Lingua 114 (2004) 327

www.elsevier.com/locate/lingua

Dont break, or do:

prosodic boundary preferences
Lyn Frazier*, Charles Clifton, Jr., Katy Carlson
Department of Linguistics, University of Massachusetts, Amherst, MA 01003, USA
Received 16 May 2002; received in revised form 20 February 2003; accepted 20 February 2003

Abstract
Four naturalness judgment experiments were conducted to test dierent hypotheses about
prosodic phrasing. The hypothesis that syntactic constituents should not be broken into distinct
prosodic phrases [as in Truckenbrodts Wrap constraint (Truckenbrodt, H., 1995. Phonological
Phrases: Their Relation to Syntax, Focus, and Prominence. Unpublished PhD Dissertation,
MIT)] was less predictive of the results of Experiments 1 and 2 than the hypothesis that constituents may be freely divided into prosodic phrases, as long as the resulting phrases are semantically coherent [Selkirk, E., 1984. Phonology and Syntax: The Relation Between Sound and
Structure. MIT Press, Cambridge, MA]. The results of two further experiments conrmed Watson and Gibsons (Watson, D. G., Gibson, E., 2001. Linguistic structure and intonational
phrasing. Paper presented at the Fourteenth Annual CUNY Conference on Human Sentence
Processing, Philadelphia, 1517 March 2001) claim that prosodic breaks are natural before long
upcoming constituents, but did not support their hypothesis that the distance between a new item
and its integration site is what motivates the presence of a prosodic phrase boundary. The results
are interpreted as further evidence that the use of high level breaks in language comprehension
is not governed by an invariant local mapping from syntax or processing considerations to
prosody/intonation, but is related to the overall pattern of intonational choices made.
# 2003 Elsevier B.V. All rights reserved.
Keywords: Prosody; Prosodic boundary preferences; Psycholinguistics

1. Introduction
While speakers of English can be quite eective in prosodically phrasing their
utterances to get the intended message across (see Schafer et al., 2000, for a recent
* Corresponding author. Tel.: +1-413-5450996; fax: +1-413-5452175.
E-mail address: lyn@linguist.umass.edu (L. Frazier).
0024-3841/03/$ - see front matter # 2003 Elsevier B.V. All rights reserved.
doi:10.1016/S0024-3841(03)00044-5

L. Frazier et al. / Lingua 114 (2004) 327

convincing demonstration), they also have a rather wide range of options about how
they can divide a spoken sentence into prosodic constituents. At lower levels of the
prosodic hierarchy, the conditions on forming a prosodic constituent are quite strict.
For example, there is little variability with respect to the formation of a phonological word. However, above the phonological word, there appears to be a lot of
optionality concerning where and if a prosodic break should occur. Linguistic analyses (Selkirk, 1984, 2000; Gussenhoven, 2000) make it clear that the syntactic
structure of a sentence aects but does not dictate its prosodic phrasing, and in
relatively natural production tasks like that used by Schafer et al. (2000), speakers
vary in the prosodic or intonational phrasings they use.
The present research examines several dierent approaches to the question of
what determines the appropriateness of a given choice of intonational phrasing. The
rst two experiments pit a syntactically-based constraint on intonational phrasing
against a semantically-based constraint. The second two experiments examine factors that are motivated by processing considerations, namely the length of a constituent and the distance over which interpretative integration must take place. The
outcome of the experiments suggests that multiple factors inuence the naturalness
of a given intonational phrasing, but also indicates that some factors are of lesser
importance than others.
Experiments 1 and 2 contrast two broadly dierent linguistic approaches to the
naturalness of prosodic breaks. One approach employs a basic constraint against
breaking up syntactic phrases, inspired by Wrap. Wrap was originally proposed by
Truckenbrodt (1995) as one of a small number of violable universal constraints in an
optimality theory (OT) grammar. In simple terms, Wrap says that prosodic phrasings that do not break up a syntactic phrase are favored over prosodic phrasings
that do, unless breaking up a syntactic phrase is required for some reason, e.g. to
satisfy some higher-ranked constraint. As phrased by Truckenbrodt, Wrap does not
apply to adjoined nodes or sentence nodes, though it is not clear conceptually why it
should not. Our concern in the present paper is not the particular details of any
prohibition against breaking up syntactic phrases, but the question of whether we
can nd evidence for a prohibition against breaking up syntactic phrases below the
level of the root sentence node. We will call this prohibition Generalized Wrap.
A semantically-based linguistic approach might claim instead that syntactic phrases
may be broken into parts freely providing that the resulting parts are semantically
sensible. In other words, given the syntactic constituents x y z, if x and y belong
together semantically but y and z do not, one may separately phrase each of x, y, and
z, but one may not place x alone and group y and z together. If only x and z belong
together semantically, one may place x, y, and z together or each alone but one may
not group y and z together without x (or x and y together, without z). Essentially, this
constraint punishes any prosodic phrasing which groups together semantically unrelated elements. We will call this the semantic coherence constraint (SCC).
The SCC was inspired by an earlier, more specic version of the semantic
approach proposed by Selkirk (1984). Her Sense Unit Condition, given in (1), places
a semantic constraint on the prosodic constituents resulting from breaking up a
syntactic unit.

L. Frazier et al. / Lingua 114 (2004) 327

(1) The Sense Unit Condition on Intonational Phrasing:

