Thursday, February 17, 2011

Q&A: The meaning of Watson

By Todd Bishop, Seattle TechFlash

The folks at the University of Washington Computer Science and Engineering Department have been watching IBM’s Watson computer on “Jeopardy!” this week with even more interest than the rest of us. They even hosted a viewing party last night on campus, followed by expert commentary.

We followed up today with one of those experts, computer science professor Oren Etzioni, the director of the UW Turing Center, to get his take on the whole thing. Continue reading for excerpts from our conversation.

Who are you rooting for? Man or machine?
My perspective is that, if Watson wins, it’s a victory for humans. I don’t think that human intelligence and human uniqueness is in any way in jeopardy here, if you pardon the pun. This is still a very narrow task. It’s much more exciting in certain ways than chess because of the ambiguity, because of the broad scope, because it’s natural language. But it’s a far cry from the day when supposedly robots may deign to keep us as pets. In that context, the question is, can this type of technology that IBM and other people in the research community have been developing, be used for better medical diagnosis, for better search, and lots of other applications? And if so, that’s fantastic. There’s a team at IBM that has worked day and night to achieve this amazing feat, so it’s a great victory for them, as well.

Watson just cleaned their clocks last night. What did you think of the result?
It didn’t surprise me that Watson won. In fairness there are some small things in the game that give it an advantage, like the time that it has to process -- it sees the clue instantly, it gets transmitted via text, and then it’s told when it can buzz in. So it has a number of these small advantages and particularly last night it was often winning on time. Ken Jennings knew the answer but he just didn’t buzz in time. And that’s a place where the mechanics give Watson somewhat of an advantage. But the thing that I’m not surprised by at all is the power of using both large amounts of computation and large amounts of memory and this statistical computation to analyze text. What I believe we’ll see over the next year or two, let alone three to five, is more and more sophisticated text processing over the bodies of text that we know and love, which is text on the web, text on Twitter, things like that.

What do you think would have happened if, instead of getting a text file, he had to listen to Alex Trebek’s voice and analyze the speech?
Actually, Alex is such a clear speaker, and speech-processing technology is so good, the accuracy of mapping from speech to text would be extremely high. It would give it very little disadvantage. But it is slower, it takes a few seconds, and so it would have cost dearly in terms of its ability to respond in a timely fashion.

You’re calling Watson “it”, and I’m calling it “he.” How do you feel about applying gender to machines?
It’s very natural in these situations to anthropomorphize it. People last night, and when people were watching and analyzing Deep Blue, did that. I know that it’s a very natural tendency. It’s important to remember that it’s an it, not a he or a she. By the way, that’s one of its advantages -- it doesn’t experience frustration, it doesn’t get distracted, it’s just mindlessly executing its computations.

Why couldn’t somebody like IBM just take Watson and turn it into the ultimate search machine?
In fact, that was a topic of discussion last night at UW -- what’s harder, playing Jeopardy! or doing search. There was some debate. I do think that the particular Watson program is tuned to Jeopardy. But what it was showing is the potential for using this kind of shallow language understanding for things like search, for question-answering. Furthermore, if you think, for example, about mobile -- we’re more and more interacting with the web and with information from mobile devices, and typically they have a small screen -- in that context, question-answering makes a lot of sense. You can get answers from people -- Quora and Aardvark, Yahoo Answers -- or you can get it automatically from machines. And I think we will see this paradigm continue to grow.

What has surprised you about Watson so far?
I do think the level of performance that it exhibited, even if you think, yeah, this should be doable, to actually see it in action is impressive. The tremendous speed is also impressive. Being able to pull in so many facts on such a range of topics in under three seconds is remarkable. The speed, the diversity and the quality were surprising even to somebody who’s immersed in it like myself.

What about last night’s Final Jeopardy question?
The thing that was quite interesting is that it was asked about a U.S. city and it came up with Toronto. That’s a great example of how it was using statistics. It knows vaguely that Toronto is loosely speaking associated with the U.S. But it’s not actually able to employ hard constraints, and say, OK, this is in the U.S., this isn’t. That’s what is happening when you’re dealing with text.

Implications for the future of technology and startups? Over the next few years, I think what we will see is more and more adept and sophisticated extraction of useful information from text.

(This article originally appeared Wednesday on TechFlash, Seattle’s Technology News Source)