Gender Bias: the Challenge of ‘He’ vs. ‘She’ for MT

John Tinsley has worked with machine translation over the last 16 years in both academia and industry. In 2009, he received a PhD in Computing from Dublin City University, after which he co-founded Iconic Translation Machines. He now leads Iconic as a dedicated technology division of RWS Group.


The worst forms of bias are those that are explicit conscious  forms of prejudice against some individual or group based on some characteristic of their being, such as race, gender, sexual orientation, or religious beliefs. Arguably though, the most insidious forms of bias are those unconscious biases — the ones we might not even realize we have or exhibit.

One common example is confirmation bias, which is the inclination to draw a conclusion about something or some-one based on your own personal desires, beliefs, and prejudices rather than on unbiased assessment. Similarly, gender bias and ageism are the tendencies to make decisions or have preferences on the basis of gender and age respectively. 

These unconscious biases can often be downplayed as “gut instincts” as opposed to a true bias. However, when it comes to artificial intelligence (AI), machines don’t have gut instincts, at least not yet — that’s when they’ll truly take over. Instead, they inherit their biases from elsewhere, namely the data used to train AI models.

 Bias in AI 

Machine learning systems can learn biases based on assumptions that are built into the algorithms used to train them, but the cause is mainly rooted in the data. Examples of biased AI are fairly common these days. They range from the relatively harmless — such as online ads being served to someone based on generic demographic information as opposed to their actual interests — to the more sinister, such as recruitment algorithms suggesting jobs with leanings based on gender, or automated healthcare risk assessments with judgments based on race.

One bias that has become more prevalent, or at least more visible, in the last few years in our industry, particularly when it comes to machine translation (MT), is gender bias.

 Gender bias and MT 

Different languages deal with gender in different ways. For instance, some languages have gendered pronouns (e.g., he/she/him/her in English) whereas others, such as Finnish, Hungarian, and Turkish, have gender-neutral pronouns. In the example below from Finnish, you can see this in action. As you can probably imagine, the challenge comes when we need to translate one of the Finnish sentences on the right — which pronoun in English do we choose?

A human translator might be able to tell what the correct translation should be based on the context in the document. However, we don’t always have that luxury, for instance when we’re translating a single sentence. Additionally, on the MT side, even when we have context at the document level, the technology doesn’t yet have the capacity to use it in this way.

The problem of gender bias then manifests itself in that the MT engine has to pick one pronoun or the other, and the driver behind the choice is the noun. In the example in figure 1, “doctor” will always be associated with the masculine pronoun “he” (he is a doctor), while “nurse” will always be associated with the feminine pronoun “she” (she is a nurse). 

This is gender bias. It is a topic that has reared its head (again) recently and you can find many more examples with a quick online search. 

Editor’s note: Google Translate, for its part, has been working on improving its algorithm to address these concerns. As of late July 2021, the above sentence in Finnish will render in English with options for masculine and feminine. It’s worth noting, however, that in translating Finnish into other languages on Google Translate the results will vary; French and Portuguese will render as masculine, Greek will render as feminine.

 Cause and effect 

We know why this happens. MT, as with other forms of AI, is trained on large datasets. In the case of MT, these are datasets (or corpora) of previously translated documents. While big data is certainly big, it is not diverse. Generally speaking, data is over-represented with male data. For example, one research study found that the masculine form of engineer in German (der Ingenieur) was 75 times more common than its feminine counter-part  (die Ingenieurin). If an MT engine is trained on this data, it will inevitably replicate this bias. MT is essentially inheriting the biases that humankind has baked into the data.

 So what can be done to avoid or mitigate this? 

Without the ability to look at wider context in the text, it’s not possible to determine what the correct choice of gender could be. To complicate things further, in some cases the source language is gender-neutral by design. However, even though the text is fundamentally ambiguous, the translator — in this case human or machine — must still produce some translation in a language that doesn’t have a gender-neutral option. How?

From the MT perspective, there are a number of ways to address this challenge, but there is no clear and universally accepted solution. Nevertheless, let’s take a look at some of them.

 How do we fix it? 

Broadly speaking, there are two main approaches to addressing this issue. The first involves dealing with bias in the data before training any models. This essentially means removing, or debiasing the data. Some of the strategies for doing this in an automated way include processes such as downsampling, upsampling, and counterfactual augmentation. However, it is generally a very challenging task and experiments have shown it has a nasty side effect of significantly impacting the overall quality of the translations, without being super effective at removing the bias anyway.

Another critique of this approach is that the data is real. It’s a reflection of an existing state and if we modify it, there’s an ethical question as to whether we are distorting reality to achieve some desired result. Where would we draw the line with this? It’s a much broader question that is probably out of scope for this column.

The second, and more popular, approach is to address the situation after the models have been trained. The goal here is to reduce bias as much as practicable, as opposed to eradicating it, and do so without disproportionately impacting the quality of the translations themselves. 

Again, there are two strategies for this. One is to look at the task of debiasing, like adapting an MT model for a particular domain. This means training the model on the original large dataset, but then finetuning the model with a specialized dataset that was created to be explicitly gender-neutral. 

The other strategy is to deal with it in real-time when translating. This involves producing the translation and then  detecting whether it is gendered or not. Once this information is available, what to do with it begs another question. The answer will depend on the user interface, the use case for the translation, and other preferences for the provider and/or user.

Some of the options include:

  1. Flagging that the translation is gendered.
  2. Presenting the user with alternative translations.
  3. Modifying the translation to make it gender-neutral. 

The first option is the easiest to implement but has relatively little impact or value — most translations will be legitimately gendered.

The second option carries significant overhead to actually produce alternative gendered translations related to the original. There is then an open question of how to best display this to a user. It is easy to see how it could look in an online translation tool, but less obvious how this might work for document translation or some enterprise workflow. 

The final option is only relevant for languages that have adequate options for gender neutrality, and the question would still remain as to whether it’s appropriate to do this in all cases. 

As you can see, it’s a nuanced topic — the positive news is that it’s on the research agenda both in academia and among commercial providers of MT. There is a quite significant volume of work in this area, not only for gender bias in MT, but also bias more generally in AI systems. It’s certainly one to watch for the coming year.