1.2.2 The Issue of Language Acquisition
A second theoretical reason for CLASPnet is more linguistically inspired. Consider these statements by Chomsky about language acquisition: "The term "learning" is, in fact, a very misleading one and one that is probably best abandoned as a relic of an earlier age, and earlier misunderstandings" (1987: 635); or "One striking feature is that, although [these behaviorist theories] have been widely believed, indeed asserted as virtual doctrinal truths they are supported by no compelling evidence. In fact, attention to the simplest facts suffices to undermine them" (1988: 162); and "there is only one proposal on the table that merits serious consideration" (1991: 35). That proposal, of course, is that a fair amount of linguistic knowledge has to be innate in order for language learning to be possible at all.
One important reference that is usually mentioned in support of Chomsky's position is Gold's (1967) article on language learnability. Gold's conclusion was that natural languages like English, Mandarin Chinese, or Warlpiri can not be learned on the basis of positive evidence alone -- i.e. it is not possible for a child to find the correct formal description for her grammar in a finite amount of time if she does not receive explicit information about which constructions are not allowed in the language. Given the complexity of natural languages, the search space of possible grammars is just too large to be scanned exhaustively (Atkinson 1986). So, only an innate endowment which constrains the search space for the language learning child can offer a solution.
As there is nothing faulty with Gold's argumentation, the conclusion seems unavoidable as well. Nonetheless, dissenting voices have been raised. Some have pointed out that children do receive negative feedback when they mistakes. It is far from clear, however, whether such feedback is always of sufficient quality and quantity to really help the children (Bavin 1992). A second line of reasoning holds that Gold was wrong in picturing language learning as trying to find the unique formal description for a language. If a language is not to be described with exact formal rules, more solutions are open to the child; and if there's no unique target for a given language, then more than one possible solution should be considered valid, as the target grammar is no longer a single point in the search space, but a much larger region. Finally, a third suggestion has recently been put forward by Quartz & Sejnowski (in press). They point out that Gold's model fails to take into account that the language learning mechanism which children have to use (i.e. their brains) develops continuously during the learning process: the absolute number of neurons is changing constantly as are the axonic and dendritic trees of each individual neuron. Consequently, it may well be that these neurophysiological changes reduce the search space which the child has to consider to the point where the target grammar can be learned without too much difficulty (cf. Clark & Thornton 1994; Elman 1993). As the three lines of suggestions do not exclude each other, it seems fair to conclude that Gold's argumentation need no longer be accepted at face value.
Against this background, it is easy to see where CLASPnet fits in. The model is an attempt to illustrate that some non-trivial aspects of language can be learned successfully by a neural network without an innate linguistic endowment. Because of connectionist simulations like those of Plunkett & Marchman (1991, 1993), Sinha & Astorp (1993), or MacWhinney & Leinbach (1991), Chomsky's claim that "no alternative is suggested, even vaguely" (1991: 32) appears to be in need of careful reconsideration.