Lost in translation?

Machine translation and gender-fair German

Abstract

Efforts to increase the visibility of women in the German language date back to the 1970s. These initiatives have been complemented by more recent attempts to enhance the visibility of non-binary individuals. Consequently, a growing number of language-specific strategies have been proposed to either neutralise the language or to linguistically render individuals of all genders visible. The use of such strategies, rather than the still prevalent masculine generic, is often a highly conscious choice. Therefore, when translating such forms of gender-fair language (GFL), an adequate translation should seek to capture the original intentions expressed in German.

However, in the context of machine translation (MT), on which we are increasingly reliant, previous studies have repeatedly identified a male bias. While most prior research has primarily focused on binary forms and often on translations from natural gender to grammatical gender languages, this study aims to close the existing research gap by analysing MT of neutralisation, binary, and non-binary strategies from a grammatical gender language, German, into another grammatical gender language, Italian, as well as a natural gender language, English. Given the growing presence of large language models (LLMs) in everyday life and their translation capabilities, the study further seeks to determine whether differences exist between commercial translation programmes (DeepL, Google Translate, Bing) and LLMs (ChatGPT, Copilot, Gemini) in translating GFL.

The findings indicate that translations into Italian capture the original GFL intentions slightly better than those into English. Further, it was shown that LLMs provide significantly better and more accurate translations of GFL, with ChatGPT performing best for Italian and Copilot for English translations. However, the primary focus of translations remains on neutralisation and binary forms. Even when specifically prompted to produce translations that include non-binary individuals, the models continue to struggle.