Microsoft’s AI can create decent Chinese poetry from just a few images

Emma Sims August 13, 2018

Microsoft’s AI wields the cool but questionably useful skill of being able to conjure Chinese poetry from a couple of images. The company’s Xiaolce chatbot is the system in question – a Chinese language conversational AI.

Microsoft’s AI can create decent Chinese poetry from just a few images

Conversation isn’t all it’s able to do, with the AI creating better poetry than I managed in my university creative writing module. And, while Xiaolce isn’t the first AI to create poetry – just look at this one recreating Donald Trump’s famously erudite speeches – it is the first that’s able interpret images as poetry.

Obviously we’re not literary critics, but it has to be said that the verse conjured by Microsoft’s Xiaolce AI isn’t atrocious. In fact, it’s capable of some elegantly sparse lines: “Wings hold rocks and water lightly / in the loneliness / Stroll the empty / The land becomes soft”. Machine learning nails enjambment – now that’s a line I never thought I’d write.

microsoft_ai

Image: Microsoft

The process is (relatively) simple: developers input rules, then engineer a neural network that splits itself into two sides – one that creates a poem, and the other assesses that poem. If the latter decides the former’s output is good enough for human consumption, scientists then assess the poetry themselves. If they’re sufficiently pleased, then great. If not, the parameters are adjusted and it’s back to the drawing board for the AI, whose next creative output will likely be of a higher standard.

While this all sounds vaguely gimmicky, the bigger picture is important; in teaching AI to generate text descriptions – or in this case, poetry – for an image, we’re teaching AI to interpret the world the same way we do. And teaching AI to see the world through our eyes catapults machine learning closer to human capabilities (and beyond) than ever.

Currently, neural networks and computer vision are about as capable as the average human child of making inferences from images. But Xiaolce marks a notable, if humourous, leap forward in the capacity of AI to undertake such tasks. The end goal is to improve the machine learning model until its work is indistinguishable from the products of its human counterparts.

That being said, Microsoft has something of a mixed record in its AI endeavours; lest we forget Tay, the firm’s Twitter AI that turned racist in just 24 hours, before retiring from active duty with a sassy: “c u soon humans need sleep now so many conversations today thx <3”.