Aspect-Based Sentiment Analysis on Webtoon Reviews Using Ensemble Learning

The rapid growth of online comic platforms such as LINE Webtoon has produced large volumes of user comments that reflect diverse opinions on storytelling elements. Analyzing these comments provides meaningful insights into reader perceptions of aspects such as plot, characters, and visuals. This study proposes an Aspect-Based Sentiment Analysis (ABSA) framework using ensemble learning to classify sentiment in Indonesian Webtoon reviews. The research follows an experimental quantitative methodology consisting of data collection, text preprocessing, manual annotation of aspects and sentiments, feature extraction using Term Frequency–Inverse Document Frequency (TF-IDF), ensemble model training with Random Forest and XGBoost, and performance evaluation. A total of 1,010 annotated comments were used, covering three aspects—plot, character and visual—and three sentiment categories: negative, neutral, and positive.

The results demonstrate that incorporating aspect information enhances sentiment classification performance. While the tuned XGBoost model using TF-IDF features achieved an accuracy of 62.38% in the text-only scenario, the best performance was obtained by the tuned Random Forest ABSA model, which combined TF-IDF and One-Hot Encoded aspect features and achieved an accuracy of 65.84% with a weighted F1-score of 0.65. Class-level analysis shows that neutral comments are the easiest to classify, while negative sentiment remains the most challenging due to informal and context-dependent expressions. A 10-fold cross-validation yielded a mean accuracy of 62.38% with a standard deviation of 0.057, indicating stable generalization. These findings highlight the effectiveness of aspect-enhanced ensemble learning for sentiment analysis in Indonesian Webtoon reviews.

Keywords: Aspect-Based Sentiment Analysis, Ensemble Learning, Random Forest, XGBoost, Webtoon Reviews