I. 0.»: 1‘0
4

 

a... 1.2. .
t. «1a. .n 1.9.2114.
Emu: “K1533... n ,
.T e a? .. :.
.5: i, ,
a a a .

 

 

‘4‘ 1 .nfi
. .l

 

 

¢ln .-
etyaa, :

...:m.. _ . ‘ . «infihm
0‘ . , , 5.1. Q‘s-c'uutl ¢ ell; {III ..

 

 

   

            
 

 

   
   

   
   
              

agafirgai 2: $3 .. 5:51;; ‘ . #5:," . that?

 

ll 'r I‘l I I III, .4 '
I D

[r

 

 

Date

MICHIGAN STATE UN resumes
l
l

mm mm: III I II I! I II

’I ,
I l
3 1293 01562 1182

ll

 

 

IVEFISI
l
l

 

 

 

 

 

 

 

 

LIBRARY
Mlchlgan State
Unlversity

 

 

 

This is to certify that the

thesis entitled

PROCESSING PARALLEL STRUCTURE IN
READING AND LISTENING

presented by

MICHAEL D. ANES

has been accepted towards fulfillment
of the requirements for

MA degree in Psychology

 

 

KW [Ml/u Aer;

Major professor

Fernanda Ferreira

 

12/16/96

 

0—7 639

MS U is an Affirmative Action/Equal Opportunity Institution

PLACE II RETURN BOX to roman thh checkout mm your moord.
TO AVOID FINES return on or baton date duo.

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

PROCESSING PARALLEL STRUCTURE IN READING AND LISTENING
By

Michael D. Anes

A THESIS
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

MASTER OF ARTS

Department of Psychology

1996

ABSTRACT
PROCESSING PARALLEL STRUCTURE IN READING AND LISTENING
By

Michael D. Anes

Following a sentence fragment such as Sherlock Holmes noticed, the
phrase the man's shoes is syntactically ambiguous. The phrase could be a direct
object ending the sentence or it could start an embedded Clause, the man's
shoes were muddy (a sentence complement construction). Experiments were
conducted to determine the beneficial effects of repeating the above temporarily
ambiguous structures in reading and listening. Contrary to previous findings,
repetition of the object construction resulted in greater processing benefit than
repetition of the sentence complement contruction in both self-paced reading
and self-paced listening tasks. In the listening task, repetition of the sound, or
prosodic, structure had no influence. Sensibility ratings by another subject group
showed that the object constructions were less sensible than sentence
complement constructions, providing a possible explanation for the processing
differences. The role of semantic, syntactic and prosodic variables in the

processing of repeated structures is discussed.

Dedicated to Louise J. Anes, whose life and work will always be a part of me.

iii

ACKNOWLEDGMENTS

I would like to thank the members of my committee for their help and
constructive criticism; Fernanda Ferreira, John Henderson, Barbara Abbott, and
particularly Rose Zacks for her expert guidance through the last stages of this
project. The love and patient support of Ann Garrett-Anes has enabled me to

complete this work, and I am very grateful.

iv

TABLE OF CONTENTS

LIST OF TABLES

LIST OF FIGURES

INTRODUCTION

I. Is Prosody Used IH'TAQIQBIGQHAQ 'és'r'rigééiiér'I's'IBHJéfff.
II. ls Prosody Used in Parsing Tasks?......
Ill. Processing Parallel Structure

EXPERIMENT1
I. Method
II. Results

Ill. DISCUSSIon ..

EXPERIMENT2
I. Method
II. Results

Ill. DISCUSSIon

GENERAL DISCUSSION ....

l. Structural Effects in the éiééiifiéaiéLlll2212'.IZ'.'.'.'.'.'.'.’.I'.'.'.'.'.2'.
ll. Prosodic Effects: Where are They?... ..

APPENDIX A

Experimental Sentences..................

APPENDIX 8
Main Verb Sentence Complement l

Noun Phrase Preference Normative Judgments ...................

APPENDIX C

Tables

APPENDIX D

Figures..................

vii

viii

.30)“;

62

65

71

ListofReferences.........................

76

LIST OF TABLES

TABLE 1
Frazier et al. (1984) - Experimental Sentence Conditions....... 66

TABLEZ
Experimental Sentence CondItIons 67

TABLES
Experimental Conditions - Experiment 2... 68

TABLE 4
Correlations of Difference Time and Sensibility Judgments... 69

TABLE 5
Correlations of Difference Time and Sensibility Judgment Parallel Benefit..70

vii

LIST OF FIGURES

FIGURE 1
SGnSIbIIItyJUdgments 71
FIGURE 2

Experiment 1 Reading Time-Second Clause Disambiguating Region ........ 72
FIGURE 3

Experiment 1 Reading Time-Second Clause lnterchanging Region .......... 73
FIGURE 4

Stimulus Sentence Construction-Experiment 2... 74
FIGURE 5

Experiment 2 Structural Parallelism Benefit-Second Clause

lnterchanging Region......... ..........75

viii

Introduction

Prosody is a set of suprasegmental phonological characteristics of
speech, such as amplitude, duration, and pitch, that speakers produce in
addition to segmental features making up the words of an utterance. These
attributes are perceived by listeners as word and sentence stress, rhythm, and
intonation. In the following sections, evidence that the prosodic characteristics
of sentences are identified by listeners is reviewed, as are the results of
research of the relationship between syntactic and prosodic structure in spoken
language. In these studies, sentences with two Clauses were manipulated so
that Clauses had the same or different syntactic and prosodic structures, and
the processing benefit of parallel syntactic and prosodic structure was
measured. The objective of the present experiments was to determine if
prosody is immediately used by comprehenders to assign syntactic structure to
what they hear.

Since Lehiste's (1970) pioneering work, investigators have been
concerned with the question of the reliabilty of prosodic information in spoken
language and the potential benefit any regular cues might confer in
comprehension. If regularities are found in the way in which speakers render a
given syntactic structure prosodically, researchers can manipulate recordings
of these structures and employ a comprehension task that, it is hoped, can
discover aspects of performance contingent upon prosodic cues in the speech
stream. In this way, we can begin to understand how and when comprehenders

register such information.

The first focus, then, has been to submit productions of speech to
acoustic analysis. These productions are obtained in various contexts.
Sometimes they are recordings of spontaneous speech, though they more
often consist of readings of sets of sentences selected specifically to be of a
certain phonological and syntactic structure. Cooper (1980), Lehiste (1970),
and Pike (1945), have shown that the fundamental frequency (intonation) of a
broad class of declarative sentences peaks early in phrases and falls at the
ends of phrases. F erreira (1993) has provided an account of speakers' phrase-
final segmental lengthening and pausing found in descriptive phrases such as
THE GREEN TABLE. These pitch patterns may be a source of information
listeners use to determine the syntactic structure of what they hear.

Some researchers (Beach, 1988; Cooper and Sorensen, 1981) have
found distinct pitch contours in different syntactic structures, including
sentences in which a phrase is temporarily ambiguous in its syntactic
constituency. In the temporarily ambiguous structure below, the bracketed
portion of the sentence could be a direct object of the verb noticed, or could be
the start of a new sentence complement. In 1b), the presence of the verb were
disambiguates the preceding Noun Phrase (NP) as the start of a complement
sentence.
1a) Sherlock Holmes noticed [the man's muddled brown shoes] NP Object

1b) Sheltock Holmes noticed [the man's shoes were muddy] Sentence Complement

Pitch and pausing cues to syntactic structure in production are intriguing, yet
precisely how comprehenders use prosodic information is still an open
question.

One major consideration in the present study was the task to be used.
Methods which reveal processing implications of stimulus changes as they
occur in real-time, with a minimum of intrusion, are considered "on-line". On-
Iine methods should show, for example, effects of integrating disparate
information types or of processing material of differing complexity. Effects
should be seen soon after presentation of the relevant input. Off-line methods,
however, may show effects of stimulus changes but do not give an accurate
account of their time course. A new, on-Iine sentence processing task, the
Auditory Moving Window (Ferreira, Henderson, Anes, Weeks, Jr., 8
McFarlane, 1996; Ferreira, Anes, & Horine,1996), is used in the present study
to examine the possible influence of both syntactic and prosodic structure in
sentence processing.

I. Is Prosody Used in Tasks of General Comprehension?

The presence of prosodic information has been shown to facilitate
predicting sentence length in a sentence gating task (Grosjean, 1983), where a
portion of a sentence was played to participants and they were asked to press
a button when they thought the sentence would end. In addition, detecting
word-initial phoneme targets is speeded as prosodic information increases with

sentence length (Shields, MCHugh, and Martin, 1974), when targets are

preceded by appropriate prosodic contours (Cutler, 1976) and as semantic
focus (in the form of prosodic cues such as lexical stress) is given to a lexical
item (Cutler and Fodor, 1979). All of these results have been taken as
evidence that prosodic information is identified and used in some general
aspects of sentence processing, such as preparing for a specified amount of
upcoming information or to become aware more quickly of semantically salient
portions of a sentence.

Speer, Crowder, and Thomas (1993) investigated a claim, originally
made by Slowiaczek (1981 ), that prosodic structure helps to package
constituents in short-term memory, and that prosodic structures are stored in
long-term memory. Speer et al. recorded two versions of structurally ambiguous
sentences, called syntactic-Change sentences, each with-appropriate prosody.
Contrastive stress was added to two different words in a new set of recordings
to create semantic focus-change sentences, where syntactic structure does not
change. A set of recognition sentences was created, holding either prosody or
words constant. Subjects listened to sentences and judged each as new or old
(they had/had not heard the words before), and gave a confidence rating for
their judgment.

Recognition accuracy in matching conditions (previously presented
words and prosody) was significantly greater than in mismatching conditions
(new words or prosody) for both syntactic— and focus-change sentences. The

cost of a mismatch was the same for both sentence types, and similar results

were found with nonsense syllable strings in presentation and recognition sets.
Speer et al. claimed that the recognition process was not based on sentence
content alone — prosody was maintained in memory, providing a distinct
recognition advantage to same prosody recognition probes. Memory for
prosody of focus-change stimuli, where syntactic structure does not change,
was as great for syntactic-Change stimuli, supporting Speer's earlier assertion
(Speer and Slowiaczek, 1990) that prosody is not a signal of syntactic structure
in one-to-one correspondence.

Speer et al.'s (1993) results are particularly important for researchers
interested in prosody in parsing. A single prosodic production of a syntactic
structure may or may not conform to a mean pattern if one exists, variability
that may be moderated by semantic attributes of the utterance. More
importantly, one prosodic structure may provide varying degrees of evidence
for differing syntactic structures.

One way to combat the inherent difficulties in investigating a processing
system that is faced with so much uncertainty is to focus on specific situations
such a system might face. In investigating prosodic effects in comprehension,
researchers have attempted to reduce uncertainty by creating rigid syntactic
environments conforming to theoretical notions of processing ease or difficulty,
to eliminate lexical differences to the greatest extent possible, and to allow
prosodic effects to have an influence. These were the goals of the research

described below, where prosodic cues to syntactic structure were manipulated.

ll. ls Prosody Used in Parsing Tasks?

For years, researchers have examined the processing of structurally-
ambiguous sentences in order to understand the mechanisms underlying
syntactic structural assignment. Parsing is the process by which
comprehenders identify constituent parts of a sentence and attach them into a
structure compatible with the grammar of the language. By forcing
comprehenders to encounter sentences with more than one syntactic stmctural
alternative at a point in processing (a common natural occurrence, not limited
to experimental situations), and observing comprehension effects, it is
assumed that fundamental linguistic processes can be discovered.

