| Summary: | YouTube is the most widely used in Indonesia, and it’s reaching 88% of internet users in Indonesia. YouTube’s comments in Indonesian languages produced by users has increased massively, and we can use those datasets to elaborate on the polarization of public opinion on government policies. The main challenge in opinion analysis is preprocessing, especially normalize noise like stop words and slang words. This research aims to contrive several preprocessing model for processing the YouTube commentary dataset, then seeing the effect for the accuracy of the sentiment analysis. The types of preprocessing used include Indonesian text processing standards, deleting stop words and subjects or objects, and changing slang according to the Indonesian Dictionary (KBBI). Four preprocessing scenarios are designed to see the impact of each type of preprocessing toward the accuracy of the model. The investigation uses two features, unigram and combination of unigram-bigram. Count-Vectorizer and TF-IDF-Vectorizer are used to extract valuable features. The experimentation shows the use of unigram better than a combination of unigram and bigram features. The transformation of the slang word to standart word raises the accuracy of the model. Removing the stop words also contributes to increasing accuracy. In conclusion, the combination of preprocessing, which consists of standard preprocessing, stop-words removal, converting of Indonesian slang to common word based on Indonesian Dictionary (KBBI), raises accuracy to almost 3.5% on unigram feature.
|
|---|