<![CDATA[Public complaint management in public services continues to face challenges because complaint data are generally presented as unstructured text containing informal language, abbreviations, and Indonesian-Javanese code-mixing. These characteristics complicate the identification of essential information and the determination of appropriate destination agencies when complaints are processed manually. This study aims to analyze the capability of the IndoBERT model for Named Entity Recognition (NER) in extracting relevant entities from public complaint texts in Semarang City to support complaint routing. A quantitative experimental approach was employed through complaint data collection, BIO-scheme annotation, text preprocessing, IndoBERT fine-tuning, and model evaluation using Precision, Recall, and F1-Score. The dataset consisted of 1,066 complaint sentences containing four target entities: service, location, problem category, and time. Strict-chunk evaluation using seqeval showed that the best IndoBERT checkpoint, obtained at Epoch 3, achieved a Macro F1-Score of 0.62, with the highest performance for LAYANAN (F1=0.68) and the lowest for LOKASI (F1=0.56). Error analysis further indicated that entity-boundary shifts, category ambiguity, and lexical variation in local and informal expressions remained important factors affecting recognition performance. The findings demonstrate that IndoBERT can transform heterogeneous complaint narratives into structured entity information and provide a foundation for improving the efficiency of public complaint management and routing within smart governance systems.]]>
Copyrights © 2026
