Log-odds ratio with a Dirichlet prior (Monroe et al. 2008), using plain word-count dictionaries. No scikit-learn, no dependencies.