Nonlocal effects of prosodic boundaries.
ABSTRACT Placing a prosodic boundary before a phrase may influence its syntactic analysis. However, the boundary's effect depends on the presence, size, and position of other, earlier, prosodic boundaries. In three experiments, we extend previous results about the effect of the position of the early boundary. In sentences in which a final phrase may modify either a local verb or an earlier verb, a boundary immediately after the first verb leads to more first-verb attachments than when the earlier boundary appears in another position between the two verbs (Experiments 1 and 2). This effect cannot be attributed to weaker effects of more distant boundaries (Experiment 2), but is likely due to the first verb being more prominent when a boundary immediately follows it, since similar effects are observed when the verb is accented (Experiment 3). The results support the informative boundary hypothesis and show that the impact of earlier, nonlocal boundaries is not fully uniform.
Non-local effects of prosodic boundaries
Katy Carlson, Morehead State University
Charles Clifton, Jr., University of Massachusetts Amherst
& Lyn Frazier, University of Massachusetts Amherst
English, Foreign Languages, and Philosophy Department
Morehead State University
Morehead, KY 40351
Running head: Nonlocal prosodic effects
Nonlocal prosodic effects 2
Placing a prosodic boundary before a phrase may influence its syntactic analysis. However, the
effect of the boundary depends on the presence, size, and position of other, earlier, prosodic
boundaries (Carlson et al., 2001, Clifton et al., 2002). Three experiments are reported that
extend previous results about the effect of the position of the early boundary. In sentences where
a final phrase may modify either a local verb or an earlier verb, a boundary immediately after the
first verb leads to more first verb attachments than when the earlier boundary is placed in some
other position between the first verb and the local verb (Experiments 1,2). This effect cannot be
attributed to weaker effects of more distant boundaries (Experiment 2), but is likely due to the
first verb being more prominent when a boundary immediately follows it, since similar effects
are observed when the verb is accented (Experiment 3). The results further support the
Informative Boundary hypothesis, and show that the impact of earlier, nonlocal boundaries is not
Nonlocal prosodic effects 3
There has been considerable interest in the question of how intonation, in particular prosodic
boundaries, influences the processing of a sentence (e.g., Beach, 1991; Kjelgaard & Speer, 1999;
Kraljic & Brennan, 2005; Lehiste, 1973; Nespor & Vogel, 1986; Price, Ostendorf, Shattuck-
Hufnagel, & Fong, 1991; Pynte & Prieur, 1996; Schafer, Speer, Warren, & White, 2000). The
vast majority of research has examined sentences in which the presence or location of a single
prosodic boundary affects the division of a sentence into syntactic phrases or affects the
attachment of one phrase to others.
Our previous work investigated the effect of multiple prosodic boundaries, and led to the
Informative Boundary Hypothesis (Carlson, Clifton, & Frazier, 2001; Clifton, Carlson, &
Frazier, 2002; cf. Schafer, 1997, for a position similar in some respects to ours). This hypothesis
states that listeners interpret a prosodic boundary before a constituent that could be attached to
either of two earlier constituents with reference to the presence and size of any relevant earlier
boundary. For example, in (1) after John visited can be attached to the phrase headed by
learned, where it modifies learned as in the paraphrase in (1a), or attached to telephoned, for the
meaning in (1b).
(1) Susie learned that Bill telephoned after John visited.
a. Susie learned (after John visited) that Bill telephoned.
b. Susie learned something, namely, that Bill telephoned after John visited.
If there are no relevant earlier boundaries, then a boundary before the ambiguously-attached
constituent discourages local (low) attachment (Price et al., 1991, and many others). That is, a
prosodic boundary immediately preceding after John visited may favor attachment to the high
attachment site learned rather than to the low attachment site telephoned. But, according to the
Informative Boundary Hypothesis, if a (relevant) earlier prosodic boundary is larger than the
Nonlocal prosodic effects 4
boundary after telephoned, then the prosody may still favor local attachment of after John visited
to telephoned, even though a prosodic boundary intervenes. The Informative Boundary
Hypothesis takes a ‘relevant’ earlier boundary to be a boundary that intervenes between the
possible low and high attachment sites.
To specify the size of a boundary, we assume a phonological system that distinguishes
word boundaries, intermediate phrase boundaries (ip), and intonational phrase boundaries (IPh).
An intonational phrase is the largest unit, and it contains one or more intermediate phrases. An
intermediate phrase must contain one or more words and at least one accented constituent. These
prosodic units (as well as prosodic words) are identified following the ToBI system, a prosodic
annotation procedure inspired by Pierrehumbert (1980) and explained in Beckman & Elam
(1997). The ends of both types of prosodic phrases (IPhs and ips) are associated with tonal
changes, increased duration, and optional pausing, with IPhs involving more extreme changes.
