<![CDATA[This study examines how text preprocessing depth and model architecture jointly affect aspect-based sentiment analysis (ABSA) of MyTelkomsel application reviews. A total of 850 cleaned Google Play reviews were weakly labeled into 1,036 aspect-sentiment instances covering package price, network signal, and customer service using a keyword lexicon and a rating-derived heuristic. Three preprocessing frameworks of increasing complexity (transformer-oriented, hybrid, and augmented hybrid with class-imbalance handling) were each evaluated on three IndoBERT-based architectures (a frozen baseline, full fine-tuning, and an IndoBERT+BiLSTM hybrid) under an identical evaluation protocol. The augmented hybrid framework combined with full fine-tuning achieved the best performance (accuracy 92.16%, macro F1-score 0.7406), ahead of the transformer-oriented (macro F1 = 0.6219) and hybrid (macro F1 = 0.6404) frameworks, and its advantage for the fine-tuned model was consistent across three random seeds, although these differences do not survive correction for multiple comparisons at three seeds. Fine-tuning consistently outperformed both the frozen baseline and the BiLSTM extension, while the customer service aspect remained the hardest to classify because of class imbalance and weak-label noise. A post-hoc audit found that 61.5% of test instances share near-identical text with training instances, an artifact of splitting multi-aspect reviews at the row level; absolute scores are therefore upper bounds, and although all conditions share one split, the interaction between this overlap and the augmented framework's oversampling remains untested. Unlike prior Indonesian IndoBERT-based ABSA studies that fix a single preprocessing pipeline, this study isolates the separate contributions of preprocessing depth and model architecture.]]>
Copyrights © 2026
