How Greg Lindsay Beat IBM's Watson at Jeopardy(fastcompany.com)
fastcompany.com
How Greg Lindsay Beat IBM's Watson at Jeopardy
http://www.fastcompany.com/1726969/how-i-beat-ibms-watson-at-jeopardy-3-times
7 comments
Even if Watson doesn't win, the fact that strong players think of him as a worthy opponent is a huge step forward for the popular perception of AI.
>So I had to steer him into categories full of what I called "semantic difficulty"--where the clues’ wordplay would trip him up.
This doesn't seem as useful as it's written here -- it's only useful under two cases: 1. Not all the clues are chosen before time runs out. 2. The money you gain here is wagered on a Daily Double, or the money Watson loses here is not able to be wagered on the Daily Double.
Contestants aren't able to make up categories, to create ones that would cause Watson trouble; they have to pick from the six on the board.
This doesn't seem as useful as it's written here -- it's only useful under two cases: 1. Not all the clues are chosen before time runs out. 2. The money you gain here is wagered on a Daily Double, or the money Watson loses here is not able to be wagered on the Daily Double.
Contestants aren't able to make up categories, to create ones that would cause Watson trouble; they have to pick from the six on the board.
After having watched the first half of game one: Another interesting category that Watson seemed to have a lot of trouble with was a category where all the answers were decades (1910s, 1920s, etc.). The top answers that Watson came up with were always years, but he was never that confident, and he never caught on that the answers were supposed to be a span of years.
I've read that Watson may ignore the category titles as they are often puns or confusing text that doesn't help him find an answer. Rather, I think, they may been seen as more of a source of confusion than help.
It did figure it out eventually, on the one where it answered the 20's incorrectly after Ken already did.
The first thing I noticed last night was Watson's strategy of sweeping the questions across instead of down like is normally played. My instinct tells me that humans get more context out of the category name and feel comfortable sticking to one category at a time.
If this machine is already so good, why doesn't IBM offer a system based on it? (honest question, not suggesting it's not good) It could decimate library research work, or legal research, by providing high-quality suggestions for simple phrases.
They are planning on doing just that. They'll be selling Watson-type QA machines to business for a few million dollars, starting this summer I think. I'm not 100% sure about the price and timeline, but they're definitely planning on making a product out of this.
I wouldn't be surprised if Watson leaves the contestants in tears this time - not because of some sadness at losing to a machine, but the sheer intellectual drain of a large number of very quick decisions based on intuition - which is what I guess it will take to beat the machine.
This was always a winning strategy. Be quick on the buzzer and use the time you have to come up with an answer. It's the way Ken Jennings kept winning.
All those games Watson played during the last year were played with the express purpose of improving Watson.