Seminar talk via Zoom.
Abstract: Initial consonant mutations are a well-known feature of the Celtic languages. I will present two models for accurate prediction of initial mutations in context, focusing primarily on Irish and Scottish Gaelic. The first is a “resource-light” model, obtained by training a suitable neural network using only plain text corpora. The second involves deeper linguistic analysis of the input text, building on existing parsing models in the Universal Dependencies framework. These models allow us to give data-driven estimates of the information content of initial consonant mutations, and have also proved useful in end-user applications such as spelling and grammar checkers.