Most computerized translators
translate a document word for
word:
they look up the first word in the
document with the help of a
dictionary file, write
the corresponding word in another
language, read the second word,
and
so on, until the entire document is
in
a different language.
This allows the
user to get a vague idea of what's
going on, but it isn't as good as a
human translation.
My idea would be to build a set of
templates for a language, as well as
a
dictionary. When asked to translate
a
sentence, it would first consult its
templates and, by determining the
possible types of each word, find
the
one that fits the sentence best. The
templates would be two part; once
the blanks in the first part were
filled
out by the sentence to be
translated,
the second part would automatically
be filled, based on what was in the
first. After making changes to word
order or the words themselves, the
computer would proceed with the
word for word translation.
For example, take the French
sentence, "Est-ce que vous parlez
français?" Translated word for word,
this means, "Is it that you speak
French?" However, say you had the
template "Est-ce que vous
*verbPhrase1* ? : Do you
*verbPhrase1* ?" The translator
would be able to see that "Est-ce
que
vous" followed by a verb phrase meant
"Do you (phrase)?" It could detect
that the phrase in the French
sentence was probably "parlez
français." There might be a second
template for verb phrases, "*verb*
*verbModifier* : *verb*
*verbModifier*" If the words in
"parlez
français" could be used this way,
the
translator could be fairly certain
that
it had found the right template.
Because both side of the template
are
the same, it would leave the word
order in
the phrase unaltered. Once the
computer was sure that it had
worked
out the phrase, these words would
THEN be substituted for their
English
counterparts and inserted into the
English half of the Est-ce template,
resulting in the better, "Do you
speak
French?"
Of course, there would be some
problems. To create templates for
every possible sentence would be
troublesome. You would have to
implement some scheme where
templates to fit any sentence could
be built out of smaller basic
templates,
which could be built of even smaller
ones, and so on. Also, the sentence
might be ambiguous, having more
than one possible meaning.
Sentences might have more than
one
template that fit it well or words
might have more than one
substitute in another language. You
might be able to get around this
with
word associations and classes of
words; the computer would be
taught
that certain words would work
together better than others. For
example, a serial port carries data,
but sequential harbors usually do
not. In this case, you would have to
associate serial ports with words
expressing data transfer with
computer hardware, so that when
the computer read that a port was
sending or receiving data to or from
a peripheral, it would understand
that the type of port you find on the
back
of a computer is the one most
likely meant. The computer would
also have to learn "serial port" as a
single noun, not just as an adjective
and noun.
You could make the project files
open source, so that if anyone
noticed
a mistranslation, they could write a
few templates to fix the problem
and
update the translation
resources. Although the translations
would probably always be below
anything a human could do, with
enough advancements in language
processing, they could be accurate
enough for almost any use. (I'm
sure
the government would still want to
double check treaties and such.)