Because there are an arbitrary (and perhaps large) number of possible
constituents intervening between ambiguity and disambiguation in sentences
with temporary structural ambiguities, Frazier (1979) proposed two principles
by which a single, simple syntactic analysis is pursued. If the parser attempts to
incorporate a constituent and the resulting structure is ungrammatical,
reanalysis will be necessary. Frazier‘s model is often called the ”garden-path"
model, because of the possibility of initial, incorrect stmctural commitment.
Frazier‘s first principle is Minimal Attachment, where the parser attaches
syntactic constituents into a phrase structure tree using as few nodes as
possible consistent with the grammar of the language. Under Minimal
Attachment (MA), an NP that could be an object of a verb or a subject of an

embedded clause should be incorporated as a direct object, because a phrase

structure tree so constructed will have fewer nodes. If an NP is ultimately
incorporated as part of a the start of a Sentence Complement (S-Comp), the
sentence as a whole is termed Non—Minimal Attachment (NMA). Frazier‘s
second principle is Late Closure, where the parser attempts to attach
constituents to the currently open phrase, rather than create a new one or
modify an old structure. Under Late Closure, a noun phrase that could be the
direct object of a verb or the subject of an embedded clause should be
assigned as an object, because the verb phrase can still accept incoming
material. The result of applying these two principles is the same; the parser
constructs the simplest possible analysis. Late Closure is potentially operable
in sentences where there is a main clause, subordinate clause relationship
between the ideas expressed. Early (sentence 2a below) or Late Closure
(sentence 2b below) sentences often begin with a subordinating conjunction,
and usually contain a comma to separate the clauses. ln Minimal Attachment,
phrases do not exhibit semantic or syntactic dependence in the same way, and
the ambiguity is not resolved with punctuation.
2a) Because Sam left, the party seemed dull. Early Closure
2b) Because Sam left the room, the party seemed dull. Late Closure

Many researchers (Ferreira and Clifton, 1986; Ferreira and Henderson,
1990; Frazier and Rayner, 1982) have tested the garden-path model and the
application of heuristics such as Minimal Attachment. Using eye-movement

monitoring, long reading times on the disambiguating word of nonminimal

attachment sentences have been found, as well as a pattern of regressions to
early sentence regions. These authors suggest that long reading times at
disambiguation are the result of misanalysis, and that regressions are the
parser’s recovery process, indicative of a restructuring of phrases. Most of the
research and theory-building done in the 1980's in parsing was based on
evidence from reading, although the goals of this work were not to characterize
modality-specific processing principles. Only recently have researchers begun
to investigate the effects of prosody in parsing.

Price et al. (1991) conducted match-to-context experiments where
subjects listened to a member of a sentence pair and chose the context from
which it came. Sentences with ambiguous prepositional phrase (PP)
attachment sites were presented auditorily to participants, an ambiguity that is
often subsumed under the principle of minimal attachment. Far attachment PPS
(MA) were disambiguated slightly better than chance. For example, in sentence
3 below, either Laura or the man could be wearing the robe. If the man is
wearing the robe, near PP attachment to the man as a modifier is a more
complex structure than far attachment to the main verb, where Laura wears the
robe.

3) Laura ran away with the man wearing the green robe.
Beach (1991) investigated the NP object I Sentence Complement

ambiguity by presenting sentence fragments in a matching task. Fragments

were of two lengths: material up to an ambiguous NP, and including part of an
ambiguous NP. An example sentence is presented below.

4a) The store manager knows/ peculiar II customers and their personal problems.

4b) The store manager knows! peculiar I/ customers can cause lots of work problems.

The first slash in sentences 4a and 4b marks the position up to which
sentences were presented in short versions, and the double slash marks the
presentation including part of the ambiguous NP. Synthesized speech was
used to control pitch rise and fall and segment duration precisely. Subjects
matched fragments they heard (which had a prosodic pattern typical of the
structure) to continuations of the sentences at a level better than chance.
Fragment length did not affect matching performance. In another experiment
duration and pitch values were independently varied (in the italicized syllables
above) along a 5-point continuum so that segments were presented with
constant pitch, or with varying degrees of pitch fall at the verb (knows) and rise
in the next stressed syllable (in peculiar). Pitch and duration values in the
middle of the continuum were presumed to provide ambiguous syntactic
structural evidence, and those on the endpoints were thought to provide clear
evidence for a single structure. Beach argued that duration and pitch effects
indicated both information sources influence structural identification. Pitch and
duration interacted such that an information source was especially effective
when the other source was ambiguous (in the middle of the continuum). Beach

concluded that "there is no 1:1 threshold mapping between one type of

10

prosodic pattern and identification" (p. 658) and that a probabilistic relationship
exists between the registration of a (possible) prototypic prosodic pattern and
matching the pattern to a syntactic representation.

In summary, testing listeners' ability to use prosody to disambiguate
structurally ambiguous material has been done largely with matching tasks (to
paraphrase or to context), but these tasks cannot uncover the time course of
registering prosodic information. Researchers see effects of duration and pitch
changes upon structural disambiguation, and even that segment duration and
pitch change prosodic components "trade-off" in the extent to which they
provide evidence for a syntactic structure, but we are unable to determine how
and when these variables are registered by the comprehension system. The
same criticism applies to other tasks recently used to study prosodic effects in
comprehension (with mixed results), such as end-of-sentence probing
(Kielgaard and Speer, 1993) and mispronounciation monitoring tasks (Watt and
Murray, 1993).

Recently, more on-line methods have been used to study the effects of
prosody in parsing. Marslen-Wilson, Tyler, Warren, Grenier, and Lee (1992)
recently investigated prosodic effects in MA and NMA sentences using a cross-
modal naming task. In this task, participants hear a portion of a sentence and
name a visually-presented continuation word that appears at the offset of the
auditory material. Fast naming times in comparison to a baseline are presumed

to reflect an ease of incorporation of the continuation word. The authors

11

recorded auditory sentences in natural prosodic renderings, such as Sa-c
below
58. The workers considered that the last offer from the management“ was a real insult.
5b. The workers considered the last offer from the management“ was a real insult.
50. The workers considered the last offer from the management“ of the factory.
With the inclusion of the complementizer that, Sentence 5a is syntactically
unambiguous. Sentences 5b and 5c are ambiguous up to the word after the
caret shown in the sentences. Sentence 5b is NMA without a complementizer,
and sentence 5c is MA. Marslen-Wilson et al. reviewed evidence of fast
reading times on and after the post-verbal NP in NMA sentences with
cornplementizers, indicating the effectiveness of overt lexical cues in reading.
The authors suggested that if prosody functioned similarly to the lexical cues,
there should be no naming time difference to an appropriate visual probe word
continuation (WAS at the caret in the sentences above) in the NMA without
Complementizer (NMA - Comp) and the NMA with Complementizer (NMA +
Comp) conditions. Naming times to targets in neutral sentences (simple
declarative sentences containing the word WAS) and to inappropriate
continuations (violations of grammatical number, eg. WERE at the caret
position) were baselines. Subjects also rated the appropriateness of the
continuation word.

Marslen-WIlson et al. (1992) found that naming times to continuation

words were significantly faster in the NMA + Comp and NMA - Comp conditions

12

compared to the neutral carrier baseline condition (10 and 11 ms faster,
respectively). In the MA condition, where the target is an inappropriate
continuation, naming time was 14 ms above the neutral carrier baseline, similar
to the grammatical number violation condition (24 ms over baseline).
Attachment conditions did not differ statistically in rated appropriateness,
although Marslen-Wilson et al. took numerically lower MA than NMA
continuation ratings as a registration of some conflict between prosody and the
continuation. The rating of grammatical number violation (inappropriate
continuation) was near zero. The prosody of NMA sentences was as effective
as a complementizer, and the authors suggested that prosodic representations
are created quickly enough to be used on-line in processing.

Another procedure that can be used to provide on-line information about
the role of prosody in comprehension is the Auditory Moving Window task
(AMW), developed by Ferreira et al. (1996).The AMW is an auditory analog of
self-paced word-by-word reading tasks (Just and Carpenter, 1980). In one of
Ferreira et al.'s experiments, high and low frequency words were embedded in
identical sentence frames, and the AMW was sensitive to lexical frequency on
the critical segment. in other sentences, immediate effects of a syntactic
garden path were seen in reduced relative “sentences compared to active
sentences that were identical until disambiguation. Mismatching prosody
conditions were created by splicing, and post-disambiguation words were

listened to longer if the preceding intonation contour was inappropriate for the

13

structure than if the contour was appropriate. All of these findings indicate the
AMW might be a good task to use when manipulating prosodic and syntactic
structure independently.
Ill. Processing Parallel Syntactic Structure

One way to study the parsing of temporarily ambiguous syntactic
structures in the visual domain has been to conduct self-paced word-by-word
and eye-movement monitoring investigations of the reading of these structures.
in spoken language comprehension work, investigators have also presented
sentences to subjects and monitored their performance on a variety of
measures, which, as discussed, differ in the extent to which they can uncover
the real-time nature of syntactic comprehension. Early on in the investigation of
syntactic comprehension, researchers hit upon the simple idea that
performance on a task may be enhanced if the syntactic structure under
investigation is repeated. While relatively few studies have had as their aim the
question of whether repeated occurrences of a syntactic structure result in
facilitated comprehension, the notion that the processing of syntactic structures
can be "primed" by prior presentations is enjoying a bit of a renaissance in the
present literature.

In the initial investigations of syntactic priming, Mehler and Carey (1967)
repeatedly presented tape-recorded sentences of the same syntactic structure

to subjects, whose task was then to perceive a target sentence in noise.

14

Perception was difficult for sentences with an unexpected structure, but not for
sentences of the same structure as the preceding items.

Frazier, Taft, Roeper, Clifton, and Ehrlich (1984) further investigated the
facilitated processing of repetitive (or parallel) structures. These researchers
used several types of conjoined constructions exhibiting two form variations, in
a first clause/second clause self-paced reading task. Variations were syntactic
(passive/active, NMA/MA, and shifted/nonshifted heavy NP) and semantic
(agent/theme, animate agent/inanimate agent). Example sentences of the
syntactic form conditions from Frazier et al. are shown in Table 1.

Frazier et al. (1984) found facilitative effects of parallel structure for all
the form variants tested. Facilitation was computed by subtracting the reading
time per word of the second clause from the reading time per word of the first
clause. In the syntactic manipulations, there were substantial parallelism
effects for conjoined active sentences but not for conjoined passives, conjoined
NMA sentences but not conjoined MA sentences, and for conjoined shifted NP
sentences but not conjoined unshifted NP sentences. Semantic manipulations
resulted in strong parallelism effects as well.

More recently, parallelism effects in a variety of sentence structures
have been examined. MacDonald and Schuster (1994) reported results from
self-paced reading studies indicating that ambiguous reduced relative clauses
are comprehended more easily following both morphologically unambiguous

reduced relative clauses (using participle forms like shown) and unreduced

15

relative clauses. Branigan, Pickering and Stewart (1995) tested a host of
syntactic structures in parallel and non-parallel environments, including
reduced relative-main Clause ambiguities (with the same results as MacDonald
and Schuster [1994]) and NP objectlS-complement ambiguities (MA/NMA). The
NP object/S-complement ambiguity did not result in any significant parallel
effects, despite the use of full-clause self-paced reading, as used by Frazier et
al. (1984). Branigan et al. suggest that although NMA structures are more
syntactically complex and thus are initially less preferred by the parser than MA
structures, NMA structures may cause too weak of a garden path for the parser
to benefit much from initially postulating an NMA analysis of a second clause.
Alternatively, reanalysis may occur rather quickly in such structures and
parallel structure benefit is observed only in more difficult cases.

The following studies are intended to address issues of concern in
previous syntactic priming research. One point of some importance is whether
the MAINMA ambiguity is fruitful to investigate given the history of equivocal
processing results. The present experiments are an attempt to extend the
reach of prior investigation to include another type of structural repetition,
namely of prosodic structure, and to fix methodological problems in Frazier et
al. (1984). Among the methodological problems in the Frazier et al. work is the
fact that full clause reading times do not provide information on how quickly
mechanisms that result in facilitation are activated. Also, in Frazier et al., MA

regions are much shorter than NMA regions, so analyses of the second Clauses

16

are conducted after comprehenders have received differing amounts of
material, and not only material of differing syntactic complexity. Finally, in
Frazier et al., lexical frequency of the constituent words in the critical regime is
not controlled. With analysis regions controlled on these factors, full region
reading times, as well as those for shorter segments, may be compared.

The present study allowed comparisons of processing time on second
clauses of two-clause sentences that were parallel or non-parallel in structure.
The first experiment is intended to replicate Frazier et al. (1984) using self-
paced segment-by-segment reading and better-controlled stimuli. The AMW
was used in the second study, using as a dependent measure the amount of
time subjects listened to a segment after its offset. Ferreira et al. (1996)
suggested that as there is no necessarily linear relationship between the length
of an auditory segment and the time after offset required to comprehend the
segment, segment duration should be subtracted from segment listening time
(the time between button presses to receive new segments) for an accurate
representation of processing ease or difficulty. Subtracting the segment
duration from the listening time results in a measure termed the Difference
Time, which were computed for each second clause segment.

