Abstract
Legal language is challenging to translate (Killman 2023, Quinci & Pontrandolfo 2023), both for humans and machines. This is due to some features of legal language (Gualdo & Telve 2021, Mattila 2018), among others, syntactic complexity and terminological density. Chromá (2008) estimated that 20% to 29% of legal texts consist of legal terminology. The latter is one of the main challenges in legal translation (Killman 2023, Prieto Ramos & Cerrutti 2021), as it must be interpreted in the context of a specific legal system, text type and moment in time for a correct translation.
The consequences of legal translation mistakes may be serious (Mattila 2018) and include financial loss, legal disputes, infringement of basic human rights (e.g. bad interpreting in court cases). In minority language situations, high-quality legal translation is essential to ensure that the minority has equal access to information and is not discriminated against compared to the majority.
We work with a local-only minority language (Edwards 2012) that is a lower-resourced standard variety of a big European language (German). In particular, we have analysed translation mistakes in a corpus of 26 decrees (36,000 tokens) (De Camillis et al. 2023) that had been machine-translated from Italian into South Tyrolean German using the neural machine translation (NMT) system ModernMT. Several of the 368 terminology mistakes identified concern designations of bodies and institutions that were not translated with their official German counterpart. Other mistakes concern the use of designations that would only be correct within other legal systems. For example, South Tyrolean legal texts are subdivided into Artikel (articles) rather than Paragraphen (paragraphs) as in Germany. Translating articolo with Paragraph with reference to a local decree is a severe issue. Some mistakes are translations that disrespect the officially standardised terminology (e.g. deliberazione, ‘resolution’, should be translated with Beschluss, not Entschließung). A small amount are translations that use obsolete designations. Finally, wrongly translated or untranslated abbreviations are a known weak point of NMT (Sánchez-Gijón & Kenny 2022).
The contribution will sketch the results of our annotations focused on terminology in local South Tyrolean decrees and reflect on the potential consequences of using NMT systems to translate into lower-resourced varieties of big international languages.