How a machine learning algorithm beat the assembled masses of Mechanical Turk.
Technology Review, Christopher Mims 12/20/2010
Amazon's Mechancial Turk is the ultimate in nearly anonymous outsourcing: any task that can be completed online can be accomplished by the combination of automated marketplace and human labor. Those who sign up to complete tasks - Turkers - are paid wages as low as pennies per chore to do everything from data entry to folk art.
Mechanical Turk is designed to complete tasks that are easy for humans and hard for machines, such as categorizing or identifying the content of images. The problem for Amazon and all its imitators, however, is that machines are getting better at many tasks, while the humans on Mechanical Turk, for reasons I'll explore in tomorrow's post, are getting worse.
Recently, for example, researchers working at the online review site Yelp released a paper ( Towards Building a High-QualityWorkforce with Mechanical Turk, pdf) on their experience matching thousands of Mechanical Turkers against a supervised learning algorithm.
The results weren't pretty: in order to find a population of Turkers whose work was passable, the researchers first used Mechanical Turk to administer a test to 4,660 applicants. It was a multiple choice test to determine whether or not a Turker could identify the correct category for a business (Restaurant, Shopping, etc.) and verify, via its official website or by phone, its correct phone number and address.
79 passed. This was an extremely basic multiple choice test. It makes one wonder how the other 4,581 were smart enough to operate a web browser in the first place.
These 79 "high quality" workers were then thrown at the problem of verifying business information three at a time. This allowed the researchers to take only the results that a simple majority of Turkers agreed were correct, or in some cases to take the result chosen by the Turker who had historically been the most accurate.
Researchers threw a "Naive Bayes classifier" at the same set of problems. This is a kind of supervised learning algorithm; one that, according to a 2006 comparison of these systems, isn't even the best kind out there.
The Bayes classifier won handily.
In almost every case, the algorithm, which was trained on a pool of 12 million user-submitted Yelp reviews, correctly identified the category of a business a third more often than the humans. In the automotive category, the computer was twice as likely as the assembled masses to correctly identify a business.