Pages in topic:   [1 2] >
Machine Translation – Post Editing should we take on these jobs, for how much?
Thread poster: Fi2 n Co

Fi2 n Co  Identity Verified
Portugal
Local time: 06:22
English to French
+ ...
Aug 28, 2017

Hello there 😊

I’ve been approached recently to perform MTPE (Machine Translation – Post Editing).
I think taking on such jobs would require some reflexion by the community, here’s why:

Many of us use MT to work, and there are some language pairs where results have improved a lot as you can see in these two threads:
http://www.proz.com/forum/translation_news/314860-machines_to_replace_human_translators_they_already_have_english_tabloid.html
http://www.proz.com/forum/machine_translation_mt/312626-google_gave_their_mt_a_"brain"_google_neural_machine_translation.html

As I explained in this video: https://youtu.be/irYnSJX3Yxw
Neural Machine Translation requires training. Although we can’t say for sure, chances are MTPE is going to be just that for the foreseeable future.
This means that the human translator is handed what the Machine Translation has guessed and has to edit and/or proofread it. In turn, the MT is presented with the optimised translation provided by the human translator. If the MT is a Neural MT, it will learn from the translation that the human translator produced. In time, the MT will then be able to get a result on its own closer and closer to what the human translator does. That’s what they’re aiming for anyway.

So, when someone approaches a translator asking for our services to do that kind of job, this raises a few questions:

_Should we accept a job that has the goal that in years to come human translators will be a lot less required?
_the last offer I had was 0,02 EUR/word on the claim of the large volume (which I turned down off course). Should I lower my rate for such jobs spending a lot of time correcting very poor translation achieved by a machine that will be competing with my services in the future
_Should I charge extra then?
_What should the translators community think of all this?

I look forward to know how the majority feels about this, since this may very well have an impact on our future!

My bests to all


 

Anton Konashenok  Identity Verified
Czech Republic
Local time: 07:22
English to Russian
+ ...
My view Aug 28, 2017

I have a rigorous quality policy, and in my experience, editing the MT output to that quality level takes longer than translating the same text from scratch. Accordingly, my post-editing rate is higher than my translation rate.

 

Fi2 n Co  Identity Verified
Portugal
Local time: 06:22
English to French
+ ...
TOPIC STARTER
Agree Aug 28, 2017

Anton Konashenok wrote:

I have a rigorous quality policy, and in my experience, editing the MT output to that quality level takes longer than translating the same text from scratch. Accordingly, my post-editing rate is higher than my translation rate.


Sounds about right, should we agree to take on these jobs.


 

Samuel Murray  Identity Verified
Netherlands
Local time: 07:22
Member (2006)
English to Afrikaans
+ ...
@Fi Aug 28, 2017

Fi2 n Co wrote:
I’ve been approached recently to perform MTPE (Machine Translation – Post Editing).


You should raise or lower your rate for this type of work based on the amount of skill, effort and time doing this kind of job takes. Unfortunately, as is the case with any new type of work that you're not familiar with, you're going to have to take a bit of a risk at first with your rate, until you have learnt what *your* speed in this process is.

By the way, the old non-neural MT also "learned" from translations submitted by the public. The new neural MT produces sentences that are grammatically much better than the non-neural machines, but they still make mistakes, and the sentences are not all stylistically consistent. The upshoot of this is that your job as a Google Translate MTPEer will in future involve less fixing grammar errors, and more checking whether the translation is correct or not.


 

Teresa Borges
Portugal
Local time: 06:22
Member (2007)
English to Portuguese
+ ...
I don't... Aug 28, 2017

I have never been asked to proofread MT, but lately I have been refusing most proofreading projects because I prefer to spend my time translating.

 

Fi2 n Co  Identity Verified
Portugal
Local time: 06:22
English to French
+ ...
TOPIC STARTER
Correct, approach? Aug 28, 2017

Samuel Murray wrote:

You should raise or lower your rate for this type of work based on the amount of skill, effort and time doing this kind of job takes. Unfortunately, as is the case with any new type of work that you're not familiar with, you're going to have to take a bit of a risk at first with your rate, until you have learnt what *your* speed in this process is.


Yes, thank you that would have to be taken into consideration. My approach on this is thread is more to ask whether that's even something experienced translators should consider. Hence where colleague's opinion is very much needed too.

Samuel Murray wrote:

By the way, the old non-neural MT also "learned" from translations submitted by the public. The new neural MT produces sentences that are grammatically much better than the non-neural machines, but they still make mistakes, and the sentences are not all stylistically consistent. The upshoot of this is that your job as a Google Translate MTPEer will in future involve less fixing grammar errors, and more checking whether the translation is correct or not.



That's correct. I didn't mention old style MTs because most of effort now is going towards NMT although PBMTs (old style) are still in use, mostly somehow connected and paired to the new ones. And Neural MT should be able to leverage more quality from the editing experienced translators are providing. That makes the topic even more relevant I think. By the way I wasn't approached fo a Google MTPE job, but that doesn't matter.

My bests


 

eyepod  Identity Verified
United States
Local time: 22:22
English to Spanish
+ ...
Many people already do post editing Aug 28, 2017

Many translators are basically already doing this since their CAT tool has a default MT, possibly even Google translate. I also don't think you can just say "I never accept those jobs" unless you want to be left behind by technology and the industry.

I wouldn't want to do post editing without the original because MT can get pretty confused at times.


 

Peter Shortall  Identity Verified
Local time: 06:22
Member
French to English
+ ...
An extra step in the process Aug 28, 2017

I don't take on MT work, I tried it once and didn't like it. Like Anton, I found that it was actually more time-consuming than translation. With ordinary translation, there are three steps to the process:

(1) read the original
(2) think of the translation in your head
(3) type the translation

With MT, there are four steps:

(1) read the MT text
(2) compare the MT text with the source text
(3) work out in your head what the correct translation should be
(4) type the translation

And at step 4, although some of the words that have already been typed can be left as they are, you end up using your mouse a lot to click on parts of words that need to be changed or replaced entirely, so your typing is interrupted a lot. It's easy to make mistakes when you have to interrupt your reading and typing all the time, so you have to read extra-carefully, and that slows you up. I find it quicker to type a sentence from scratch without stopping and starting all the time, and I feel the likelihood of errors is lower (not zero, of course, but lower).

This is why I feel that anyone thinking of doing MT should consider charging a higher rate than for translation from scratch if they want it to be worth their while. Though of course, persuading clients to pay that higher rate is another thing!


 

Samuel Murray  Identity Verified
Netherlands
Local time: 07:22
Member (2006)
English to Afrikaans
+ ...
@Peter Aug 29, 2017

Peter Shortall wrote:
With ordinary translation, there are three steps to the process:
(1) read the original
(2) think of the translation in your head
(3) type the translation


What you describe is is how super translators do it (i.e. translators with super short-term memory). I can do that only for very short segments.

For me, moving the cursor back and forth in the partially completed segment, and adding/deleting/moving text, is a normal part of translating.

With MT, there are four steps:
(1) read the MT text
(2) compare the MT text with the source text
(3) work out in your head what the correct translation should be
(4) type the translation


The process for MT should be the same as the process for proofreading another translator's work. And when you're proofreading, you're not just typing your own translation while using the translator's words merely as a typing aid, right?

...you end up using your mouse a lot to click on parts of words that need to be changed or replaced entirely, so your typing is interrupted a lot.


I never the mouse inside a segment. Good grief. (-: Adding, deleting and moving text, and moving the cursor, inside a segment, should all be done using the keyboard.



[Edited at 2017-08-29 09:01 GMT]


 

Michael Wetzel  Identity Verified
Germany
Local time: 07:22
German to English
The worst teachers that money can buy? Aug 29, 2017

In the discussions here, most people clearly seem to feel that most of the rates offered for most MT projects are too low when converted into an hourly income based on a reasonable level of translation quality. (Feel free to substitute "almost all" for "most" in each case, but my argument doesn't depend on it.)

Some people state that they or we need to do MTPE, but no one seems happy about it or even willing to do it - unless they consider the market to be forcing them to do so.

My question is: Is neural MT getting worse? If it is learning from editors' input and that input is being provided by the worst editors that money can buy, wouldn't it be plausible to think that a lot of what it is being taught is nonsense?


Thierry Guitton
 

Sharon Sarah Schmitz  Identity Verified
Germany
Local time: 07:22
English to German
Time to move on? Aug 29, 2017

Samuel Murray wrote:

The upshoot of this is that your job as a Google Translate MTPEer will in future involve less fixing grammar errors, and more checking whether the translation is correct or not.



Google Translate is no longer top of the bill. Linguee just launched its new DeepL platform at https://www.deepl.com/translate - and I gotta say that I'm rather impressed by some of the translations it offered when I checked it with medium-level financial stuff and snippets from some machine tool user manuals. Of course, it's not perfect, and it doesn't provide consistent quality with everything you throw at it, but it's already good enough to cover significant sections of the market. I think it offers 13 languages at the moment. So, MTPE might actually be the present, not the future.

And no, I'm not a DeepL advertising bot.


 

José Henrique Lamensdorf  Identity Verified
Brazil
Local time: 02:22
English to Portuguese
+ ...
If we take them, it should be at the same rate as traditional translation Aug 29, 2017

Michael Wetzel wrote:

My question is: Is neural MT getting worse? If it is learning from editors' input and that input is being provided by the worst editors that money can buy, wouldn't it be plausible to think that a lot of what it is being taught is nonsense?


