Opinion Mining and Sentiment Analysis of Reviews for Shopping Websites

Abstract

Before buying a product it is essential to perform a research on its features and functionalities and web is a vast platform to obtain the information regarding it. One of the subset of the obtained information is customer review where we get the first hand user experience of a product. There is a need for a system that summarises the opinions of all the reviews present on the web. Also when we read customer reviews on the web we encounter few difficulties like mismatch of reviews and stars/ratings, review not so specific to the feature, review-research time for the customer being too high or ease of judging the product without the overhead of reading all the reviews is not possible.  Hence to address these issues we are build a website that fetches review of any given product from the internet and display a collective opinion on the products features using machine learning techniques. In this project we discuss the importance of having a generic framework to fetch data from the web and perform sentiment analysis on it and extract the opinions on that specific product and present the opinions based on each subset of that product specification based on user reviews. Major focus was given towards pre-processing the review data, as it is fetched directly from the web. Methods were implemented to clean and store the data from the internet. Encoding methods such as count, tfidf and hash vectorization were used and results were compared. A study was made on the frameworks to work with for supervised and unsupervised data for sentiment analysis. Vader, text blob, lexical analysis were used for unsupervised data and SVM, random forest and gradient boost for supervised. In our work, the Textblob approach performed the best in comparison to other methods for sentiment analysis. To achieve opinion mining POS tagging was used.