AI can now produce passable parody song lyrics
The system is called Weird AI Yankovic. Really.
The coronavirus pandemic has caused many outbursts of creativity, but few have the potential to be as meaningful as Mark O. Riedl’s lockdown project. The academic at the Georgia Institute of Technology’s school of interactive computing has spent the last several long months producing an artificial intelligence system that’s able to produce lyrics for parody songs, similar to those produced by Weird Al Yankovic.
The system – which Riedl has christened Weird AI (with a capital “I) Yankovic – started out as a joke and a way to pass the time. “Normally I do a project over the summer where I keep my skills up to date, but I largely did it as a distraction from everything going on,” says Riedl, who unsurprisingly has outed himself as “a big fan” of Weird Al.
Reidl took two off-the-shelf neural networks, GPT-2 and ExcelNet, and combined them to produce rhyming lyrics for songs matching melodies that follow some simple rules. “I didn’t touch them at all,” he admits. “I didn’t do any additional training or fiddling with the internals. I just said: ‘I’m going to use these things to generate a line’, and I’m going to write some extra functions on top of that to control what’s coming out of the network so it fits the right rhyme scheme and syllable structure.”
The results can be surprisingly good: a parody of Michael Jackson’s ‘Beat It’ produced by the neural networks, tells the story of the consumption of junk food.
“The best part is that each taco contains a small/-to-medium-sized piece of sliced chicken Nepal/I don’t think the food in question lasted a while/I promise, just promise”, the neural network ‘sings’.
Unlike several examples of AI-generated text, including The Guardian’s recent trials at getting GPT-3, a successor to one of the neural networks Reidl used for Weird AI Yankovic, to write an op-ed for the paper, the results of Reidl’s experiment weren’t cut and pasted together by a human afterward. (The Guardian’s extensive editor’s note about how the essay came about, producing eight versions that a human editor compiled together into one coherent argument, was lambasted on social media when revealed.)
The academic set several constraints for the AIs, which ran freely in their pre-programmed way. He put constraints on the fact that the AI had to use a rhyming dictionary in order to produce the parodic lyrics, and limited the number of syllables each line could contain so they would fit the melody of the song. “These neural networks don’t know they’re producing lyrics, so it generates 10 or 15 different possible things until it finds one that meets those two constraints,” he says.
What Riedl discovered was surprising: a playful project has become an academically-rigorous project that is part of a workshop at NeurIPS, one of the world’s most prestigious neural processing academic conferences, and that neural networks are more advanced than we think.
“Even some of what are now previous generation neural networks are rather quite good at the things they need to do,” he explains. “We’re at the point where research has given us off-the-shelf technologies that can be applied to interesting, non-trivial problems.” They’re able to do more than the menial tasks most people assume they’re best suited for, such as developing chatbots, answering simplistic questions or editing barely-legible copy for online ads.
“Creativity comes from constraints,” says Riedl. “As soon as you provide constraints, it says there’s an external force forcing you to meet some set of expectations, and we can apply those constraints to these neural networks. That’s when it triggers our perceptions as humans that something humorous is happening here.”
There were some hiccups in the process: Riedl admits in the paper that he programmed the rhyming dictionary used by the neural network to rhyme words ending with “r” and those with vowels, which he says “gave my system a bit of a British sound, which is fine by me because I listen to a lot of pop music from the United Kingdom”.
But the AI doesn’t just create work that would pass for Weird Al: Riedl tried the system on classics like Simon & Garfunkel’s ‘Sound of Silence’. Giving it the rhyme, metre, and a prompt of “Hello darkness, my old friend”, the neural network produced something the singers could plausibly present:
“But darkness never explained,” begins the AI, in place of “Hello darkness, my old friend”. “But once gained, darkness never gained./My shadow has vanished from/This plane of existence. See ‘em/On the horizon and look at darkness in/The between./See ‘em, in the silence.”
Riedl has pressed pause on the project for the time being as his real job and life have taken precedence, but the work he’s done with it during lockdown has fed into his ‘proper’ research. “The lessons I learned about constraints and you can have these black-box systems and wrap constraints around them led to some interesting insights an assistant of mine in my lab is going to pursue,” he says.