Cowdsourcing.org The Industry Website

Register Login
or sign in with

Web's Largest listing of crowdsourcing and crowdfunding events

Events

Advertise
 
Web's Largest Directory of Sites

2,412 crowdsourcing and crowdfunding sites


Duolingo and the Power of Human Computation
© Image: Eric Blattberg / Duolingo
editorial

Duolingo and the Power of Human Computation

Luis von Ahn wants to translate the web — all of it. To call him ambitious is an understatement. In a TED Talk that was originally uploaded to YouTube in April 2011, von Ahn introduced Duolingo, a crowdsourced translation project, and boldly proclaimed that with one million users, the site could help to convert the entirety of Wikipedia into Spanish in 80 hours. Free of charge. Even with a slightly more modest prediction of 100,000 users, the task would be completed within five weeks. What von Ahn, an entrepreneur and computer science professor at Carnegie Mellon, was proposing, and what Duolingo is now beginning to offer in a private beta, is a crowdsourced translation service that provides every volunteer with a service of their own. What he envisions is a tool that will not just revolutionize the Internet, but education itself.

 

By reaching out to those who are trying to learn a language, a pursuit which von Ahn estimates 1.2 billion people around the globe are already engaged in, Duolingo is seeking to harness this motivation and put it to good use. As he points out, the fact that there isn’t a single foreign language Wikipedia that comes close to the size and comprehensiveness of the English-language Wikipedia is something that the power of the web can and should challenge. Until this point, translation has been approached from the professional sphere. Those who want their website translated will usually have to pay, by the word, to have a professional translate it for them.

Duolingo, by contrast, puts its faith in the power of crowds to not only do the work of professional translators but — in harmony with one another — to do it faster, more affordably and just as well. Writing about the social tools that the internet have given us, media critic Clay Shirky points out that we have reached a point where we can “remove older obstacles to public expression and thus remove the bottlenecks that characterized mass media.” Put simply, Duolingo is removing the bottleneck of professional translation services. When everyone becomes a translator, and the tools for translation are entrusted to the crowd, the professional is no longer necessary.

Critically, too, the focus for the users of Duolingo will presumably be more on the process of translation than the final product of their efforts. The mass amateurization of publishing, which Shirky describes giving all writers their own space to publish stories online, provides a final product that is tied to the individual who created it. Your blog, your stories, your material. The Duolingo model, on the other hand, provides a final product that is the sum of the crowd’s efforts. As a beginner, you will translate simple, short sentences with suggestions for words that you might not know. As you continue, you will be able to rate other users’ translations of the same sentence. By combining the many — sometimes imperfect — translations of the same sentence, Duolingo comes up with an optimum translation for the sentence.

 

Michael Nielsen, whose book Reinventing Discovery looks at how crowds can transform our intellectual capabilities, argues for the importance of amplifying collective intelligence. On the surface of things, Duolingo would not necessarily fit his model, which presents the power of the crowd relying on “designed serendipity” to take advantage of individual expertise. Duolingo, first and foremost, seems to be designed to take advantage of the motivation to learn, not a pre-existing expertise.

Yet the system will not, presumably, run without some degree of expert oversight. To start with, the ability of the system to present beginners with clues and suggestions would not be possible without some base level of pre-translated, screened material. This was the same thinking that went into turning von Ahn’s previous project, CAPTCHAs, into such an enormous success. This project, sold to Google in 2009 and still in use today primarily under the name reCAPTCHA, utilizes the brain potential of Internet users who would otherwise be waiting idly for 10 seconds at a time. By asking humans to verify that they are human by typing words that appear distorted on their screen, CAPTCHAs serve the double purpose of increasing web security and digitizing books. What von Ahn realized in creating the system was that because the need for web security and human verification was not going to go away, the CAPTCHA program could positively use that potential.

