Modularity and Artificial Grammar Learning

People have an impressive capacity for storing information about particular events. This "episodic" memory allows us to recall the context of specific experiences, such as what we did on our last holiday, and is now fairly well understood both psychologically (Tulving, 1983) and at the neural level (McClelland, McNaughton, & O'Reilly, 1995). We also have an ability, however, to acquire knowledge about generalities, that is, properties of classes of objects or events. We can judge the grammaticality of a novel sentence, read a word in an unfamiliar script, perform arithmetic operations, and so on. These seem to require representations of abstract, general properties such as the rules of a grammar that transcend and are separate from knowledge of specific objects or events. Cognitive psychology has traditionally dealt with this distinction by assuming separate processes for acquiring specific and general knowledge. Under various terms (e.g., episodic, explicit, declarative), knowledge of specific events is assumed to be distinct from knowledge about general properties (e.g., semantic, implicit, procedural, nondeclarative). A puzzle, however, is to explain how we acquire general knowledge as abstract properties themselves are never directly observed (see Whittlesea, 1997a, b). Instead, such properties must be induced from multiple experiences with specific objects or events. Hence, the separate-systems account assumes that there exists a mechanism for creating abstractions across specific experiences. Moreover, as we are not normally deliberately intending to perform such abstraction, it must be largely an incidental and unconscious process. The aim of this chapter is to evaluate this modular, separate-systems account in one particular branch of implicit learning research. Undoubtedly, there is a wealth of evidence consistent with the separate-systems account with a good deal of that evidence coming from artificial grammar learning (AGL) research (see Meulemans and Van der Linden, this volume, for a review). For example, Shanks et al. 3 Knowlton, Ramus, and Squire (1992) trained normal participants and amnesic patients by asking them to memorise strings of letters generated from a finite-state grammar. One of the two grammars Knowlton et al. used is shown in Figure 1. This grammar specifies rules, similar to those that exist in natural languages, for ordering string elements. Grammatical strings are generated by entering the diagram at the leftmost node and moving along legal pathways, as indicated by the arrows, collecting letters, until an exit point is reached on the right-hand side. The letter string XXVXJJ is grammatical as it can …