| |
Abstract:
Recent controversy has surrounded the domain specificity of
resources available to sentence processing [e.g. Just, et al.
(1992), Waters & Caplan (1996)]. In a set of four experiments,
we examine the effect of different kinds of increased processing
load on the ability to use a semantic cue for disambiguation. When
well-known ambiguities are embedded in such a manner that syntactic
processing load is significantly increased, semantic cues to
disambiguation cannot be used; when almost identical embedding
contrasts are used which do not significantly increase the
syntactic processing load, semantic cues can still be used. This is
compatible with a competition-based theory of resource
availability, where semantic and syntactic processes vie for a
common pool of processing resources.
The Main Verb/Reduced Relative (MV/RR) ambiguity has been shown
to be sensitive to the animacy of the noun which introduces the
ambiguity. Examples with an animate subject NP are disambiguated at
the post-verbal PP, but in examples with an inanimate subject NP
semantic information can be used to disambiguate already at the
verb (Trueswell et al. 1994).
1) The defendant/evidence (that was) examined by the lawyer was
unreliable. [simple conditions, all expts]
In each of our self-paced reading experiments, subjects read
simple MV/RR ambiguities as well as ambiguities which were embedded
in a larger context (2). In Experiments 1-3, the MV/RR ambiguity
was center-embedded such that it interrupted an incomplete
Subject-Verb dependency (3); in Experiment 4, the ambiguity was
embedded as part of a simple right-branching structure (4),
although the embedding context contained an identical number of
arguments/predicates to the earlier experiments. In all examples
the embedding does not change the nature of the ambiguity. While
they seem difficult, all embedded sentences are within
experimentally defined limits of comprehensibility [see Gibson
(1998)], and comprehension questions were answered with above 81%
accuracy.
2) The judge remembered that the document... [embedded
conditions, all expts.]
3) ...stating that the [ defendant/evidence (that was) examined by
the lawyer was unreliable ] had been stolen from the filing
cabinet. [embedded conditions, expts 1-3]
4) ...had stated that the [ defendant/evidence (that was) examined
by the lawyer was unreliable] and should be withdrawn from the
testimony. [embedded conditions, expt 4]
Each experiment contained 48 MV/RR ambiguities and 152 fillers,
presented in two blocks, with at least 50 subjects per experiment.
Our analyses focus on subjects classified as high-span (upper 50%
of subjects), since only this group consistently showed immediate
use of the semantic cues in the syntactically simple conditions, a
key premise for our manipulation (cf. Just & Carpenter 1992).
Reading times at by-phrase show the expected animacy x ambiguity
interaction in syntactically simple conditions in all experiments.
However, in the embedded conditions in which the ambiguity
interrupts an incomplete subject-verb dependency (experiments 1-3),
we find consistent failure to exploit the semantic cue, as
indicated by a main effect of ambiguity (F1(1,28)=12.78,
p<0.001; F2(1,47)=9.79, p<0.005). In the right-branching
embedded conditions (experiment 4), we observe the same animacy x
ambiguity interaction found in the syntactically simple conditions
(F1(1,22)=6.19, p<0.02; F2(1,47)=4.76, p<0.05).
Taken together, these results indicate that a specifically
syntactic load prevents the parser from exploiting semantic cues to
disambiguation. If load blocks *detection* of semantically
anomalous main verb reading, then it suggests syntactic and
semantic processing compete for a common resource-pool;
alternatively, if load blocks ability to exploit semantic anomaly
to search for more plausible reduced-relative parse, then results
support a more modular approach to resource-allocation.
|
