Duolingo: The Next Chapter in Human Computation by Luis von Ahn (Full Transcript)

But the best part is the message that the Yahoo Help got about 20 minutes later. [“Help! I’ve been waiting for 20 minutes, and nothing happens!”] The person thought — this of course is not as bad as this poor person who… [R E S T A R T].

OK, now, this captcha project is something that we did here at Carnegie Mellon about ten years ago and it’s been used everywhere. Let me now tell you about a project that we did a few years later, which is sort of the next evolution of captchas. This is the project that we called Recaptcha, which is something that we started here at Carnegie Melon, then we turned it into a startup company, and then about a year and a half ago Google actually acquired this company. So, let me tell you what this project started.

So this project started from the following realization: it turns out that approximately 200 million captchas are typed every day by people around the world. When I first heard this I was quite proud of myself I thought, “Look at the impact that my research has had.” But then I started feeling bad, so here is the thing: each time you type a captcha, essentially, you waste ten seconds of your time. Because it takes ten seconds to type a captcha — and if you multiply that by 200 million you get that humanity as a whole is wasting about 500 thousand hours every day typing these annoying captchas.

So then I started feeling bad, and then I started thinking, well, of course we can’t just get rid of captchas, because the security of the Web sort of depends on them. But then I started thinking, “Is there any way we can use this effort for something that is good for humanity?” so here is the thing, while you’re typing a captcha, during those ten seconds, your brain is doing something amazing. Your brain is doing something that computers cannot yet do. So, can we get you to do useful work for those ten seconds?

Another way of putting it is: Is there some humongous problem that we cannot get yet computers to solve, that somehow we can split into tiny ten second chunks such that each time somebody solves a captcha they solve a little bit of this problem? And the answer to that is yes, and this is what we are doing now. So what you may not know is that nowadays, while you’re typing a captcha not only are you authenticating yourself as a human but in addition you are actually helping us to digitize books. Let me explain how it works.

There is a lot of projects out there trying to digitize books. Google has one, the Internet Archive has one, Amazon, with the Kindle, is trying to digitize books. Basically the way this works is, you start with an old book, like a physical thing, you’ve seen those things right? Like a… book.

So, you start with a book, and then you scan it. Now scanning a book is like taking a digital photograph of every page of the book. It gives you an image for every page of the book. This is an image with text for every page of the book. The next step in the process is that the computer needs to be able to decipher all of the words in this image. That is done using a technology called OCR, Optical Character Recognition, which takes a picture of text and tries to figure out what text is in there.

Now, the problem is that OCR is not perfect, especially for older books, where the ink has faded, and the pages have turned yellow, OCR cannot recognize a lot of the words. For example, the things that were written more than 50 years ago, the computer cannot recognize about 30% of the words. So what we are doing now is we are taking all of the words that the computer cannot recognize and we are getting people to read them for us, while they are typing a captcha on the Internet. So next time you type a captcha, these words that you’re typing are actually words that are coming from books that are being digitized that the computer could not recognize.

Now the reason we have two words nowadays instead of one, is because, you see, one of the words is a word that the system just got out of a book, it didn’t know what it was and it’s going to present it to you, but since it doesn’t know the answer for it, it cannot grade it for you. So what we do is we give you another word, one for which the system does know the answer, we don’t tell you which one is which, and we say, please type both. And if you type the correct word for the one which the system already knows the answer, it assumes you are a human, and it also gets some confidence that you typed the other word correctly. And if we repeat this process to like 10 different people, and all of them agree on what the new word is, then we get one more word digitized accurately. So this is how the system works.

And basically since we released it about three or four years ago, a lot of websites started switching from the old captcha, where people wasted their time, to the new captcha, where people are helping to digitize books. So, for example, Ticket Master, so every time you buy tickets on Ticket Master, you help to digitize a book. Facebook, every time you add a friend, or poke somebody, you help digitize a book. Twitter, and about 350,000 other sites are all using recaptcha. And in fact, the number of sites that are using recaptcha is so high that the number of words that we’re digitizing per day is really, really large. It’s about 100 million a day which is the equivalent of about 2.5 million books a year. And this is all being done one word at a time by just people typing captchas on the Internet.

Of course, since we are doing so many words per day funny things can happen, and this is especially true because now we’re giving people two randomly-chosen English words next to each other. So funny things can happen, for example, we presented this word, the word “christians”, there’s nothing wrong with it, but if you present it along with another randomly chosen word, bad things can happen: [“bad christians”] But it’s even worse because the particular website where we showed this actually happened to be called The Embassy of the Kingdom of God.

