谷歌的人工智能翻译

GUY RAZ, host: If you're looking for final proof that machines are taking over even the most human of tasks, our producer Brent Baughman says he's found it in a new challenge being taken on by artificial intelligence.

BRENT BAUGHMAN: Artificial intelligence already exists in a lot of places. It's what recommends movies for your Netflix queue, and even some cars now have automatic parallel parking. But the basic definition of artificial intelligence is when a machine learns to do something.

Mr. DMITRIY GENZEL (Research Scientist, Google): Basically something that a machine cannot do well, but a human can.

BAUGHMAN: That's Dmitriy Genzel. He's a researcher at Google. And maybe you've used Google Translate to transform, say, English into Chinese. Well, Dmitriy Genzel is part of a team at Google working with that same technology to tackle what he says is one of the biggest challenges in artificial intelligence: translating poetry. And it's what computer scientists call AI complete.

Unidentified Woman: AI complete.

Mr. GENZEL: Which means that it's as hard as any problem that's considered to be an artificial intelligence.

BAUGHMAN: And that's because as you go deeper into poetry, you begin to run into all sorts of things that are a real challenge for artificial intelligence to think about: length, meter and rhyme.

So take Edgar Allan Poe's "The Raven."

(Soundbite of music)

Unidentified Woman: Once upon a midnight dreary, while I pondered weak and weary...

BAUGHMAN: But, say, it's in Spanish.

Unidentified Woman: (Spanish spoken)

BAUGHMAN: So Dmitriy Genzel says if you want to translate this to English, here's what his program does.

Mr. GENZEL: If you imagine this kind of a huge space of sentences, and there's all these English sentences that you could pick, some of those sentences have nothing to do with the original.

Unidentified Woman: Once upon a time, in a galaxy far, far away...

Mr. GENZEL: So this is those that are bad. Some of them maybe have some of the right words, but they don't make any sense because they are not grammatical.

Unidentified Woman: Dreary, pondered and weak.

Mr. GENZEL: So those also we don't want. So what we do is we search through this huge space of billions and billions and billons of possibilities, and we do it in a smart way so that we pick ones which are both accurate and fluid.

Unidentified Woman: Once upon a midnight dreary, while I pondered weak and...

BAUGHMAN: But just a quick reality check here. Translating the actual language isn't even the hardest part. Where things get more complicated is with meter. Take Shakespeare.

Unidentified Woman: Shall I compare thee to a summer's day?

Mr. GENZEL: I'd say Shakespeare is usually in iambic pentameter(抑扬格五音步), which means you get five what they call feet, and each foot has two syllables, one unstressed and one stressed.

BAUGHMAN: One Dmitriy's program can do is look through a pronunciation dictionary, sort of like you would use to learn a foreign language.

Mr. GENZEL: A dictionary that says, okay, the first sound is A, and then second sound is M, you know?

Unidentified Woman: Compare thee, compare thee.

Mr. GENZEL: So you know exactly how many syllables a word has and where the stress is on the word.

Unidentified Woman: Shall I compare thee to a summer's day?

BAUGHMAN: But even that, Dmitriy Genzel says, is not the hardest thing for artificial intelligence to do.

Mr. GENZEL: The hardest thing to do is rhyme because rhymes are not just in one place in the sentence. It connects to different places.

BAUGHMAN: So let's go back to "The Raven." We're translating this line from Spanish...

Unidentified Woman: (Spanish spoken)

BAUGHMAN: ...to English.

Unidentified Woman: Once upon a midnight dreary...

BAUGHMAN: But when they get to the word that has to rhyme...

Unidentified Woman: While I pondered weak and...

Mr. GENZEL: We don't make one decision. We make a thousand.

BAUGHMAN: That means for every translation that's close...

Unidentified Woman: While I pondered weak and teary...

BAUGHMAN: ...there's another sentence in this thousand that's not even close to a rhyme.

Unidentified Woman: While I pondered weak and orange...

Mr. GENZEL: So it will pick orange just as happily, and it will have to then throw it away afterwards.

BAUGHMAN: Eventually, the computer gets it right, but it takes some time. Google's poetry translation program is not public yet, and full working translations of poetry are probably pretty far off.

But, Dmitriy Genzel says, working to improve translation software of any kind is still useful.

Mr. GENZEL: Most of the content on the Internet is not in English anymore. Even for English speakers, if you read a news article, you know, about some country, you know, you bet if you open their news site, which may not be in English, and read that, you will get huge understanding that people don't really see things in the same way as, you know, that you might read in the foreign press, for example.

RAZ: That's researcher Dmitriy Genzel, who works for Google. He spoke to our producer Brent Baughman. And in our podcast this week, you can hear our credits(工作人员) translated into amphibrachic trimeter（抑扬抑格三音步） by Google's poetic translation program.