« Midweek Mystery | Main | Don't shake hands! (Revised) »
October 23, 2003
Confuse-a-bot
Steve over at Little Tiny Lies tried to hold a conversation with a bot. In the comments, Aaron suggests a couple of ways to confuse bots (or at least to detect them) including the 'mis-information trick' and the 'English complexity trick' both of which can be resolved by (very) careful coding or a restriction on the subject matter. However, the article (and comments) reminded me of something I read many years ago in Cognitive Science where the discussion was about methods for defeating a turing test. The idea was to come up with something that was simple but effective and impossible to code against. The solution (sorry but I can't recall the authors of the paper) was to use comparative terms where you get ratings. One example went like this:
On a scale from 1 to 10 (1 is terrible, 10 is brilliant) rate the name "FLUGLY" as a screen name for an actress. On the same scale rate the name "FLUGLY" as applied to a childs' soft stuffed toy.
The deal is that humans give very consistent ratings (negative for the actress, positive for the toy) whereas the computer will have to make wild-ass guesses. Because of possible scoring problems, the authors went on to expand the idea and derive another test that involves categorisation.
There are two categories: "ping" and "pong". You need to decide which category the following words belong to: "needle", "basketball", "Eiffel tower", "wet dog".
Note that shape itself is not a valid solution but surprisingly humans get this right while the poor old computer is once again guessing.
Any other suggestions?
Posted by Ozguru at October 23, 2003 09:10 AM
Comments
Posted by: Kathy K at October 23, 2003 09:10 AM
Posted by: ozguru at October 23, 2003 09:10 AM
Posted by: Paul Jané at October 23, 2003 09:10 AM
Posted by: Paul Jané at October 23, 2003 09:10 AM