CHUNKS ARE NOT ENOUGH - THE INSUFFICIENCY OF FEATURE FREQUENCY-BASED EXPLANATIONS OF ARTIFICIAL GRAMMAR LEARNING

Two experiments tested chunk frequency explanations of artificial gram mar learning which hold that classification performance is dependent o n some metric derived form the frequency with which certain features o ccur within the letter string stimuli. Experiment 1 revealed that clas sification performance was affected by close graphemic similarity betw een specific training (e.g., MXRVXT) and test strings (e.g., MXRMXT), despite the fact that similar strings did not contain frequently occur ring features (e.g., bigrams or trigrams). This effect was replicated in Experiment 2a and Experiment 2b demonstrated that substituting lett ers to make the consonant strings pronounceable (e.g., substituting X, R, and T, in the consonant string MXRMXT with Y, A, I, to produce MYA MYI) affected classification performance, despite the fact that object ive measures of feature frequency were not altered. It is argued that models of classification that focus entirely on the frequency of featu res within the literal stimulus are insufficient, and that some allowa nce must be made for how the stimulus is encoded.