As of this moment, there are more than 2 trillion posts on Facebook—a figure that expands by a billion each day.
Alan Packer, the director of Facebook’s Language Technology Group, wants to understand and, ultimately engage with, these messages. But the task is more difficult than it sounds. Of the 1.6 billion people who use Facebook, he said, most don’t speak English. Most, in fact, don’t speak each other’s language. Jason Pontin, editor of MIT’s Technology Review, called Facebook the internet’s “first planetary platform.”
The beauty of the internet, Packer said on Monday at EmTechDIGITAL 2016 in San Francisco, is that we now have “access to almost all the knowledge in world, from almost anywhere in the world.” The “proliferation of languages and cultures on the internet,” however, has created a unique problem: “There’s a disconnect between the languages people speak and the content they want to access,” he said.
“We need AI to solve the problem.”
So, Facebook is using machine learning to grapple with issues like natural language understanding, translation, and interpreting slang.
“Translation is about removing language as a barrier,” Packer said, “so people can connect.”
Facebook has done studies with people who have access to translation and people who don’t, Packer said. In fact, people who used translation had twice as many friends-of-friends than those who didn’t. “People who have access to translation have more friends,” he said.
And, in conversational understanding, Packer said, Facebook wants to “find out what people want, and either connect them to other people or to the information they want.”
Why hasn’t Facebook just applied another machine learning algorithm to figure it out? It’s because people on Facebook don’t talk the way they do on the rest of the internet.
Most existing systems are trained using academic data sets mined from the web. But while useful in some situations, the text of a product manual, for instance, doesn’t apply to the language of Facebook. In contrast, said Packer, conversation on the social media site is casual, full of slang, and “is often riddled with (intentional) misspellings.”
The importance of understanding posts is also connected with the “changing nature of the internet and the devices we use to access it,” said Packer.
As more and more Facebookers have transitioned to mobile devices, the problems have become more pronounced. Content is easier to digest, said Packer, but harder to produce. The move towards wearables will only amplify this problem, he said, when keyboards become “smaller, worse, or go away,” creating a real need for vocalization tools.
The reason Packer said that Facebook wants AI to understand conversation is to engage with conversations between users. For example, he said, you might want Facebook to post a photo, tag friends, and add a caption—all through voice-recognition.
Packer also sees Facebook as assisting users who have questions they pose to friends. For instance, recommendations about hotels or restaurants. Facebook could potentially have the ability to step in and boost that post on the wall of a friend who was recently tagged in the destination in question.
“We have the ability to participate and enhance the conversations happening on Facebook today,” Packer said.
But the problem is more difficult than it sounds. The ability for Facebook to understand meaning would require, Packer said, Facebook to “track conversation, to have a memory, to be contextually aware.”
Still, Facebook has a key advantage: An extraordinary volume of user data.
“We see how people are communicating, we see how they are having conversations,” said Packer. “We know you.”
And it’s clear that “knowing you,” and connecting people socially isn’t Facebook’s only motivation in improving communication tools. Facebook makes money selling ads,” said Packer, “but that’s already Facebook’s business model.”
When Packer was asked about the potential privacy concerns that may arise when Facebook users discover that the company wants to access their personal conversations online.
“Facebook takes privacy more seriously than any company I’ve ever worked at,” Packer said. “You can’t write code that violates privacy.”