In this short post, we’ll discuss what gender bias in machine translation looks like, its causes, and what we do at Cadenza Academic Translations to address it.
Firstly, where do we see gender bias in machine translation at Cadenza?
As well as editing full academic articles and other publications, we also edit academic metadata (the title, abstract, and keywords of an article). While some of it is translated by a human before it reaches us, much of the metadata we receive for editing has been through machine translation.
As quick and convenient as it may be, machine translation has its limitations and blind spots, such as the reinforcement of gender bias. In practice, translated text might default to the masculine whenever a gender is not specified, or a masculine pronoun might appear for a job title such as “doctor,” whereas a female pronoun might appear for a job title such as “nurse.” You can find endless examples like these online!
Why does gender bias come up in machine translation?
Well, the truth is that machine translation is simply reflecting back the biases that are already present in society. This is because machine translation models are trained on huge volumes of pre-existing text which unfortunately contain a high proportion of gender-biased language.
As a result, these models quickly learn that “doctor,” for example, is most frequently linked with the pronoun “he,” a lesson which is then repeated in every situation where there is any contextual ambiguity over the intended grammatical gender. In this way, machine translation can end up reproducing and reinforcing the gender bias we see in society.
What can we do about this issue?
At Cadenza, we’re constantly on the lookout for gender bias and aim to nip it in the bud whenever we see it. For example, when we’re evaluating the linguistic quality of an article, “gender-biased language” features on our list of serious errors to flag to journals.
What’s more, our editors have plenty of techniques to avoid gender bias in a text, from simply using a gender-neutral pronoun or term instead, to rewording the whole sentence if needed.
You can therefore be confident that no article or abstract leaves our hands without first being checked for gender bias. We hope this practice of ours will make its own small difference to reducing the presence of gender-biased language in our society.
Below are a couple of articles that inspired this blog post. Feel free to give them a read to find out more about this topic!
https://www.rws.com/blog/gender-bias-machine-translation-language-weaver/
