|
Abstract:
Current constraint-based models, most explicitly MacDonald,
Pearlmutter, and Seidenberg (1994), propose that syntactic
ambiguity resolution is a form of lexical ambiguity resolution.
We discuss data from two experiments that challenge this
view.
Evidence from lexical ambiguity resolution suggests that
balanced ambiguous words in neutral contexts are more difficult
to read than unambiguous control words (Duffy, Morris, and
Rayner, 1988; Rayner and Duffy, 1986). An explanation of these
results is that equally favored meanings compete for activation.
If syntactic ambiguity resolution is a form of lexical ambiguity
resolution, balanced syntactic ambiguities should behave
similarly. However, our experiments show that ambiguous syntactic
structures are easier to read than disambiguated structures.
In one study, we contrasted (1a-c):
1a. The hunter killed only the poacher with the rifle not long
after sunset. (ambiguous)
1b. The hunter killed only the leopard with the rifle not long
after sunset. (VP attachment)
1c. The hunter killed only the leopard with the scars not long
after sunset. (NP attachment)
A pre-test demonstrated that (1a) was equally plausible on NP
and VP attachment analyses, and that both (1b) and (1c) had only
a single plausible analysis. Another pre-test showed that the
ambiguity in (1) was balanced: Readers chose VP and NP attachment
equally often in (1a). Thus, both the VP and NP attachment
analysis were equally favored in (1a), but not in (1b) and (1c)
where plausibility supports only one analysis.
In contrast to the predictions of lexical models, eye-tracking
data showed that (1a) was reliably easier to process on a variety
of measures than (1b) and (1c), which did not differ. In
contrast, another experiment showed that ambiguous sentences in
biased ambiguities pattern very similarly to sentences that are
disambiguated toward the favored analysis.
We interpret these results in terms of an unrestricted race
account, in which the processor uses any relevant information to
select an initial analysis, but where alternative analyses do not
compete.
|
