New Translation Tool Will Help Facebook Master International Slang
A feature that helps you post updates in multiple languages will also provide data to train software on colloquial speech.
More than one billion people use Facebook every day. If you picked two at random, they most likely wouldn’t speak each other’s language. But the company thinks they should still be able to socialize with each other.
A new feature uses automatic translation software to help people post Facebook updates in multiple languages at the same time. Users viewing a post made that way are shown the version most likely to be readable to them in light of their own past language use and settings.
The new feature is also designed to provide Facebook with valuable new data to help its translation software convert slang and other colloquial language from one language to another. That could improve the company’s automatic translations of short posts and comments from friends not using your language.
Fazil Ayan, who leads work on Facebook’s translation software, says the multilingual-post option that starts rolling out today was inspired by the fact that many people already write posts in multiple languages. But they had to either write multiple versions of a message in one post or create separate posts—strategies that tend to reduce the chance anyone would respond.
When you compose a new Facebook update, the new feature offers the option of clicking to create additional versions in different languages. You can write those from scratch, or edit text that Facebook’s translation software generates from your original post. An earlier version of the tool was offered to people and companies operating public Facebook pages earlier this year.
Ayan thinks the new feature will inspire more people to post in multiple languages, and he also expects it to help Facebook’s software do a better job translating the type of language people tend to use when posting.
The company already analyzes posts written out in multiple languages to instruct its software. As a result, Facebook’s system knows to translate the birthday salutation “HBD friend!” into “Feliz cumpleaños amigo!” in Spanish, while Google Translate leaves “HBD” as it is. Earlier this year Facebook’s system learned that some people writing in French use unconventional variations on the English word “wow,” such as “uau” (see “Facebook Plans to Boost Its Translations Using Neural Networks This Year”). The new feature should provide more such data, as well as very explicit feedback when people edit the automatic translations of their posts.
Diana Inkpen, a professor at the University of Ottawa, says that data could provide a unique opportunity to build translation software with a good colloquial touch.
Most translation systems are trained with collections of business, political, legal, and news documents that have been translated by professionals, but they tend to use formal language. Google’s system lets people edit its automatic translations, but that text also leans toward the formal, says Inkpen. “I think the Facebook idea is very good,” she says.
Christopher Manning, a professor at Stanford University, agrees. “A good quantity of parallel Facebook-post-style data would allow much, much better and more colloquial translations,” he says.
But Manning adds that although free, crowdsourced data isn’t perfect. Facebook users may not themselves be very good at translating slang, and they might even try to teach the system false translations. The company says it doesn’t automatically start using new translations picked up from user data.
Keep up with the latest in machine learning at EmTech MIT.
Discover where tech, business, and culture converge.
September 11-14, 2018
MIT Media Lab