https://nor.proz.com/forum/machine_translation_mt/290404-how_do_you_use_mt.html

Pages in topic:   [1 2 3 4] >
How do you use MT?
Thread poster: ..... (X)

..... (X)
Local time: 01:53
Sep 1, 2015

Hello everyone,

I'm interested in learning more about how (or if) translators use machine translation in their translation workflow.

- Do you use machine translation?

If no:
- Why not?
- Have you tried it? If so, what did you try? How did you use it?

If yes:
- How does it fit into your translation workflow? How do you use it? Do you use it as a reference? Do you "post-edit" the results? Do you use it on whole segments or to se
... See more
Hello everyone,

I'm interested in learning more about how (or if) translators use machine translation in their translation workflow.

- Do you use machine translation?

If no:
- Why not?
- Have you tried it? If so, what did you try? How did you use it?

If yes:
- How does it fit into your translation workflow? How do you use it? Do you use it as a reference? Do you "post-edit" the results? Do you use it on whole segments or to search terminology / phrases?
- Do you use a general machine translation engine or have you trained your own machine translation engine?

If you have never trained your own machine translation engine:
- Are you interested in doing so? What has held you back (i.e. don't have time, seems complicated, cost, etc.)?

If you have trained your own machine translation engine:
- How did it go? How did you find the process (easy, difficult, etc.)?
- Did you use an open source or commercial option?
- Do you use one trained engine, or have you trained different engines for different fields?
- About how much work did you train it with (10,000 segments, 50,000 segments, 100,000 segments, etc.)?
- How were the results? In your opinion was there a big enough gain in the quality of the results to make it worthwhile over a general machine translation engine?


- any other thoughts on the topic?

Thanks!
Kevin

[Edited at 2015-09-01 22:55 GMT]
Collapse


 

Christine Andersen  Identity Verified
Denmark
Local time: 17:53
Member (2003)
Danish to English
+ ...
I don't use it and have promised several of my clients not to use it Sep 1, 2015

I have tried post editing, and we are in fact going to discuss MT at the coming Danish powwow this Friday.

For reasons of confidentiality, I have signed agreements with several of my agency clients that I will not use MT or any cloud-based resources in any way that makes their clients' material accesible to anyone else. That means not contributing to TMs or allowing their texts and my translations to be used to train engines.

I dislike MT - it is often more trouble than
... See more
I have tried post editing, and we are in fact going to discuss MT at the coming Danish powwow this Friday.

For reasons of confidentiality, I have signed agreements with several of my agency clients that I will not use MT or any cloud-based resources in any way that makes their clients' material accesible to anyone else. That means not contributing to TMs or allowing their texts and my translations to be used to train engines.

I dislike MT - it is often more trouble than it is worth in the languages I work with, Danish to English (and English to Danish).

Although the languages are closely related, the results in Danish are sometimes incomprehensible, and if the results are usable in English, they are either heavily dependent on human translation, or the syntax in the sentences is often unnatural and hard to read.

MT has trouble with homonyms.

I am a translator, not an IT specialist, and would not know how to begin training a translation engine. One thing is for sure - I do not have the volumes of material available to make it meaningful.
I do use a CAT tool, but critically...

I don't dismiss MT out of hand, but I don't find it useful in my workflow. I don't trust it for reference - there are far more reliable dictionaries and resources in my languages, based on human intelligence, not a statistical lottery!



[Edited at 2015-09-01 23:13 GMT]
Collapse


 

Soonthon LUPKITARO(Ph.D.)  Identity Verified
Thailand
Local time: 23:53
Member (2004)
English to Thai
+ ...
In my case Sep 1, 2015

Kevin Dias wrote:
I'm interested in learning more about how (or if) translators use machine translation in their translation workflow.
Kevin

[Edited at 2015-09-01 22:55 GMT]


- Do you use machine translation?
Yes

If yes:
- How does it fit into your translation workflow?
It keys in texts for me.
How do you use it?
To seek idiom.
Do you use it as a reference?
Yes
Do you "post-edit" the results?
Yes
Do you use it on whole segments or to search terminology / phrases?
Yes
- Do you use a general machine translation engine or have you trained your own machine translation engine?
My own

If you have trained your own machine translation engine:
- How did it go?
Good
How did you find the process (easy, difficult, etc.)?
Easy
- Did you use an open source or commercial option?
No
- Do you use one trained engine, or have you trained different engines for different fields?
Different
- About how much work did you train it with (10,000 segments, 50,000 segments, 100,000 segments, etc.)?
May be 50,000 segments
- How were the results?
Acceptable
In your opinion was there a big enough gain in the quality of the results to make it worthwhile over a general machine translation engine?
Yes

- any other thoughts on the topic?
I believe MT is going to command the translation industry of the near future.

