Improving Neutral Sentiment Classification in Indonesian E-Wallet Reviews Using Word2Vec and Easy Data Augmentation (EDA)
TL;DR
It is demonstrated that semantic representations and controlled lexical variation can jointly enhance minority-class recognition in short informal Indonesian text and highlight the importance of aligning embedding strategies with sequence architectures.
Abstract
The rapid expansion of digital payments has produced massive volumes of user-generated reviews, making manual analysis impractical. This study focuses on the challenge of neutral sentiment classification in Indonesian e-wallet reviews, where neutral comments often contain ambiguous language and are underrepresented relative to positive and negative classes. A total of 26,537 preprocessed DANA application reviews were used to evaluate whether Word2Vec embeddings and Easy Data Augmentation (EDA) can improve neutral sentiment detection when combined with Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) architectures. Experiments comparing eight model configurations showed that the combination of Word2Vec, EDA, and LSTM achieved the best performance, with 0.861 accuracy, 0.841 macro-F1, and 0.749 F1-score for the neutral class. These findings demonstrate that semantic representations and controlled lexical variation can jointly enhance minority-class recognition in short informal Indonesian text and highlight the importance of aligning embedding strategies with sequence architectures.