The Implicit Learning of Structure: Analogic and Abstractive Strategies in Artificial Grammar Learning

Abstract

<p>For years, psychologists and psycholinguists have described human learning and memory in terms of abstract knowledge structures such as grammars, orthographies, prototypes, scripts and rules. These abstractions often have been thought to be implicit, and acquired through some implicit process. Recently, some theorists have suggested that at least some of the behaviour that has been attributed to implicit, abstract knowledge may actually arise from people's memory for individual cases. In this thesis, this issue is investigated by addressing the recent claims of Arthur Reber and his associates for the rapid, implicit abstraction of artificial grammars. In a series of six experiments, subjects were trained with a sample of items generated from an artificial grammar, and then given a surprise classification test in which they were asked to sort new items with respect to grammatical status. In each experiment, it was demonstrated that what had previously been attributed to implicit abstraction of grammaticality was actually a function of the similarity between transfer items and specific training experiences. It was also shown that variations in how subjects are asked to learn the training items affects their sensitivity to the specific similarity of the transfer items rather than, as Reber and his associates had suggested, their sensitivity to the grammaticality of the items. These results are interpreted in terms of "breadth of transfer" - the degree to which subjects may generalize around their memory for specific experiences. Other results suggest that breadth of transfer is a function of the specific encoding of events, and that variables that affect the encoded similarity between events affect the likelihood of item-specific transfer. All of the results are interpreted in terms of a model of structural learning that suggests that subjects' knowledge of complex domains is represented in memory as multiple traces, and that the close relationship found in these experiments between original item learning and subsequent tests of item recognition and categorical transfer is a function of the memorial distribution of these traces. Finally, the results of post-testing questionnaires that asked the subjects to attribute the bases of their transfer decisions indicated that the subjects knowledge in these tasks is less implicit than previously suggested. In affirmation of the objective analyses of their behaviour, subjects attributed their transfer decisions primarily to the similarity between training and transfer items. The results are discussed in terms of a general framework that suggests that the knowledge underlying people's performance in complex domains consists of a mixture of episodic memory for individual instances and limited, explicit abstractions.</p>