In 2023 we took a long look at how we could make our machine translations more inclusive and above all less binary. Even more so now as AI continues to be in the spotlight and considered a new way of working more efficiently. However, even with these advances and increased efficiency we are still left with the same challenges when it comes to inclusion; limited data sets.

The data sets being used continue to exclude large parts of society, whether that be women, racialized people or non-binary folk. In this session we focus on the non-binary exclusion in particular and the ideas we had to overcome this lack of inclusion in AI. We considered different options of how to include non-binary translations and in the end the most viable option was, we’ll do it ourselves!

We then set about creating our own data sets, to facilitate non-binary language training internally. It was a massive task, which needed the collaboration of many teams and detailed planning.

Did we make huge steps towards ending in-house binary machine translations? Yes. Have we still got more work to do? Also yes. In this session I’ll talk you through how we went about training our machines to stop being binary, specifically:

-how we curated data sets

-people involved in the process

-the essential tools we used

-the challenges we faced and how we overcame them

-learning points for future inclusive machine learning

Host organization: Globalization and Localization Association