In the second experiment, prosodic and syntactic Characteristics of the
first clause were fully crossed. Both the structure and prosody of the first
clause, and the structure of the second clause, were varied independently. This

design allowed the determination of several effects concerning the way in

17

which prosody is registered and used in spoken language processing. The cost
(or benefit) of a mismatch (or match) of first clause prosodic and syntactic
structure can be computed by subtracting the processing time for the parallel
cases from the corresponding non-parallel cases. Prosodic parallelism effects
can also be determined, for example, if the prosody of the first Clause produces
an advantage in same prosody (and structure) second clauses, even if the first
clause has a conflicting syntactic structure. Parallelism effects may only
become apparent if prosodic and syntactic structure match, e.g. only if all
sources of information in the first clause "point to" a single analysis of the
second Clause. An interaction of first clause prosody, first clause type, and
second clause structure in the disambiguating region would constitute strong,
on-line evidence of the immediate use of prosodic structure in parsing.
Self-paced reading and listening times were compared to show
similarities in syntactic and prosodic contributions to processing time. Because
presentation segments were the same, and both tasks required identical motor
movements and used the same comprehension task, visual and auditory
comprehension were more directly compared than in other work, where
modality comparisons are confounded with substantial differences in task

demands.

18

Experiment 1
I. Method

Subjects. Subjects were 48 Michigan State University students enrolled
in undergraduate psychology courses, who were given course credit for their
participation. All subjects were unaware of the specific purposes of the
experiment, had normal or corrected eyesight and normal hearing, and were
native speakers of American English.

Stimuli. Twenty-four experimental sentences were constructed, each a
two clause conjoined sentence (with the conjunction and). Twelve sentences
were modifications of Frazier et al.'s (1984) sentences. Each clause was either
minimally or nonminimally attached with respect to the attachment of a post-
verbal noun phrase. Within each sentence, all portions of the sentence after
the main verb (the interchanging region) of MA and NMA clauses were equated
in the number of words and syllables by adding prenominal adjectives. Across
all sentences, first and second clause interchanging regions were matched in
character length and summed lexical frequency of component words.
Perrnuting clause position and the attachment variable resulted in four
experimental conditions: first clause MA/second clause MA, first clause
MAIsecond Clause NMA, first clause NMA/second clause MA, and first clause
NMA/second clause NMA. Table 2 shows an example sentence in each of the
four conditions of the experiment, with interchanging regions delimited by

carets (“). Experimental sentences are presented in Appendix A.

19

All of the main verbs for the experimental sentences were taken from the
S I NP-complement norms obtained by Trueswell, Tanenhaus, and Kello
(1993). Eleven were equally biased in their tendency to take an NP or sentence
complement, that is, for each of them the ratio of one type of completion to the
other was no more than 1.67 to 1. These equi-biased verbs had an average S-
completion I NP-completion I other completion percentage distribution of
47/49/4. Nine other biased verbs were selected that had a completion (S I NP I
other) average of 40/41/18. These twenty verbs were combined with repetitions
of the four most equi-biased verbs so that the overall average completion
preference ratings for the verbs was 45I46/9. Sentence Complement I Noun
Phrase preference normative judgments are presented in Appendix 8.

Because experimental sentences in this study were greater than 80
characters long, line breaks were necessary. Breaks were placed immediately
before the conjunction. Thus the first clause was always the first line of the
display and the second Clause was always the second line of the display,
preceded by the conjunction. This display was chosen specifically so that when
a reader's eyes swept to the new line, eye movement and button pressing
disruptions would be confined as much as possible to the conjunction region,
and would not affect processing the subject noun phrase of the second clause.

The experimental sentences were segmented for presentation in self-
paced reading. The first Clause was presented whole, then the conjunction on

the next line of the display, and then segments which were either single content

20

words, or a determiner plus the following content word. Prepositions were
presented singly, so that three-word segments (eg. with the umbrella) were not
presented. Sentences ranged from seven to twelve segments in length. There
were five sentence regions: the first clause, the conjunction, the second clause
subject NP, the second clause main verb, and the second clause
disambiguating word. In addition, reading time for segments was summed for
two aggregated regions, the second clause interchanging region (which
contained material after the main verb including the disambiguating word), and
the full second clause. Analysis was conducted for all seven regions.
Seventy-eight filler sentences were constructed. Twenty-six filler
sentences were a single clause, constructed to be similar in length to the
experimental sentences by adding prepositional and infinitival phrases. The
remaining fillers were conjoined, two-clause sentences that did not exhibit the
parallel structure of the experimental sentences. Comprehension questions for
each of the experimental sentences and 16 of the filler sentences were made,
to ensure that subjects carefully read the sentences in the study.
Comprehension questions were questions about an action or participant in the
sentence, or about some attribute of the environment described in the
sentence. For example, the comprehension question for the sentence shown in
Table 2 was 'Was the meeting place well hidden7". Half of the comprehension

question were asked about some aspect of the first part (or first clause) of the

21

sentence and half were asked about some aspect of the second part (or
second clause) of the sentence.

Plausibility norms were collected for each clause of the two-Clause
sentences to ensure that they would be understood by the undergraduate
subject population, and that there were no semantic differences attributable to
plausibility among the conditions. The experimental and filler sentences were
rated by 72 undergraduate subjects on a five—point plausibility scale.
Plausibility was defined as how likely or believable the events depicted in the
sentences would be. A rating of 5 was used to indicate a highly implausible
sentence, 3 a moderately plausible sentence, and 1 a highly plausible
sentence. To eliminate possible effects of parallel structure on the normative
judgment, each clause was presented separately. In some second clauses, the
subject NP was elaborated because of a previously existing anaphoric
relationship to the subject of the first clause. Clause position was a between-
subjects variable, and clause type was within subjects. Thirty six subjects
received the 24 first clause sentences for judgment (half of which were MA and
half were NMA) and 36 subjects received the second clause sentences (half of
which were MA and half were NMA). Twelve lists of sentences were prepared
for the naming study, with sentences occurring in a different random order in
each list. Two examples each of highly implausible and plausible sentences

were shown to subjects in the instructions.

22

First clauses were judged marginally more plausible (mean=2.28) than
second clauses (mean=2.46), but only by items, F2(1,47)=3.51, p<.10,
MSe=7921‘. Second clause plausibility may have been adversely affected
because of the increased length required in some second clause items to
elaborate what had been an anaphoric reference to the first clause. Mean NMA
clause plausibility (2.38) did not differ from mean MA clause plausibility (2.36),
and did not interact with Clause position.

Because registering the believability of an event may be secondary to
constructing a coherent, sensible sentence representation, plausibility
judgments may not reflect the kind of semantic processing required in normal
comprehension; For this reason, the two clauses were conjoined to create
syntactically parallel and nonparallel sentences and another norming study was
done. Subjects were asked to rate "how easy it is to make sense of the
sentences", using a five-point scale. A rating of 5 indicated a sentence was
very difficult to make sense of, a rating of 3 indicated moderate sensibility, and
1 high sensibility.

Four lists of the 24 experimental items were constructed and presented
without fillers. Each list contained six items in each of four conditions created
by crossing the first and second Clause type variable. individual sentences
occurred in one condition per list. Thirteen participants rated each list.

Sentences with NMA first clauses were judged more sensible than those with

 

' in all reported results, F1 refers to analyses treating subjects as random
effects, and F2 refers to analyses treating items as random effects.

23

MA first clauses (2.28 vs. 2.55), F(1,12)=29.03, MSe=.032., p<.05 There was
no effect of second clause type. For sentences with NMA second clauses,
there was a significant sensibility increase (rating difference=0.69), t(23)=5.56,
p<.05, when the first clause was NMA compared to MA first clauses, but no
sensibility increase for MA parallel sentences, resulting in an interaction of first
and second clause type, F(1,12)=65.50, MSe=.036. These results are shown in
Figure 1, with each structural version shown with 95% confidence intervals.

Eguigment. The experiment was run on an IBM-compatible 486-based
microcomputer, with an SVGA color monitor. A button box was used to obtain
self—paced presentation keypresses and question answers. The box contained
a button labeled NEXT for proceeding to the next display segment, and keys
labeled YES and NO to answer comprehension questions.

We. Subjects were seated and the experimental procedure was
explained to them. Subjects sat with their preferred hand resting on the pacing
key of the button box. Subjects were instructed to read each segment normally,
and that comprehension questions would appear after a random subset of the
sentences. Subjects read twelve, two-line sentences for practice. Practice
sentences were not parallel in structure, and were provided to give subjects
practice pressing the button to display reading segments. After the practice
session the group of 102 sentences was randomized for presentation for each

subject by the experiment-running program. Each subject saw only one version

24

of each experimental sentence, but across items, received all experimental
conditions.
II. Results

Results from reading time analyses on seven regions are reported: (1)
the first clause, (2) the conjunction, and, in the second clause, (3) the subject
NP, (4) main verb, (5) disambiguating word, (6) interchanging region (including
all material after the main verb, including the disambiguating word). In addition,
processing time for all segments of the second clause were summed and are
also reported (7). Reading time for all experimental sentences is reported,
including those in which there were comprehension question errors. Data
including only sentences with correctly answered comprehension questions
were analyzed and are reported separately.

In all of the processing time analyses that follow, F values and
significance levels are reported for the interaction terms of first and second
clause structural factors. Where the interaction term is indicative of processing
benefit resulting from parallel structure, contrasts on the simple effects were
conducted to determine if a significant parallelism benefit for each structure
(NMA and MA) was obtained. Unless noted, all p values for statistical
comparisons are less than .05.

Question Answering Accuracy. Analysis of comprehension question
accuracy revealed a marginal effect of second clause type, such that questions

about sentences with MA second clauses were easier to answer (89.1 percent

25

correct) than questions about sentences with NMA second clauses (85.4
percent correct), F1(1,47)=3.18, p<.10, MSe=.0201. No other effects
approached significance.

1, First Clause. Reading times for the first clause as a function of the two
experimental variables (first and second clause type), irrespective of question
answering accuracy, were submitted to an ANOVA. Surprisingly, MA first
clause reading times were 300 ms longer than NMA first clause reading time
(4740 ms vs. 4440 ms), resulting in a main effect of first clause type, by
subjects, F1(1,47)=5.41, MSe=798,284, and by items, F2(1,23)=5.18,
MSe=481,419. This finding runs contrary to a body of results with NMA
structures showing them to be more difficult to process than corresponding MA
structures, and may have to do with the addition of words used to equate MA
structures to their NMA counterparts in length and lexical frequency. There was
no effect of second clause type and no interaction of the two variables.

Reading time analyses for the first clause were also conducted after
removing sentences associated with incorrect comprehension question
responses. There was a marginal effect of first clause type, with an NMA
advantage of 243 ms (4473 ms vs. 4716 ms), F1(1,47)=2.70, p=.10,
MSe=1,051,278 and F2(1,23)=2.99, p<.10, MSe=534,917. Again, there was no
effect of second clause type and no interaction of the two variables.

2, Conj'umion. The conjunction and constituted the initial region of the

second clause for all experimental sentences. There was a marginal reading

26

time advantage if the first'clause was NMA of 13 ms (370 ms vs. 383 ms),

F1 (1 ,47)=2.77, p<.10, MSe=3019, F2(1,23)=3.81, p<.10, MSe=1159. There
was no effect of second clause type and no interaction of the variables. After
removing errors, a marginal NMA first clause advantage of 12 ms was
observed in the items analysis (371 ms vs. 383 ms), F2(1,23)=3.24, p<.10,
MSe=1701. There was no effect of second clause type and no interaction of the
variables.

S. Subject NP. There were no effects of first or second clause type and
the factors did not interact, either with comprehension question error items
included or removed.

4. Main Verb. Analysis of all data revealed a marginal effect of first
clause type. If the first clause was NMA, a marginal reading time advantage of
27 ms over MA first clauses (487 ms vs. 514 ms) was seen at the main verb of
the second clause, F1(1,47)=2.71, p<.10, MSe=12,932, and F2(1,23)=3.06,
p<.10, MSe=6177. There was no effect of second clause type and no
interaction of the factors. In the analysis without errors, there were no
significant effects.

