Machine-translating legal language: error analysis on an Italian-German corpus of decrees

The paper analyzes the most frequent error categories in a bidirectional corpus of machine-translated decrees in the language combination Italian-South Tyrolean German. The aim is to assess translation issues when using a fine-tuned machine translation (MT) system to produce legal texts in an Italian province where German is an officially recognized minority language, and the local legal language differs from that used within other German-speaking legal systems. Our fine-tuned MT system struggles with features that are typical for the legal language, e.g., legal phraseology, legal terminology (especially the specific local South Tyrolean terminology), and gender-sensitive language. The latter is a requirement for local legislation. The errors identified shed light on the need to feed MT systems with terminological information, especially for low-resource language varieties such as South Tyrolean German. We consider our results key information for the training of post-editors, professional translators, and non-professional translators working in multilingual public administrations.