Speaker
Chundra Cathcart
Description
Bayesian mixed-membership models are popular in linguistics, as they explicitly model contact between languages (Reesink et al 2009, Syrjänen et al 2016). Most linguistic applications use the biological Structure program (Pritchard et al 2000) with default presets, fixing the concentration parameter of the population-level Dirichlet prior over allele frequency (treated as an analog for the language-level prior over features) at 1. We show, using a crosslinguistic typological database, that there are linguistically meaningful consequences for the choice of this hyperparameter (either fixed at different values, or inferred from the data) using a series of posterior predictive checks designed for mixed-membership models (Mimno et al 2015).
Primary authors
Chundra Cathcart
Gerd Carling