Analysis of Labeling and Class-Balancing Effects on Clash of Champions Sentiment Using LSTM and BERT

Advances in digital technology have changed the way people interact and access information, including in education. One educational event that has caught the public's attention is Clash of Champions by Ruangguru, designed to increase young people's interest in learning through an interactively presented competition. The purpose of this study is to use posts on X social media to examine public opinion on the event. Using TweetHarvest, 1,891 tweets were gathered and preprocessed (cleaning, case folding, normalization, tokenization, stopword removal, stemming, and English translation). A total of 12 experimental scenarios were created by combining VADER and TextBlob labeling strategies with class balancing techniques (undersampling and SMOTE), and the LSTM and BERT models were evaluated for each scenario. The best results were achieved by combining VADER, SMOTE, and BERT, yielding an accuracy of 97.73%, with precision, recall, and F1-scores of 98%, 98%, and 96% (positive), 99% (neutral), and 98% (negative), respectively. These findings highlight the efficacy of transformer-based models like BERT in addressing class imbalance and improving sentiment classification. The integration of SMOTE effectively mitigated class imbalance, providing consistent and accurate performance across all sentiment categories.