Analisis Performa Model LSTM, GRU Dan Varian IndoBERT Untuk Deteksi Emosi Pada Teks Bahasa Indonesia

This study evaluates gating mechanisms, specifically Long Short-Term
Memory (LSTM) and Gated Recurrent Unit (GRU), in comparison with
attention-based models utilizing IndoBERT variants (Base, Large, and
Lite) for Indonesian emotion detection across six emotion labels. The
evaluation examines accuracy, efficiency, and robustness using both in-
distribution and out-of-distribution (OOD) datasets collected from social
media. Statistical significance is assessed through confidence interval
estimation and bootstrap paired tests, and a detailed error analysis is
conducted to identify model limitations. The results indicate that
IndoBERT Large achieves superior performance, with a Macro F1-Score
of 80.05% and greater robustness to domain shifts, whereas gating
models exhibit substantial performance degradation on unseen data. In
contrast, GRU outperforms LSTM and achieves the lowest inference
latency, with training times up to 131 times faster than IndoBERT Large.
Statistical tests confirm that the performance gap between IndoBERT
variants and RNN-based models is significant. These findings highlight
a key trade-off: attention mechanisms provide state-of-the-art accuracy
and robustness, while GRU offers a practical and efficient solution for
resource-constrained settings.