Google has had long-standing issues with gender parity, from admitting it hires too many white guys to debates over whether the women that is does hire are underpaid compared to men. With the issue of gender discrimination receiving a significant amount of public attention this year, Google has been making efforts to be more gender-inclusive through efforts like the Change the Game initiative to encourage young women to get into game design.
Now Google has taken another step towards reducing gender bias by changing the way that Google Translate offers translations. As Translate learns from content that is already on the web, it tends to reproduce gender-based assumptions in language. This is a problem in many languages, like Spanish, Portuguese, French, and Italian, which have nouns which can be in either masculine or feminine. When Google translates into one of these languages, it tends to carry assumptions about gender based on stereotypes.
For example, previously when you translated a word like “strong” or “doctor,” Translate would tend towards a masculine interpretation of the word. Conversely, if you translated a word like “nurse” or “beautiful,” it would tend towards a feminine interpretation. Now, Translate will offer you both masculine and feminine versions of the phrase as a translation.
So, formerly if you typed “o bir doktor” into Translate, it would translate from Turkish into “he is a doctor.” Now you are offered two options instead: “she is a doctor” as well as “he is a doctor.” This feature is currently only available in some languages and on the web version of Translate, though Google is planning to roll it out to other languages and to other versions of the software like the iOS and Android apps. Google also mentions plans to address non-binary gender in translations at some point in the future as well.
Equality in terms of gender and race has been particularly tough to achieve in situations which use machine learning, as these systems are trained on existing content, by creators who are often themselves not demographically representative. This has lead to facial recognition software which is less accurate at working with non-Caucasian faces, for example, and automatically filled in search results which were derogatory towards black, Asian, and Latina women. Hopefully this change to how Translate works with gender is a step towards lessening this bias.