Positional Neutralization in an Exemplar Model: The Role of Unique Inflectional Bases

This paper explores whether an exemplar-based framework is capable of modeling positional neutralization (specifically, final devoicing) within a morphological paradigm. In an initial simulation of a 10-item lexicon, a statistical bias towards final devoicing and a similarity bias for words in the same paradigm to resemble each other combined to yield morphological leveling: word-final voiceless obstruents in one form spread even to morphologically related forms in which they were not word-final. These simulations were incapable of modeling a language with voicing alternations, as attested in German and other languages with final devoicing. However, the simulations were successful when the similarity bias was modified such that words were under pressure to resemble their morphologically related counterparts only when those counterparts were informative (i.e., not neutralized). I conclude that in order for exemplar models to be descriptively adequate, they must implement paradigm uniformity among surface forms that is crucially asymmetrical.