A Sentiment Analysis of Customer Product Review Based on Machine Learning Techniques in E-Commerce (original) (raw)

The use of customer evaluations as a basis for purchasing decisions has grown in importance with the meteoric rise of e-commerce in the last several years. Reviews provide businesses with valuable information and foster trust, in addition to helping potential customers make informed decisions. A comprehensive examination of a dataset including reviews from Amazon that spans many product categories was part of this research. This SA project's primary goal is to categorise Amazon reviews as neutral, negative, or positive using a BERT model. A study involves several key stages, including data preprocessing, collection, splitting, model training, and evaluation. The JSON-formatted reviews of items sold by Amazon were compiled, including devices like mobile phones, laptops, televisions, tablets, and video security systems. Preprocessing steps included lowercasing, removal of stop words, punctuation, contractions, tokenization, and part-of-speech tagging. Sentiment scores were generated using an opinion lexicon, and word embeddings were applied for numerical vectorization. Employing a cross-entropy loss function trained inside the PyTorch framework, a BERT model was used for sentiment categorisation. To measure how well the model performed, we employed evaluation criteria like recall, F1-score, precision, and accuracy. BERT achieved superior results compared to logistic regression and decision tree models, demonstrating its ability to capture long-term dependencies in textual data. Empowering e-commerce platforms to make educated judgements and expand their service offerings, these results have substantial practical consequences and will give them confidence in their strategy.