Theories of "Sexuality" in Natural Language Processing Bias Research

Sexuality is poorly defined in NLP bias research: The paper found that a majority of the 55 articles surveyed did not clearly define what they meant by "sexuality," instead relying on assumed or normative understandings. This lack of clarity undermines the reliability of bias measurements and recommendations.

Gender and sexual identity are being conflated in bias testing: Current methods for detecting sexuality-based bias often mix up gender identity with sexual orientation, creating oversimplified and inaccurate representations of queer identities. This conflation means that actual biases against LGBTQ+ communities may be missed or mismeasured.

Current bias research produces "monolithic conceptions of queerness": By not distinguishing between different sexual identities and practices, NLP bias research treats all queer people as a single category rather than recognizing the diversity within LGBTQ+ communities, leading to incomplete and potentially misleading findings about how AI systems treat different groups.

Queer communities are largely absent from the research process: The paper emphasizes that bias researchers have not adequately engaged with queer communities themselves or drawn on relevant interdisciplinary scholarship, meaning the people most affected by these biases have had little input into how bias is defined and measured.

Improving sexuality bias research requires fundamental methodological changes: To address these gaps, researchers need to more carefully define sexuality, avoid conflating different identities, and actively involve queer communities and interdisciplinary perspectives in designing bias studies.