The relationship between lexical semantic, word-level, and phonological processes in the production of single words Matthew Goldrick There has been considerable debate concerning proposals to increase interaction between processes within the speech production system. The evidence supporting claims of increased interactivity is reviewed to identify (a) points of empirical and theoretical consensus and (b) continued points of debate. The results of the review are then discussed within a proposed architecture for the production system, with the aim of providing a functional motivation for the observed patterns of interaction. Background. Three major production processes are distinguished here: conceptual preparation (activation of the concept to be named); lexical selection (selection of a word to express the concept); and phonological encoding (retrieval of the word’s sounds). Two ways of increasing interactivity beyond that of a discrete system are discussed: cascading activation and feedback. Cascading activation. Unlike a discrete system, where each process yields a single output, cascading activation allows processes to send on information about multiple representations. This is present throughout the production system. Activation cascades from conceptual preparation to lexical selection processes, introducing semantically related words as lexical competitors. This has been shown both in chronometric studies (e.g., Damian & Bowers, 2003) and studies of the errors of aphasic individuals (e.g., Caramazza & Hillis, 1990). Activation also cascades from lexical selection to phonological encoding. Lexical competitors can be phonologically activated, influencing naming times (e.g., Peterson & Savoy, 1998), spontaneously occurring tip-of-the-tongue states (Gollan & Acenas, 2004). and speech errors of individuals with deficits to phonological encoding (e.g., Rapp & Goldrick, 2000). The debate regarding cascading activation concerns the degree to which this activation is limited. Limitations are clearly needed, as many lexical competitors do not substantially active their phonological representations (Levelt et al., 1991). Feedback. Unlike a discrete system, feedback allows “later” process to influence the output of “earlier” ones. The strongest evidence for feedback from phonological encoding to lexical selection is the presence of phonological effects on lexical selection processes. This has been shown in chronometric experiments (e.g., Damian & Martin, 1999), experimentally elicited speech errors (Ferreira & Griffin, 2003), and studies of the semantic errors of individuals with deficits to lexical selection (Rapp & Goldrick, 2000). Two issues remain open with respect to this type of feedback. First, what is its strength? Second, is this feedback mediated directly (via reciprocal connections within the production system) or indirectly (via cascading activation from perceptual processes)? Finally, no current data suggest that later processes influence conceptual preparation. Studies of aphasic speech errors (Rapp & Goldrick, 2000) suggest that there is no substantial phonological activation of competitors within conceptual preparation processes. At issue is whether such feedback is present, but so weak as to be “functionally invisible.” Functional motivation for the pattern of interactivity. Consistent with interactive activation theories of spoken production, the empirical consensus is that there is cascading activation throughout the production system. From this perspective, it is surprising that there are limitations on feedback. This pattern may derive from the mechanisms implementing production processes. Suppose the production system is composed of two attractor network modules (one network associating lexical concepts and semantic features, the other word-level and phonological representations). Following Mathis’ (1998) proposal, assume that each of these networks remains in a neutral (“rest”) state until sufficiently strong input drives it into some other state. Because of this rest state, lexical and phonological representations remain inactive until the semantic network has selected a strong output; this blocks any effect of feedback. In contrast, feedback from phonological to word-level representations is a natural consequence of the structure of attractor networks (which rely on bidirectional activation flow). Differences in feedback may therefore be a reflection of deeper structural divisions within the production system.