Enhanced LSTM and BiLSTM-Based Sentiment Analysis of Public Opinion on Ibu Kota Nusantara (IKN)

Abstract

The establishment of Ibu Kota Nusantara (IKN) has triggered extensive public discourse on the social media platform X (formerly Twitter), highlighting the importance of systematically evaluating public sentiment toward this national initiative. However, sentiment classification in this context remains challenging due to heterogeneous opinions, informal linguistic usage, and the large-scale, unstructured nature of textual data. This study examines public sentiment related to IKN using deep learning approaches, specifically Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM), and compares their performance with a conventional Support Vector Machine (SVM) baseline. Indonesian-language tweets collected from April 2024 to April 2025 via Tweet Harvest were processed through comprehensive preprocessing steps, sentiment labeling using the Indonesia Sentiment Lexicon (InSet), data balancing using SMOTE, and semantic representation via Word2Vec Skip-gram embeddings trained on a domain-specific corpus. The deep learning models were trained for 100 epochs with an embedding dimension of 64, while key hyperparameters such as batch size, neuron units, and dropout rates were systematically optimized. Experimental results show that LSTM achieved a validation accuracy of 0.8934, while BiLSTM performed slightly better with 0.8994, indicating improved contextual learning capability. In contrast, SVM obtained the highest accuracy of 0.91, demonstrating strong performance as a classical machine learning baseline. Both LSTM and BiLSTM models achieved ROC-AUC scores above 0.97, reflecting excellent class separability. Overall, although performance differences among models were relatively marginal, BiLSTM provided more consistent sentiment discrimination across classes, suggesting its effectiveness for analyzing Indonesian social media sentiment regarding the IKN development, while SVM remains a computationally efficient and competitive benchmark model.