One of the classic puzzles of language is posed by the phenomenon of discontinous dependency, in which the form of an element at one point in an utterance depends on the form of a noncontiguous controlling element. How do speakers use the information carried by the controller to implement the correct form of the dependent element? We contrasted two accounts of this process that differ in their assumptions about the organization of language formulation. The serial account, patterned after an augmented-transition-network model of the parsing of discontinuous dependencies, suggests that the controller is held in working memory until the point in the string at which the dependent appears. A second hypothesis, derived from a hierarchical model of language production, predicts that controllers and dependents within the same clause are specified concurrently, even when they are eventually separated in the utterance. Using a procedure to elicit verb-agreement errors in speech, we found that agreement errors were more frequent after phrases than after clauses that separated the verb from its head noun, reversing the direction of a related effect in language comprehension. When length varied, longer phrases led to more errors; longer clauses did not. These results support the hierarchical hypothesis. © 1992.
