Is human sentence parsing serial or parallel? Evidence from event-related brain potentials

In this ERP study we investigate the processes that occur in syntactically ambiguous German sentences at the point of disambiguation. Whereas most psycholinguistic theories agree on the view that processing difficulties arise when pat-sing preferences are disconfirmed (so-called garden-path effects), important differences exist with respect to theoretical assumptions about the parser's recovery from a misparse. A key distinction can be made between parsers that Compute all alternative syntactic structures in parallel (parallel parsers) and parsers that compute only a single preferred analysis (serial parsers). To distinguish empirically between parallel and serial parsing models, we compare ERP responses to garden-path sentences with ERP responses to truly ungrammatical sentences. Garden-path sentences contain a temporary and ultimately curable ungrammaticality. whereas truly ungrammatical sentences remain so permanently-a difference which gives rise to different predictions in the two classes of parsing architectures. At the disambiguating word, ERPs in both sentence types show negative shifts of similar onset latency, amplitude, and scalp distribution in an initial time window between 300 and 500 ms. In a following time window (500-700 ms), the negative shift to garden-path sentences disappears at right central parietal sites, while it continues in permanently ungrammatical sentences. These data are taken as evidence for a strictly serial parser. The absence of a difference in the early time window indicates that temporary and permanent ungrammaticalities trigger the same kind of parsing responses. Later differences can be related to successful reanalysis in garden-path but not in ungrammatical sentences. (C) 2003 Elsevier Science B.V. All rights reserved.