision/recall of 0.81/0.81 for sarcasm and 0.92/0.92 for non-sarcasm. Compared with TF-IDF + SVM, BiLSTM, and IndoBERT, RoBERTa improved the F1-Score by 24.29%, 12.99%, and 3.57%, respectively. These findings indicate that RoBERTa-based auto-labeling can support a more controlled expansion of sarcasm corpora while reducing reliance on fully manual annotation.]]>
Copyrights © 2026
