<![CDATA[This study explores the use of the BERT model for sentiment analysis of the SIREKAP system, focusing on preprocessing and model optimization. Lexicon-based sentiment labeling revealed that 8% of the data was positive, 85.1% negative, and 6.9% neutral. These results were refined through manual evaluations by two judges, integrated with lexicon-based labels using majority voting. To address class imbalance, SMOTE was applied, leading to notable performance improvements. The BERT model's accuracy showed substantial enhancement after SMOTE was implemented. A lower learning rate of 1×10⁻⁵ consistently produced strong results, especially with larger batch sizes. The highest average accuracy of 98.9% was achieved at epoch 1 with a learning rate of 2.5×10⁻⁵ and a batch size of 8. This result indicates a significant improvement in model performance compared to initial results.]]>
Copyrights © 2026
