Key points are not available for this paper at this time.
Recently there has been a growing concern about machine bias, where trained models grow to reflect controversial societal asymmetries, such as or racial bias. A significant number of AI tools have recently been to be harmfully biased towards some minority, with reports of racist behavior predictors, Iphone X failing to differentiate between two people and Google photos' mistakenly classifying black people as. Although a systematic study of such biases can be difficult, we that automated translation tools can be exploited through gender languages to yield a window into the phenomenon of gender bias in AI. In this paper, we start with a comprehensive list of job positions from the. S. Bureau of Labor Statistics (BLS) and used it to build sentences in like "He/She is an Engineer" in 12 different gender neutral such as Hungarian, Chinese, Yoruba, and several others. We translate sentences into English using the Google Translate API, and collect about the frequency of female, male and gender-neutral pronouns in translated output. We show that GT exhibits a strong tendency towards male, in particular for fields linked to unbalanced gender distribution as STEM jobs. We ran these statistics against BLS' data for the frequency female participation in each job position, showing that GT fails to a real-world distribution of female workers. We provide experimental that even if one does not expect in principle a 50: 50 pronominal distribution, GT yields male defaults much more frequently than what be expected from demographic data alone. We are hopeful that this work will ignite a debate about the need to augment statistical translation tools with debiasing techniques which can be found in the scientific literature.
Prates et al. (Thu,) studied this question.