A Comparative Study of BERT and Traditional Machine Learning Models in E-commerce Review Classification

This paper presents a comparative analysis of sentiment classification methods applied to a dataset of Amazon reviews. Three methods are evaluated: the pre-trained BERT model, Naive Bayes with TF-IDF, and the lexicon-based TextBlob. The dataset consists of 20,350 samples, with 5,869 positive and 14,481 negative reviews, derived by categorizing ratings of 4 and 5 as positive, and 1 and 2 as negative. Results show that Naive Bayes achieved the highest accuracy (86.62%), excelling in negative sentiment classification, while BERT demonstrated a balanced performance with 84.81% accuracy and superior contextual understanding. TextBlob, with an accuracy of 69.79%, struggled to classify positive sentiments due to its reliance on predefined lexicons. The study highlights the strengths and trade-offs of each method, emphasizing BERT’s potential for achieving state-of-the-art results in sentiment analysis, albeit with higher computational costs.