Hi, 23/2/09
At 20:27 -0500 22/02/09, Job M. van Zuijlen wrote:
I'm wondering: what's exactly the purpose of this discussion?
Job van Zuijlen
F.Kovacs is simply spamming MT-list with his "MT-hater" e-mails,
- disseminating false examples of MT added to a Jews-hater picture,
which is not only bad taste, but ethically very incorrect,
- and not reading anything he gets in answer -- or so blinded by his
prejudices he can't listen to reason and goes on an on.
His last e-mail is so beyond the point I prefer to stop trying to let
him understand that MT is a tool and does not compete with human
translators.
What follows is more for our MT community -- arguments against MT
bashing may come in handy in some situations like this one.
I might give examples such as METEO, where professional translators
(from the Bureau of Translation) actually begged TAUM to produce an
MT system to relieve them of that "purgatory" -- then the system
actually replaced the junior translators producing the first drafts,
and the post-edition time (by seniors) was divided by 6 or 7. Since
more than 30 years now, that system (first TAUM-meteo, written in
Q-systems, then METEO, written in GramR) has regularly translated 30
M words/year (20 E-F, 10 F-E), and has saved the equivalent of 17
translators -- simply assigned to more interesting tasks than to
translate a bulletin which will be obsolete less than 4 hours later).
Or the METAL system (first developed through a contract of Siemens to
Slocum's team at Austin in 1981):
- already in 1984, after a few months of use (G-E), postedition time
went down to 10-15 min/page.
- in 2005 (EAMT-05, Budapest), Comprendium reported the actual use of
a version of METAL (grammars, dictionaires, of course)
Spanish-Catalan and Spanish Galician, to translate newspapers
everyday, with a post-edition time of 5 min/page, compared to 1 hour
for purely human work (measured before the system was put to use) and
30 mn for humans working with a translation memory (my estimation for
this last figure). METAL is now owned by LucySoftware and used for
these and other applications.
Well, I hope that what I wrote in answer to these e-mails can
convince young translators that there are MT systems which can really
help them, and, even more, that MT systems can be developed or tuned
with their particular data (lexicons, translation memories), to help
them more.
Or: what is important for humans is the usefulness of MT for the task
*they* have to or want to perform.
2 simple mottos:
- there can be and are (linguistically) very good MT systems, better
than human translators in the sense that professional revision is
possible and takes less time, but ONLY for applications on restricted
tasks/sublanguages, such as Nikkei flash reports (ALTFLASH) or
weather bulletins (METEO) or identified technical typologies/domains
(METAL) or on very similar languages.
- there are a lot of (linguistically) bad or even very bad MT systems,
BUT no MT output is bad enough not to be used with profit by real
bilinguals to augment their productivity -- more than by using a
translation memory if less than 20% exact matches and 40% fuzzy
matches can be found.
Best regards to the MT community,
Ch.Boitet
_______________________________________________
Mt-list mailing list
