University of California, Los Angeles, 1999, -128 pp.These notes were prepared for a UCLA seminar on computational proposals in recent phonology. Very many corrections and contributions were made by the seminar participants, especially Adam Albright, Dan Albro, Marco Baroni, Leston Buell, Bruce Hayes, Gianluca Storto, Siri Tuttle. Thanks also to Ed Keenan for some corrections and suggestions. The notes are still rough (many typos are left, I’m sure). I hope to improve them! They are intended to be an accompaniment to the literature, not a replacement; they presuppose an acquaintance with the original sources that are discussed.
One of the main traditions in computational phonology is based on finite state models of phonological constraints. This is perhaps surprising, since finite state models, at least at first blush, seem to be too strong and too weak. They seem too strong because phonological relations seem to be local for the most part, in a way that dependencies in finite state languages are not. (For example, it is easy to define a finite state language with strings that have either a single a or b followed by any number of c’s, followed by a repetition of the first symbol: (ac∗a) ∪ (bc∗b). The final symbol can depend on a symbol that occurred arbitrarily far back in the sequence.) And on the other hand, finite state models are too weak in the sense that some phenomena exhibit dependencies of a kind that cannot be captured by these devices: notably, reduplication. These issues come up repeatedly in these notes.
These notes go slightly beyond what is already in the literature in only a couple of places. We are perhaps clearer about the one-level/two-level distinction in §§5.2,7.3 than the literature has been. And rather than restricting attention to finite state compositions as is sometimes done, we take the perhaps less practical but scientifically more promising route of emphasizing the prospects for composing finite state models witht he grammars of larger abstract families of languages in §§7.4,10.2.
Formal, computational models are important in linguistics for two main reasons. First, the project of making our vague ideas about language elegant and fully formal is a useful one. It improves our understanding of the real claims of the grammar, and it enables careful comparisons of competing ideas. Second, the best models we have of human language acquisition and use are computational. That is, they regard people using language as going through some changes which can be modeled as formal derivations. The idea that the relevant changes in language learning and language use are derivations of some kind is an empirical hypothesis which may well be false, but it is the best one we have. In my view, the main project of linguistic theory is to provide this computational account.
Since theoretical linguistics provides formal generative models of language, it implicitly treats human language learners and language users as computers. The existence of the artifacts we usually call computers is really beside the point. Computers are useful in the development of linguistics in just the way that they are useful in physics or biology: they sometimes facilitate calculations. These calculations are not the reason that our pursuit is called computational. The reason the subject at hand is called computational phonology is that we adopt the programmatic hypothesis that the abilities we are modeling are computational.Preface
Finite recognizers of languages
Some early proposals
Using non-deterministic machines
One level phonology
Optimality theory: first ideas
OTP: Primitive optimality theory
Lenient compositions: the proper treatment of OT?
Acquisition models
Exercises and speculations
