University of Southern California Dissertations and Theses

Effects of working memory and semantic impairments on speech in Alzheimer's disease.

(USC Thesis Other)

Effects of working memory and semantic impairments on speech in Alzheimer's disease.

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content INFORMATION TO USERS
This manuscript has been reproduced from the microfilm master. UMI
films the text directly from the original or copy submitted. Thus, some
thesis and dissertation copies are in typewriter face, while others may be
from any type o f computer printer.
The quality of this reproduction is dependent upon th e quality of the
copy submitted. Broken or indistinct print, colored or poor quality
illustrations and photographs, print bleedthrough, substandard margins,
and improper alignment can adversely affect reproduction.
In the unlikely event that the author did not send UMI a complete
manuscript and there are missing pages, these will be noted. Also, if
unauthorized copyright material had to be removed, a note will indicate
the deletion.
Oversize materials (e.g., maps, drawings, charts) are reproduced by
sectioning the original, beginning at the upper left-hand comer and
continuing from left to right in equal sections with small overlaps. Each
original is also photographed in one exposure and is included in reduced
form at the back of the book.
Photographs included in the original manuscript have been reproduced
xerographically in this copy. Higher quality 6” x 9” black and white
photographic prints are available for any photographs or illustrations
appearing in this copy for an additional charge. Contact UMI directly to
order.
UMI
A Beil & Howell Information Company
300 North Zeeb Road, Ann Arbor MI 48106-1346 USA
313/761-4700 800/521-0600
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Effects of W orking Memory and Semantic Impairments
on Speech in Alzheimer's Disease
by
Lori Jean Altmann
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(Linguistics)
August 1998
Copyright 1998 Lori Jean Altmann
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
UMI Number: 9919007
Copyright 1998 by
Altmann, Lori Jean
All rights reserved.
UMI Microform 9919007
Copyright 1999, by UMI Company. All rights reserved.
This microform edition is protected against unauthorized
copying under Title 17, United States Code.
UMI
300 North Zeeb Road
Ann Arbor, MI 48103
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
UNIVERSITY OF SOUTHERN CALIFORNIA
THE GRADUATE SCHOOL
UNIVERSITY PARK
LOS ANGELES. CALIFORNIA 90007
This dissertation written by
L o r i J . Altmann
under the direction of h.5E Dissertation
Committee, and approved by all its members,
has been presented to and accepted by The
Graduate School in partial fulfillm ent of re
quirements for the degree of
DOCTOR OF PHILOSOPHY
Date Au^us t . 181 998
DISSERTATION COM MITTEE
Chairperson
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Table of Contents
List of Tables iv
List of Figures V
Acknowledgments vi
Dedication vii
Abstract viii
1 Introduction 1
1.1 Theories o f speech production and Working Memory 4
1.2 Speech Production in Alzheimer’s Disease 6
1.3 Overview of the dissertation 8
2 Literature Review: Speech Production and Semantic Representation 12
2.1 Theories of speech production— Garrett and Levelt 12
2.1.1 Open and closed class words 13
2.1.2 Garrett’s model of speech production and word class differences 14
2.1.3 Levelt and speech production and word class differences 16
2.1.4 Word class differences as contrasts in semantic richness 19
2.1.5 Evidence for modular speech production theories 20
2.2 Two views of semantic representation 24
2.3 Research that supports compositional semantic representations 30
2.3.1 Cross-linguistic studies of aphasia 30
2.3.2 Imageability effects 34
2.3.3 Syntactic priming and closed class words 37
2.3.4 Frequency effects in recognition memory tasks 38
2.3.5 Frequency effects on lexical access 39
2.3.6 Summary 41
2.4 Conclusions and predictions 42
3. Literature Review: Working Memory, Speech Production and
Alzheimer’s Disease 45
3.1 Theories of working memory 46
3.1.1 Baddeley and the “central executive” 47
3.1.2 Engle and the General Capacity Model 49
3.1.3 Just and Carpenter and the effects of limited Working Memory 51
3.2 Working memory and speech production 52
3.2.1 Daneman and Green (1986) 54
3.2.2 Daneman (1991) 54
3.2.3 Power (1985) 56
3.2.4 Valencia-Laver (1992) 57
3.3 Speech production in Alzheimer’s Disease 59
3.4 Conclusions and predictions 62
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
4. Methods 65
4.1 Subjects 65
4.2 Procedures 66
4.2.1 Working Memory and Cognitive Assessment 66
4.2.2 Semantic Ability Assessment 68
4.2.3 Independence of Measures 74
4.2.4 Speech Production Assessment 77
5 Analyses of Spontaneous Speech 80
5.1 Methods 81
5.2 Dysfluencies in spontaneous speech 83
5.3 The distribution of lexical choice errors 86
5.4 The etiology of lexical errors 90
5.4.1 Analyses of the etiology of NC speech errors 90
5.4.2 Analyses of the etiology of AD patients’ speech errors 91
5.5 Discussion of results 94
6 Analyses of Constrained Production 98
6.1 Method 99
6.2 Subjects' Analyses 103
6.2.1 Difficulties with the task 103
6.2.2 Dysfluency rates and rejected responses 107
6.2.3 The distribution and etiology of lexical errors 112
6.2.4 The distribution and etiology of misused stimuli 124
6.2.5 The syntactic complexity of responses 128
6.2.6 Summary of subjects’ analysis results 132
6.3 Item analyses on CP stimuli 136
7 Discussion and conclusions 141
7.1 Discussion of results 143
7.2 An Interactive, Distributed Model of Speech Production 154
7.3 Unaccounted For Results 166
7.4 Summary 169
Appendix A. Demographic information on subjects 174
Appendix B. Naming errors. 175
Appendix C. Raw numbers of errors by type of error and group. 180
Appendix D. The CP stimuli 181
Appendix E. Samples of CP responses showing dysfluencies 183
Appendix F. Samples of CP responses showing rejected responses 184
Appendix G. Samples of CP responses showing omitted closed class items 185
Appendix H. Samples of CP responses with substituted or omitted stimuli 186
References 187
iii
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
List of Tables
Table 1. Comparison of AD and Control Groups.
Table 2. Scores on the three WM measures (proportion correct).
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
List o f Figures
Figure 1. An example of a spreading activation pattern in semantic space from
Duchek and Balota (1993). 25
Figure 2. False Start Rates in spontaneous speech. 84
Figure 3. The frequency of errors in spontaneous speech by word type. 88
Figure 4. Mean omission and substitution rates, combined across groups. 89
Figure 5. Sample of the stimuli used in the CP task. 101
Figure 6. Distribution of difficulties with the CP task across AD subjects. 105
Figure 7. Frequency of two types of speech dysfluency in two tasks. 108
Figure 8. The number of AD subjects’ fresh starts in the CP task was highly
related to the proportion of their Within-Category naming errors. 110
Figure 9. Comparison of AD and NC error frequencies between tasks. 114
Figure 10. Subjects’ closed class error rates across tasks.. 115
Figure 11. The relative error rates in the PN of NC subjects and AD subjects
who were high or low omitters of closed class words in the CP task. 117
Figure 12. Closed class omissions as a function of semantic impairment in the
high and low omitter groups. 1! 9
Figure 13. Distribution of AD subjects’ errors across CP stimulus conditions. 121
Figure 14. The relative frequency of the four types of stimulus misuse errors
in the CP by AD subjects. 126
Figure 15. The proportion of passives in each CP condition by group. 130
v
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Acknowledgments
Those people who have had the most influence on me over the last several years are the
participants of Coglab ‘93 - ‘95. I would like to thank you all for hours of stimulating
discussions and lots of fun. In particular, I need to say thank you and au revoir to Laura
Gonnerman. Mike Harm, Joe Devlin, Joe Allen, and Donna LaVoie. I’ll also miss Sarah
Schuster, Karen Marblestone, Laila Lalami, Karen Stevens, Lynn Stallings. Robert
Thornton and Mariela Gil.
I would especially like to thank Elaine Andersen, my adviser, for her patience and her
insistence on clarity of thought, as well as her friendship over the last six years. Dan
Kempler has also been a major help, particularly in learning the value of clear, concise
writing. In addition, I would like to thank Maryellen MacDonald for her unique ability to
quickly grasp what our real questions were, and then proceed to answer them, or to lead
us to answer them ourselves! Mark Seidenberg had a major influence also in guiding the
formation of new connections and new ideas, and encouraging us all to think for ourselves.
Going back to beginnings of my graduate school odyssey, I send my gratitude to Angela
Della Volpe and Pam Ballinger from the California State University at Fullerton who
encouraged me to continue my education.
I could not have made it through these years at USC without the unfailing support of my
husband, Randy, who thought that two years at Cal State Fullerton would lead directly to
a “real job,” and was rather surprised (and dismayed, perhaps) when it led to five more
years of graduate study! My two wonderful daughters. Sharon and Becky, also did
everything they could to allow me to accomplish this feat, including lots of housework
and cooking! I have been very encouraged by their attitude that it was "Cool” that their
mom was doing this, when they easily could have resented how busy and distracted I
have been for the past several years. My parents, Shirley and A 1 Petronis, also were
continually supportive and encouraging throughout the process. They always taught my
siblings and me that dreams were for pursuing, not sighing after, and without this early
training I would never have been able to consider going back to school at my (advanced!)
age.
Finally, I would like to thank Vem Bengtson, for conceiving of and continuing the Multi-
Disciplinary Training in Gerontology, which allowed me pursue my dream here at USC.
The multi-disciplinary seminars are completely responsible for giving me an appreciation
of the discipline of Gerontology and the plight of the elderly. Linda Hall and Chris
Hilgeman were unfailingly helpful and cheerful in making everything to do with this grant
run smoothly.
This research was supported by NIA grant #2t32ag00037to Vem Bengtson and
Department of Health and Human Services Public Health Service grant # 5348-28-2560.
Maryellen MacDonald, principal investigator.
v i
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Dedication
This work is dedicated to
the subjects who willingly participated
in the study and their families.
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Abstract
Effects of Working Memory and Semantic Impairment on Speech in Alzheimer's
Disease
This dissertation evaluates the relative adequacy of modular versus interactive
theories of language processing to explain speech production patterns exhibited by
patients with Alzheimer’s Disease (AID) patients. Two aspects of modularity are
explored: 1) the view that language processing is encapsulated from the effects of
individual differences in extra-linguistic resources, such as working memory (WM): and
2) the claim that syntax and semantics are independent or autonomous components of
grammar. Because claims of selective impairment of semantic and/or memory abilities in
AD have been cited as support for modular accounts, the current study investigated
whether speech production data support such claims by comparing the performance of
mild-moderate AD patients and healthy elderly controls on a range of language and
cognitive tasks.
The findings indicate that AD subjects exhibit impairments in several aspects of
speech production usually considered part of syntactic processing, (e.g., closed class
word use and sentence construction ability), and that the frequency of both the syntactic
and semantic errors in their speech was related to the severity of their semantic
impairment. In addition, individual differences in WM accounted for significant portions
of the variance in several types of errors by both AD patients and healthy elderly people.
These results, therefore, are incompatible with any theoretical account that
assumes a completely modular language system in which semantic and syntactic
processing proceed independently, and in which WM has no effect on speech production
processes. Instead, they support a more interactive account of language processing in
which semantic and WM abilities interact during the production of speech. In order to
account for the range of findings reported in this thesis, I propose a model of speech
production based on connectionist principles, which specifies that: WM plays a major
role in the activation of the words that comprise an utterance, all words (both open and
closed class) are semantically represented, and the syntactic information associated with
words is either part of or activated from their semantic representations.
Key words: working memory, semantic impairment, speech production. Alzheimer's
Disease, aging, closed class words, dysfluencies.
viii
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Effects of W orking Memory and Sem antic Im pairm ent
on Speech in Alzheim er's Disease
Chapter 1. Introduction
O ne of th e m ost controversial issues in psycholinguistics concerns
the m odularity of language processing. H ow ever, the term "m odularity"
has tw o distinct m eanings. It may refer to the theoretical structure of the
language faculty itself, in which each type of language processing, (e.g.
semantics, syntax, and phonology) is accom plished by non-interacting,
inform ationally-encapsulated mechanisms. A lternatively, it can refer to
the separation of language processing from the effects of other cognitive
processes, such as w orking memory (WM) (e.g., Blumstein, 1988; Pinker,
1988, 1994). In the first definition, im pairm ents of sem antic processing, for
example, should have little or no effect on aspects of language provided by
the syntactic processor, such as grammatical form atives and sentence
forms. The latter definition of m odularity predicts that impairments of
WM should n o t have m easurable effects on speech production. This
thesis explores aspects of both of these issues by exam ining the effects of
im paired WM and sem antic ability on the fluency and accuracy of speech
production. To accom plish this, I compare the perform ance of a group of
m ild-to-m oderate A lzheim er's Disease (AD) patients and a group of age-
and education-m atched healthy elderly subjects on both language and
m em ory tasks. These two groups are especially suitable populations for
this study because AD patients at the same general level of cognitive
im pairm ent often dem onstrate different degrees of im pairm ent in
sem antics and W M (M artin, 1990), while healthy elderly people have no
I
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
semantic im pairm ent and show a w ider range of WM abilities than
younger subjects (Salthouse, 1994; Kempler & Zelinski, 1994).