S. Disambiguating Word. A structural change is apparent at the
disambiguating word, so this is the first place to expect effects of second
clause structure. Second Clause structure may be consistent or inconsistent
with first clause structure here; if either (or both) NMA and MA consistency

(parallel structure) results in facilitated processing, there will be an interaction

27

of first and second clause type. Main effects of first and second Clause type are
not interpretable, though, because the words to be compared are different. For
example, in the example sentence in Table 2, reading time for wouldn’t in the
first sentence and new in the second sentence are likely to differ. Thus, the
interaction was examined to determine if benefit was significant for NMA and
MA parallel sentences. in these comparisons reading times are compared for
the same word as a function of the preceding clause type.

In the disambiguating region, the interaction of first and second clause
structure was significant, F1 (1 ,47)=4.54, MSe=15,157 and F2(1,23)=6.80,
MSe=4646. At disambiguation in NMA second clauses, there was a
nonsignificant 24 ms advantage if the first clause was parallel, while in MA
second clauses there was a 52 ms advantage if the first clause was parallel,
F1(1,47)=5.25, MSe=12,351, and F2(1,23)=6.59, MSe=5107, for the MA
parallel case contrast. Figure 2 shows the pattern, with 95% confidence
intervals for each condition.

With question errors removed, the interaction of first and second clause
structure was significant, F1(1,47)=5.61, MSe=20,225 and F2(1,23)=13.41,
MSe=4615. At NMA second clause disambiguation there was no apparent NMA
parallelism effect (9 ms). At MA second clause disambiguation, there was a 88
ms MA parallelism effect, F1 (1 ,47)=16.70, MSe=11,132, and F2(1,23)=5.24,
MSe=4367.

28

S, lnterchanging Region. Controls over length and lexical frequency in
this region allow main effects to be analyzed. Reading time after NMA first
clauses was 87 ms slower than after MA first clauses (2304 ms vs. 2217 ms),
F1(1,47)=3.25, p<.10, MSe=113,073, and F2(1,23)=3.39, p<.10, MSe=59,181.
There was no effect of second clause type. The interaction of first and second
clause structure was significant, F1 (1 ,47)=20.15, MSe=116,243 and
F2(1,23)=18.00, MSe=65,066. In contrast tests, an NMA parallelism effect of
133 ms was marginal by subjects and items (2182 ms vs. 2315 ms),

F1 (1 ,48)=3.20, p<.10, MSe=133,353, and F2(1,23)=3.07, p<.10, MSe=65,531,
and an MA parallelism effect of 308 ms was significant (2426 ms vs. 2118 ms),
F1(1,48)=23.79, MSe=95,957, and F2(1,23)=19.93, MSe=58,712. Figure 3
shows this pattern, with 95% confidence intervals for each condition.

Analysis without question errors showed that reading time after NMA
first clauses was 89 ms slower than after MA first clauses (2302 ms vs. 2213
ms), resulting in a first clause type effect, by items, F2(1,23)=4.38,
MSe=56,960. There was no effect of second clause type. The interaction of first
and second clause structure was significant, F1(1,47)=18.69, MSe=143,795
and F2(1,23)=12.57, MSe=112,948. An NMA parallelism effect of 148 ms was
marginal in both subjects and items analyses (2171 ms vs. 2319 ms),
F1(1,48)=2.82, p<.10, MSe=187,001, and F2(1,23)=3.16, p<.10, MSe=75,855.
An MA parallelism effect of 325 ms was significant (2433 ms vs. 2108 ms),

F1 (1 ,48)=23.83, MSe=106,506, and F2(1,23)=15.20, MSe=94,054.

29

Z, Second Slause. Analyses were conducted combining all segments of
the second clause. No effects of first or second clause type were seen. The
interaction of first and second clause structure was Significant, F1(1,47)=7.87,
MSe=301,350 and F2(1,23)=7.97, MSe=162,477. A 159 ms advantage for NMA
second clauses preceded by NMA first Clauses, as opposed to MA first clauses
(4057 ms vs. 4216 ms) was marginal in the subjects analysis, F1(1,47)=2.98,
p<.10, MSe=204,274. An effect of 286 ms for MA second clauses when
preceded by MA first clauses (4013 ms vs. 4299 ms) was significant,

F1 (1 ,47)=6.92, MSe=282,643, and F2(1,23)=8.40, MSe=124,255.

No main effects were found in the analysis with errors removed. The
interaction of first and second clause structure was again significant,
F1(1,47)=6.89, MSe=394,028 and F2(1,23)=6.57, MSe=256,740. Comparisons
showed that a 142 ms advantage for NMA second clauses preceded by NMA
first clauses (as opposed to MA first clauses, 4053 ms vs. 4195 ms) was
marginal by items, F2(1,23)=3.30, p<.10, MSe=124,280. An effect of 334 ms for
MA second clauses when preceded by MA first clauses (as opposed to NMA
first clauses, 4007 ms vs. 4341 ms) was significant, F1(1,47)=8.78,
MSe=305,370, and F2(1,23)=6.29, MSe=227,570.

Pogt-hgg Reading Time Analyses. After inspecting the experimental
sentences, I determined that not all of the MNMA sentences were parallel in
the structure of the interchanging region from the first clause to the second

clause. For example, in sentence (6) below, the predicate NP of the first clause

30

is a simple noun followed by a prepositional phrase with a left-branching
modified NP (that is, the adjectives modify the rightmost noun). The second
clause also contains a simple noun followed by a prepositional phrase with a
left-branching modified NP. The full sentence structure in sentence (6) is
identical, node-for-node. In sentence (7) below, the beginning three words of
the first clause modified NP are right-branching, while the final three words of
the second clause modified NP are left-branching. Parallel NMA sentences,
conversely, always had a consistent structure from first to second clause.
Given the irregularity of the MA experimental sentences, I conducted analyses
including a subset of nine experimental sentences in which the MA forms were
parallel from the first to second clause, in other words, sentences like (6)

below, but not like (7).

(6) NP V Modified NP

Friends in Washington mentioned the news about young man Clinton

CONJ NP V Modified NP

and other friends mentioned the rumors about old crook Nixon.

(7) NP V Modified NP

The CIA guessed the terrorists' very well hidden meeting place
CONJ NP V Modified NP

and the FBI guessed the location of the top secret new computer network.

1 First lause. The NMA advantage (121 ms), though present, was no
longer statistically significant. There was no effect of second clause type and

no interaction of the variables.

31

2, Sonjunction. The NMA advantage over MAS was robust (366 ms vs.
399 ms), F1(1,47)=4.48, MSe=11,412, and F2(1,8)=7.91, MSe=1066. No effect
of second clause type and no interaction of the two variables was apparent.

S. Subjeg NP. An effect ofisecond clause type reached significance in
the items analysis, in that the subject NP was read faster (by 37 ms) if the
second clause turned out to be MA, F2(1,8)=5.91, MSe=2010. No effect of the
first clause type was found and there was no interaction of the variables.

4, Main Verb. A 42 ms NMA first clause advantage (473 ms vs. 515 ms)
was reliable in the subjects analysis, F1 (1 ,47)=4.71, MSe=18,001. There was
no effect of second clause type or interaction of first and second clause type.

S, Qi§§mbiguating Word. First and second clause structure did not
interact in this region, so contrasts were not analyzed.

S. lnterchanging Rggion. There were no significant effects of first or
second clause type. An interaction of first and second clause structure
obtained in this region, F1(1,47)=15.30, MSe=144,353, and F2(1,8)=29.92,
MSe=14,544. Contrast testing showed a nonsignificant NMA parallelism effect
of 145 ms. An MA paralleliSm effect of 294 ms was reliable (2433 ms vs. 2139
ms), F1(1,47)=10.46, MSe=198,065, and F2(1,8)=8.68, MSe=55,536.

Z. Smnd Slause. There was no effect of first or second clause type. An
interaction of first and second clause structure was seen in this region,
F1(1,47)=6.56, MSe=290,860, and F2(1,8)=7.57, MSe=48,280. Comparisons

showed that a 216 ms advantage for NMA second Clauses when preceded by

32

NMA first clauses (as opposed to MA, 3850 ms vs. 4066 ms) was marginal in
the subjects analysis , F1 (1 ,47)=2.96, p<.10, MSe=379,776. An effect of 183
ms for MA second clauses when preceded by MA first clauses (3939 ms vs.
4122 ms) was nonsignificant.

Eost-Hoc Sorrelational Analyses. The direct objects in MA sentences in
this experiment were quite complex. Because the internal object NP structure of
the MAIMA sentences make them far less parallel than the NMA sentences
(only 9 of 24 MNMA sentences had the same internal direct object structure
from the first to the second clause), repetition of the exact structural tree cannot
be the only level at which parallelism benefits are observed. Otherwise, strong
syntactic parallelism benefit would be seen for the MAIMA parallel subset, and
not necessarily for all of the MNMA sentences.

in order to investigate the relationship between the processing difficulty
of MAIMA structures and their inherent complexity, a series of correlational
analyses were undertaken. First, reading time data for regions of the second
clause (the main verb and the disambiguating and interchanging regions) and
the raw sensibility norms were entered into analyses for each structural
condition of Experiment 1. in this way, the relationship between the rated
sensibility of a particular sentence and its reading time can be seen. Bonferroni
corrections for multiple comparisons were applied based on the total of twelve

computed correlations (four structures and three regions).

33

The amount of rated sensibility benefit for conjoining each sentence with
its same structure second clause was also computed, and these data were
entered into an analysis with the computed reading time benefit for each
conjoined sentence in second Clause regions, starting at the main verb. This
procedure allowed a determination of the relationship between ratings benefit
from parallelism in individual tokens of each structural type and any resultant
processing time benefit for those same sentences. Bonferroni corrections for
multiple comparisons were applied based on the total of 6 computed
correlations (two parallel structure conditions and three regions).

The rated sensibility of a sentence did not significantly correlate with
reading time for any of the four sentence types; In the sensibility norming
experiment, only NMA parallel structures showed a mean sensibility benefit;
that is, subjects rated sentence tokens as being more sensible in their
NMAINMA versions than in their MNNMA versions. Overall, MAIMA sentences
were not rated more sensible than their corresponding, non-parallel NMA/MA
versions. Benefit in rated conjoined sensibility could conceivably be associated
with processing time advantages, resulting in a relationship between the
individual conjoined sentence sensibility benefit and the reading time
parallelism benefit for each sentence. This relationship might be seen for MA
parallel structures, even though as a set they did not show a mean parallel
sensibility benefit in the norming study results. For each sentence, second

clause reading times of the main verb, disambiguating region and

34

interchanging region of parallel sentences were subtracted from their non-
parallel counterparts. Sensibility judgments of each parallel sentence were
subtracted from that of the non-parallel version. There was no relationship
between sensibility increases in NMA parallel sentences and any of the
segment reading time benefits. For MA parallel sentences, however, sensibility
judgment and reading time parallelism benefits mirrored each other, showing a
strong positive correlation in the disambiguating region, Pearson r = 0.528,
(x’=7.022, p<.01).

ill. Discussion

Two normative judgments were collected on the experimental sentences.
In one of these judgments, single NMA and MA clauses were seen as equally
plausible. in a second judgment, of the sensibility of the conjoined structures,
sentences with NMA first clauses were rated more sensible than those with MA
first clauses, and conjoined NMA clauses were rated more sensible than
conjoined MA clauses.

Reading times for the first clause showed unexpected difficulty for MA
structures and relative ease for NMA structures. Early in the second clause, the
conjunctions of second Clauses were read faster after NMA first clauses, a
possible "spillover" effect of first clause ease. Benefit of MA structural
parallelism was seen in reading time at disambiguation of the second clause,
and for both structures in the aggregated reading times of second clause

regime. Given the difficulty of MA first clauses in reading time and in

35

sensibility, it may be that if processing a clause is quite difficult, parsing
another of the same type is easier. Parallel structure benefit is still observed in
cases where structural assignment is not as difficult, as in NMA/NMA
sentences, although effects are not necessarily immediate and not always
robust. Interestingly, in analyses of full second clause processing for only those
sentences with identical syntactic structure in the MA interchanging regions,
the MA parallelism advantage disappears and the NMA parallel sentences
show marginal benefit.

