Tag Archives: computer-generated wine reviews

“Crisp and fresh:” AI wine writing strikes again

ai wine writingWill AI wine writing eventually make wine tasting irrelevant?

This is a mineral-driven wine that’s crisp and fresh, with a flinty edge. It is very tangy, with zesty citrus, giving a bright character. It needs time to mature, so wait until late vintage.

That’s a review of one of my favorite $10 wines – the Chateau Bonnet Blanc, a white French Bordeaux. But I didn’t write it, and neither, technically, did any other human wine critic.

Instead, it was written by an artificial intelligence – the Wine Review Generator created by long-time wine industry executive Michael Brill. Brill, who also does tech, software, and AI consulting, wanted to find out if he could “teach” a machine to write tasting notes.

And, for the most part, that’s what he did.

Brill left a comment about last week’s blog post about the future of AI wine writing. That led to our phone conversation this week, where Brill said improved technology has made it possible to create the Chateau Bonnet review with a minimal amount of human programming. All you need, he said, is a database of wine terms, wine regions, grape varieties, and so forth. That information, combined with advances in neural network research that have helped scientists better understand how to program machines to “think,” led to the review software and to the Bonnet review.

In this, Brill said, a machine’s ability to “write” longer and more coherent sentences has improved tremendously. Before, he explained, an AI story might be half readable and half nonsense, and the most it could create was a 10-word sentence. Today, those numbers are 90 and 10 percent, and it can write a readable 10-sentence paragraph.

How the machine does this, needless to say, is incredibly complicated. It makes predictions about what comes next in a sentence based on the words that came before, a process that is much more like writing than previous AI efforts; those were more like filling in a template. Here, the AI has “learned” that a mineral-driven wine is crisp and fresh, and not oaky and flabby, so it picks the former phrase to follow mineral-driven instead of the latter.

Which is why the Chateau Bonnet Blanc effort is not a bad tasting note. It’s mostly accurate (save for the bit about aging) and it conforms to the rules of grammar and the sensibilities of wine. That the machine wrote the review without tasting the wine is impressive, and knowing only the cost and some characteristics, is impressive. And more than a little spooky.

And not just because an AI is cheaper to hire than I am. Brill said advances in machine writing could eventually make product reviews useless. Some of that happens today on Amazon, where it’s not uncommon to see badly written AI reviews praising a product. But the situation could get even worse as AI writing improves.

A top-notch AI could flood Amazon with machine-generated positive (or even negative) reviews, with the resulting effect on sales. Or it might be possible for one restaurant to force another out of business with an AI-written campaign on Yelp.

And who would know the difference?

do-it-yourself-wine review

Coming soon to the WC: Computer-generated wine reviews

computer-generated wine reviews

“Damn. … where is that neural network when I need it?”

A little Python, some neural network command line work, and the blog can start posting computer-generated wine reviews

Sooner rather than later, I’m going to post computer-generated wine reviews on the blog. Thanks to the Lifehacker website, all I need are some basic Python programing skills. Or, even better, find a Python-savvy volunteer from among the blog’s sophisticated and erudite audience who wants to help “write” them.

We’ve followed the advances in artificial intelligence that makes these reviews possible for several years. Barbara Ehrenreich, writing in the New York Times, said that “the business of book reviewing could itself be automated and possibly improved by computers.” And writing one kind of review isn’t all that different from writing another kind.

The point here? That wine has become so mechanized and so predictable that we can probably get acceptable Winestream Media-style reviews from an artificial intelligence. It’s probably even possible to teach the machine to give scores – a delicious irony that is reason enough to make this work.

Lifehacker’s Beth Skawrecki writes that machine-written reviews are more possible than ever thanks to advances in neural networks. A neural network is “a type of [artificial intelligence] modeled on the network-like nature of our own brains. You train a neural network by giving it input: recipes, for example. The network strengthens some of the connections between its neurons (imitation brain cells) more than others as it learns. The idea is that it’s figuring out the rules of how the input works: which letters tend to follow others, for example. Once the network is trained, you can ask it to generate its own output, or to give it a partial input and ask it to fill in the rest.”

For our purposes, we would tell the computer what chardonnay is supposed to taste like, where the grapes were grown, information about the vintage and the winemaker’s style, and the price. Then, we can “teach” it to interpret that information to write the review – that chardonnay from California is different in certain ways from chardonnay from France, for example.

Now, to brush up on my Python.

More about computer-generated wine reviews:
Winecast 30: Arty, the first artificial intelligence wine writer
Do we really need wine writers?

Winecast 30: Arty, the first artificial intelligence wine writer

artificial intelligence

Arty, the first artificial intelligence wine writer

“Wine drinkers want to be reassured that what they are drinking is worth what they paid for it. That’s the goal of the post-modern wine business and premiumization, and I was created to do that.”

Computer-generated wine writing has arrived, if this interview is any indication. I talked to Arty, the world’s first artificial intelligence wine writer, for this edition of the podcast.

Arty and I discussed why he was created, his goal as a critic — “We’ll always need quality wine writing, human or otherwise. But I think I can offer consumers wine criticism that they can’t get anywhere else” — and why his kind may be the future.

Click here to download or stream the podcast, which is about 3 minutes long and takes up 2 1/2 megabytes. The sound quality is almost excellent.

A tip of the Curmudgeon’s fedora to the Mary Text to Speech system, which made it possible to create this interview. Maybe what I’m joking about is more possible than we know.

Let the computer write the wine reviews

computer-generated wine reviews

How am I supposed to know if there’s too much oak?

Could artificial intelligence make writers obsolete? Because I’m not the only one who wonders. Barbara Ehrenreich, writing in the New York Times, firmly believes that “the business of book reviewing could itself be automated and possibly improved by computers.”

So why not wine writing — computer-generated wine reviews?

This would solve any number of problems, not the least of which is that winemakers wouldn’t have to deal with people like me. I had a brief email discussion recently with an annoyed producer who insisted that her wines didn’t taste the way I described them; she certainly would have been better off with WineNet than what Ehrenreich calls a “wet, carbon-based thinking apparatus” with self-awareness and a sense of obligation to its readers.

The last time I wrote about this, a company called Narrative Science had made significant inroads in taking disparate facts and turning them into a readable narrative. Unfortunately, it seems to have veered elsewhere, developing a product that “creates new revenue opportunities by transforming data into engaging content that can be productized and monetized.” This approach has little to do with writing, since there is money involved.

Still, much work has been done. TechCrunch reported last month that robot writers are all the rage in Silicon Valley, while a data scientist named Tony Fischetti has written that Markov chains can be used to simulate what he calls the “exercise in pretentiousness” that is a wine review. The concept of a Markov chain, which deals with probability, is far beyond my math skills, but Fischetti used 9,000 reviews from the Wine Spectator to write a program that came up with tasting notes that are no worse than most, including: “Quite rich, but stopping short of opulent, this white sports peach and apricot, yet a little in finesse” and “this stylish Australian Cabernet is dark, deep and complex, ending with a polished mouthful of spicy fruit and plenty of personality.”

Meanwhile, a wine producer in France, using N-Gram analysis (also beyond my math skills, but apparently related to word order) also thinks it’s possible to generate wine reviews without a wine writer. Both approaches seem to jive with what I wrote last time, that an artificial intelligence, working with a wine term database and the proper algorithm, could scrape together effective reviews. Probably even scores, too.

I just hope, if and when this puts me out of business, that someone will remember that I saw it coming. Maybe I can monetize the blog that way.

Image courtesy of Techbrarian, using a Creative Commons license