Prove You're Human Online With New Language Test
CREDIT: Public domain
Anyone who signs up for an online service or buys something over the Internet usually takes a tiny Turing test along the way, to distinguish him from an automated computer program. Just before users click “Submit,” many websites ask people to type the contents of a smudgy-looking series of letters. Those letters, called a CAPTCHA, are more difficult for spam programs to decipher than for people.
Over the past few years, however, programs have gotten better at solving CAPTCHAs, allowing spammers to sign up for multiple email accounts and post dubious links in the comments sections of websites. So a team of computer scientists is proposing a new CAPTCHA system that takes aim at a different computer vulnerability – programs’ awkwardness in translating between languages. At the same time, FluTCHA records data about how people improve translations, which scientists can use to improve translating programs.
Three computer scientists from the University of Maryland wrote a program that gathers Japanese news stories that Google Translate has rendered in English.
The results are awkward sometimes: “After the incident the captain, was placed under local government oversight, not even fishing.”
FluTCHA highlights the awkward phrase – “not even fishing” – and asks users who are native English speakers to fix it. In this example, someone might write, “and was not even allowed to fish.”
The answer goes to a group of human graders on Amazon’s Mechanical Turk service, an online job exchange where members post and perform small tasks. Graders evaluate how fluent the fixes are and how well the fixes match the original sentence in meaning, using a nine-point scale.
If the grade matches, FluTCHA lets the user through.
In a study with five users and 83 workers, FluTCHA’s creators found their system lets humans through 86 percent of the time and lets computer-generated fixes through 18 percent of the time. Ninety-eight percent of humans should be able to pass FluTCHA after two tries, the researchers calculated.
Theoretically, spam programs can solve regular CAPTCHA systems 100 percent of the time, though in actual practice, their success rate is much lower, the scientists wrote in a paper they presented at the Association for Computing Machinery’s computer interaction conference in early May. They did not say how often computer programs solve CAPTCHAs in practice.
News websites generate new content everyday, so they provide an endless source of puzzles for FluTCHA, the scientists wrote. At the same time, because the correct answers come from humans, a spam program can’t cheat by finding the original source of correct, fluent phrases, which can happen with another fluency-based CAPTCHA system called SS-CAPTCHA.
One drawback is that FluTCHA needs graders who are ready to work 24 hours a day. Website customers will get frustrated if they can’t get through a service because there’s no one there to grade their fixes. Mechanical Turk and similar services draw workers from all over the world, however, so there often is someone up and working at all times. The researchers also suggested using Quik Turkit, a service for getting nearly real-time responses on Mechanical Turk.
As FluTCHA works, it gathers data about how humans edit machine translations. Other scientists can use that data to train programs such as Google Translate to work better, FluTCHA’s makers wrote. That makes FluTCHA a deliberately temporary solution. It compiles the data it needs to make itself obsolete as it works. When that day comes, FluTCHA will have succeeded in a different way, the FluTCHA scientists wrote. It’ll have contributed to making computer programs act more naturally and more human and ever more difficult for a Turing test to distinguish.