Complementing Bart Geurts' (2000) review of the same book, Ken Drozd's Learnability and linguistic performance points out several serious problems with the L1 acquisition research programme of Crain & Thornton (1998). Drozd's objections are so clearly stated that there is little I can add. Nonetheless, I will attempt to flesh out some of his more general points by examining more closely a specific case, namely the experiments in Crain, Thornton, Boster, Conway, Lillo-Martin & Woodams (1996) which focused on normal preschool children's understanding of simple universally quantified sentences.

Much of Drozd's penetrating review of Investigations in Universal Grammar (IUG), beyond noting countless lamentable misattributions of ideas and inconsistencies, evaluates its methodology. The standardization of methodology in IUG can, I believe, undermine our understanding of acquisition more than it supports it. It is a pity if we lose the real value of their useful experimental refinements of existing experiments (and several novel ones in semantics).

I would like to begin by clarifying some technical points regarding learnability and Gold's theorem and will work my way around to advocating the power of distributed connectionist networks. I may wind up repeating much of what MacWhinney has already said, but I hope to do so in a manner that clarifies the status of and relationship between his seven solutions to the language acquisition problem.

Evaluating Brian MacWhinney's multiple ‘solutions’ to the logical problem of language acquisition requires delineating exactly what the alleged problem is. He takes it to stem from the theorems of Gold (1967), the most celebrated of which is G1.

Radford (1988: 2) and others before and after have argued that the core theoretical question (i) ‘How are human languages structured?’ is logically prior to another core theoretical question (ii) ‘How are human languages acquired?’ This is said to be so because the answer to (i), the structure of human language, is the real target of question (ii) – language acquisition is about acquiring the structure of a human language. Thus, we can't ask about how a thing is learned before we know what it is. Bearing this dictum in mind, consideration of both Crain & Thornton (1998) (C&T) and the Drozd critique (present volume) raises an interesting question: is linguistic theory and the data on which it is established sufficiently developed to inform ‘external’ empirical studies in areas such as language acquisition? At this point, it isn't clear that it is. The data of theoretical linguistics is often idealized and sometimes rather narrow. A given construction may be narrowly investigated (cited) for its apparent interest as a side-light on another area or a more general theoretical issue rather than being investigated more fully/paradigmatically in its own right. Further, the latter sort of investigation might shed a very different light on a phenomenon only investigated narrowly. So paradoxically, it is often only in the context of doing experimental work such as acquisition or variation research that such fuller investigation takes place, possibly revealing new and crucial facts which may prove relevant to theory construction.

In his review of Crain & Thornton's (C&T) (1998) ‘Investigations in Universal Grammar’ (IUG), Drozd raises the question of how models of children's linguistic competence relate to their linguistic performance. He highlights the use of processing limitations (PLs) to explain children's non-adultlike performance in most models of acquisition that adopt a Universal Grammar (UG) framework. Although the Modularity Matching (MM) model claims to avoid the need for performance-based constraints by assuming that children and adults operate with the same processing resources, Drozd argues that the model lacks predictive power.

These twelve thoughtful commentaries demonstrate interesting shifts in our collective understanding of the ‘logical problem of language acquisition.’ The bulk of the commentary supports the multiple process approach to the logical problem. At the same time, there is strong and productive disagreement regarding the ways in which conservatism, competition, probabilistic identification, indirect negative evidence, item-based learning, cue construction, monitoring, and constraints make their separate and related contributions to the learning of specific target structures.

Three important themes emerge from the commentaries offered in response to ‘Learnability and linguistic performance’. In this reply, I would like to critically but briefly review these themes.