Skip to main content

UNDERSTANDING AI TRANSLATION EDITING DOUBLE SPEAK


By Carlos de Paula


AI seems to have taken the world by storm, including the translation business.


One thing you might have noticed is that translation companies have surrendered to the fact that a lot of what will remain of the translation business may be editing of translations done by AI. However, you should know exactly what you might be buying from a vendor, even one you share a great working history. A large part of the industry is in survival mode, and everything might be done for a buck.


In a nutshell, you might be offered different levels of service, from less to more expensive.


These may be referred to as:

RAW MACHINE TRANSLATION

LIGHT HUMAN POST-EDITING TRANSLATIONS

FULL HUMAN POST-EDITING

 

I will start with the least expensive, something often referred to as RAW MACHINE TRANSLATION.


Believe it or not some companies are actually offering this to clients for a cost. It is the type of thing you should run away from. Why should you pay a company to run huge texts on a translation platform, when you can do it for free? Sure, they will tell you the platforms they have are state of the art, not your regular google translate, but they are just as bad. They might tell you they have industry specific databases, proprietary glossaries for every conceivable subject and all types of stories to convince you, but they might do what you would: run the text and hope it is good. More often than not it is bad, but they will charge you for it anyway.



The longer a text is, the worse the result. AI translation is very faulty when it comes to context, and normally you must break down the text in sections, because it will not translate 30k words in a single session. Heavy formatting will result in worse output. Text from one section might be moved to another, and whole sections might be lost in coding. 



If the original text comes from a PDF conversion, results might be even worse. Machine translation often skips whole sections of long text, without explanation or warning, and sometimes repeats whole sections, without reason or warning. Additionally, OCR has severe trouble with accents, formatting, and often fails to convert or hide large sections of huge files. Conversions of power point presentations are even more disastrous.



Using Raw Machine Translation is asking to be fired, lose a case, or anger/offend a client. With some luck, you will come out unscathed. Unless you are well versed in both languages, and have the time to thoroughly edit yourself, you should avoid it at all costs. It is indeed a raw deal.



LIGHT HUMAN POST-EDITING TRANSLATIONS.

 

This is more expensive than Raw Machine Translation, and what they promise is light human editing,  in other words, a qualified translator who knows well both languages, and will make corrections to the text without going crazy. More than likely, you will get text that has been machine proofread and machine edited. With some luck someone with different skill sets might read the entire document and make a couple hundred changes in 30k words, most in the beginning of the text, and you will feel you got a great deal, maybe even great praise from the boss.  



When translations are done into English, editing is often done in India, where the editor might have a fair command of English, but understand little to nothing of the source language. In other words, the English might sound good, but entire sections of it will not correspond to what the original text says.



Then there are the “little AI helps”, over translation that can bring dire consequences. I mentioned the other day that AI had over translated ML Account, in a long petition, as “Money laundering account” into Spanish. This was done  several times, although ML was clear a reference to a well-known Wall Street, incidentally, mentioned in the text more than once. In the interest of speed someone might have been indicted or investigated for money laundering, if I did not pick up the faux pas.



Such errors are unlikely to be detected by light human translators who are being pressured to deliver huge output for little pay. Remember that agencies are middlemen, and although most say they have specialized inhouse staff to do this kind of job, it is often subcontracted.



 At the end of the day, Light Human Post Editing is just slightly better than Raw Machine Translation.



FULL HUMAN POST-EDITING



There should not be no question in my mind, full human post editing is the only way to go for text translated by AI. No business translates text for sport: there is always a reason, and as much care should be used in the translation process as in the production of the original. Originals sometimes take weeks to be signed-off, myriad versions, rewrites, and sizable budgets in terms of person-hours. Perfect translations are expected to emerge in a brief time, for pittance. It has always been so, but now, in this age of AI and Machine Translation the scenario is gloomy. There is a perception that translation is nothing more than a mathematical operation that can be solved in seconds.



Maybe AI translation will be perfect one day. Who knows. But here are a few things that it does very badly that warrant having translated text edited by a qualified, experienced human:

- it messes up dates.

- it over translates things. Remember the ML account translated as Money Laundering account.

- it often includes words out of the blue that are nowhere in the original.

- it messes up formatting

- it is inconsistent changing commas and periods in numbers.

- it is inconsistent with terminology. It may translate X as Y in one line, in the next line it translates it as Z.

- it often uses the wrong translation, not only for technical terms. Even words that a 5-year-old uses can be mistranslated by AI.

- it makes positive statements negative, and vice versa. A definite danger in contracts.

- it excludes text that should not be edited out.

- it repeats whole sections of text.

- amazingly, it commits occasional spelling errors.

- it often fails to read irony, exaggerations and it is bad for context as well.

- it often translates words literally when it should not.

- it often does not identity errors made by OCR in PDF conversion

- it translates names when it should not (unbelievably bad in documents).

- last, but not least, it gets exceptionally low grades in syntax, grammar and style. 



There are some more issues, but I do not want to bother the reader. You get the idea.


In other words, considering the cost of having AI translations professionally edited is not that high, and considering that until recently corporate clients were willing to have text proofread and edited by someone other than the translator, it seems a good idea to have someone flying the plane for take off, other than leaving the job for the automatic pilot.


Think about that the next time you have a translation project.



Carlos de Paula is one of the top Brazilian Portuguese translators in the USA since 1982. And now a top Portuguese AI Translation editor as well. 


(877) 626-0642

WhatsApp (786) 543-1687

https://legaltranslationsystems.com


RETURN TO ARTICLES