The immediate constituents of an intonational phrase must together form
a sense unit.
Two constituents, Ci, Cj form a sense unit if (a) or (b) is true of the
semantic interpretation of the sentence:
a:
Ci modies Cj (a head)
b:
Ci is an argument of Cj (a head).
Either the Sense Unit Condition, or the more general semantic coherence constraint discussed above, will limit the ways in which a sentence may be broken into
intonational parts. Although (1) refers to an intonational phrase, it is clear from the
examples discussed by Selkirk that smaller prosodic boundaries must also be regulated.
Assuming Pierrehumbert and Beckmans prosodic system (Beckman and Ayers, 1993;
Pierrehumbert, 1980; Pierrehumbert and Beckman, 1988), the condition must apply to
intermediate phrases (ip) as well as full intonational phrases (IPh).
Neither of these semantic constraints is fully satisfactory in its present form: the
SCC needs more explicitness in its description of semantic belonging, while the
Sense Unit Conditions specic predictions are occasionally too strong.1 Additionally, Generalized Wrap and the semantic coherence constraint (SCC) based on
Selkirks (1984) Sense Unit Condition do not exhaust the space of relevant constraints. For example, constraints favoring units of particular (relative) sizes will
likely also be needed to fully capture the placement of phonological boundaries.
Nevertheless, the two proposed constraints make clear and in some cases conicting
predictions on their own, and we believe that empirical testing of these hypotheses
will increase understanding of the prosodysyntax mapping.
Experiments 1 and 2 contrast our constraints predictions in four sentence types,
illustrated below in (2), (4), (5) and (6), with predictions summarized in the columns
to the right of the examples.2 Consider (2), an extraposed relative clause sentence
(Type I).
1
To accommodate examples like (i), Selkirk allows just the head of a phrase and the head of its
argument to form a sense unit. (i) (This is the cat) (that ate the rat.) In cases like (i) this will suce. But
with prenominal modiers it would allow (pretty)(girls left), which does not seem fully natural. Further,
Hirotani (2000) has shown that semantically unrelated moved constituents may at times prefer to occur in
the same prosodic phrase in complex Japanese sentences, if the constituents are moved from the same
clause. This prosodic phrasing would allow them to be held in memory as one unit, Hirotani proposes,
until the perceiver receives the clause from which they have been extracted. In this case, prosodic phrasing
may aid interpretation and memory, but not in ways explicitly captured by the Sense Unit condition or
the semantic coherence condition.
2
Caution: the columns after the examples may look similar to the tableaux of OT grammars, but that
is not their status. The Gwrap column simply summarizes the predictions of the Generalized Wrap
principle for the corresponding example, e.g., (2a) violates Wrap, hence an asterisk following (2a), but
(2b) does not violate it and therefore receives a @. (Note that the
indicates that the principle is violated one or more times. Unlike OT, more violations do not result in more
s.) The predictions of the
semantic constraint are summarized to the right. We treat the two principles as if they represent distinct
hypotheses, e.g. distinct grammars, not dierent principles in the same grammar (although in fact they
could co-exist in an OT grammar).

(2)

L. Frazier et al. / Lingua 114 (2004) 327

a.
b.

(Every girl laughed)(who was smiling)

(Every girl)(laughed who was smiling)

GWrap


@

SCC
@

The parentheses in (2) and the other examples in this paper stand for prosodic
boundaries above the level of phonological words. In our materials, these prosodic
boundaries are always intonational phrase boundaries, which end with a phrase tone
(H-, L-) followed by a boundary tone (H%, L%; Pierrehumbert, 1980; Pierrehumbert and Beckman, 1988). Assuming that extraposed relatives are adjoined to
VP,3 the structure of (2) is given in (3):
(3)

The phrasing in (2a) breaks up the VP, violating Generalized Wrap. Thus GWrap
predicts that (2a) should be unacceptable but (2b) should be acceptable. The
semantic coherence constraint predicts the opposite. In (2a) every girl and laughed
can be related semantically, though questions could arise concerning the presence
(and interpretation) of a trace left by the extraposed relative. But in (2b) the unit
laughed who was smiling clearly violates the SCC, because who was smiling semantically belongs with every girl and not directly with laughed. In other words, this violates the constraint that in x y z if x and z belong together and y and z do not, y and
z cannot form a unit on their own.
In (2) one might claim that a preference for constituents that are approximately
equal in size (cf. Gee and Grosjean, 1983; Fodor, 1998), which we will call Balance,
explains the intuitive preference for (2a), instead of the SCC. But consider (4), a
separated particle sentence (Type II):

(4)

a.
b.

(Jason)(called all the participants over)

(Jason called)(all the participants over)

GWrap
@

SCC
@

Bal


@

3
Examples with VP ellipsis, like (i), suggest that an extraposed relative clause may adjoin to either VP
or IP, but that VP is preferred. (i) A girl left who was laughing and so did a boy. Perceivers prefer to
interpret the second conjunct of (i) as and a boy left who was laughing, with the relative inside the copied
VP (instead of as a boy left, in which only the verb is in the VP and thus copied).

L. Frazier et al. / Lingua 114 (2004) 327

In (4), Balance would favor (4b). In contrast, Generalized Wrap predicts a preference for (4a), because the phrasing in (4b) breaks up the VP. The semantic coherence constraint also favors (4a), because the phrase all the participants over in (4b)
groups together the object of the verb with the verbs particle. This is a case of y and
z being grouped together when x and z (or x and y) are the elements that belong
together. Consequently if both (2a) and (4a) are judged more acceptable than their
counterparts, only the SCC could explain the overall pattern of data.
Next consider (5), a modied subject sentence (Type III).

(5)

a.
b.

(A lady)(from the accounting oce)(visited.)

(A lady)(from the accounting oce visited.)

GWrap

SCC
@

In (5a) and (5b), Generalized Wrap is violated because the subject DP/NP has been
broken up. So (5a) and (5b) are both predicted to be disfavored. By contrast, the
semantic coherence constraint predicts that it is acceptable to break up the sentence
into the parts in (5a). However, (5b) is predicted to be unacceptable because the PP
should not appear in a unit that contains the verb but not the head noun, as it belongs
with the head of the subject NP. Thus this is a case of x y z when only x and y belong
together semantically, but y and z have been grouped together without x. (Similarly,
the Sense Unit Condition would disfavor (5b) because a head (the verb) is grouped
together with a PP that is neither its argument nor an adjunct modifying it.)
Turning to (6), a sentence complement sentence (Type IV), it is clear that the VP is
broken up in both (6a) and (6b). Thus Generalized Wrap does not predict dierences between them but might predict both to be disfavored.

(6)

a.
b.

(John promised Mary)(that he would go)

(John promised)(Mary that he would go)

Gwrap

SCC
@

However, the semantic coherence constraint does make a distinction between

them. Specically (6b) should be unacceptable because Mary, which is an argument
of the verb promise, should either occur on its own or in a unit containing the head
verb promise rather than being phrased together with the verbs CP complement.

2. Experiment 1
Ten sentences of each of the above four types (extraposed relative clause, separated particle, modied subject, and sentence complement sentences) were constructed. Each was recorded twice: once with the a-prosody and once with the bprosody (see below, under Materials). Both Experiments 1 and 2 were designed to

L. Frazier et al. / Lingua 114 (2004) 327

measure the prosodic naturalness or acceptability of the utterances. In Experiment

