I'm looking for translators willing to answer 2 simple questions. Thread poster: Gert Van Assche
|
|---|
| | Tina Vonhof (X)
Canada
Local time: 07:02
Dutch to English
+ ...
Maybe, if you could give at least a hint what the questions are about and why you are asking.
| | | | Henry Hinds 
United States
Local time: 07:02
English to Spanish
+ ...
In memoriam | Machine Translation | Mar 26, 2014 |
It has to do with how we deal with machine translation in editing, etc. Since I never have and never will do such work, my replies will not be useful. He probably should state the subject so that those of us not involved in it can stay away.
| | | | | what we'd like to know... | Mar 26, 2014 |
Tina,
it is about how translators approach a translation job when the text has been pre-translated by a memory or a machine. We observed 2 techniques, but for sure there are more approaches that make sense. We would like to know how many translators opt for on of these 2 approaches, or for another one.
thanks
gert
| | |
|
|
|
| Even when you don't post-edit, your answers are useful. | Mar 26, 2014 |
Henry, this is not only about post-editing. We use the the first question to know what type of translator you are. Only if you never work on pre-translated text at all (so even not TM fuzzies), you cannot help me with this survey.
thanks
gert
| | | | Steven Segaert
Estonia
Local time: 16:02
Member (2012)
English to Dutch
+ ...
| Never happens | Mar 26, 2014 |
The theory is that some machine translated segments have a correct content and are fluent, so that they don't need to be touched. That almost never happens, except for numbers-only segments or fairly standard formulas. Even with "trained" engines.
The schematics in the survey quite clearly show why PE doesn't work. When I translate, I have two steps: read the source, compose the target. Compare that to the five steps in PE.
So, as "good" segments almost never materialis... See more The theory is that some machine translated segments have a correct content and are fluent, so that they don't need to be touched. That almost never happens, except for numbers-only segments or fairly standard formulas. Even with "trained" engines.
The schematics in the survey quite clearly show why PE doesn't work. When I translate, I have two steps: read the source, compose the target. Compare that to the five steps in PE.
So, as "good" segments almost never materialise, using PE instead of translating from scratch is a waste of time, effort and (mostly the translator's) money.
As I see it, the only way in which automation can be valuable is in offering terms and sub-segments and in guarding style rules. And the only way that can be made to work for a translator is to offer these as stand-alone items, not as a composition that needs to be edited.
As the machine-craze seems to be subsiding a bit, I guess I'll know soon enough whether or not I'm on the right track with that personal analysis. ▲ Collapse
| | | | | this is not just about MT | Mar 26, 2014 |
Steven,
It all depends on what the MT system is translating, what language pair, what content... But today MT technology is also used for instance on sub segment matching on TM systems, and the tools that exist today are getting "almost hybrid". There are plenty of opportunities to improve the algorithms and the training corpus used. A big weak spot is good support of new terminology... But also in this area progress can be expected in the next 5 years.
This small s... See more Steven,
It all depends on what the MT system is translating, what language pair, what content... But today MT technology is also used for instance on sub segment matching on TM systems, and the tools that exist today are getting "almost hybrid". There are plenty of opportunities to improve the algorithms and the training corpus used. A big weak spot is good support of new terminology... But also in this area progress can be expected in the next 5 years.
This small survey is not on post-editing MT output only, but on how translators approach pre-translated text in general. I did a similar "survey" 10 years ago when we were trying to understand why some of our best translators were simply not good in reviewing work of more junior translators.
About the quality of MT systems: they cover the full spectrum from black to white. I have seen output that was so terrible that not one single sentence out of 1000 sentences was of acceptable quality, and on 90% all of the meaning was lost. I have seen an engine where the output was close to perfect, and the translators refused to work on TM output and preferred the MT output. It really depends on how good your training data matches the new texts you have to translate, on what kind of pre- and post-processing you can put in place. But even when all looks great, things can still go wrong in production when the translators don't get the right support, the correct payment for their work, or when no error-centric QC is put in place.
