Supporting Gender-Neutral Writing in German

Das Departement Biologie (D-BIOL)

“…Biased models are useful for de-biasing. At least in the subfield of gender-fair rewriting, debiasing research has focussed extensively on human annotation (Qian et al, 2022) and rule-based processing and training data creation (e.g., Sun et al, 2021;Jain et al, 2021;Alhafni et al, 2022;Diesner-Mayer and Seidel, 2022). Conversely, our work demonstrates that robust de-biasing rewriters can be implemented by leveraging inherently biased NLP models.…”

“…Both of these works only focus on a binary interpretation of gender. Non-sequence-to-sequence approaches have also been explored (Zmigrod et al, 2019;Diesner-Mayer and Seidel, 2022) but required extensive linguistic tools such as morphological, dependency and co-reference analysis, named entity recognition and word inflexion databases.…”

“…As an alternative to handcrafting rules, we propose to exploit the fact that current MT models generate inherently biased text: we create pseudo source segments via round-trip translation through a pivot language that (mostly) does not mark gender (Figure 1c). We use this method to train a gender-fair rewriter model for German, and benchmark it against a highly engineered fully rule-based baseline (Diesner-Mayer and Seidel, 2022).…”

“…Test Set Diesner-Mayer and Seidel (2022) evaluate their system on a subset of the TIGER Treebank (Brants et al, 2004) with German news articles from the 1990s. Since masculine forms are prevalent in these articles, we create a random set of 1,200 TIGER sentences with gendered forms which we balance for masculine, feminine, singular and plural forms, and a random set of 300 non-gendered sentences.…”

“…Second, while already intricate for English, it is likely even harder to define de-biasing rules for more morphologically complex languages with grammatical gender (Figure 2). An approach proposed by Diesner-Mayer and Seidel (2022), for example, requires morphological, dependency and co-reference analysis, as well as named entity recognition and a word inflexion database.…”