<![CDATA[The rapid growth of e-wallet adoption in Indonesia has generated large volumes of user reviews on Google Play Store, yet document-level sentiment analysis fails to capture that a single review may express different sentiments toward different service aspects. This study applies Aspect-Category Sentiment Analysis (ACSA), a sub-task of Aspect-Based Sentiment Analysis (ABSA) combining Aspect Category Detection (ACD) and Sentiment Classification (ASC), to classify sentiment across five predefined aspects of ShopeePay reviews: Security, Ease of Use/UI, Performance, Features, and Customer Service. From 42,145 scraped reviews, 22,756 unique reviews remained after preprocessing, of which 12,525 (55.0%) contained at least one relevant aspect. Labels were generated using Gemini 3.5 Flash-Lite and validated against 1,500 manually annotated samples via Cohen's Kappa (ACD: κ = 0.744; sentiment: κ = 0.903), with ACD further evaluated against the gold-standard (macro-F1 = 0.794). Using 15,986 binary-labeled instances, IndoBERT was fine-tuned via stratified 5-fold cross-validation and compared against SVM, Naive Bayes, and IndoRoBERTa. IndoBERT consistently outperformed all baselines across the five aspects (F1 = 0.970–0.984), with the largest gap for Ease of Use/UI (0.970 vs. 0.788), and results were confirmed on a human-annotated subset (F1 = 0.971–1.000). Knowledge discovery on the model's predictions showed Features and Performance were the most frequently complained-about aspects (3,035 and 1,455 negative reviews, respectively), suggesting these areas require urgent technical attention, while Ease of Use/UI received the highest volume of positive sentiment (3,819 reviews), indicating its interface design should be preserved rather than redesigned, offering developers a prioritized, data-driven basis for improvement.]]>
Copyrights © 2026
