[ExI] Watson on NOVA

[Kelly Anderson]

On Sun, Feb 13, 2011 at 2:14 AM, Amara D. Angelica <amara at kurzweilai.net> wrote:
> Kelly, I had similar questions, so I interviewed an IBM Watson research
> manager. Please see if this helps:
> http://www.kurzweilai.net/how-watson-works-a-conversation-with-eric-brown-ib
> m-research-manager. I would be interested in any critiques of this, or
> questions for a follow-up interview.

"open, pluggable architecture of analytics" sounds like it has an
engine, and can add heuristics. If that's the case, then this is a
pretty powerful core technology, but it requires that it be "built and
tuned to play Jeopardy!" So if I were going to ask follow up
questions, I would ask some along these lines...

On the NOVA show it talked about adding gender information... is this
one of the pluggable pieces you are referring to?

When you say "open" do you mean open source? Or open for purchasers of
the system to augment?

Is this going to be available in a cloud configuration anytime soon?

Tell us more about "building" and "tuning"... It appears from the NOVA
show that it took 4 years to build and tune the system for Jeopardy,
how much effort would it take to build and tune a system for medical
diagnosis? Or build a technical support database for say Microsoft
Word.

It seems that the natural language processing of the questions and
categories is very extensive and uses a kind of search tree technology
reminiscent of AI search trees used in games such as chess. Is that
correct?

Tell us more about the index that is build a priori of the raw data
that the answers are sought from. Is it indexed, or is there just a
brute force algorithm based on keyword searches and then further
statistical processing of the results of the keyword search. In other
words, what's done prior to the question being asked on the index side
of the equation?

(I'm sure you could make that question shorter... :-)

You talk about Watson "learning", is the learning on the side of
understanding the question, finding the answer or both? Are you using
neural networks, statistical approaches, or some new approach for
that?

If developers wanted to build and tune their own solutions on this
architecture, how soon do you think it will be available? Is there a
business unit working on this yet?

Are there going to be any papers published by the Watson team?

I'm sure I could come up with more questions... but those would be
among the ones I would ask first I think.

-Kelly