1, one version of each sentence was played to participants who were asked to rate
the acceptability of the prosody, described as the melody and rhythm, of the
sentence. In Experiment 2, the same items were tested but the critical sentences were
presented as pairs. Participants were asked to make a relative judgment about which
member of the pair sounded better and then indicate whether the less good member
was or was not acceptable. Experiment 2 was conducted both as a replication and to
compare the absolute judgments in Experiment 1 with presumably more natural
(though more complicated) relative judgments.
2.1. Method
2.1.1. Materials
The 10 sentences of each of the four types described above appear in Appendix A.
Each was recorded in a sound-deadened chamber by a trained phonologist, who
placed an intonational phrase boundary (IPh) at the points indicated by the parentheses in examples (2), (4), (5), and (6), and by the / marks in Appendix A. Each
sentence was digitized (16 bit, 22 kHz). Pitchtracks of typical utterances of sentences
like those in example (2) appear in Fig. 1. The intonation contours and timing they
exhibit at the claimed IPh boundaries were generally typical of utterances of all four
types of sentences. The prosodic analysis was veried by measuring several acoustic
parameters of each word immediately preceding a boundary and the corresponding
word when it did not precede an intended boundary [e.g., girl and laughed in (2)].
Specically, we measured the duration of the word, its minimum F0 (fundamental
frequency) value, and the F0 value at the end of the word. An IPh boundary in these
materials is typically characterized by durational lengthening of a preceding word
and by a fall then a rise in F0. Table 1 provides the means of the measured parameters and indicates that the sentences generally had just this pattern where an IPh
boundary was intended (compared to shorter durations and no fall-rise when no
boundary was intended).
2.1.2. Participants and procedure
The resulting 40 sentences (with two versions of each) were separated into two
counterbalanced lists, each with ve instances of each version of each sentence type.
They were supplemented with 44 additional unrelated sentences and two-sentence
discourses (28 of which will be described in Experiments 3 and 4).
Thirty-two undergraduate students at the University of Massachusetts participated in individual half-hour sessions for course credit. Each was tested in a sounddeadened chamber. A session began with instructions that the participant was to
rate each sentence or discourse for its rhythm and melody, deciding whether it was
pronounced with a natural, normal, acceptable rhythm and melody. They were
instructed to use a ve point scale, where 5 is the best and 1 is the worst, and were
told not to rate how much sense the sentence makes apart from its rhythm and
melody. The subject announced the rating to the experimenter who recorded it into
the computer. A six-sentence practice list was rst presented by playing the digitized

L. Frazier et al. / Lingua 114 (2004) 327

Fig. 1. Pitchtracks of typical utterances of the extraposed relative clause sentences [as in (2)] used in
Experiments 1 and 2.

les, and then the experimenter answered any questions the participant had. Following that, the participant received an individually-randomized presentation of one
of the two counterbalanced lists (16 participants to each list).
2.2. Results
The mean ratings are presented in Table 2. The version of the extraposed relative
clause sentences (Type I) favored by the Semantic Coherence Constraint was rated
signicantly higher than the version favored by Generalized Wrap [(2a) vs (2b),
F1(1,31)=20.03, F2(1,9)=20.27, P < 0.001] (F1 designates the results of an analysis
permitting generalization to participants, and F2 the results of an analysis permitting generalization to items). The version of the separated particle sentences (Type
II) favored by generalized Wrap and the SCC was rated higher than the version
favored by balance [(4a) vs (4b), F1(1, 31)=52.37; F2(1, 9)=95.63, P < 0.001]. The

10

L. Frazier et al. / Lingua 114 (2004) 327

Table 1
Acoustic measurements on critical words (with example words), Experiment 1; mean durations (ms) and
F0 values (Hz). Boldfaced values indicate a word with an assumed LH% intonational boundary
Sentence group and version

Word and measure

Word 1

I. Girl laughed
A.
B.
II. Jason called
A.
B.
III. Lady oce
A.
B.
IV. Promised Mary
A.
B.

Word 2

Duration

F0 min

F0 end

Duration

F0 min

F0 end

432
645

193
169

197
232

630
391

178
193

232
197

561
372

168
199

229
200

265
509

184
167

187
226

584
572

175
172

231
226

512
305

170
204

231
208

358
572

180
171

180
247

448
246

175
184

243
190

Table 2
Mean sentence ratings (Experiment 1) and mean percentage preference (Experiment 2)

I.
II.
III.
IV.

a
b.
a.
b.
a.
b.
a.
b.

Sentence type with example

Expt 1

Expt 2

(Every girl laughed)(who was smiling)

(Every girl)(laughed who was smiling)
(Jason)(called the participants over)
(Jason called)(all the participants over)
(A lady)(from the accounting oce)(visited)
(A lady)(from the accounting oce visited)
(Josh promised Mary)(that he would go)
(Josh promised)(Mary that he would go)

3.30
2.87
3.93
3.05
3.70
3.70
3.94
2.99

70%
80%
66%
80%

two versions of the modied subject sentences (Type III) did not dier in prosodic
acceptability [(5a) vs. (5b), F < 1]. GWrap made no predictions about the relative
acceptability of the sentence complement sentences (Type IV), but the version
favored by the SCC was rated higher than the other [(6a) vs (6b), F1(1, 31)=101.48,
F2(1, 9)=49.69, P < 0.001].
2.3. Discussion
The ratings supported the predictions of the semantic coherence constraint in
three cases of four, all except the Type III modied subject sentences. For the
extraposed relative clause sentences and sentence complement sentences, a prosodic
rendition which broke up a syntactic constituent was favored or highly acceptable,
as long as the phrases thus created were semantically coherent, contrary to Gen-

L. Frazier et al. / Lingua 114 (2004) 327

11

eralized Wrap. The separated particle sentences (Type II) provided no support for
Balance. It is possible that participants found each version of the Type III sentences
fairly acceptable (for reasons to be discussed later) and that the rating task was not
rened enough to pick up subtle dierences in acceptability. Experiment 2 was conducted using the same materials but a dierent, two-choice, technique for determining acceptability.

3. Experiment 2
3.1. Method
Forty-eight undergraduates were tested in the same environment used in Experiment 1, and the materials used (including the ller and practice items) were the same
as those used in Experiment 1. However, on each trial, a participant heard the two
versions of each sentence in succession and indicated by pulling one of two response
triggers whether the rst or the second was the better one in terms of its rhythm and
melody. Which version occurred rst was randomly determined on each trial. Following the participants response, a message appeared on a video screen asking the
participant to indicate by pulling one of two triggers whether the unpreferred version was acceptable or not. The participant could re-hear the unpreferred version as
often as needed by pressing a thumb button. The primary measure was the percentage of times each version was initially preferred. A secondary measure was the
percentage of times each version was classied acceptable when it had been initially the unpreferred version.
3.2. Results
Table 2 presents the percentages of time the version favored by the SCC was
chosen as the more preferred one. In each case, the percentage was signicantly
greater than 50%, with F1 and F2 ranging from 43.26 to 212.83, P < 0.001, for
sentences of Types I, II, and IV. While the two prosodic versions of the Type III
sentences had not diered from each other in the rating data of Experiment 1, the
forced choice response of Experiment 2 proved more sensitive, with the version
favored by the SCC (6a) chosen as preferred 66% of the time, signicantly greater
than 50% [F1(1, 47)=21.80, P < 0.001; F2(1, 9)=38.61, P < 0.001].
Table 3 presents the percentages of times that participants indicated that the version that they dispreferred on a given trial was nonetheless acceptable. For each type
of item except for Type III modied subject sentences, participants were more frequently willing to accept the version favored by the Semantic Coherence Constraint
than the version favored by Generalized Wrap or Balance. Varying numbers of
participants (between 5 and 13) had to be eliminated from the by-subjects analyses
because they never indicated that one version of a sentence type was unacceptable,
which is reected in the d.f. of the following tests: Type I: F1(1, 40)=3.78, P < 0.06;
F2(1, 9)=6.34, P < 0.04; Type II: F1(1, 41)=6.32, P < 0.02; F2(1, 9)=32.30,

12

L. Frazier et al. / Lingua 114 (2004) 327

Table 3
Mean percentages with which initially unpreferred version was indicated to be acceptable
Experiment 2

I.
II.
III.
IV.

Sentence type with example

Percentage acceptable
when initially disfavored

a
b.
a.
b.
a.
b.
a.
b.

