November 1, 2013 § Leave a comment
Luis von Ahn, computer-science professor at Carnegie Mellon University, has a goal. It’s to translate the entire Web into every major language, for free. Sound impossible? Not to von Ahn. But he does see two obstacles: not enough bilinguals and not enough translator motivation.
So when it comes to translation, what can turn those obstacles from mountains into molehills? Von Ahn is working on an answer, and so is Chang Hu.
It Takes a Crowd
The Guatemalan-born von Ahn is best known for helping to invent CAPTCHAs. If you don’t know what a CAPTCHA is, it’s that image of distorted letters you see on a lot of Website forms. You’re required to type in those letters to prove that you’re a human, which keeps computer programs from fooling the system.
As he told the crowd at a TEDx Talk in 2011 (embedded below), Von Ahn estimates that each day, about 200 million CAPTCHAs are typed around the globe. With every CAPTCHA taking about 10 seconds to key in, that’s around 500,000 hours a day. Von Ahn wondered how he could redeem this “wasted” time and came up with reCAPTCHA.
Now owned by Google, reCAPTCHA replaces the often random characters of a CAPTCHA with actual words from books that are being digitized. The reason this is a good thing is because the text-scanning software used to digitize printed text can’t recognize every word, especially when dealing with books over 50 years old. But these hard-for-computers-to-read words aren’t hard for human’s at all. So when you’re typing in a CAPTCHA on one of over 350,000 sites using reCAPTCHA—including Facebook, Twitter, and Ticketmaster—you’re helping digitize books.
So what does this have to do with translation? Well, another of von Ahn’s projects, based on the same kind of crowd-sourced “human computing” as reCAPTCHA, is Duolingo. It’s a free language-learning site, currently teaching six languages. What makes Duolingo unique is that while you’re learning a language, you’re joining 10 million other users in translating text on the Web, because the phrases used by Duolingo come from real Websites.
For instance, after you learn some basic Spanish vocabulary, you’ll be able to test your skills by translating simple phrases to and from Spanish. And as you do so, you’ll be helping translate some English Websites into Spanish, or vice versa. Success earns you “skill points,” unlocking new lessons, while mistakes take away one of your hearts. Lose all of your hearts and you have to redo the level. As you learn more, you translate more-complex sentences, and, as your attempts are compared with those of others, useful, accurate translations are produced.
According to von Ahn, two great things about Duolingo are, “People really can learn a language with it, and they learn it about as well as the leading language-learning software,” and, “The translations that we get from people using the site, even though they’re just beginners . . . are as accurate as those of professional language translators.”
Oh, yeah, and did I mention it’s free? That’s possible because the sites that submit their text for translation are paying the tab—sites like Buzzfeed and CNN, which, von Ahn announced just a couple weeks ago, are the first to come on board.
Of course, even when there’s no monetary cost, not everyone wants to invest his time into the hours that are required for learning a language. If there could be a way for monolinguals to help out with just a few seconds—kind of like with the reCAPTCHAs—that might bring more people in.
The Power of Widgets
MonoTrans (named MonoTrans2 in its newer version) is a process that combines machine translation with help from monolingual humans to produce accurate translations. A team from the University of Maryland’s Department of Computer Science, led by Chang Hu—a PhD candidate at UMD—proposed the process in 2010 to overcome the problem of not having enough bilingual translators to work on (a) texts in rare languages, and (b) huge amounts of text that would require enormous amounts of human effort.
MonoTrans starts with a computer translation of a passage, which is notorious for producing flawed (and often humorous) results. The output is then passed on to a person who speaks the target language. She then makes a guess as to the correct meaning and phrasing of the sentence, and her efforts are back-translated into the source language. Then a speaker of that language compares the results to the original passage, and the process between the two speakers is repeated until a satisfactory translation is produced. Along the way, the two monolinguals can help each other by including annotations, such as images and Web links, and multiple participants can vote on results.
While the process doesn’t necessarily take a large number of steps, it can be complicated and time consuming. MonoTrans2 addresses this problem by breaking the process into smaller, individual “microtasks,” so that many more people will take part in a translation, with each one handling only a small part of the whole process.
This new method was tested using children’s books at the International Children’s Digital Library. Visitors to the Website were presented with “widgets,” windows on a page that run a simple program. These widgets allowed users to edit or paraphrase a sentence, identify errors, or vote for the sentence they think is best.
The results of the trial show that using the MonoTrans Widgets in conjunction with Google Translate is a significant improvement over using Google Translate alone. And while this method also introduced some inherent problems, it indicates that the future of crowd-based computation by monolingual humans is very promising.
A Match Made in Cyberspace
Luis von Ahn coined the term human computation to describe using people to accomplish tasks that computers usually perform. Hu, in a blog post, sums up the relationship of human computation to translation in this way:
[H]uman computation presents a unique opportunity to significantly lower the threshold to do translation. At the same time, translation provides a set of interesting problems for human computation.
It sounds as if the relationship is something like a dance, with the dancers figuring out the steps as they go. Or maybe it’s more like a marriage, where both partners aid and challenge each other at the same time.
It’s a good union, and I’m glad there are people like von Ahn and Hu to serve as matchmakers.
(Luis von Ahn, “3,2,1 Takeoff! And We’re Translating the Web!” Official Duolingo Blog, October 14, 2013; Chang Hu et al., “Translation by Iterative Collaboration between Monolingual Users,” University of Maryland Department of Computer Science, July 25, 2010; Chang Hu et al., “Deploying MonoTrans Widgets in the Wild,” University of Maryland, May 2012)