<![CDATA[This study evaluates a pipeline that separates rule-based aspect extraction from IndoBERT-based binary sentiment classification for Indonesian Ruparupa mobile application reviews. Google Play reviews were collected on 31 July 2026, anonymized, deduplicated before aspect expansion, and cleaned by lowercasing, removing URLs, emails, and special characters, and normalizing whitespace; no slang normalization, stop-word removal, or stemming was applied. Ratings 1-2 and 4-5 provided weak negative and positive labels, while three-star reviews were excluded. A 331-entry aspect dictionary mapped 1,495 unique reviews into 2,873 aspect-review pairs across six aspects. Across five repeated leakage-free group hold-out splits, IndoBERT achieved mean accuracy 0.9179 ± 0.0214, macro F1 0.9178 ± 0.0214, and ROC-AUC 0.9719 ± 0.0103; a calibrated TF-IDF + linear SVM baseline achieved 0.8765 ± 0.0124, 0.8759 ± 0.0127, and 0.9452 ± 0.0102, respectively. A McNemar test on run 1 showed a significant paired difference (p = 0.00013). Performance measures agreement with rating-derived weak labels rather than human-validated aspect sentiment. Because results from system-assigned aspects lacked independent human validation, aspect frequencies are descriptive rule-system outputs. Within this dataset, IndoBERT performed consistently across the five splits; supervised aspect extraction and human aspect-level annotation remain priorities.]]>
Copyrights © 2026