Data analysis targets predictions based on two established theories of
speech production, those proposed by Garrett and Levelt, which are
explicit in their characterization of speech production regarding those
processes w hich theoretically should or should not be vulnerable to
semantic im pairm ent. For example, both assert that deficits in semantic
processing should have no effect on the accuracy of operations carried out
by the syntactic processor. Thus, in either theory an im pairm ent of
semantic ability should have no effect on the grammaticality of speech
output, but the theories of these two researchers differ with respect to the
degree of separation that exists betw een semantic and syntactic processing,
a difference w hich will be clarified in Chapter 2. Regardless, other
psycholinguistic researchers have questioned the assertion that semantics
and syntax do not interact (e.g., MacW hinney & Bates, 1989; Bates &
Wulfeck, 1989; M cDonald, Bock & Kelly, 1993; MacDonald, Pearlm utter &
Seidenberg, 1994). The current study explores this issue by exam ining the
errors m ade by AD patients on aspects of speech production typically
defined as syntactic and comparing these error rates to subjects' degree of
sem antic im pairm ent.
W ith respect to the m odularity of language processing from WM
resources, G arrett ignores this possibility in m ost of his early w riting (1975,
1982,1984), b u t discusses it briefly m ore recently:
A related and particularly difficult issue concerns the role of
attentional and memory factors in norm al error processes.
There is excellent reason to identify certain of the lexical error
processes in norm al language perform ance with failures of
i
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
attention a n d /o r the w orking m em ory system s that control
discourse. H ow those factors m ay enter into aphasic error
production is largely unknow n, though it is certain that they are
of considerable significance. (Garrett, 1992, p. 150)
Thus, he adm its the probability of WM deficits leading to speech errors in
aphasics, b u t does n o t pursue the issue further. Levelt, however, limits
the possible influences of im paired WM to post-production processes,
specifically to self-m onitoring (1989). Several other researchers, however,
have explicitly explored the relationship betw een language use and WM
ability (Danem an & C arpenter, 1980, 1983; Power, 1985; Danem an & Green,
1986; D anem an, 1991; M iyake, Carpenter & Just, 1995), w ith results
suggesting that WM m ay play a crucial role in both language
com prehension and production.
This thesis addresses these conflicting hypotheses and makes its
prim ary goal the determ ination of whether different types of speech
production difficulties, such as inaccuracy of lexical choice, ungram m atical
speech and reduced fluency, are related to sem antic im pairm ent, WM
lim itations or both.
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
1.1 Theories of Speech Production and W orking M em ory
As discussed briefly above, the tw o prom inent theories of speech
production described in this thesis are based on the assum ption that
different levels of linguistic processing occur in separate, inform ationally
encapsulated m odules (Garrett, 1975, 1982; Levelt, 1989). Researchers
w orking in this paradigm attribute all the speech errors of norm al and
pathological populations to m alfunctions of different linguistic m odules;
thus, errors in gram m ar are blam ed on faulty m orphosyntactic processing,
while errors of lexical choice are attributed to problem s in the sem antic
processing m echanism (Fromkin, 1973; G arrett, 1980, 1984). G arrett (1975,
1982) proposes that speech production relies on tw o separate pathw ays
dedicated to sem antic and syntactic processing respectively, that process
inform ation sim ultaneously. In his m odel, closed class w ords are
classified as gram m atical, or syntactic, elem ents; consequently, errors
involving these w ords are attributed to syntactic deficits. Similarly, w ord
omissions are labeled as syntactic errors involving the sentence fram e
(Garrett, 1982; Levelt, 1989). W ithin this fram ew ork, any speech errors
involving closed class w ords or lexical om issions are particularly
im portant in distinguishing between sem antic and syntactic deficits.
Levelt, on the other hand, distinguishes betw een closed class w ords w ith
sem antic content, such as m odals and locational prepositions, w hich, he
proposes, are accessed by semantics and purely grammatical closed class
words, such as auxiliaries and determ iners, that are accessed by syntactic
building procedures. In contrast, other theorists suggest that all closed
class w ords are sem antically represented, and, as such, m ight be
vulnerable w hen there is a semantic im pairm ent (Goodglass & M enn,
4
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
1985; Bates & Wulfeck, 1989). Consequently, closed class errors and
om issions of all types of w ords are of particular interest in this
dissertation, because they will help to distinguish betw een these three
theories.
The encapsulation of language processing from other cognitive
resources has been challenged repeatedly in recent years by researchers
w ho have found strong associations betw een many different language
processing abilities and WM capacity. W orking m em ory capacity has been
found to predict overall vocabulary size, w ord learning ability, and reading
com prehension in young adults (D anem an & Carpenter, 1980; Daneman
& Green, 1986), as well as vocabulary size and reading ability in children
(Yuill, Parkin, Oakhill, 1986; Gathercole & Baddeley, 1993). In addition,
several studies using elderly subjects have shown that lim itation of WM
capacity can also affect discourse processing and the establishm ent of
reference for pronouns in running text (Ulatowska, H ayashi, Cannito, &
Fleming, 1986; Pratt, Boyes, Robins, & M anchester, 1989). A lthough most
of the research exploring the relationship between WM and language use
has focused on the effects of WM on language com prehension, a few
researchers have also examined speech production. These researchers
have found that higher WM scores are associated with faster lexical access,
richer vocabulary, more creative speech, and better recall of one's own
utterances. They have also speculated, but not tested, the possibility that
other aspects of speech, such as the accuracy of lexical choice and the
overall fluency of production, m ight be vulnerable if WM w ere limited
(D anem an & Green, 1986; Danem an, 1991; Power, 1985). The current
research continues this line of research by examining the lexical errors and
5
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
speech fluency of AD patients and healthy elderly speakers for effects
associated w ith lim ited WM, as well as sem antic deficits.
1.2 Speech Production in A lzheim er's Disease
Speech production is often described as being "preserved" in AD
patients, although m ost researchers have focused prim arily on the fluency
of production and the range of syntactic constructions used in
spontaneous speech (Bayles, 1982; Kempler, C urtiss & Jackson, 1987; Nebes,
1989; Kempler, 1991; Bates, Marchman, H arris & W ulfeck, 1994).
However, there is an acknowledged deficit in the accuracy or specificity of
lexical choice w hich does not appear to affect fluency (Hier, Hagenlocker,
& Shindler, 1985; Nicholas, Obler, Albert & Helm-Estabrooks, 1985;
Kempler & Zelinski, 1994). It is a common assum ption that AD subjects'
errors of lexical selection and their use of "em pty speech" in spontaneous
conversation are the result of a semantic deficit which im pairs the use of
more specific w ords first, leaving access to m ore general w ords intact. In
truth, this assum ption has not been experim entally tested. By doing so,
this thesis adds to the knowledge of speech production processes, in
general and the speech production ability of A D patients, in particular.
