A Neural Network Has Generated Hilariously Insane Candy Heart Phrases


Janelle Shane

Looking for a way to impress that tech-savvy special someone in your life? This Valentine’s day, introduce a little machine learning into your romance. A new project by research scientist Janelle Shane offers a novel take on the candy heart, replacing standard Hallmark phrases with messages generated by a neural network.

“It is so charming to watch this algorithm try to charm humans, partly because its only motivation is to do a good job, and partly because a ‘good job’ is so arbitrary when you think about it,” Shane told Inverse in an email.

“It doesn’t know what ‘sweet’ means — or that when you change one of the e’s to an a, it becomes no longer so appealing. Why love bugs and not love bogs?”

Following a post about her project Friday, the Twittersphere went off about the candy hearts’ phrases, which give new meaning to the term “sweet nothings.” At the time of writing, her post had been liked more than 2,400 times and retweeted over 1,500 times.

This isn’t Shane’s first foray into the world of computational humor. In the past, she has trained neural networks to write pick-up lines, draft recipes, and invent Pokemon names.

You had me at hole.

Janelle Shane

Shane’s project has exemplified upon one of the defining attributes of generative neural nets. Because they operate at the character level, they don’t deal with the meaning of words and phrases. Rather, they notice patterns in a dataset (in this case, an assortment of candy heart phrases), and arrange letters according to those patterns. This leads neural networks to come up with phrases that wouldn’t occur to a human, while still sounding weirdly real.

The results are comical. Tired candy heart stalwarts like, “BE MINE” were replaced by nonsensical phrases like, “STANK LOVE,” and “LOVE 2000 HOGS YEA.” The generated phrases exist somewhere in the space between doggo-speak and childlike gibberish, so they fit right in with the internet joke zeitgeist. Slight misspellings, like “doge” and “smol” are a staple of internet humor, and nobody does inane typos quite like a neural network.

“Humans are weird,” Shane said. “And the algorithm points that out to us unavoidably.”

You may not swoon over the neural network’s romantic messages, but were you really going to be swept off your feet by a candy that said, “CALL ME?” Besides, beauty is in the eye of the beholder.

“Frankly, I would find ‘ALL HOVER’ to be very charming,” Shane said.

To read more about Shane’s projects, check out her blog.

Related Tags