Corrections or additions?
Published in U.S. 1 Newspaper on June 14, 2000. All rights
Artificial IQ: Is it a Fair Test?
Karin Verspoor and other computer scientists at
Intelligenesis in New York City are discovering that creating
intelligence is a little bit like child-rearing. There a team of
are raising "WebMind," a software program that will
in the same abstract, relational, and creative way that humans do.
Like a parent coaching a child, Verspoor, who directs natural language
engineering, tries not to program WebMind on what to "think,"
but to give it the tools essential to thought itself. "You can
hard-code a lot of knowledge but then you’re not giving it the tools
to find and learn new information," says Verspoor. With a true
artificial intelligence system, says Verspoor, "you can ask
that can’t be answered just by spitting back an answer that you’ve
heard before, but are done by making connections, generalizing, the
kind of things that you and I take for granted."
Verspoor speaks on "A Distributed Architecture for Artificial
Intelligence," on Wednesday, June 21, at 6:30 p.m. at the
for Women In Science at Princeton University’s Engineering Quad, Room
217. Call 732-235-5048.
A speaker of both Dutch and English, Verspoor has a unique
with language. She earned her BA in computer science at Rice
Class of 1993, while simultaneously completing a degree in cognitive
science so she could incorporate elements of psychology, philosophy,
and linguistics into her work. She later earned a PhD in cognitive
science from the University of Edinburgh in Scotland.
Verspoor began working with Intelligenesis after a year-long residency
at the Microsoft Research Institute in Sidney, Australia, the
institute funded by Microsoft Australia. "The experience at the
research institute was great," says Verspoor, "because they
supported free research, no strings attached. We focused on natural
language generation problems, which is going from an underlying
of meaning to a surface form, to a sentence or paragraph that conveys
Talking linguistics gets tricky, but if you can imagine how long it
takes to completely verse a person in the English language — from
birth to early adulthood, let’s say — you get an idea of how
it is to get a machine to understand a human language. One of the
major challenges for programmers is getting the computer to move from
concrete to abstract concepts. "If you’re interacting with a child
and point to a green cup and say `cup’ over and over again, they will
figure out it’s a cup," says Verspoor. "But eventually they
move to the abstract notion of a cup, that it doesn’t have to be a
green cup in order for it to be a cup."
Programmers at Intelligenesis are attempting to endow WebMind with
that kind of human reasoning. "You could teach it that a cup is
distinct from a glass, and that’s done in traditional artificial
work," says Verspoor, "but there are two problems with that:
it’s actually hard to figure out all the important relationships,
and also it’s a lot of work to encode that by hand. Eventually you
hit a dead end unless you give the system the tools for learning
Once the concepts are in place, the next challenge for programmers
is getting the machine to translate those into language. "If you
have an abstract representation of a relationship, like you know that
a cat is similar to a lion," says Verspoor, "you still need
to work out how to basically use the grammatical constraints of the
language to produce a sentence."
Although still in its infancy, WebMind will be used for advanced
retrieval and numerical predictions that could change the way stocks
are traded. "It could predict whether the stock market is going
to go up or down on more than just numerical modeling — it will
look at what has happened in the past, psychological factors,"
Intelligenesis signed its first commercial partnership on June 7 with
NetCurrents, a company that analyzes communications from a universe
of targeted Internet locations and provides real time information
to its clients. The company chose to add WebMind’s artificial
functionality to its AgencyFacts Internet monitoring and analysis
service, which serves public and investor relations agencies with
information on most talked about topics, online sentiment, most active
posters, real-time percentage breakdown of volume on the most active
message boards, and the 10 most recent messages within minutes of
Even though WebMind will eventually think in much the same way humans
do, says Verspoor, its capacity will far exceed that of humans.
could keep track of every single word in a newspaper, which humans
obviously don’t do, since we abstract the knowledge and combine it
with other knowledge. It also has access to a lot more techniques
for finding new knowledge than any human. It can be an expert in
modeling, physics, and linguistics because it’s being programmed by
people who are experts in those areas."
But the problem with artificial intelligence, says Verspoor, is every
time you think you’ve achieved "intelligence" somebody else
redefines it. "It’s called the moving frontier," she says.
One of the first scientists to prophesy that computers would one day
think, Alan Turing, defined an "intelligent" machine as one
that could not be deciphered from a human in an interrogation.
experts today consider that definition naive. "It’s really not
a good test for an intelligent system because it doesn’t require the
system to learn," says Verspoor. "It just requires the system
to be able to simulate human conversations."
By those standards, people are already using artificial intelligence
today, says Verspoor. Speech-recognition systems, for example.
the technology is going to improve with time," she says.
by little, we’ll see more artificial intelligence coming into the
applications we’re using. But at some level, people always want to
see themselves as more intelligent than a machine."
Corrections or additions?
This page is published by PrincetonInfo.com
— the web site for U.S. 1 Newspaper in Princeton, New Jersey.