New Research Explores the Social Signals in Yelp Reviews

Nov 11, 2014

By Saeideh Bakhshi, David A. Shamma, and Partha Kanuparthy, There are 139 million people that use Yelp every month and, to date, Yelpers have written over 67 million reviews where they voice their opinions, share their stories, and otherwise rate experiences they had in the physical world.  But there’s more to the Yelp story than the reviews and the reviewers.  People on Yelp also log in and express their opinions, not as reviews, but as votes on reviews.  In effect, it’s a higher granularity than a Flickr “favorite” or a Facebook “like,” as Yelpers cast their votes with the distinct sentiments of cool, funny, and/or useful.  These votes are three kinds of “likes”; they are a minimal social signal that many online sites use for communication and recommendation.  The three options that Yelp offers lets one investigate the implied meanings carried by these sentiments more accurately than many other social networks.  But there’s something more here.  In aggregate, a random person on Yelp might carry a running total of votes they have cast, including 469 useful votes, 192 cool votes, and 260 funny votes.  The same could hold true for a venue.  We began to wonder if we could understand something more from these votes; are they indicative of particular emotions? Do the votes represent some fingerprint of a Yelper or of an establishment? We used over 200,000 Yelp reviews available for research to answer these questions.  First, we found the way people vote on reviews, including the sentiment of the text, has a relationship with the tone of the text and the text’s rating, depending on the vote type. The findings of our research, which appear in a paper called “If It Is Funny, It Is Mean: Understanding Social Perceptions of Yelp Online Reviews” and published in this week’s Proceedings of the International ACM Conference on Supporting Groupwork (ACM GROUP 2014), suggest that there is a deeper meaning and engagement associated with the three signals cool, useful, and funny than their labels suggest. While many would be correct in associating the useful and funny votes as representing reviews with the most amount of information or humor they contain, these signals are actually a proxy for negativity in reviews. A cool vote is more ambiguous in its meaning, but clearly associates with more positive reviews. Understanding these votes, or signals, and how they affect ratings can better inform customers as they come across reviews and take them into account for their own purposes; ultimately, they could alter one's perception of a business, for better or worse. image Screen Shot 2014-11-04 at 10.33.06 AM.png

Two reviews, a 4-star funny review and a 1-star funny review; each carries a negative sentiment despite the rating of the review itself.  Reviews marked cool (not pictured) were often related to a high-scored rating.

Beyond useful, cool, and funny, we wanted to know more about the collections of signals.  We began by asking if these votes are actually meaningful and representative signals since many factors, like weather, can impact a review.  This follow-up research, entitled “Understanding Online Reviews: Funny, Cool or Useful?” will appear in the Proceedings of the ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW 2015) in March 2015.  Here, we studied a variety of circumstances from businesses, reviewers, and reviews that might impact votes. In our analysis (modeled using a zero-inflated negative binomial regression), we verified these votes are in fact representative and found that reviews written by members who are active for longer periods of time tend to receive more votes. We also found that readers tend to prefer long and objective reviews.  This research is just the first step toward understanding sentiment and signals in a new way that starts to uncover the social DNA inherent in online communities.