These results are different than those reported in Frazier et al. (1984)
and in Branigan et al. (1995). Frazier et al. do not report first clause reading
time but we can presume, in line with F razier's parsing principles, the authors
would have expected NMAs to be difficult. in Branigan et al. there were no
differences in first clause reading time. Here, NMA first clauses were read for
much less time than MA first clauses. Frazier et al. report significant second
clause parallelism effects in NMA parallel sentences but not MA parallel
sentences. The converse of Frazier et al.'s results is seen in Experiment 1, with
parallelism effects for MAIMA sentences and trends for NMA/NMA sentences.
Branigan et al., on the other hand, did not find significant parallelism benefit for
either structure in NP objectIS-complement constructions. It is possible that the
MA clauses were unusual in the present study as a result of equating sentence
length in the object NPs with the addition .of prenominal adjectives. These

additions led to variability in the internal object NP syntactic structure of

36

interchanging regions of MA parallel sentences, making them more difficult
than their more consistent NMA parallel counterparts.

in summary, the results from Experiment 1 indicate immediate facilitative
effects of parallel structure in the reading time of second clause disambiguating
regime. The effect is reliable only for the more difficult MAIMA structures. The
processing time advantage of MA parallel sentences over NMA parallel
sentences remains constant in larger analysis regions of the second clause.
importantly, reading time benefit in the disambiguating region for MA parallel
sentences was strongly correlated with the sensibility rating benefits for those
sentences, but this relationship was not apparent in NMA parallel structures.

Experiment 2

Experiment 2 was designed to determine whether parallel syntactic and
prosodic structure affect listening, using the Auditory Moving Window task.
Acoustically recorded versions of the stimuli used in Experiment 1 were used,
with the prosody of each first clause manipulated to be consistent or
inconsistent with the syntactic structure. As in Experiment 1, the full first Clause
of each experimental sentence was presented as a single segment when the
subject initiated atrial. By presenting the first clause in its entirety, l hoped to
provide the best environment for subjects to use both the prosodic and
syntactic structure available to them. Subjects continued to press the pacing
button to hear the conjunction, and to hear each succeeding second clause

segment. The conjunctim was a separate segment to allow greater within-

37

sentence practice so that subsequent button presses would be minimally
dismpted.

Presentation segments were a modified version of the prosodic word
(Nespor and Vogel, 1987). Segments were content words unless preceded by a
determiner, in which case the determiner was included with the following
content word. Prepositions and the infinitival to were individual segments.

Two tokens of each sentence were recorded, so that all first clauses of
the experimental sentences consisted of two waveform portions combined into
me clause. For mismatching conditions, an NMA syntax first clause was
presented with MA prosody, or a MA syntax first Clause was presented with
NMA prosody. By crossing first clause prosody and first clause syntax, I hoped
to determine if and when prosodic and syntactic structure matching benefits are
apparent, and if specific benefits are observed in second clauses with parallel
syntactic and prosodic structure. The dependent measures were the listening
time for a segment (the time between button presses) and the difference time,
computed by subtracting the duration of each presentation segment from the
listening time. Only difference time results are discussed here.

I. Method

Subjggs. Subjects were 48 Michigan State University students enrolled

in undergraduate psychology courses. All subjects were given course credit for

their participation, and were unaware of the specific purposes of the

38

experiment. Participants had normal hearing and were native speakers of
American English.

m. A male native speaker of American English with a neutral accent
was recruited through an advertisement posted in the theater department. The
speaker was unaware of the specific purposes of the experiment. The speaker
was presented with the materials to be spoken, and was asked to familiarize
himself with each sentence before reading it. Only one sentence was visible to
the speaker at a time. The speaker produced each sentence and then listened
to the recorded version. if there were no errors or hesitations, the speaker
immediately recorded a second token of the same sentence. Comprehensim
questims were also recorded for each experimental sentence. Two versions of
each experimental sentence in each of the two parallel structure conditims
(MAIMA and NMA/NMA) were recorded and digitized (at a 20 kHz sampling
rate) using the Computerized Speech Laboratory software and hardware
system (CSL, Kay Elemetrics, Pine Brook, NJ). Filler sentences and their
cornprehensim questions were the same as those of Experiment 1. The
speaker recorded all sentences in a random order.

Waveforms for the eight experimental conditions were manipulated. Two
splices were made into one token of each originally-recorded sentence, one at
the beginning of the interchanging region of the first clause, and the second
before the conjunction which initiated the second clause. Several steps were

then taken to create the four conditions with NMA first clause prosody. A

39

pictorial representation of the experimental manipulations for NMA/NMA and
MAIMA sentences is reproduced in Figure 4, and the experimental conditims
are shown in Table 3.

lnaudible presentation markers (tags) were placed in the waveform of
each sentence at the end of the first clause, before the conjunctim, and at the
end of each succeeding presentation segment until the end of the sentence.
These tags demarcated sections of the waveform that the experimental
program played to the subject, consecutively, with each button press.
Questions were played out to the subjects as one segment. A 500 ms tone was
appended to the end of each sentence waveform as part of the final segment,
so that subjects were aware that the end of the sentence had occurred.
Comprehension questions were played after‘the button press at the end of the
last segment.

Presentation tags were placed at low amplitude transitims between the
last phmeme of a word and the first phoneme of the next word to create
natural-sounding transitions. Cursor control of placement in CSL is very
precise: each cursor movement can be as short as one-half of one millisecond.

. The location of the first tag was manipulated in the filler sentences. By
moving the location where segment-by-segment listening began, l hoped to
reduce the predictability of segment lengths in the full set of sentences, and to
maintain subject vigilance. if subjects could‘predict when self-paced listening

would begin, they might not be as attentive to the content of the first segment.

40

Eggipment. Subjects' button presses were obtained with the computer
and buttm box used in Experiment 1, and the experiment was conducted in the
same location. The computer stored all of the digitized stimuli, and contained
all of the digital-to—analog signal conversion hardware needed to present the
stimuli auditorily. Subjects listened to the signal through headphones.

Erocedure. Subjects were seated and the experimental procedure was
explained to them. Subjects listened to fifteen sentences to gain practice in
pacing through sentences with the pace key. Listening volume was adjusted for
comfort during practice. The experimenter listened to the sentences through a
second set of headphones. After practice, the 102 experimental sentences
were randomized for presentation by the experiment-running program. Each
subject heard only one version of each experimental sentence, but across
items, received all conditions of the experiment. Subjects answered
comprehension questions by pressing the appropriately marked buttm m the
button box.

M93. As can be seen in Table 3, the experiment was a 2 (Second
Clause Structure) X 2 (First Clause Structure) X 2 (First Clause ProSody)
factorial design. Given this design, several interactions may be obtained. First,
a first clause prosodic matching benefit may be seen. This effect is the
interactim of first clause structure and first clause prosody, collapsed over

second clause type. if first clause structure and prosody are both NMA or both

41

MA, one might expect processing benefits (compared to mismatching cases) for
these items in the first clause results, and perhaps in succeeding segments.

Second, there is the possible beneficial effect of structural parallelism,
compared to non-parallel cases, irrespective of first clause prosody. These
effects (for NMA and/or MA parallel structures) are reflected in the interactim
of first and second clause structure, collapsed over first clause prosody.

Third, in this design I can determine whether the prosodic structure of
the first clause exerts an independent influence on the processing of similarly-
structured second clause material. If the first Clause prosody of an item is NMA,
second clause structures that exhibit NMA prosodic characteristics in the
second Clause may show a benefit over non-parallel (MA first clause prosody)
structures. The same pattern might hold for second clause MA structures with
MA first clause prosody. If either (or both) of the structural types exhibits the
effect, an interaction of first clause prosody and second clause type, collapsed
over first clause structure, will result.

Finally, both syntactic and prosodic structure may influence processing
of segments of parallel second clauses in comparison to non-parallel cases.
Reliable effects in the three-way interaction of the experimental factors would
indicate that processing of a parallel second clause (whether after NMA or MA
first clauses) benefits from the joint influence of an introductory prosodic

structure and a continuing syntactic structure.

42

II. Results

Question flswering Accuracy. Comprehension question accuracy for
the experimental sentences was computed and analyzed in an ANOVA Grand
mean accuracy for the experimental sentence comprehension questions was
86.6%, and ranged from 81.3% for MA syntax/MA prosody/NMA second clause
sentences to 91.0% for NMA syntax/NMA prosody/MA second clause
sentences. There were no first or second cause main effects, no interactions of
first clause and second clause type, of first clause type and first clause
prosody, or of first clause prosody and second clause type. The interactim of
the three experimental variables was reliable, F(1,23)=5.39, MSe=.0068. This
interactim was the result of marginal prosodic matching benefits when second
clauses were MA (F[1,23]=3.16, p<.10, MSe=.0058) and not when second
clauses were NMA For MA second clauses, matching MA first clause prosody
and structure resulted in a 1.4% question accuracy increase, matching NMA
first clause prosody and structure resulted in a 4.2% accuracy increase. For
NMA second clauses, however, matching NMA first clause prosody and
structure resulted in a 0.7% accuracy increase, and a match of MA first clause
prosody and structure resulted in a 6.2% accuracy decrease. This interactim is
a reversal of a parallel structure effect, and is likely spurious. Evidence was not
found for specific benefits in the case of all three variables in their parallel

instances.

43

nggim Time Results: Difference Times. Results are reported below for

the difference between the duration of the segment and the listening time. First
clause duration did not differwith the structural or prosodic manipulations (all
ts<1, grand mean=3343 me), so first clause and subsequent region analyses
should be unaffected by the amount of time required to listen to the first clause.

Acoustic analyses of the stimuli were undertaken to determine the
prosodic characteristics of each clause type. Measurements were taken of both
the subject NP and main verb of first clauses, which were presented whole and
had a consistent prosody for a structure through about half of the first clause
duratim (until the end of the main verb). This manner of first clause
presentation was chosen to set up a prosodic representation without
interruption from button-pressing, so that subjects would have the best chance
to use any available prosodic cues in second clause processing. However, the
comparisons of NMA and MA clauses at (presumably) the two most
prosodically salient sentence regions showed no differences in duratim, mean
fundamental frequency (F0), F0 range, or F0 standard deviation, all t
statistics<1.

The seven analysis regions and the types of analyses conducted (with
queetim errors, errors excluded, and parallel item subset) are the same as
those reported for Experiment 1. Analyses of the parallel subset are presented
following the data with and without errors, in a separate section. Subject

analyses are not reported for the post-hoe analyses of the fully parallel item

44

subset, because only nine items could be analyzed, and as a result each
subject did not receive the full set of conditions.

1. First Slagse. Difference times of all data were analyzed. No effects of
first type, second clause type, or of first clause prosody were seen, and no
interactims approached significance. Likewise, there were no significant
effects of any variable or interactions in the difference time analyses without
errors. .

z. Sonjunctim. The conjunction and was the initial region of the second
clause for all experimental sentences. An analysis of difference times showed
no main effects of the variables. A first clause type by second clause type
interaction was observed, so that if the second clause (well after this region)
turned out to be NMA, an NMA parallel structure effect of 49 ms was observed,
and if the second clause turned out to be MA, an MA parallel structure effect of
68 ms was observed, F1(1,47)=4.41, MSe=74,813, and F2(1,23)=4.01, p<.10,
MSe=41,206, for the interaction. Conjunctions at the beginning of NMA second
clauses did not differ in duration, fundamental frequency mean or fundamental
frequency variability from those in MA second clauses, all ts<1. Second clause
structure is not yet known to the listener, and in the absence of any physical
differences in the signal the apparent parallel structure effect seems to be
spurious. Contrast testing was not done. First clause prosodic matching benefit

was not reliable, and no other interactions reached significance.

45

Difference time analyses without errors showed no main effects of the
variables. The first clause by second clause type interaction reached
significance, F1 (1 ,47)=3.76, p=.05, MSe=91,667, and F2(1,23)=5.01, p<.05,
MSe=34,543, reflecting an NMA parallel structure effect of 30 ms and an MA
parallel structure effect of 90 ms. Contrast testing was not done. The
interactim of first clause structure and prosody was marginal by subjects,
F1(1,47)=2.77, p<.10, MSe=92,629, indicating some facilitative effects of
matching first clause NMA prosody and syntax of 33 ms (621 me vs. 654 ms)
and of matching first clause MA prosody and. syntax of 76 ms (572 me vs. 643
me), though contrasts showed both NMA and MA prosodic matching to be of no
statistically significant benefit. No prosodic parallelism was seen, and the
three-way interaction of the variables was unreliable.

