<![CDATA[The development of digital public services such as the Sapawarga application generates a large volume of user reviews, most of which are unstructured, making manual analysis inefficient. In addition, the use of mixed languages, including Sundanese and local slang, increases linguistic complexity that is difficult to handle using conventional methods. This study aims to compare the performance of Support Vector Machine (SVM) and IndoBERT using a dual-pipeline preprocessing approach to preserve data context. The dataset consists of 4,277 reviews classified into positive, negative, and neutral sentiments. The results show that IndoBERT outperforms SVM with an accuracy of 80.02% and an F1-Macro of 0.6517, while SVM achieves an accuracy of 75.82% and an F1-Macro of 0.6161. This advantage is supported by IndoBERT’s ability to understand context through self-attention and subword tokenization. This study supports the development of an automated public opinion monitoring system (Digital Ombudsman) for the Government of West Java to improve the quality of bureaucratic response.]]>
Copyrights © 2026
