<![CDATA[The rapid development of internet technology globally, including in Indonesia, has drastically changed communication and interaction patterns between individuals. One impact is seen in the increasing use of text-based social media applications, such as Threads, developed by Meta. Within a short time, Threads managed to attract millions of users. However, the large number of user reviews on the Google Play Store presents its own challenges, particularly in manual sentiment analysis, which is very time-consuming and prone to bias. This research aims to overcome these challenges by implementing a variant of bidirectional encoder representations from transformers (BERT), the robustly optimized BERT pretraining approach (RoBERTa) model, which has been optimized for natural language processing. The research process followed the cross-industry standard process for data mining (CRISP-DM) framework, including several main stages: understanding the business context, data exploration and model building preparation, performance evaluation, and model deployment. Data were obtained directly from the Google Play Store and then cleaned through deduplication, normalization, and tokenization stages. The RoBERTa model demonstrated strong performance, with an accuracy of 88%. Precision was recorded at 92% for positive sentiment and 81% for negative sentiment, while recall was at 88% and 87%, respectively. The F1 score was also high, at 90% for positive and 84% for negative sentiment. When compared to algorithms like naïve Bayes and support vector machine (SVM), RoBERTa proved superior. This research opens opportunities for exploring other transformer models or using ensembles to improve performance in the future.]]>
Copyrights © 2025