Soonthon L.


 

Meta Arkadia
Local time: 23:53
English to Indonesian
+ ...
Yes, but... Sep 1, 2015

... only as a last resort, and for the Auto-Complete feature. MyMemory only, and it runs "in the background." I disabled "Repair with MT" in my settings. It just isn't good enough, it's not worth the trouble to post-edit the results, but every once in a while it can be useful.

Some time ago, I downloaded Moses, and started to RTFM. It's way too difficult for elderly Luddites who are against any change, but the main reason why I stopped studying it and didn't even try it, is that mod
... See more
... only as a last resort, and for the Auto-Complete feature. MyMemory only, and it runs "in the background." I disabled "Repair with MT" in my settings. It just isn't good enough, it's not worth the trouble to post-edit the results, but every once in a while it can be useful.

Some time ago, I downloaded Moses, and started to RTFM. It's way too difficult for elderly Luddites who are against any change, but the main reason why I stopped studying it and didn't even try it, is that modern CAT tools imply MT technology. I already use Auto-Assemble and Subsegment Matching (for certain jobs) in CafeTran, and I doubt if setting up an MT and training it will offer additional benefits. True or false?


Cheers,

Hans
Collapse


 

Meta Arkadia
Local time: 23:53
English to Indonesian
+ ...
Me too Sep 1, 2015

Soonthon LUPKITARO(Ph.D.) wrote:
I believe MT is going to command the translation industry of the near future.


And I expect a "revolution" in the technology very soon. I think the current common approach (statistical) is wrong and its evolution too slow, and although it may lead to acceptable results in the near future, it can never be better than "acceptable." Revolutions in either hardware of software, or - more likely - both are on their way. Beware!

Cheers,

Hans


 

Oliver Pekelharing  Identity Verified
Netherlands
Local time: 17:53
Dutch to English
Sure Sep 2, 2015

I mostly use it as a reference, but just now I populated a segment with this MT suggestion: "Dimensions, materials, manufacturing processes". Why would I want to type that out if I can grab it with a key combination or click of the mouse? I find MT to be useless for standard prose (although I may scan an MT suggestion for a word or term that is escaping me), but it can be great for lists. And use MT as one of my autosuggest dictionaries: plenty of useful suggestions.

Olly

[Ed
... See more
I mostly use it as a reference, but just now I populated a segment with this MT suggestion: "Dimensions, materials, manufacturing processes". Why would I want to type that out if I can grab it with a key combination or click of the mouse? I find MT to be useless for standard prose (although I may scan an MT suggestion for a word or term that is escaping me), but it can be great for lists. And use MT as one of my autosuggest dictionaries: plenty of useful suggestions.

Olly

[Edited at 2015-09-02 06:36 GMT]
Collapse


 

Andrea Halbritter  Identity Verified
France
Local time: 17:53
French to German
+ ...
Not at all Sep 2, 2015

I do not use it at all. In my language pair the result is not understandable.

 

..... (X)
Local time: 01:53
TOPIC STARTER
Interesting question Sep 2, 2015


Meta Arkadia wrote:
I already use Auto-Assemble and Subsegment Matching (for certain jobs) in CafeTran, and I doubt if setting up an MT and training it will offer additional benefits. True or false?


I think this is a very interesting question and would be curious to know if any translators have tried anything similar.

For example, if you had a 100,000 translation unit TM in a certain field. You use that TM to train a machine translation engine. Then in your CAT tool you use the original 100,000 translation unit TM with auto-assemble and subsegment matching. Does feeding each segment of the new source text you are translating through the trained MT provide you any benefit as an additional reference point, or would its output be unnecessary as you already have subsegment matching?

What about subsegment matching and then machine translating the "unmatched" portion of the source segment? I think I have seen a few CAT tools that claim to do this, but probably that is only with a general MT (please fill me in if I'm wrong). I wonder if anyone has tried this with a MT engine trained on a specific field?

Kevin


 

Oliver Pekelharing  Identity Verified
Netherlands
Local time: 17:53
Dutch to English
Auto-Assemble and Subsegment Matching Sep 2, 2015

I have found auto assembly and subsegment matching to be useless as 'pre-translation processes', not because the suggestions aren't useful, but because it takes too much time to rearrange the order. I prefer to just start typing and select any useful words and sub-segments that are offered by autosuggest as I go.

Olly


 

Dan Lucas  Identity Verified
United Kingdom
Local time: 16:53
Member (2014)
Japanese to English
Confidentiality concerns Sep 2, 2015

Kevin Dias wrote:
I'm interested in learning more about how (or if) translators use machine translation in their translation workflow.
- Do you use machine translation?

I would be happy to use machine translation on my own terms. If nothing else it would be an instant gisting function. The result might be awful or wonderful, but given that I can delete or populate text in a target segment with a keystroke, what's the downside?

However. With most of my clients I have signed conditions prohibiting me from using MT. And even with their permission, I would be very wary of using anything that passed sensitive information - such as, for example, a financial report that has not yet been released - back to the MT provider.

There was a case widely reported in Japan early in 2015 in which it was found that people (including employees of government ministries and major banks) had entered various strings into an MT site based in China. Those strings were, unsurprisingly, recycled by the MT engine and all kinds of confidential information came up, including email addresses that nobody wanted to be made public.

I cannot have Google (or SDL, or anybody else) leaning over my shoulder and taking notes as I translate unless they promise never to use what they have seen. Would you trust them? My first condition for use, then, would be absolute, guaranteed privacy at a reasonable price.

Regards
Dan


 

Meta Arkadia
Local time: 23:53
English to Indonesian
+ ...
Settings and algorithms Sep 2, 2015

Olly Pekelharing wrote:
I have found auto assembly and subsegment matching to be useless as 'pre-translation processes', not because the suggestions aren't useful, but because it takes too much time to rearrange the order. I prefer to just start typing and select any useful words and sub-segments that are offered by autosuggest as I go.


Auto-assemble doesn't necessarily include "insert in target field." It means that - based on the CAT tool's algorithms, and your settings - your tool shows you the best matches from your available resources, including the best suggestions for Auto-Complete (Auto-Suggest) typing. You are right about the order for most if not all language pairs, but I think you're wrong the usefulness of AA and SM. SM is an interesting subject anyway, if you are that way inclined (I'm not).

Cheers,

Hans


 

Meta Arkadia
Local time: 23:53
English to Indonesian
+ ...
General vs subject specific, and more Sep 2, 2015

Andrea Halbritter wrote:
I do not use it at all. In my language pair the result is not understandable.


That goes for most if not all language pairs, I think (and hope). But I think you're referring to the MTs you can consult on/via the Internet. They have o.a. the following problems:

  • They are too general, out of necessity
  • They don't care about style, and cannot care about it
  • As far as I know, ALL of them use English as a "conversion"/"bridge" language, in other words, your original French text is translated into English, and the resulting English text is translated into German (which is why you'll find English words in the German text). problems? Here's how to get them
  • They offer one suggestion per segment only. That's their result, and you'll have to be happy with it

    But what if you only use subject specific corpora, in the right style, several suggestions, and not translated via English?
    I do EU BS, and for that, I use MT technology - Auto-Assemble, Subsegment Matching, Virtual Matches - on the DGT and IATE resources in my CAT tool (CafeTran) to arrive at very decent results. Okay, I have to rearrange the order, but most essential stuff is there. I can't use all of the technology for some other jobs, though.


    Cheers,

    Hans

     

  • Samuel Murray  Identity Verified
    Netherlands
    Local time: 17:53
    Member (2006)
    English to Afrikaans
    + ...
    As for me Sep 2, 2015

    Kevin Dias wrote:
    If yes:
    - Do you use a general machine translation engine or have you trained your own machine translation engine?


    There are only three MT engines available for my target language, namely Google, Apertium, and Autshumato. I'm not a government contractor, so I don't have access to Autshumato, and Apertium's engine is not really good enough for professional use (and is not being developed further at this time). Fortunately, my source and target languages are culturally, grammatically, and syntactically sufficiently similar for Google Translate (GT) to be mostly useful to me.

    I use MT whenever I'm allowed to.

    - How does it fit into your translation workflow? How do you use it? Do you use it as a reference? Do you "post-edit" the results? Do you use it on whole segments or to search terminology / phrases?


    I don't use MT to search for the translation of terms or phrases, or for reference, although I do find that technical terms are often translated correctly by GT. Still, whenever there is a technical term (e.g. a chemical compound name) in a segment for which I used GT, I double-check the translation of that term using independent sources. Sometimes GT gets it wrong. Sometimes most people get something wrong, and then GT also gets it wrong.

    My workflow is this: essentially, I turn my source text into a machine translated TM.

    First, I extract the segments of the current job to a text file (file A), then use qTranslate to translate those segments into the target language (save as file B), and then I use PlusTools to align the source and target texts (files A and B) to create a machine translated TM. I then change the "user ID" for all segments to "GT". Then, in my CAT tool (WFC), I penalise the user "GT" by 30%. This means that all segments from my other TMs that have a match percentage of higher than 70% will show up before the MT segment, but if the match percentage is lower than 70%, then the MT segment is shown first (and the lower matches are shown only if I cycle to them, for that is how WFC works). Oh, I also do some find/replaces on the GT output to "fix" it before adding it to the TM.

