[Mahh]

Some cool stuff here: http://www-943.ibm.com/innovation/us/watson/

The real challenge behind Watson is the natural language parsing. Instead of abstracting information away from their sources(like a graph), sources seem to have been left intact in sentences in Watson's memory. Watson would read through this information in a way alike to how it interprets a question, and it would try to create links and possible answers based on connections in sentences from many sources(this gives thought on why pun questions are difficult for Watson). I can't speak on behalf of the mathematical implementation of the answer choices, but this is the high level way that Watson finds answers. Those videos talk about the cool stuff behind the algorithmic challenges of Watson.

[spitfire]

You aren't joking. I took a few minutes and wrote up a bayesian engine in mathematica. I've got a pretty good start on that already, and as the IBM stuff notes it's embarrassingly parallel. It seems to me the entire problem is parsing. If you can parse well and feed a well formed input to your data layer (and you've fed it enough data) you're golden.

So who wants to build a real Q/A site based on this? Call it hal-18000.

[paganel]

> So who wants to build a real Q/A site based on this? Call it hal-18000.

You'd have to learn it to deal with thick accents like this one: http://www.youtube.com/watch?v=5FFRoYhTJQQ . Honestly, I don't know if that's possible, no matter how much training you'd put into the machine.

[khafra]

Natural Language Processing != Voice Recognition.

[paganel]

> Natural Language Processing != Voice Recognition

This is what I don't get, why should be "language processing" tied to written text? Part of the answer I know, because it's easier for computers to parse, but other than that it doesn't make sense.

[gloob]

Speech recognition is speech recognition. Different problem entirely.