<![CDATA[Online marketplaces in Indonesia generate massive unstructured consumer reviews laden with slang and informal language, posing challenges for standard sentiment analysis systems. This Systematic Literature Review (SLR), following PRISMA 2020 guidelines, synthesizes 32 primary studies from Google Scholar, Scopus, and IEEE Xplore (2014–2025) to address four analytical research questions on slang normalization effectiveness, LSTM performance relative to dataset characteristics, trade-offs of imbalanced data handling methods, and contexts where F1-Score outperforms accuracy. Results show dictionary-based normalization achieves 87.3% accuracy versus 59% for automated methods. LSTM consistently outperforms traditional models on large, long-text datasets but remains sensitive to preprocessing quality. SMOTE is the most adopted technique yet underperforms at imbalance ratios exceeding 10:1. F1-Score is confirmed as the most representative metric for skewed datasets. This study proposes an integrated conceptual framework and identifies future research gaps including sarcasm detection, code-mixing, and Transformer-based approaches.]]>
Copyrights © 2026
