ALBERT has been crowned the world鈥檚 most thoughtful computer program. Written
by Robby Garner, a programmer from Georgia, it has won the 1999 Loebner prize.
The contest is based on the most famous computer intelligence challenge, the
Turing test.
Since the dawn of computing, people have wondered whether machines could
think. In 1950, Alan Turing, one of the founders of computer science, dreamt up
an objective test he called the 鈥渋mitation game鈥�. A person and a machine each
communicate with an inquisitor, who must decide which is the human and which is
the computer. If they guess wrong half the time, the computer is essentially
indistinguishable from a human and passes Turing鈥檚 test.
At Flinders University in Adelaide last week, programs were pitted against
humans in a series of Turing tests sponsored by the New York philanthropist Hugh
Loebner. Eleven judges were asked to distinguish between six programs and five
humans by conducting brief conversations on terminals. But none of the programs
was mistaken for a human more than half the time, as a 鈥渢hinking鈥� computer
should. Even Albert was judged human only 11 per cent of the time. This was a
worse result than that achieved by the person whose behaviour was judged to be
the most 鈥渞obotic鈥� of the five, who was deemed more human than a human partner
34 per cent of the time.
Advertisement
There are two possible approaches for these programs, says Thomas Whalen of
the Communications Research Center in Ottawa, who won the Loebner prize in 1994.
鈥淥ne is the Eliza approach, where you try to have fairly clever responses based
on syntactic tricks like rearranging sentences, and relying heavily upon
non-sequiturs and smart-alecky responses,鈥� he says. Eliza is a classic program
designed to mimic a psychotherapist by picking out key words from a person鈥檚
answers and using them as the basis for the next question. The second approach
is 鈥渢rying to figure out what the question is actually saying and returning
颈苍蹿辞谤尘补迟颈辞苍鈥�.
Whalen complains that the judges seemed to be impressed by smart-alecky
responses, and were 鈥渓ess interested in actually answering questions鈥�. Albert
showed much of this smart-alecky behaviour (see Panel).
鈥淟anguage really depends upon knowing about the world,鈥� says Whalen. 鈥淥ur
brains do that analysis so well and so easily that we miss how hard it is.鈥�
Others dismiss the Turing test. Marvin Minsky, an artificial intelligence
expert at the Massachusetts Institute of Technology, says it鈥檚 simply 鈥渁 test to
see how easily a person can be fooled鈥�.