When Luis von Ahn returned to Duke last month to give a Provost’s lecture in our “Information Futures” series, it was both a homecoming and a glimpse of the future of learning. Luis received a B.S. in mathematics from Duke and, in fact, was one of many young alums we consulted with back in 1998 when we were creating our Program in Information Science and Information Studies (ISIS), the local Duke predecessor to the international network that became HASTAC (hastac.org). He is now Associate Professor in the Computer Science Department at Carnegie Mellon University.
reCAPTCHA and Crowdsourcing
Luis is the person who created CAPTCHA and reCAPTCHA. You may not know it, but those annoying squiggles and blurs you have to translate into words in order to gain entry to some websites aren’t just tests of your patience but serve a function in the world. When you get one right, you not only gain entry into the website you want to enter but you are translating a word that computers cannot. Some of these are words in books that cannot be translated computationally. Others are misspellings. But all are in need of human correction. Von Ahn created an algorithm that can generate tests that it cannot pass—but that we humans can. So far 950,000,000 people (13% of the world’s population) have offered a reCAPTCHA. We’re translating about 200 million words a day.
Von Ahn’s genius, though, isn’t just in the algorithm but in the concept behind it: there are things neither humans nor computers can do alone but that, together, can be accomplished beautifully. He called this “human computation” originally but now “crowdsourcing” is the acceptable term for humans working together online to solve problems. For me, what is inspiring about this concept is that it breaks the technodeterminist model of machines solving everything while at the same time emphasizing that some things require computational machines.
Education, to my mind, has not fully absorbed this lesson, this paradigm shift, or this model—and, before MOOCs (Massive Online Open Courseware) really become about learning (not scale), they have to really absorb, on an epistemological and pedagogical not just a financial level, the potential new ways of learning online that will help us all live better, connected lives.
I’ll write more about this in the coming weeks but, for now, I want to talk about a few more wise “human computation” or crowdsourced learning insights in von Ahn’s lecture.
His new project is called Duolingo and it adopts the same reCAPTCHA principles to learning a foreign language. In this, users can take free online language learning lessons and contribute to the World Wide Web at the same time. Instead of pointless translation exercises, everyone enrolled in a Duolingo course learns that they will be translating as part of their lessons some kind of text that machines cannot translate (very much like the reCAPTCHA principle). Here it might be translating Wikipedia pages better than Google Translator or other translation programs. The way this works is several people unknown to one another who are taking a language lesson through Duolingo might be translating the same passage. If their translation is the same, the computer accepts it as accurate. If not, then a human translation expert is brought in to do the job. Interestingly, the crowdsourced translations have been rated in blind tests by translation experts against texts translated by translation experts and have achieved the same ratings. Totally automated translators fare about 30% worse in the same rating comparisons.
Right now about 125,000 people are translating every major language for free while they are learning a language. This method has been called a “cyclical tool”—the tool that allows you to translate a language helps you learn one. It is not “extractive” like a standardized test—where all your efforts go to the task of being measured and the only result is a measurement. It is more like a bicycle, where riding one helps protect the environment and also makes you stronger so you can enjoy the environment more.
All learning, in my opinion, should be cyclical and, if we do our job right, all formal education should be a cyclical tool. Our students should enjoy learning not simply for learning’s sake but because learning, knowledge, skills, insights, wisdom, and all the rest increase one's power to live a satisfying, happy, productive, and generous life. Learning, in all its forms and methods, helps you do what you want to do in the world you inhabit beyond school.
This is one reason why I now hold all my classes in public, on Word Press blogs, so my students aren’t just sharing their thoughts with me as their professor but with others in the class, and anyone in the world who wants to tune in. I want them to respect their own thinking and skills enough to care about how they contribute to the world beyond the classroom. By making their knowledge public, they are not only learning how to get an “A” in a course (an extractive exercise if ever there were one) but they are learning to communicate complex insights effectively to people they do not know—a life skill in communication, persuasion, critical thinking, creativity, eloquence, and, of course, basic digital skills and digital lessons: about intellectual property, privacy, security, open access versus proprietary access, the benefits of one browser that keeps and commercializes your personal data versus another [like Mozilla’s open source Firefox] that does not, and a host of other intertwined technical and social issues. Posting your ideas online, you don’t just get an A but, in the process of recording your lessons, you enhance many skills you will use for a lifetime.
