Using Text Classification Models to Determine Review Helpfulness

Abstract

In an increasingly online world, it has become progressively difficult to manually parse all the reviews that we interact with on a daily basis. From products to even scholarly work, the amount of data available has accumulated exponentially. Businesses like Amazon and Yelp have attempted to provide insight by adding voting and reporting systems that reward reviews that actually help people, such as through a "helpful" tick box or by voting up or down on reviews. However, this still requires manual effort. Past research has looked into judging review sentiment in order to more succinctly inform customers, but little work has been done in judging the "helpfulness" of a review. My work looks to improve the process of judging how helpful a review is using automated text analysis and classification. I compare a bag of words representation and word vector representation, and train these representations on a variety of machine learning models to attempt to classify a review as helpful or not helpful. The goal of this work is to gather insight into what makes a given review useful to the reader, and in testing a multitude of models and representations, find the most efficient and accurate method of classification. Using these automated techniques will allow for faster and more efficient parsing of the multitudes of review data that the average person is exposed to on a frequent basis.