Publications

Bonatti et al. (Psych Sci, 2007) showed that adult learners picked up on statistical properites of consonants but not vowels in an artificial language learning study. They claimed that this finding could not be explained by a general statistical learning mechanism. We showed that the asymmetry reflects properties of French, the native language of the subjects. (Their response was to abandon the original claim--that the effects were not due to any statistical property of French--in favor of the idea that UG is responsible for the observed statistical asymmetry.)

This article reports research (dating from the mid-90s, actually) addressing whether there is a distinct, structurally defined level of morphological structure (a la Marslen-Wilson, Tyler, Waksler, & Older, Psych Rev, 1994) or whether morphological units arise from correlations among semantics, phonology, and (in literate people) orthography.

Think the DRC model does better at reading nonwords than connectionist models? Read this paper and decide.

Much of the debate about the proper treatment of inflectional morphology has focused on English, which has a dreadfully impoverished inflectional system. More complex system, such as the one for Serbian, which encodes number, gender, and case, are difficult to even describe in rules. We present a connectionist model in which acquiring this system is treated as a statistical learning problem. Gender, in this system, is a graded property of words, which is acquired through exposure to a large vocabulary. This contrasts with other theories which employ explicit gender nodes.

fMRI evidence bearing on theories of the past tense. "Irregular" past tenses such as SLEPT actually pattern with "regular" past tenses such as STEPPED. Hence the brain does not organize this knowledge in terms of rule-governed forms and exceptions.

A brief nontechnical overview of our reading models. Includes "The First Law of Modeling".

A brief paper reporting an elegant experiment that was part of Anne Sperling's USC thesis. Strong evidence against the magnocellular deficit account of dyslexia; strong evidence that dyslexics are impaired in suppressing perceptual noise, in both magno and parvo channels.

A major extension of the "triangle" model to the computation of meaning; addresses controversy about role of direct vs. phonologically-mediated mechanisms in reading.

The fact that an eater of rats is a rat-eater, not a *rats-eater has been repeatedly cited as evidence for the level-ordering theory of the lexicon and the Pinker theory of the past tense. We show that the facts about rat-eating are not as commonly portrayed and that constraints on plurals as modifiers in compounds are graded in nature, arising from phonological and semantic factors.

The Seidenberg, MacDonald, and Saffran (2002) article elicited a response from Marcus and Berendt, included here along with our response.

This article presents computational modeling evidence linking a phonological deficit to impaired processing of an aspect of grammar, pronominal anaphora. The results provide a causal demonstration of how a perceptual deficit that affects phonological representation could have secondary effects on grammatical processing. This is consistent with the conclusion that "grammatical" deficits in SLI do not reflect anomalous development within grammatical modules but rather are sequelae of more basic information processing deficits.

Evidence that impairments in generating regular past tenses such as WAITED derive from phonological deficits, not an impairment in the "rule" module.

Using a computational model of reading to understand why some types of reading remediation work and others don't.

Can neuroimaging provide decisive evidence concerning competing theories of past tense morphology? And what about German, anyway?

A "perspective" on an article by Pena et al. claiming to have demonstrated distinct statistical and grammatical mechanisms in learning language. Summary: not so fast.

An accessible discussion of the controversies over how reading should be taught.

Provides a new analysis of age of acquisition effects, in the context of a more general theory of how experience limits plasticity.

American Psychological Society. A review of psychological research bearing on how reading should be taught; illustrates the deeply unfortunate disconnect between reading research and educational practice that has occurred over the past 25 years.

Modeling and behavioral research that contradict the sometime claim that acquired phonological dyslexia also involves an orthographic deficit. Our models show how the critical data could arise with intact orthographic processing and the behavioral data show the same effects in normals.

A theoretical discussion piece in which we propose that derivational morphology can be treated as a statistical phenomenon resulting from the convergence of orthographic, phonological, semantic, and other codes across words, rather than a discrete level of representation.

Because their little networks tell them so.

This is the preprint version of the article that appeared in the journal. It provides evidence for a speech perception deficit in children whose dyslexia is secondary to a more general language impairment ("SLI") but not in dyslexics whose spoken language is apparently normal. Thus the "phonological" deficit apparent in many dyslexics may arise from something other than a sensory deficit.

A spelling-sound model applied to issues concerning learning to read and dyslexia. Some cool simulations of the effects of phonological impairment on reading and illustrations of attractor dynamics.

This review article discusses two competing hypotheses concerning developmental language impairments (SLI): one holds that SLI is caused by impairments in components of UG; the other that SLI is secondary to information processing deficits which affect the acquisition of several aspects of language. In this article we discuss how a phonological deficit could account for the impairments in both inflectional morphology and syntax (e.g., the resolution of pronouns) that are typically observed in such children. This account also correctly predicts that children with SLI will tend to be phonological dyslexics. The Joanisse and Seidenberg (2003) Brain and Language article presents some relevant modeling results and Joanisse et al. (2000) provide relevant behavioral data.

This article describes an initial attempt to model the word segmentation problem using a conjunction of probabilistic cues. The model is a demonstration of an approach more than a realistic solution to the problem: to keep the simulations manageable, it was given capacities that infants lack (e.g., knowledge of segmental phonology). Also, the trained model correctly identified only about 45% of the words in the speech stream--but then again how accurate is your average 4 month old?

A short overview of the bigger picture.