IBM computer to compete against humans in Jeopardy

Soldato
Joined
1 Jan 2008
Posts
11,046
AI has been doing this for the last 60 years!
E.g., some origional work from 1943!

McCulloch, W. and Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics, 7:115 - 133.

It may have been, but it can't be that great can it, or i'd have a robot doing my reports for me!
 
Soldato
OP
Joined
9 Nov 2003
Posts
9,510
Location
The Motor City
Theres no voice recognition, they are giving the computer text versions of the question.
Yes. But how do you interpret a question such as the one youtubed above? Even Ken Jennings, the most winning contestant on any gameshow ever (75 straight wins, 2.3 million dollars), blew it. So many of Jeopardy's questions are like that. Would the computer ever hit the buzzer? If it waited for a losing response, would it correctly identify the winning response? That would require two-times, at least, the research on it's own database to come up with an answer.

Personally, I don't see it happening. Not in our lifetime.

It is quite the challenge, for sure. I think Google's team will find it first.
 
Associate
Joined
25 Aug 2005
Posts
2,166
having a quick look at the question structure. It would seem the most logical way to program the computer would be similar to the 20 questions thing. A big database of information that determines an object from the answers to its questions.

www.20q.net

So i'd assume the computer would pick out key words from the questions and link those words into a database to find the most likely match. So if a question starts with "She" it would automatically narrow the answers down to female person/character. And so on and so forth.
 
Last edited:
Soldato
Joined
7 Jun 2004
Posts
3,534
So i'd assume the computer would pick out key words from the questions and link those words into a database to find the most likely match. So if a question starts with "She" it would automatically narrow the answers down to female person/character. And so on and so forth.

Getting the complicated questions right is much more difficult than matching keywords. If they want the computer to win against a good opponent it will need to have a pretty good semantic understanding of the questions. Translating a natural language question into a database query, and then a document that contains the answer into the correct single word - that's the really challenging problem.

I'm not sure that voice recognition or OCR is going to come into this - it's a completely different task to the AI buisness of answering the questions and if the question is shown in a fixed font on the screen then it's already a solved problem anyway.
 
Back
Top Bottom