<![CDATA[Public sentiment toward government digital services in Indonesia is increasingly expressed through application reviews and social media, yet the choice of natural language processing (NLP) model for Indonesian-language text remains inconsistent across studies. This study aims to identify the most effective NLP models for Indonesian sentiment analysis published in 2021–2026, with a particular focus on the public service domain. A systematic literature review was conducted following the PRISMA 2020 statement and the Kitchenham and Charters guidelines, covering Scopus, IEEE Xplore, ScienceDirect, SpringerLink, ACL Anthology, arXiv, SINTA, Garuda, and Google Scholar. Eighteen primary studies met the inclusion criteria and were synthesised descriptively and thematically along five research questions. The results show a shift from classical machine learning (Naive Bayes, SVM) in 2021–2023 toward Indonesian-specific transformer models (IndoBERT, IndoBERTweet) and hybrid architectures in 2024–2025. In within-study comparisons, IndoBERT and its variants generally outperformed classical and recurrent models, reaching accuracy or F1-scores above 0.90 on public service datasets such as Mobile JKN and Coretax. Large language models are emerging but have not consistently surpassed fine-tuned encoders on standard sentiment benchmarks. The main gaps are reliance on automatic lexicon- or rating-based labelling, accuracy reporting on imbalanced data without macro-F1, and the scarcity of aspect-based and shared public service benchmarks. IndoBERT-family models are recommended as the default baseline, together with manually annotated public service corpora and macro-F1 reporting.]]>
