| Higher-order syntactic structure is a hallmark of language, contributing to its expressive power, but also making it more difficult to learn. One such example is the nonadjacent dependency, where one element predicts another element, but at a distance. The present work reviews behavioral studies of nonadjacent dependency learning in artificial grammars, focusing on a few critical studies that highlight important properties of language. In this work, we explain behavioral performance in those studies within a single theoretical framework, the connectionist simple recurrent network (SRN). Five simulations were performed, showing that the SRN captures the qualitative patterns of human performance in these learning tasks. Further, examination of why the SRN succeeds provides additional insights into how people may learn language. One important conclusion is that the SRN, like people, performs dramatically better in situations where the learning situation is made more realistic, incorporating cues like meaning and variability that children have when learning natural languages. |
