The rapid growth of e-commerce applications in Indonesia has resulted in a large volume of user reviews. The review contains important information that can be used to understand user satisfaction, complaints, and needs. Therefore, sentiment analysis of e-commerce app reviews is important to support future decision-making. This study aims to explore and compare the performance of the Transformer model and various word embedding methods in analyzing the sentiment of reviews of Indonesian e-commerce applications. The methods used involve extracting review data from the Google Play Store, text preprocessing, and text representation using Word2Vec, FastText, and IndoBERT. Next, this combination of embedding was tested using the Gradient Boosting Classifier as a prediction model. The evaluation was carried out by comparing the accuracy, precision, recall, F1-score, as well as the visualization of the confusion matrix and word cloud for each model. The results of the tests that have been carried out show that all three models have a fairly good ability to recognize positive reviews, with the highest accuracy score of 88% achieved by Word2Vec and FastText. While IndoBERT produces a lower accuracy value of 86%, IndoBERT shows a better balance in recall values and f1-scores for minority classes compared to Word2Vec and FastText. In conclusion, the application of the IndoBERT-based Transformer model is more effective in capturing the context and meaning of sentiment in Indonesian-language e-commerce reviews. These findings are expected to be a reference for the development of a more accurate sentiment analysis system for e-commerce applications in Indonesia.
Kadarsih et al. (Thu,) studied this question.