Modelling autosegmental phonology with multi-tape finite state transducers

Phonology may be briefly defined as the study of sound patterns in spoken language. One of the most well-known computational models of phonology, Koskenniemi's two-level phonology, is based on an underlying linguistic theory that has been superseded by autosegmental phonology, which began with the work of Goldsmith. There is a need for computational models that are faithful to this more recent theory. Such a model can form the basis of a computational tool that can quickly and accurately check the validity of a phonological analysis on a large amount of phonetic data, freeing the linguist from the tedious and enor-prone task of doing this by hand. This-thesis presents a new computational model of phonology that is faithful to standard autosegmental theory, that has clearly adequate expressive power, and that is suitable as the basis for a tool for phonological analysis. It follows on very recent efforts by Kornai and Bird & Ellison to model autosegmental phonology. The model is based on a view of phonology that sees phonological representations as data and phonological rules as procedures that manipulate them. It models rules using multi-tape state-labelled finite transducers (MSFTs), a natural extension of finite state transducers obtained by adding multiple input and output tapes. MSFTs are shown to be powerful enough to express a wide range of autosegmental rules. We also investigate the class of formal languages accepted by multi-tape state-labelled finite automata (MSFAs) when their input tapes are considered to encode a single word in parallel. This class is quite large, including some languages that are not context free. Given that our model is faithful to autosegmental theory, this gives an upper bound on the computational power required to model autosegmental phonological rules.