YouTube has the potential to become a propaganda medium that can disrupt national security stability. So comments on YouTube videos can be used as a medium for assessing public opinion which is formed from acts of propaganda. The presence of Artificial Intelligence technology makes it easier to assess public opinion but is hampered by accuracy and datasets. This research proposes modeling using Bi-Directional Long Short-Term Memory (Bi-LSTM) and Latent Dirichlet Allocation (LDA) using quantitative methods to assess sentiment categories and topics from comments on YouTube videos. This research aims to design a model that is novel in the form of combining a sentiment analysis model with the addition of topic category output to support intelligence data collection on social media. The research results show that the Bi-LSTM model with Word2Vec has the highest performance compared to other models, such as Logistic Regression (LR), Multinomial Naïve Bayes (MNB), K-Nearest Neighbors (KNN), and Random Forest (RF), with an average value -average accuracy, precision, recall and F1-Score reaches 98%. The results of this research contribute to the assessment of public opinion and categorization of topics currently being widely discussed by the public on YouTube as an open source of intelligence information to support defense strategies.