Three eye-tracking experiments investigated two frequency-based processing
accounts: the serial lexical-guidance account, in which people adopt the an
alysis compatible with the most likely subcategorization of a verb; and the
serial-likelihood account, in which people adopt the analysis that they wo
uld regard as the most likely analysis, given the information available at
the point of ambiguity. The results demonstrate that neither of these accou
nts explains readers' performance. Instead people preferred to attach noun
phrases as arguments of verbs even when such analyses were unlikely to be c
orrect. We suggest that these results fit well with a model in which the pr
ocessor initially favors informative analyses. (C) 2000 Academic Press.