Sentiment Analysis and Topic Modeling of Ruangguru Application User Reviews Using IndoBERT and BERTopic on a Kaggle Dataset

Abstract

User reviews provide valuable information for evaluating digital learning platforms because they contain direct expressions of user satisfaction, complaints, and expectations. This study analyzed user reviews of the Ruangguru application using a Kaggle dataset by combining IndoBERT for sentiment classification and BERTopic for topic modeling. The study aimed to classify user sentiment, compare IndoBERT with a TF-IDF and Logistic Regression baseline, identify dominant discussion topics, and examine whether the model performance difference was statistically significant. The dataset originally contained 100,000 reviews, and 76,699 valid reviews were used after data cleaning and preprocessing. Sentiment labels were generated from rating values and grouped into negative, neutral, and positive classes. IndoBERT achieved an accuracy of 0.8920 and an F1 macro score of 0.6347, outperforming the baseline model with an accuracy of 0.8092 and an F1 macro score of 0.5666. McNemar’s test confirmed that the performance difference was statistically significant (chi-square = 594.30, p < 0.001). BERTopic generated 41 topic groups, including one outlier group. Positive sentiment dominated most topics, particularly those related to learning support, material comprehension, and video-based learning. Negative sentiment was concentrated in topics related to payment, application updates, advertisements, and account issues. These findings show that integrating IndoBERT and BERTopic provides a comprehensive understanding of user perceptions toward educational applications.