Analisis Sentimen Terhadap Layanan Kereta Api Indonesia Menggunakan Fine-Tuned Indobertweet

Abstract

The increasing use of digital public transportation services has led to a growing volume of user feedback on social media platforms, particularly Twitter. Such feedback provides valuable insights for service evaluation; however, the large volume and unstructured nature of the data make manual analysis inefficient. Therefore, this study aims to analyze public sentiment toward Indonesian Railway services using a fine-tuned IndoBERTweet model. This research adopts a case study approach by collecting Indonesian-language tweets through a crawling process and manually labeling them into positive, neutral, and negative sentiment classes. The IndoBERTweet model is trained and evaluated under several preprocessing scenarios, including a baseline without stopword removal and stemming, the application of stopword removal, stemming, and their combination. In addition, the impact of data imbalance is examined by evaluating several imbalance handling techniques, namely random oversampling, class weighting, text augmentation, and random undersampling, which are applied to the bestperforming scenario. Model performance is evaluated using accuracy, precision, recall, macro F1-score, and confusion matrix metrics. The results indicate that the baseline scenario without stopword removal and stemming achieves the best performance, with a macro F1-score of 0.7702 and an accuracy of 0.8333. Additional preprocessing techniques do not significantly improve performance and tend to reduce effectiveness on the minority class. Although imbalance handling techniques increase sensitivity toward positive sentiment, they do not outperform the baseline configuration overall. These findings suggest that minimal preprocessing, combined with careful consideration of data distribution, is more suitable for sentiment analysis using IndoBERTweet on Twitter data.