New Methods for Detecting Deceptive Product Reviews

Abstract

With the explosion of online shopping sites, there has been a proliferation of businesses offering to post positive product reviews in exchange for payment. The presence of these deceptive reviews transforms a product's online reviews from a source of candid customer feedback into a forum for surreptitious advertising. Given this trend, there is a demand for effective methods of detecting fake reviews and/or products for which reviews have been purchased. These tools could enable companies to identify and crack down on reviewers and vendors engaging in these deceptive practices. In this work, we present new statistical methods for the detection of deceptive product reviews, focusing specifically on the detection of products for which a high percentage of the positive reviews are fake. Using a new hand-built corpus of online product reviews, we show that there is a correlation between textual features of a product's 5-star reviews and product metadata features that have been demonstrated to suggest the presence of deceptive product reviews, such as star rating distribution. Drawing from the literature on advertising language, the literature on deception, and a series of human performance experiments, we then propose a model that makes use of discourse structure to classify individual reviews as suspicious or trustworthy. While there have been numerous studies in fake review detection, there has been relatively little work in identifying deception on the product level. Our work is also novel in that it draws connections between fake review detection and advertising, a domain in which there has been little computational linguistics research. Furthermore, our work contrasts with much of the previous work in fake review detection in that we explore methods for developing a gold standard in a corpus of voluntarily posted product reviews, as opposed to soliciting deceptive reviews for the purpose of our research.