I just created an account at IMDB (because irritatingly they don't let you even read user comments unless you sign up) and noticed their image based sign-up challenge went one up on other systems I tried. I faced the usual distorted image of some letters and maybe numbers, the point was I couldn't really tell, it looked kind of like mobonna or it could have been mo6onna. However they sneakily relied on the fact that the word was a movie or movie star name. So in fact it was "madonna" - something no automated OCR code could get right first time.
Its clever to use the context and human domain knowledge to disambiguate possible interpretation of the image. Not that this is fool proof - I can imagine if someone really wanted to hack into IMDB and create false accounts they would quickly code up something to do a fuzzy match of possible character strings against movie and movie star names. In addition you know that IMDB can only use a small number of move names and movies stars because they want to make it as universal as possible. That may actually make the job of the sign-up spammer easier.
Still, it is an interesting direction to move in. I'm sure there are other more reliable fuzzy match tactics that could be used in this never ending war against spammers.
Subscribe to:
Post Comments (Atom)
0 comments:
Post a Comment