Semantic im pairm ent is also blam ed for the difficulties that AD
patients have w ith confrontation nam ing tests in which they comm only
produce category coordinates (items from the sam e semantic category) or
superordinate term s (the category name) instead of the target w ord (Huff,
Corkin & G row don, 1986; Nebes, 1989; Smith, Faust, Beeman, Kennedy &
Perry, 1996). D uring picture naming tasks AD patients' responses are often
hesitant and slow w ith m any false starts and w o rd substitutions, m aking it
6
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
quite different from their fluent spontaneous speech. This is som ew hat of
a paradox--that sem antic im pairm ent in A lzheim er's patients should
cause hesitant, often dysfluent speech in single-word production tasks like
picture nam ing, b u t lead to fluent b u t em pty speech in a conversation. It
is not clear w h y or how semantic im pairm ent should have such different
effects in different situations. One possible explanation is that there exists
a trading relationship between lexical accuracy and fluency for these
speakers (Kem pler, 1991). According to this hypothesis, in spontaneous
speech AD patients sacrifice lexical accuracy in order to m aintain fluency,
but in tasks requiring the production of specific words, such as nam ing,
their fluency suffers in the quest for lexical accuracy. To test this
hypothesis, the current study com pares spontaneous speech to utterances
m ade in a constrained production task w hich requires subjects to use
specific, relatively comm on w ords in a sentence. I examine error and
fluency rates for effects of semantic (and WM) im pairm ent to see if the
requirem ent of lexical accuracy im pairs fluency.
Also w ith respect to fluency issues, Levelt has proposed that certain
types of self-correction and self-m onitoring m ay require WM, especially
those in w hich subjects m ust backtrack and “retrace" part of an utterance
(Levelt, 1983, 1989). This claim leads to the prediction that the frequency
of retracings m ight actually decrease as a result of limited w orking
memory. An alternative view point states that WM is necessary for
m aintaining the plan for a sentence as it is being articulated (Valencia-
Laver, 1992). In the latter case, lim ited WM w ould lead to increases in
speech dysfluency, since incomplete activation of the set of w ords to
comprise the utterance could lead to the abandonm ent of the sentence in
7
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
progress. This issue is especially interesting in that recent research has
found little evidence of W M effects on the frequency o f speech
dysfluencies of healthy elderly subjects (Valencia-Laver, 1992) or of AD
patients (Lyons, Kemper, LaBarge, Ferraro, Balota & Storandt, 1994).
H ow ever, these two studies used a limited set of WM m easures and did
not take into account sem antic ability, either of w hich m ay account for
their lack of significant results. To more fully explore these questions, I
investigate the possibility that individual differences in WM and semantic
ability m ay affect the fluency of AD speakers and healthy elderly speakers
in both spontaneous and constrained speech settings using a variety of
WM and sem antic m easures.
1.3 O verview of the Dissertation
As just described, this dissertation investigates the effects of
im paired sem antic ability and w orking m emory lim itations on speech
production. In particular, I com pare the speech of AD patients and
healthy elderly subjects w ith respect to the distribution of lexical errors
and dysfluencies and seek to determ ine the etiology of these difficulties.
To foreshadow m y results, this study provides no su p p o rt for the
assertion that closed class w ords are accessed differently from open class,
content w ords, (i.e. nouns, verbs and adjectives). In fact, the error rates of
AD subjects on closed class w ords are as high as their open class error
rates and they show the sam e relationship with sem antic im pairm ent as
errors on open class w ords do. This type of finding is easily explained in a
theory of language representation in which both open and closed class
w ords are sem antically represented, but is much m ore difficult to account
8
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
for if sem antic and syntactic processing occur sim ultaneously in
noninteracting m odules. In addition, this study finds evidence that
individual differences in WM may have an im pact on the frequency of
certain errors in the speech of healthy elderly subjects who, presum ably,
have no sem antic im pairm ent (Kempler & Zelinski, 1994). Further, WM
deficits are found to increase the probability of certain kinds of errors in
the lexically constrained speech of AD patients. To explain these findings,
I describe a m odel of the semantic system based on that described in
Gonnerm an, A ndersen, Devlin, Kempler & Seidenberg (1997) and sim ilar
to that presented in Tippett, McAuliffe and Farah (1995). In this model,
the featural representations of both open and closed class words consist of
both intercorrelated and distinguishing features, and m ore general words,
like category nam es, are represented by the m ost frequent intercorrelated
semantic features shared by the more specific w ords in that semantic
category.
The interdisciplinary nature of this research requires the
integration of a variety of topics, as well as several levels of analysis.
Consequently, the thesis includes two literature review and two results
chapters. In C hapter 2 , 1 present the two dom inant speech production
theories introduced above, along w ith the kinds of evidence they have
relied upon for validation. Also, I contrast the particular type of semantic
representation that is usually accepted by researchers adopting these
theories w ith a com peting version and present the results of several
psycholinguistic studies supporting this latter conceptualization.
9
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
C hapter 3 presents various theories of WM and discusses the few
studies that have looked for WM effects on speech production. In
addition, it presents an overview of the research on speech production in
Alzheim er's Disease. Chapter 4 describes the subjects who participated in
the study, as w ell as the semantic and WM m easures which are used in
later analyses an d the types of errors and dysfluencies that are of interest.
C hapter 5 exam ines the spontaneous speech of the AD and NC
groups to com pare their lexical and gram m atical error and dysfluency
rates. In addition, w here the sam ple size is large enough, I explore the
frequency of both errors and dysfluencies for effects of sem antic and WM
ability. Similarly, in Chapter 6, I analyze subjects' perform ance on a
constrained speech production task in w hich utterances m ust be
generated using three specified stim ulus w ords. AD subjects have several
types of difficulty in this task, including m any om itted closed class items,
increases in dysfluency rates, and difficulties including all three stim ulus
w ords in a single sentence. All types of difficulties are analyzed for effects
of sem antic an d WM im pairm ent and for distribution across stim ulus
conditions. The subjects' analysis is followed by an item analyses to
determ ine if differences in frequency or num ber of m eanings m ight
account for the difficulties that AD subjects had w ith irregular past
participle stim uli.
C hapter 7 presents a discussion of the results and their im portance
for speech production research. The m ost im portant findings contradict
the traditional m odels of speech production; therefore, I outline an
alternative m odel of speech production. In this model WM interacts
10
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
with sem antic activation processes to account for lexical errors in speech
and differences in fluency. I ad o p t a feature-based sem antic netw ork
sim ilar to th at described in C hapter 2 that employs m any of the principles
of connectionist models, including distributed representations, the
im portance of frequency, and graceful degradation of perform ance
(McClelland, Rum elhart & H inton, 1989).
11
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Chapter 2.
