In both the triangle and the dual-process models, the output of the phonological assembly network interacts with the phonological code produced by a lexical (or lexical-semantic) pathway. However, it is worth pointing out that the way in which Zorzi et al. (1998a) simulated the contribution of retrieved phonology (i.e. the lexical route) is very similar to the way in which PMSP simulated the contribution of a “semantic” pathway. In both cases, what it really amounts to is a further (frequency-weighted) input to phonology. As argued by Bernstein & Carr (1996), there is nothing compellingly “semantic” in the simulation described by PMSP: rather, these properties can be ascribed to a lexical nonsemantic route. One limitation of the PMSP simulation is the lack of assumptions about the time course of the semantically-driven activation of phonology; that is to say, the input from the two routes to the phonological output begins simultaneously, and this seems rather implausible. Given that the phonological and the semantic pathways are thought to operate in parallel, the phonological pathway should be substantially faster than the semantic pathway, since the latter contains two additional layers of nodes that must be traversed (i.e. the semantic layer and a further layer of hidden units; see Figure 13.1). Full implementations of the triangle model (e.g. Harm & Seidenberg, 2001) are necessary to shed more light on the issue of the “distribution of labour” between the phonological and semantic routes.