A major distinction among models of the recognition of spoken words concerns whether or not a model allows feedback from logically later to logically earlier levels of processing. One of the leading current models is TRACE (McClelland & Elman, 1986). TRACE is an interactive model which allows feedback from words to phonemic representations. Other models (e.g. Shortlist; Norris, 1994) do not allow such feedback.
A challenge to TRACE was posed by Marslen-Wilson and Warren (1994), who conducted experiments in which listeners had to perform phonetic decisions or lexical decisions on words and nonwords, some cross-spliced so that they contained acoustic-phonetic mismatches. Compared to unspliced words, both words cross-spliced with words and words cross-spliced with nonwords produced equivalently slower decisions. However, performance on nonwords was poorer for nonwords cross-spliced with words than for nonwords cross-spliced with nonwords. Marslen-Wilson and Warren were unable to simulate their findings with the standard version of TRACE. They used this failure as a reason to reject the TRACE model, and they attributed the problems in particular to the inter-word competition process in the model and to TRACE's use of phonemic representations.
These cannot be the reasons why TRACE could not simulate Marslen-Wilson and Warren's data, however. Norris, McQueen and Cutler (in press) have shown that the Merge model, a model of phonemic decision-making which is integrated with Shortlist, can simulate the same data. Merge has both phonemic representations and inter-word competition.
Merge and Shortlist differ from TRACE in a number of ways. TRACE incorporates feedback between processing levels, Merge and Shortlist do not. Furthermore, in Merge/Shortlist, in contrast to TRACE, the inter-word competition process produces a continuously optimal lexical parse of the input. We therefore decided to explore the reasons why TRACE could not simulate the Marslen-Wilson and Warren data. Using a small-scale version of TRACE we systematically altered features of the model and compared the performance of each version with the successful Merge simulation.
In line with Marslen-Wilson and Warren, the small-scale version, unaltered, failed to simulate the pattern of findings correctly. It wrongly predicted a difference between the two types of cross-spliced word, and it grossly exaggerated the difference between the two types of cross-spliced nonword. However an adapted version of the model did produce an acceptable simulation. In effect, the more we made the adapted model resemble Merge, the better it performed. However, even the very best version never produced as close a simulation of the data as Merge did.
The best-performing TRACE version differed from Merge principally by having feedback. Thus the presence of feedback could not be the reason for the failure of the simulation by the standard version of TRACE. Instead, the crucial feature which the TRACE adaptation required was the addition of the Shortlist optimisation procedure. There is no reason to prefer the model with feedback given that the model without feedback performs better, and given also that there are other data which crucially challenge the feedback assumption (Pitt & McQueen, 1998).