Statistical approaches to language acquisition and the self-organizing consciousness: a reversal of perspective

The shift from Universal to idiosyncratic grammar has obvious consequences with regard to an innatist position. A specific rule needs to be learned. However, there is no clear explanation of how a rule can be learned. For instance, when referring to a monograph of Marcus et al. on over-generalization, Bybee (1995, pp.449–450) noted that the authors “do not address the question of exactly how the rule is extracted and deposited in a separable module. Thus, despite the fact that the relevant section in their paper is named “How might a regular rule be learned?”, the question of how the restructuring from lexical schema to symbolic rule takes place is actually not addressed!”

This is reminiscent of the traditional Chomskyan distinction between (inferred) competence and (observed) performance, and its associated requirement for researchers to account for the extraneous limitations that prevent unlimited competence from fully translating into performance.

Consider the following story for a short illustration of the ambiguities inherent in the notion of “computational irrelevance.” A cognitive scientist builds a psychological model that implies at some location the computation of 2 × 2 = 4. The scientist notes (rightly) that nothing has changed in his program according to whether or not this computation is thought to be accompanied by a conscious experience in the actual subjects he intends to simulate. From this observation, he coins the notion of “computational irrelevance of consciousness.” In some later location of his program, he requires that his model now computes the cubic power of 6371, and he is first baffled by the fact that it is highly implausible that actual subjects are able to perform this computation mentally. Fortunately, he realizes that this limitation holds only if it is assumed that this computation is performed consciously. Because the issue of consciousness has been declared computationally irrelevant, this slight glitch is automatically removed, and even turns out to be an advantage: His program succeeds at simulating behavior without ever involving the need for consciousness! This story allows a contrast of two ways of conceiving constraints and parsimony: The conventional one, in which actual subjects are assumed to be able to compute the cubic power of 6371, but which calls for no conscious counterpart for this computation, and the one to which I subscribe, in which actual subjects are assumed to be able to compute only what they do consciously, which is hardly more than 2 × 2.

Although they do not directly address the word segmentation issue, Boucher and Dienes (2003) also explore the possibility that the sensitivity to statistical regularities is not the result of statistical computations on individual elements, but rather the by-product of local representations of chunks of individual elements. In their view, to borrow the title of their paper, there are “two ways of learning associations,” one in which chunking is an emergent property of statistical analyses, and the other in which chunking is a primitive process, the result of which amounts to simulating statistical computations. However, they do not equate their chunks with the focal content of phenomenal consciousness, as I do.

This presentation is obviously an over-simplification. I do not intend to mean that learning a language amounts to successively isolating the words from the speech flow, to map the word to its meaning, and so on. Those processes are considered here as independent steps for the sake of clarity, but it would be in keeping with a dynamic approach to consider them as closely interconnected processes.