Quantitative Analysis of Indonesian Sentiment Classification Model Robustness Against Typographic Noise via Word-Level and Character-Level Feature Fusion

Informal Indonesian text on social media frequently contains typographic noise such as slang substitution, character repetition, and abbreviations, which can significantly degrade the performance of sentiment classification models. This study proposes a Hybrid Word-Character Feature Fusion architecture combining a Bidirectional Long Short-Term Memory (BiLSTM) word-level branch with a dual Convolutional Neural Network (CNN) character-level branch. The model is evaluated on the Twitter PPKM Indonesia dataset (23,644 tweets, three sentiment classes) under six linguistically motivated noise categories at three noise levels (10%, 20%, 30%). Results demonstrate that the Hybrid model reduces accuracy drop rate by 58.82% at noise level 30% compared to the BiLSTM baseline (3.06% to 1.26%), with Robustness Score improving from 0.9694 to 0.9874. Although performance differences did not reach statistical significance (Paired t-test, Wilcoxon Signed-Rank test, α = 0.05), the consistent directional trend and negligible Cohen's d values confirm that the hybrid architecture provides meaningful robustness gains while maintaining competitive classification accuracy.