Earlier this week the quiz gameshow Jeopardy! had a new champion – a computer named Watson. It defeated former champions Ken Jennings and Brad Rutter with ease.
Watson created a wave of good PR for IBM, its developer. And rightfully so. I think it is indeed a big deal.
Before I get to why I think Watson is a big deal, take a minute and read this Wired article on how Watson was set up for Jeopardy!
The computer is fed the answer in text form at the same time the answer panel appears to the two human players. Watson then queries its database for an appropriate question response, a process that doesn’t involve using the internet at all.
…Watson then must push a physical buzzer to respond, just like its human competitors.
They went to great pains to level the playing field. So it would seem that Watson truly bested its human competitors. So how did it win?
I believe speed was the biggest contributing factor to Watson’s performance. There were many questions to which it got the wrong answers. But the one’s that it knew, it just beat Rutter and Jennings to the buzzer.
Jeopardy! allows contestants to hit the buzzer only after a light comes on. They are penalized by a fraction of a second’s delay if they hit the buzzer before the light comes on. Champion players like Rutter and Jennings, anticipate the light and hit it almost immediately after. But they still have a disadvantage over Watson.
Watson, on the other hand, does not anticipate the light, but has a weighted scheme that allows it, when it is highly confident, to hit the buzzer in as little as 10 milliseconds, making it very hard for humans to beat. When it was less confident, it took longer to buzz in. In the second round, Watson beat the others to the buzzer in 24 out of 30 Double Jeopardy questions.
I couldn’t tell you why IBM made Watson take longer to buzz if Watson was less confident. Perhaps, taking a bit longer would help it arrive at a better answer. Or perhaps IBM wanted to handicap Watson just a little so that it appeared like it was a fair contest.
Watson would also have read the question faster than its human competitors. However, human beings process as they read. Watson would have read the question instantaneously and then processed it. So its not strictly comparable.
But Watson’s speed advantage would have been worth nothing without its ability to understand the question and find the right answer. And that’s where the magic of Watson lies.
Watson could understand natural language. It could understand the question. It could also understand the encyclopedias and other content that went into its database. To make that happen is very, very difficult computer science.
If you look at the Algorithms team at IBM that worked on Watson, you will sense that a lot of natural language processing techniques went into the development of Watson. Probably a whole bunch of statistical methods as well. And some serious brains.
Watson is a big deal because it changes the paradigm of the human/computer interface. Today, human beings interact with computers by giving it very precise instructions in a limited syntax. Like clicking on a button on Facebook, or running a program in Python. Even if you say something in a natural language (like English, not Python) like what you might type into a search box, the computer does something reasonably well defined with the words. It doesn’t try to understand what you said.
Take Google for instance. You type in keywords into Google search and it will return links to pages with those keywords.
You might enter a question “Is Preity Zinta married?” and Google will return links to pages which have these words on the page. After a few clicks you might conclude that she was expected to marry Ness Wadia, but then broke up. But you would have click on some of the links to figure that out for sure.
Google search does not understand your question. It simply takes it as a “bag of words” and pulls up pages that contain those words. It does many clever things to determine how to rank the results. But to Google “Zinta Preity Married”, which doesn’t mean much, is the same (more or less) as our question above. It does not try to understand the question.
Nor does it answer your question directly. It makes no inferences from the content it puts on the search results page. It just brings you as close as possible to the answer by pulling up the most relevant pages and within them the most relevant snippet on the search page. Good as it is, it’s no Watson.
Would we benefit if Google search understood our question, or at least our intent a little more? Sure. And Google has been making steady progress on this front already.
If you enter “time in India” in the search box, Google returns links to pages with those key words. But right at the top, it just gives you the time in India. It assumes that you are probably looking for the current time in India. Similarly if you look for CX870, the top result will be the flight status of a Cathay Pacific flight from Hong Kong to San Francisco. If you enter 2+2, it doesn’t even bother looking for the keywords. It just returns the answer viz. 4.
These were use cases where you were trying to “find” some information. What if you wanted action, rather than information. On my Macbook I use Spotlight to type in the first few letters of an application and hit enter. The application fires up. You might call this a key word based action. Pretty basic.
In Google Calendar, there is a neat feature called Quick Add. In Quick Add, I can write “Lunch with Gaurav at 12pm on Monday” and the Calendar will add an event called Lunch with Gaurav, from 12 to 1 on my calendar for next Monday. It accepts input in a variety of forms and has a lot of flexibility. Just as my human assistant would understand me if I told her to schedule me with Gaurav at 12pm next Monday, in any of a dozen ways.
But that is as far as it gets as far as natural language processing goes in every day applications. Watson takes us well beyond that.
Human language is one of the most complex things about human beings. They are biologically wired for language. All languages have basic underlying constructs that are universal. Our language separates us from other living beings. And from machines.
And that is why Watson is a big deal. It is an important step in the journey of getting a machine to understand human language. And yes, making it work in the limited context of Jeopardy! is a far cry from making it a general purpose semantic search engine. And maybe one shouldn’t even expect that. I don’t understand half the things I see on Twitter, myself, so how can Watson? Nevertheless, it shows us the possibilities. And aspiration is what drives human endeavor, doesn’t it?
thanks, basab, for this. ben zimmer has a take in the atlantic, and later, on language log.
We stumbled over here different page and thought I might as well check things out. I like what I see so now i am following you. Look forward to looking over your web page yet again.