NETtalk is a connectionist network model that learns to convert English text into phonemes. While the network performs the task with considerable accuracy and can generalize to novel texts, httle has been known about what regularities the network discovers about English pronunciation. In this paper, the structure of the internal representation learned by NETtalk is analyzed using two varieties of multivariate analysis, hierarchical clustering and factor analysis. These procedures reveal a great deal of internal structure in the pattern of hidden unit^w;tivations. The major distinction revealed by this analysis of hidden units is vowel/consonant. A great deal of substructure is also apparent.For vowels, the network appears to construct an articulatory model of vowel height and place of articulation even though no articulatory features were used in the encoding of the phonemes. This interpretation is corroborated by an analysis of the errors or confusions produced by the network; The network makes substitution errors that reflect these posited vowel articulatory features. These observations subsequently led to the discovery that articulatory features of place of articulation and, to some extent, vowel height, are largely present in first-order correspondences between vowel phonemes are their spellings. This work demonstrates how the study of language may be profitably augmented by models provided by connectionist networks.

NETtalk is a massively-parallel network that leams to convert English text to phonemes. In NETtalk, the memory representations are shared among many processing units, and these representations are learned by practice. In humans, distributed practice is more effective for longterm retention than massed practice, and we wondered whether learning in NETtalk had similar properties. NETtalk was tested on cued paired-associate recall using nonwords as stimuli. Retention of these target items was measured as a function of spacing, or the number of interspersed items between successive repetitions of the target. A significant advantage for spaced or distributed items was found for spacings of up to forty intervening items when tested at a retention interval of 64 items. Conversely, a significant advantage for massed items was found if testing immediately followed study. These results are strikingly similar to the results of many experiments using human subjects and suggest an explanation based on distributed representations in massively-parallel network architectures.