In both cases, with CAPTCHA and Duolingo, von Ahn is trying to fill a gap that computers are currently incapable of filling automatically. A system such as CAPTCHA, which verifies that someone is a human, rather than a computer program, will by definition be unable to translate text that is only comprehensible by the human eye. The 99.5% accuracy of transcription that millions of humans provide is simply beyond the reach of even the most sophisticated computer system. Similarly, pure computer translation is still some way from an acceptable level of accuracy: von Ahn estimates “15 to 20 years,” which more accurately means we really have no idea. Computer translations are plagued by basic grammatical errors, limited vocabulary and a failure to appreciate the complexities of idioms, rhetorical devices and informal speech.

With an audience that von Ahn hopes will enter enthusiastically into this system, Duolingo can potentially use mass human potential to outsmart a computer system. Their motivation, however, will be harder to pin down. The private beta is still being slowly rolled out, with a high demand meaning that invites are in short supply. Spanish and German are currently offered as the only options while French, Italian and Chinese are listed as “coming soon.” Clearly, however, von Ahn sees the financial accessibility of the program as essential to its success. By keeping it free, he is making a case for a kind of educational initiative that will span not only countries, but also socioeconomic levels. With a more connected world, Duolingo will be providing a service that challenges a current business model where — as von Ahn puts it — “the student pays Rosetta Stone $500.” Shirky neatly sums up why a system such as Duolingo could not survive any kind of fee-paying model in his essay on micropayments. With his idea that “the competitive edge of free content is increasing,” you can see how persuasive Duolingo will be once it has rolled out fully to a world that wants to learn languages but has — until now — expected to pay money to learn them whether through tuition, books or CDs.

With that said, signing up for Duolingo still represents an investment of time, rather than an investment of money. If the product you create with that investment — a translated sentence — is not identifiably “yours,” then what is the value of your time? The simplicity of the model, von Ahn suggests, comes from the immediate reward for your time being an understanding of a new language. Furthermore, the Duolingo model allows for collaboration on a massive scale without direct communication between users. In other words, just by participating at the most basic, beginners’ level you will be collaborating with every other beginner who has been assigned the same sentence for translation as you. You will be provided with a wide range of activities, including audio transcription, to exercise the different parts of your brain necessary for language acquisition. Michael Nielsen notes that the scale of online collaboration often means that it is most beneficial to “direct participant’s attention to places where their talents are best suited to take the next step.” On Duolingo, every participant would by default be directed to whatever the next chain of basic text, or audio clip, or image caption in need of translation was. By sheer scale, the model would eliminate the redundant aspects of collaboration.

As Duolingo goes public, what remains to be seen is how quickly the service will be utilized and how the end product will be used. As impressive as the concept of translating the entirety of Wikipedia into Spanish within 80 hours is, Duolingo has yet to announce exactly which websites it will be working to translate directly, and how it will go about choosing the material that needs to be translated. Assuming that Wikipedia is a natural starting point for this massive collaboration due to its in-built support for user input, a pure translation of the English Wikipedia would only be up to date for the first second in which it began to translate. Translation the Wiki Way, written as far back as 2006, notes that a mass translation project “is optimized for translating a frozen and finalized document.” Wikipedia, by definition, is neither of these things. It constantly evolves, with new information and edits being added on a second to second basis.

It’s not easy to see how Duolingo intends on working around this problem, yet it could potentially open up an exciting prospect if the project is utilized consistently by as many people as von Ahn is hoping. Assuming that at any moment in time, at least a certain percentage of Duolingo users will be logged into the system, a live edit of the English language Wikipedia could potentially be translated into Spanish within seconds. Say, for example, that a sentence is added to reflect a recent development on Barack Obama’s education policy. Immediately, that sentence would be shared with the Duolingo servers and offered as a “learning” sentence for users who had signed up to learn Spanish. Immediately, and automatically, the sentence would be processed by English speakers learning Spanish, and by Spanish speakers learning English. This opens up the possibility of live translation of ongoing news stories, offering a global, interconnected audience for news stories that would not rely on the inefficient, expensive practice of professional translation. Likewise, English readers would benefit from immediate translation of stories about, for example, the European debt crisis, from respected publications across the continent.

