Neural Machine Translation of Legal Texts: Why You need a Human Translation Expert

In 2016, a new form of machine translation upended expectations about what this technology can achieve. This new form is known as neural machine translation and even the most sceptical translator would – perhaps grudgingly – acknowledge that, compared with the previous state of the art, neural produces smoother translations that read more naturally.

This increased readability has resulted in the widespread adoption of neural machine translation. Now the industry standard, used by most translation agencies, it is also widely available, free of charge, through websites such as Google Translate and DeepL.

Is Google Translate Accurate?

You don’t have to be an expert translator to have noticed that Google Translate sometimes produces comically bad translations; indeed, it has become the butt of online jokes over the years. To be fair to Google Translate, its reputation for clangers was partly earned in the pre-neural days, but even the more modern version of its machine translation can’t hold a candle to DeepL.

DeepL vs. Google Translate

Take a look at these two translations of a clause, taken from the terms and conditions of sale from a French website: anyone with legal training – even if they have no knowledge of French – can easily spot some problems with the Google version. The most obvious mistake is that it uses “defender” instead of “defendant”, but another standout is that the final sub-clause (“the seller… buyer’s seat”) is nonsensical.

By contrast, DeepL has, by and large, produced a fairly reasonable translation of the French clause, correcting the obvious errors identified in the previous paragraph, along with some that require a knowledge of French to identify. For example, it has corrected the mistranslations of “visées” (“envisaged” in the Google version, “referred to” in the DeepL) and of “différents” (“different matters” in the Google version, “disputes” in the DeepL). The machine translation has done particularly well in the latter case because “différents” is actually a mistake in the French; its homophone “différends” should have been used instead. Impressive.

So, DeepL can take care of all your translation needs now, right? Wrong.

Is DeepL Accurate?

While DeepL produces extremely fluent texts, I have found, when trying it out, that it can suffer from accuracy problems.

Like most expert translators, I don’t use machine translation for my work. Nevertheless, in the course of teaching postgraduate translation technology courses at the University of Bath, I have had the chance to do quite a bit of study of the quality of machine translation output. (After all, few things make translation students sweat more than the fear that their brand new Master’s degree could be rendered obsolete by the rise of the machines.) While DeepL has style, it lacks substance, as shown in the table below.

Even though it has handled some of the legal terms included in the French original better than Google, there is a mistranslation. In French law, appel en garantie is a process whereby the defence counsel can seek to transfer liability for a debt to a guarantor; DeepL’s nonsensical rendering as “appeal in warranty” does nothing to bring across this meaning to the reader.

DeepL does well where there are functional equivalents (such as “exécution”, which it correctly translates as “performance”). However, it struggles with challenging situations, like terms of art that are not a perfect match between the two languages or – even harder – that have no equivalent at all.

This sort of linguistic problem is very common in legal translation because it’s vanishingly rare for two legal systems to be identical. With this sort of country-specific term, in order to eliminate any potential confusion, readers need to be sure that they know the proper name in the original language; after all, it could come up in court. For ease of reference, however, readers also need some sort of explanation of the concept in terms that they are familiar with and in a language that they know. That is why, in my version, I have included the original French term with a brief explanation in brackets. Only a person with an excellent understanding, not just of the languages involved, but also of their respective legal systems can translate foreign terms of art in this way.

What’s more, the need for excellent understanding of the languages is compounded by the very fact that the machine-translated text is more fluent and contains fewer errors; it lulls readers into a false sense of security. This is one of the main reasons why I don’t use machine translation, even DeepL.

No Style or Substance

A second problem with the DeepL translation is that it does not read as a person with legal training would expect. Translated legal documents have to make sense to lawyers, judges, and other participants in the legal system of the country where the translation is to be used.

Of course, as we saw above, with appel en garantie, parts of a legal translation will inevitably look alien to target-language readers. However, the text is easier for them to understand if it uses familiar wordings, where possible. In this case, since it is an “Applicable Law” clause of the sort found in most contracts, it contains elements also present in English-law contracts. I have therefore drafted those parts in a way more familiar to English-speaking lawyers, while being careful to retain any aspects specific to the French wording. Again, artificial intelligence is not capable of doing this.

Human Translation Vs. Machine Translation

There is no doubt that good-quality machine translation can be put to good use. It does a great job of giving you the gist of what a foreign language text says. What it can’t do is draw on any knowledge of different legal systems to help you bridge the sometimes-yawning gap between them. That’s where an expert legal translator comes in.

Why Work with Me?

As a legal translator with well over a decade’s experience, who is used to meeting some of the most stringent quality requirements in the world, I am perfectly placed to tackle the challenges of bridging that yawning gap. I’ll make sure that nothing whatsoever is lost in translation: where unfamiliar terms from a foreign legal system are used, I’ll make sure you know the actual term being used, while providing an explanation; where there is a functional equivalent, I’ll couch the translation in familiar language.