Literature Review:
Speech P roduction and Semantic Representation
The m ost prom inent theories of speech production today, those
developed by Levelt (1989) and G arrett (1975,1984), are founded on beliefs that
sem antic and syntactic processing are fundam entally distinct, non-interactive
m odules. The following chapter describes the theories offered by these tw o
researchers, some of their assum ptions, and the type of data that has been
cited in support of their theories. It also discusses the model of
noncom positional sem antic representations on w hich these two theories rely
and contrasts it w ith a distributed, com positional sem antic representation
which, in Chapter 7, is expanded into a m odel of speech production that can
better account for the results of the current study. The last section of this
chapter presents findings from other researchers w hich are difficult to
account for w ithin G arrett and Levelt's theories, b u t which are m uch m ore
easily accounted for using a com positional theory of semantic representation.
2.1 Theories of Speech Production— G arrett and Levelt
An underlying them e in the theories of both G arrett and Levelt is
that the production process includes two stages, sem antic access followed by
phonological access. Both also agree that speech production must include a
certain degree of parallel processing. In addition, they both ascribe to a non
com positional semantics; that is, a word, or w ord lem m a, is the basic unit of
12
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
m eaning (Garrett, 1982,1992; Levelt, 1989). However, Garrett and Levelt
differ on other im portant issues, such as how the syntactic form of a
sentence is generated and how different types of words are accessed. The
two theories and their differences are highlighted below after another
distinction im portant to both theories is introduced, the difference betw een
open and closed class words.
2.1.1 O pen and Closed Class W ords
O ne of the crucial assum ptions in linguistic theory and in the
theories of both G arrett and Levelt is that w ords are processed and accessed
differently depending on w hether they are open class, "content" w ords or
closed class, "function" words. Linguistic theory addresses this distinction
as a difference in function, vulnerability to historical change, and set size.
O pen class w ords (i.e. nouns, verbs, adjectives, and some adverbs) provide
the content of the sentence— who or w hat is involved in the event and w hat
is happening. New items are frequently added to these classes via coining,
relexicalization or borrowing, hence the nam e "open class" words. These
also represent the vast majority of a person's vocabulary. In contrast, closed
class item s, such as prepositions, determ iners, conjunctions, auxiliary verbs
and m odals, and adverbs of degree (e.g. very, rather, quite, som ewhat),
provide the fram ew ork for the sentence that allows speakers to express the
tim e and duration of the event being described as well as, in some
languages, the roles of the people or objects involved. In addition, closed
class item s typically include both individual w ords and affixes. There are
relatively few closed class items in a language com pared to open class words,
13
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
and additions to closed class categories are very rare, w hich is why these are
called "closed" classes.
M uch of the evidence for the open an d closed class distinction comes
from different kinds of speech errors that are associated w ith one of the two
word types (G arrett, 1975). In addition, one of the m ost well-known
findings concerning open and closed class item s is that closed class items are
highly vulnerable w hen language areas in the brain are dam aged. In fact,
certain types of aphasia have been defined as reflecting specifically
gram m atical im pairm ents, because they prim arily seem to affect
gram m atical item s from the closed classes. Thus, several theorists,
including G arrett and Levelt, have postulated that open and closed class
words, and, by extension, semantic and gram m atical inform ation, are
accessed and processed in very different w ays and in distinct parts of the
brain. This evidence is discussed briefly in section 2.1.4. On the other hand
ideas about how closed class word are represented and accessed differ
som ew hat betw een these two theorists; this difference is described below.
2.1.2 G arrett's M odel of Speech Production and W ord Class Differences
In G arrett's m odel of sentence production (1975, 1982, 1984), the
kernel of m eaning to be expressed is generated at the non-linguistic,
message level. The m essage level then activates two independent
processing m echanism s, one which deals solely w ith lexical-semantic access
and one w hich accesses the syntactic form w hich best fits specifications sent
from the m essage level. Both pathw ays have essentially two "levels" of
processing, a functional and a positional one. A t the functional level,
semantic processes first access the sem antic representations, or lemmas, of
14
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
the w ords w hose m eanings correspond m ost closely to the message.
Sim ultaneously, syntactic processes identify w hich them atic roles will be
required, and the tw o m echanism s pool their results to form a "Functional
Level R epresentation" w hich consists of sem antic representations "tagged"
w ith them atic role inform ation. Subsequently, this functional
representation serves as input to positional level processors. At this level,
the sem antic processor accesses the phonological form corresponding to
each lem m a, and the syntactic processor accesses a sentence frame that is
com patible w ith the them atic roles expressed in the functional
representation and th a t contains a "slot" for each w ord (based on Garrett,
1984). The o u tp u t of the semantic and syntactic processors are then
com bined into a "Positional Level Representation," w hich feeds into a
phonological encoding mechanism. Despite the pooling of inform ation at
each level, how ever, there is no interaction betw een the sem antic and
syntactic processing pathw ays at either level, consequences, choices m ade
during sem antic processing, such as the use of a particular verb, should not
affect the determ ination of the thematic roles to be used or the specific
sentence structure chosen to encode those them atic roles.
In this theory, open class words, which provide the m eaningful
content of a m essage, are retrieved in response to the message from the
non-linguistic, "conceptual level;" this m essage sum m ons interm ediate
w ord forms or lem m as1 from the lexicon (Garrett, 1975, 1982). The sem antic
portions of lem m as th a t are related in m eaning are posited to be located
'The lemma is a construct introduced by Kempen & Huibers (1983) and developed by Levelt
(1989) which includes the semantic meaning of a word, pointers to its morphological form,
phonology and information about its syntactic requirements. Though he didn't using the
terminology in earlier papers, by 1992, Garrett had adopted the "lemma" construct.
15
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
close together in semantic space; thus, speech errors in w hich a closely
related w ord is substituted for a target are attributed to a slightly off-target
access attem pt (Garrett, 1984). Further, semantic closeness also is used to
explain semantic prim ing, in th at activation is posited to spread from the
prim e to other w ords in the sam e vicinity.2 On the other hand, closed class
items, including both w ords and bound, inflectional affixes, are
conceptualized as having no sem antic representation at all, an d are not part
of the conceptual message or included in the lexicon, but instead, are
accessed as an integral part of a sentence frame. In sum , G arrett defines
closed class words and inflectional m orphology as purely syntactic elements,
accessed by syntactic mechanisms, while only open class item s are accessed
by lexical-semantic processes. A crucial implication of this theory for the
current study is that any difficulty w ith closed class items is attributed to
difficulties with syntactic processing.
2.1.3 Levelt. Speech Production and W ord Class Differences
Levelt (1989) theorizes that all sentence level speech production is
lexically driven, or m ore accurately "lemm a-driven," and, th at nothing in
the speaker's message will, by itself, trigger a particular syntactic form. He
suggests that it is a function of the particular lexical items chosen and their
order of activation, which together trigger a specific syntactic form, such as
an active or passive, or a prepositional or double object dative.3 A key
concept here is the "lem m a" w hich can be pictured as an organizing node
that anchors pointers to four types of linguistic knowledge about a word: its
2 See the discussion of spreading activation models of semantic representation in section 2.2 for
more detail.
