I will talk about probabilistic models of language change, and how we used these models to reconstruct proto-languages and to understand the statistical regularities of the language change process itself. I will illustrate the potential of our methods with our results on the so called functional load hypothesis. This conjecture has eluded classical models for decades, but here I will show how we obtained compelling evidence for it by using our probabilistic models.
If time permits, I will also talk about new Sequential Monte Carlo algorithms for the related problems of phylogenetic tree and cognate inference. The technique we used to construct these algorithms also has applications in other NLP inference tasks over combinatorial spaces, for example alignment and parsing.