At 15:45 +0200 13/10/03, Anneleen Pareyn wrote:
Dear all,
this year I have to write my dissertation. The purpose is to compare a translation by a human translator (that'll be me, as I intend to graduate as a translator) and two machine translation systems. I also have to write some chapters about MT and everything that has to do with it. One of the subjects will be Controlled Language.
My question now is: Is Controlled Language really necessary to have a "perfect" MT output?
No! See below.
And� what is a "perfect" (human or machine) translation output anyway? Did your translation teacher give you some objective criteria for that?
Are there any books or special articles about this subject?
Thanks, Anneleen
I take it you mean "100% automatic MT output".
Then, A RESTRICTED ENOUGH SUBLANGUAGE, NOT CONTROLLED but resulting from some (ever slightly shifting) conventions and more or less formal rules of "well writing" IS ENOUGH TO GET NEAR PERFECT MT OUTPUT --- with a lot of ingenuity and elbow grease, that is.
Example: the sublanguage of the weather bulletins handled by METEO is NOT a controlled language. But METEO was last reported to be "97% correct" (in 1985, the figure was "only" 85%), or rather "3% less than perfect" by the measure of how many text editor operations (for 100 words) are performed in average by the human posteditors to get what they judge to be "professional" (perfect?) output. That corresponds to 1 mn for a typical bulletin of 100-150 words or so.
By comparison, the same posteditors (I mean, posteditors with the same level of excellence) had to spend about 7-10 mn before the first version of METEO, TAUM-METEO, was introduced (in 1976?). That is because translators producing the draft translations were junior translators assigned to this "purgatory" just a few months, the time to master this difficult kind of translation� and to become fed up with it. (By the time it was translated and sent back to wherever it came from, a bulletin had only 2 hours to live before becoming obsolete. Not very motivating!)
By that measure, METEO is about 7-10 times better than an average, learning, sweating and swearing junior translator just put on the job.
I think such task-oriented metrics are the best one for judging QUALITY-oriented MT, that is, "MT for revisors" as opposed to "MT for watchers".
Maybe John Chandioux could correct and complete what I sketched above, as well as give you pointers to some papers on the topic.
Best,
Ch.Boitet
-- For MT-List info, see http://www.eamt.org/mt-list.html