This theory gives us a 3-way distinction between prosodic boundary sizes: a word boundary is
smaller than an ip boundary, which is smaller than an IPh boundary. Syntactic boundaries also
differ in size, and that primarily depends on what syntactic phrases can be contained inside
others: the edge of an entire clause is a larger syntactic boundary than the edge of one of its
constituents, such as a noun phrase or a verb phrase.
The detailed motivation for the Informative Boundary Hypothesis and some discussion of
possible mechanisms are laid out in our earlier papers (Carlson et al., 2001; Clifton et al., 2002;
cf. Frazier, Carlson, & Clifton, 2006, for general discussion), but briefly, we view the presence
of a prosodic boundary as evidence that a listener must interpret with respect to the local context,
including the presence, size, and location of certain other prosodic boundaries. Our favored
justification for this context-sensitivity is that a listener assumes that speakers generally make
Nonlocal prosodic effects 5
their prosodic phrasing decisions conform to the intended structure and meaning of a sentence,
differing primarily only where phonetic considerations dictate. For example, if a speaker reaches
a syntactic boundary that is smaller than a previous syntactic boundary, the speaker will not
capriciously produce a prosodic boundary that is larger than the prosodic boundary at the end of
the larger constituent.
The Informative Boundary Hypothesis has successfully predicted processing effects in
various syntactic structures, shown in (2), with the relevant early and late boundary positions we
tested marked as #1 and #2 (Carlson et al., 2001; Clifton et al., 2002):
(2) a. Susie learned #1 that Bill telephoned #2 after John visited.
b. Old men #1 and women #2 with very large houses
c. Johnny #1 and Sharon’s #2 in-laws
d. The daughter #1 of the Pharaoh’s #2 son
e. I met the daughter #1 of the colonel #2 who was on the balcony.
In these sentences, a prosodic boundary of a given phonological size at position #2 produced
different results depending on whether this boundary was equal to, smaller than, or larger than a
prosodic boundary at position #1.
We propose that these effects arise not because prosodic boundaries send special signals
to the syntactic processor (e.g. “end a syntactic constituent here;” cf. Marcus & Hindle, 1990),
but because they create prosodic phrases that affect the likelihood of attachment. We claim that
the processor prefers to syntactically group nearby constituents and constituents that are
prosodically grouped together, if possible. If a constituent is in a different prosodic phrase from
the nearest attachment site, it is less likely to attach locally than a constituent that is in the same
prosodic phrase as the potential attachment site. The constituent is especially likely to attach
Nonlocal prosodic effects 6
non-locally if the boundary that marks the end of the prosodic phrase containing the attachment
site is larger than any other prosodic boundaries in the relevant area. In (2c), for example, an ip
boundary at position #2 would usually inhibit attachment of in-laws to Sharon’s and favor
attachment to the higher phrase Johnny and Sharon’s. If an IPh boundary is also present at
position #1, though, then Sharon’s and in-laws are more closely grouped prosodically than either
one is to Johnny: [(Johnny)ip]IPh [(and Sharon’s)ip (in-laws)ip]IPh and low attachment becomes
more attractive again (making the in-laws relatives only of Sharon). However, an earlier prosodic
boundary that does not intervene between the potential attachment sites does not modulate the
effect of the later boundary. Clifton et al. (2002) showed this for sentences like (2a), where a
prosodic boundary after Susie (instead of after learned) did not reduce the effect of a boundary
The Informative Boundary Hypothesis denies the simple and attractive suggestion that a
prosodic boundary is a local cue, signaling a listener not to attach an upcoming phrase to the
immediately preceding material (Marcus & Hindle, 1990; Watson & Gibson, 2005). Marcus and
Hindle state that “…major intonation boundaries are introduced into the lexical stream fed to the
parser and…the parser knows nothing about them…Upon encountering an utterly unkown
‘lexical’ item [the boundary], a deterministic parser simply terminates the constituent at hand….”
(page 495). Watson and Gibson explicitly propose the “Anti-Attachment Hypothesis (AAH):
Listeners prefer not to attach an incoming word to a lexical head that is immediately followed by
an intonational boundary” (page 285).1 In the work we report here, we explore a limitation on the
1 Lee and Watson (2009) have recently presented evidence that in directive sentences like Click on the candle below
the triangle that’s in the blue circle, the effect of a prosodic boundary before that’s on the blue circle is not affected
by the size of a prosodic boundary after candle. They argued that this finding contradicts the Informative Boundary
Hypothesis. We are not yet able to explain why Lee and Watson’s results differed from the results we obtained with
a similar relative clause attachment ambiguity (2e, above), and we view the Lee and Watson finding as a not-yet-
understood limitation on the hypothesis.
Nonlocal prosodic effects 7
Informative Boundary Hypothesis, namely, the effect of placing an earlier prosodic boundary at
different positions between the possible attachment sites. Even though these different positions
are still “relevant,” according to our hypothesis, it turns out that they have somewhat different
effects, which we explore in the present research.
