Anugerah Mohamad Setiawan – Poster

Amidst the escalating operational and cyber risks within the digital financial ecosystem, monitoring public opinion on payment services has become crucial for maintaining the stability of Indonesia’s payment system. Therefore, this research aims to develop and evaluate an optimal sentiment classification model to support the implementation of supervisory technology (SupTech) in the Indonesian digital payment domain. To facilitate this evaluation, a new public dataset was compiled from the social media platform X through a hybrid labeling method, which combines automated annotations (from Brand24, ChatGPT, and Gemini) with manual validation. A hybrid IndoBERT-BiGRU architecture was then proposed and comparatively evaluated against several conventional machine learning and deep learning models, and it was further optimized using the back-translation augmentation technique. The experimental results demonstrate that the IndoBERT-BiGRU model consistently achieved the highest performance, with an F1-Score of 87.96%, significantly outperforming other deep learning models such as IndoBERT (85.11%) and FinBERT (68.43%), as well as classical models like Naïve Bayes (76.73%) and Support Vector Machine (SVM) (75.02%). Following the application of back translation, the model’s performance was further enhanced to an F1-Score of 88.46f%, proving the effectiveness of augmentation in improving the model’s generalization capability. In conclusion, this study successfully confirms that the optimized IndoBERT-BiGRU architecture provides a reliable framework for sentiment-based supervisory technology.