Consider the following wine review: “This is a sound Cabernet. It’s very dry and a little thin in blackberry fruit, which accentuates the acidity and tannins. Drink up.”
It reads like the kind of descriptive, evocative writing you’d expect from a wine critic. But there’s a catch: it was actually written by an algorithm.
A group of researchers from Dartmouth University have developed an artificial intelligence algorithm that can write wine (and beer) reviews that are almost indistinguishable from those written by a human critic. It’s not just for giggles — this could help beer and wine producers aggregate reviews and give human reviewers a template to work from, they argued.
“Using artificial intelligence to write and synthesize reviews can create efficiencies on both sides of the marketplace,” Prasad Vana, one of the authors of the study, said in a statement. “The hope is that AI can benefit reviewers facing larger writing workloads and consumers who have to sort through so much content about products.”
Reviews and machine learning
The researchers took on two challenges for this goal. First, writing an algorithm that could write original and human-quality product reviews using a set of product features. The next task was to see whether the algorithm could be adapted to generate new reviews for products from a large number of existing reviews. To narrow their work, they focused only on wine and beer reviews.
As the code couldn’t actually taste the beer or the wine, the researchers trained the algorithm on hundreds of thousands of published wine and beer reviews. And it turns out it was a big success. For example, one of the results read: “Pretty dark for a rosé, and full-bodied, with cherry, raspberry, vanilla and spice flavors. It’s dry with good acidity” — which sounds pretty enticing.
The algorithm first used previous reviews to learn the structure and style of one. Then, to generate its own, it was given the specifics of the wine or the beer, such as the brewery’s name, the style, the alcohol percentage, and price. Based on these, the AI found existing reviews, took commonly used adjectives, and used them for its reviews.
The researchers also tested the performance of the program. They picked one human and one AI-generated review each for 300 wines and human reviews and one AI reviews each for 69 beers. Then they asked a group of human test subjects to read both versions and check whether they could tell which was which. In most cases, they couldn’t.
However, the algorithm has significant limitations. For example, it may not be able to predict the flavor profile of a drink that wasn’t sampled by humans and described by human writers. That’s why the team insists in the paper that code shouldn’t replace professional product reviewers and instead make the tasks of producing and reading the material more efficient.
“It’s interesting to imagine how this could benefit restaurants that cannot afford sommeliers or independent sellers on online platforms who may sell hundreds of products,” Vana said.
The study was published in the International Journal of Research in Marketing.
