On 11/01/12 18:19, Cristian Consonni wrote:
ReCAPTCHA already works in a way similar to this. Two words are presented but only one is known and actually serves to filtrate accesses. They then collect answers for both words and if the test on the first is passed (which indicates a human) then the answer for the second is recorded. When a certain number of people agree on the transcription of a previously unknown word then that transcription is taken as good and used in future as a filter word.
I know. I was thinking on the basis that being so different, people could cheat by giving a bad 'learning' word. Originally, recaptcha showed two images apparently equal, although recently it seems to have changed, and it's clear which is the test and which the unknown. Although we could even plainly make the second one optional, letting the users choose if they want to help or not. We would get less trains but of higher quality.
I believe the trickiest part is creating a system to put results back in Wikisource in a semi-automated way, but having "captcha reviewers" may help.
I was just 'defering' that :)