А вот что пишет один очень умный товарищ (Jost Zetzsche, очень рекомендую подписаться на его рассылку):
I spent this last week training a rules-based machine translation engine (in this case, PROMT -- Premium subscribers with access to the archives can read about this in the 184th Tool Kit), and I finally understood a rather fundamental truth about MT (I've never claimed to be particularly particular quick-witted!).
The translation results out of the box (language direction English into German) were really rather pitiful. And while they got better after changing some basic settings (such as the text genre and form of address for the user group), they still were not good and not much of a productivity boost for any translation effort. Plus, they were decidedly worse than what Google Translate would have given me.
But then I set out to work on it. I imported large and high-quality translation memories that the client had given me, I extracted terminology from the source texts that I translated based on the data in the TM and defined grammatically, and I imported and fine-tuned glossaries. Now the result was better, but still far from what it was supposed to be. So the rest of the week was spent "translating" the source texts with the MT engine at its current state, hunting down terms that were not at all or incorrectly recognized and entering them into the dictionary (and "entering them" here does not just mean sending source and target translation, but defining them grammatically as well), then retranslating, hunting down more terms, and on and on.
At some point it was fun to see the results getting better each day, and at the end of the week the result was certainly better (though maybe less idiomatic) than that of Google Translate. Assuming that the final product will be a translation of "high" quality, I would guess that the post editor will save somewhere between 20% to 40% of the time that it would take to translate the text within the traditional TEnT environment.
Clearly this is not something that would make sense for a smallish kind of project, but for a very targeted project with a few hundred thousand words this can be large productivity boost. (And if the goal is to deliver "only usable" translation, the saving would naturally be higher.)
But here is what I learned. I know that more and more translators are using tools like Google Translate or Bing Translator as productivity boosts. Aside from the confidentiality issues (your clients probably will not be too thrilled to see you uploading their data to Google or Microsoft), the results might look like decent productivity boosts at first glance (this obviously depends on your language combination and kind of translation you're doing). But if you are inclined to use MT as a productivity tool and you generally work in the same domain, you will achieve greater boosts with MT tools that you can train. In turn, however, this means that you will have to make a heavy investment of time before you see any results.
(And if tools like PROMT ever become widely used translation productivity tools, they might also change some things about their user-friendliness and the efficiency with which external data, in particular TM data, is being used.)
I did have a talk with a representative from the market-leading rules-based machine translation tool Systran this week -- more on that in the next newsletter.

Ответ на комментарий # Интересно. Подписалась.
Когда-то тоже пыталась ради интереса "учить промт". На много меня хватило, но все равно, результат был заметен. Но душа так и не легла.

Душа к такому лечь и не может, но использование в качестве очередной маленькой подпорки - почему бы и нет? Это и правильно, мне кажется.