Sentiment Analysis for Movie Review in Bahasa Indonesia Using BERT
Dwi Fimoza, Amalia Amalia, T. Henny Febriana Harumy
Abstract
This study aims to analyze the sentiment in Indonesia Language towards the Gundala movie reviews on YouTube. However, sentiment analysis on YouTube comments are varying from positive, negative, and neutral comments which requires some automation in terms of classifying comments based on the polarity of sentiment. Sentiment analysis using traditional machine learning algorithms such as Naïve Bayes, SVM, etc cannot understand the context of comments in depth about the semantic of words because it only learns the given patters such as the frequency of occurrence of words. We need a transfer learning approach such as BERT (Bidirectional Encoder Representations from Transformers) which produces a bidirectional language model. The dataset used to do sentiment analysis goes through a pre-processing step which consists of case folding, data cleaning, tokenization, stop words removal, stemming, and normalization, using libraries from NLTK and Sastrawi. In this study, the hyperparameters used were 10 epochs, learning rate of 2e-5, and a batch size 16 based on experiments of hyperparameters used in another studies. In sentiment analysis, we will be using a multilingual-cased-model BERT <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BASE</inf> model and it was carried out with three experiments. During this experiment, the accuracy gained in first experiment is 66%, while the second experiment was 68%, and the third experiment was 66%. So, the average accuracy obtained is 66,7%.