|Comment:||Invited talk at CLS 50 (2014); version submitted to Proceedings volume|
|Abstract:||The phonotactic learner of Hayes and Wilson (LI 2008) discovers what could be called absolute phonotactics: using a maxent framework, it selects and weights constraints so as to maximize the predicted probability of the set of existing words against a backdrop of all possible strings. The same apparatus can be used for comparative phonotactics: given two populations of strings, A and B, we seek a grammar whose output probabilities accurately indicate the likelihood that any given novel string will belong to A or B.
Do language-acquiring children learn comparative phonotactics? I think it likely that they do, and indeed that they do so for multiple purposes. Such would include part-of-speech prediction (work of Christiansen), prediction of gender (work of Lyster and others), and two areas I focus on here. I put forth a comparative phonotactics that singles out words of the Latinate vocabulary stratum of English (Chomsky and Halle 1968), distinguishing it from the native stratum, and ponder how the presence of such strata in a language could be detected by a bootstrapping process. I will also use comparative phonotactic analysis to carry out stem sorting in the sense of Becker and Gouskova: the population of stems in a language can be sorted according to which affix allomorphs they take. The pattern predicted by stem-sorting-cum-allomorph-selection is often indistinguishable from the result of ordinary GEN+EVAL phonology, but in one area of Hungarian vowel harmony, the evidence is quite clear that the pattern must be the result of stem sorting.
|Area/Keywords:||phonology, phonotactics, lexical strata|