Intersectional Identities and Machine Learning: Illuminating Language Biases in Twitter Algorithms

Intersectional analysis of social media data is rare. Social media data is ripe for identity and intersectionality analysis with wide accessibility and easy to parse text data yet provides a host of its own methodological challenges regarding the identification of identities. We aggregate Twitter data that was annotated by crowdsourcing for tags of “abusive,” “hateful,” or “spam” language. Using natural language prediction models, we predict the tweeter’s race and gender and investigate whether these tags for abuse, hate, and spam have a meaningful relationship with the gendered and racialized language predictions. Are certain gender and race groups more likely to be predicted if a tweet is labeled as abusive, hateful, or spam? The findings suggest that certain racial and intersectional groups are more likely to be associated with non-normal language identification. Language consistent with white identity is most likely to be considered within the norm and non-white racial groups are more often linked to hateful, abusive, or spam language.