    In WFC, whenever you edit a segment with someone else's user ID, it is updated in the TM with your own user ID, and the other user's segment is marked for deletion, so at the end of a translation session, most of the GT segments in the TM are marked for deletion. What's more, in WFC, new segments are added to the bottom of the TM, so if you start with a purely MT'ed TM, all MT segments will be at the top of the TM where you can easily delete them in a text editor, to leave only your own segments. Also, in WFC, concordance searches (i.e. searches through the TM) are performed from the bottom upwards, to GT translated segments tend to appear at the bottom of a concordance search if the GT segments are at the top of the TM. Not that it matters, because GT segments are clearly marked in the concordance search result as having the user ID "GT".

    If you have never trained your own machine translation engine:
    - Are you interested in doing so? What has held you back (i.e. don't have time, seems complicated, cost, etc.)?


    Yes, I'd be very interested in doing so, but the freely available MT engines are too complex for me to set up. I have tried a few times, but I think the free MT engines all assume the engine will be set up by someone who is an expert at setting up MT engines.


     

    Philippe Etienne  Identity Verified
    Spain
    Local time: 17:53
    Member
    English to French
    Desktop MT vs. online MT Sep 2, 2015

    Dan Lucas wrote:
    ...
    However. With most of my clients I have signed conditions prohibiting me from using MT. And even with their permission, I would be very wary of using anything that passed sensitive information - such as, for example, a financial report that has not yet been released - back to the MT provider.

    There was a case widely reported in Japan early in 2015 in which it was found that people (including employees of government ministries and major banks) had entered various strings into an MT site based in China. Those strings were, unsurprisingly, recycled by the MT engine and all kinds of confidential information came up, including email addresses that nobody wanted to be made public.

    I cannot have Google (or SDL, or anybody else) leaning over my shoulder and taking notes as I translate unless they promise never to use what they have seen. Would you trust them? My first condition for use, then, would be absolute, guaranteed privacy at a reasonable price.

    This is indeed the main issue when using online tools, and I have recently signed my first agency form that explicitely prohibits the use of online MT.
    But here I think the OP means desktop, local MT systems, and not online tools such as GT. With local MT processing, I would think that the outcome is just as safe as any other file on a computer connected to the internet.

    To reply to the OP's question, I have post-edited MT through agencies in suitable fields, with suitable MT custom setups and with suitable raw quality.
    I don't use MT for commercial purposes, but I am aware that buying a standalone MT program and fine-tuning it could be translated into new, profitable "streams of revenue". Besides, fitting an MT program to my needs would likely be much less boring than post-editing its outcome. However, learning new tools and customising them is certainly time-consuming, for an outcome and a market that I may have overestimated.

    Therefore, since my time is currently all used with translations for which machines are of no use whatsoever, any MT investment remains hypothetical in the foreseeable future.

    Philippe


     

    Samuel Murray  Identity Verified
    Netherlands
    Local time: 17:53
    Member (2006)
    English to Afrikaans
    + ...
    @Dan Sep 2, 2015

    Dan Lucas wrote:
    There was a case widely reported in Japan early in 2015 in which it was found that people (including employees of government ministries and major banks) had entered various strings into an MT site based in China.


    If you're talking about the article in The Japan News of February 2015, well, it was reported in only one newspaper (The Japan News) and then reposted verbatim on several blogs, but I was unable to find any other report about it in any newspaper in English. Do you happen to know the name of the MT system that they used? You seem to know that the MT system was based in China, but this fact is not reported in the The Japan News article, as far as I'm aware, so can you remember where you read that piece of information?

    I would have really liked to double-check the journalist's comment that the users of that web site were unaware that their texts would be published on the internet.


     
    Pages in topic:   [1 2 3 4] >


    To report site rules violations or get help, contact a site moderator:


    You can also contact site staff by submitting a support request »

    How do you use MT?

    Advanced search






    CafeTran Espresso
    You've never met a CAT tool this clever!

    Translate faster & easier, using a sophisticated CAT tool built by a translator / developer. Accept jobs from clients who use SDL Trados, MemoQ, Wordfast & major CAT tools. Download and start using CafeTran Espresso -- for free

    More info »
    PerfectIt consistency checker
    Faster Checking, Greater Accuracy

    PerfectIt helps deliver error-free documents. It improves consistency, ensures quality and helps to enforce style guides. It’s a powerful tool for pro users, and comes with the assurance of a 30-day money back guarantee.

    More info »