Daniel Grodner,1 Edward Gibson1 & Vered Argaman2
1 Massachusetts Institute of Technology, 2 Northeastern University
Structural reanalysis is generally assumed to be representation- preserving, whereby the initial analysis is manipulated to arrive at a new structure. This paper argues that the theoretical and empirical basis for such approaches is weak. A conceptually simpler alternative is that the processor reprocesses (some portion of) the input using just those structure-building operations available in first pass parsing.
First, reprocessing is a necessary component of any realistic model. Intuition suggests that extreme garden paths require reprocessing (1). Further, representation-preserving mechanisms are not useful when the candidate readings of an ambiguity have little or no structural or semantic overlap, as in syntactic category ambiguities (2).
Second, repair demands more powerful structure revision operations than is warranted by the abilities of the first-pass parser. To be sure, structural revision is needed to process unambiguous left-recursive structures ((3a) --> (3b)) (Clifton et al., 1991). However, this is a restricted operation that does not alter dependency relations between lexical items. All second-pass parsing operations necessarily revise such dependencies.
Third, the evidence for repair is explained by independently motivated factors. Sturt et. al. (1999) claim that disparate repairs explain why an NP initially misanalyzed as a direct object of a transitive verb is more difficult to reanalyze as the subject of a matrix clause (4b) than as the subject of a sentential complement (SC) to the verb (4a). However, the effect arises straightforwardly if locality differences affect ambiguity resolution (as in e.g., Altmann et al., 1998). In the subordinate SC reading of (4a), the NP is the subject of a clause that is linked with the immediately preceding verb ("understood"). In the subordinate intransitive reading of (4b) the new clause must be linked with the clausal connective ("because") --- a non-local integration that spans an entire clause. The favored transitive reading in each (4a) and (4b) requires only a local integration to the preceding verb. On a race-based serial or constrained ranked parallel model, there should therefore be more difficulty recovering the intransitive analysis of (4b) than the SC reading of (4a). Hence, the repair-based explanation is superfluous.
We will also present results from a self-paced reading experiment designed to tease apart the two explanations. We manipulated the presence of a modifier on the initial subject in sentences like (4a) and (4b). The modifier increases the integration distance between the ambiguous NP and the connective in the intransitive reading of (5b), but it does not affect integration distance in (5a). The locality-based account therefore predicts that the misanalysis effect will be larger for modified versions of the intransitive (5b) than for unmodified versions (4b), but modification should not affect misanalysis difficulty in (5a). The repair account does not differentiate between the modified and unmodified conditions.
Examples
(1) The horse raced past the barn fell.
(2) The desert trains (people to be tough) / (are tough on people).
(3) a. [IP [NP John's] ...]
b. [IP [NP [NP John's] neighbor's] ...]
(4) a. NP/SC: The employees understood the contract would be changed very soon.
b. NP/Intrans: Because the employees negotiated the contract would be changed very soon.
(5) a. NP/SC: The employees (who went on strike) understood (that) the contract would be changed very soon.
b. NP/Intrans: Because the employees (who went on strike) negotiated (,) the contract would be changed very soon.