- People have the answers (Image: Reiner Riedler/Anzenberger/ Eyevine)
- mg22029464.400-1_300[1].jpg (18.41 KiB) Viewed 2404 times
DataSift is new kind of search engine that uses crowdsourced human intelligence to answer vague, complex or visual questions, even when the users are not sure what they are searching for.
"Take someone like my mother," says its developer Aditya Parameswaran of the University of Illinois at Urbana-Champaign. "She has recently purchased this piece of electronic equipment which has a port that she has no idea what to do with. With DataSift, she can just take a photo of the port and say 'find me the cable that works with this'."
DataSift works by breaking down a query into components that can be answered easily and quickly by human workers on crowdsourcing sites such as Amazon's Mechanical Turk. Humans are much better than machines at answering questions based on images, complex language or that are ambiguous.
If you were looking for a new flat that was pet-friendly and within walking distance of a train station, for example, your internet search would take a few hours at least, says Parameswaran. "You can send it off and DataSift will give you results in 20 minutes," she says.
But humans are unreliable. That's why DataSift runs a web search on every human answer. Say a user asks the software to identify the building in an image – for example, the Eiffel Tower. Humans make suggestions, the software runs a web search to checks these suggestions and then feeds the search results back for a new set of humans to confirm.
This confirmation step means that in a crowd of thousands, the majority can be wrong and DataSift will still filter the correct answers to the top. Parameswaran presented DataSift at a human computation conference in Palm Springs, California, last month.
"Crowdsourcing is essential because people are still much better than machines at interpreting natural human language," says Yubin Kim of Carnegie Mellon University in Pennsylvania, who is working on a similar system.
This article appeared in print under the headline "Google can't find what you're looking for? Ask a human"