Working-memory capacity and island effects: A reminder of the issues and the facts

1. Understanding the Terms of the Debate.

In their response to our article (Sprouse et al. 2012), Hofmeister, Staum Casasanto, and Sag (2012, henceforth HSS) level three primary criticisms at our studies: (i) that working-memory (WM) correlations are not predicted by reductionist theories, (ii) that the WM tasks that we chose are inappropriate for showing the predicted correlation, and (iii) that our results do, in fact, show the predicted correlation. We note that these three criticisms appear to be contradictory—if any one were true, the other two would very likely be false—but here we focus on an attempt to clarify why all of them are unjustified: (i) the WM correlation is exactly what a viable account of the island effects predicts; (ii) the diversity of noncorrelations that have now been documented forces the reductionist account to abandon independently motivated memory mechanisms in favor of currently undefined 'mystery' mechanisms; (iii) HSS's characterization of our results is selective and misleading, and it inaccurately describes the predictions of the reductionist account of island effects. But before addressing these specific arguments it is worthwhile to briefly remind the reader of the central question of this debate. This is already discussed at length in our original article, but a number of the issues raised by HSS appear to be orthogonal to the central question.

Island effects are the sharp declines in acceptability that are found when WH dependencies and similar linguistic dependencies span certain structural domains, such as relative clauses, interrogative clauses, complex subjects, adjoined phrases, and many others. These structural domains are commonly known as syntactic islands. The empirical fact to be explained is uncontroversial: extraction from island structures is much less acceptable than would be predicted by the simple summation of the acceptability cost of (i) WH-extraction and (ii) the presence of island structures. We refer to this as the SUPERADDITIVE property of island effects, and the focus of the debate is on the source of this property. A long-standing view is that the superadditivity reflects a linguistic constraint that specifically targets the combination of extraction and islands (i.e. island violations). A huge literature has arisen from this starting point. An equally long-standing competing view is that island effects reflect limits on language processing resources. According to this REDUCTIONIST view, island effects can be explained without appeal to domain-specific linguistic constraints. The superadditive property of island effects cannot, of course, be explained by simply summing the processing cost of extraction and island structures, so the simplest reductionist view is a nonstarter. But a variant of this reductionist hypothesis is more promising. Perhaps the superadditivity is a consequence of resource overload: to the extent that simultaneously processing extraction and an island structure exceeds the available memory resources, the extra acceptability cost (i.e. the superadditivity) might be explained as the cost for exceeding the available memory resources (Kluender & Kutas 1993). This account makes a straightforward prediction: the degree of superadditivity should vary as a function of the WM resources available to an individual. We tested this prediction in two experiments with over three hundred speakers, and the results were not encouraging for the reductionist approach: individual differences in WM capacity accounted for an average of only around 1% of the variance in the size of superadditive effects, and in many cases the correlations were zero or the opposite of what the reductionist approach predicts. Similar conclusions have been reached in studies by Tokimoto (2009) and Michel (2012).

Crucially, the central question of this debate is whether linguistic constraints are needed in order to account for island effects, specifically the superadditive effect of combining extraction and island structures. The reductionist account of a given island constraint succeeds only if it eliminates the need for linguistic constraints entirely. What this means is that demonstrations that island-violating sentences are more or less difficult to understand under different circumstances, such as those in HSS's response and several of their other papers, are irrelevant to the central question. Showing that factor X modulates an...