Yelp random reviews

Titipat Ahakulvisut, Zaw Htet Aung

Abstract

Reviews by fellow customers are an essential part of deciding to which doctor, restaurant, or other general establishment to go to. While it is easy to recognize a good review from its language, it is much harder to understand the reverse relationship: how the rating and other quantities, such as helpfulness, are related to language. In this work, we developed a new technique that can generate reviews using as inputs review stars, business type, coolness, usefulness, and humor level. Our technique is based on a continuous review sentence model using word vectors. We use a Markov chain model where sentence word vectors depend on a previous sentence word vector and the inputs. The model generates a sentence by predicting the next sentence vector and sampling one from a database of sentences weighted inversely proportional to the distance to the prediction. We developed a website and free software that can generate an unbounded number of reviews on the fly based on user inputs (http://yelp.scholarfy.net). Our work promises to help Yelp understand how features of the review relate to their language and how reviews can be randomly generated, helping to tell apart real from fake reviews.