The Architecture Underlying Syntactic Processing

In this thesis, I report five eyetracking experiments that tested current sentence processing theories. So far, most research has attempted to discriminate between various sentence processing theories by investigating whether non-syntactic sources of information can be employed immediately in syntactic ambiguity resolution. Two-stage theories such as the garden-path theory claim that the use of non-syntactic information is delayed, whereas interactive or constraint-based theories claim that all sources of information can be employed immediately. The experiments in this thesis focussed on a different aspect of current sentence processing theories, which has been largely ignored. They investigated whether the architecture of the sentence processor involves reanalysis or competition. Two-stage theories claim that processing difficulty occurs when an initially adopted syntactic analysis has to be revised, whereas most current constraint-based theories stipulate that competition between two or more syntactic analyses that are activated in parallel makes a sentence difficult to process. The experiments in this thesis investigated reanalysis and competition by testing globally ambiguous syntactic structures and contrasting them with structures that are disambiguated (either to one analysis or the other). Constraint-based competition theories predict that competition occurs in globally ambiguous sentences which do not have a bias for one structure over another, because two syntactic analyses are about equally activated. No such competition should occur in disambiguated sentences, because only one analysis is supported by the disambiguating information. In contrast, traditional two-stage theories such as the garden-path theory predict that the processor initially adopts the structurally preferred analysis. When the disambiguation is inconsistent with this analysis, reanalysis should occur. Reanalysis should not occur when a sentence is disambiguated toward its preferred analysis, or when a sentence is globally ambiguous. The eyetracking experiments in this thesis showed that disambiguated sentences are more difficult to read than globally ambiguous sentences. These results are incompatible with competition as a mechanism of syntactic ambiguity resolution, and therefore disconfirm the predictions of most current constraint-based theories. They are also problematic for some two-stage theories, because processing difficulty occurred in sentences where the disambiguation was toward the structurally preferred analysis. In this thesis an alternative model, the unrestricted race model, is proposed, which explains the results in a straightforward manner. The unrestricted race model claims that the alternative analyses of a syntactically ambiguous sentence are engaged in a race. The analysis that is constructed fastest is adopted. The model stipulates that the analysis that receives most support from both syntactic and non-syntactic sources of information usually wins the race. When two analyses are about equally supported, as in balanced ambiguities, each analysis is adopted about half the time. Consequently, when the sentence is disambiguated (toward one analysis or the other), it is inconsistent with the analysis on half the trials, and therefore reanalysis should occur on those trials. Thus, the disambiguated sentences are more difficult than the ambiguous sentences, where reanalysis does not occur. Balanced ambiguities contrast with biased ambiguities, where there is a preference for one analysis. The unrestricted race model predicts that in such ambiguities, the processor adopts the preferred analysis on nearly all trials. Therefore, reanalysis should occur on very few trials when the disambiguation is consistent with this preference