In this paper, we study a structure with two relevant early boundary positions, illustrated
(3) Jimmy comforted #1 the girl #2 that he had insulted #3 at the end of the party.
In this structure, the temporal adjunct phrase at the end of the party could attach low, to the verb
insulted inside the relative clause, or high, to the first verb comforted. Syntactic structures for
these options are shown in Figure 1.
Syntactic structures for high (a) and low (b) attachment of the adjunct
V DP PP
comforted at the end of the party
he had insulted
Nonlocal prosodic effects 8
had insulted at the end of the party
On any theory that considers only a local prosodic boundary, such as Watson & Gibson’s (2005)
Anti-Attachment Hypothesis, a boundary at position #3 should discourage local attachment of
the adjunct regardless of the prosodic boundaries that might occur elsewhere in the utterance.
According to our Informative Boundary Hypothesis, the impact of a prosodic boundary at
position #3 (preceding the adjunct phrase) should be reduced by an equal or larger prosodic
boundary at either of the two preceding positions, #1 or #2. The presence of either early
boundary should thus reduce the frequency of high attachments.
We report three experiments designed to understand the effects of boundaries at these
different positions. Experiment 1 compares the impact of two relevant early boundaries: one
immediately following the high attachment site (the first verb), and one two words later, after the
noun phrase that is the direct object of the first verb (the experiments in Carlson et al., 2001, and
Clifton et al., 2002, examined only early boundary positions directly after the high attachment
site, with sentences that did not contain two relevant prosodically-acceptable positions for
Nonlocal prosodic effects 9
boundaries). Experiment 1 indicated that the impact of the two relevant early boundaries
differed, a finding not predicted by the Informative Boundary Hypothesis. Experiment 2 was
designed to explore one account of this finding, namely, that the distance between the early
boundary and the later boundary affects the impact of the later boundary. Experiment 3 tested the
hypothesis that the effect of the boundary immediately following the high attachment site was
due to making that site prominent and salient.
To anticipate the results, we claim that our evidence indicates that a boundary
immediately after an attachment site makes that site more salient and more attractive as an
attachment site. This claim lets us explore the joint effects of prosodic boundaries and
prominence on attachment decisions. Previous research has shown that the attachment of a
constituent can be influenced not only by prosodic constituency, but also by prosodic
prominence as manipulated by pitch accents (Beckman & Elam, 1997). Schafer, Carter, Clifton,
& Frazier (1996) studied the effects of pitch accents in the resolution of relative clause
attachment, using phrases such as the propeller of the plane that was so shiny…. The relative
clause could attach to the nearest noun, plane, or the more distant noun, propeller. Accenting
either noun led to increased attachment of the relative clause to that noun, and the increase was
larger if the accent was contrastive (i.e. a particularly high and steep accent, labeled as L+H*).
Because prosodic boundaries often follow the most important accent in a phrase (Ladd,
1996; Selkirk, 1995), we hypothesize that the association between boundaries and accents might
encourage a listener to interpret a prosodic boundary immediately after an attachment site as
highlighting that site. That is, the boundary might give the impression of prominence even
without a preceding accent. If this is true, then the Schafer et al. (1996) results suggest that when
the first verb in our sentences (comforted in (3)) is immediately followed by a prosodic boundary
Nonlocal prosodic effects 10
(position #1), it will be perceived as having increased prominence, and will attract attachment.
Therefore, more high attachments will be observed when an early prosodic boundary follows the
first verb (#1) than when it occurs after a non-attachment site (#2).
In this first experiment, we explored the effects of single and multiple prosodic boundaries on the
attachment of temporal adjuncts in ambiguous sentences with restrictive relative clauses. We
predicted that a single early prosodic boundary would result in numerous low attachments, and a
single late boundary would result in frequent high attachments, while equal early and late
boundaries would lead to intermediate results.
We created sixteen sentences, all ending in final temporal or locative phrases that could attach
high or low. A full list of the stimuli is in Appendix 1. These sentences were produced in four
prosodic conditions, as in (4), with Intonational Phrase boundaries at differing locations
(indicated by //). IPh boundaries can involve different final tonal movements. The boundary
tones we used are those that are most consistent with sentence-internal phrases and form what is
called a continuation rise, consisting of a low pitch target followed by a higher one (labeled as L-
(4) a. EarlyN: Jerry kissed the actress // that he had recently met at the Oscars ceremony.
b. Late: Jerry kissed the actress that he had recently met // at the Oscars ceremony.
c. EarlyV+Late: Jerry kissed // the actress that he had recently met // at the Oscars
d. EarlyN+Late: Jerry kissed the actress // that he had recently met // at the
Nonlocal prosodic effects 11
L+H*!H* L-H%H* L+H*L-L%
Jerrykissedthe actressthat he hadrecently met at the Oscarsceremony
The speaker, a native English speaker with training in ToBI analysis and experience recording
experimental sentences, recorded the sentences and transferred them into a computer, digitized at
16 bits at a 22050 Hz rate. They were examined in PRAAT (Boersma, 2001) to see that the
intended prosodic contours had been achieved. Any anomalous sentences were re-recorded and
analyzed. Pitch tracks for an example sentence can be seen in Figure 2, and average F0 and
duration measurements for critical sentence positions are in Tables 1-2.2 All sentences showed
the L-H% contour diagnostic of an IPh boundary at the appropriate locations.
