Multitask Aspect-Based Sentiment Analysis of Indonesian Tweets on Mandalika Circuit using CNN and IndoBERTweet Embeddings

Abstract

This study proposes a multitask Aspect-Based Sentiment Analysis (ABSA) model for Indonesian tweets related to the Mandalika Circuit, using IndoBERTweet embeddings and Convolutional Neural Networks (CNN). The model simultaneously predicts aspect categories and sentiment polarities. Two experimental setups were evaluated: one using raw tweets (Scenario 1) and another with preprocessed text (Scenario 2). The results show that Scenario 1 consistently outperforms Scenario 2, highlighting the ability of IndoBERTweet to handle informal tweet structures without requiring standard text cleaning. A paired t-test was conducted to evaluate statistical differences in performance between scenarios. While Scenario 1 showed higher average F1-scores, the p-value (0.7178) suggests no statistically significant improvement across all classes. Further analysis reveals that certain classes, primarily neutral and positive sentiments, tend to perform worse than negative sentiments. Data augmentation was shown to improve recall and help the model handle underrepresented classes, particularly for “Ekonomi-Negative” and “Fasilitas-Negative” labels. The study highlights the importance of preserving informal language structures and utilizing data augmentation to enhance ABSA performance on real-world tweet data.