[ The Poor
Conversationalist | The Construction Worker ]
ne
of the holy grails of AI and to pass the Turing test is understanding what a person means
through his language. The history of AI has be wrought with programs that seemingly
understand language but are really using "tricks" to appear intelligent to those
that have a genuine understanding because it draws from a knowledge base. The following programs are
representatives of what the field of language processing is working on.
As an attempt to get a computer to communicate in English, Joseph Weizenbaum in the
1960s had noticed how earlier attempts in this endeavor produced systems who could only
talk about specific topics. It was this main problem that he developed tricks in his
program called Eliza. Eliza was a conversational program that was able to load
modules(a.k.a. 'scripts') that would allow it to talk about anything. For instance,
if one wanted Eliza to talk about baseball, then just load in the baseball module. A
change of topic from baseball to gardening would merely require exchanging the baseball
module with the gardening one.
Weizenbaum had named Eliza as an allusion to Eliza Doolittle from George Bernard Shaw's
play, Pygmalion because the program became better at conversation the more it
practiced. On the other hand according to Weizenbaum, his program was like the
Doolittle character because one could not be sure whether or not either Elizas were
actually getting smarter. An example of Eliza can be found on this web page.
Some have argued whether Eliza can be considered even somewhat intelligent because it
employed tricks that covered up its lack of understanding. When confronted with
sentences it could not parse and understand, Eliza would
either ignore the last statement or it would produce the all-encompassing "I'm not
sure what you mean." Eliza employed other tricks like attaching key words
produced by the user to questions like "How do feel about [blank]?"
Curiously, programs that were off-springs of Eliza were a little more sophisticated but
that was good enough to fool people into thinking that they were talking to a real
person. Some people even formed enough attachment to the program that they shared
personal experiences with the computer too! It begs the question whether who has the
most intelligence, the computer or the duped people!
The Construction Worker--Shrdlu
Shrdlu. In 1968, the unusual name was a joke by its inventor, Terry Winograd, a
disciple of Seymour Papert who picked up the name from Mad Magazine. The
purpose of the program was to communicate in English with the user to manipulate blocks.
The computer would start with knowledge of what blocks were there to be manipulated
and what it could do with the blocks. The user would instruct the program in plain
English to, for example, stack the red pyramid on the blue block. From there, the
program would search for the blocks, pick it up, and place it on the blue cube. Of
course, the blocks all existed in the computer's virtual world.
What was significant about Shrdlu was that it represented the first, if limited,
program to have a meaningful conversation with the user. Not only could it
understand what to do when instructed in natural English, Shrdlu could answer back in
English why it created a tower using the blocks by backtracking its memory of the steps
that led to the building's construction. The program did have limitations because it
could not comprehend idioms like "stack all the blocks on top of each other."
Nevertheless, weak AI advocates would be quick
to point out that ultimately, Shrdlu would always answer "Because you told me to do
that" as the final reason why it manipulated the blocks in that manner.
Obviously, the computer is not yet intelligent enough to have "a mind of its
own."
