Emergent feature structures: harmony systems in exemplar models of phonology

In exemplar models of phonology, phonotactic constraints are modeled as emergent from patterns of high activation between units that co-occur with statistical regularity, or as patterns of low activation or inhibition between units that co-occur less frequently or not at all. Exemplar models posit no a priori formal or representational properties to the phonological units or sound patterns that emerge from the statistical regularities of speech, in contrast to analyses in the generative phonology tradition, including Optimality Theory, where sound patterns are determined by well-formedness constraints on phonological structures. This paper focuses on the analysis of long-distance assimilation, i.e., harmony systems, evaluating the predictions of generative analyses based on constraints on representation against typological and experimental evidence. Representational approaches model assimilation with constraints that favor extended feature structures. The question addressed here is whether and how the feature structures of harmony systems can be modeled as emergent structure. It is shown that an exemplar account can model the co-occurrence patterns of harmony systems in the transitional probabilities between segments that share the harmony feature, without invoking feature structure, but that the domain properties of harmony feature structures emerge due to the associations between phonological units (the harmonizing segments) and the morphological units that delimit harmony domains. This association grounds the sound pattern in the lexicon, and provides a “convergence of regularities” [Frisch, S., 2007. Levels of representation in acquisition (Commentary). In: Cole, J., Hualde, J.I. (Eds.), Laboratory Phonology, vol. 9. Mounton de Gruyter, Berlin, pp. 339-352] which facilitates learning.