On Watson's Eyes and Ears

by Edward Goldstick

A reader focuses on my observations:

Submission to Watson by text isn't really the advantage it might appear, or at least it's a "fair advantage."

First, it's essential to "see" the text for clues that aren't available by hearing. For example, if an item includes "the '20s" then the apostrophe is an important clue to the meaning. Or there might be quotes around words to signal titles, puns and wordplay. Humans receive these clues from seeing the written question. If Watson were forced to use speech-recognition, it would be at a clear disadvantage. So, text submission seems only fair. If Watson were required to play via voice recognition, then the written text should be removed from the game for everyone. But that's not how the game is set up.

While I agree that speech-recognition would be a burden for the machine, that is probably so for the human players as well; however, I had presumed the machine would be "seeing" the text optically rather than parsing it in text form directly (and none of this was explained clearly on NOVA ScienceNow or in the background clips during the show which were my only references...). On the other hand, it is conceivable that the pattern/form recognition advantage offered by visualizing the entire text in an instant is, in that case, an advantage for the human contestants, so it might well be a wash.

Second, I've read comments that the text transmission should be delayed. This also is wrong. If it were delayed until the host finished reading the text aloud, for instance, this ignores the human players' ability to read and begin processing the item far more quickly than it's spoken -- so this would be an unfair delay. As it is, and should be, humans and Watson both get the complete text at the same moment. Sure, Watson is presumably analyzing the text in its entirety while the humans are still reading, but isn't that part of the point?  Watson has advantages (in this case, speed) while humans have others and the challenge is how they match up. If in an all-human game it were somehow discovered that one player could read items nearly instantaneously while others typically needed 2 seconds, it would be ridiculous to suggest that the faster player be penalized.

I'm not as curious as you about the idea of imposing on Watson a fixed half-second response delay. Let's assume that a half-second is a fair/accurate measure of the time between the instant a player decides to buzz in, and the recording of his button-push (which it might not be, but let's go with it). Can a human player reassess his decision in the time it takes to press the button, and stop? Certainly, there's nothing preventing him. So then, how would Watson be offered the same ability to rescind the decision to answer, given another half-second of processing time?

The notion of a 'delay' that I offered was only with regard to the latency in the button push, not in that leading to the "Aha!" reflex that triggers the process. In fact, my sense was that all three contestants knew the answer to all the questions except those very few that none of them got very quickly [thanks Jim for making this obvious because I hadn't realized that I perceived it until I got your note.] ... and in the end, I would have been interested to see if Ken and Brad had pushed their buttons just moments behind Watson.

Here's what I'd be curious about: You surmise that Ken Jennings is adapting his strategy to risk quicker "buttons" and then using the time available for his answer to finish his analysis. He'd first assess that he's strong in a category, and then assess his chances on a particular question from a quick impression of key words, and buzz in. (Perhaps he'd also, on some level, mitigate this strategy based on the dollar value/penalty of the question and his standing in the game at the time.) I'd be curious to know if Watson does the same, or could. That is -- buzz in to answer before having found an answer of sufficient probability to do so, but based on an initial assessment of success within the time allowed actually to respond.

Anyway, it's fun stuff.

Agreed, though not without a bit of disappointment, if I may, because it was truly impressive how outmatched these two fact-filled fellows seemed to be ... so for now, bravo IBMers*!!!

... but now, onto Day Two!

*I wonder if Watson could recognize a word like that?

Edward Goldstick is a veteran of the high-tech, software, defense, and energy-technology worlds in the U.S. and France.