Aidan is a senior majoring in math and public policy. Outside of the major, Aidan spent his (pre-pandemic) time working at the Women’s Center, on the Model UN team, and as a facilitator for the Common Ground retreat. Aidan enjoys spending time with his friends and family, who have helped him immensely in getting across the finish line with his honors thesis. He does not yet know what he will be doing after graduation, but he is certainly excited to have a break!

Honors Thesis:

Intersectional identities and machine learning: illuminating language biases in twitter algorithms

Faculty Advisor: Professor Jay A. Pearson

Abstract: “Intersectionality,” described first by Kimberlé Crenshaw, has expanded in the past 20 years from a legal framework to a broad theoretical framework to describe people’s lived experiences. The research field was largely defined by Dr. Leslie McCall in 2005, when the categories of quantitative and qualitative intersectionality research emerged. Today, social media data is ripe for identity and intersectionality analysis with wide accessibility across the globe, easy to parse text data, and wide use by a diversity of folks. We specifically consider Twitter data (N = 99,534 tweets) that was annotated by crowdsourcing for tags of “abusive,” “hateful,” or “spam” language (Founta et al., 2018). We run natural language prediction models to predict the tweet’s author’s race (Blodgett et al., 2016) and gender (Sap et al., 2014). We investigate whether these tags for abuse, hate, and spam have a meaningful relationship with the gendered and racialized language predictions we’re making. Specifically, we ask if certain gender and race groups are more likely to be predicted if a tweet is labeled as abusive, hateful, or spam. We find interesting effects with race and intersectional race gender interaction using chi-squared tests, logistic regressions, and ANOVAs using each of our racial groups as the baseline for a different regression. Most notably, we report ordered probabilities for each racialized language model being associated with each of our labels with statistical significance. Finally, we make broad policy recommendations based on our findings and discuss the implications of an algorithm’s creators and trainers.