3 Levelt does not, however, specify what will determine the order of activation of different
lexical items.
16
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
semantics, m orphology, phonology, and syntax. Levelt considers the
lem m a level to be equivalent to the lexical level in other speech production
m odels that include a separate w ord-node level (e.g., M ackay, 1987; Dell,
Juliano & Govindjee, 1993). In his formulation of the speech production
process, the message level activates the lemmas of the specific w ords which
best capture the m eaning to be communicated. Once a lemma is chosen, it
first activates its sem antic and syntactic representations and, subsequently,
the other two aspects of w ord representation, m orphology and phonology.
Processing in each of four levels of representation then proceeds w ith no
direct interaction betw een them. In this theory, m orphological variants of
w ords, such as plural nouns or the past tense forms of verbs, are included
w ithin the m orphological sector of the lemma and are distinguished from
each other by diacritics indicating, for example, the tense, m ood, aspect,
person and num ber for a verb. The usage of a particular form is determ ined
either directly, by characteristics of the original message from conceptual
processes (e.g., tense, m ood, and aspect), or indirectly, via pointers from
other portions of the m essage, (e.g., num ber and person diacritics from the
sentence subject). In this theory, not all w ords are activated from their own
m eanings, instead some rely on "syntactic building procedures" to activate
them or specify their m orphological forms. A syntactic building procedure
m ay be activated either from the message level itself, as w ould be the case
for an embedded sentence, or via the syntactic representation of another
lemma. To clarify this, Levelt suggests that activating the lemma for a verb
will activate that lem m a's syntactic representation, w hich will, in turn,
activate the syntactic building procedure for a verb phrase typically used
w ith the verb in question that includes slots for argum ent noun phrases.
17
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
O ther syntactic building procedures m ight call up a prepositional phrase
structure, the determ iner for a noun, or an em bedded sentence
representation, com plete w ith a pointer to the lexical item "that" to signal
the start of the subordinate clause, if the m essage contains inform ation best
expressed using these structures.
Regarding closed class items, Levelt draw s a distinction betw een those
w hich are activated via a semantic representation and those w hich are
activated by "pointers" from syntactic building procedures triggered by the
syntactic representations of other lexical items. Closed class items that are
accessed from their ow n semantic representations include prim arily
locational prepositions, adverbs of degree and m odals. How ever, w ithin
this theory the m ajority of closed class w ords, such as determ iners, verb
particles, com plem entizers, and the "by- phrase" for a passive sentence are
accessed via syntactic building procedures. For example, determ iners are
described as being accessed by a noun phrase building procedure w hen the
syntactic representation of a common noun is activated, and verb particles,
such as the FOR in WAIT FOR, are accessed w hen an intransitive verb is
chosen in conjunction w ith a theme, w hile a com plem entizer, such as THAT
introducing a relative clause, would be accessed by a syntactic procedure
specific to relative clauses (Levelt, 1989, C hapter 6). In sum , this theorist
sees sentence production as being prim arily lem m a-driven, w ith closed class
w ords having an am biguous status, depending on their use in a given
sentence. M oreover, apparently the only interaction betw een the four parts
of the lem m a is a t the lemma level itself, so that, for example, the syntactic
representation of a w ord m ight be dam aged w ithout any harm to the rest of
the lem m a's com ponents. Thus, according to this theory, closed class w ords
18
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
accessed via syntactic building procedures should be unaffected in a
population w ith a sem antic im pairm ent, while an im pairm ent of specific
syntactic building procedures should affect every use of a particular syntactic
form, (i.e. all noun phrases, all relative clauses, or all verb phrases
accom m odating a particular argum ent structure.)
2.1.4 W ord Class Differences as Contrasts in Semantic Richness
An alternative conceptualization of the differences betw een open
and close class w ords, and the one preferred in this study, attributes this
distinction to differences in the richness of m eaning and frequency of
occurrence. This conceptualization of closed class w ords depends crucially
on the use of a com positional sem antic system, rather than the non-
compositional one suggested by Garrett and Levelt (see section 2.3 below).
In this alternative view, closed class items are postulated to be
sem antically encoded, b u t to have many fewer sem antic features than
open class w ords (Goodglass & Menn, 1985; Plaut & Shallice, 1993). In
addition, the types of features associated w ith closed class items m ay be
qualitatively different than those associated with open class items, in that
the m eanings of the form er w ord class are largely relational and abstract
(e.g., tense, aspect, case), rather than perceptual (e.g., HAS STRIPES, HAS a
MANE) or functional (USED-TO-INSTALL-NAILS, USED TO TURN SCREWS)
(Bates, Friederici & W ulfeck, 1987; Bates & Wulfeck, 1989). O ther
differences between the tw o w ord types are attributed to frequency
differences that affect the speed of activation and activation decay of very
high frequency w ords differently than w ords of low er frequency (Gordon
19
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
& Carram azza, 1982, 1983; Hirshm an & Palij, 1992). This theory and its
ramifications are discussed below in section 2.3.
2.1.5 Evidence for M odular Speech Production Theories
As m entioned above, the two prim ary sources of evidence for
G arrett's and Levelt's theories are speech errors from both healthy speakers
and from persons w ith certain kinds of brain dam age. Garrett has been
responsible for m uch of the research of speech errors and believes that
healthy speakers in norm al conversation m ake qualitatively different types
of errors on open and closed class w ords (1975, 1980,1982). He asserts that
open class w ords participate in "sound shift" errors and "exchange" errors,
but closed class w ords do not. Instead, closed class elements, both free w ords
and bound m orphem es, are only susceptible to "shift" and "m orphem e
stranding" errors. The distinction betw een exchange and shift errors has to
do with the num ber of elements that are o u t of place. In an exchange error,
two open class w ords, usually from the sam e gram m atical class, exchange
places in the output, i.e. "give my bath a hot back" instead of "give m y back
a hot bath" (this an d the following exam ples are from Garrett, 1975). In a
shift error, the w ord order of the utterance is affected in that a single closed
class w ord is out of place, for instance, saying "W ho w ould else like one?"
instead of "W ho w ould else like one?" Even bound m orphem es m ay
appear to shift, such as in the example, "he gets w eird everier day," instead
of "he gets w eirder every day." In som e errors, open class words are
exchanged, apparently leaving the bound m orphem es where they should
have been in the correct utterance, a process referred to as "m orphem e
stranding." Some exam ples of this from G arrett (1975) are:
20
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm ission .
