Sentiment polarity classification of Turkish product reviews for measuring and summarizing user satisfaction

Abstract

This study aims to classify the polarity of sentiments in customers' product reviews. By exploiting the features of Turkish, the analysis focuses on the customer language characteristics. Initially, negative and positive product reviews are collected from two different web sites, extracted and pre-processed to select the essential features. Additionally, some negativity and strengthening rules that are specific to Turkish are employed to fully reflect the polarity of the reviews. Basically, three types of feature selection methods are used: Binary Method, Frequency Method and Weighted Frequency Method. The analysis of the polarity of the documents is achieved by using three well-known machine learning classifiers, namely Naive Bayes, VotedPerceptron and J48, a decision tree algorithm. Promising accuracy results are achieved, ranging from 84.4% to 96.5% by applying these methods. The best results are obtained from Binary method and Naive Bayes, obtaining 96.5% accuracy. This study is significant since it enables the efficient and automatic summarization of large amounts of reviews on products, which is very precious in today's market to make profitable choices.