Given ambiguous representations, the MSC hypothesis predicts that the processor will commit to one interpretation
In an experiment designed to evaluate these hypotheses with respect to the representation of distributivity, participants’ eye movements were recorded as they read sentences containing distributive or collective predicates that were either disambiguated by a preceding adverb or left locally ambiguous by delaying the disambiguating adverb until the end of the predicate
The results suggested that a semantic commitment is made in locally indeterminate cases as evidenced by a significant interaction of ambiguity and distributivity in first pass times, total times, and regressions
Hence we argue that the distributive/collective distinction is treated as a matter of ambiguity rather than as one of vagueness
In the absence of evidence for a distributive reading, the processor commits itself to a collective reading sometime during the processing of the predicate (before the disambiguation in our late disambiguation examples)
By the MSC hypothesis, this will predict that a premature decision, one made early on the basis of little information, should not occur during immediate processing
The point is that no commitment will be made in the absence of supporting evidence
By contrast, given the MSC hypothesis, the notion that a string is grammatically ambiguous predicts that the processor encounters a choice point on initial analysis, adopting one representation rather than another
In our experiment we only examined collective/distributive subjects in single clause sentences
The ambiguity hypothesis predicts an interaction
the vagueness hypothesis doesn’t
T
In order to test for possible pragmatic biases in the predicate, participants were asked to rate the ‘naturalness’ of the locally ambiguous collective or distributive predicates contained in the experimental sentences
when the actual conjoined NP subject in the experimental materials was replaced by a pronominal subject
Thirty-two students native English speakers
Each version of the questionnaire contained eight sentences in which the locally ambiguous predicate was subsequently disambiguated towards a distributive interpretation using each and eight sentences in which the predicate was subsequently disambiguated towards a collective interpretation using together
7-point scale
They were told that there were no right or wrong answers and that we were only interested in their opinions of the naturalness of the sentences
another rating study was conducted using truncated versions of the locally ambiguous items such that each item ended after the word each or together
examined first pass reading times, total reading times, and the pattern of regressions for different regions of the target sentences
Sixty undergraduate students bite bar
Experiment
dealt with where readers look during reading and that they should read each sentence for comprehension
Sixteen experimental sentences were embedded in 107 filler sentences
Each sentence appeared in one of four versions, as in (7) above: in two versions (a and c), the predicate received a distributive interpretation, and in the other two (b and d) the predicate received a collective interpretation
region
he first region consisted of the words preceding the predicate (a conjoined NP)
the second region consisted of the predicate itself
the third region comprised the next three words, or next two words if the third word was the last word in the sentence
and the fourth region was the remainder of the sentence
The fourth region was created solely so that results from the third region would be free of sentence wrapup effects
In the first region, none of the effects were significant
In the second region, there was an effect of ambiguity, wherein the locally ambiguous versions were read faster than unambiguous versions, possibly due to a preference for adverbs to appear in verb-phrase final position
he raw reading times analyses and analyses of residuals were consistent in
indicating that the interaction did not approach significance
In the third region, the different versions of any given sentence were once again identical
Here, locally ambiguous versions were read somewhat slower overall than unambiguous versions
However, this difference failed to approach significance in the subjects analysis (F1(1,59) = 1
91, P � 0
and was only marginally significant by items
Crucially, each of these marginally significant main effects was qualified by a highly robust interaction between ambiguity and predicate type, suggesting that distributive predicates were read more slowly than collective predicates in the locally ambiguous versions
Pairwise comparisons confirmed that while distributives were read more slowly in the ambiguous versions (F1(1,59) = 10
84, P � 0
01
F2(1,14) =20
48, P � 0
001), there was no reliable predicate effect in the unambiguous versions
The results of this analysis for first pass times revealed striking differences between ambiguous and unambiguous forms
In the residuals analyses, distributives (each) were read slower than collectives (together) in both ambiguous and unambiguous forms
However, the effect was much larger in ambiguous forms than unambiguous forms (223 ms vs
43 ms), and while in ambiguous forms the effect was significan
Distributives were significantly faster than collectives in unambiguous forms (means: 844 ms vs
925 ms
F1(1,59) = 4
72, P � 0
05, F2 (1,14) = 10
39, P � 0
01), but significantly slower than collectives in ambiguous forms
n the third region, where the different versions were again identical, there was a main effect of ambiguity, with ambiguous versions being read slower overall than unambiguous versions
There was also a robust predicate effect (F1(1,59) = 9
53, P � 0
01
F2(1,14) = 10
88, P � 0
01), which indicated that distributives were read more slowly overall
In the second region, there were no significant differences across conditions in the percentages of trials on which regressions occurred out of the region (Fs � 1)
In the third region, the mean percentage of trials on which regressions occurred out of the region was 19%, 8%, 4% and 4% for the ambiguous-distributive, ambiguouscollective, unambiguous-distributive, and unambiguous-collective conditions, respectively
There were significantly more trials involving regressions out of the third region in the ambiguous than the unambiguous versions
The predictions of the vagueness hypothesis were clearly disconfirmed
Given the MSC hypothesis, the vagueness hypothesis predicts no interaction between ambiguity and sentence form: in both the ambiguous distributive (7a) and the unambiguous distributive (7c) the processor should postulate a distributive operator when each is encountered
ounter to this prediction, the ambiguous distributive form was substantially more difficult to understand than the other sentence forms
This may be seen in the significant interaction of ambiguity and sentence form in first pass and total times in region three, as well as in the regressions out of region three and regressions into region two
Readers clearly exhibited a preference for the collective reading of the ambiguous portion of the sentences in our experiment
These results make it difficult to maintain the assumptions needed to salvage the vagueness hypothesis
Instead, given the MSC hypothesis, they support the assumption that the correct grammatical account of collective/distributive differences treats the distinction as one of ambiguity rather than as one of vagueness, at least in cases like those tested, where subject-predicate relations are involved
We turn now to alternative interpretations of our results
The question is whether the results can be attributed directly to the complexity of the distributive reading
We think not
It may be true that distributive readings, even unambiguous ones, are slightly more complex than collective readings
This suggests that readers may not simply add information to the current representation of these locally ambiguous forms when each is encountered
Similarly the results are difficult to reconcile with a parallel processing view unless the processor has computed both a collective and a distributive representation and then abandoned the distributive representation before each is encountered