Multi-classification Sentiment Analysis from sentences in Bahasa is a challenging process due to problems in slang, local language combined with many English words. Current state-of-art methods rely on feature extraction using unsupervised treatment. A research to solve this problem was conducted using LSTM and CNN that are capable of learning complex features from the lower level. The objective of this study was to investigate the results of the sentiment analysis based on the extraction of aspects that were carried out with attention models and several deep learning methods. Research data was collected from Zomato comments in Bahasa for any Indonesian restaurants. The data was annotated manually based on four subjects namely place, taste, location, and service. The result of this study showed that Bi-LSTM with attention model and CNN without attention model had the best performance compared to other methods, while CNN without attention model for sentiment analysis using deep learning showed the best accuracy.