<![CDATA[This study evaluates the effects of heavy and minimal preprocessing on static and contextual text representations for sentiment classification of Indonesian Shopee reviews. The dataset comprised 29,618 Google Play Store reviews collected in January 2026 and weakly labeled from star ratings into negative, neutral, and positive classes. The data were divided into training, validation, and test sets using an 80:10:10 stratified split. Heavy preprocessing included case folding, text cleaning, slang normalization, stopword removal, and stemming, whereas minimal preprocessing applied case folding only. Four experimental configurations combined the two preprocessing strategies with Word2Vec–Logistic Regression and fine-tuned IndoRoBERTa. Model performance was evaluated using accuracy, weighted precision, weighted recall, weighted F1-score, macro F1-score, and confusion matrices. Heavy preprocessing produced a marginal weighted F1-score improvement for Word2Vec–Logistic Regression, from 86.0% to 86.2%. Minimal preprocessing improved IndoRoBERTa’s weighted F1-score from 88.4% to 89.7% and achieved the best accuracy of 90.7% and macro F1-score of 65.8%. The difference between weighted and macro F1-scores indicates that the neutral minority class remained difficult to identify. The findings show that preprocessing effectiveness depends on model architecture. Minimal preprocessing better preserves contextual information for IndoRoBERTa, whereas aggressive preprocessing provides only a marginal benefit for the static baseline.]]>
