The problems in the field of translation crop up in the revision stage – which was the case before neural translations and remains so after them – and in the quality of the original translation. This will be discussed primarily in the outline of the two problem zones:

5. Problems with neural translations and their revision

6. Small agencies: “Know Your Translator/Quality”

In the following, we will briefly describe the key aspects of these five points:

1. Highest quality, reasonable price: neural or manual translation, 4-eye principle, small agency processing

The preparation of a translation according to this method involves the following steps:

i. Neural translation + post-editing or manual translation by translator

ii. First revision of translation by translator

iii. Second revision of translation by translator

iv. Proofreading of translation by translator

v. Revision by second translator/revisor

A translation according to this method should be flawless. The use of neural translations is between 20 and 30% faster, based on our experience and depending on the text’s level of difficulty (our prices take this into account). The reason that the price with neural translations is not even cheaper is due to the rounds of revision and proofreading that are still necessary manually: The machines only speed up the first of five steps (see Section 3 for workaround and Section 4 on the problems with neural translations). The use of a neural translation with artificial intelligence is preferable because a professional translator has two options, essentially two minds, for each sentence/segment. This lets the translator choose the best: the machine-produced option or the one in their head. In most German texts, at least one and often two or three edits are necessary for a sentence/segment, even if the translation produced neurally is good. At more than three changes every twenty words or so, it is generally faster to translate manually (based on our experience), but then the “two-mind” approach is lost.

2. High quality, affordable price: neural translation, small agency

In the teens, it became quite popular to offer and prepare translations according to the DIN 15038, DIN 9001 or DIN 17100 standard, which is tantamount to the 4-eye principle or having someone other than the original translator revise/review the translation. The problem here was and is that a person reading the translation once often misses minor mistakes made by the original translator, mistypes their own corrections here and there and repeats the misinterpretations of the original translator. The mistake often – nearly universally made – is that native speakers of the target text do the revision as well (we use native speakers of the source language) and most of what takes place is simply a rewriting of correctly or incorrectly translated sentences, but a failure to catch objective mistakes or misinterpretations (in 19 years, the failure rate is 100% in our experience).

This dynamic has changed somewhat with the neural translations. Now the translator acts as a decision-maker choosing between the pretranslated version and the version produced in their mind. This not only reduces the amount of brain capacity allocated to interpretations and formulations in the target language. It also means that the first review is taking place during the translation phase – especially since machine translations are now better than most mediocre translators (the ones hired by large translation agencies). So the 4-eye principle is not as necessary as it used to be.

Here, the cost is reduced by the elimination of the external revision.

3. Client-based quality, inexpensive: client handles and revises neural translation, small agency proofreads

With this approach, the translation agency solely handles the proofreading part of the translation process. This is by far the most affordable solution for clients, and it is welcomed by translation agencies. As mentioned and described in more detail below, this translation method absolves the translation agency from responsibility for the time-consuming revision parts, i.e. ensuring correspondence, correct terminology, etc. This approach is particularly well suited for internal documentation, translations by a client’s near-native speakers, less formal translations where the English language text should be correct, and transpositions (rough translations).

The only cost that is important to understand here is the one for proofreading (reviewing of the pure target language text). Similar to revision, one round of proofreading is insufficient if the text has not been handled multiple times prior to proofreading. For example, in a regular translation, the text has been translated, revised twice and then is proofread once. In that context, one round of proofreading is enough. In the case of a client-provided translation, the first proofreading is also the first encounter with the text. Consequently, the text must be read twice and priced accordingly.

4. Postediting or Revision

Translation agencies do not fundamentally want to provide low-cost revised translation of neurally produced texts because of the myriad problems described in Section 3 and 5. It is nearly impossible to define a sufficient amount of post-editing under circumstances where perfection is not the objective. A client and small agency have the best chance of agreeing on the parameters because a small agency will have the same lead editor processing all the work.

5. Problems with neural translations and their revision

As most clients must know from experience and we have observed repeatedly from working in the field for 19 years, the detection of minor mistakes, primarily misreadings of the original and consequently erroneous translations, is the biggest problem in the field of translation. The primary cause of this, as mentioned above, is the practice of having one target language revisor review line-by-line the translation of another target language translator, and furthermore do this in one round. Even if the translation is nearly perfect, the revisor effectively never discovers all the errors with this approach (we have never seen this with a reasonable length translation).

This problem continues to exist with neural translations. It is for this reason that large translation agencies have not substantially reduced the price of translations, and often ignore this new development altogether. It is this problem that accounts for our remarks in Section 3. When the client requests a neural translation and expects to receive a target language text corresponding perfectly to the original at the proofreading cost because the machines do all the work, they ignore the time-consuming revision process and the translation agency’s need to defend the quality of its work by pointing to the lack of objective mistakes. The most insidious aspect of the machines is that basic and complicated sentences can be translated very well, even lull an untrained revisor/translator into assuming that everything is correct. But not only consistency and terminology can be wrong, but neural machines often make minor, but critical interpretative mistakes, not to mention their general lack of linguistic feel that is, at least for translations into English, critical for readability.

6. Small vs. large agencies

If you have a project that requires texts to be translated into 6 languages in a short span of time, the only option is a large translation agency. If you need 30,000 words translated within a week, you must contact a major language service provider. A small office like Intertexting GmbH cannot handle such projects. But most organizations do not require translations under these conditions and cannot afford the extravagant costs and mediocre quality they receive. This is where small agencies like Intertexting GmbH are ideally positioned to handle your language service needs: You get much more affordable prices (due to lower overhead) and consistently outstanding quality from translators you know and work with regularly. As a result of automation and the corona crisis, large agencies have been forced to drastically reduce the price paid for translation. While machines ensure a base level of quality today, these price cuts have drastically increased the likelihood that minor interpretative mistakes by machines will go undetected, the consistency of translated terms will be lacking, and the rough stylistic result of neural translation will not be smoothed out or revised (e.g. too many sentences starting with prepositional phrases in the case of German-English translations). Either low-paid, inexperienced translators without a firm grasp of the source language now prepare the translations for large agencies or experienced translators have accepted a price cut, but also reduced the quality they deliver.

Niche language service providers focused on a limited number of services are the best option in the short and long term.

Version: December 2021