<![CDATA[To improve the accuracy of sentiment classification in CapCut app reviews, this study tested a hybrid model built from a combination of RoBERTa and Word2Vec. A total of 5,000 reviews from the Google Play Store were used as a dataset, which was then processed through data cleaning, tokenization, and stopword removal stages. Next, the EDA oversampling technique was used to address the issue of class distribution imbalance. The proposed model architecture works by combining the concatenation of vector features from Word2Vec for local word meaning representation and RoBERTa for overall sentence context understanding. Model evaluation showed an accuracy of 80%, a higher result compared to the 79% accuracy obtained by the single RoBERTa baseline model. This study concludes that combining contextual and semantic feature representations effectively results in better sentiment classification performance.]]>
Copyrights © 2025
