Klasifikasi Emosi Netizen Menggunakan IndoBERT pada Media Sosial X Berbahasa Indonesia

Abstract

Social media users or what we often call Netizens often express their expressions and thoughts on social media as a place to "confide" without any limits. Collecting Expressions and opinions from netizens poured out through a tweet can contain various types of emotions or feelings, according to Shaver and Murdaya the structure of emotional vocabulary has five classes, namely love, anger, sadness, happiness, worry/fear. Previous research using the social media dataset X with the LSTM-RNN (Recurrent Neural Network) algorithm, which was initially labeled into five classes of emotions, namely love, anger, sadness, joy, fear, into two classes of depression and normal with an accuracy value of 86%, precision 86%, recall 86% and F1-score 86%, while researchers who used the same dataset without changing the initial labeling used the CNN algorithm with an accuracy value of 89.8%, precision 90.1%, recall 90.3% and F1-score 90.2%.. With the difference in the use of preprocessing at the stopwords and stemming stages, it can affect the evaluation results with the best results without going through the stopword and stemming stages with the acquisition of accuracy, precision, recall, and f-1 score sequentially for IndoBERT obtaining 88.8%, 89.4%, 88.4%, 88.8%. While IndoBERTweet obtained 92.5%, 92.7%, 92.2%, 92.4%. Therefore, this study uses a dataset originating from Twitter, which will be used to train the IndoBERT model so that with the dataset used in previous studies, satisfactory results can be obtained using different models, and resulting in the conclusion that the IndoBERTweet model is superior to IndoBERT, while for each emotion classification the IndoBERT model is only superior in the emotions of fear and love in the Recall section 79.36% compared to 75.26% and 86.46% compared to 78.30%. while overall IndoBERTweet is still superior.