U-Shaped Learning in Phonological Development

(1999). "Regressions" ("u"-shaped learning) in the acquisition of prosodic structure. This paper addresses learning of constraint rankings by children. We argue that constraint demotion cannot account for all developmental patterns observed in child phonology. Learning algorithms that allow for constraint promotion fare better, but some still cannot account for all developmental changes. Child language contains many inaccuracies relative to the target adult language. Over the course of development, these inaccuracies disappear. Most developmental changes bring the child's speech closer to the target language. Theories of learning focus on how such improvements are accomplished. However, a minority of developmental changes are U-SHAPED: a developmental change in which there is a decrease in accuracy relative to the adult target. Historically, much of the debate over different theories of learning has focused on how each theory deals with these relatively rare developmental occurrences (e. Regressions are also the best test of the adequacy of learning models proposed for OT. We argue that constraint promotion must be a possible change during-PAGE 2-learning, contra Tesar and Smolensky (1998), and Prince and Tesar (1999). Some learning algorithms that allow for constraint promotion, such as that of Boersma and Hayes (1999), may nonetheless have difficulty accounting for some regressions, because of restrictions on which constraints may be promoted. These models fail to account for the particular set of regressions examined here, and hence are not adequate to describe learning by human beings. 1. A Typology of Phonological Regressions U-shaped learning appears to be especially prevalent in the development of phonology (the pronunciation of words), where it is known as regression. We have found it useful to describe phonological regressions using two dimensions that describe how the system has changed: TRADE-OFF vs. PURE: whether along with the regression there is some improvement in the same output CORRELATED vs. UNCORRELATED: whether the regression correlates with a change made to the system that leads to improvements in other outputs than the one showing the regression ALL theories of learning predict regressions only when there is a motivation: trade-off: Due to limitations in the range of outputs, altering the system to achieve accuracy on one dimension of a form may be at the expense of accuracy on another dimension. correlated: Changes that increase the range of possible outputs or that improve the efficiency of processing may make errors possible that before were impossible, so the system was not set up to …