Abstract
Gender is widely discussed in the context of language tasks and whenexamining the stereotypes propagated by language models. However, currentdiscussions primarily treat gender as binary, which can perpetuate harms suchas the cyclical erasure of non-binary gender identities. These harms are drivenby model and dataset biases, which are consequences of the non-recognition andlack of understanding of non-binary genders in society. In this paper, weexplain the complexity of gender and language around it, and survey non-binarypersons to understand harms associated with the treatment of gender as binaryin English language technologies. We also detail how current languagerepresentations (e.g., GloVe, BERT) capture and perpetuate these harms andrelated challenges that need to be acknowledged and addressed forrepresentations to equitably encode gender information.