Mitigating Gender Bias in Machine Translation with Target Gender Annotations

Proceedings of the 5th Conference on Machine Translation (WMT), pages 629–638 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 629 Mitigating Gender Bias in Machine Translation with Target Gender Annotations Art¯urs Stafanoviˇcs*†‡ and Toms Bergmanis*†‡ and M¯arcis Pinnis†‡ †Tilde / Vien¯ıbas gatve 75A, Riga, Latvia ‡Faculty of Computing, University of Latvia / Rain¸a bulv. 19, Riga, Latvia {firstname.lastname}@tilde.lv Abstract When translating “The secretary asked for de- tails.” to a language with grammatical gen- der, it might be necessary to determine the gender of the subject “secretary”. If the sen- tence does not contain the necessary infor- mation, it is not always possible to disam- biguate. In such cases, machine translation systems select the most common translation option, which often corresponds to the stereo- typical translations, thus potentially exacerbat- ing prejudice and marginalisation of certain groups and people. We argue that the infor- mation necessary for an adequate translation can not always be deduced from the sentence being translated or even might depend on ex- ternal knowledge. Therefore, in this work, we propose to decouple the task of acquiring the necessary information from the task of learn- ing to translate correctly when such informa- tion is available. To that end, we present a method for training machine translation sys- tems to use word-level annotations containing information about subject’s gender. To prepare training data, we annotate regular source lan- guage words with grammatical gender infor- mation of the corresponding target language words. Using such data to train machine trans- lation systems reduces their reliance on gender stereotypes when information about the sub- ject’s gender is available. Our experiments on five language pairs show that this allows im- proving accuracy on the WinoMT test set by up to 25.8 percentage points. 1 Introduction Most modern natural language processing (NLP) systems learn from natural language data. Findings of social sciences and corpus linguistics, however, indicate various forms of bias in the way humans *First authors with equal contribution. use language (Coates, 1987; Butler, 1990; Fuertes- Olivera, 2007; Rickford, 2016). Thus the result- ing NLP resources and systems also suffer from the same socially constructed biases, as well as inaccuracies and incompleteness (Jørgensen et al., 2015; Hovy and Søgaard, 2015; Prates et al., 2019; Vanmassenhove et al., 2019; Bordia and Bowman, 2019; Davidson et al., 2019; Tan and Celis, 2019). Due to the prevalent use of NLP systems, their sus- ceptibility to social biases becomes an increasingly significant concern as NLP systems not only reflect the biases learned but also amplify and perpetuate them further (Hovy and Spruit, 2016; Crawford, 2017; HLEG, 2019). This work concerns mitigating the manifes- tations of gender bias in the outputs of neural machine translation (NMT) systems in scenarios where the source language does not encode the information about gender that is required in the target language. An example is the translation of the English sentence “The secretary asked for de- tails.” into Latvian. In English, the gender of “sec- retary” is ambiguous. In Latvian, however, there is a …

Content truncated to 500 words