Pages in topic:   [1 2] >
Translation platforms cannot replace humans-- but they are still astonishingly useful

This discussion belongs to Translation news » "Translation platforms cannot replace humans-- but they are still astonishingly useful".
You can see the translation news page and participate in this discussion from there.


Michele Fauble  Identity Verified
United States
Local time: 01:12
Member (2006)
Norwegian to English
+ ...
More hype May 2, 2017

"The new system still makes mistakes, but these are now relatively rare, where once they were ubiquitous."

 

Siegfried Armbruster  Identity Verified
Germany
Local time: 10:12
Member (2004)
English to German
+ ...
Usefull? May 2, 2017

Translation is a complex task. I just copied a standard text from my market segment, English into German.

English original
Read all of this leaflet carefully before you start taking this medicine because it contains important information for you.
• Keep this leaflet. You may need to read it again.
• If you have any further questions, ask your doctor or pharmacist.
• This medicine has been prescribed for you only. Do not pass it on to others. It may harm them, even if their signs of illness are the same as yours.
• If you get any side effects, talk to your doctor or pharmacist. This includes any possible side effects not listed in this leaflet.

German translation by Google
Lesen Sie die gesamte Packungsbeilage sorgfältig durch, bevor Sie mit der Einnahme dieses Arzneimittels beginnen, da es wichtige Informationen für Sie enthält.
• Bewahren Sie diese Packungsbeilage auf. Möglicherweise müssen Sie es noch einmal lesen.
• Wenn Sie weitere Fragen haben, fragen Sie Ihren Arzt oder Apotheker.
• Dieses Arzneimittel wurde nur für Sie verschrieben. Gib es nicht an andere weiter. Es kann ihnen schaden, auch wenn ihre Anzeichen von Krankheit die gleichen sind wie deine.
• Wenn Sie Nebenwirkungen haben, sprechen Sie mit Ihrem Arzt oder Apotheker. Hierbei handelt es sich um eventuelle Nebenwirkungen, die in dieser Packungsbeilage nicht aufgeführt sind.

It looks like German, many people will even believe that it is a good translation. NO it is not a good translation, the sample is part of a regulated document, the wording is fixed by a regulatory body. Not one of the "translated" sentences is even close to the required wording, i.e. the translation is useless and completely wrong. The official translation is:

Lesen Sie die gesamte Packungsbeilage sorgfältig durch, bevor Sie mit der Einnahme dieses Arzneimittels beginnen, denn sie enthält wichtige Informationen.
- Heben Sie die Packungsbeilage auf. Vielleicht möchten Sie diese später nochmals lesen.
- Wenn Sie weitere Fragen haben, wenden Sie sich an Ihren Arzt oder Apotheker.
- Dieses Arzneimittel wurde Ihnen persönlich verschrieben. Geben Sie es nicht an Dritte weiter. Es kann anderen Menschen schaden, auch wenn diese die gleichen Beschwerden haben wie Sie.
- Wenn Sie Nebenwirkungen bemerken, wenden Sie sich an Ihren Arzt oder Apotheker. Dies gilt auch für Nebenwirkungen, die nicht in dieser Packungsbeilage angegeben sind.

A modern CAT tool with a specialized TM will give you better results. Sure, these new systems will get better and they might help to translate general text, but specialized translator don't need to worry in my opinion, The maschines still don't understand a thing, and understanding the source and the target content is in my opinion key to produce quality translations.


 

neilmac  Identity Verified
Spain
Local time: 10:12
Spanish to English
+ ...
Translation: May 3, 2017

MT (which is what I understand "translation platforms" to mean) can be a useful tool for an accomplished human translator. However, it is not a magic wand and isn't going to make Harry Potters out of numpties.

 

Michael Wetzel  Identity Verified
Germany
Local time: 10:12
German to English
Granted, the German translation is useless, but why does it look so good? May 3, 2017

If the German text were not part of a regulated document, the translation would be shockingly good for MT.

