Every time I think I might write about this it seems a drop in the ocean, but then the topic comes up again. Machine translation has become very much better since it was first based on neural networks. In fact, I thought we had departed from rule-based MT and arrived at statistical MT, but we are now on neural MT, excuse my ignorance.
In my experience of using DeepL and DeepL Pro and Google Translate a few times, these systems are very good but not 100% reliable. Which means that sometimes a negative sentence may be rendered as a positive. I have no experience of revising MT output or preparing texts for MT.
But what strikes me specifically about legal texts is that when I put a German text through DeepL, the standard or ‘official’ translations of court names and statute titles are missing, although in Linguee they are present. In the old days, a law firm wanting to use rule-based MT was able to adapt its MT system by filling it with the standard translations into English. Now it is not so, and I would spend a lot of time revising the versions.
This was remarked on in a short article in MDÜ 6/2019 by hans Christian von Steuber. He refers to a talk by Patrick Mustu “Was DeepL & Co. im Zeitalter von 4.0 (noch) nicht können”:
Ein Beispiel aus eminer Erfahrung: Die beliebte “Datenschutzgrundverordnung” wird als “basis data protection regulation” übersetzt, “DSGVO” überhaupt nicht”.
This is the GDPR. It’s always surprising when this is not recognized. And when the translation of a statute title varies within one text.
Another reference was mentioned by colleagues this week. The Swiss Federal Chancellery (Bundeskanzlei) was going to buy licences for 2000 DeepL users in 2019 (UEPO December 2018) and had a review of the effectiveness done (PDF Bericht DeepL-Test). Here an example:
Original text: loi fédérale sur les prestations
de sécurité privées fournies à
l’étranger (LPSP)DeepL translation: Federal Law on Private Security
Services Abroad (LPSP)Post-edited text: Federal Act on Private Security
Services Provided Abroad
(PSSA)
I always wonder when I see English versions of abbreviations of statute names recommended. Leaving it in German is odd and so is creating an English version such as PSSA (obviously the work of a human post-editor).
That’s all I have to say about MT and legal translation, but the Swiss test contains a classification of MT problems.
MT is awful for legal translation, if quality translation is the desired outcome. I suppose it’s ok for gisting, but I would hesitate even there. In my experience, MT software fails to follow basic rules of composition in English and in German, especially when it comes to legal texts. To name but a few issues with MT legal translation:
— subjects and verbs are placed too far apart;
— the organization of thought gets lost, provided there was any in the first place;
— perfectly normal legalese is rendered in run-on sentences;
— commas are lacking; occasional nonsense crops up;
— a lack of uniformity in the translation of standard terms of art;
— close tracking of the source text’s syntax;
— overliteral translations of words;
— etc.
As for abbreviations of statute names: I have a very definite opinion on this issue, with a host of caveats. I wrote about this over at Beck: https://community.beck.de/2018/06/11/gesetzesverweise-in-englischer-uebersetzung