Languages and linguistics for linguaphiles, polyglots, language lovers and language geeks!
Click HERE for the start page

Note/Disclosure: This blog contains affiliate links to third-party products and services, which means I may receive a small commission if you make a purchase after clicking on any of these links

A State of Flux - The Difficult (and Controversial) Task of Integrating Machine "Translation" Tools into Professional Translation

[tldr?: Machine translation plus low paid "editors" = content that is too unreliable for serious use; MT no justification for reduced rates since translators achieve little to no time savings from using MT]
 
The world of translation is in a state of flux. To the untrained, monolingual eye, machine translation seems like a miracle and it’s difficult to make clients understand just what could go wrong when they see what to them looks like an almost-finished product. But the machine doesn’t translate in the sense that it comprehends the input and reprocesses the text into another language with the same meaning. Instead it provides what looks like a translation.

It’s important to understand how the machine arrives at this illusion in order to properly understand the effort required to turn this illusion into a true translation and why reduced rates don’t make sense for a freelance translator.

In addition, statistical-based MT will probably get worse over time instead of better (see link to article below). It works by collecting a corpus of on-line human bilingual translations and finds matches between words and phrases based on their frequency and thus predicting which statistical match is the correct translation.

It was, at the time, an ingenious invention, and markedly better than the word-by-word MT that was previously available. If anyone is old enough to remember, that software frequently left many words and phrases untranslated because the words and terms were simply not in its database.

Modern MT never seems to “not know” the translation because it either leaves it out or “makes something up” or “finds the closet match” which may mean the exact opposite of the translation (which is why you still have to be able to read and understand the source language and not just quickly skim the ‘translation’ for stuff that ‘doesn’t sound quite right’.)

However, what the inventor did not predict was that as more and more people post machine translated content on-line and as more and more raw MT output - or poorly translated or "edited" text - is put on the web, the (possibly incorrect) MT output may eventually have a higher statistical frequency than the correct translation resulting in a cascade failure where the translation actually gets worse instead of better. (Those of you who use the LINGUEE dictionary tool, which is based on the same principle, have probably already seen this happen when you search for terminology, and this is just with human translation errors).

The website ProZ has a feature where you can ask fellow translators about suggestions for the translation of difficult terms. I remember that I always ended up being more confused than before I posted the question since every translator seemed to have a slightly different suggestion. Indeed, there are many different ways of translating the same content. I can't help but wonder, when the computer does the translating, who makes the decision about which translation to use and which one best fits the present context in order to program the machine?

As a result, some MT systems have adopted a system whereby they try to use AI to identify when a text has been machine translated before it is used in its MT corpus, but it’s not entirely successful.

In other words, you still have to have the knowledge and experience to understand when and where the translation needs correcting and be able to conduct the terminology research that MT does not. Even when a sentence is 100% correct, you still have to have the language and subject knowledge to recognize this.

In short, MT does all the easy bits, leaving all the hard work to the translator. You have to carefully read and parse the output or you could miss a crucial translation mistake or omission. Working for less money means you will have to work faster in order to pay your bills, resulting in crucial errors, wrong terminology, incorrect nuances, omissions, insensitive language, gender biases, etc…

So it definitely can be a useful tool and in some cases it may provide clients with a translation that is “good enough”, but this doesn’t mean that translators should be paid less for their work (especially when rates have already been relatively stagnant for 20 years). And the idea that machine translation will allow you to work faster is a myth (see the Bonus Story below under the links).

The combination and use of MT and translation automation with professional human translators should be promoted and marketed as a benefit that improves the quality and consistency of translation rather than as simply being cheaper and faster. It’s not an exception to the holy trinity of “cheap / fast / good - pick two”.

I have even seen companies that pay translators based on the percentage of the MT output that was changed. Under this system, the translator’s ability to read and understand the source and target text in order to identify the fact that the MT was suitable is not compensated. As a result, translators are forced to make unnecessary changes to the output in order to ensure that they will get paid.

More recently, the trend has been to claim that whatever MT system the company uses is “better” than out-of-the-box products or free on-line tools because it uses something called “neural translation” or “AI” or “we use an enhanced proprietary MT system.”

I recently completed a 40,000 word project where the client provided a “neural” machine translation. It still took me two weeks to do the terminology research and complete the translation (with a CAT tool). I couldn’t imagine doing all that work for less money just because I had that initial rough MT template to start with. In that case, I would probably have been making less per hour than minimum wage.

Millions of dollars have and will be invested in machine translation and these systems are marketed as the panacea to the world’s communication problem. The profession of translator is in a state of flux right now as people are trying to figure out how to best integrate the new tool of MT into the translation process - especially when MT is available for practically free to anyone (most are trying to do so to reduce costs which, in my opinion, is not the way to go). 

Yes, machine translation is the future of translation, but it’s a future tool for translators not a replacement for human beings and not a substitute or excuse for not paying professional rates. In fact, translators will have to be even more professional in the future since with the help of MT, anyone can pass themselves off as a “translator”, as many do on some other sites.

The idea that instead of paying someone a living wage, you can pay someone a low rate to edit MT and then pay a second person an even lower rate to verify the first person’s work and that you will somehow end up with the same quality is fallacious. You are just replacing one person who cares about the quality of their work because they are adequately compensated with two "gig workers" or "Mechanical Turks" who are not morally, financially and emotionally invested in their work.

Related articles:

Google admits ‘garbage in, garbage out’ translation problem: https://www.theregister.com/2014/02/06/google_translate_issue/

The Shallowness of Google Translate: https://www.theatlantic.com/technology/archive/2018/01/the-shallowness-of-google-translate/551570/

MTPE takes more time and energy than human translation for poorer results: https://www.at-it-translator.com/i-dont-offer-machine-translation-post-editing-here-is-why/

Bonus Story:

Last year I was contacted by an agency with a “post editing” job. It was a 200,000 word project (to be split up in a million tiny pieces of course). This company sent me a list of their four levels of machine translation editing quality that went something like this.

Level 1: Raw MT output you will receive:
“After product on market launch patients complained sporadically allergic reactions.”

Level 2: Light editing (only correct serious errors, ignore incorrect syntax and what is otherwise understandable):
“After product launched on market, patients complained sporadically about allergic reactions.”

Level 3: Detailed context-level post editing:
“After the product was launched on the market, patients complained sporadically about allergic reactions.”

Level 4: Full complete translation:
“After the product was launched on the market, there were sporadic complaints from patients regarding allergic reactions.”

The goal, they said, was to provide a post-edited translation of at least level 3, but preferably level 4.

Does it save you any time to rewrite Level 1 to look like Level 4 rather than simply just translating from the source yourself? Not if you understand the language.

Then they included the actual document to be “post edited”, which looked like this:

“The agreed according to paragraph 7 deemed value guaranteed is done in accordance with indexation clause which when beyond this threshold, represents the change to the full extent and forms the initial basis for the calculation of further excedances.”

and

“Turning work includes 4-axis precision bearing machines on-turning milling vertical machining for drive standings”

Seriously?


 

No comments: