Automated Sentiment Analysis Of Product Reviews Using Machine Learning Techniques

Abstract

Over the last few years, the rapid growth of e-commerce has increased the significance of customer reviews in influencing decisions about what to buy. Reviews not only help potential customers make informed decisions, but they also give businesses invaluable insight and lend your business credibility and trust. For this research, we conducted an analysis of a dataset containing reviews from Amazon that covers a wide range of product categories. This SA project's goal is to use a BERT model to categorize reviews from Amazon that are either positive,  negative, or neutral. Important steps in an investigation include preprocessing, gathering, dividing, training, and assessing data. Amazon-sold items, written in JSON, were amalgamated as devices which included Video security systems, tablets, laptops, televisions, and cell phones. For data preparing, we applied lowercasing, stop word removing, punctuation removal, contraction removal, tokenization and tagging of parts of speech. An opinion lexicon was used to generate sentiment scores, and word embeddings were used of numerical vectorization. A BERT model was then implemented for sentiment categorization using a training a cross-entropy loss function in the PyTorch framework. We used evaluation criteria for performance measurement including recall, F1-score precision, and accuracy. In October 2018, BERT was released, and it outperformed logistic regression and decision tree model, thanks to its effectiveness in capturing long-term dependencies of text. These findings have significant real-world implications and will instill confidence in their strategy by  enabling e-commerce sites to make informed choices and solidify their service provisions.