Postscript: Plaut and Booth's (2006) New Simulations--What Have We Learned?

Abstract

Plaut and Booth’s (2006) first simulation shows that there is essentially perfect discrimination between word and nonwords sharing the same orthographic structure when the simulation is carried out in the way we suggested. We are therefore satisfied that their model can (now) be characterized as a parallel distributed processing (PDP) model of lexical processing (with the caveat that although it is beyond the scope of the present work, it is important to show at some point that the model “scales up” when it has a more realistically sized vocabulary of 40,000 words or so). We voiced concern that their model makes lexical decisions at the semantic level but there are patients with severe damage to semantics who are nonetheless as accurate at lexical decision as control patients without such damage (Coltheart, 2004). Can a model that makes lexical decisions at the semantic level simulate these data? In reply, Plaut and Booth (2006) reported new simulations showing that increasingly severe “lesioning” of their model at the semantic level impairs lexical decision accuracy in a monotonic way, but it is a remarkably small effect (see their Figure 1). They argue that “distinguishing the semantic activation of one word from that of another requires far more detailed information— and, thus, is less robust to damage—than distinguishing either from the much weaker activation produced by a nonword” (Plaut & Booth, 2006, p. 198). The results of these simulations not withstanding, a substantive difficulty for their account of lexical decision remains. Blazely, Coltheart, and Casey (in press) reported a detailed analysis of two patients (EM and PC) with semantic dementia, both of whom had significant impairments of semantic memory. EM performs slightly worse than PC on semantic tasks, but her visual lexical decision performance (two-alternative forced choice) was virtually perfect (97% correct), whereas PC’s visual lexical decision performance was significantly impaired (75% correct). It is difficult to see how this pattern can be simulated by Plaut and Booth’s (2000) model if lexical decision is carried out at the semantic level, but it is easy to understand if the decisions are carried out at the lexical level and PC’s lexical processing abilities are impaired. Blazely et al. provide converging evidence in support of this conclusion. Our third major concern was that Plaut and Booth’s (2000) PDP model’s use of the sigmoid function relating activation to reaction time (RT) would not permit the joint effect of two factors on RT to be additive when one of these factors and a third factor produce an interaction that lies within the same range of RTs. They responded to this by reporting a new simulation that purports to accomplish this. These data are reported in their Table 1. Some of these data are reproduced in Figure 1 for illustrative purposes. One point here is that this simulation is an unusual way of looking at the joint effects of word frequency and stimulus quality because it entails collapsing across the significant Context Stimulus Quality interaction in the right-hand panel. The joint effects of stimulus quality and word frequency should be examined when there is a neutral prime rather than a related or unrelated word, as argued by Neely (1991; see also Borowsky & Besner, 1993, who used nonword primes, and see Plaut & Booth, 2000, who also used a nonword baseline in their experiments and simulations but have now dropped it in their 2006 simulation). The more important question is whether the model has produced additive effects of stimulus quality and word frequency (as claimed for the data in the left hand panel). The size of the nonsignificant (underadditive) interaction in the left panel (.011 units) is almost double the size of the significant interaction in the right-hand panel (.006 units). Plaut and Booth (2006) do not comment on this. Even if Plaut and Booth were able to provide a more convincing simulation of both additive and interactive effects within the same range of RTs, we would like to know what principled reason makes it possible for the simulation to produce these results given that they now agree that the sigmoid-based explanation “only approximates the actual behavior of the model” (Plaut & Booth, 2006, p. 199). We take the view that Plaut and Booth’s (2006) new simulation work settles little beyond the fact that their model can discriminate 450 500 550 600 650 700