How machine learning can contribute to gender-equal technology

Subtitles for the hearing-impaired, automatic translations, and ChatGPT, great contributions have been made to all by researchers in language technology, researchers who study how computers can interact with human languages. While these smart tools facilitate our everyday lives, there are also pitfalls and drawbacks, such as unequal access to the results of language technology tools, and these tools reproducing and perpetuating harmful biases they learn from us.

“We have seen with the introduction of tools like ChatGPT and similar large generative models, how language technology has the capacity to influence the ways we interact with language and technology. Language technology has the capacity to provide us with useful tools that would make many of our tasks easier. These tools, however, also have the potential to learn and reproduce harmful biases, potentially increasing already existing inequalities”, says György Kovács, Senior Lecturer at Luleå University of Technology.

With the boom of AI tools – especially ChatGPT – many users quickly started pushing boundaries: What can you get the tools to say? Quite a bit of harmful things, it turned out. One reason for this is how the information is gathered, namely large amounts of user-generated data.

“While modern tools like Large Language Models – ChatGPT and Gemini being two famous examples – can pick up a lot of beneficial information from large amounts of text that they can use to help us, they can also learn and reproduce harmful biases hidden in these texts.”

These biases can come from users, developers, and data sources from which the tools collect information. During the spring of 2024, György Kovács was part of a team organizing a workshop on the use of Language Technology for Equality, Diversity, and Inclusion, to bring together interested researchers and practitioners. The topic of accessibility was also part of the questions tackled in the workshop:

“Research in language technology is largely focused on the English language, which means that many languages with fewer speakers or fewer resources are at risk of falling behind, not being able to enjoy the same benefits. Another important aspect of accessibility is access to information regardless of mental ability or reading capacity.”

Machine Learning researchers contribute to a wide variety of efforts in the field of language technology. This includes exploring how existing solutions could be used in various tasks. Additionally, they examine language technology solutions on a fundamental level to understand how they could be improved. Detecting hate speech and offensive language is one research area that, deservedly, received much attention. Another current issue is the use of language technology in education to both alleviate the workload of teachers and assist students – and how language technology can more easily be applied to specialised languages is a question that ties all together.