<![CDATA[Sentiment analysis of mobile application reviews supports service evaluation in the fast-growing Indonesian digital finance sector. This study compared an English lexicon-based method (VADER) with a fine-tuned Indonesian RoBERTa model on 1,000 reviews of the DANA e-wallet application collected from Google Play. Following best practice for transformer models, reviews were classified primarily from their original (raw) text, and a preprocessed variant was used to assess the effect of text cleaning. On raw text, VADER labelled 903 reviews as neutral, 78 as positive, and 19 as negative, whereas Indonesian RoBERTa produced 494 positive, 385 negative, and 121 neutral, a positive-leaning distribution consistent with the sample's mean rating of 4.06. Inter-method agreement was negligible (Cohen's kappa = 0.008). Removing stopwords and numbers before classification changed 30.2% of the RoBERTa labels and flattened its distribution, confirming that such preprocessing degrades the contextual input transformers rely on, while leaving VADER almost unchanged. Because no human-annotated gold standard was used, the reported figures reflect inter-method agreement rather than absolute accuracy against human labels. The findings show that VADER is unsuitable for Indonesian reviews and that language-specific transformers should be applied to unprocessed text.]]>
Copyrights © 2026