58
46
66
49
78
69
68
47

(Every girl laughed)(who was smiling)

(Every girl)(laughed who was smiling)
(Jason)(called the participants over)
(Jason called)(all the participants over)
(A lady)(from the accounting oce)(visited)
(A lady)(from the accounting oce visited)
(Josh promised Mary)(that he would go)
(Josh promised)(Mary that he would go)

P < 0.001; Type IV: F1(1, 47)=6.99, P < 0.02; F2(1, 9)=46.98, P < 0.001. Both
versions of the Type III modied subject sentences were generally rated as acceptable even when they were initially dispreferred, and Version a, favored by the SCC,
was rated as acceptable signicantly more than Version b only by items, but not by
subjects: F1(1, 42)=2.31, P > 0.13; F2(1, 9)=14.65, P < 0.01.
3.3. Discussion
In general, both the absolute and the relative judgments supported the predictions
of the semantic coherence constraint. Breaking up syntactic constituents is ne as
long as the resulting unit is semantically sensible. The data do not provide any support for Generalized Wrap. In particular, (2a), (5a), and (6) show that breaking up
syntactic constituents such as a VP or DP can be perfectly acceptable. The data also
suggest that if balance plays a role in prosodic phrase, it must be a minor role, subordinate to the semantic conditions.
The one unexpected nding is the high absolute rating received by (5b): (A
lady)(from the accounting oce visited). The relative judgment data of Experiment
2 did show a clear (66%) preference for the a-form. However, the absolute rating
data of Experiment 1 did not. The oddity may be due to acoustic characteristics of
the materials: (5b) is so odd that the speaker may have often introduced a very
subtle break before the verb [visited in (5b)]. However, the break was clearly not an
IPh boundary in (5b), as can be seen in Table 1. There was little or no tonal movement on the word before the break (oce in Table 1) and no evidence of lengthening
of this word. Further, listening to the sentences gave little or no evidence of an ip
boundary that might have been enough to save sentences like (5b) when a better
alternative was not available. We conclude that while the dierence observed in
Experiment 2 does support the Semantic Coherence Constraint, other unidentied
factors may also inuence the acceptability of dierent prosodic phrasings.
The data from Experiments 1 and 2 support the spirit of Selkirks Sense Unit
Condition, if not necessarily the specics, and certainly support the more general

L. Frazier et al. / Lingua 114 (2004) 327

13

Semantic Coherence Constraint. We take these results as evidence that the theoretical notion of semantic coherence is important in the evaluation of prosodic phrasing and should be considered in future processing work. Now we turn to the eects
of processing factors on prosody.

4. Experiment 3
Experiments 3 and 4 examine the eect that two factors motivated by processing
considerations may have on the naturalness of dierent prosodic phrasings. In particular, the experiments examine the approach that Watson and Gibson (2001) take
to intonational phrasing. Watson and Gibson are primarily concerned with where a
speaker places an intonational phrase break and suggest that just two factors are
important. One is the distance between the syntactic head of the upcoming phonological phrase and the heads syntactic integration site. Longer integration distances
create processing diculty and this increases the probability of an intonational
boundary between two phonological phrases.4 The second factor is the eventual size
(in phonological phrases) of the upcoming constituent. The larger the constituent
the higher the probability of an intonational phrase boundary. The two factors are
summed and the sum is claimed to predict the probability of an intonational phrase
boundary.
In a production study, Watson and Gibson (2001) varied the integration distance
of the main verb and the head of its subject NP by manipulating the length of an
intervening relative clause. The longer the relative clause, the more likely the speaker
was to place an intonational phrase boundary before the main verb. In a comprehension study, Watson and Gibson (2001) expected listeners to be sensitive to the
same factors that had mattered in production. Watson and Gibson placed an intonational boundary at point (a) in sentences like (7), before a long distance integration of to the client to showed, or at point (b), before a long relative clause.

(7)

The detective showed the blurry picture of the diamond (a) to the client
(b) who was in the oce that was on the fourteenth oor.

Listeners rated sentences with a boundary at either (a) or (b) as less complex than
sentences without the boundary. Although Watson and Gibson did not test it, presumably having both breaks would also be predicted to aid comprehension.
To test Watson and Gibsons hypotheses, we investigated sentences like those in
(8) where there is a long distance integration between the trace (t) and the interrogative wh-phrase.
4
Later work has persuaded these authors to give up the claim that integration distance aects the
appropriateness of a prosodic boundary (personal communication, D. G. Watson, March, 2002; Watson
and Gibson, 2002). The results we report here are consistent with this change of position.

14

L. Frazier et al. / Lingua 114 (2004) 327

(8)
a. (Which lm ) (did Anthony say Timothy wanted Mary) (to see t?)
integration:
0
0
0
0
0
5
size:
6
5
4
3
2
1

b.
(Which lm did Anthony say) (Timothy wanted Mary to see t?)
integration:
0
0
0
0
0
5
size:
6
5
4
3
2
1

c.
integration:
size:

(Anthony said Timothy wanted Mary)( to see the lm.)

0
0
0
0
0
0
6
5
4
3
2
1

d.
integration:
size:

(Anthony said)(Timothy wanted Mary to see the lm.)

0
0
0
0
0
0
6
5
4
3
2
1

The numbers below specic words in (8) designate the integration and size indices.
The integration index indicates how many phonological words the specic word
must be integrated across to relate it to other material in the sentence.5 Watson and
Gibsons integration factor predicts that each phonological word has an integration
cost of zero until see, since no integration across intervening phonological phrases is
required until to see is encountered. At this point the trace in the position of the
object of see must be integrated across ve intervening items. The large integration
index of 5 below see in (8a, b) increases the probability of a boundary immediately before see; such a boundary should be helpful to a listener, because it occurs
before a dicult integration.
In contrast to the integration index, which is associated with the word after the
boundary of interest, the size index appears under the word before the boundary of
interest. In indicates how many phonological words the upcoming syntactic constituent contains. A boundary after which lm, which appears in (8a), should help
processing because the constituent did Anthony say Timothy wanted Mary to see is
long, containing 6 phonological words. Therefore the two boundaries in (8a) are at
just the right places, and (8a) should be preferred to (8b). By contrast, the boundary
in (8b) is at a position where the integration index is zero, instead of 5 at the second
break in (8a), and the size index is 4 instead of 6.
We also tested the declarative counterparts of (8a and b), namely, (8c and d). The
integration index does not choose between the phrasings in (8c) and (8d), since no
long-distance integration is necessary in either sentence. However, the size index will
5

The term phonological word in this context designates each word except did, to, the and possessive
pronouns, following Watson and Gibson.

