In all of the experiments we have described, there are two classes of categories to be learned, nouns and adjectives, differing in one or more ways. The task of the network is to learn the categories, and we have shown how certain differences between classes of categories can affect the rate of and ultimate level of learning. The network's task is not, however, to learn that there are two classes of categories and to discover how these classes are distinguished. Ultimately children do learn to make this distinction. Does our simple model have anything to say about how this is accomplished?
While the network starts the task without the knowledge that there are two classes of categories, it does have access to a much more direct indication of the distinction: the linguistic contexts associated with the two classes of words. More precisely, what these inputs tell the network is simply that there is a distinction to be made. But does the network use the linguistic context inputs in this way? The explicit task of the network is to map input objects, accompanied by linguistic contexts, onto one label or another. However, if the linguistic context is informative for this task, then we would expect the network to also learn to associate particular contexts with particular words. These associations, in a sense, would constitute the beginnings of syntactic categories. In this final experiment, we ask what the network can learn when the meta-categories associated with specific linguistic inputs, that is, noun and adjective, are more arbitrarily defined than the classes of categories thus far examined. If noun and adjective are just arbitrary collections of categories, the network will have to rely on the linguistic context input if it is to learn anything about these meta-categories.