S, Subjggt NP. Effects of second clause structure, collapsed over the
other variables, began to emerge in this region although the structure of the
second clause was not yet apparent. Difference time data for all sentences
showed that if the second clause was going to be MA, a marginal difference
time advantage of 29 ms was seen over NMA second clauses (594 me vs. 623
ms), F1(1,47)=3.35, p<.10, MSe=23,814, and F2(1,23)=3.58, p<.10,
MSe=11,131. No other main effects were apparent. Processing was facilitated
by 32 ms for matching MA prosody and structure first clauses, and by 31 ms for
matching NMA prosody and structure first clauses, resulting in an interactim of

first clause type and first clause prosody by subjects, F1 (1 ,47)=7.95,

46

MSe=12,015. Contrasts indicated that the NMA prosody and structure match
benefit was not significant. The MA prosody and structure match contrast was
marginal in the subjects analysis, F1(1,47)=3.17, p<.10, MSe=15,148, and
reached significance in the items analysis, F2(1,23)=4.47, MSe=5753. No other
interactims approached significance. NMA second clause subject NPe did not
differ in duration from those in MA clauses (431 me vs. 441 ms), t(23)<1.
Difference time analyses without errors showed no main effects or interactions
of the variables.

4. Main Verb. Difference time analyses revealed no main effects, and a
trend toward a first claUSe prosody and structure matching effect, with an 8 ms
advantage at the verb if the first clause prosody and structure were both NMA
(423 me vs. 431 ms) and a 20 ms advantage (416 ms vs. 436 ms) if the first
clause prosody and structure were'both MA, by subjects, F1(1,47)=2.86, p<.10,
MSe=7078, for the interaction. Contrasts showed no effect for matching NMA
prosody and structure, but a marginal effect for MA prosody and structure
matches in the subjects analysis, F1 (1 ,47)=2.97, p<.10, MSe=6883. No other
interactims approached significance. Main verbs in NMA second clauses did
not differ in duration from those in MA clauses (370 me vs. 367 ms), t(23)<1.

Analyses without errors showed no significant main effects of the
variables. Processing benefits for matching first clause NMA prosody and
syntax of 14 ms (422 me vs. 436 ms), and for matching first clause MA prosody

and syntax of 30 ms (405 me vs. 435 ms) were seen, resulting in a marginal

47

interaction of first clause type and first clause prosody in the subjects analysis,
F1 (1 ,47)=3.80, p<.10, Mse=11,997, though neither the NMA or MA contrasts
were significant. There were no interactions of first and second clause type, of
first clause prosody and second clause type, or of all three variables.

S. Disgmbiguating Word. Second clause structure is determined at this
point, but main effects of first clause type, first clause prosody, and second
clause type are uninterpretable because the main effect collapses over
different words. Facilitated processing of parallel structure is indicated by any
interactions of first clause type, first clause prosody and second clause type. If
significant interactions were obtained in the overall analyses, simple effects
contrasts were done on the three analyses (all data, errors out, and parallel
item subset) to see whether parallelism effects were significant within each
structural type.

The difference time analysis of all data showed no reliable interactims
among the three experimental variables. In the analysis without comprehensim
questim errors, all interactions were also unreliable.

S. lnterchanging Region. Because character length, number of syllables
and lexical frequency were controlled over the interchanging regim of the
second clause, all main effects and interactions in the processing time data can
be analyzed. Difference time results with all data included revealed a main
effect of first clause type. Subjects listened to interchanging regions after NMA

first clauses 98 ms longer than after MA first clauses (2274 me vs: 2176 ms),

48

F1(1,47)'=4.87, MSe=189,032, and F2(1,23)=5.79, MSe=79,457. No effects of
first clause prosody were found, or of second clause type. A first clause by
second clause type interaction was significant in the subjects analysis,
reflecting a 16 ms processing time increase for NMA/NMA sentences and 180
ms facilitation for MAIMA sentences, F1 (1 ,47)=5.1 1, MSe=125,309. Cmtrasts
showed no significant effect of NMA parallel structure and a significant MA
parallelism effect, F1 (1 ,47)=1 1.70, MSe=132,230, and F2(1,23)=5.63,
MSe=137,294. No other interactions approached significance. These results
are shown in Figure 5, with 95% confidence intervals for each condition.

inspection of difference times with error data excluded revealed a main
effect of first clause type, with processing time of the interchanging region of
second clauses after NMA first clauses 114 ms slower than the times after MA
first clauses (2267 me vs. 2153 ms), F1 (1 ,47)=6.60, MSe=191,009, and
F2(1,23)=8.23, MSe=99,078. No effects of second clause type were found. The
interactim of first and second clause structure was not significant, yet contrast
tests showed structural parallelism benefit for MAIMA sentences,
F1 (1 ,47)=9.63, MSe=176,971, and F2(1 ,23)=6.76, MSe=159,300. No other
interactims approached significance.

7, Sgcong Slguse. Analyses were also conducted for the summed
difference times for all second clause segments. Difference times showed an
effect of first clause structure in the subjects analysis: second clauses after

NMA first clauses were listened to for 140 ms longer than after MA first clauses

49

(4326 me vs. 4186 ms), resulting in a main effect of first clause type, by
subjects, F1 (1 ,47)=5.02, MSe=373,161. There were no effects of first clause
prosody or of second clause type. No other interactions approached
significance. Although the structural parallelism interaction did not reach
significance, contrast testing was done for each structure to investigate
possible parallelism benefit. Difference times for NMA second clauses after
NMA first clauses were 10 ms slower than after MA first clauses: Clearly, there
was no NMA structural parallelism benefit across the full second clause for
NMA conjoined structures. Difference times of MA second clauses after MA first
clauses were 269 ms faster than after NMA first clauses, resulting in significant
parallel structure benefit for MA second clauses by subjects, F1 (1 ,47)=9.46,
MSe=367,888, and F2(1,23)=3.17, p<.10, MSe=578,849.

In the data without errors, there was an effect of first clause type, such
that second clauses after NMA first clauses were listened to for 181 ms longer
than after MA first clauses (4297 me vs. 4116 ms), F1 (1 ,47)=7.00,
MSe=448,022, and F2(1,23)=5.51, MSe=479,154. There were no effects of first
clause prosody or of second Clause type. Structural parallelism effects were
marginal in the subjects analysis, F1 (1 ,47)=3.05, p<.10, MSe=882,753.
Cmtrasts showed a nonsignificant 13 ms processing time increase for NMA
conjoined structures, but 395 ms of benefit for MA second Clauses preceded by
MA first clauses, F1 (1 ,47)=6.09, MSe=612,413, and F2(1,23)=12.51,

MSe=465,147. First clause prosody and second clause structure did not

50

interact. First clause prosodic matching benefit was observed in the subjects
analysis, F1 (1 ,47)=4.61, MSe=521,848. with a 163 ms benefit for matching
NMA first clauses. and a 149 ms benefit for matching MA first clauses. Neither
structure yielded significant levels of prosodic matching benefit in contrast
tests.

Beactim Time Results: Parallel Subset Difference Times. A subset of
sentences in this experiment (nine of 24) were identical in the structure of the
first and second Clauses. Data from this subset are reported below.

1. First Slause. Difference time analyses showed no effect of first or
second clause type, or of first clause prosody. No two-way interactions were
significant. The interaction of all three variables was reliable, F(1,8)=8.89,
MSe=13,390, though likely spurious because none of the second clause had
yet been heard.

2. Smjunctim. In difference time. no advantage was found at the
conjunctim following either structure. If the second clause turned out to be
NMA rather than MA. an advantage of 67 mswas seen in difference time for
the conjunction (596 me vs. 663 ms), F(1.8)=3.49, p<.10, MSe=22.881. As in
the full item set. there was a first clause structure by second clause structure
interaction, F(1,8)=4.88. p<.10, MSe=10.413, with an NMA structural
parallelism effect of 24 ms (584 me vs. .608 ms) and an MA structural

parallelism effect of 82 ms (622 ms vs. 704 ms). Contrasts were not tested.

51

S, 4, S, S, Z. Subject NP. Main Verb, Disambiguating and lntermanging
Rggiggg, ang Sflnd Slause. There were no main effects or interactions of the

 

three experimental variables in difference time in these regions.

Pgst-Hoc Sorrelatimal Analyses. Post-hoc correlational analyses of the
relatimships between sensibility and processing measures were conducted on
data from Experiment 2. In contrast to Experiment 1, there were statistically
significant relationships between difference time and raw sensibility judgments
in sentences with MA second clauses, whether parallel or not. These cases
showed increased difference times for those sentences judged less sensible,
across the main verb, disambiguating region, and interchanging region of
second clauses. Table 4 shows these correlations.

Correlational analyses were also conducted for the data transformed to
show parallel sensibility judgment and difference time benefits. The amount of
rated sensibility benefit for conjoining each sentence with its same structure
second clause was computed, and these data were entered into an analysis
with the computed difference benefit for each conjoined sentence in three
second clause regions (the main verb, disambiguating region and
interchanging region). This procedure allowed a determination of the
relatimship between ratings benefit from parallelism in individual tokens of
each structural type and any resultant processing time benefit for those same
sentences. Bonferroni corrections for multiple comparisons were applied based

on the total of 6 computed correlations (two parallel structure conditims and

52

three regions). in these analyses, the relationship between an individual item's
savings in sensibility and savings in difference time for conjoined structures
can be seen. There was no relationship between the two types of parallel
benefit (sensibility and processing) for NMA parallel structures, but a robust
relatimship for conjoined MA structures. This strong relationship between MA
parallel sentences that benefit in the sensibility judgments and their difference
time is apparent at the main verb and disambiguating region, and is shown in
Table 5.
III. Discussion

In Experiment 2, subjects answered comprehension questions with
about the same accuracy as in Experiment 1. In none of the region analyses
did removing data from inaccurately answered questions qualitatively affect
processing time results.

in contrast to Experiment 1, subjects did not show differential difficulty
with the clause types presented in Experiment 2. Difference times for NMA and
MA first clauses were statistically equivalent. Not surprisingly, then, there were
no spillover effects as a result of processing ease or difficulty of first clause
structure seen in the initial regions of the second clause. Accounting for the
discrepancy between Experiment 1 and 2 in first clause processing is taken up
in the General Discussion.

There were no significant effects of prosody in this experiment. Over the

first clause, and all subsequent regions. there were no reliable effects of first

53

clause prosody that would indicate an intrinsic advantage for one prosodic
structure over another. While there were trends toward prosodic and syntactic
structure matching effects early in the second clause, they were never reliable.

The present prosodic null results may be explained in part by the results
of the acoustic analyses. There were no differences in the overall prosodic
characteristics of initial regions of the first clause. One possible reason for this
lack of difference may be that adding prenominal adjectives to increase MA
clause length may have made MA Clauses less prosodically distinct than they
might be naturally. Also, the speaker who recorded the sentences used in the
study may not have realized that the sentences he was recording were of two
structures and may have felt no reason to accentuate or prosodically
disambiguate the structures; after all, he was not instructed to do so. in fact,
sentences were randomized specifically to ensure that the speaker did not read
different structural tokens of the same sentence close together in time, and
prosodic cueing may have been de-emphasized as a result.

Other factors may have contributed to finding no prosodic influence.
Vifrth the level of experimental control over item characteristics in the study, it
was very difficult to create the 24 items. in an eight condition design, each
subject heard a given condition on three trials. Condition means for subjects,
then, are likely to be unstable, reducing the statistical power necessary to
detect higher-order interactions. Also. prosodic parallelism is defined by the

matching of first and second clause prosodic structure, and in the present

experiment, there were two inherent obstacles to finding prosodic effects. F irst,
a first clause prosodic structure that is consistent with one syntactic structure is
mly partially presented: at the beginning of the interchanging region, the
splicing manipulation ensured that the next segments (and their prosodic
characteristics) were from a sentence token of the other syntactic etmcture, in
four of the eight conditions. In this way, only about the first half of the duration
of the first clause contained the appropriate prosody that could "set up"
prosodic effects in the second clause. Second, the AMW, by its nature, disrupts
the flow of prosodic information to subjects. If more of the second clause had
been presented with the initial second clause keypress, determination of
second Clause prosodic structure might have been more successful and
prosodic parallel structure benefits may have resulted. Structural and prosodic
parallelism are each defined by interactions of two design factors (both
interaction terms are collapsed over one other variable, doubling the number of
observatims to six). yet consistent parallel effects are found for syntactic and
not prosodic structure. In this study, it is Clear that syntactic structural factors
were more influential than prosodic structural factors.

