The evidence consists of eye-tracking data, recorded while the participants read sentences such as (1)-(3). Sentence (1) is Globally Ambiguous (GA) because the relative clause can be plausibly attached both high (to ''bodyguard'') and low (to ''governor''). Variants (2) and (3) are only Locally Ambiguous (LA) as ''retiring'' fits semantically with only one attachment site. The crucial observation was that GA sentences were consistently read faster than their LA counterparts.
The authors interpreted this as evidence against competition-based models of syntactic parsing, which predict the opposite pattern of reading times. In GA sentences, the two attachments match in strength, whereas in the LA variants one of them is much stronger. As balanced competitions take more time to settle down than unbalanced ones, GA sentences will consume more processing time than their LA counterparts.
We argue that this argument only holds in conjunction with a specific assumption concerning the availability of semantic/pragmatic information (determining the plausibility of alternative attachment sites), viz. that all semantic/pragmatic information associated with an input word becomes available at (nearly) the same time as the syntactic information of the word, and immediately affects attachment decisions.
However, such "synchrony" is by no means the only theoretical option. The syntactic parser may select an attachment site prior to (i.e. asynchronously with) the arrival of semantic or pragmatic information, on the basis of recency, at random, or otherwise. (For a recent ERP study indicating that semantic/pragmatic information may lag behind syntactic information, see De Vincenzi, Job et al., 2003.) Late arrival of semantic/pragmatic information may trigger reanalysis, thereby decelerating reading. In the LA (but not GA) sentences, this may happen frequently, which explains the experimental outcomes.
We propose a model in which several parallel processors (syntactic, semantic, pragmatic) work at their own pace and inform each other of their decisions. Such a model may very well contain a competition-based syntactic parser (Stevenson, 1994; Vosse & Kempen, 2000; Tabor & Hutchins, 2004) without contradicting the above findings.
| (1) | The bodyguard of the governor | \ |
| (2) | The governor of the province | | retiring after the troubles is very rich |
| (3) | The province of the governor | / |