L. Frazier et al. / Lingua 114 (2004) 327

15

favor the phrasing in (8d) because the break in (8d) coincides with the beginning of a
larger constituent than the break in (8c)
Subject questions like those in (9) were also tested.

(9)
a.
(Which assistant did Marcia announce) (t hated his job?)
integration:
0
0
2
0
size:
4
3
2
1
b.
(Which assistant) (did Marcia announce t hated his job?)
integration:
0
0
2
0
size:
4
3
2
1

The trace (subject of hated) in (9) must be integrated across two phonological
words and thus a break before the verb (hated) should be helpful. For the size
parameter, however, the longest upcoming constituent is the C0 constituent (did
Marcia announce hated his job). If both size and integration contribute to boundary
preferences (as claimed by Watson and Gibson, 2001), then summing the indices
under (9) makes it clear that (9b), where a boundary occurs at a position with a
summed index of 4, should be similar to (9a), where the boundary occurs at a position which also has an index of 4.6 However, if only size and not integration aect
the appropriateness of a boundary (see footnote 4), then (9b) should be preferred to
(9a).7
Although sentences (8) and (9) were not designed to provide strong tests of the
principles examined in Experiments 1 and 2 (Generalized Wrap, SCC, and balance),
these principles do apply to the sentences. Sentences (8a) and (8c) can be argued to
violate Generalized Wrap, in that they both contain a prosodic boundary that places
part of a sentence complement into one prosodic unit and the remainder into a different unit. Sentence (8b) seems to violate the SCC in that it separates the verb of the
complement sentence from its (wh-moved) object without separating this moved
phrase from material it is unrelated to. In other words, which lm did Anthony say
does not form a coherent unit. Sentence (8b), however, is most consistent with the
balance principle. Sentence (8d) is consistent with both Generalized Wrap and SCC.
A similar analysis indicates that (9a) is consistent with Generalized Wrap but not
SCC, while the opposite holds for (9b).
6
Six of the eight items in this subexperiment contained four phonological words in the C0 . One item
contained ve phonological words in C0 and one contained six. The predictions of Watson and Gibsons
summed index do not dier for these items.
7
Nagel et al. (1994) proposed that wh-traces, including subject traces, are marked prosodically by long
durations and large pitch excursions on the word preceding the trace. However, subsequent research
strongly suggests that the eects they observed are in fact due to the presence of intonational phrase
boundaries that happened to occur at the position of the trace. In a production study, Straub et al. (2001)
showed that traces give rise to long durations and large pitch excursions only when they co-occur with an
intonational phrase boundary, in which case the boundary alone could explain the eects. When a trace
occurred internal to an intonational phrase, neither long durations nor large pitch excursions were
observed on the pre-trace word.

16

L. Frazier et al. / Lingua 114 (2004) 327

4.1. Method
Twenty sentences like (8) were constructed and recorded in four versions each
(question vs. declarative, each with two prosodic phrasings). Eight sentences like
those in (9) were recorded, in two versions each. The same speaker who recorded the materials of Experiments 1 and 2 recorded the Experiment 3 (and
Experiment 4) materials in the same setting. She introduced an IPh boundary at
the points marked by parentheses in (8) and (9) and by / in Appendix B. Figs. 2
and 3 present pitchtracks of typical utterances of all six sentence types. The presence
of fall-rise contours and lengthening/pausing at IPh boundaries can be clearly
observed.
The experiment was run concurrently with Experiment 1 and the procedures
used were identical except that Experiment 3 required four dierent counterbalanced lists to permit each version of each sentence like (8) to be tested equally
often.
4.2. Results
The results are presented in Table 4. Turning rst to the object questions (8a,b)
and their declarative counterparts (8c,d), the declarative with a prosodic boundary
after the matrix verb (8d) was rated much better than the other three forms (8a,b,c),
resulting in a signicant interaction between question/declarative and boundary
position [F1(1, 31)=24.92; F2(1, 19)=27.83, P < 0.001] as well as signicant main
eects of both factors. Although the Watson and Gibson size hypothesis is consistent with the high rating for (8d), their integration measure is not consistent with
the comparable ratings for (8a) and (8b). (8a) was predicted to be better than (8b),
with breaks at two positions with summed indices totaling 6 vs. a break at one
position with a 4 index.
Turning to (9), both (9a) and (9b) received a relatively high rating (3.60 and 3.84
respectively) and they diered signicantly from each other only in the analysis by
participants [F1(1, 31)=4.54, P < 0.05; F2(1, 9)=1.80, P > 0.20). The direction of
the dierence suggests that boundaries are favored before long constituents, supporting the size hypothesis, but this result cannot independently disconrm the prediction of equal acceptability for (9a) and (b) which was given by the integration
plus size theory.

5. Experiment 4
5.1. Method
The nal experiment used the relative naturalness judgment task of Experiment 2
to test the sentences examined in Experiment 3. The experiment was conducted
concurrently with Experiment 2, and used the materials from Experiment 3,
excluding the declarative versions of (8).

L. Frazier et al. / Lingua 114 (2004) 327

Fig. 2. Pitchtracks of typical utterances of Experiment 3 and 4 sentences as in (8).

17

18

L. Frazier et al. / Lingua 114 (2004) 327

Fig. 3. Pitchtracks of typical utterances of Experiment 3 and 4 sentences as in (9).

5.2. Results
Table 4 presents the percentages of times that (8a) was chosen over (8b), and that
(9a) was chosen over (9b). The former percentage was not signicantly dierent
from 50% [F1(1, 47)=2.11, P > 0.15; F2(1, 19)=1.94, P > 0.15]. The latter percentage, however, was signicantly less than 50%, both by participants and items
[F1(1, 47)=22.44; F2(1, 7)=32.91, P < 0.001]. This preference for the phrasing in
Table 4
Mean sentence ratings (Experiment 3) and mean percentage preference (Experiment 4)

(8)

(9)

a.
b.
c.
d.
a.
b.

(Which lm)(did A say T wanted M)(to see ____?)

