[ From
Translation... | ...To Understanding | Thought-Language ]
"We cannot think what we cannot say."
--Ludwig Wittgenstein, philosopher(Kurzweil 32)
ust as Wittgenstein
implies, thought is inextricably linked with language. Conscious thinking usually
is done by saying things in one's head because to know how to express something means that
one has an adequate understanding of what is being said. Using this fact,
researchers have approached artificial intelligence by trying to design a computer
that can process natural language--i.e. the language that is commonly
used by people such as English, French, Russian, Chinese, etc. Researchers
hope that if a machine can understand a person's language, then it is exhibiting a kind of
intelligence.
From Translation...: The Groundwork for Understanding
Language
Drawing upon the design specifications on boolean
algebra founded by George Boole, today's
computers understand only one language: binary. All of its processes can be reduced
to electricity turning a computer's components on or off. This problem has aptly
solved by programmers who eventually codified the meanings simple words(usually in
English) to facilitate programming. This new kind of computer language is called
high-level computer language. Nevertheless, computers with this high-level language
have a very limited understanding of the commands it was programmed. It may be
countered, as philosopher John Searle did in his Chinese
Room Argument, that the computer does not really understand what it is told to do, but
it is merely matching up pre-existing words to actions and performing the commands.
Of course, high-level programming
languages are by no means a conscious attempt to produce artificial intelligence, but it
was a precursor to it.
After the successful development of code-deciphering machines used to find out secret
German plans in World War II, the American mathematician Warren Weaver set out to invent a
machine to translate natural language. In theory, the computer would have a sentence
entered into its memory, find comparable words in a translation dictionary, rearrange the
newly translated words to form a grammatically-correct sentence, and then output the
translated sentence.
Since the United States Department of Defense loved the idea of translating Russian
journals into English to find out what the Soviets were up to in the Cold War, it had
Anthony G. Oettinger actually develop the word-for-word translation device in 1954.
What Oettinger and other scientists and engineers found out was that unlike humans, a
machine was not sophisticated enough to be a sensible translator.
For most people, it may only require a direct translation-dictionary to grasp the idea
of a sentence written in another language. Even if the translated sentence does not
make sense literally, the gist of the sentence can usually be deducted because the
thinking process of people are usually the same no matter what language or culture.
This literal-translation approach, however, does not work for computers.
Computers may be able to process mathematical problems with great accuracy yet they
cannot translate language as well as a person does because it does not understand what it
is translating. In the English language alone, most words have multiple dictionary
definitions. This does not include slang or other idioms that are culturally-based
and thus usually require the background knowledge of a particular culture to understand
the figures of speech.
So what happened to Oettinger's translation computer? Historically-speaking, it
always returned gibberish when it tried to translated a sentence. From such
incompetence sprung tales of translation mishaps like when the computer was said to have
translated "The spirit is willing but the flesh is weak" from English to Russian
and then back to English. At the end of its processing, the computer returned,
"The vodka is good but the meat is rotten."(Herndon 55)
[ Top ]
Today, programmers have developed more sophisticated language-processing techniques.
One of the first steps to understanding a sentence is to understand its words
individually. An approach to this problem is to build a dictionary for the computer
to access. However, a problem in this method is that the definitions of a word may
contain words the computer doesn't understand as well so it must look those words up.
Soon, the computer will run into problems when either it can't find a word without
a definition or the word refers to a term that had directed the computer to the current
word in the first place.
Unfortunately, this method is not fool-proof. Even though parsing a sentence
usually yields word relationships that are clear, there are other times when sentences can
be parsed in different ways that affect its meaning. This is especially true when
the sentence is complexly structured. There is an example
of the multitude of ways to parse a complex sentence that produces a variety of meanings.
[ Top ]
What's the difference between "The cook is baking" and "The cake is
baking?" Obviously, in the former the cook is a person whose action is
"baking" while in the latter the cake is an object that the action
"baking" is being performed on. From this example, linguist Charles
Fillmore set out in the late 1960s to distinguish between the literal definition of a word
and the deeper meaning the word represents. Fillmore proposed eight "deep
cases" like an "agent"--the instigator of an event, an
"instrument"--the physical cause of an event, and the
"experiencer"--the object in which the instrument acts on. Somewhat like
intuition, a computer would have to know that any agent must be an active entity in order
to do something while the experiencer is inanimate in the sentence. Thus, the cook
is an agent so he is doing the baking while the cake is an experiencer so it is being
baked.(Crevier 164-165)
Roger Schank, director of Northwestern
University's Institute for the Learning Sciences(ILS), essentially took sentence parsing
and Fillmore's deep cases another step so that a computer could understand more about the
sentence when it is broken up. People tend to gain knowledge not from the words
composing a sentence, but from what the sentence means. That would explain why it is
easier for a person to paraphrase then to recite something said ten minutes ago.
According to Schank, it is the meaning of a sentence that is interpreted and translated
into a person's language of thought--a kind of "mentalese." Henceforth,
Schank develop a system that breaks down English into elementary concepts that can be
represented by a machine. Surprisingly, Schank realized that grammar wasn't as
important in translating sentences into representative concepts. Like Fillmore,
Schank broke down English into what he refers as "semantic primitives."
In the case of actions, eleven basic acts represent the entire spectrum of English verbs
in which a few can be seen below:
