A HYBRID BERT-BILSTM-ATTENTION MODEL FOR PUBLIC SENTIMENT ANALYSIS ON A NATIONAL SOCIAL INSURANCE PROVIDER IN INDONESIA

Abstract

Sentiment analysis in Indonesian social media presents significant challenges due to informal language, code-mixing, and sentiment ambiguity in multi-clause expressions. While transformer-based models such as IndoBERT effectively capture contextual semantics, they may struggle to model sentiment transitions and resolve conflicting polarity in noisy and heterogeneous text. To address this limitation, this study employs a hybrid BERT–BiLSTM–Attention model that integrates contextual, sequential, and attention-based representations. Although the architecture itself is not novel, the contribution lies in its systematic integration and empirical evaluation under realistic conditions. Experimental results show that the proposed model achieves an accuracy of 0.85 and a Macro-F1 score of 0.85, outperforming the IndoBERT baseline (0.83) by approximately 2.4%. This improvement is statistically significant (p = 0.027) and supported by effect size analysis, indicating meaningful performance gains. Robustness evaluation under controlled perturbations—including slang injection, character elongation, emoji usage, code-mixing, and class imbalance (up to 30% minority downsampling)—shows only minor performance degradation (0.02–0.03 Macro-F1), demonstrating stable generalization under noisy conditions. These findings provide empirical evidence that the integration of sequential modeling and attention improves the handling of sentiment transitions and multi-clause structures beyond transformer-only approaches, offering practical value for real-world sentiment analysis in Indonesian social media.