<![CDATA[Class imbalance and the use of non-standard language in football supporters’ opinions on social media constitute major obstacles to producing accurate sentiment classification for evaluating federation performance. This study aims to identify the most effective bidirectional recurrent architecture for capturing public opinion after applying data balancing techniques. Using a primary dataset of 1,039 instances (604 positive and 435 negative samples), the proposed method integrates a pre-trained NusaBERT model with hybrid Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Unit (BiGRU) layers. To address data imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied to the training data, with dataset partitioning using a stratified split ratio of 70:30. The results indicate that the NusaBERT-BiLSTM model achieves the best performance, with a testing accuracy of 70.83% and an F1-score of 0.6990, outperforming the BiGRU variant, which attains an accuracy of only 64.74%. Furthermore, NusaBERT-BiLSTM demonstrates greater reliability in detecting negative sentiment, achieving a recall value of 0.6336 compared to 0.4504 for BiGRU. In conclusion, combining NusaBERT's semantic strength with SMOTE-based balancing and BiLSTM layers significantly enhances the model’s sensitivity to minority opinions without causing data leakage. This study contributes a more objective classification model for national team management to accurately map public criticism and aspirations on social media.]]>
Copyrights © 2026
