LEARNING ARTIFICIAL GRAMMARS WITH COMPETITIVE CHUNKING

When exposed to a regular stimulus field, for instance, that generated by an artificial grammar, subjects unintentionally learn to respond efficiently to the underlying structure (Miller, 1958; Reber 1967). We explored the hypothesis that the learning process is chunking and that grammatical knowledge is implicitly encoded in a hierarchical network of chunks. We trained subjects on exemplar sentences while inducing them to form specific chunks. Their knowledge was then assessed through judgments of grammaticality. We found that subjects were less sensitive to violations that preserved their chunks than to violations that did not. We derived the theory of competitive chunking (CC) and found that it successfully reproduces, via computer simulations, both Miller's experimental results and our own. In CC, chunks are hierarchical structures strengthened with use by a bottom-up perception process. Strength-mediated competitions determine which chunks are created and which are used by the perception process.