| |
Abstract:
A model of sentence processing needs an account of the
processes involved in changing from one analysis of an ambiguity to
another. This abstract describes two self-paced reading experiments
which investigate these processes, contrasting with most other
recent work in this area, which examines the difficulty of
different types of reanalysis (Pritchett, 1992, Sturt et al 1999).
Consider (1a):
1a. The movie man who discovered the actress was always
injecting herself/himself with drugs from Afghanistan......
At `was', the attachment of `the actress' as the object of
`discovered' can be reanalysed (low attachment, consistent with the
`herself' continuation), or alternatively the initial attachment
can be maintained, with `was' being attached as the main verb (high
attachment, consistent with the `himself' continuation). Previous
work (Schneider & Phillips; Sturt et al, CUNY 1999) has shown
that the processor prefers high attachment, which does not involve
reanalysis.
We examined two alternative explanations of this preference. The
preference could be the result of the processor avoiding
reanalysis, and following a structure preservation principle (e.g.
`Reanalysis as Last Resort', Fodor and Frazier, 1978).
Alternatively, a length-based account could explain these results
if the processor assumes that the main clause subject is as short
as possible, and therefore avoids the long and ungainly main clause
subject produced by low attachment (cf. Hawkins, 1991, 1995).
In experiment 1, we crossed high vs. low disambiguation
(himself/herself) with length of the main clause subject (presence
vs. absence of the string bracketed in 1b):
1b. The movie man who discovered the actress [of the historical
drama] was always injecting himself/herself with drugs from
Afghanistan......
If reluctance to attach low increases with length, the high
attachment preference should be increased in the long conditions.
However, length had no effect on the size of the high attachment
preference. Rather, it influenced the time course of the effect.
While the short conditions showed an immediate effect at
disambiguation (himself < herself), the long conditions only
exhibited this effect in the following region (`with drugs').
Additionally, regions following `was...' showed a reading time
advantage for the long conditions, contra length-based theories of
integration difficulty (cf. Gibson, 1998). Hence the high
attachment preference appears to be due to structure preservation
rather than length.
In experiment 2, we tested modified sentences like (1c), where
structure-preserving theories make a contrasting prediction:
1c. The tourists noticed the movie man who discovered the
actress was always injecting herself/himself with drugs from
Afghanistan.
Here, both high (the movie man was...) and low (the actress
was...) attachments of `was' involve reanalysis. As reanalysis is
unavoidable, structure preserving principles do not apply. Instead,
Sturt and Crocker (1996) claim that the choice of reanalysis should
follow recency, a standard attachment principle. This prediction
was confirmed; in the experiment (which also included unambiguous
controls using `that' complementizers), we found a low attachment
preference at the post- disambiguating region (`with drugs'),
contrasting with the high attachment preference found in (1b). This
contrast was confirmed by a between-experiments ANOVA comparing the
relevant conditions of (1b) and (1c), which established a reliable
Experiment * Attachment interaction. To conclude, the parser seems
to preserve structure when it can, and otherwise follows standard
ambiguity resolution principles, like recency, when reanalysis is
unavoidable.
|