Pitch tracks for a sample item from Experiment 1
2 Although a prosodic boundary after the head noun of a relative clause (as in conditions a and d) is required for
non-restrictive relative clauses, it is also possiblefor restrictive relative clauses. The relative pronoun that, found in
most of the relative clauses in these sentences, unambiguously indicates that they must be restrictive. It’s also
unclear how attachment of a later constituent would be affected by the status of the relative clause in any case.
L+H*!H*H* H* L-H%L+H*L-L%
Jerrykissedthe actress that he had recentlymetat the Oscarsceremony
Nonlocal prosodic effects 12
L+H* L-H%H*H*L+H* L-H%L+H* L-L%
Jerry kissedthe actressthat he hadrecently metat the Oscarsceremony
L+H*!H* L-H%H*L+H* L-H% L+H*L-L%
Jerry kissedthe actressthat he had recently met at the Oscarsceremony
Nonlocal prosodic effects 13
Average Fundamental Frequency (F0) measurements, in Hz, for boundary positions,
V1 peak V1 L-H% N1 peak N1 L-H% RC peak RC L-H%
EarlyN 279 (252) 240 141, 191 214 (198)
Late 276 (259) 243 (221) 214 143, 189
EarlyV+Late 231 147, 200 218 (195) 244 143, 194
EarlyN+Late 267 (253) 238 141, 188 242 140, 185
Short EarlyV 256 156, 209 313 (253) 247 151, 207
Short EarlyN 294 (262) 256 147, 196 259 146, 203
Long EarlyV 275 150, 202 306 (243) 233 146, 203
Long EarlyN 300 (269) 258 144, 196 246 145, 199
Note--Parenthetical numbers indicate F0 averages over the few sentences which showed a discernible drop from the
peak to the F0 at the end of the word or constituent. In the other conditions, the two numbers indicate the successive
low and high targets of the continuation rises (L-H%, in ToBI transcription). V1 = first verb; N1 = noun modified by
the relative clause; RC = relative clause.
Nonlocal prosodic effects 14
Average duration measurements, in ms, for boundary positions, Experiments 1-2
V1 N1 V2
EarlyN 388 763 483
Late 396 397 933
EarlyV+Late 776 366 918
EarlyN+Late 360 760 920
Short EarlyV 721 373 796
Short EarlyN 367 690 818
Long EarlyV 720 370 813
Long EarlyN 360 687 823
Note—Duration measurements include the duration of the indicated words (the first verb, the noun modified by the
relative clause, and the second verb) plus any following pause.
For each sentence, a follow-up question and two possible answers were created. The
answers clarified the attachment of the adjunct phrase, as shown in (5) with added parentheticals:
(5) What happened at the Oscars ceremony?
a. Jerry met the actress. (low attachment)
b. Jerry kissed the actress. (high attachment)
The same speaker who recorded the experimental sentences also recorded 104 other sentences
that the subjects were to hear. These sentences were a mixture of sentences with ambiguously-
interpreted ellipses and sentences with lexical ambiguities.
Nonlocal prosodic effects 15
The participants were 28 native English-speaking students at Northwestern University who
reported normal hearing and vision. They received course credit in an introductory linguistics
class for participating in the experiment.
Participants were seated wearing headphones at computer stations within a sound-proof booth.
They pressed a button on a keyboard to hear each sentence as it was played by a computer. After
the sentence ended, they read a corresponding question and answers on a written questionnaire
sheet. They were to circle the answer that best fit how they understood the sentence, and then
press a button to hear the next item. A short practice round familiarized them with the procedure
and allowed them to adjust the volume, if necessary. The entire procedure took less than ½ hour.
The 16 experimental sentences (together with the 104 other sentences) were divided into
four sets and assigned to four counterbalanced lists. Each counterbalanced list contained the four
sentences of a set in one of the four prosodic conditions illustrated in (4), so that each list
contained four sentences in each condition and each sentence appeared in all four forms across
the lists. Seven subjects heard each of the counterbalanced lists in a different pseudo-randomized
order with no consecutive trials containing sentences of the same type.
The results are shown in Figure 3.
Percentage of high attachments, Experiment 1, with by-items Standard Error bars