(Which lm did A say)(T. wanted M to see ____?)
(A said T wanted Mary)(to see the lm)
(A said)(T wanted Mary to see the lm)
(Which assistant did Marcia announce)(____ hated his job?)
(Which assistant)(did Marcia announce ____ hated his job?)

Expt 3

Expt 4

2.89
2.86
2.82
3.58
3.60
3.84

47%

36%

L. Frazier et al. / Lingua 114 (2004) 327

19

(9b) over that in (9a) reinforces the tendency observed in Experiment 3 for a
boundary to be preferred before a long constituent.
The secondary response measure gathered in Experiment 4, the percentage of
times the version that a participant indicated to be dispreferred was nonetheless
rated as acceptable, did not dier between items (8a) and (8b) [49 vs 54%, F1(1,
47)=2.27, P > 0.10; F2(1, 19) < 1]. However, there was a suggestion that participants more frequently indicated accepting (9b) when they had initially preferred (9a)
than the opposite, 79 vs. 66% of the time [F1(1, 43)=6.52, P < 0.02; F2(1, 7)=3.91,
P < 0.10], providing some additional evidence that (9b) was favored over (9a).
5.3. Discussion
The clear superiority of (8d) over (8ac) is congruent with the fact it is the only
version that is consistent with both Generalized Wrap and the semantic coherence
constraint, but it also ts with Watson and Gibsons (2001) claim that length considerations favor breaks before long upcoming constituents. The tendency for (9b)
to be preferred over (9a) is also consistent with the claim that boundaries are
favored before long constituents. The results give no support to the proposal that
prosodic boundaries are also favored before a point where integration is required
over a long distance. They also suggest that Balance, which had favored (8b), is at
most a weak principle.

6. General discussion
We have examined the role of several principles in characterizing the intonational
phrasing of sentences: Generalized Wrap (dont break up constituents); semantic
coherence constraint (dont form semantically incoherent groups); balance (favor
equi-sized units); integration (break before long distance integrations); and length of
upcoming constituent (break before long upcoming constituents).
The data do not provide evidence supporting Generalized Wrap or the Integration
principle, and Balance must be a low ranked principle as indicated by the results for
both (4b) and (8). The semantic coherence constraint and the importance of length
of an upcoming constituent are clearly supported. However both principles need to
be formulated more precisely. The semantic coherence constraint depends on a
notion of what goes together semantically, without providing an explicit denition
of the notion. The length principle also requires further attention. Intuitions suggest
that the relative length of prosodic phrases may play a role in determining the
eective weight of a phrase, and that focus may contribute to weight (Arnold et al.,
2000). Further, the interaction of the SSC and length principles must be explored.
It is possible to take our data as arguing against the existence of any prosodic
constraint (like Wrap) which prohibits breaking a syntactic constituent into prosodic phrases. Limiting attention to the higher level prosodic categories (intermediate
phrases and intonational phrases), though, there is nothing intrinsically better about
avoiding phrase breaks than having phrase breaks. In our experimental sentences,

20

L. Frazier et al. / Lingua 114 (2004) 327

pronunciations in which breaks occurred were rated as being very natural (e.g. 34
on a 5 point scale) even if a break occurred internal to a constituent. This holds even
for sentences presented out of the blue, without narrow focus which might have
motivated a (following) phrase break, and even for relatively short sentences that
might in principle have occurred without a higher level boundary. What matters is
not the simple presence or absence of a break, but whether the resulting prosodic
phrases are semantically coherent. Sentences in which strings of constituents which
do not semantically belong together are prosodically grouped together are rated as
less natural than sentences with breaks creating semantically coherent groups.
The results of auditory comprehension studies (Carlson et al., 2001; Clifton et
al., in press) have led us to argue that absolute break size (intermediate phrase
versus intonational phrase) is not important to the listeners interpretation of a
sentence. Instead, the listener is concerned with relative break size: a break is
informative to the extent that it is, say, larger than some other break in the sentence. The present results t well with the conclusions of this earlier work: regularities in prosodic phrasing appear to be concerned not with purely local properties,
such as this particular placement of a boundary, or this particular kind of boundary,
etc., but instead they are concerned with the patterns resulting from particular
intonational choices. Whats informative about optional higher level prosodic
breaks is not whether they occur or what size the break is, but where else they do or
do not occur (and thus what semantically related or unrelated elements are grouped
together) and the relative size of those breaks. This level of abstract informativity,
interpreting a speakers intonational decisions relative to the speakers other intonational decisions in the utterance, makes sense in light of the fact that length is
important to the articulatory system (which is parasitic on our respiratory mechanisms) but not to the syntactic/semantic systems. In other words, invariant intonational translation of particular syntactic phrases would be dicult given that the
phrase in question might be one syllable long or 25. Instead of an invariant intonational interpretation of a particular syntactic or semantic structure, what human
languages seem to have is a exible system allowing considerable optionality in the
prosodic phrasing of a sentence (see Schafer et al., 2000). But exibility does not
necessarily tolerate irrationality or arbitrariness. The particular options chosen by
the speaker must be coherent in light of that speakers other choices and the speakers intent.
The current results together with already known facts about low level phonological phrasing (into phonological words or phonological phrases) suggest that a break
is prohibited not by any general constraint against breaking up syntactic constituents, but simply by constraints on phonological words/phrases as well as constraints favoring semantically coherent groups. Otherwise a break may freely occur,
and it will have interpretive consequences relative to the existence, size, and location
of other breaks. Long phrases are of particular interest in this regard because on the
one hand they may motivate the speaker to take the option of breaking in some
location, but on the other, length per se may lessen the informativity of a break,
since phonological rather than syntactic or semantic reasons may justify the existence of the break. We are currently testing this reduced informativity hypothesis.

L. Frazier et al. / Lingua 114 (2004) 327

21

This approach to prosodic phrasing makes the prediction that the results of
almost any prosodic boundary may be salvaged (unless it violates low level phonological constraints, for example by forcing a function word to be a phonological
phrase by itself). Basically, a break should be bad only if it creates semantically
incoherent groups, which may be salvaged by further breaks, or if the break results
in a misleading pattern, e.g., a break at a small syntactic boundary but no break at a
larger or higher syntactic boundary, which could also be salvaged by further breaks.
This is illustrated below for example (2b), repeated as (10).

(10)

(Every girl) (laughed who was smiling)

If the second prosodic phrase is divided into (laughed) and (who was smiling), the
sentence is predicted to be natural. Our intuitions support the prediction.
Let us consider some of the other dispreferred prosodic conditions, such as
example (3b), repeated as (11):

(11)

(Jason called) (all the participants over).

Presumably the oensive prosodic phrase in (12) is all the participants over. This in
principle might be broken up. However, a short function word like over does not
readily stand as a phonological phrase by itself. Consequently, this is one prosodic
phrasing that may not easily be salvaged.
In contrast, though, see sentence (12), which repeats (5b).

