A Proposal for a Computational Model of Phonemic Acquisition

Abstract

In the field of computational linguistics, computational modeling of linguistic behavior has been motivated not only by the creation of practical language-related tools such as machine translation, automatic speech recognition, speaker identification, and natural language search, but also by a desire to deepen our understanding of how language works, either in an abstract mathematical sense or in the more literal sense of describing human behavior at various levels of analysis. These models take various forms, some derived from mathematical models of electronic transmission of information (Shannon, 1948), others from abstract models of neural behavior (McCulloch and Pitts, 1943; Rosenblatt, 1958). In computational neuroscience, computer models are developed to mimic the behavior of brains, with a greater degree of biological realism. These models focus on neural behavior ranging from single neurons to large-scale networks of neurons. Typically, the behavior of interest is the relative activation of groups of neurons, the emergence of synchronized or otherwise patterned activation, and the propagation of signals across networks. The elaboration of relatively high-level cognitive behavior is, at best, secondary to the exploration of low-level physical and electrical interaction (Zednik, 2018). The growing field of computational cognitive neuroscience has as a goal the development of computational models that are biologically plausible and that exhibit cognitive behavior of interest (Ashby, 2011). Linguistic models of this type are intended to exhibit the kind of linguistic behavior that is observed in humans, but with an underlying structure and behavior that closely parallels the human brain. Marr (1982) offers a three-level framework for analyzing models of the brain that has become a standard in neuroscience (Bechtel, 2014). Applying this method of analysis to broad classes of computational models yields insights into the strengths and weaknesses of each. While the ideas in this dissertation may ultimately find broader relevance, it is presently concerned primarily with the modeling of phonemic acquisition in infants. Application of Marr’s analysis to the actual system being modeled – the human infant – suggests an approach to the development of acquisition models that departs significantly from traditional computational linguistics models and computational cognitive neuroscience models.