Hybrid accounts
An alternative to both the garden path GP model and the constraint-based approaches introduced above is the unrestricted race model (for further discussion see Van Gompel & Pickering, 2007). This model proposes that when an ambiguous sentence is encountered, the various possible analyses are involved in a race, with the winner being the analysis that is built fast est. The speed with which an analysis is constructed depends on the strength of the information supporting it, and so from this point of view it is similar to constraint-based models and unlike the GP model, which assumes a single syntactic analysis is pursued. Like the GP model, however, it assumes that a syntactic analysis is built before it is evaluated on the basis of plausibility. The model also differs from the constraint-based models in that it assumes no competition between the different analyses.
These three model types GP, constraint-based, unrestricted race make different predictions about the set of sentences in (11.64)–(11.66), where the structural ambiguity concerns whether e modifies the first or second full NP (e.g. the bodyguard or the governor in (11.64)). First, the GP model predicts a preferred syntactic analysis of attaching the modifying phrase to the second NP, just as was seen earlier for relative clause attachment ambiguities see p. 192. This would mean that both sentence (11.64) and sentence (11.66) would be straightforward, since this syntactic preference is not contradicted by the subsequent semantic interpretation. In (11.65), however, the attachment of retiring to the province will present an anomaly to the semantic interpreter. Second, constraint-based models predict that since both resolutions of the ambiguity remain possible in (11.64), these will be in competition with one another, making this sentence more difficult to process than (11.65) and (11.66), which are semantically disambiguated only the governor can retire, not the province.
Finally, however, the unrestricted race model predicts that the globally ambiguous sentence in (11.64) will be easier than the sentences in (11.65) and (11.66). In all three cases, the model predicts that each syntactic analysis is developed, but the fastest one will be selected. In (11.64) the processor will sometimes come up with one analysis, and sometimes with the other, but in neither case is there any conflict with a subsequent plausibility analysis as both interpretations are possible. In the case of the disambiguated sentences in (11.65) and (11.66), the syntactic processor similarly sometimes reaches one analysis and sometimes the other, and some of the time this will be contradicted by the plausibility analysis, so that overall (11.65) and (11.66) should throw up more processing problems than (11.64). In a study measuring eye movements during the reading of such sentences, Van Gompel et al. (2005) found precisely this result – (11.64) was easier to read than the others.
