Analisis Sentimen Berbasis Aspek untuk Ulasan Aplikasi Investasi Online dengan Pendekatan Bidirectional Encoder Representations from Transformers (BERT)

Abstract

The rapid growth of investment applications in Indonesia requires a deep understanding of user reviews. This study aims to analyze the sentiment and aspect found in user reviews of the AJAIB application, collected from the Google Play marketplace, totaling 11,004 data. Aspect identification was conducted using the Latent Dirichlet Allocation (LDA) method, which successfully extracted three main topics with a coherence score of 0.465, interpreted as interface, performance, and transaction. Furthermore, sentiment analysis is performed using the Bidirectional Encoder Representations from Transformers (BERT) model. In this research, the type of BERT used is IndoBERT, which is a pre-trained model optimized specifically for Indonesian. The IndoBERT model is then used to classify sentiment into two categories, namely positive and negative. The BERT model was trained with a batch size of 32, a learning rate of 5e-5, and over 3 epochs, achieving an accuracy of 91%, with a precision of 78%, recall of 88%, and F1-score of 82% for negative sentiment and a precision of 96%, recall of 92%, and F1-score of 94% for positive sentiment. As a benchmark, the Support Vector Machine (SVM) model was used as a baseline. The evaluation results show that the BERT model generally outperforms the SVM model across nearly all evaluation metrics, except for precision in the positive sentiment class and recall in the negative sentiment class, where both models achieved equal scores. These findings indicate that the combination of LDA and BERT approaches can be effectively used to understand user opinions and needs in online investment applications.