How Facebook Learns What Languages You Speak
And how you can change it
Since 2011, Facebook has been rolling out translation features backed by artificial intelligence algorithms. The A.I. reads the post or comment, parses what’s being said, and then translates it into a hopefully natural-sounding translation at the click of the button. Last year, instead of asking users if they wanted to have a post translated, Facebook started automatically showing translations, and asking if users want to see the original.
Now, more and more Facebook posts are automatically translated; 800 million users per month use the translation feature. But behind the scenes, Facebook is running low-level artificial intelligence on all the text uploaded to Facebook, and documenting how you interact with each language.
First Facebook has to recognize the language of every post before it can translate. From there, based on the posts that you like and the languages that you post or comment in, Facebook quantifies how much you speak other languages. From this, it can show those who speak multiple languages untranslated posts in all of them.
“The ingredients in deciding whether [users] see “See Translation” or auto-translate is based on two things: One, do they speak that language, and two, did we get the translation right?” says Joaquin Quiñonero Candela, Facebook’s head of Applied Machine Learning. “It’s all based on interactions that you’ve had that we’re able to observe.”
But if you want to manually tell Facebook that you speak a second language, you can do that too. Candela said that a few people have already found this by themselves, but in Language Settings (Settings>Language) users can set their preferred language, which languages they can understand, and which they specifically don’t want automatically translated.
Of course, all this helps the user, but it also helps Facebook understand more about language. Candela says that in a perfect world, he would like to be able to translate everything seamlessly. And although we’re not there yet, the data from 800 million users can’t hurt.