Public complaint tweet data feature analysis for sentiment classification

Abstract

The perception of the public regarding a government's performance significantly impacts a city's advancement. This research involved analyzing complaint tweets from Jambi City residents directed at the government to gauge sentiment. In the testing phase, 500 Twitter accounts were examined to categorize sentiment as positive, negative, or neutral. Training data was prepared by extracting tokens through feature selection techniques such as information gain (IG) and mutual information (MI). For testing, all tokens are entered as data in the input layer in the recurrent neural network (RNN). From the tests carried out, the average use of feature selection can achieve a good value compared to no feature selection. But more specifically the use of IG produces better accuracy compared to the use of MI. From the research conducted, Twitter data is classified using a RNN and several tests by adding feature selection to produce differences. The results are proven to improve classification performance. With a recall value of 92.243%, it shows the system's success rate in sentiment classification and a precision of 92% indicates a level of accuracy that is sufficient to support the government's sentiment assessment.