Clever Guy, Very Cool Job

Google do Tech Talks and publish them on Google Video. I found a very interesting talk called (video link) "Human Computation" about how they are going about the task of tagging / labelling images so that image searches can retrieve accurate results. The clip is over 50 minutes long, so I'll give a summary, but if you're interested it is quality stuff, definitely worth watching the whole thing. The solution they've come up with is to get people playing games that use image tagging as part of the game. Ingenious idea, well actually there are several creative solutions to problems in this talk. Clever guy Luis von Ahn, but they still spell his name wrong (Louise!?) in the subtitles.

In the video Luis talks about "captchas", which are those images of text that you have to read and copy in some web forms in order to prove that you're not a software agent trying to hack the form. He says that spam hackers have found a way to hack the forms, which is they pass the image of the text back to a porn website and interrupt users with a message saying "you must copy this text before you can continue watching". When the user enters the text it can be passed back to the form. So that's a hacker solution that gets humans cooperating with computers, albeit unawares. Luis uses games to involve people in the task.

So anyway, Luis and his team have built a game where an image is displayed on the screen. Two players are teamed together by the system and they get points when they both tag the image with the same word - thus tagging the image. The only communication between players is when they win a round because then they know the other player used the same tag. Several tags can be useful for one image, so Luis and co. have started to make some tags "taboo" for each image after it has been labelled with the same tag several times.

Luis says that the game has been very popular and a lot of images have been tagged. His team can now also use tagged image as a second level check that players are genuine. In theory a group of sabateurs could join the game together and respond with the same tag to every image. That would damage the accuracy of the tags if it was successful, so the team have started to include test pictures; images that most people label with the same tag; if players gets all these wrong, their tags may be treated as suspect.

There are several other cool solutions to sub-problems of this general area in the clip. That is one clever guy.