A Connectionist Simulation of Structural Rule Learning in Language Acquisition

According to a dual-mechanism hypothesis, although statistical computations based on nonadjacent transitional probabilities may suffice for speech segmentation, an additional rule-following mechanism is required in order to extract structural information out of the linguistic stream. We present a neural network study that shows how statistics alone can support the discovery of structural regularities, beyond the segmentation of speech, disconfirming the dualmechanism hypothesis.