Abstract

This research was conducted to apply the Bidirectional Encoder Representation from Transformer (BERT) method to multi-aspect sentiment analysis of film reviews. The review data was obtained using the scraping method. The dataset used consists of 1899 data to 3245 data having a positive sentiment, 4825 data with a neutral sentiment, and 1424 data with a negative sentiment. The proposed approach includes the aspects such as acting, plot, cast, animation, and music. The aspect with the most positive sentiment is music with a total of 631 data, the neutral sentiment is found in the animation aspect with a total of 1146, and the negative sentiment is found in the plot aspect with a total of 362. The dataset used went through cleaning data, including case folding and removing HTML tags, punctuation, numbers, and special characters. This research uses the BERTBASE-UNCASE model with four experiments using hyperparameters max_epoch 10, batch size 16, and learning rates of 1e-4, 5e-5, 3e-5, and 2e-5. The research results show that, from all experiments, the best accuracy value is achieved in the third experiment using a learning rate of 3e-5, which is 82,32%. Meanwhile, the best precision, recall, and f1-score values for the “animation” aspect are 86%, 85%, and 85%.