Abstract

Information on the internet is very diverse, yet so many opinions make it difficult for other users to get information. Sentiment analysis is the process of analyzing or identifying a person's opinion on a particular subject or product that falls into positive, negative, or neutral categories. Aspect-level sentiment analysis shows better performance than document-level and sentence-level. This study aims to determine the accuracy performance of feature optimization using Information Gain with word normalization in aspect-based sentiment analysis. Therefore, this research uses Support Vector Machine as a classification algorithm with a polynomial kernel as well as non-standard word repair using Slang Word and Abbreviation (SS) dictionary followed by Spelling Corrector using Peter Norvig algorithm with additional Information Gain feature selection to optimize the number of features. Based on the test results that have been carried out using K-fold Cross Validation and Confusion Matrix on test data, the accuracy results vary according to the testing process flow. The best accuracy is obtained from the use of Information Gain without Peter Norvig's word normalization resulting in an average accuracy of 83%. Errors are often found when changing words. This error occurs because the word that can be changed can only correct one wrong letter.