Here is another really bad one: johnedwards.com [“damn liberal”]

So we keep on insulting people left and right every day. Now of course we are not just insulting people, see, here is the thing. Since we’re presenting two randomly chosen words, just interesting things can happen. So this actually has given rise to a really big Internet meme that tens of thousands of people have participated in, which is called Captcha Art. I’m sure some of you have heard about it.

Here’s how it works: imagine you’re using the Internet and you see a captcha that you think is somewhat peculiar. Like this captcha: [“invisible toaster”] Then what you are supposed to do is, you take a screenshot of it, then of course you fill out the captcha because it helps us digitize a book — but then, first you take a screenshot and then you draw something that is related to it. That’s how it works. There are tens of thousands of these, some of them are very cute, [clenched it] some of them are funnier, [stoned founders] and some of them, like “paleontological shvisle”, they contain Snoop Dog.

OK, so this is my favorite number of recaptchas, this is the favorite thing that I like about this whole project. This is the number of distinct people that have helped us digitize at least one word out of a book through recaptcha: 750 million, which is a little over than 10% of the world’s population has helped us digitize human knowledge. And it is numbers like these that motivate my research agenda.

So the question that motivates my research is the following: If you look at humanity’s large scale achievements these really big things that humanity have gotten together and done historically, like for example, building the pyramids of Egypt, or the Panama Canal, or putting a man on the moon, there’s a curious fact about them. And it is that they were all done with about the same number of people. It’s weird. They were all done with about 100,000 people. And the reason for that is because before the Internet, coordinating more than 100,000 people let alone paying them, was essentially impossible.

But see now with the Internet, I’ve just shown you a project where we’ve gotten 750 million people to help us digitize human knowledge. So the question that motivates my research is if we can put a man on the moon with 100,000 what can we do with 100 million? So based on this question we’ve had a lot of different projects that we’ve been working on. Let me tell you about one that I’m most excited about. This is something that we’ve been sort of semi-quietly working on for the last year and a half or so. It hasn’t yet been launched, it’s called Duolingo, since it hasn’t been launched, shhh! Ah, yeah, I can trust you.

OK, this is the project. So here is how it started. It started with me posing a question to my graduate student, Severin Hacker — this clicker is not working so well – OK, that’s Severin Hacker, so I posed a question to my graduate student. By the way, you did hear me correctly, his last name is Hacker. So I posed this question to him: How can we get 100 million people translating the Web into every major language for free? There’s a lot of things to say about this question. First of all, translating the Web. So, right now the Web is partitioned into multiple languages, a large fraction of it is in English. If you don’t know any English you can’t access it, but there’s large fractions in other different languages, and if you don’t know those languages you can’t access it. So I would like to translate all of the web or at least, most of it into every major language. So that’s what I would like to do.

Now, some of you may say, well, why can’t we use computers to translate, why can’t we use machine translation? Well , machine translation nowadays is starting to translate some sentences here and there, why can’t we use it to translate the whole Web? Well, the problem with that is, it’s not yet good enough, and it probably won’t be for the next 15 to 20 years. It makes a lot of mistakes, and even when it doesn’t make a mistake, since it makes so many mistakes you don’t know whether to trust it or not.

So let me show you an example of something that was translated with a machine. It was actually a forum post with someone who was trying to ask a question about Javascript. It was translated from Japanese into English. Let’s see if this works. I’ll just let you read, this person starts apologizing for the fact that it’s translated with the computer. [This is question, English is faulty. Thank computer to translate. Sorry!]

The next sentence is going to be the preamble to the question. He’s just explaining something, remember, it’s a question about Javascript: [At often, the goat-time install a error is vomit.]

Then comes the first part of the question: [How many times like the wind, a pole, and the dragon?] Then comes my favorite part of the question: [This insult to father’s stones?]

And then comes the ending, which is my favorite part of the whole thing: [Please apologize for your stupidity. There are a many thank you.]

OK, so computer translation, not yet good enough.

OK? So back to the question, we need people to translate the whole Web. Now the next question you may have is “Why can’t we just pay people to do this, we could pay professional language translators to translate the whole Web”. We could do that. Unfortunately, it’d be extremely expensive. For example, translating a tiny tiny fraction of the whole Web, Wikipedia, into one other language, Spanish. You know, Wikipedia exists in Spanish, but it is very small compared to the size of English, it’s about 20% the size of English. If you want to translate the other 80% into Spanish it would cost at least $50 million. This is even at the most exploited outsourcing country out there. So, it’d very expensive.

