Exploiting Biased Models to De-bias Text: A Gender-Fair Rewriting Model

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4486–4506 July 9-14, 2023 ©2023 Association for Computational Linguistics Exploiting Biased Models to De-bias Text: A Gender-Fair Rewriting Model Chantal Amrhein1∗ Florian Schottmann2,3 Rico Sennrich1,4 Samuel Läubli1,2 1University of Zurich, 2Textshuttle, 3ETH Zurich, 4University of Edinburgh {amrhein,sennrich}@cl.uzh.ch, {schottmann,laeubli}@textshuttle.ai Abstract Natural language generation models reproduce and often amplify the biases present in their training data. Previous research explored us- ing sequence-to-sequence rewriting models to transform biased model outputs (or original texts) into more gender-fair language by cre- ating pseudo training data through linguistic rules. However, this approach is not practi- cal for languages with more complex morphol- ogy than English. We hypothesise that creat- ing training data in the reverse direction, i.e. starting from gender-fair text, is easier for mor- phologically complex languages and show that it matches the performance of state-of-the-art rewriting models for English. To eliminate the rule-based nature of data creation, we instead propose using machine translation models to create gender-biased text from real gender-fair text via round-trip translation. Our approach allows us to train a rewriting model for German without the need for elaborate handcrafted rules. The outputs of this model increased gender- fairness as shown in a human evaluation study.1 1 Introduction From facial recognition to job matching and medi- cal diagnosis systems, numerous real-world appli- cations suffer from machine learning models that are discriminative towards minority groups based on characteristics such as race, gender, or sexual orientation (Mehrabi et al., 2021). In natural lan- guage processing (NLP), gender bias is a particu- larly significant issue (Sheng et al., 2021b). While research in psychology, linguistics and education studies demonstrates that inclusive language can in- crease the visibility of women (Horvath et al., 2016; Tibblin et al., 2022) and encourage young individu- als of all genders to pursue stereotypically gendered occupations (Vervecken et al., 2013, 2015) without ∗Work done during an internship at Textshuttle. 1We publicly release our data and code here: https:// github.com/textshuttle/exploiting-bias-to-debias S Any Text T Artificial De-biased Text augment (complex rules, parsing, ...) (a) Forward Augmentation (Vanmassenhove et al., 2021; Sun et al., 2021) Any Text T Real Unbiased Text filter (simple rules) S Artificial Biased Text augment (complex rules, parsing, ...) (b) Backward Augmentation (this work) Any Text T Real Unbiased Text filter (simple rules) S Artificial Biased Text translate Biased MT P Artificial Biased Text (Pivot Language) translate Biased MT (c) Round-trip Augmentation (this work) Figure 1: De-biasing rewriters can be implemented as neural sequence-to-sequence models trained on source (S) to target (T) text examples. Previous work creates artificial T from real S through complex augmentation (a). We propose to use real T and generate artificial S to accommodate morphologically complex languages and avoid target-side noise (b). Furthermore, we show that by leveraging biased off-the-shelf machine translation (MT) models, complex rules can be avoided altogether to generate training data for de-biasing rewriters (c). sacrificing comprehensibility (Friedrich and Heise, 2019), state-of-the-art text generation models still overproduce …

Content truncated to 500 words