A ctual -> Target
it just so u n d ed to start -> it just started to sound
orange juice is quench thirsting -> orange juice is thirst quenching
the sm ash light he flashed -> the flashlight he sm ashed
G arrett explains the difference betw een these patterns by postulating
that open class w ords are assigned to their slots in a sentence fram e and
realized phonologically at an earlier stage of the production sequence than
closed class w o rd s are. Consequently, exchange errors and m orphem e
stranding occur w hen open class w ords are assigned to the w rong sentence
frame slots w hich are already m arked w ith pointers to closed class
morphology. Sim ilarly, sound shift errors are said to take place as open
class w ords are being phonologically coded, which again is at the earlier
stage of the production process before closed class items are phonologically
realized. In sum m ary, in G arrett's theory closed class w ords are seen only as
syntactic m arkers w ith no m eaning w hich are encoded phonologically very
late in the production process, after open class w ords have already been put
into position and phonologically specified.
Levelt distinguishes betw een speech errors that involve the
misplacem ent of w ords in a sentence and those that involve errors at the
different linguistic levels, such as the phonological level (i.e. sound
exchanges), the lem m a level (word association substitutions), and the
conceptual level (sem antic substitutions). In general, he attributes speech
errors to glitches in the parallel processing m echanism due to the
sim ultaneous activation of several lem m as, m eanings, or w ord forms. He
does not address errors that involve bound m orphem es, such as the
stranding errors described above.
21
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
The second type of evidence for G arrett's and Levelt's theories comes
from different patterns of language im pairm ent d u e to brain dam age. Two
syndrom es are traditionally used to highlight the apparent processing
distinction betw een open and closed class item s, agram m atism (traditionally
called Broca's aphasia) and paragram m atism (also known as W ernicke's
aphasia). In the agram m atic speech of English-speaking aphasics, utterances
are nonfluent w ith hesitant and often labored w ord production; open class
w ords such as nouns and m ain verbs are produced (though w ith difficulty),
but closed class w ords, such as determ iners, verb auxiliaries, and
prepositions, are often om itted (Davis, 1983). A gram m atism is considered
prim arily a disorder of speech production (G arrett, 1992), although these
subjects also often have deficits in the com prehension of com plex syntactic
constructions (e.g., Caram azza & Zurif, 1976; Caplan, Baker & Dehaut, 1985;
Caplan & H ildebrandt, 1988; Blumstein, 1988). Additionally, m any
agram m atics have difficulty reading closed class items aloud, w ith the m ost
com m on error type being their omission; this variation of agram m atism is
called deep dyslexia (Garrett, 1992). Thus, in this subtype of agram m atism
the difficulty w ith the production of closed class item s is found in two
modalities, reading and spontaneous speech. The agram m atic pattern of
speech production is often cited as evidence that open and closed class
w ords are processed differently in the brain, based on the reasoning that, if a
certain type of brain dam age can disproportionately affect only closed class
w ords, the dam aged part of the brain m ust be specialized for the production
of these w ords. In contrast to agram m atics w ith their sparse, labored speech,
paragram m atics are described as fluent speakers w ho m ake large num bers of
both open and closed class w ord substitutions, and who also suffer from
22
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ith out p erm issio n .
significant im pairm ents in the com prehension of every day language
(Garrett, 1992; Bates, Friederici & Wulfeck, 1987; Davis, 1983, am ong others).
A fair description of the speech production im pairm ent in paragram m atism
m ight be, "the inappropriate use of both open and closed class vocabulary
[with] a loss of the ability to select those elements that correspond to a target
meaning" (Garrett, 1992, p .168). Importantly, the loss of the ability to
identify the appropriate w ord corresponding to a target m eaning extends to
both com prehension and production in this group of aphasics, so overall,
paragram m atics are m uch m ore severely im paired than Broca's aphasics.
These tw o aphasic syndrom es have several characteristics in
common besides the fact that they involve dam age to language centers in
the left hem isphere of the brain. For example, both types of aphasic patients
suffer from som e degree of word-finding difficulty w ith open class w ords in
spontaneous conversation and tests of confrontation nam ing (Davis, 1983;
Butterworth, 1992). In addition, they make errors on both open and closed
class w ords (G arrett 1992), though the pattern of errors in term s of
substitutions an d om issions m ay differ depending on language and
syndrom e (Bates, Friederici & Wulfeck, 1987). M oreover, as just pointed
out, closed class item s are especially vulnerable to error in these patients
(Bates & Wulfeck, 1989). The two groups also exhibit varying degrees of
com prehension deficits, w hich range from relatively m inor difficulties
decoding com plex syntactic structures in agram m atics to pronounced
deficits of even single w ord com prehension in severe paragram m atics. This
com prehension im pairm ent is not all-or-nothing; that is, to qualify as
paragram m atic on the W estern Aphasia battery, a patient m ust score
between 0 and 6.9 on the com prehension section of the battery, w hile to
23
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
qualify as an agram matic, the patient m ust score between 4 and 10 (Davis,
1983). Thus, despite traditional descriptions of preserved com prehension in
agram m atism and im paired com prehension in paragram m atism , there may
be considerable overlap in com prehension skills between these two groups.
A nother im portant sym ptom th at these syndrom es have in comm on is that
the language deficit is accom panied by some loss of working memory (WM)
(Davis, 1983). This is im portant because, as pointed out in C hapter I, WM is
thought to play an im portant role in language comprehension (Davis, 1983;
Just & C arpenter, 1992; Gathercole & Baddeley, 1993; Small, 1994;
H enderson, Kratz, Smith, & Buckwalter, 1995), and there is evidence that it
may affect speech production as well (Daneman & Green, 1986; Power, 1985).
In sum , there are several com m on characteristics shared by these aphasic
populations, any of which m ight contribute to their language difficulties.
2.2 Two Views of Semantic Representation
Both G arrett and Levelt ascribe to the theory of noncompositional
sem antics; that is, the w ord itself is the sm allest division of meaning. This
is the view of semantic representation that is also adopted by most
researchers w ho study sem antic prim ing. In this theory, each w ord is
associated w ith a single node w hich has connections to all the w ords which
share elem ents of m eaning or use w ith it (Mackay, 1987; Balota & Duchek,
1993). C onsider the model in Figure 1 taken from Duchek and Balota (1993).
Like other m odels of its type, this one has single nodes for each word and
the connections between the nodes represent shared characteristics or
sim ilarities in m eaning, so TEACH and TRAIN share a verbal m eaning and so
are connected, b u t TEACH and TRANSPORTATION do not, so there is not a
24
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
connection betw een them . In addition, the connections are weighted
according to the closeness in m eaning of the tw o w ords and the frequency
w ith which the two w ords are associated (Nebes, 1989; O ber & Shenaut,
1995). In system s such as this, access of a w ord refers to the activation of a
specific node.
Figure 1. An example of a spreading activation pattern in semantic space
from Duchek & Balota (1993).
