The Promise Of ‘They’: Reducing Gender Stereotyping in Text-Based AI Models With Gender-Inclusive Language

The language we use shapes our perception of the world. Therefore, adapting language in a way that counters gendered expectations and stereotypes, as well as making non-binary and queer gender identities visible, has the potential to promote gender equality in society. My research takes this idea and applies it to Artificial Intelligence-based language technology. Pre-trained textual AI models, which are used to represent text in a machine-readable way, have been shown to absorb gender stereotypes contained in the large volume of documents they are trained on. This is problematic, because pre-trained models serve as the foundation for a variety of language technology such as automatic translators or resume filtering systems, which can then replicate gender discrimination and stereotyping at application level. Our research uses gender-inclusive language to adapt the training data of pre-trained models by using the gender-neutral pronoun ‘they’ as well as substitutions of terms like ‘flight attendant’ instead of ‘stewardess’ or ‘fire fighter’ instead of ‘fireman’, among others. Experiments showed that training on gender-neutral English in fact reduces stereotyping learned by the model, and that the reduction was more stable than for other, algorithm-based methods. Further, harmful gendered associations within the model, such as the association of men with violence, and women with sexuality, were disrupted after training on gender-inclusive English. This research thus underscores the transformative potential of language adaptation in reshaping the landscape of Artificial Intelligence, offering a promising path towards fostering equality, eradicating stereotypes, and creating a more just society.