ASPECT BASED SENTIMENT ANALYSIS OPTIMIZATION USING VADER AND SMOTE FOR MOBILE GAME REVIEWS
HAYYU ILHAM WICAKSONO, Dr. Azhari, MT.; Dr. Sigit Priyanta, S.Si., M. Kom.
2023 | Tesis | S2 Ilmu Komputer
Automated sentiment analysis encounters challenges in recognizing sentiment in long texts, primarily due to multipolarity and imbalanced training sentiment datasets. To overcome these obstacles and enhance performance, a proposed model based on aspect-based sentiment analysis employs Support Vector Machine (SVM) combination method with sentiment segmentation VADER and SMOTE oversampling.
This model deploys in a sequential process, including aspect extraction, preprocessing, sentiment segmentation, feature extraction, SMOTE oversampling, and SVM classification. The research utilizes a dataset from Google Play, scraping reviews of five popular mobile games. The proposed model predicts sentiment for each review across aspects like visual, gameplay, audio, and characters.
Evaluation results reveal superior performance compared to base and alternative methods. SVM achieves accuracies of 86.4% (visual), 82.2% (gameplay), 76.1% (audio), and 82.9% (characters). In contrast, Naive Bayes with accuracies of 72.4% (visual), 68.8% (gameplay), 68% (audio), and 67.4% (characters). On average, the proposed model improves accuracy by 8.9% with the VADER method. In conclusion, the proposed model excels in deep sentiment recognition, outperforming the original and alternative methods. By addressing multipolarity and imbalanced datasets, it achieves higher accuracy across various sentiment aspects. Aspect-based sentiment analysis combined with synthetic oversampling, and VADER, shows potential for advancing sentiment analysis in lengthy texts, particularly in mobile game reviews.
Kata Kunci : Natural language processing, aspect based sentiment analysis, fine-grained, VADER, SMOTE, mobile game reviews