Learnability and the Statistical Structure of Language: Poverty of Stimulus Arguments Revisited

1. Introduction Statistical learning, and " any account which assigns a fundamental role to segmentation, categorization, analogy, and generalization " is rejected in Chom-skyan linguistics as " mistaken in principle " (Chomsky, 1975). Acquisition is viewed, rather, as a search through the set of possible grammars for natural language , guided by successive inputs; or alternatively, as a parameter setting process in which the inputs serve as triggers. The stochastic nature of the input is thus ignored supposedly the learner is oblivious to the distributional frequencies of lexical items, grammatical constructions, and utterance types. Recent acquisition research, however, has shown that children, and even infants , are sensitive to the statistical structure of their linguistic input (Saffran et al. The situation with respect to learnability is thus significantly different from that which has been assumed. Stochastic languages may be learnable from positive examples alone, while their non-stochastic analogues require negative evi-observed, distributional information can provide " a kind of 'negative evidence' " in that expectations can be formed which may then be violated. And so, in at least some cases, the so-called 'logical problems' associated with the no negative evidence hypothesis may be solved by admitting the stochastic information. Thus, if UG is to account for all and only those properties of language " that can reasonably be supposed not to have been learned " (Chomsky, 1975) we must adopt a learning theory which is sensitive to the statistical properties of the input, and reassess poverty of stimulus arguments under those theoretical assumptions. This paper illustrates this by showing that the " parade case of an innate constraint " (Crain, 1991) i.e., Chomsky's (1975) poverty of stimulus argument that structure dependence must be a principle of UG fails to hold once stochas-tic information is admitted; the property of language in question is shown to be learnable with a statistical learning algorithm. Chomsky (1975) suggests that it is reasonable to suppose that aux-questions are derived from declaratives, and so children, presumably exposed to only simple forms of either type, should be free to generate either of two sorts of rules: a structure-independent rule i.e. move the first 'is' or the correct structure-dependent rule. Chomsky argues that since " cases that distinguish the hypotheses rarely arise, " (Piatelli-Palmarini, 1980) at least some children can be assumed not to encounter the relevant evidence for a considerable portion of their lives. Thus, since the …