Where Stochastic OT fails: A discrete model of metrical variation

In a remarkable confirmation of OT in an empirical domain for which it was not originally intended, phonological and morphological variation has been successfully modeled by partially ranked categorical constraints (Anttila 1997, 2002). Poetic meter is a good place to extend and test this approach to variation, because there is abundant and diverse quantitative data available for it, and because it is typically governed by a relatively small number of well-understood constraints. I report the results of four such studies here. They confirm that choices among metrical options are governed by the interaction of partially ranked constraints, in each case constraints that are grounded, and motivated independently of variation data by related systems in which they have a fixed rank. The partially ranked constraint systems turned out to predict not only the relative preferences among metrical options, but also their actual frequencies in the corpora, with surprising accuracy. These findings support the partial ranking model of variation, and provide an explanatory benchmark beyond the reach of intrinsically weaker stochastic approaches that posit a statistical component for metrical competence (Hayes & MacEachern 1998). We derive the distribution of verse types by constraint systems which are formally analogous to grammatical constraint systems in the following respects. (1) a. The constraint systems accept any input and generate well-formed verse types as out- puts. b. Constraints are partially ranked. We consider the set of fully ranked constraint systems consistent with the permitted rankings. c. M is METRICAL if it is the optimal output in at least one such constraint system for at least one input.