Analisis Sentimen Berbasis Aspek dengan LDA dan IndoBERT pada Ulasan Aplikasi Stockbit

Abstrak

This study aims to analyze sentiment in user reviews of the Stockbit application using a topic modeling approach combined with IndoBERT-based sentiment classification. Aspect extraction was carried out using Latent Dirichlet Allocation (LDA), and the experimental results indicate that selecting five topics (n_components = 5) provides the most optimal representation, as evidenced by a topic coherence score of 0.6191. These five topics reflect semantic structures that are highly relevant to the content of the reviews. For the sentiment classification stage, the IndoBERT-base model achieved an accuracy of 90.86%. The best performance was observed for the positive class, with an F1-score of 93.73%, while the negative class yielded an F1-score of 83.12%. This performance gap is attributed to the imbalanced data distribution, where positive sentiments are more dominant. Nevertheless, the macro-average F1-score of 88.43% demonstrates that the model is still capable of classifying both classes in a relatively balanced manner.