Boris Smus

interaction engineering

Crowdsourcing articles with mechanical turk

Last semester at CMU, I was involved in a research project involving Mechanical Turk. The goal was to get Mechanical Turk users (turkers) to collaborate on creating online wikipedia-style articles. Prior to my team's involvement, an undergraduate created a mediawiki-based platform to allow turkers to collaborate on articles. Despite a high compensation, few turkers completed the task. My team tackled the problem and came up with some interesting videos on the way.

We began by conducting contextual interviews with turkers living in Pittsburgh, all of whom rather unexpectedly, were female. The general takeaway was clear: turkers are used to very short and repetitive tasks, but article creation requires a prolonged period of concentration. Our solution was to significantly tweak the task, making it seem less arduous. In addition to simplifying the HIT's flow, we switched from mediawiki to etherpad as the article editing and collaboration platform. As a result of these changes, we were able to churn out turker-created articles on a given topic for under ten dollars. Here's a video of turkers collaborating on an article about Halloween:

We started out by creating an etherpad instance with a simple paragraph about the topic, as well as some article quality guidelines. Next, we created a series of Mechanical Turk HITs referencing the etherpad instance's URL. We paid our turkers a quarter up front for accepting the task, and provided a nickel (up to one dollar) every time they returned to edit the pad. We had no good way to verify the bonus mechanism, so we generally gave out the maximum bonus to every active collaborator. Here's the evolution of an article on Windows 7:

Watching the etherpad explode in color as multiple turkers simultaneously edit the same article is still mesmerizing. Though the quality of the articles was quite low, it generally increased with each turker's successive pass. Also it's worth noting that errors that we deliberately inserted in the starting paragraph as well as in real time were swiftly edited out. Not much quantitative analysis of this collaboration data has been done yet, though there are plans to conduct more scientific experiments in the future.

Several other researchers have been conducting interesting studies on mturk. Greg Little's work at MIT generated an interesting project called TurKit, intended to simplify setting up experiments such as the one outlined above. Panos Ipeirotis at NYU runs a variety of turk experiments as well as an mturk statistics monitor, which continually scrapes Mechanical Turk and generates summary data. Most recently, Jennifer Boriss surveyed the Turk community about their browser preferences projecting a growing interest in Chrome.

These varied Mechanical Turk projects represent only a small fraction of the potential of crowd-sourced marketplaces. It's especially interesting to take complex tasks, break them down into turk-sized morsels, and recombine them again. To improve the article collaboration scenario discussed here, one could provide an outline of an article and task turkers to elaborate on each point. This seems to be exactly what Greg's group is doing in an collaborative essay writing experiment. Such an approach may also be effectively applicable to crowd-sourced software development, which I hope to explore in the near future. Do you know other interesting projects and resources related to Mechanical Turk? If so, please respond below!