MA parallelism effects in the larger analysis regions show a strong
similarity to those seen in Experiment 1. In contrast to Experiment 1, however,
NMA structural parallelism effects were not reliable in the disambiguating or
interchanging regions. or over the entire second clause, but MA structural

parallelism effects were strong in the two larger analysis regions.

55

General Discussion
. The major findings in these experiments were that MA structures were
more difficult to process than NMA structures, that there was greater
parallelism benefit for theusually simpler version of a temporarily ambiguous
structure, rather than for the more complex structure, and a lack of significant
prosodic effects, either as a cost of mismatching or as an indication of prosody
as a guide to parsing decisions. These major findings are all in some sense
contrary to those seen in the existing literature, and are each discussed in turn.
Structural Effects in the Experiments
The structural findings can be accounted for by considering syntactic
parallelism effects as arising from three distinct, co-occurring sources.
Branigan et al. (1995) have suggested that parallel structure effects can arise
when there is repetition of a single, high-level syntactic node, that is. a
syntactic choice point in temporarily ambiguous structures. According to
Branigan et al., coincidence of this choice point does not entail a node-for-node
correspondence in syntactic subconstituents. The authors suggest that the
effect is competitive; only when two alternative structures are competing to be
assigned by the parser will the effect be found, and furthermore, not all
ambiguities will Show the effect. Only those that cause significant initial parse
difficulty or engender a difficult reanalysis will show benefits of repetitim.
Some evidence for a single-node choice point component to parallelism

benefit is found in the present experiments. In Experiment 1, larger second

clause analysis regions show marginal NMA parallelism effects, and these
reach significance in the second clause analysis that includes only the fully
parallel MA sentences and their NMA counterparts. Of course, the NMA/NMA
sentences share not only the single-node attachment site for the sentence
complement. but also all subconstituent nodes. A single-node choice point
contribution to syntactic priming is Unlikely to be the only source of effects in
the NP objectIS-Complement ambiguity because it is found in some studies but
not others, yet the structural choicepoint defines the ambiguity in all instances.
Branigan et al. (1995) do not find parallelism benefit for either MA or NMA
sentences in their experiment, perhaps because the NP objectIS-Complement
ambiguity is relatively weak, as they suggest. and more susceptible to change
as a result of lexical or semantic stimulus properties, a point which I will
address below.

Branigan et al. (1995) also discuss the possibility of a more global
contributim to syntactic parallelism benefit. Here, in the extreme case, each
syntactic node would have to be repeated for benefit to occur. Effects are likely
in more moderate situations. however, with syntactic priming occurring for
portims of sentences containing more limited node-for-node correspondence.
Note that this source of parallelism benefit is not necessarily tied to structures
that exhibit temporary structural ambiguities, but may be more general.

in many parsing experiments (not only those intended to investigate

parallel structure effects). the single-node choice point and node-for—node

57

similarity of an ambiguity are not independent of each other, that is, structures
are relatively homogeneous. In the NMA structures presented here, a similar,
single-node syntactic choice point is shared among items, and so is the node-
for-node structure. In the MA sentences, though, there is a dissociatim in the
single-node attachment site and the node-for-node structure. Node-for-node
syntactic descriptions of the NP objects are quite different, though they retain
the single-node NP object attachment site at a similar point in the sentence. if
parallelism benefits are based solely on node-for-node coherence of syntactic
descriptims of the first and second clauses, then MA parallelism effects should
have been strongest in the full second clause analysis of fully parallel items
and weakest in the full set of items, which is the converse of the results
obtained in Experiment 1.

I suggest that there is a third, semantic component to parallel structure
effects: the lexical and conceptual similarity of the two clauses of a sentence.
Subjectiveiy, it seems that stimuli used in previous research present a
continuum along this dimensions. The present experiments use sentences with
some overlap of lexical and cmCeptual structure, though not nearly as much as
the original Frazier et al. (1984) study Upon which they were based. Branigan
at al. (1995), m the other hand, made a concerted effort to reduce lexical and
conceptual overlap even more than I have here. in terms of semantic (and

syntactic) complexity, Frazier et al.'s stimuli are the least complex, Branigan et

58

al.'s lie in the middle, and the present stimuli (the MA sentences only) appear
to be rather difficult.

In the present experiments. the sensibility norming judgments can be
seen as a metric for quantifying this semantic subcomponent of parallel

structure effects. Sentences with MA first clauses and conjoined MA clause

I;
sentences were judged less sensible than sentences with NMA first Clauses or .
conjoined NMA clause sentences. The NMA conjoined sentences may be :
sensible in such a way that they do not show the same gain from repetition as

in cases with more difficult adjectival and adverbial sequences. Evidence for L

 

this account is seen in two results of the present experiments. If we assume
that the fully parallel subset of MA sentences are also the least complex MA '
sentences (possibly because the more complex a syntactic structure is, the
harder it is to create parallel clauses). then the lack of an NMA advantage in
the first clause, fully parallel analysis of Experiment 1 is a result of equalizing
the complexity of the two versions of the ambiguity. The correlational analyses
of sensibility and processing benefit are more striking. At the disambiguating
regim in Experiment 1, and at the main verb and disambiguating regim in
Experiment 2, relative processing time benefit for a given MA parallel sentence
is strmgly correlated with the sensibility benefit of that sentence. but there are
no significant processing and sensibility benefit correlations for the NMA

parallel sentences.

59

Both the correlational analyses of parallel sentence processing benefit
and sensibility benefit, and the processing time results, show differences
across the two tasks used in this study. MA parallelism effects are robust in
both experiments, but NMA parallelism is only found in reading (and there mly
marginally). Why should the tasks differ in their sensitivity to sensibility effects?
The answer may lie in the size of the segments used in the AMW task. In
Ferreira et al. (1996), participants rated how natural the task was for them. how
difficult it was to comprehend segments, and how often difficult to understand
segments became clear to them by the end of the sentence. Participants rated
the task as fairly natural, and did occasionally have difficulty understanding
segments. These difficulties were almost always resolved by the end of the
sentence. First clauses in Experiment 2 were presented whole, and in MA
clauses, subjects were faced with segmenting the speech stream of a complex
object NP. Several short segments (adjectives, typically) are then paced
through, one after another, in the second clause MA interchanging regim. it is
likely that these short segments were less intelligible to subjects, placing
different processing demands on subjects than when reading the same short
segments in Experiment 1. The demands of the listening task may have made
MA sentences even more difficult to comprehend than NMA sentences and
inflating the parallel structure benefit observed.

A more systematic investigation of how structural facilitation is

moderated by structural difficulty is certainly warranted. Strong MA parallel

60

effects may be a result of global-level procedures that can be primed to attach
a large number of similar structural nodes quickly when faced with complex .
NPs. Other investigations have not used NPs as structurally complex as the
mes presented here. and NP structure should be a manipulated variable in
future studies.

Prosodic Effects: Where are They?

Several key questions remain regarding Iisteners' reliance on prosodic

I“ It

structural characteristics as cues to parsing. The environment of

 

comprehenders is filled with informative (and perhaps not so informative) __~~
prosodic features, and both garden-path and constraint-based parsing models

need to account for findings of prosodic effects in the literature. The present

experiments were not designed to adjudicate between these two models of

parsing - they were intended to examine syntactic parallelism effects in a new

way. and to see if repeating a prosodic structure can provide the kind of benefit

previously seen in the processing of repetitive syntactic structures.

The acoustic analyses of the stimuli indicate that the structures used
here were not realized differently by the speaker. Analyses were conducted on
the first portions of the first clause, in regions where Beach (1991) showed
effects of manipulations in off-line matching. Whether prosody can be used to
guide an initial parse or to aid in reanalysis is still an open question, but it is
possible that in a matching task like Beach's, effects are late in the stream of

processing, in reanalysis of the presented fragment. While Ferreira et al.

61

(1996) have argued convincingly that the AMW is a sensitive, m-line
technique. we should only expect to see effects if prosodic structural cues are
actually present.

Researchers have concluded that prosodic cues to a given structure are
not perfectly predictive, but that a probabilistic relationship holds in the
mapping of prosodic structure to syntactic structure. It is likely also that
different temporary syntactic ambiguities vary in the extent to which they are

prosodically realized by Speakers in natural contexts. and thus how

 

dependable the prosodic characteristics would be for comprehenders to rely

9%

on. The present results suggest that the NP objectIS-Complement ambiguity
may be particularly variable in its prosodic cueing. Determining which
structures are produced with strong and stable prosodic characteristics is an
important endeavor for researchers in the field. We can then conduct
experiments utilizing parametric manipulations of these characteristics if we
hope to understand the relationship between the processing of prosody and

syntactic structural assignment.

APPENDICES

 

APPENDIX A

 

IT

62

APPENDIX A

Experimental Sentences

Jim believed Tom's $ stories were fantastic | fantastic new stories 5 and Sue
believed Jim's 8 stories were fictitious | very boring stories 8.

The explorer learned the route to the S Pacific Coast was difficult I very rainy
Pacific Coast 8 and his guides learned the route to 8 South America was
treacherous | very remote South America 8.

Sally doubted the solution to the S physics problem was easy I really hard
physics problem $ and Ervin doubted the solution to the 8 math problem was
trivial | one difficult math problem 8.

Johnny claimed the old basketball at the $ garage sale was deflated | very
crowded garage sale $ and his mother claimed the 8 volleyball net was torn |
torn old volleyball net 8.

The newspaper admitted atrocities in $ Asia weren't yet over | war-stricken
southeast Asia 3 and Amnesty international admitted increased torture of 8
prismers had just started I new and younger prisoners 8.

Tim guessed the terrorists' 5 meeting place was hidden nearby | very well
hidden meeting place 3 and the inspector guessed the Iocatim of their 8
computer network wouldn't be found I top-secret new computer network 8.

Friends in Washington mentioned the news about 8 Clinton was true I young
man Clintm S and other friends mentioned the rumors about 8 Nixon were true
| old crook Nixon 8.

The spy suspected the man with the S umbrella killed Harold j wonderful old
umbrella S and the cop suspected the 8 woman witnessed everything |
impatient female witness 8.

Developers predicted the construction of the $ skyscraper would continue | new
gigantic skyscraper $ and the bankers predicted the decay of 8 nearby
neighborhoods would quicken | quiet old nearby neighborhoods 8.

.- 1:."

 

r3 -'-. t -

63

The skiers felt the S strong winds die down | fierce new strong winds S and the
instructor felt the 8 dropping temperature stabilize | dangerous dropping
temperature 8.

The museum's tour guide denied the S girls should leave | bright young girls S
and the security guard denied the 8 tour guide should hurry | busy new tour
guide 8.

The clever boy asserted the truth about the S Earth's age wasn't well
understood j poorly-designed deep space probe S and he asserted the 8 I:
implications were interesting I radical new implications 8.

Bugs Bunny confessed the time of his S business meeting was changed I lmg-
planned business meeting S and Daffy Duck confessed the stealing of the 8
briefcase was wrong | wrong black briefcase 8. l

 

The instant replay revealed the call of the S referee was incorrect | most hated
referee S and the league office revealed the 8 new technology was faulty I
quite costly new technology 8.

The woman maintained the S sculpture was truly beautiful | beautiful and
modern sculpture S and the art dealer maintained the 8 client was wealthy |
wealthy young client 8.

The scientist implied the S theory was crazy | greatest new idea S and the lab
assistant implied the previous 8 results were fabricated | year's undocumented
results 8.

People at last year's Rose Bowl noticed the S streaker was wearing tennis
shoes I very thin and smiling streaker S and the announcers noticed the 8
quarterback was watching closely | well-conditioned young quarterback 8.

Students in the advanced philosophy class observed the S teacher was
unshaven | unshaven new teacher S and the teacher observed the 8 clock
hands moved slowly I slowly-moving clock hands 8.

The insurance agent inferred the S procedure's cost would rise | new
procedure's high cost S and the patient inferer the 8 surgery would take
months | long-term surgical risks 8.

The plane crash survivors realized their S rescue wouldn't happen soon |
quickly-fading rescue hopesS and journalists around the world realized the 8
story would be powerful I bold story's positive impact 8.

64

The prisoner asserted his S rights had been abused | own civil rights views S
and the police department asserted its 8 innocence was now questioned | quite
lmg-questimed innocence 8.

The network president denied the S news story was fabricated | widely-reported
news story S and the company spokesperson denied the 8 recent accusatims
were serious | recent and serious accusatims 8.

