Machine Translation
‘Women are beautiful, men are leaders’, this is the introductory title of a new study on the presence of gender biases within automatic translation engines. A lingering issue, far from being resolved even today.
These gender biases occur when automatic translation or writing systems embed gender-related prejudices within their textual outputs, reflecting biases typical of certain social contexts.
Namely, when, for instance, the feminine gender is associated with terms related to emotions, fragility, family, and beauty. Simultaneously, these biases occur when the masculine gender is attributed characteristics associated with professionalism, leadership, strength, and rationality.
Related: Schwa (Ə) and Inclusive Language
As reported by Slator, this new study conducted by researchers from the Kempelen Institute of Intelligent Technologies (accessible through the link provided in the conclusion) aims to identify fundamental patterns leading to the propagation of these biases, attributing over 3,500 examples to 16 core prejudice models.
How Gender Biases Describe Women and Men
To conduct the analysis in question, researchers collaborated with a team of gender bias experts, embedding instances of these biases in short sentences closely resembling those that might arise in everyday language.
The list, comprising a total of 3,565 examples, was then condensed into a few universal macro-stereotypes, considered as the foundational basis for prejudiced reasoning.
In detail, these gender biases portray women as ’emotional and irrational’ in 254 examples, ‘gentle, kind and submissive’ in 215, ’empathetic and caring’ in 256, ‘neat and diligent’ in 207, ‘social’ in 200, ‘weak’ in 197, and ‘beautiful’ in 243.
Conversely, men are depicted as ‘tough and rough’ in 251 examples, ‘self-confident’ in 229, ‘professional’ in 215, ‘rational’ in 231, ‘leaders’ in 222, ‘childish’ in 194, and ‘strong’ in 221.
These unacceptable stereotypes are capable of portraying a distorted and clearly misleading reality. In an attempt to definitively address this issue, researchers have introduced the GEST dataset, a tool for measuring gender bias within automatic translation systems.
In conclusion, to quote the researchers: “If we are to understand the stereotypes in these models, we need to have them properly defined. […] Our results show a pretty bleak picture of the state of the field today. Different types of models have seemingly very similar patterns of behavior, indicating that they all might have learned from very similar poisoned sources. At the same time, as we now have a more fine-grained view of their behavior, we can try and focus on specific issues, e.g, how to stop models from sexualizing women. This might be more manageable compared to when gender bias is taken as one vast and nebulous problem.”
For those interested in delving deeper, we recommend reading the entire paper:
