The nature of regressions in the acquisition of phonological grammars

Under one typical view, children’s acquisition of their L1 phonological grammar is understood as a gradual progression from an initial universal state towards a language-specific one, in which learners respond to mismatches between their outputs and the ambient language (i.e. their ‘errors’) by changing their grammars incrementally to better approximate the target (i.e. ‘resolving’ their errors). One challenging problem for this view are the many reports of ‘U-shaped development’ in which production temporarily regresses, diverging further from the target rather than drawing closer: see e.g. Menn (1971) et seq; Macken (1980); Vihman and Velleman (1989), (2002); Bleile and Tomblin (1991); Bernhardt and Stemberger (1998); Stemberger, Bernhardt and Johnson (1999); Becker and Tessier (2011). To what extent do such regressions cast doubt on the view of phonological acquisition as a gradual process of grammatical error resolution? Based on existing and novel analyses of longitudinal data, this paper argues that phonological regressions should not be captured directly within the normal workings of children’s error-driven mechanisms for grammar learning. Section 2 defines the crucial, problematic type of U-shaped development – grammatical backtracking – and claims that grammatical backtracking is restricted to childspecific processes, suggesting an exceptional treatment of these regressions via child-specific constraints that are induced over the course of learning (in the spirit of Becker and Tessier, 2011; see also Inkelas and Rose, 2008). Beyond this limited grammatical backtracking (and other types of regressions which are argued not to be grammatical in nature, see 2.4) section 3 identifies the kind of regression that seems plausible but is nonetheless apparently unattested: one in which markedness constraints flip-flop over time, so that improvement on one marked structure entails regression on another. With this initial empirical base, section 4 demonstrates that an error-driven OT-like learner which stores its errors and imposes certain persistent biases can in fact easily regress in the unattested way. Section 5 discusses how OT’s grammatical parallelism is in part responsible for creating the unattested regression pattern, and how a serial constraintbased grammar like Harmonic Serialism (McCarthy 2007 et seq) avoids this regression.