Word embedding association tests: A methodological autopsy
Alex Kindel, who recently joined the médialab will present his study of the Word embedding association tests methodology, within its historical, interprofessional, and theoretical context.
Rendez-vous, Séminaire de recherche
"Word embedding association tests (WEATs) are a increasingly popular method in computational social science for measuring cultural aggregates (e.g. concepts, biases, sentiments, schemas, stereotypes). The underlying model family estimates the difference in mean cosine similarity between sets of vectors of word association statistics indexed by keyword lists (i.e. word embeddings). Unfortunately, WEAT models are invalid whenever the keyword lists contain more than three words each, because the resulting similarity metric does not guarantee that every linear subspace spanned by at least three words is closest to itself (which is impossible). In the talk, I will conduct an autopsy of the WEAT methodology; in addition to explaining its mathematical shortcomings, I will situate the technique in its historical, interprofessional, and theoretical context. I will also discuss a well-established multidimensional similarity metric with preferable geometric consistency properties."
Alex Kindel is an Assistant Professor in the médialab and the Department of Sociology at Sciences Po. He received his PhD in Sociology from Princeton University in 2023. Alex's research examines the historical and methodological dimensions of measurement in the social sciences. His current line of work focuses on the measurement of word associations.
The seminar will be held on the Sciences Po campus, located 1 place Saint Thomas d'Aquin (Room K.011), 75007 Paris.