Penggunaan LSTM dan Fast-Text dalam Analisis Sentimen Berbasis Aspek pada Ulasan Aplikasi dengan Seleksi Fitur Information Gain

Abstract

This study aims to perform aspect-based sentiment analysis on user reviews of the GoTube application from the Google Play Store. The large volume of user reviews makes manual analysis inefficient, requiring an automated approach to extract meaningful insights. Data were collected through web scraping, resulting in over 280,000 reviews, which were further processed through cleaning, labeling, and data balancing. To improve label consistency, a pseudo-labeling approach using IndoBERT was applied. The proposed method combines Information Gain for feature selection, FastText for word representation, and Long Short-Term Memory (LSTM) for sentiment classification. In addition, Latent Dirichlet Allocation (LDA) was used for aspect extraction. The experimental results show that the sentiment classification model achieved an accuracy of 95.37% and F1-score of 0.9537 using an optimal threshold of 0.59, with balanced precision (0.9539) and recall (0.9534). Meanwhile, the aspect classification model achieved an accuracy of 87.12% with a macro F1-score of 0.8697. These findings indicate that the combination of feature selection, subword-based representation, and sequential modeling is effective in producing accurate and informative aspect-based sentiment analysis