Some observations from watching the Science channel special and the three Jeopardy episodes:
Watson had two great advantages: access to vast amounts of information and extreme processing speed to parse it all. Given that, I don't think its performance was all that impressive and certainly not surprising. At the heart of the software is simply looking up keywords from the question and finding what other keywords pop up in the resulting documents most frequently that might correspond to answers. Now, identifying the keywords and figuring out what subject the answer should be (person, title, etc) is no easy task and they did use some machine learning techniques (I wonder which?) to train the computer to find patterns using old questions. That's neat, but not exactly revolutionary. Machine learning using neural networks, genetic algorithms, etc. have been around for decades. And even without training, I think Watson would do fairly well on some of the simpler questions (the ones with lots of obvious keywords).
It was easy to tell right away which questions it would have no problem with and which questions would be problematic. And despite all the training, it still struggled with some of the categories or unusually phrased questions. The US city with the airports is a good example. How do you look that up? The key there was to associate the two airports with the war hero/battle (using "named after" which is a problematic key phrase) and then to associate the two airports with the city they were in. All that requires some substantial logic.
But the computer won not because it could find a lot of the answers, but because it could find them faster than a human could read and understand the question, recall the answer, and press the buzzer. And that is directly related to its parallel processing power. Had they put even more processors in there, it would have probably won by even a wider margin.
