Machine translation from 1948 to 2003
John Hutchins, who has been working on machine translation
systems for many years, has posted online versions of his
books and papers as well as a great deal of source material
on his web site. People working on natural language processing
and related fields may find it useful:
http://ourworld.compuserve.com/homepages/WJHutchins/
For an overview of MT from the early days to the present,
see the paper
http://ourworld.compuserve.com/homepages/WJHutchins/HasMTimproved-exp.pdf
The title of this paper "Has Machine Translation Improved?" is
significant for many reasons other than just MT. In his 50-year
overview, Hutchins shows that there are indeed improvements, but
they are not the kind of spectacular revolutions that one might
have hoped for.
This snail-like progress raises questions for any area related
to natural language -- including the current work on ontology
and the semantic web.
And by the way, people often say that progress in hardware is
much greater than progress in software. Actually, however,
there is very little difference: the latest and greatest CPUs
from Intel, AMD, IBM, etc., are just cheaper, faster versions
of architectures that are 20 or 30 years old. The register set
and the basic instructions of the latest and greatest Pentiums,
for example, are based on Intel's first CPU chip, the 4004,
from the early 1970s.
Even the details that make the machines faster are extensions
of the pipelined, parallel, superscalar, and vector-processing
innovations designed for the supercomputers of the 1960s.
Everything else is the result of Moore's law, which says the
amount of circuitry per silicon area doubles every 18 months.
That makes all the old stuff smaller, faster, and cheaper -- but
no different in basic architecture.
John Sowa