NPR's Podcast 'Invisibilia' Begins A New Season NPR's Invisibilia podcast brings you a story about a new Facebook effort to build a modern kind of Rosetta Stone, using artificial intelligence.
NPR logo

NPR's Podcast 'Invisibilia' Begins A New Season

  • Download
  • <iframe src="https://www.npr.org/player/embed/813577442/813577447" width="100%" height="290" frameborder="0" scrolling="no" title="NPR embedded audio player">
  • Transcript
NPR's Podcast 'Invisibilia' Begins A New Season

NPR's Podcast 'Invisibilia' Begins A New Season

NPR's Podcast 'Invisibilia' Begins A New Season

  • Download
  • <iframe src="https://www.npr.org/player/embed/813577442/813577447" width="100%" height="290" frameborder="0" scrolling="no" title="NPR embedded audio player">
  • Transcript

NPR's Invisibilia podcast brings you a story about a new Facebook effort to build a modern kind of Rosetta Stone, using artificial intelligence.

NOEL KING, HOST:

There's a sophisticated piece of technology that offers a new understanding of language. NPR's podcast Invisibilia is reporting on it, and host Alix Spiegel is going to start more than 200 years ago.

ALIX SPIEGEL, BYLINE: It was a summer day, hot and dusty, and the soldiers were supposed to be cleaning, clearing up the tumbled-down rocks in the foundation of the fort they'd been ordered to rebuild. But then one of the soldiers, this man named Bushar (ph), noticed what looked like a gray slab peeking out from the dirt.

JOHN RAY: And there was an old stone covered with writing.

(SOUNDBITE OF MUSIC)

SPIEGEL: An old stone - specifically, the Rosetta stone. This was 1799. And Egyptologist John Ray says because the tablet they pulled from the rubble had the same declaration written in both hieroglyphs - this strange picture script which no one then understood - and ancient Greek - which was, at that point, still known - the stone unlocked a mystery. For more than 2,000 years, the world of ancient Egypt had been largely lost. But now...

RAY: Can suddenly read the jokes that they're saying to each other as they work in the fields.

SPIEGEL: And what researchers discovered was a world that actually felt remarkably similar to our own.

RAY: They fell in love. They fell out of love. They went to their gods with troubles and worries and say, gods, help me. They wrote down humorous stories. They wrote down some stories that were only for an adult audience; let's put it that way.

(SOUNDBITE OF MUSIC)

SPIEGEL: The British carried the Rosetta stone off. Today, it sits in the British Museum in London. Ray says it is hands down the most popular thing there.

RAY: More people stand in front of the Rosetta stone than any other thing in the museum.

SPIEGEL: There's just something about the idea of a key that unlocks a whole system, that compels people, which brings me to the real thing I want to tell you about - a new Rosetta stone that isn't made of marble or granite, but electricity, and was discovered not in the foundation of an old fort, but in a much more modern institution.

So why don't we just start with, can you say your name and what you do at Facebook?

ANTOINE BORDES: My name is Antoine Bordes. I'm a research director at Facebook AI Research.

SPIEGEL: Several years ago, two machine-learning experts in Antoine Bordes' lab at Facebook approached him with a strange-sounding idea. They wanted to see if they could get a computer to translate from one language to another - say, English to German - even if the computer was never given any kind of dictionary or any other information about the actual words in those two languages.

BORDES: And so I told them, give it a try, but I don't think this is going to work.

SPIEGEL: Now, this wasn't completely insane. And to understand why, you need to know a little about this interesting advance in machine learning that happened in 2013, when some clever computer scientists created a program which took all the words in the English language and had the computer use them to create this image. Picture in your head a three-dimensional cloud of tiny dots hovering in space.

BORDES: A cloud of points where each of the points correspond to one of the words of the dictionary. That's what the machine produced.

SPIEGEL: But here's the thing. The placement of each of the tiny dots hovering in the cloud - it isn't arbitrary.

BORDES: They are not random in this cloud. The position depends on the meaning.

SPIEGEL: Their position depends on the meaning and how concepts relate to each other. So for example, there's a point in the cloud that represents the word king...

BORDES: King.

SPIEGEL: ...And a point in the cloud that represents the word man.

BORDES: Man.

SPIEGEL: Of course, since this is the whole English language, there's also a point in the cloud representing the word queen and a point in the cloud representing the word woman. And what you see when you zoom in and look really closely is that the placement of the man point relative to the king point is the exact same as the placement of the woman point relative to the queen point. The distance and direction of man to king is the same as the distance and direction of woman to queen, and girl to princess, et cetera, et cetera. So you start to see analogies represented geometrically, which means that you can do actual arithmetic on language.

BORDES: You can do operation on words now. And you could say, if I take man and I add to it king and if I remove queen, then what I should obtain is woman.

SPIEGEL: Or another example.

BORDES: France minus Paris plus Washington, D.C., equal United States.

SPIEGEL: When language is represented geometrically in this way, you can do all sorts of interesting things - a new kind of math with words. So dozens of languages were turned by computers into these clouds.

BORDES: I would say more than 200 languages have been turned into these point clouds - French, Italian...

SPIEGEL: German, Japanese, Swahili.

BORDES: ...Even Klingon.

SPIEGEL: #computerprogrammers, shrug emoji.

BORDES: Yeah.

SPIEGEL: Which brings me back to those two machine-learning experts at Facebook who approached Bordes.

BORDES: Guillaume Lample and Alexis Conneau.

SPIEGEL: They wanted to take the French cloud shape and line it up with the English cloud shape just to see if you might be able to maybe somehow use this underlying shape to do translation, because, you know, if all the dots in the clouds were in the exact same place, then you wouldn't necessarily need a dictionary.

Again, Antoine thought this was extremely unlikely to succeed.

BORDES: You would need at least to know a few words. Maybe you don't need, like, millions of them, but you need at least a few hundred.

SPIEGEL: And he wasn't alone in his skepticism. Many people thought that languages are so different and cultures are so different and the way we see the world is so different that the underlying shapes wouldn't match up. But you know what? They did. Though it wasn't perfect, underneath, the shapes were similar enough that you could do translation. You've created a kind of Rosetta stone, right?

BORDES: Yeah. I mean, we wanted to call the project Rosetta at some point, actually.

SPIEGEL: Now, there are plenty of practical applications for this. There's even a theory that you might be able to use it to decipher animal communication.

BORDES: I think we could to some extent, yes.

SPIEGEL: But to me, there's something about this work that feels deeper than any practical application. It feels important that underneath all of our massive and very real cultural differences, we're so similar that all you have to do is blur your eyes a little bit to translate from one language to another.

(SOUNDBITE OF MUSIC)

Copyright © 2020 NPR. All rights reserved. Visit our website terms of use and permissions pages at www.npr.org for further information.

NPR transcripts are created on a rush deadline by Verb8tm, Inc., an NPR contractor, and produced using a proprietary transcription process developed with NPR. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.