<![CDATA[Public opinions expressed on social media are often multidimensional, making conventional single-label sentiment analysis insufficient for capturing the complexity of public discourse. Multilabel sentiment analysis enables a text instance to be associated with multiple categories simultaneously, providing a more comprehensive representation of public perceptions. This study presents a comparative analysis of FastText and Bidirectional Long Short-Term Memory (Bi-LSTM) for multilabel sentiment classification using Indonesian social media data related to the Asset Confiscation Bill (RUU Perampasan Aset). Data were collected from X (formerly Twitter) and YouTube and annotated into twelve predefined labels encompassing legal, political, economic, administrative, and governance dimensions. The proposed framework consisted of data collection, text preprocessing, multilabel annotation, model development, and performance evaluation using Hamming Score. Three experimental scenarios were conducted for each model to evaluate the impact of Label Attention and architectural optimization. Experimental results demonstrated that FastText consistently outperformed Bi-LSTM across all scenarios. FastText combined with Label Attention and Label Normalization achieved the highest Hamming Score of 0.987450, while Bi-LSTM attained its best performance of 0.967833 using Dense Layer Optimization and Label Attention. The findings indicate that FastText is more effective in handling noisy Indonesian social media texts due to its subword embedding capability. This study contributes empirical evidence regarding the effectiveness of lightweight embedding models for multilabel sentiment analysis and provides insights for future applications in public policy monitoring and social media analytics.]]>
