Data Representation in Machine Learning-Based Sentiment Analysis of Customer Reviews

摘要

In this paper, we consider the problem of extracting opinions from natural language texts, which is one of the tasks of sentiment analysis. We provide an overview of existing approaches to sentiment analysis including supervised (Naive Bayes, maximum entropy, and SVM) and unsupervised machine learning methods. We apply three supervised learning methods-Naive Bayes, KNN, and a method based on the Jac-card index - to the dataset of Internet user reviews about cars and report the results. When learning a user opinion on a specific feature of a car such as speed or comfort, it turns out that training on full unprocessed reviews decreases the classification accuracy. We experiment with different approaches to preprocessing reviews in order to obtain representations that are relevant for the feature one wants to learn and show the effect of each representation on the accuracy of classification.