Before Duolingo can master that potential, however, it will have to convince its primary wave of users that the system is actually providing them with a tangible benefit in the form of learning a language. This is one part of von Ahn’s vision that remains, at the moment, somewhat vague. Without tangible evidence that those who are using the system are — individually, rather than collectively — mastering a language, Duolingo will struggle both to attract new users and retain the users it has already attracted. Current literature on the subject has leaned heavily towards critiquing online language learning services for lacking “communication and confidence.” Most obviously, the lack of face-to-face, verbal interaction has limited the potential of online language tuition. Synchronous learning management systems (SLMSs), such as virtual whiteboards and videoconference teaching seminars, have been singled out as essential to any online language learning enterprise.

The alternative that Duolingo presents, however, defies the basic premise of all previous virtual learning environments — that a professional teacher must impart knowledge to the learner. What von Ahn calls “learning by doing” does not eliminate the need for some professional oversight entirely. Rather, the role of the professional bilingual on Duolingo would be to provide the suggestions that help to guide a beginner, or to provide the basic translation services that help the program to get off the ground. Once a beginner has started to use the system, the power of the crowd ensures accuracy and provides consistency. It is up to the Duolingo user to put in the hours necessary to consistently improve at the language they have chosen.

This leads to the potential way in which the increasing availability of mobile networks can aid Duolingo’s quest to rapidly translate the web into foreign languages. A free Duolingo app, available for all mobile users, would allow those who had registered for Duolingo to continue to learn while they were on the go. Because the ethos of Duolingo is built upon learning by doing, there is no reason to believe that those who sign up for the service would not try to maximize their time spent improving their grasp of a foreign language. On the go — while commuting, relaxing, waiting — users could continue to participate in the massive online collaboration that Duolingo could potentially usher in. Irma Borst’s comprehensive Understanding Crowdsourcing illuminates, among other things, that intrinsic motivation is an important driver of participation and performance for crowdsourcing projects. What this means is that extrinsic motivation — rewards, recognition or monetary gain — is not necessarily relevant for activities that provide both “pleasure and challenge.” (132) The satisfaction of learning a language combined with the challenge of mastering it suggests that participation in Duolingo would likely follow onto mobile networks.

Borst’s paper also recommends that those trying to crowdsource “rely on the minority of [their] online community members.” (144) This, in essence, is the Wikipedia model that takes The Long Tail into consideration. For all of von Ahn’s optimism, it is hard to argue with the idea that even if a fraction of the 1.2 billion people trying to learn a language today connect to Duolingo, and even a fraction of those who connect to Duolingo use the service consistently enough to make a difference, the service will still be able to rely on a large enough minority of members to provide a consistent stream of translated material. There would be no need to add a rewards system to Duolingo, because there is no need for a specific level of performance on Duolingo. Even those who deliberately “trolled” the system and provided misleading translations would be weeded out due to the massive scale of the translation aggregation and the fact that you are constantly rating your fellow translators’ translations. In essence, you perform the double duty of both translator and screener for material that is clearly inaccurate.

Ultimately, the crowd stands to gain the most by participating the most in a service like Duolingo. The National Science Foundation, who funded von Ahn with $480,000 in March 2011 to build Duolingo and roll it out to the world, clearly sees the merit of the project. In the abstract for the grant, von Ahn notes that language education on Duolingo will be both the “vehicle and the incentive” for human computation. Rather than entrusting the task of translation to a slice of bilingual translators who have professionalized the trade, von Ahn is asking us to do ourselves a favour by doing the web a favour. Though we are still some way off a full translation of Wikipedia, let alone live translations of rolling news stories, the model that von Ahn is proposing may do more to unlock the potential of the internet — where every user is a learner — than any other development in its previous existence.

Alistair Mackay is a journalism and history double major enrolled in the Arthur L. Carter Journalism Institute at New York University. Follow him on Twitter @alimackay.

Flag This

0

Comments

Guest
 Join or Login
 Optional