Psychological Review Copyright 1996 by the American Psychological Association, Inc. 1996, Vol. 103, No. 1, 56-115 0033-295X/96[$3.00 Understanding Normal and Impaired Word Reading: Computational Principles in Quasi-Regular Domains David C. Plaut and James L. McClelland Carnegie Mellon University and the Center for the Neural Basis of Cognition Mark S. Seidenberg University of Southern California Karalyn Patterson Medical Research Council Applied Psychology Unit A connectionist approach to processing in quasi-regular domains, as exemplified by English word reading, is developed. Networks using appropriately structured orthographic and phonological rep- resentations were trained to read both regular and exception words, and yet were also able to read pronounceable nonwords as well as skilled readers. A mathematical analysis of a simplified system clarifies the close relationship of word frequency and spelling-sound consistency in influencing nam- ing latencies. These insights were verified in subsequent simulations, including an attractor network that accounted for latency data directly in its time to settle on a response. Further analyses of the ability of networks to reproduce data on acquired surface dyslexia support a view of the reading system that incorporates a graded division of labor between semantic and phonological processes, and contrasts in important ways with the standard dual-route account. Many aspects of language can be characterized as quasi-reg- ular--the relationship between inputs and outputs is systematic but admits many exceptions. One such task is the mapping be- tween the written and spoken forms of English words. Most words are regular (e,g., GAVE, MINT ) in that their pronuncia- tions adhere to standard spelling-sound correspondences. There are, however, many irregular or exception words (e.g., HAVE, PINT ) whose pronunciations violate the standard corre- spondences. To make matters worse, some spelling patterns have a range of pronunciations with none clearly predominating (e.g., _OWN in DOWN, TOWN, BROWN, CROWN vS. KNOWN, SHOWN, GROWN, THROWN, or _OUGH in COUGH, ROUGH, BOUGH, THOUGH, THROUGH). Nonetheless, in the face of this complexity, skilled readers pronounce written words quickly David C. Plaut and James L. McClelland, Department of Psychology, Carnegie Mellon University, and the Center for the Neural Basis of Cog- nition; Mark S. Seidenberg, Neuroscience Program, University of Southern California; Karalyn Patterson, Medical Research Council Ap- plied Psychology Unit, Cambridge, England. This research was supported financially by National Institute of Men- tal Health Grants MH47566, MH01188, and MH00385, National In- stitute on Aging Grant Agl0109, National Science Foundation Grant ASC-9109215, and McDonnell-Pew Program in Cognitive Neurosci- ence Grant T89-01245-016. We thank Marlene Behrmann, Derek Besner, Max Coltheart, Joe Devlin, Geoff Hinton, and Eamon Strain for helpful discussions and comments. We also acknowledge Derek Besner, Max Coltheart, and Mi- chael McCloskey for directing attention to many of the issues addressed in this article. Correspondence concerning this article should be addressed to David C. Plaut, Department of Psychology, Carnegie Mellon University, Pitts- burgh, Pennsylvania 15213-3890. Electronic mail may be sent via In- ternet to [email protected]. 56 and accurately and can also use their knowledge of spelling- sound correspondences to read pronounceable nonwords (e.g., MAVE, RINT ). An important debate within cognitive psychology is how best to characterize knowledge and processing in quasi-regular do- mains in order to account for human language performance. One view (e.g., Pinker, 1984, 1991) is that the systematic as- pects of language are represented and processed in the form of an explicit set of rules. A rule-based approach has considerable intuitive appeal because much of human language behavior can be characterized at a broad scale in terms of rules. It also pro- vides a straightforward account of how language knowledge can be applied productively to novel items (Fodor & Pylyshyn, 1988). However, as illustrated above, most domains are only partially systematic; accordingly, a separate mechanism is re- quired to handle the exceptions. This distinction between a rule-based mechanism and an exception mechanism, each op- erating according to fundamentally different principles, forms the central tenet of so-called "dual-route" theories of language. An alternative view comes out of research on connectionist or parallel distributed processing networks, in which computation takes the form of cooperative and competitive interactions among large numbers of simple, neuron-like processing units (McClelland, Rumelhart, & the PDP Research Group, 1986; Rumelhart, McClelland, & the PDP Research Group, 1986). Such systems learn by adjusting weights on connections be- tween units in a way that is sensitive to how the statistical struc- ture of the environment influences the behavior of the network. As a result, there is no sharp dichotomy between the items that obey the rules and the items that do not. Rather, all items coex- ist within a single system whose representations and processing reflect the relative degree of consistency in the mappings for different items. The connectionist approach is particularly ap-UNDERSTANDING NORMAL AND IMPAIRED WORD READING 57 propriate for capturing the rapid, on-line nature of language use as well as for specifying how such processes might be learned and implemented in the brain (although still at a somewhat ab- stract level; see Sejnowski, Koch, & Churchland, 1989, for discussion). Perhaps more fundamentally, connectionist mod- eling provides a rich set of general computational principles that can lead to new and useful ways of thinking about human performance in quasi-regular domains. Much of the initial debate between these two views of the lan- guage system focused on the relatively constrained domain of En- glish inflectional morphology--specifically, forming the past tense of verbs. Past-tense formation is a rather simple quasi-regular task: There is a single regular "rule" (add -ed; e.g., WALK ~ "walked") and only about 100 exceptions, grouped into several dusters of similar items that undergo a similar change (e.g., SrNG ~ "sang~" DRINK ~, "drank") along with a very small number of very high- frequency, arbitrary