Umm… I’m pretty sure that Watson is using natural language speech recognition.
I’ve read the following type of comment consistently:
“…Watson is processing vast amounts of context-dependent information presented in natural human speech to arrive at an answer that is based on any number of factors. This is, in a word, absolutely groundbreaking. On a speech and context recognition front, in particular – turning questions that require multifaceted layers of knowledge into answers that rely on a number of variables from any number of sources. The analytics involved are mind-boggling.”
You are seriously underestimating the difficulty of the problem. Try Googling the questions, or anything, and see how often the “feeling lucky” button gives you something useful.
There was an article in the Times Magazine about this program. The two difficult pieces are first understanding the wordplay in the question, and second, deciding how confident it should be and ranking several possible answers. The raw data itself is not much of a problem, except in that you have to have a good way of organizing it.
I’ve been on Jeopardy - made it the first time I tried - and it had nothing to do with intelligence. I was born with very good data organization, so I can pull up obscure facts very quickly. That was the key, not brains.
In a similar way, computers don’t play good chess because of pure processing power, but rather because of good board evaluation functions and heuristics. The combinatorial complexity of chess is still way to big for brute force approaches.
The hard part of speech recognition is not recognizing the words - my Droid does that pretty well. It is understanding the semantics of the words. That is what the program is doing, and if it is using speech or text as input the problem is no different.
“…context-dependent information presented in natural human speech…”
I addressed a poster who was ignorant of the fact that natural-language speech recognition was used. The issue at hand was not to discuss nuances, context, etc. It was about pointing out that it wasn’t done via text input.
I’ve gone ahead and grabbed a quote that specifically mentions that it’s absolutely groudbreaking… on a speech and context recognition front, in particular.
Yes, your Droid recognizes words, but not natural-language speech. Just what Watson does here alone is amazing, but understanding the context – the nuances of Jeopardy answers – is one more thing that contributes to it being groundbreaking.
And the problem IS different if you are using natural-language speech recognition, because it WAS a problem for IBM to get Watson to work against all the problems simultaneously – again: groundbreaking.
Surely that gives Watson an advantage, if Watson is fed the text at the same time as it appears for the contestants then he has a head start whilst the contestants read the question. Clearly the longer the question the more chance Watson has of winning, he begins processing immediately and once 3 seconds have passed he is good to go, provided his confidence is high enough. Speaking of which I wonder what limit they have placed on the confidence level before an answer is given, i.e. will he answer at 50%, 60% or as low as 30%.
It doesn’t seem likely that Watson would be using speech recognition to capture the data, if he was it would take 3 seconds from the question being read out before he could answer, he would be beaten by the humans every time (provided they knew the answer).
Watson definitely does not get the questions using speech recognition, he gets them as plain text, as others have alluded to. I think some of the confusion on this issue has arisen because a great deal of research into natural language processing (not necessarily spoken language) went into the project.
For all those pointing out that Watson doesn’t get the questions by processing spoken language, neither do the human contestants, right? They see the question printed on the screen, and they can surely read much faster that Alex manages to say the words.
The state of OCR is such that requiring Watson to have a camera pointed at a television screen would not really add anything interesting to the achievement.
It seems to me more and more that Jeopardy should allow contestants to buzz in immediately after the question is started. We can still delay the answer until the question is read (for the home viewers), but making the competition be about who can buzz in the soonest past some arbitrary point isn’t that interesting.
One thing we have to nip in the bud right now. At some point they’ll tell us that this will lead to “thinking computers” that will do our work and give us more leisure time.
Ain’t gonna happen. They’ve been saying that automation will give us more leisure time for decades. Instead we have longer work hours and Blackberrys that connect us to the job 24-7.
We used to have a high school quiz show in Canada called Reach For The Top, where the rule was that you could buzz in before the question was finished, but if you did that and your answer was wrong, the penalty was doubled. So whether you buzzed in early or not would depend on how confident you were that you had the right answer.