Roblox is a community-based gaming platform that is extremely popular among users of various age groups. Millions of user reviews available on the platform contain valuable information regarding user satisfaction, expectations, and criticisms of the gameplay experience. To extract insights from these reviews, a reliable natural language processing (NLP) approach tailored to the local language context is essential. This study aims to classify sentiments in Indonesian-language user reviews of Roblox into three categories: positive, negative, and neutral. The model used is IndoBERT, a transformer-based model specifically trained to understand the structure and vocabulary of the Indonesian language. One of the main challenges in this study is the imbalance in the number of data points across sentiment classes. To address this, the SMOTE (Synthetic Minority Over-sampling Technique) method is applied to strengthen the representation of minority classes. The dataset consists of thousands of reviews that have been manually labeled by annotators. Model performance is evaluated using accuracy, precision, recall, and F1-score metrics. The results show that the combination of IndoBERT and SMOTE provides significant improvements compared to the baseline approach without oversampling. This research contributes to the development of automated sentiment analysis systems in the Indonesian language, which can be applied across various digital platforms. The implementation of this model can assist game developers and product analysts in efficiently understanding user opinions, thereby driving improvements in service quality and user satisfaction in the future.
Ansyah et al. (Sat,) studied this question.