Connectionist Models of Language Processing

Abstract

Connectionist models of language resemble, in one way or another, the central nervous system of humans. Classical neurological models proposed by nineteenth-century neurologists specify centers and pathways that are analogous to cortical areas and fiber bundles most relevant to language, and modern neurobiological approaches propose intensely connected cell assemblies with different cortical distributions as the brain basis of language. Symbolic connectionist models suggest single artificial neurons corresponding to linguistic units (language sounds, words, etc.), and distributed connectionist models represent such linguistic entities by activity vectors involving numerous neuronal elements. A recent debate raises the question whether rules of language and their exceptions can be modeled by a single distributed network of two or three layers of artificial neurons. As an alternative, it has been proposed that a more complex network structure is necessary where subcomponents specialize in the storage of knowledge of rules and irregulars, respectively. This ongoing debate documents fruitful multidisciplinary interaction between the linguistic, cognitive, computational, and brain sciences.