CHAT: Machine translation (was Re: translation)

On 6/19/06, Roger Mills <rfmilly@...> wrote:
> Machine Translation....well what did we expect anyway?
Not much.  MT has gotten cheaper and more widely available, but the
quality hasn't really improved much in the last 30 years.

Sometime in the 80s or thereabouts the AI community mostly stopped
working on what we might think of as "real"  or "pure" AI - by which I
mean, the attempt to produce software that can think (or, as some
would say, give the impression of thinking, but I don't want to get
into the fallacy of the philosophical zombie on here) like a human.
After decades of  funding, that sort of research had consistently
yielded not much in the way of profitable applications, and
consequently said funding dried up.  As a result the AI community
shifted its focus to narrowly-scoped, rules-based "expert systems"
instead.

Such systems led to commercial success.  The ubiquitous "driving
directions" software provided by mapping sites and GPS devices is a
good example.  In more specialized domains expert systems are
everywhere;  there is software that can do a preliminary medical
diagnosis, or help police and military modify an attack plan on the
fly based on enemy activity.  Google's indexing engine is essentially
an expert system for classifying documents - although it's also
another example where software that could actually *understand* what
it was reading would do a better job.

Unfortunately, translation doesn't lend itself well to the
expert-system approach.  My personal belief is that reliable MT is
tantamount to real AI - that the two problems are, in fact, actually
the same problem.  There is a widespread belief (held by many in the
AI community itself, not just by laymen) that MT is somehow not only
more tractable than real AI, but orders of magnitude simpler.  But
IMHO, that's just another example of humans underestimating the
difficulty of something that we happen to be optimized for.

--
Mark J. Reed <markjreed@...>

Reply