The young boy mentioned the S school fight had bothered him I cruel and
vicious school fight S and the principal mentioned the 8 parents were very
concerned | concerned and upset parents 8.

Sherlock Holmes noticed the S man's shoes were muddy | men's muddied
brown shoes S and Watson noticed the 8 suspect was quite upset | quite upset
young suspect 8.

in the sentences above, lexical material that is swiched to create first
clause conditions is shown within dollar sign symbols, and is delimited by the
piping symbol (|). The lexical material that is swiched to create second clause
conditims is shown within ampersands, and is delimited by the same symbol.

 

APPENDIX B

«I-

 

65

APPENDIX B

MAIN VERB SENTENCE COMPLEMENT I NOUN PHRASE PREFERENCE
NORMATIVE JUDGMENTS

11 neutral verbs ' 9 biased verbs
S-Completim I NP Completion / Other S-Completim I NP Completim I Other

1. infer 4315017
2. maintain 7912110
3. suspect 2917110
4. confess 43143114
5. admit 3615717
6. assert 5014317
7. deny 50/4317
8. doubt 3615717
9. mention 4315710
10. notice 50/5010
1 1. predict 5714310

. realize 7193/0
believe 21129150
. learn 64121115
claim 7171/22
.guess 43121136
.feel 36121143
. reveal 79/2110
. observe 8611410
. imply 2117910

toooxroaotgscogo-s

average 4714914 average 40141/18

average of
6, 7, 9, 10 4814814

WEIGHTED AVERAGE FOR 11 NEUTRAL VERBS + 4 NEUTRAL VERBS
+ 9 NON-NEUTRAL VERBS: 45/46/9

APPENDIX C

 

Table 1

Frazier et al. (1984) - Experimental Sentence Conditims

Active/Passive

1a. The tall gangster hit John and the short thug hit Sam.

1b. John was hit by the tall gangster and Sam was hit by the short thug.
1c. The tall gangster hit John and Sam was hit by the short thug.
1d. John was hit by the tall gangster and the short thug hit Sam.

Minimal Attachment/Nonminimal Attachment

2a. Jim believed all Tom's stories
2b. Jim believed all Tom's stories were literally true

2c. Jim believed all Tom's stories

2d. Jim believed all Tom's stories were literally true
NonshiftedIShifted Heavy NP

3a. Mary wrote a long note about her predicament
to her mother

3b. Mary wrote to her mother a long note

about her predicament

30. Mary wrote a long note about her predicament
to her mother

3d. Mary wrote to her mother a long note

about her predicament

and Sue believed Jim's stories.

and Sue believed Jim's stories were
fictitious.

and Sue believed Jim's stories were
fictitious.

and Sue believed Jim's stories.

and Sue wired a telegram

requesting more money to her father.
and Sue wired to her father a telegram
requesting more money.

and Sue wired to her father a telegram
requesting more money.

and Sue wired a telegram

requesting more mmey to her father.

67

Table 2

Experimental Sentence Conditions

1) The CIA guessed the terrorists' “meeting place was hidden nearby“
and the FBI guessed the location of “their computer network wouldn't be found.“ (NMA/NMA)

2) The CIA guessed the terTm'sts' “meeting place was hidden nearby“
and the FBI guessed the location of “their top-secret new computer network.“ (NMA/MA)

3) The CIA guessed the terrorists' “very well hidden meeting place“
and the FBI guessed the location of “their computer network wouldn't be found.“ (MA/NMA)

." 'm- fly

4) The CIA guemd the terrorists' “very well hidden meeting place“
and the FBI guessed the location of “their top-secret new computer network.“ (MAIMA)

 

First Clause
Syntax/Prosody

NMA/NMA
NMA/MA
MAINMA

MAIMA

68

Table 3

Experimental Conditions - Experiment 2

Second First Clause Second
Clause Syntax/Prosody Clause
NMA NMA / NMA MA
NMA NMA I MA MA
NMA MA I NMA MA
NMA MA / MA MA

69

Table 4

Correlations of Difference Time and Sensibility Judgments

Main Verb Disambiguation lnterchanging Region
Structure .
NMA/MA n.s. n.s. 0.202
x2=7.746, p<.01
MAIMA 0.503 0.381 n.s

x2=13.255, p<.01 12:7,151, p<.01

70

Table 5

Correlations of Difference Time and Sensibility Judgment Parallel Benefits

Main Verb Disambiguation lnterchanging Region
Structure
MAMA 0.574 0.518 n.s.

x2=8.593. p<.01 12:6.703, p<.01

APPENDIX D

<_ (Most) Sensibility Judgment (Least) ->

 

71

Figure 1

Sensibility Judgments

 

 

 

NMA/NMA MAINMA NMA/MA MAIMA

Sentence Type

 

Reading Time (ms)

72

Figure 2
Experiment 1
Reading Time
Second Clause Disambiguating Region

 

800

 

750 —

700 -

650-

600-—

550—

 

50° / J“

“Ll-1.1

NMA/NMA MAINMA NMA/MA MAIMA

 

 

Sentence Type

Reading Time (ms)

73

Figure 3
Experiment 1

Reading Time
Second Clause lnterchanging Region

 

2500 -

2000-

 

 

 

/
50° / /
.__-_-j_-__
NMA/NMA MAINMA NMA/MA MAIMA
Sentence Type

74

Figure 4

Stimulus Sentence Construction - Experiment 2

 

(8a) The prismer assertedlhis rights had been abusedl

oooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooo

eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee

 

 

ooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooo

oooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooo

 

 

.........................................................

 

 

ooooooooooooooooooooooooooooooooooooooooooooooooooooooooo

eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee

Material within the solid boxes in Be and 8b was switched to create the NMA first
clause structure I NMA first clause prosody I NMA second clause structure condition. Material
within the solid boxes in 93 and 9b was switched to create the MA first clause structure / MA
first clause prosody 1 MA second Clause structure condition. Material within the dashed box in
9a then replaced the dashed box in 8a to create the NMA first clause structure I NMA first
clause prosody 1 MA second clause structure condition. Material within the dashed box in 8a
replaced the dashed box in 93 to create the MA first clause structure I MA first clause prosody!
NMA second clause structure condition.

Mismatching prosody conditions were then constructed by replacing material within the
solid box in 8a with material within the solid box in 9b. to create the MA first clause structural
NMA first clause prosody 1 NMA second clause structure condition. Material within the solid box
in 9a was replaced with material within the solid box in 8b, to create the NMA first clause
structure I MA first clause prosody I MA second clause structure condition. Material within the
dashed box in 9a then replaced material within the dashed box in 8a to create the MA first
clause stnicture 1 NMA first clause prosody 1 MA second clause structure condition. Material
within the dashed box in 8a then replaced material within the dashed box in 9a to create the
MA first clause structure I NMA first clause prosody I NMA second clause stntcture condition.

Difference Time (ms)

75

Figure 5
Experiment 2
Structural Parallelism Benefit
Second Clause lnterchanging Region

 

2750 -

 

 

 

MIA/MA MAIMA MIA/MA MAIMA

First Clause Structure I Second Clause Structure

LIST OF REFERENCES

76

LIST OF REFERENCES

Beach, CM. (1988). The influence of higher level linguistic information on
production of duration and pitch patterns at syntactic boundaries.
Journal of the Acoustical Society of America, 75, 8(1).

Beach, CM. (1991 ). The interpretation of prosodic patterns at points of
syntactic structure ambiguity: Evidence for cue trading relations. Journal
of Memory and Language, 30, 644-663.

Branigan, H., Pickering, M., 8. Stewart, A. (1995, March). Syntactic priming in
language comprehension. Paper presented at the Eighth Annual CUNY
Sentence Processing Conference, Tuscon, AZ.

Cooper, W.E.. (1980). Syntactic-tophonetic coding. In B. Butterworth (Ed),
Language production. Vol. 1: Speech and talk. (pp. 297-334). San
Diego, CA Academic Press.

Cutler, A (1976). Phoneme monitoring reaction time as a function of preceding
intonation contour. Perception & Psychophysics, 20, 5560.

Cutler, A, 8: Fodor, J.A. (1979). Semantic focus and sentence comprehension.
Cognition, 7, 49-59.

Ferreira, F. (1993). Creation of prosody in sentence production. Psychological
Review, 100, 233-253.

Ferreira, F., Anes, M.D., 8. Horine, MD. (1996). Exploring the use of prosody
during language comprehension using the Auditory Moving VVIndow
technique. Journal of Psycholinguistr‘c Research, 25, 273-290.

Ferreira, F., Henderson, J.M., Anes, M.D., Weeks, P.A., Jr., 8. McFarlane, DK
(1996). Effects of lexical frequency and syntactic complexity: Evidence
from the Auditory Moving Window Paradigm. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 22, 324-335.

77

Ferreira, F ., 8 Clifton, C. (1986). The independence of syntactic processing.
Journal of Memory and Language, 25, 75-87.

Ferreira, F ., 8. Henderson, J.M. (1990). The use of verb information in syntactic
parsing: A comparison of evidence from eye movements and word-by-
word seIf-paced reading. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 16, 555-568.

Frazier, L. (1979). On comprehending sentences: Syntactic parsing strategies.
Bloomington, IN: Indiana University Linguistics Club.

Frazier, L., & Rayner, K (1982). Making and correcting errors during sentence
comprehension: Eye movements in the analysis of structurally
ambiguous sentences. Cognitive Psychology, 14, 178-210.

Frazier, L., Taft, L., Roeper, T., Clifton, C., & Ehrlich, K (1984). Parallel
structure: A source of facilitation in sentence comprehension. Memory &
Cognition, 12, 421-430.

Grosjean, F. (1983). How long is the sentence? Prediction and prosody in the
on-line processing of language. Linguistics, 21, 501-529.

Kjelgaard, MM, 8 Speer, SR. (1993, March). Prosodic structure, syntactic
structure, and lexical verb preferences in the comprehension of
temporary syntactic ambiguity. Poster presented at the Sixth Annual
CUNY Sentence Processing Conference, Amherst, MA.

Lehiste, I. (1970). Suprasegmentals. Cambridge, MA MIT Press

MacDonald, M., Perlmutter, N., & Seidenberg, MS. (1994). The lexical nature
of syntactic ambiguity resolution. Psychological Review, 101 ,676-703.

MacDonald, M., & Schuster, S. (1994, March). Paper presented at the Seventh
Annual CUNY Sentence Processing Conference, New York, NY.

Marslen-Wilson, W.D., Tyler, LK, Warren, P., Grenier, P., & Lee, CS. (1992).
Prosodic effects in minimal attachment. Quarterly Journal of
Experimental Psychology, 45A, 73-87.

Mahler, J., 8. Carey, P. (1967). The role of surface and base structure in the
processing of sentences. Journal of Verbal Learning and Verbal
Behavior, 6, 335-338.

Nespor, M., 8. Vogel, l. (1987). Prosodic phonology. Dordrect: Foris.

78

Pike, KL. (1945). The intonation of American English. Ann Arbor, MI: University
of Michigan Press.

Price, P.J., Ostendorf, M., Shattuck-Hufnagle, S., & Fong, C. (1991). The use
of prosody in syntactic disambiguation. Journal of the Acoustical Society
of America, 90, 2956-2970.

Shields, J.L., McHugh,A., 8 Martin, JG. (1974). Reaction times to phoneme
targets as a function of rhythmic cues in continous speech. Journal of
Experimental Psychology, 102, 250-255.

Slowiaczek, ML. (1981 ). Prosodic units as language processing units.
Unpublished doctoral dissertation, University of Massachusetts,
Amherst.

Speer, S.R., Crowder, R.G., & Thomas, L.M. (1993). Prosodic structure and
sentence recognition. Journal of Memory and Language, 32, 336-358.

Speer, S.L., & Slowiaczek, ML. (1990). The influence of prosodic structure on
comprehending sentences with temporary syntactic ambiguities.
Unpublished manuscript.

Trueswell, J., Tanenhaus, M., & Kello, C. (1993). Verb-specific constraints in
sentence processing: Separating effects of lexical preference from
garden paths. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 19, 528-553.

Warren, P. (1985). The temporal organisation and perception of speech.
Unpublished doctoral dissertation, University of Cambridge, UK

Watt, SM, 8 Murray, W.S. (1993, March). Complement sentence parsing in
the auditory modality. Poster presented at the Sixth Annual CUNY
Sentence Processing Conference, Amherst, MA