The German target text also departs from the English source text in ways that I would not expect MT to do unless there are hundreds or thousands of very similar translations floating around in the Internet.

Probably six months ago, I fed a few random short paragraphs from the New York Times into GT neural or premium or plus (or whatever the name of their latest marketing gag was). In the direction German>English, the results were far more plausible than those of whatever the standard GT was called, but they were nowhere near acceptable in terms of language or reliably conveying content.

I accept your point that the MT translation is completely useless, but my question is: What is it that makes your example special? Why was MT able to produce (outwardly) very plausible results at several points in this case?


 

Matthias Brombach  Identity Verified
Germany
Local time: 10:12
Member (2007)
Dutch to German
+ ...
Shared intelligence May 3, 2017

Michael Wetzel wrote:
I accept your point that the MT translation is completely useless, but my question is: What is it that makes your example special? Why was MT able to produce (outwardly) very plausible results at several points in this case?


To answer this question you would have to ask all our colleagues who use Google Translate for their daily professional working processes via Google APIs, i.e. in Studio, memoQ etc.: You will not only be given the results, but in return you also feed the system with your solutions.

[Edited at 2017-05-03 10:00 GMT]


 

John Fossey  Identity Verified
Canada
Local time: 04:12
Member (2008)
French to English
Not feeding the system May 3, 2017

Matthias Brombach wrote:
To answer this question you would have to ask all our colleagues who use Google Translate for their daily professional working processes via Google APIs, i.e. in Studio, memoQ etc.: You will not only be given the results, but in return you also feed the system with your solutions.


I don't believe the last piece is true - nothing feeds the MT engine with the translator's solutions. There is no data going back to the MT API engine about how the translator used the translation (unlike the browser-based version, where corrections are taken note of).

I will sometimes have the GT window running off to one side, when allowed, and it can occasionally suggest a turn of phrase I hadn't thought of, somewhat like a sophisticated but not very accurate dictionary.

[Edited at 2017-05-03 11:47 GMT]


 

Mirko Mainardi  Identity Verified
Italy
Local time: 10:12
Member
English to Italian
+1 May 3, 2017

John Fossey wrote:

Matthias Brombach wrote:
To answer this question you would have to ask all our colleagues who use Google Translate for their daily professional working processes via Google APIs, i.e. in Studio, memoQ etc.: You will not only be given the results, but in return you also feed the system with your solutions.


I don't believe the last piece is true - nothing feeds the MT engine with the translator's solutions. There is no data going back to the MT API engine about how the translator used the translation (unlike the browser-based version, where corrections are taken note of).

I will sometimes have the GT window running off to one side, when allowed, and it can occasionally suggest a turn of phrase I hadn't thought of, somewhat like a sophisticated but not very accurate dictionary.


+1 to that, but I was wondering about "unlike the browser-based version, where corrections are taken note of". Is that really the case?

If you're using GT, for instance, you can only edit the source, not the translation. It does have a "Suggest an edit" feature, but you have to go out of your way to use it.


 

John Fossey  Identity Verified
Canada
Local time: 04:12
Member (2008)
French to English
GT: Browser vs. API May 3, 2017

Mirko Mainardi wrote:

+1 to that, but I was wondering about "unlike the browser-based version, where corrections are taken note of". Is that really the case?

If you're using GT, for instance, you can only edit the source, not the translation. It does have a "Suggest an edit" feature, but you have to go out of your way to use it.


When you use the "Improve this translation" feature in the browser version, a message is displayed: 'Your contribution will be used to improve translation quality and may be shown to users anonymously". So evidently with the browser-based GT the edit is being resent back to Google.

But in the case of the MT API engine, a string of characters is sent to the API, which then sends a string of characters back in response. That closes the dialogue with the API, there's no further communication regarding that particular translation.

[Edited at 2017-05-03 17:02 GMT]


 

Mario Chavez  Identity Verified
Local time: 04:12
English to Spanish
+ ...
Astonishingly? May 3, 2017

I'm not surprised that this platform (older than The Matrix movies) subscribes to useless and uninformative hype. What's so astonishingly useful? What hacks write that sh1t?

