Examining Customer Satisfaction Through Transformer-Based Sentiment Analysis for Improving Bilingual E-Commerce Experiences

As e-commerce continues its rapid expansion, accurately gauging and interpreting customer sentiment has become critical for optimizing user satisfaction and predicting purchasing behavior. This study advances feature-level sentiment analysis by examining Chinese e-commerce product reviews, translated into English, to uncover the contextual impact of sentiment-bearing keywords across diverse product categories. The research intends to discover trends, reasons, and broad conclusions associated with the volume, co-occurrence, and contextual use of keywords across various product types. Using sentiment intensity analysis, N-gram modeling, and machine learning, this research attempted to evaluate the performance of these models: ELECTRA, BERT, BiLSTM, Random Forest with XGBoost, and SVC with SGD. Empirical results demonstrate that ELECTRA yields the highest accuracy (98.09%) and F1 score (0.9711), illustrating superior performance in capturing nuanced sentiment cues, with low loss of 0.062. BiLSTM followed with 96.09% accuracy and F1 score of 0.9583. Although BERT performed strongly with an accuracy of 95.28% and an F1 score of 0.9528, it is extremely resource demanding. Ensemble methods like Random Forest with XGBoost and SVC with SGD had moderate performance, finding 82.43%, and 88.84% accuracy, respectively, but had better runtimes. The results supports the advantage of using transformer based models in attempting more sophisticated contextual analysis by uncovering paramount sentiment patterns, product specific issues, and customer needs. Furthermore, by translating Chinese reviews into English, the study highlights the potential for cross-lingual frameworks to enrich global e-commerce strategies. Future research directions include integrating cultural factors, extending analysis to additional languages, and refining computational approaches for resource-limited environments.