What Crowdsourcing Cannot Translate
Here are some other insights from Luis von Ahn. First, not every text can be translated through Duolingo’s crowdsourcing. Legal language cannot be: it is too complicated, convoluted, and site-specific to be translated by non-experts (including those who speak the language). The other kind of language that does not admit crowdsourced translation: poetry. Beauty, suggestion, inference, allusion, metaphor, alliteration and assonance, figurative language, cadence, rhythm, meter: that form of translation doesn’t require human computation. It requires a muse.
Two Ways of Motivating Learning
All this work with learning online has also given von Ahn some insights into human motivation that those of us experimenting with next generation learning can take to heart. Duolingo keeps track of every keystroke by all those hundreds of millions of language learners, and researchers study the data and find out patterns in our learning that are too subtle for us, as teachers, to understand in our micro-interactions with students. They are learning a lot about how we learn. And von Ahn shared two really clear, simple lessons that any teacher can adapt tomorrow, to face-to-face learning as well as to online learning:
1. List a student’s learning accomplishments in a course before testing them on a new skill or insight. Everyone knows if you are making a MOOC you should include a lot of quizzes. If you have ever taken a class online (I’ve taken several), the quizzes are pretty dreary and lame but they work better than learning without them. But von Ahn has found they can also be a dis-incentive to continue. Unless (and this is a good insight) the quiz is preceded by a list of what you’ve learned so far in the course. That simple device (another cyclical tool) of listing all the things that past quizzes have helped you master is not only a great way of revving up our cognitive motor, it is also a motivator that inspires you to do well on the quiz to add to your accumulated store of knowledge. Everything about human computational research shows us this is true. We like learning. We like sharing what we know. We like getting credit. We like keeping track of our learning. It’s all part of motivation, whether it is concretized or not, but, when it is, we enjoy it. Pride helps us to learn better and we take pride in that learning. Cyclical tool.
2. Offer a student the chance to do better before telling them they’ve failed a test. The Duolingo team has also learned from the data that if you just tell a student that they didn’t pass a test, it is a big de-motivator. People drop the course when they think they are failing it. We know that holds true for school drop out rates too. Failure breeds defeat and retreat. But simply by framing a test failure as an opportunity and a challenge, it snatches victory from the jaws, as the old saw goes, of defeat. Example: “You didn’t get that last question right but let’s try again with X in mind.” Cyclical tool. The failed test doesn’t extract censure, shame, and wasted effort—but gives you an insight into what you don’t know yet in order that you can do better on the next challenge. It takes more effort to build in this kind of specificity but if the goal is retention and success and motivation to continue, the effort is hugely important.
Plus an Owl
Luis von Ahn is a marvelous speaker, who presents his work with warmth and humor. He ended with a delightful final story about the mystery of learning that not only made us all smile but reminded us that there is so much about the human psyche that we do not know, do not understand, and no amount of data will ever tell us.
If you drop out of a Duolingo course, you receive an email that includes a bemused-looking owl beckoning you to return. This owl, it turns out, was arrived at by a combination of artistry and data. They kept trying out different renderings by different artists of the owl and kept track of which one worked best. There were many owls and none did better than any other.
The Duolingo owl they now use, it turns out, motivated a higher rate of return than any other owl.
Someone in our audience asked “Why?” What about this owl set it apart from all the other cartoon owls they tried?
Von Ahn’s answer is the most important one for anyone dedicated to learning. “We do not know. All we know is that it worked better than any other so we use it.”
Bingo. Some times we know exactly what works best and why. Other times, we just know it works—and we go with it. Great teachers know that, as humans, we are very bad at tracking our micro or even our macro gifts of timing and interaction, but sometimes we get it right and a great teacher goes with that—and learns to give feedback (yes, cyclical tool again) to students when they do something right, even if no one quite knows why. It’s like stand-up comedy or other performance where slight micro-adjustments are the key to success or failure, even if our data and our critical thinking and our analysis don’t get to the heart of the untranslatable poetry that is learning. Or a adorable green innocent-looking owl bidding us to return.
It’s like the Mona Lisa. We don’t know why she is smiling or why that smile captivates us for centuries. We just know, in our heart, it’s a masterpiece. So we go with it. And, if we want to make it a cyclical tool, we remind our students that sometimes, with all our machines, and all our tools, and all our data, in the end, trusting ourselves to “go with it” is a key part of learning.
It was the perfect note on which to end a lecture. Thanks, Luis, for a great one!