| This thesis uses the continuous representation space of a Connectionist network to make predictions about syntactic innovation in natural language. It proposes to replace current, categorically organized accounts of linguistic structure with a metrically organized model in which innovation is treated as a quantitative interpolation process.; Current theories of grammar, following Chomsky 1957, are based mainly on data gathered by introspection. Their generalizations tend to be sweeping. The inventories of representational and derivational devices they posit are large. Such theories are not very useful for making constrained predictions about historical grammar change for there is no domain-independent structuring of the representation space that reveals which changes are probable and which are not. Nevertheless, recent work in the field of gramaticalization (e.g., Traugott and Heine 1991, Hopper and Traugott 1993) indicates strong constraints on grammar change. These constraints are especially apparent in sequences of closely-spaced historical texts, for the distributional properties of such texts change gradually. The gradualness is evident both at a categorical level, in that successively emerging types tend to be similar to one another, and at a quantitative level, in that constructions appear and disappear via long periods of probabilistic alternation, with the probabilities varying gradually.; I propose to unite probabilistic and categorical gradualness in a single mechanism. I replace categories with clusters and let quantitative differences in usage rates manifest themselves as contrasting distances from a prototype. The theory is implemented in a recurrent Connectionist network trained on the next-word prediction task studied by Elman 1990 and 1991. Three empirical phenomena provide evidence in its favor: linked frequency changes in grammatically related constructions, quantitative changes which anticipate events of categorical reanalysis, and emergence of hybrid structures during periods of transition. Case studies of degree modifier sort/kind of and of future be going to are presented. The model permits simplification of the theory of language change by replacing reanalysis and analogy with a single mechanism. It also permits simplification of the theory of language structure by allowing the same interpolative mechanism to handle both normal productive syntax and the hitherto problematic hybrid cases. |
