| Various predictions of the Competition Model, a theory of on-line language acquisition and processing (Bates and MacWhinney, 1989) are tested, using the MAL (Miniature Artificial Language) paradigm with both humans and multi-layer neural networks.; Humans were tested in a series of different dialects of an MAL with regular syntactic and morphological rules which could also be used as cues to the meaning of the sentence. The variables manipulated included the frequency of the cue, the reliability of the information it offered, and the surface form of the cue (i.e., word order, agreement morphology, or animacy). In some conditions, subjects' performance followed the predictions of the Competition Model and profiles seen in child language acquisition, in that their performance initially showed sensitivity to those cues which were more frequent, and then converged later in training upon those cues which were more reliable. However, this effect interacted with the form of the cue in that subjects overall had a more difficult time using agreement morphology than word order, a finding also seen in natural language processing.; Neural networks were tested with languages that were similar in their underlying structure to the MALs used with humans; some versions of the networks showed the same effects as predicted and as seen in normals in that they initially showed sensitivity to those cues which were more frequent, and then converged later in training upon those cues which were more reliable. This effect interacted in an interesting way with the type of network, in that networks with an additional "hidden" layer of processing units were closer to the predicted performance than those with only one layer, even though the one layer networks were well able to solve the problem, suggesting an additional constraint on the types of models that can be used if they are to be psychologically valid.; Implications for the Competition Model and for further research with MALs and their relevance to natural language acquisition are discussed, as well as the useful parallels between MAL research with humans and neural network models. |
