Mitigating Class Imbalance in Indonesian Sarcasm Detection: A Cross-Platform Transformer Study

Abstract

Sarcasm detection in Indonesian social media remains challenging due to implicit pragmatic expressions, severe class imbalance, and strong domain variation across platforms. Unlike prior Indonesian sarcasm studies that predominantly focus on in-domain accuracy using conventional balancing methods, this study provides the first systematic cross-platform analysis of generative data balancing under domain shift. We empirically examine whether GPT-4o based generative balancing improves robustness rather than accuracy-centric evaluation in Transformer-based sarcasm detection. Models trained on Twitter data are evaluated across Twitter, Reddit, and TikTok as an unseen domain. The results show that generative balancing yields limited gains in in-domain evaluation but consistently improves cross-domain robustness by increasing sarcasm recall, particularly for Base models. Notably, XLM-R Base achieves an absolute F1-score improvement of +10.8 points on TikTok, while IndoBERT-Large attains the highest in-domain F1-score of 0.7444. These findings indicate that generative augmentation partially mitigates class imbalance by enhancing robustness under domain shift, thereby repositioning sarcasm detection as a robustness-oriented problem and highlighting generative balancing as a complementary strategy rather than a substitute for larger Transformer models in cross-platform NLP settings.