Although many people could only laugh or shake their head at the poor quality of translations produced by free online translation tools for many years, a new company has been shaking things up on the market since late 2017. DeepL offers free translation into and out of nine European languages with its neural machine translation tool (you can find a good explanation of how neural machine translation works here.
Here at Textworks Translations, our highest priority is producing top-quality translations of your academic texts. In doing so, it is important to us that we use all the tools available on the market to produce our translations for you in the most efficient way possible. After testing out DeepL, we have come to the conclusion that it can represent a sort of “compromise” between a full translation and proofreading, and we would therefore like to start offering the service. First, however, we feel it is important to discuss the advantages of using machine translation tools like DeepL but also the potential pitfalls so you can make an informed choice about which of our services best suits you.
Academic translators have been reporting an uptick in requests to “proofread” English texts that turn out to be translations produced by DeepL. In essence, there is nothing wrong with having a human read and edit a translation produced by machine. The task is, however, much different than when a human reads and edits a translation produced by another human. In fact, it is so different that in specialist language it is not even called “proofreading” but instead referred to as “post-editing.” We would therefore like to discuss here what these differences are and how you can best make use of machine translation compared to the alternatives.
Important information about DeepL
One very important thing to note about the free version of DeepL is that any texts entered into DeepL and any corrections you make to the translations using the DeepL interface are saved by the company to train their algorithms. Especially when dealing with unpublished research data, a great deal of care should be used here. In the Pro version of DeepL (currently €20 a month for an unlimited amount of translation), according to the data protection guidelines, the company only saves the data for as long as is necessary to run it through the machine to produce the translation and then deletes it. Here authors need to decide whether or not they trust the company to do as they say and to ensure that the server is safe from hackers who could still have access to the texts for that short period of time. In addition, it is very important to note that in the data protection guidelines, DeepL states that even in the Pro version you should not send any personal data with the texts. This means that you may need to anonymize texts before you enter them into the DeepL interface.
Advantages and disadvantages to using DeepL for social science texts
Professional, high-quality translations of entire articles will unarguably give you the highest quality and save you the most time, but they can be quite expensive. If you have a tight budget in your project or department, you may not be able to afford it. Translating an article yourself and then having it sent to be proofread is typically much less expensive than having it professionally translated, but this can be very time-consuming and, depending on your level of fluency, quite frustrating.
Machine translation followed by post editing done by a qualified, experienced translator can be a good compromise. In terms of the budget, it is significantly less expensive than a full translation, although it is significantly more expensive than proofreading a human translation. When it comes to the time you as an author need to invest, post editing involves more work for you than simply sending an article to be fully translated by a professional translator, but less work than if you attempt to translate the entire article by yourself before sending it out for professional proofreading.
Is DeepL any good for academic translations?
DeepL produces incredibly fluent-sounding translations, and in our experience the quality is quite good. However, a translation produced by DeepL should never be taken at its face value and submitted directly to a journal, for example. Professional post-editors who are native speakers in the language into which the article is translated are familiar with the problems of the neural machine translation tools and do not allow themselves to be led astray by fluent-sounding text. Another issue, however, is that DeepL (and all other neural machine translation tools currently on the market) produces errors that even an experienced translator cannot necessarily recognize as an error if they are only looking at the translated text.
- Inconsistent terminology: Using the correct terminology is absolutely essential when translating social science texts. DeepL is a huge improvement over previous statistical machine translation engines because it looks at the entire sentence, meaning that the context is to a certain extent taken into consideration. However, it does not look beyond sentence boundaries, which means that a term may be translated one way in the first sentence but a different way in the next. This proves to be a serious problem for proofreaders who are only given the English text, because we assume that you as an expert in your field will always use the same English term when talking about a particular German term, and if you use a different English term, it is because you are referring to a different concept or term in German.
- Missing parts: When the neural algorithm cannot find a good match for a particular phrase or word, it sometimes drops it entirely. Sometimes this is noticeable because a sentence ends abruptly or something essential is missing. Often, however, it is a word or phrase that a person reading the text for the first time does not even notice is missing. And this can be especially problematic if it is a very important word such as “not”. Translators have reported DeepL leaving out “not” in translations. You only need to imagine the sentence “the results were not statistically significant” to understand what it could mean in practice if it were left out and the proofreader was concentrating solely on the language, overlooking the fact that in the tables the findings were *not* significant.
These two types of errors are typically not ones made by people. In addition, they are extremely difficult to see and correct even for professional proofreaders if they are only looking at the English text.
How to use DeepL
If you are interested in trying out our post editing services, the following workflow will lead to the smoothest process:
- When your original article is complete, translate it using DeepL or another neural machine translation tool (both Google and Microsoft offer tools for a fee). In all cases, please read the companies’ data protection policies first and decide whether and which policies fulfill any confidentiality requirements you have regarding your data. If you have any doubts, please consider sending the article to us for a standard human translation instead of using online translation tools.
- Once you have used DeepL, carefully read the translation to ensure that specialist terminology has been translated correctly each and every time. Just because DeepL translates it correctly in the first sentence does not mean that it will translate it correctly each time.
- Carefully read the translation again to make sure that no phrases or keywords have been deleted in the translation process.
Send us both the original text and the translation so that our translators and post editors can carefully compare the texts to make sure that DeepL’s translation doesn’t just sound good but is also a correct translation of your thoughts and research.
If you found this article useful, here’s how you can share it with your colleagues: