<![CDATA[User-generated Indonesian text frequently exhibits code-mixing with English (“Indonglish”), informal spelling, elongation, and keyboard typos. These phenomena break subword tokeniza tion assumptions and may degrade multilingual Transformer performance in deployment. This paper studies a hybrid representation that fuses XLM-R sentence features with a character-level CharCNN branch designed to capture orthographic patterns and mitigate character noise. We evaluate (i) a standard XLM-R fine-tuning baseline, (ii) an ablation that removes the character branch (NusaX only), and (iii) the proposed hybrid model on two datasets: NusaX-Senti (12 regional languages) and Indonglish (Indonesian–English code-mixed sentiment). Beyond clean test performance, we introduce a controlled robustness protocol by injecting character-level perturbations with probability p=0.18 and measuring performance drop. Results show that the XLM-R baseline achieves the best clean Macro-F1 on both datasets, while the hybrid model substantially improves robustness on Indonglish by reducing Macro-F1 drop from 0.030 to 0.007 under noise. We analyze common error confusions and discuss when character-aware features help or harm across languages.]]>
Copyrights © 2024
