Speech Errors and Language Production: Neuropsychological and Connectionist Perspectives

We examined two working hypotheses arising from Garrett's (1975) analysis of speech errors: the independent-levels and the function-content hypotheses. Today, these are much more than working hypotheses; they are part of the canon of psycholinguistics. The functional and positional levels have been spelled out in theories of sentence and word production (Bock, 1982; Dell, 1986; Hartsuiker, 2002; Kempen & Hoenkamp, 1987; Levelt, 1989; Stemberger, 1985), particularly in discrete two-step theories of lexical access (Levelt et al., 1999; Roelofs, 1997). Similarly, the function-content distinction and the related distinction between content items and inflectional affixes are present in modern theories of production and production deficits (e.g., Alario & Caramazza, 2002; Bock & Levelt, 1994; Ferreira & Humphreys, 2001; Segalowitz & Lane, 2000). The thrust of this article, however, is that recent connectionist models of production blur Garrett's sharp distinctions. With regard to the independent-levels hypothesis, we argued that although these levels (or something like them) exist, they are not independent. Lemma selection is influenced by phonological information; such information is normally part of a later step in the process. This influence can be modeled by interactive or feedback connections from phonological to word units. The interactive two-step model, in particular, offers an account of a variety of phenomena that may arise from interaction, while retaining the staged character of lexical access emphasized in discrete approaches. The function-content distinction also becomes blurred through the application of connectionist principles (Gordon & Dell, 2003). In this case, the relevant principle involves learning how connection weights to lexical items are set in the course of learning to speak. The function-content distinction is present in the relative strengths of semantic and syntactic-sequential inputs. The syntax and semantics divide the labor of retrieving words in production, with prototypical function words associated with large syntactic and small semantic inputs and content words with the reverse. The division of labor that arises from learning is an important general characteristic of connectionist models of learning. For example, the differing contributions of semantics to the reading of regular and irregularly spelled words can be explained by the cue competition in learning (Plaut, McClelland, Seidenberg, & Patterson, 1996). In the Plaut et al. reading model, the regular-irregular distinction becomes a continuum as different representations make quantitatively varying contributions to the processing. We are asserting the same for the function-content distinction. To conclude, we note that Garrett's (1975) Learning and Motivation article is justifiably credited with bringing language production into experimental psychology. Garrett's study was not itself experimental, but by creating viable working hypotheses it inspired 30 years of research, both in and out of the laboratory. If production started out as Levelt's (1989) "stepchild of psycholinguistics," it has grown up to make its family proud.