I. General information
There are many cases in which machine or neural translation services can be used.
However, the complexity of language, context and details continue to cause problems for machine-based translations and require extensive human involvement.
Please note that the term neural is merely being adopted today for the purpose of distinguishing the new machine translations from the old ones.
In the meantime, no leading service providers of machine translations are operating with the old system, as it is vastly inferior to the new so-called neural translations.
The terms are used interchangeably here.
II. Cases where machine or neural translation can be used
i. Only the big picture is needed
For example, if you hear about an article in a language you cannot read or if you receive correspondence from a customer who writes in a language you do not understand, the use of a machine such as google translate can be very helpful.
ii. An imperfect translation can be used for information purpose
If a company needs to generally inform its staff about the latest developments in the company, e.g. in a newsletter or a company newspaper, and provides the material in two or more languages, it might make sense to use a machine for the translation and have the resulting text only proofread.
iii. Example and analysis of good machine or neural translation
Here is an example:
Überwachung der Einhaltung des geltenden Datenschutzrechtes sowie der Strategien der Atlas Copco für den Schutz personenbezogener Daten einschließlich der Zuweisung von Zuständigkeiten, der Sensibilisierung und Schulung der an den Verarbeitungsvorgängen beteiligten Mitarbeiter und der diesbezüglichen Überprüfungen.
Supervision (corr: Monitoring) of compliance with applicable data protection legislation and Atlas Copco’s policies (corr: strategies) for the protection of personal data, including the assignment of responsibilities, awareness and training of employees involved in the processing operations and related checks.
This translation by the machine is fairly good. The English version looks flawless if you do not know the German. In isolation, that is, without regarding the rest of the text and the harmonization of terms, a comparison of the source and target text reflects extensive correspondence. In the larger context (and possible in any context), Überwachung should be translated as “monitoring” and no matter what, especially with German clients, you cannot translate Strategien as policies.
On the whole, however, this neural translation is understandable and conveys the content of the original. If the details mentioned are not particularly important and inconsistencies throughout the larger text can be tolerated, then a company could use such a translation and essentially incur no costs.
III. Problems with machine or neural translation
i. Inconsistency in translated terms
If a term has two common translations, it is still not possible to get a consistent result. For example, the German term Kunde can be translated as client or customer. If you run a text through a machine, you will get a mixture of these translations in the result. Naturally, you can use “search and replace”, but this happens with many different terms and in many contexts. “Search and replace” is only sufficient for harmonization in the rough draft. If the text is a contract or agreement, you will not want identical terms translated incorrectly and will have to go through all the usual steps even after the “search and replace” procedure.
ii. Weird results and errors
Recently, we were asked to work with a machine and revise the text. The neural translation produced by the machine looked like this:
Confidentiality is ensured by measures for access control, access control, access control and separation control.
This unfortunate translation came from this German sentence: Die Sicherstellung der Vertraulichkeit wird durch Maßnahmen zur Zugriffskontrolle, Zutrittskontrolle, Zugangskontrolle und Trennungskontrolle umgesetzt.
As Zugriff, Zutritt and Zugang can all be translated as access in many cases – in fact, this is often the best translation for each term – the machine used them in each individual case. However, this is impossible here since something different is meant by each of these terms in the context of the EU General Data Protection Regulation.
The machine does not know these individual differences and cannot produce the correct translation:
Confidentiality is ensured by measures for data access control, physical access control, system access control and separation control.
iii. Mistakes due to context
Although machine and neutral translations have improved substantially since the early days of google translate, they still encounter substantial difficulties (at least in German) if the text has nested sentences, complex syntax, idiomatic expressions, oblique references or tags. Any of these characteristics, along with others, confuse the machine and lead to an incorrect or meaningless result.
In translation from German to English, machines often have trouble with a word or formulation that can be translated in two (or more) different ways depending on the context.
Unserer Mandantin steht des Weiteren eine über den unter Ziff. I. dieses Schreibens genannten Betrag hinausgehende Forderung über € 21.237,68 brutto zu, welche noch nicht im Mahnantrag enthalten ist.
In addition, our client is entitled to a gross claim in excess of the amount stated under Section I. of this letter in excess of (corr: and totaling) €21,237.68, which is not yet included in the request for payment.
The machine gets confused here because of the formulation “darüber hinausgehen” + “Betrag über”, possibly thinking that there is a mistake here and guessing at the intended meaning (something the machines are not bad at in many cases, but incorrect on here). This sentence turns out categorically wrong in legal correspondence between lawyers. This cannot be permitted.
IV. The problem with revising machine/neural translation
At first glance, it seems that the translation can be done by machine with a human revising and proofreading. Instead of paying for the translation and revision, it should now only be necessary to pay for revision and proofreading.
For imperfect results, this is certainly the case, and one can proceed accordingly.
But if you need perfection, for example in academic papers, marketing material, contracts, lawsuits, case law or court rulings, the situation is much more complicated.
Firstly, should there be more than two or three errors every 15 to 20 words, it actually takes longer to correct the errors in the machine-produced sentence than to type up the sentence yourself.
Secondly, when you correct the machine-produced translation, you are still reading the text for the first time. This first reading while producing the translation is never perfect. There is too much complexity and detail to take in everything in that reading. Hence, proofreading is critical. However, when you correct a machine-produced translation, you do not quite read as intensively and accurately as you do when you actually prepare the translation. Furthermore, you are distracted by all the simple alterations that must be made, such as ensuring the Kunde is translated consistently as client or customer and numerous other simple details that do not demand much concentration if you prepare the translation yourself. This shifts more of the internalization of the text to the second and third round – the revisions.
The revisions with a comparison to the original become even more critical with machine-produced translation. This third factor changes quite a bit with machine translations. Previously, you were confident about the general translation and looked for minor mistakes in your reading of the original, e.g. if you read a word incorrectly. Now you have to look for that as well as the general meaning and sufficient mirroring. Essentially, three tasks fall upon the reviser of machine-translations: i) Is the translation generally correct? ii) Does it sufficiently mirror the original? iii) Where are the mistakes? That demands a large quantity of brain capacity and naturally makes another round of revisions indispensable. These two rounds of revision against the original continue to be necessary with machine-based translation.
Fourthly, you still need to proofread the final version without comparison to the original. A clean proofreading is still necessary. Even if fewer mistakes are found with machine-based translations, the large quantity of corrections means that typos, double words, etc. slip in.
Finally, there remains the question of whether an external reviser is desirable. Despite these technical advances and many years of experience, it is still likely that a word or phrase will be misread every 5-10,000 words.
V. Machine and neural translations at PTS
i. Specializing in complex texts
We specialize primarily in fairly complex texts where clients need perfect results. Whether these are academic papers, contracts or court decisions, financial statements/annual reports or other material, the work generally requires perfection. This makes machine and neural translations worthless.
The potential benefit from machine translation desired by customers – namely, a reduction in costs – is achieved with us at PTS because we are a small company that only offers translations from German into English. This lets us keep down the costs while providing high-quality service.
ii. Openness, especially in the case of pure proofreading
If specifically requested by a client, we are willing to discuss the use and post-editing of machine or neural translations. We are particularly open to this if a customer provides us solely with an English text prepared by a machine without requiring any comparison to an original. In such a case, we would only be responsible for the proofreading, a service usually provided for 0.04-0.05 EUR/word.
iii. Prohibition of machine and neural translations
Should it be desired, we can also include a provision in our Service Level Agreement or in an agreement provided by the customer to guarantee that you text will not be uploaded or processed by a machine.
This will not result in additional costs for you.