I have PROVEN that "neural" MT is actually worse than its nameless predecessor, plain MT, at https://goo.gl/ReRQrZ

Let's make an analogy, assuming that I run a water treatment plant. I pump the water Mother Nature gives me for free from a nondescript river, treat it (it's a SERVICE, like translation) adequately, and sell potable water by the gallon. A prospect comes to me with a tanker carrying so many thousand gallons of water he got elsewhere, and asks me how much cheaper I'd charge to treat HIS water into potable.

Why should I charge cheaper for the same service? In fact, to avoid fumbling with hoses and pumps, the most sensible thing to do is to tell him to dump his tanker's contents on 'my' river, and - since his tanker is there - sell him 'my' potable water at regular price!

The Law of Least Effort has been there longer than us, humans. Presumably water already ran downhill before our species came to be.

So when we face MT that is barely good enough to be left as it is, we'll take no steps to improve it, thus causing a general downgrade to the final translation quality, overall.

It's a sign of the times. The automotive industry has been through it. Previously, any skilled mechanic could troubleshoot any malfunction by examining the various car parts in operation. I used to carry tools and some spare parts in the trunk, so I could fix many problems myself. Nowadays, with all that on-board electronics, all a 'mechanic' has to (and can) do is to is to plug a computer to the interface under the vehicle dashboard, and that computer will tell him to 'replace the blue box under the hood', just that. (Now I don't even have a screwdriver in my car.)

Every week I read reports from customers who bought expensive cars from renowned brands, had severe trouble during the warranty period, and neither the authorized shop mechanics nor the factory engineers could figure out what's causing the malfunction to fix it. They spend MONTHS taking their new car to and fro the shop, to no avail, until they give up, sell it, and try their luck with a different brand. It's all over both the online and offline media.

This is what MT/PEMT promoters are trying to do with translation.


 

Mirko Mainardi  Identity Verified
Italy
Local time: 07:22
Member
English to Italian
Interesting consideration Aug 29, 2017

Michael Wetzel wrote:

My question is: Is neural MT getting worse? If it is learning from editors' input and that input is being provided by the worst editors that money can buy, wouldn't it be plausible to think that a lot of what it is being taught is nonsense?


There would be some kind of poetic justice in that...


 

Fi2 n Co  Identity Verified
Portugal
Local time: 06:22
English to French
+ ...
TOPIC STARTER
Just for fun Aug 29, 2017

José Henrique Lamensdorf wrote:

I pump the water Mother Nature gives me for free from a nondescript river, treat it (it's a SERVICE, like translation) adequately, and sell potable water by the gallon.


Just for fun I took your above statement, which is a rather complicated sentence with "nondescript" making it even tougher to translate. I pasted it and ran it in the DeepL website mentione by Sharon https://www.deepl.com/translate

That's what I was handed in French : Je pompe l'eau que Dame Nature me donne gratuitement à partir d'une rivière non descriptive, je la traite correctement (c'est un SERVICE, comme une traduction) et je vends de l'eau potable au gallon.

It doesn't look too bad at all! And I didn't check the other languages offered. But maybe it's only french that works well in this NMT, I don't know.

My bests


 

eyepod  Identity Verified
United States
Local time: 22:22
English to Spanish
+ ...
DeepL does look good Aug 29, 2017

[quote]Sharon Sarah Schmitz wrote:

Samuel Murray wrote:

The upshoot of this is that your job as a Google Translate MTPEer will in future involve less fixing grammar errors, and more checking whether the translation is correct or not.

Google Translate is no longer top of the bill. Linguee just launched its new DeepL platform at https://www.deepl.com/translate - and I gotta say that I'm rather impressed by some of the translations it offered when I checked it with medium-level financial stuff and snippets from some machine tool user manuals. ...


Thanks for the link to DeepL, a quick look makes me want to explore more. And this is the point that I was trying to make, although I see I am in the minority. I just want to remind everyone that computational solutions in many fields are the future. There is an incredible amount of money being spent on MT research and development, so it's very possibly going to be getting better and better.


 
Pages in topic:   [1 2] >


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


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

Machine Translation – Post Editing should we take on these jobs, for how much?

Advanced search







SDL MultiTerm 2019
Guarantee a unified, consistent and high-quality translation with terminology software by the industry leaders.

SDL MultiTerm 2019 allows translators to create one central location to store and manage multilingual terminology, and with SDL MultiTerm Extract 2019 you can automatically create term lists from your existing documentation to save time.

More info »
BaccS – Business Accounting Software
Modern desktop project management for freelance translators

BaccS makes it easy for translators to manage their projects, schedule tasks, create invoices, and view highly customizable reports. User-friendly, ProZ.com integration, community-driven development – a few reasons BaccS is trusted by translators!

More info »



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