BERT-Based Sentiment Analysis of Turkish e-Commerce Reviews: Star Ratings Versus Text

This study examines sentiment analysis in Turkish e-commerce product reviews by comparing two distinct approaches: classification based on star ratings and textual sentiment using a BERT-based model. Two models were fine-tuned for this purpose: Model 1, trained on numerical star ratings, and Model 2, trained on manually labeled sentiment in review texts, to evaluate their performance in accurately capturing customer sentiment. The results reveal that star ratings often fail to reflect true sentiment, as many users assign high ratings despite expressing negative opinions in the text. Model 1 tended to overclassify reviews as negative, while Model 2, which used direct text sentiment labels, provided a more balanced classification across sentiment categories. Chi-square tests confirmed a statistically significant difference between the predictions of the two models, highlighting the impact of labeling methods on model behavior. Furthermore, our findings reinforce the value of deep learning approaches, particularly transformer-based models like BERT, in processing Turkish-language texts, which pose challenges for traditional dictionary-based methods due to their complex morphology and syntax. From a business perspective, relying solely on star ratings may lead to an inaccurate interpretation of sentiment. Incorporating text-based analysis can offer more precise insights into customer satisfaction. Future research may explore multimodal sentiment analysis by integrating visual or video data and examining how AI-driven sentiment systems influence decision-making processes across different sectors.