As before, stimuli for this experiment were generated randomly, given the constraints which defined each of the categories. Two classes of categories (noun and adjective) were defined that were identical with respect to all of the variables of interest (volume, compactness, lexical dimensions, overlap). They differed only in terms of where the member categories were located in the representational space. The categories, 18 in each class, were defined in such a way that in the representational space, each noun category was surrounded by adjective categories and vice versa. The pattern of noun and adjective categories resembled a multi-dimensional checkerboard. Thus at the level of the meta-categories, there was no generalization whatsoever to be made about the nature of the member categories or the particular regions associated with nouns or adjectives. In a sense, the meta-categories had no semantics associated with them. Each category took up .003 of the space; this left uncategorized regions of representational space separating adjacent categories. There was no overlap between categories. As in experiments other than Experiment 1, there were four input dimensions defining the perceptual properties of the object, but in this case, there were only two linguistic context inputs, one for one class of words and the other for the second class.
As in all of the experiments, the network was trained on randomly generated instances of the categories. In this case, the network was tested, however, on a set of 18 pre-defined object input patterns which did not belong to any of the noun or adjective categories; that is, these inputs fell in the gaps between the categories which the network had been trained on. Each of these 18 patterns was tested once together with a noun linguistic context and once with an adjective linguistic context. The relevant dependent variable in each case is the relative activation over the noun and the adjective output units. If the network has begun to divide the words into meta-categories on the basis of the linguistic context, we should see higher mean activations on the adjective units when the adjective linguistic context is presented and higher activations on the noun units when the noun linguistic context is presented.