Enhancing Aspect-Based Sentiment Analysis for Radio Station Public Opinion: Evaluating Preprocessing Strategies and Imbalanced Data Handling

Radio has played a crucial role in disseminating information, significantly contributing to spreading nationalism and values of struggle in Indonesia. In the digital era, radio stations face new challenges as they integrate broadcasting with new media platforms to maintain audience engagement. This study aims to develop an aspect-based sentiment analysis system for public opinions on a radio station, focusing on evaluating preprocessing strategies and imbalanced data handling methods to enhance model performance. The research addresses the significant data imbalance observed in sentiment and aspect classes, which impacts classification model performance. Various data balancing methods, including Random Oversampling, SMOTE, Undersampling, Hybrid Methods (SMOTE + Tomek Links), and Penalizing Algorithms (SVM with Class Weights), were evaluated for their effectiveness in improving classification metrics such as accuracy, precision, recall, and F1-score. Our findings reveal that the Penalizing Algorithm consistently achieved the highest accuracy and recall across most aspects, with the highest accuracy of 0.870 observed in the infrastructure aspect, while Random Oversampling and SMOTE reached accuracies of 0.861 and 0.855, respectively. The combination of SMOTE and Tomek Links yielded similar results to SMOTE alone. Additionally, stemming and its combination with stopword removal consistently improved classification accuracy across various aspects, with the highest accuracy of 0.892 observed in the program and infrastructure aspects. These results underscore the importance of data balancing and appropriate preprocessing in developing robust aspect-based sentiment analysis models, particularly for imbalanced datasets, to ensure accurate and reliable public opinion analysis.