(12)

(Josh promised) (Mary that he would go.)

The break which creates the semantically incoherent phrase Mary that he would go
should be salvageable if another break is introduced after Mary. This may be somewhat
unnatural given the shortness of Mary but seems considerably better if we replace Mary
with a long name: (Josh promised) (Signora Carminati) (that he would go.). Again intuitions seem to support the prediction. In short, even with our dispreferred sentences, it is
not one particular break that is unnatural but the overall prosodic pattern it creates.

Acknowledgements
This research was supported in part by NSF Grant BCS-0090674, Prosody in
Sentence Comprehension, to the University of Massachusetts. Katy Carlson is now
at Northwestern University.

22

L. Frazier et al. / Lingua 114 (2004) 327

Appendix A. Sentences used in Experiments 1 and 2

Note: A / indicates an intonational phrase boundary. In each case, the version
predicted in the text to be preferred appears rst.
I. Extraposed relative clause sentences

1. Every girl laughed / who was smiling; Every girl / laughed who was smiling.
2. No student complained / who was passing; No student / complained who was
passing.
3. Each customer replied / who lived in the area; Each customer / replied who
lived in the area.
4. Lots of patients returned / who didnt respond to the treatment; Lots of
patients / returned who didnt respond to the treatment.
5. Many players retired / who had serious injuries; Many players / retired who
had serious injuries.
6. Some scientists quit / who objected to the project; Some scientists / quit who
objected to the project.
7. An assistant lied / who was afraid of the police; An assistant / lied who was
afraid of the police.
8. Every child tripped / who was jumping rope; Every child / tripped who was
jumping rope.
9. Many actors starved / who were very talented; Many actors / starved who
were very talented.
10. No employee protested / who had children; No employee / protested who had
children.
II. Separated particle sentences
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.

Jason / called all the participants over; Jason called / all the participants over.
Alice / asked every applicant back; Alice asked / every applicant back.
Martin / paid several lenders back; Martin paid / several lenders back.
Sally / invited the friendly neighbors over; Sally invited / the friendly neighbors over.
Molly / rowed the campers across; Molly rowed / the campers across.
Anthony / brought the lawn mower back; Anthony brought / the lawn mower
back.
Lydia / passed the parsley-potatoes around; Lydia passed / the parsleypotatoes around.
All the kids / carried the grocery bags in; All the kids carried / the grocery
bags in.
The angry actor / pushed the props over; The angry actor pushed / the props
over.
The director / sent the notice around; The director sent / the notice around.

L. Frazier et al. / Lingua 114 (2004) 327

23

III. Modied subject sentences

1. A lady / from the accounting oce / visited; A lady / from the accounting
oce visited.
2. The electrician / from the physical plant / stopped by; The electrician / from
the physical plant stopped by.
3. The contractor / for the housing project / was sued; The contractor / for the
housing project was sued.
4. A waitress / at the coee shop / ran away; A waitress / at the coee shop ran
away.
5. A tailor / at the boutique / quit; A tailor / at the boutique quit.
6. The nurse / at the clinic / helped Maria; The nurse / at the clinic helped Maria.
7. A lawyer / from the big law rm / was indicted; A lawyer / from the big law
rm was indicted.
8. A garbage man / in San Francisco / was killed; A garbage man / in San
Francisco was killed.
9. The mayor / of New York / should resign; The mayor / of New York should
resign.
10. The swimmers / near the pier / shouted; The swimmers / near the pier
shouted.

IV. Sentence complement sentences

1. Josh promised Mary / that he would go; Josh promised / Mary that he would go.
2. Max assured Megan / that hed show up; Max assured / Megan that hed
show up.
3. Anita persuaded Lou / that she could dance; Anita persuaded / Lou that she
could dance.
4. Angela convinced Freddie / that she would help; Angela convinced / Freddie
that she would help.
5. Martin promised Lisa / that he would change; Martin promised / Lisa that he
would change.
6. Ian told Sue / that he resigned; Ian told / Sue that he resigned.
7. Marta warned Fritz / that she was sick; Marta warned / Fritz that she was
sick.
8. Paula promised Ken / that shed excel; Paula promised / Ken that shed excel.
9. Niki persuaded Tom / that she worked hard; Niki persuaded / Tom that she
worked hard.
10. Daniel assured Lucy / that he told the truth; Daniel assured / Lucy that he
told the truth.

24

L. Frazier et al. / Lingua 114 (2004) 327

Appendix B. Sentences used in Experiments 3 and 4

Sentences of type (8)

1. Which lm / did Anthony say Timothy wanted Mary / to see?

Which lm did Anthony say / Timothy wanted Mary to see?
Anthony said Timothy wanted Mary / to see the lm.
Anthony said / Timothy wanted Mary to see the lm.
2. Which house / did the realtor report that Misha intended / to buy?
Which house did the realtor report / that Misha intended to buy?
The realtor reported that Misha intended / to buy the house.
The realtor reported / that Misha intended to buy the house.
3. Which bicycle / did Ma say Fred thought Billy / should ride?
Which bicycle did Ma say / Fred thought Billy should ride?
Ma said Fred thought Billy / should ride the bicycle.
Ma said / Fred thought Billy should ride the bicycle.
4. Which book / did Lucy indicate that Tom said Maria / hated?
Which book did Lucy indicate / that Tom said Maria hated?
Lucy indicated that Tom said Maria / hated the book.
Lucy indicated / that Tom said Maria hated the book.
5. Which experiment / did Tim hear Chuck say Anita /designed?
Which experiment did Tim hear / Chuck say Anita designed?
Tim heard Chuck said Anita / designed an experiment.
Tim heard / Chuck said Anita designed an experiment.
6. Which trip / did the travel agent complain that she hated / to plan?
Which trip did the travel agent complain / that she hated to plan?
The travel agent complained that she hated / to plan the trip.
The travel agent complained / that she hated to plan the trip.
7. Which rug / did the salesman pretend he wanted his wife / to keep?
Which rug did the salesman pretend / he wanted his wife to keep?
The salesman pretended he wanted his wife / to keep the rug.
The salesman pretended / he wanted his wife to keep the rug.
8. Which coee / did Barbara claim Bill said Tim / wouldnt drink?
Which coee did Barbara claim / Bill said Tim wouldnt drink?
Barbara claimed Bill said Tim wouldnt / drink the coee.
Barbara claimed / Bill said Tim wouldnt drink the coee.
9. Which report / did the chemist announce the editor indicated the journal /
wouldnt print.
Which report did the chemist announce / the editor indicated the journal wouldnt
print.
The chemist announced the editor indicated the journal / wouldnt print the
report.
The chemist announced / the editor indicated the journal wouldnt print the
report.

