Can connectionist models of phonology assembly account for phonology?

Abstract

Connectionist models have gained considerable success as accounts of how printed words are named. Their success challenges the view of grapheme-to-phoneme correspondences (GPCs) as rules. By extension, however, this challenge is sometimes interpreted also as evidence against linguistic rules and variables. This inference tacitly assumes that the generalizations inherent in reading (specifically, GPCs) are similar in their scope to linguistic generalizations and that they are each reducible to token associations. I examine this assumption by comparing the scope of generalizationsrequired for mapping graphemes to phonemes and several linguistic phonological generalizations. Marcus (1998b) distinguishes between two types of generalizations: those that fall within a model's training space and those that exceed it. The scope of generalizations is determined by the model's representational choicesspecifically, the implementation of operations over mental variables. An analysis of GPCs suggests that such generalizations do not appeal to variables; hence, they may not exceed the training space. Likewise, certain phonological regularities,such as syllable phonotactic constraints and place assimilation, may be captured by an associativeprocess. In contrast, other phonological processes appeal to variables; hence, such generalizations potentially exceed the training space. I discuss one such case, the obligatory contour principle. I demonstrate that speakers conform to this constraint and that their behavior is inexplicable by the statistical structure of the language. This analysis suggests that, unlike GPCs, phonological generalizations may exceed the training space. Thus, despite their success in modeling GPCs, eliminative connectionist models of phonology assembly may be unable to provide a complete account for phonology. To the extent that reading is subject to phonological constraints, its modeling may require implementing operations over variables.

Large‐Scale Modeling of Wordform Learning and Representation

“…It has been argued that connectionist models are fundamentally limited in their ability to account for realistic language phenomena (G. F. Marcus, 2001). The biggest problem in this regard is the limited ability of many connectionist models to generalize much beyond the items on which they are trained (Berent, 2001;G. F. Marcus, 2001;Berent, Marcus, Shimron, & Gafos, 2002).…”

“…F. Marcus, 2001). The biggest problem in this regard is the limited ability of many connectionist models to generalize much beyond the items on which they are trained (Berent, 2001; G. F. Marcus, 2001; Berent, Marcus, Shimron, & Gafos, 2002).…”