<![CDATA[The rapid growth of social media, particularly X (Twitter), has generated massive volumes of informal and highly contextual opinion data, making sentiment analysis a challenging task. This study aims to systematically evaluate the performance of Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and two hybrid architectures (CNN BiLSTM and BiLSTM–CNN) in classifying Indonesian-language Twitter sentiment related to the Makan Bergizi Gratis (MBG) policy using frozen IndoBERT embeddings. A quantitative comparative experimental design with stratified 10 fold cross-validation was employed on a dataset of 1,569 tweets that underwent case folding, cleansing, and normalization. Model performance was assessed using accuracy, precision, recall, F1score, ROC-AUC, and computational efficiency (training and inference time), while latent feature quality was analyzed through t-SNE visualization. The results show that hybrid architectures provide more stable and competitive performance than single models. CNN achieved the fastest computation, whereas BiLSTM and hybrid models were superior in capturing sequential context. tSNE visualizations further indicate clearer class separation for hybrid models. These findings confirm that under a uniform IndoBERT embedding, hybrid CNN–BiLSTM and BiLSTM CNN offer the best trade-off between accuracy and efficiency for Indonesian Twitter sentiment analysis.]]>
Copyrights © 2026