L. Frazier et al. / Lingua 114 (2004) 327

25

10. Which tea leaves / did the palmist say she was sure the gypsy / wouldnt read?
Which tea leaves did the palmist say / she was sure the gypsy wouldnt read?
The palmist said she was sure the gypsy / wouldnt read the tea leaves.
The palmist said / she was sure the gypsy wouldnt read the tea leaves.
11. Which assignment / did the students announce they didnt have enough time /
to complete?
Which assignment did the students announce / they didnt have enough time to
complete?
The students announced they didnt have enough time / to complete the assignment.
The students announced / they didnt have enough time to complete the assignment.
12. Which billboard / did the worker think the manager was wrong / to approve?
Which billboard did the worker think / the manager was wrong to approve?
The worker thought the manager was wrong / to approve the billboard.
The worker thought / the manager was wrong to approve the billboard.
13. Which cat / did Elizabeth know Josh saw Meghan / feed?
Which cat did Elizabeth know / Josh saw Meghan feed?
Elizabeth knew Josh saw Meghan / feed the cat.
Elizabeth knew / Josh saw Meghan feed the cat.
14. Which rumor / did Max claim the reporter found out / Sam started?
Which rumor did Max claim / the reporter found out Sam started?
Max claimed the reporter found out Sam / started the rumor.
Max claimed / the reporter found out Sam started the rumor.
15. Which swing set / did Ana say the neighbor claimed Fred / xed?
Which swing set did Ana say / the neighbor claimed Fred xed?
Ana said the neighbor claimed Fred / xed the swing set.
Ana said / the neighbor claimed Fred xed the swing set.
16. Which quartet / did the newspaper report the University refused / to hire?
Which quartet did the newspaper report / the University refused to hire?
The newspaper reported the University refused / to hire the quartet.
The newspaper reported / the University refused to hire the quartet.
17. Which scandal / did Kathy announce the supervisor told her / to believe?
Which scandal did Kathy announce / the supervisor told her to believe?
Kathy announced the supervisor told her / to believe the scandal.
Kathy announced / the supervisor told her to believe the scandal.
18. Which medicine / did Lisa say her doctor claimed Mr. Wu / invented?
Which medicine did Lisa say / her doctor claimed Mr. Wu invented?
Lisa said her doctor claimed Mr.Wu / invented the medicine.
Lisa said / her doctor claimed Mr.Wu invented the medicine.
19. Which chair / did Lynne proudly announce her husband said Timmy /
repaired?
Which chair did Lynne proudly announce / her husband said Timmy repaired?
Lynne proudly announced her husband said Timmy / repaired the chair.
Lynne proudly announced / her husband said Timmy repaired the chair.

26

L. Frazier et al. / Lingua 114 (2004) 327

20. Which concert / did the report indicate the promoters now refused / to back?
Which concert did the report indicate / the promoters now refused to back?
The report indicated that the promoters now refused / to back the concert.
The report indicated / that the promoters now refused to back the concert.
Sentences of type (9)
1. Which assistant did Marcie announce / hated his job?
Which assistant / did Marcie announce hated his job?
2. Which report did Maxwell claim / was false?
Which report / did Maxwell claim was false?
3. Which wounded bird did Ellen think / would live?
Which wounded bird / did Ellen think would live?
4. Which recording did Rachel claim / was lost?
Which recording / did Rachel claim was lost?
5. Which pen did Karen think / lasted longest?
Which pen / did Karen think lasted longest?
6. Which neighbor did Mrs. Grimshaw indicate / was sick?
Which neighbor / did Mrs. Grimshaw indicate was sick?
7. Which oak tree did Thomas assert / was 200 years old?
Which oak tree / did Thomas assert was 200 years old?
8. Which guest did Carey suspect / hated cherries?
Which guest / did Carey suspect hated cherries?

References
Arnold, J.E., Wasow, T., Losongco, A., Ginstrom, R., 2000. Heaviness vs. newness: the eects of structural complexity and discourse status on constituent ordering. Language 76, 2857.
Beckman, M.E., Ayers, G.M., 1993. Guidelines for ToBI Labelling, Version 2.0. Ohio State
University.
Carlson, K., Clifton, Jr., C., Frazier, L., 2001. Prosodic boundaries in adjunct attachment. Journal of
Memory and Language 45, 5881.
Clifton, Jr., C., Carlson, K., Frazier, L., 2002. Informative prosodic boundaries. Language and Speech 45,
87114.
Fodor, J.D., 1998. Learning to parse? Journal of Psycholinguistic Research 27, 285319.
Gee, J., Grosjean, F., 1983. Performance structure: a psycholinguistic and linguistic appraisal. Cognitive
Psychology 15, 411458.
Gussenhoven, C., 2000. The lexical tone contrast of Roermond Dutch in optimality theory. In: Horne, M.
(Ed.), Prosody: Theory and Experiment. Kluwer Academic Publishers, Dordrecht, pp. 129167.
Hirotani, M., 2000. Prosodic constraints on Japanese scrambling. Poster presented at the 13th Annual
CUNY Conference on Human Sentence Processing, San Diego, 30 March1 April, 2000.
Nagel, H.N., Shapiro, L.P., Nawy, R., 1994. Prosody and the processing of ller-gap sentences. Journal of
Psycholinguistic Research 23, 473485.
Pierrehumbert, J. B., 1980. The Phonology and Phonetics of English Intonation. Unpublished PhD dissertation, MIT.
Pierrehumbert, J., Beckman, M., 1988. Japanese Tone Structure. MIT Press, Cambridge, MA.

L. Frazier et al. / Lingua 114 (2004) 327

27

Schafer, A.J., Speer, S.R., Warren, P., White, S.D., 2000. Intonational disambiguation in sentence production and comprehension. Journal of Psycholinguistic Research 29, 169182.
Selkirk, E., 1984. Phonology and Syntax: The Relation Between Sound and Structure. MIT Press, Cambridge, MA.
Selkirk, E., 2000. The interaction of constraints on prosodic phrasing. In: Horne, M. (Ed.), Prosody:
Theory and Experiment. Kluwer Academic Publishers, Dordrecht, pp. 231262.
Straub, K., Wilson, C., McCollum, C., Badecker, W., 2001. Prosodic structure and wh-Questions. Journal
of Psycholinguistic Research 30, 379394.
Truckenbrodt, H., 1995. Phonological Phrases: Their Relation to Syntax, Focus, and Prominence.
Unpublished PhD Dissertation, MIT.
Watson, D. G., Gibson, E., 2001. Linguistic structure and intonational phrasing. Paper presented at the
Fourteenth Annual CUNY Conference on Human Sentence Processing, Philadelphia, 1517 March
2001.
Watson, D. G., Gibson, E., 2002. When does prosody inuence parsing? Paper presented at the Fifteenth
Annual CUNY Conference on Human Sentence Processing, New York, 2123 March 2002.