Do gender neutral affixes naturally reduce gender bias in static word embeddings?

Do gender neutral affixes naturally reduce gender bias in static word embeddings? Jonas Wagner and Sina Zarrieß Bielefeld University Faculty for Linguistics and Literary Studies {jonas.wagner,sina.zarriess}@uni-bielefeld.de Abstract In German, substituting gendered role nouns with gender neutral versions, known as gen- dergerechte Sprache, has rapidly been gaining ground, with the primary aim being the inclu- sion of non-male people. Its effectiveness, how- ever, has not been conclusively demonstrated. Previously, word embeddings have been shown to contain gender biases similar to natural lan- guage. They thus can be used to measure whether this practice impacts gender associ- ation of role nouns. Methods of debiasing pre- trained word embeddings have been devised, but their effectiveness in German, especially compared to gendergerechte Sprache, has not been tested. In this paper, we systematically compare two methods of gender neutral affix- ation to a base corpus to examine the effect on gender bias of role nouns. We also com- pare the gender biases of analogy resolutions generated with embeddings trained on the base corpus, on the base corpus after undergoing an established post-hoc debiasing method, and the corpus after introduction of gender neutral affixation. Our results show a mixed picture: affixation leads to increased gender bias of role nouns, but decreased gender bias of generated analogy resolutions, even outperforming post- hoc debiasing methods. 1 Introduction Gender bias in word embeddings and its reduction have received significant attention from compu- tational linguists and NLP researchers over the past years, and a substantial body of research around the topic has accumulated (Bolukbasi et al. 2016; Caliskan et al. 2017; Ethayarajh et al. 2019; Kaneko and Bollegala 2019 among others). Given the wide use of word embeddings and the result- ing danger of perpetuating and reinforcing gender stereotypes (Hansen et al., 2015; Musto et al., 2015; Dastin, 2018; Schnitzer et al., 2019), this is a press- ing concern. But existing research has failed to address two aspects of the issue: firstly, as is a common problem in NLP, it mostly investigates English (but see Sahlgren and Olsson 2019 and Katsarou et al. 2022 for investigations of Swedish, Ch´avez Mulsa and Spanakis 2020 for Dutch, and Basta et al. 2020 for Spanish), which, in contrast to German, does not have regular gender marking on nouns. Secondly, and possibly as a result of this, it ig- nores societal efforts to mitigate gender bias in nat- ural language. Blodgett et al. (2020, p. 5458) criti- cise this detachment from such societal processes, instead calling for researchers to “[e]xamine lan- guage use in practice by engaging with the lived ex- periences of members of communities affected by NLP systems”. One way in which language users are addressing gender biases in their languages is by changing these gender markings, such as the -e suffix in Spanish or the addition of the female suffix -in to German role nouns,1 which is the subject of the present study. Instead, research has focused on post-hoc debiasing of pre-trained word embeddings rather than the impact of these societal …

Content truncated to 500 words