<![CDATA[Bengkulu Malay is an under-resourced regional language used in daily digital communication, including public comments on Instagram and TikTok. Its informal vocabulary, spelling variation, and limited language resources make sentiment classification more challenging than for standard Indonesian. This study compares Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (BiLSTM) for sentiment analysis of Bengkulu Malay comments. The dataset was collected from three local issues, namely Q1, Q2, and Q3, with 22,888 raw comments and 3,961 labeled comments after selection. The research pipeline includes data collection, labeling, text preprocessing adapted to Bengkulu Malay, Word2Vec Skip-gram embedding, data splitting into training, validation, and testing sets, model training, and evaluation using accuracy, precision, recall, F1-score, and confusion matrix. The final labeled dataset was split into 80% training and 20% testing data, and 20% of the training data was used for validation. The results show that BiLSTM achieved the best accuracy on all datasets, with 91% on Q1, 85% on Q2, and 70% on Q3. CNN achieved 69%, 60%, and 64%, while LSTM achieved 49%, 55%, and 64%. These results indicate that BiLSTM is more effective in capturing bidirectional context in Bengkulu Malay comments, although performance decreases when class imbalance becomes stronger.]]>
Copyrights © 2026