Last but not least: MT will not replace humans, but I'm sure it can help some humans on some texts they need to work on. Just like TMs are also helpful sometimes.
gert ▲ Collapse
| | | | Steven Segaert
Estonia
Local time: 16:02
Member (2012)
English to Dutch
+ ...
| Not really my point | Mar 26, 2014 |
I'm not saying technology and algorithms can't offer useful results. In fact, there is a lot of scope for automating the translation workflow (and especially concerning QA aspects), and I am not at all opposed to that.
What I'm saying is that *editing* MT results is more work than using them as you would use a terminology or TM hit. There is a big difference between editing a composition that is "about right", but has an unnatural syntax, and using the parts of the segment that are... See more I'm not saying technology and algorithms can't offer useful results. In fact, there is a lot of scope for automating the translation workflow (and especially concerning QA aspects), and I am not at all opposed to that.
What I'm saying is that *editing* MT results is more work than using them as you would use a terminology or TM hit. There is a big difference between editing a composition that is "about right", but has an unnatural syntax, and using the parts of the segment that are right to compose your own (correct syntax).
Or applied to a "classic" CAT translation grid, between having to edit a pre-filled target segment and filling it yourself with the aid of relevant terms and sub-phrases that are offered as suggestions.
The latter can save a translator from having to look up terms or style rules, while the first is more often than not a frustrating experience that will probably never result in much quality work. The workflow in these scenarios is very different.
In the language pair English to Dutch, for example, I'm sure you come across many MT compositions that are simply the wrong way around, even if the words are right. A calculation of what has changed in comparison to the machine result might show you only a small percentage, while the effort made by the translator is actually quite big: read the source, read the target, realise that the syntax is wrong, think of the right syntax, copy/paste in the right order or re-type the whole thing.
In the end, I might have "just" changed the word order, while this in fact can easily take about three times as long as to start from an empty slate. And even longer compared to translating from scratch while having the benefit of a good term base.
A PE setup is only truly useful to the extent in which it can offer these right words or groups of words, without adding the burden of having to edit a wrong syntax. ▲ Collapse
| | |
|
|
|
Tim Drayton
Cyprus
Local time: 16:02
Turkish to English
+ ...
| I wish people who devise surveys would apply a little logic to them! | Mar 27, 2014 |
My answer to question one was:
"Never -- I reject jobs that contain machine translated segments."
So, logically, question two does not apply to me, but when I try to complete the survey without answering the second question, I am prompted to answer it.
Therefore I cannot complete the survey.
| | | | Giles Watson
Italy
Local time: 15:02
Italian to English
In memoriam | Neither do I | Mar 27, 2014 |
Tim Drayton wrote:
My answer to question one was:
"Never -- I reject jobs that contain machine translated segments."
So, logically, question two does not apply to me, but when I try to complete the survey without answering the second question, I am prompted to answer it.
Therefore I cannot complete the survey.
Well, I don't use MT either but I do use my own TMs.
You can answer "C: None of these two describes the way I translate", although Gert might want to substitute that "None" with the dual form "Neither".
| | | |
Tim, I think answer C would be the best answer in your case. If you also never use translation memory output, you don't have to fix a pre-translated target and your strategy is different. I also know about translators who receive pre-translated jobs (pre-translations coming from a translation memory), and who simply delete all and always start from scratch.
Giles, thank you for post-editing my text. I'll fix it right away.
gert
| | | | | | To report site rules violations or get help, contact a site moderator: You can also contact site staff by submitting a support request » I'm looking for translators willing to answer 2 simple questions. | Trados Business Manager Lite | Create customer quotes and invoices from within Trados Studio
Trados Business Manager Lite helps to simplify and speed up some of the daily tasks, such as invoicing and reporting, associated with running your freelance translation business.
More info » |
| | Anycount & Translation Office 3000 | Translation Office 3000
Translation Office 3000 is an advanced accounting tool for freelance translators and small agencies. TO3000 easily and seamlessly integrates with the business life of professional freelance translators.
More info » |
|
| | | | X Sign in to your ProZ.com account... | | | | | |
Login to reply/comment 
Translation news