And why are we even discussing the man vs. machine false dichotomy again?

I stopped by here because my Déjà Vu X3 project was running a pretranslate operation. Tools, CAT or otherwise, are simply extensions of our writing.


 

Mirko Mainardi  Identity Verified
Italy
Local time: 10:12
Member
English to Italian
Misunderstanding May 3, 2017

John Fossey wrote:

Mirko Mainardi wrote:

+1 to that, but I was wondering about "unlike the browser-based version, where corrections are taken note of". Is that really the case?

If you're using GT, for instance, you can only edit the source, not the translation. It does have a "Suggest an edit" feature, but you have to go out of your way to use it.


When you use the "Improve this translation" feature in the browser version, a message is displayed: 'Your contribution will be used to improve translation quality and may be shown to users anonymously". So evidently with the browser-based GT the edit is being resent back to Google.

But in the case of the MT API engine, a string of characters is sent to the API, which then sends a string of characters back in response. That closes the dialogue with the API, there's no further communication regarding that particular translation.


I think there was a misunderstanding. I was not talking about API vs Browser behavior. I was referring to your "unlike the browser-based version, where corrections are taken note of" to say that I was under the impression that there is in fact no automatism by which the browser version does "improve" translation results while you're simply using it. That is all.


 

Siegfried Armbruster  Identity Verified
Germany
Local time: 10:12
Member (2004)
English to German
+ ...
Domain knowledge is king May 3, 2017

Michael Wetzel wrote:

What is it that makes your example special?


The sample is special because it is a regulated text. This means that the translator has to identify the text as being a regulated text, he/she has (for this example) to know the about 30 documents and at least 3 different databases from at least 5 regulatory bodies. These documents specify certain parts (paragraphs, terminology) in each language. And even worse, these "reference documents/databases" do change, meaning the huge databases required to train the NMT systems will be full of outdated data.
I have been involved in the development of clinical solutions using neural nets since 1996. There is no magic and there is no AI.
We should step back a bit:

I do not know professional using an axe or a two-men saw to cut a tree
I do know a lot of professsionals using chainsaws to cut down trees
I know a few professionals using timber harvester

None of them talks about AI or NeuronalTimber harvesters, they all use and control their tools - and this is how we should do it too. There are a lot of tools that can speed up/automate certain aspects of (specific) translation workflows. NMT is, as SMT was for the last few years, just the flavor of the month, it will help laymen to understand foreign texts.

BUT do you realy want to have surgery by a surgeon who follows a Google NMT translated manual on how to implant your new whatever implant, or would you enter a plane where all pilot checklists were translated by Google NMT? As soon as somebody has to take responsibility, MT/SMT/NMT is out of the game. It can help to produce a professional translation, but it is the specialist who is doing it in the end and not the MT system.

[Edited at 2017-05-03 22:07 GMT]


 

Mirko Mainardi  Identity Verified
Italy
Local time: 10:12
Member
English to Italian
"Regulated" vs. creative texts May 4, 2017

Siegfried Armbruster wrote:

The sample is special because it is a regulated text. This means that the translator has to identify the text as being a regulated text, he/she has (for this example) to know the about 30 documents and at least 3 different databases from at least 5 regulatory bodies. These documents specify certain parts (paragraphs, terminology) in each language. And even worse, these "reference documents/databases" do change, meaning the huge databases required to train the NMT systems will be full of outdated data.


And yet, it seems to me a (tailor trained) MT engine would fare better in such a scenario, with set terms, expressions and phrases, than it would where (trans)creativity is required.


 

Siegfried Armbruster  Identity Verified
Germany
Local time: 10:12
Member (2004)
English to German
+ ...
CAT - Computer Assisted Translation versus (S/N)MT May 4, 2017

Mirko Mainardi wrote:

And yet, it seems to me a (tailor trained) MT engine would fare better in such a scenario, ......


