Information-theoretic Methods for Analysis and Inference in Etymology

We introduce a family of minimum description length models which explicitly utilizes phonetic features and captures long-range contextual rules that condition recurrent correspondences of sounds within a language family. We also provide an algorithm to learn a model from this family given a corpus of cognates, sets of genetically related words. Finally, we present an imputation procedure which allows us compare the quality of alignment models, as well as the goodness of the data sets. Our evaluations demonstrate that the new model yields improvements in performance, as compared to those previously reported in the literature.