TAUS Blog

  • Home
    Home This is where you can find all the blog posts throughout the site.
  • Categories
    Categories Displays a list of categories from this blog.
  • Tags
    Tags Displays a list of tags that have been used in the blog.
  • Bloggers
    Bloggers Search for your favorite blogger from this site.
in Quality

In the Shoes of Amélie, Post-editing Pioneer

Font size: Larger Smaller
Rate this blog entry:
0

We have interviewed Amélie Blumerel, Lead Linguist at Version Internationale, one of the LSPs contributing to the TAUS Post-editing course and certification. Amélie Blumerel, has been coordinating Version Internationale’s efforts to improve the post-editing skills of both internal and external teams. You will be able to listen to Amélie and ask her questions in our Post-editing webinar for French on 18 June 4PM CET. For more information on this new TAUS webinar series, please click here. Enjoy the interview!

When did you first start post-editing?
It must have been just over three years ago. One day, I was allocated a post-editing project although I'd never actually done one before. It involved translating an internal training programme raising awareness about corruption – around ten slides explaining to employees that it was wrong, for example, to give branded goodies in exchange for commercial services.

And how did it go?
I remember that the machine translation results were really bad at that time that it wasn't an easy job. On another project – this time a catalogue that had a lot of single terms and not many sentences – there were a lot of sentences in English that hadn't been translated by the machine, which hardly ever happens nowadays, and which needed translating directly in addition to the post-editing. I was concentrating so hard on the translation itself (meaning, terminology and sentence structure) that when my colleague re-read the work, she spotted really big errors in agreements that I never thought I would make!

At the time, what was your approach to this new way of working?
Well, it would be a lie to say that I was really enthusiastic! I wasn't really at ease and the excitement of being a "pioneer" was in practice replaced by anxiety at not being able to meet all my post-editing deadlines! With hindsight, I think that the first post-editing tasks I completed were not that good, but that just goes to show that you can really make progress! And then translation engines have got so much better that today's post-editing projects bear little similarity to those I did at the beginning.

What's your worst post-editing memory?
This involved a technical manual for a highly specific machine tool for the logging industry. What's more, it was from Italian to French, but the text was full of terms in English. The machine translation was really bad and it all had to be done again, involving a lot of extra research to ensure that the terminology was correct. It was a real nightmare…

What do you think of post-editing today? 
I think you have to look at the facts: post-editing enables you to work more quickly and to deal with high-volume projects in relatively short times. The financial stakes apply at all levels.

And in practice?
On a practical level, I've become more experienced and therefore more comfortable with post-editing. I know what mistakes I expect to find and where I have to pay special attention. For example, I know that the main issue with French involves the gender of words and that you have to be more vigilant than with human translation. The same applies for past participles, for example.

You are a senior editor at Version internationale. Do you prefer editing or post-editing?
Paradoxically, as post-editing involves editing texts produced by a machine rather than by human beings, I can actually do a bit more translating within a project. I have more control over the project, which I can handle in its widest context. Relatively speaking, it's a little more creative than pure editing.

What's the most complicated thing about post-editing?
It's not really "complicated", but it's sometimes fastidious to correct the endlessly repeated minor errors, such as problems with tags, spaces before a double sign, apostrophes, etc. That's why feedback on Machine Translation is crucial, to improve the engines wherever possible.

And what's the easiest thing?
Working alone… ☺ There's less interaction with other translators, which, in the end, simplifies the task.

Do you have any funny stories about inappropriate machine translations?
There's always one that makes you smile… For example, bars where you listen to rock ("rock bars") can become "barres rocheuses", which in French is a kind of rocky cliff! Another classic is when "disabled", as in "the default value is disabled", is translated in French as "handicapped".  

It adds enjoyment to the daily routine.

Is there a mistake that all beginners make?
Yes, there are mistakes that occur quite frequently, because when they begin, post-editors sometimes find it difficult to keep the overall vision of the project in mind. You must never forget that the machine-translated text is just a "suggestion". The machine translates the word "A" by "B", but sometimes in the following paragraph, in another context, it may translate it by "C". You therefore always have to be on the lookout and keep a step back throughout the post-editing to ensure consistency in terminology.

What advice would you give to a translator looking to start out in post-editing?
To do some. To do some without preconceptions and without becoming discouraged. To do some intensively, whilst meeting deadlines. In my opinion, you need between one and two years to become an experienced post editor. Above all, you need to persevere. 

Lastly, what advice would you give to a post-editor looking to improve? 
I think that the secret is to spot in a micro-second (no longer!) if the Machine Translation suggestion is worth keeping or not. If it's good, you keep it (and make sure you don't do too many preferential changes as that's not the aim!). If it's no good, you go directly back to the source chain. And then of course, opinions and corrections from people outside of the project are always welcome – they help you to spot recurring errors and get rid of them.

Attila Görög is responsible for the translation quality product line at TAUS. Attila’s challenge is to convince translation buyers and vendors about the flexible nature of quality by promoting tools, metrics and best practices on quality evaluation. He's been involved in various national and international projects on machine translation, terminology management and semantic web.

People in this conversation

  • Guest - Aveitos

    Way cool! Some extremely valid points! I appreciate you penning this article and also the rest of the website is also really good.

  • Guest - Kirti Vashee

    I think this type of feedback is much more useful if it is related to the specific MT solution that has been used. It is well known that RbMT and SMT make different kinds of errors, but even amongst SMT solutions some are much more capable of being fine tuned and responding to error patterns than others.

    Additionally I think some additional questions need to be asked e.g.

    Do you have experience with various kinds of MT or just the one described?
    Do you think other MT systems would give the same errors or different kinds of errors? e.g. the errors described here are very easily handled by the Language Studio platform in an automated way.
    Do you know how to apply corrective feedback to the MT system (not all MT can handle this) so that dumb, repetitive mistakes are quickly eliminated?
    Characterizing MT output as "bad" is not useful as corrective strategies are very closely linked to the proper identification of kind of error i.e. word order, word choice, writing style, punctuation, tag handling, wrong term etc... Each error category would have a different automated solution strategy.
    Do you do any kind or analysis of the translated content as many problems can be avoided with very small efforts?

    I think we have yet to reach a point where post-editors have seen how various MT platforms deal with the common translation challenges.

  • Guest - Dion Wiggins

    This is an interesting Q&A session but leaves many gaps. There are many different kinds of MT systems and many different suppliers of MT who have different technologies and approaches to MT. Articles such as these would be infinitely more useful to readers if they specified which MT tool is being referred to and the actions taken to improve the MT engine quality.

    Some MT systems such as Google are generic and you have no control. Some MT systems provide a basic level of control, while platforms like Language Studio provide a very deep level of control. Nearly all the problems described in this Q&A session are the typical issues that are occur with generic or often DIY customized MT systems. These issues are much less likely to occur in an expert customized system as work such as terminology normalization has been performed.

    Here are some additional questions that if answered I think will provide additional value to readers:

    1. What was the language pair(s) and domain(s)?
    2. What was the MT engine that you were were post editing MT output from?
    3. Have you tried other kinds of MT or just this one?
    4. Do you think other MT systems would give the same errors or different kinds of errors?
    5. You say the MT results were really bad. What specifically was bad about them? Word order? vocabulary choice? writing style? something else?
    6. When you saw errors, were you able to find solutions to the errors so that they did not occur again in future?
    7. Did you customize the engine? If yes:
    a) What kind (TMs, dictionary/glossary, monolingual data, etc.) and how much data did you provide to customize the engine?
    b) What quality was the data that was provided to customize engine?
    c) Did you perform any terminology work on the engine to improve term consistency?
    d) Did you apply any kind of rules to further control the translations?
    e) Did you provide post editing feedback to the engine to improve quality progressively?
    f) How mature was the custom engine? (i.e. had it had a lot of feedback, no feedback at all,
    8. Did you see the MT output prior to accepting the job from the LSP?
    9. Did the LSP identify the source of the MT?
    10. Were you able to negotiate a fair rate for post editing with the LSP based on the quality of the MT?
    11. Would you post edit the same kind of MT again?

Add comment

Blog Archive

Recent Comments

Guest - Aveitos
Way cool! Some extremely valid points! I appreciate you penning this article and also the rest of th...
Guest - NAZAR HUSSAIN KAZMI
Dear Friends

I am greatly impressed to know about your professional activities especially in the ...