So what we want to do is we want to get 100 million people translating the Web into every major language for free. Now, if this is what you want to do you pretty quickly realize you’re going to run into two pretty big hurdles, two big obstacles. The first one is a lack of bilinguals. I don’t even know if there exists 100 million people out there using the Web who are bilingual enough to help us translate. That’s a big problem.

The other problem that you’re going to run into is the lack of motivation. How are we going to motivate people to actually translate the Web for free? Normally, you have to pay people to do this, so how are you going to motivate them to do it for free?

Now, when we were starting to think about this we were blocked by these two things but then we realized that there’s actually a way to solve both of these problems with the same solution. There’s a way to kill two birds with one stone. And that is to transform language translation into something that millions of people want to do and that also helps with the problem of a lack of bilinguals. And that is language education.

So, it turns out that, today, there are over 1.2 billion people learning a foreign language. People really really want to learn a foreign language, and it’s not just because they’re being forced to do so in school. For example, in the United States alone there are over 5 million people who have paid over $500 for software to learn a new language. So people really really want to learn a new language.

So what we’ve been working on for the last year and half is a new website. It’s called Duolingo. Where the basic idea is as people use Duolingo, people learn a new language for free while simultaneously translating the Web. Basically they are learning by doing. The way this works is whenever you’re just a beginner we give you very very simple sentences, there’s of course a lot of very simple sentences on the Web. We give you very simple sentences along with what each word means. And as you translate them, and as you see how other people translate them, you start learning the language, and as you get more and more advanced, we give you more complex sentences to translate. But at all times, you are learning by doing.

Now the crazy thing about this method is that it actually really works. First of all, people are really really learning a language, we’re mostly done building, and we are now testing it. People really can learn a language with it. And they learn it about as well as the leading language learning software. So people really do learn a language. And not only do they learn it as well, but actually it’s way more interesting. Because, you see, with Duolingo people are actually learning with real content. As opposed to learning with made-up sentences people are learning with real content, which is inherently interesting.

So people really do learn a language, but perhaps, more surprisingly, the translations that we get from people using the site, even though they’re just beginners, the translations that we get are as accurate as those of professional language translators, which is very surprising. Let me show you one example. This is a sentence that was translated from German into English, the top is the German, the middle is an English translation that was done by somebody who was a professional language translator, who we’ve paid 20 cents a word for this translation, and the bottom is a translation by users of Duolingo, none of whom knew any German before they started using the site. As you can see, it’s pretty much perfect.

Now, of course we played a trick here to make the translations as good as professional language translators. We combined the translations of multiple beginners to get the quality of a single professional translator. Even though we’re combining the translations the site actually can translate pretty fast. Let me show you, this is our estimates of how fast we could translate Wikipedia from English into Spanish. Remember, this is $50 million worth of value. If we wanted to translate Wikipedia into Spanish we can do it in five weeks with 100,000 active users, and we can do it in about 80 hours with 1 million active users. Since all of the projects my group has worked on so far have gotten millions of users, we’re hopeful that we would be able to translate extremely fast with this project.

Now, the thing that I am most excited about Duolingo is that I think this provides a fair business model for language education. So here’s the thing: the current business model for language education is: the student pays. And, particularly, the student pays Rosetta Stone $500. That’s the current business model. The problem with this business model is that 95% of the world’s population doesn’t have $500. So it’s extremely unfair towards the poor, this is totally biased towards the rich.

In Duolingo, because while you learn you’re actually creating value, you are translating stuff which, for example, we could charge somebody for translation — this is how we could monetize this. Since people are creating value while learning, they don’t have to pay with their money, they pay with their time. But the magical thing here is that they’re paying with their time but that is time that would have to be spent anyway, learning the language. The nice thing about Duolingo is I think it provides a fair business model, one that doesn’t discriminate against the poor people.

So, here is the site: www.duolingo.com. Thank you. Here’s the site. We haven’t yet launched but if you go there you can sign up to be part of our private beta which we’re probably going to start in about three or four weeks. By the way, I’m the one talking here, but actually Duolingo is the work of a really awesome team some of whom are here. So, thank you.

Related Posts

• How to Teach Students to Write With AI, Not By It
• Why Simple PowerPoints Teach Better Than Flashy Ones
• Transcript: John Mearsheimer Addresses European Parliament on “Europe’s Bleak Future”
• How the AI Revolution Shapes Higher Education in an Uncertain World
• The Case For Making Art When The World Is On Fire: Amie McNee (Transcript)