Transformer-Based Classification of Indonesian Public Service Complaints

A Case Study of a Municipal Online Reporting Channel

Abstract

Municipal reporting channels receive thousands of free-text complaints every month, and manual routing to the responsible agency delays follow-up. This study fine-tunes a multilingual transformer model to classify Indonesian complaints into twelve service categories. A dataset of 18,400 anonymised complaints was split into training, validation, and test sets, and the transformer was compared with TF-IDF logistic regression and a bidirectional LSTM. The transformer reached a macro F1-score of 0.87, against 0.78 for logistic regression and 0.82 for the LSTM, with the largest gains in categories that share vocabulary, such as roads and drainage. Error analysis shows that most remaining mistakes involve complaints that mention several services at once. The results indicate that transformer-based routing can shorten the time a complaint waits before it reaches the right agency.