Gender Bias in Machine Translation
There are various factors that may lead to biased MT output. One factor, of course, is the way the source and target languages deal with the concept of “gender” from a linguistic perspective. But we also have to take socio-cultural and technical aspects into account when trying to mitigate gender bias in machine translation.
Linguistic Factors
Gender is encoded in different languages to different degrees. Three language groups have been defined: Genderless languages such as Turkish and Finnish, notional gender languages such as English and Danish that have lexical gender and pronominal genders, and grammatical gender languages such as Spanish and Arabic in which nouns carry gender and other parts of speech can carry gender inflections as well. When translating from a genderless language to a gendered language, the MT engine needs to make a choice. Often, the default choice is male. This is also true for ambiguous terms and phrases.
Nonbinary options are usually not considered yet. As developing nonbinary language is an ongoing social process, there is no consensus yet (and likely never will be). MT will have to adapt to the changes in language over time.
Socio-Cultural Factors
All machine translation engines rely on large corpora of existing translations. This data is naturally subject to the historical and socio-cultural factors that shape language. The data used to train the MT systems are usually not demographically well-balanced, with an strong dominance of male representation and the barely non-existing inclusion of nonbinary people. Also, gender stereotypes stemming from earlier decades will be reinforced and propagated this way.
Technical Factors
Datasets are the crucial factor in creating, training and evaluating MT output. Research shows that assymetries in gender distribution in the training and testing samples do not only cause the machine to learn this biased representation, but this behaviour will also be rewarded in the evaluation, leading to reinforcement. Certain sampling methods used by MT analysts have also shown to have a tendency to favour the generation of male forms. To mitigate gender bias in MT, it is therefore crucial to provide well-balanced, non-biased datasets to train and improve the MT engines.
Gender-Fair Language
The goal of gender-fair language is to reduce gender stereotyping, discrimination, and erasure. There are two common strategies to achieve this: feminization and neutralization. Feminization makes the inclusion of women explicit by presenting terms where the ‘male generic’ would be used as a pair, containing both the masculine and feminine version of the term (e.g., ‘Elektrikerinnen und Elektriker’ instead of just ‘Elektriker’). While feminization can go some way to reducing the imbalance between men and women in society, it has been shown that feminine versions of words are often valued less. Feminization also assumes a binary perspective of gender, which does not reflect societal reality. Neutralization tries to replace gender-marked terms with gender-indefinite terms (e.g., ‘sportsperson’ instead of ‘sportsman’) and as such also acknowledges the existence of genders beyond the binary.
Research on gender-fair language is ongoing, and has only recently begun to include non-binary genders. There does seem to be evidence so far that changing language can, in fact, change society:
- gender neutral forms in job advertisements can cause more women to apply to positions that would otherwise mostly draw male applicants
- changing pronouns in a murder trial lead to a higher number of participants deciding it was a case of self-defense
- in the political context, the use of gender-fair forms when asking participants which politician of different parties should run for the office of chancellor raised the number of female politicians mentioned in response.
- people who encounter gender-inclusive forms use them more themselves and, in turn, have more gender-balanced mental representations of social roles.
