fonolog

One of the most pressing problems for current phonological theory is how to model the statistical knowledge of word frequency etc. that speakers seem to possess. One approach is defended by Bruce Hayes and Zsuzsa Cziráky Londe in their thought-provoking article in the latest issue of Phonology (23:59-104): in their view the statistics are modeled in a stochastic OT grammar; I suppose the most widespread alternative view is one in which the modeling is done in terms of exemplars.

Hayes and Cziráky Londe illustrate their view with facts from Hungarian vowel harmony. If a stem ends in a neutral vowel, the quality of the suffix vowel is distributed in a certain way over the lexicon (depending among other things on the other vowels in the stem). Hungarian speakers seem to know this, because they show the same distribution in a wug test. The authors also show how a stochastic OT grammar can be acquired that gives the right results for the wug test.

A (big) advantage of such a model is that it is fairly conservative: we do not have to throw decades of work on our previous models, since the stochastic part is just an extension of what we already know about OT. The problem is that it is not clear why natural language would work this way: why would the language learner go through the trouble of acquiring the ranking of some fairly obscure constraints, when the only gain is that she can pass a wug test afterwards — not really an evolutionary advantage, one would think. It is not true, for instance, that this detailed knowledge brings a large advantage when encountering a new word: of course it might be (a little bit) helpful to be able to predict the most likely outcome, but why would one have to know the statistuc distribution? And further, one still has to learn the real suffix at some point. This thus seems to be an advantage of an exemplar based model: there it is just assumed that the learner stores all the words, since that is what language learning is under that view anyway; it means that the learner could do a wug test afterwards, but that is just an epiphenomenon.