Redington, M. (1996). What is learnt in artificial grammar learning? Unpublished doctoral dissertation, Department of Experimental Psychology, University of Oxford.



Abstract

In artificial grammar learning, participants memorise items which (unknown to them) obey the rules of a finite state grammar. Participants are then informed of the rule-governed nature of the items, and asked to judge whether each of a new set of items obeys or violates the rules. Participants' performance on this classification task reliably exceeds chance and control levels.

This thesis is concerned with the nature and representation of the knowledge acquired in this task.

I present a family of fragment-based "toy" models, that can account for participants' classification performance with a wide range of materials. These toy models can explain transfer performance (where the surface form of the stimuli is changed between training and test), by positing processes which abstract across surface forms during the classification task. I report studies using the guessing game paradigm, showing that grammaticality judgement and the ability to predict successive elements of a test item are reliably associated across a wide range of conditions, consistent with a unitary fragment learning account. I also report simulations with connectionist models, showing that models which are sensitive to arbitrary (nonadjacent) structure do not capture crucial aspects of human performance.

I argue that fragment learning provides a more parsimonious account of artificial grammar learning than accounts based on abstract rules (Reber, 1967), exemplar learning (Brooks & Vokey, 1991), or combinations of these kinds of knowledge.


Back to publications ...


Last modified: Jan 10, 1999