This dissertation presents a connectionist theory of how infinite languages may fit within finite minds. Arguments are presented against the distinction between linguistic competence and observable language performance. It is suggested that certain kinds of finite state automata - i.e., recurrent neural networks - are likely to have sufficient computational power, and the necessary generalization capability, to serve as models for the processing and acquisition of linguistic structure. These arguments are further corroborated by a number of computer simulations, demonstrating that recurrent connectionist models are able to learn complex recursive regularities and have powerful generalization abilities. Importantly, the performance evinced by the networks are comparable with observed human behavior on similar aspects of language. Moreover, an evolutionary account is provided, advocating a learning and processing based explanation of the origin and subsequent phylogenetic development of language. This view construes language as a nonobligate symbiant, arguing that language has evolved to fit human learning and processing mechanisms, rather than vice versa. As such, this perspective promises to explain linguistic universals in functional terms, and motivates an account of language acquisition which incorporates innate, but not language-specific constraints on the learning process. The purported poverty of the stimulus is re-appraised in this light, and it is concluded that linguistic structure may be learnable by bottom-up statistical learning models, such as, connectionist networks.
Dissertation Advisor: Nick Chater (now at the University of Warwick).
External Examiner: Noel Sharkey, University of Sheffield.
Please email suggestions/errors to mhc27@cornell.edu |