We now arrive at a point where both the results of empirical
data and the computational modelling environment support an
associative learning account of the digit span superiority effect
based on the number of encounters with random sequences of digits
within the experience of the rememberer. If this is the case, then
a natural next step is to interrogate the model so that we can construct
digit lists containing pseudo-random digit sequences that
occur frequently versus those that rarely occur. In this way, any
difference in performance cannot be because digits hold some
special characteristic that other stimuli such as words do not;
rather any difference in performance would arise from the
frequency with which the sequences occur in natural language
(under the assumption that associative learning is more likely to
occur for frequent than infrequent sequences).
However, for reasons of tractability, the computational model
was only trained on one-tenth (half a million utterances/
sentences) of the BNC. Given that we have already concluded that
even the full set of BNC data is a fraction of that encountered by
children and adults, rather than interrogate the model we
extracted sequence frequencies from the BNC in full. Experiment
4 therefore compares recall performance for random digit
sequences that occur frequently in the BNC with those that occur
less frequently, the prediction being that the former should facilitate
serial recall more than the latter