Don’t get me wrong, I am a huge fan of automation and „my computerized translation process“. I love chainsaws, harvesters, GPS based machines etc.

My point is, there is too much HYPE about the (N/S)MT aspect (and too much risk capital and Government money involved).

I see much higher gains in productivity and quality assurance in other process steps. My process (in which translation is just one step) allows me to translate 7-10.000 words/day in SPCs and PILs in high quality with validated results (i. e. I can prove that I checked the translation and that it is correct). Show me any MT system that can do the same thing and I’ll buy it.

And I have to admit, I am into MT, but on my machine - I am using SLATE to produce a dedicated MT system that feeds my CAT (it is a complex project and I am not yet using it in my production environment). Specialized MT engines will be the specialized TMs of the future and they will be worth real money.


 

Michael Wetzel  Identity Verified
Germany
Local time: 10:12
German to English
Other way around ... May 4, 2017

I understand what makes the GT translation wrong in your case. That's not the special case I was referring to. (And I wouldn't want to read a press release or catalogue essay or scholarly article translated by GT either.)

I was referring to the special case of GT delivering very plausible and apparently (though nor actually) adequate results. Usually GT does not do anywhere near this well.
I'm also interested by GT's willingness to go out on a limb and bet the farm on the assumption that the English "leaflet" in this short sentence almost has to mean "Beipackzettel". It makes sense for SLATE to do that on your machine, but I'm wondering why GT is programmed to take this risk and how often these kinds of decisions backfire.
And like Mirko said, the fact that we are dealing with a highly regulated text type makes it seem all the more surprising that it could get things so close to right while simultaneously getting them absolutely wrong.


 

Anton Konashenok  Identity Verified
Czech Republic
Local time: 10:12
English to Russian
+ ...
Agree with Siegfried May 5, 2017

MT of any flavour will inevitably keep making stupid mistakes until it learns to actually understand the meaning of the text being translated. There are fields where this understanding is truly necessary to avoid serious risks - medicine, aviation, nuclear energy, etc. - and human translators without any background in the given field do make stupid mistakes, too. Sigfried's example of regulated text is actually one of the simpler cases - after all, MT could have possibly provided a correct translation if it "looked" at the whole paragraph and this paragraph were in the database as a single entity. Here is a trickier example from the aviation field: in some languages, one and the same word can be translated into English as "altitude", "height" or "elevation" depending on what it denotes: if it is about the vertical distance from the aircraft to the mean sea level, it's altitude; if it is about a vertical distance between the same aircraft and the ground it is currently overflying, it's height; and if it is about the vertical distance between the point on the ground and the mean sea level, it's elevation (in other words, elevation + height = altitude). Problems start when several of these appear within the same sentence (which is totally normal). Confuse "decision height" with "decision altitude", and the aircraft may fly into the ground, killing everyone aboard. I have personally seen an officially approved version of the airline operations manual written in Russian with some English terms given in parentheses for reference, where a Russian phrase about the decision height contained "(decision altitude)" as an English comment. (Please don't ask me what airline it was, it has already gone out of business anyway.)

 
Pages in topic:   [1 2] >


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

Moderator(s) of this forum
Jared Tabor[Call to this topic]

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

Translation platforms cannot replace humans-- but they are still astonishingly useful

Advanced search







TM-Town
Manage your TMs and Terms ... and boost your translation business

Are you ready for something fresh in the industry? TM-Town is a unique new site for you -- the freelance translator -- to store, manage and share translation memories (TMs) and glossaries...and potentially meet new clients on the basis of your prior work.

More info »
Déjà Vu X3
Try it, Love it

Find out why Déjà Vu is today the most flexible, customizable and user-friendly tool on the market. See the brand new features in action: *Completely redesigned user interface *Live Preview *Inline spell checking *Inline

More info »



Forums
  • All of ProZ.com
  • Term search
  • Jobs
  • Forums
  • Multiple search