<![CDATA[Government complaint management systems often suffer from extreme class imbalance, where a few public service categories accumulate most reports while many others remain under-represented. This research examines whether simple class weighting can improve fairness in multilingual transformer models for automatic routing of Indonesian citizen complaints on the LaporGub Central Java e-governance platform. The dataset comprises 53,877 Indonesian-language complaints spanning 18 service categories with an imbalance ratio of about 227:1 between the largest and smallest classes. After cleaning and deduplication, we stratify the data into training, validation, and test sets. We compare three approaches: (i) a linear support vector machine (SVM) with term frequency inverse document frequency (TF-IDF) unigram and bigram and class-balanced weights, (ii) a cross-lingual RoBERTa (XLM-RoBERTa-base) model without class weighting, and (iii) an XLM-RoBERTa-base model with a class-weighted cross-entropy loss. Fairness is operationalised as equal importance for categories and quantified primarily using the macro-averaged F1-score (Macro-F1), complemented by per-class F1, weighted F1, and accuracy. The unweighted XLM-RoBERTa model outperforms the SVM baseline in Macro-F1 (0.610 vs 0.561). The class-weighted variant attains similar Macro-F1 (0.608) while redistributing performance towards minority categories. Analysis shows that class weighting is most beneficial for categories with a few hundred to several thousand samples, whereas extremely rare categories with fewer than 200 complaints remain difficult for all models and require additional data-centric interventions. These findings demonstrate that multilingual transformer architectures combined with simple class weighting can provide a more balanced backbone for automated complaint routing in Indonesian e-government, particularly for low- and medium-frequency service categories.]]>
Copyrights © 2026
