Product Rating Prediction with Online Reviews Using Support Vector Machines

Abstract

More and more shopping websites allow customers to post online reviews on products, allowing customers to share opinions and information on specific products. Reviews can be expressed in text, ratings, or both. Text-based reviews give detailed information on a product while ratings can be quickly understood. Numerical ratings are especially important when screen size is limited. However, not all customers assign ratings to text reviews, and some text-based reviews are inconsistent with their corresponding numerical ratings. In this paper, we outline a method of mapping text-based reviews to numerical ratings using an SVM classifier. Three linguistic feature types are employed in our SVM-based classifier. Given the very large number of product reviews, only features that can be efficiently extracted are employed. Since it is difficult for customers to distinguish adjacent ratings (e.g. 4 and 5), we have adopted relaxed criteria for evaluating our system precision. According to experimental results, our method achieves a precision of over 76.6% using the relaxed criteria, which is sufficient to automatically annotate text reviews with numerical ratings.