The general acceptance of the idea that young children distinguish between nouns as name for things and adjectives as labels for the properties of things is based on the facts of the noun advantage and the pattern of within-adjective confusions that characterize children's slow and errorful acquisition of dimensional terms. The simple network that we have studied distinguishes nouns and adjectives in the very same way that young children do: It learns noun categories faster than adjective categories and during the protracted course of learning adjectives, its errors consist of confusing one adjective with another and not of confusing an adjective with a noun. Thus, our network, like children, ``knows'' that nouns and adjectives are different.
The processes that make up this ``knowing'' by the network, however, are not of the kind one usually thinks of as knowledge about the different meanings of nouns and adjectives. All that appears to be known when the network in Experiment 1 makes these errors is (1) the noun categories, (2) the linguistic context that specifies nominal outputs, and (3) the fact that the linguistic contexts that specify adjective categories are not associated with nominal outputs. Apparently this is enough to get a behavioral distinction between nouns and adjectives in the course of learning. The network knows about nouns and ipso facto ``knows'' a class of items that are not nouns. The results remind us that the internal processes that comprise some external pattern of behavior may be simpler than the external behavior itself.
The present network is a very simple model that leaves out much of what children probably do know about nouns and adjectives. While our approach is unabashedly grounded in the semantics of nouns and adjectives, we have tried to show in Experiment 6 how purely form-to-form learning can also play a role in the emergence of syntactic categories. In fact learners appear to have access to a wealth of purely formal information to guide them in learning, and a large body of recent work has focused on the extent to which linguistic categories can be learned on the basis of distributional information [Elm90,FC92] or the formal properties of the words themselves [Kel92]. As in the present model, these approaches are statistical and associative; in fact, many are implemented in the form of connectionist networks. However, given the nature of the inputs and the restricted architecture, the present network obviously cannot make use of the phonology of the words or of the detailed pattern of co-occurrences with other words. We have only sought to demonstrate that syntactic categories can begin to emerge as a kind of side-effect as the system learns to label objects. Note what distinguishes these syntactic categories from the conventional ones, however; because they are directly associated with objects and their properties, they have a semantic force. Although this may not be what is usually meant by theorists who write about children's understanding of the differences between nouns and adjectives, this could be pretty much what the differences amount to in the early stages of acquisition.