TEACH TRANSPORT COLOR RACE PART OF DRESS SOOT
2.5
2.5
9.5
TRAIN BLACK
In sem antic prim ing experim ents, activation is posited to
autom atically spread from the prim ed node to all the other nodes with
w hich it is connected. This increases the activation of all of these other
nodes, and subsequently allow s m ore rapid activation and retrieval of any
of them (Kempler & Zelinski, 1994). This is exem plified in Figure 1, in that
w hen TRAIN is heard, the nodes connected to it are also activated, in
proportion to the weights on the connections betw een them. In fact, this
diagram implies that the total activation value of a prim ed word (the
num bers in the circles) is num erically divided betw een all the other nodes
w ith which it is connected, so that each related node is activated to a m uch
low er degree than the original item s themselves. This conceptualization of
25
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
spreading activation is related to the hypothesis shared by G arrett and Levelt
that proxim ity in sem antic space accounts for prim ing, in that closeness in
sem antic space is instantiated as th e strength of the connections betw een
w ords. M oreover, this model captures the notion that, in order for
activation to sp read , there m ust be som e conduit or connection through
w hich it can pass.
This theory has difficulty explaining certain im portant phenom ena.
For exam ple, if the node for a w ord, e.g. TRAIN, or the connections to it are
dam aged, this should limit the am ount of prim ing that could be
accom plished from the dam aged node. For instance, it w ould seem likely
that dam age to a node would block the spread of activation to other nodes
and result in reduced or absent prim ing of a related target (W arrington &
Shailice, 1979; Shallice, 1988). H ow ever, several prim ing studies have
found evidence of relatively norm al prim ing in patients w ho have
docum ented sem antic im pairm ents, such as people w ith A lzheim er's
Disease (e.g. N ebes, 1989, 1992; O ber & Shenaut, 1995). Furtherm ore, this
theory w ould predict that subjects w ould be unable to produce a response in
a nam ing task w hen the target w o rd 's semantic node was dam aged.
How ever, the m ost comm on type of nam ing errors m ade by both normal
and sem antically im paired populations is the substitution of a closely
related w ord from the same sem antic category (Albert, H eller & Milberg,
1988; Nebes, 1989) rather than another associated word. A long these same
lines, dam age to a semantic representation would be expected to result in
consistent errors on the same w ord across test times a n d /o r tasks
(W arrington & Shallice, 1979; Shallice, 1988). However, this criterion has
also led to am biguous findings (Rapp & Caram azza, 1993), since m ost
26
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
sem antically im paired subjects make both consistent and inconsistent
errors. M oreover, this m odel cannot explain the finding that sem antically
related w ords are activated sim ultaneously w ith and to the same extent as
the prim e, w ith no time required for activation to spread to other nodes
(e.g., Experim ent 6 in Levelt, Schriefers, Vorberg, Meyer, Pechmann, &
Havinga, 1991).
There is an alternate conceptualization of the semantic system that
offers a m ore felicitous explanation of the above phenomena. Instead of
each w ord being represented by a single node which has connections to
other related concepts, w ords may be accessed and represented by the
activation of a constellation of semantic features (Bates & Wulfeck, 1989;
Butterworth, 1989; Farah & McClelland, 1991; Dell & O'Seaghdha, 1992; Plaut
& Shallice, 1993; G onnerm an, A ndersen, Devlin, Kempler & Seidenberg,
1997). In this type of system, hearing or reading a word activates the set of
semantic features associated with it, and, w hen two words are sem antically
related, they share semantic features so their patterns of activation are
similar. As a result, w hen the target is encountered in a semantic prim ing
task, several of its features have already been activated due to exposure to
the related prim e, and this facilitates the subject's response. In this theory,
the prior activation of any group of shared features will facilitate access of
the target, though the am ount of facilitation m ay vary depending on the
am ount of feature overlap between the tw o meanings. If there is dam age to
the connections to some of the features of the prim e making them difficult
to activate, it will not affect prim ing of the target as long as there are still
som e features in com m on between the prim e and target that can be
activated. This theory explains the discrepancies in the patterns of results
27
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
different betw een on-line prim ing tasks and off-line sem antic tasks in
which on-line results suggest preservation of semantic ability, but off-line
results indicating a semantic im pairm ent, as the result of task differences.
To illustrate, w ithin this theory off-line semantic tasks require a different
degree of precision in the activation of features com pared to prim ing tasks.
In order for an object to be nam ed correctly, for instance, the set of features
which is activated m ust not only describe the item in question, it m ust also
be sufficient to distinguish it from other similar items. Consequently, if the
features HAS 4 LEGS, 2 EARS, TAIL, and TEETH were activated by a picture of
HORSE, these four features w ould not be sufficient to determ ine a discrete
output in nam ing or w ord-picture matching; there are too m any animals
that share these features. It w ould also be necessary to activate some other
distinguishing feature, such as CAN BE RIDDEN or USES A SADDLE, in order for
the correct w ord to be produced. However, in a prim ing task, this same set
of four features (i.e., HAS 4 LEGS, 2 EARS, TAIL, and TEETH) w ould facilitate the
retrieval of any items sharing these features, e.g. SKUNK, CAMEL, DOG and
COW. In sum m ary, in this theory of semantic representation it is
contrasting task dem ands that account for different levels of performance in
on- versus off-line semantic tasks, rather than different types or locations of
dam age w ithin the semantic system .
An interesting advantage of a feature-based model of semantics is
that it also accounts for two apparently anomalous results in the priming
literature (Levelt et al., 1991): 1) there appears to be no time separating the
activation of the target item and a category coordinate, (i.e. there is priming
of category coordinates as soon as there is priming of the target item) and 2)
category coordinates of the target w ords were primed to the sam e extent as
28
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
the target itself. Levelt et al. do not address the im plications of this finding
for their theory of noncom positional sem antics, since the goal of the study
w as to determ ine w hether there was phonological activation of category
coordinates of a target w ord during lexical access. Nevertheless, adopting a
theory of com positional semantics avoids the rath er em barrassing situation
of having to posit 1) that activation takes no tim e to spread to related w ord
nodes and 2) that activation can spread to any num ber of category
coordinates w ith o u t being diminished.
One draw back to this theory of sem antic representation is that, to
date, it has only been systematically applied to single w ord production.
Therefore, predictions about single w ord, lexical substitution errors, for
example, are easy to infer, but predictions about sentence length utterances
are m ore difficult. A com patible approach w ould be to define speech
production as the quintessential constraint satisfaction task, in which the
sentence produced m ust satisfy the constraints provided by (at the very
least): the conceptual message itself, the lim itations of the speaker's
vocabulary, the gram m atical requirements of the particular language being
spoken, and the pragm atics of the situation. In addition, other probabilistic
inform ation m ay affect the final sentence structure, such as the relative
frequencies of different sentence forms associated w ith the particular type of
verb chosen for use, or the animacy of the argum ents of the verb. As a
result, several alternative ways of expressing the sam e thought m ay be
initially activated and then rapidly restricted by various constraints
(MacDonald, 1993). In theories based on these ideas, the particular syntactic
structures w hich are frequently used w ith a w ord are considered to be part of
a its lexical representation (Tanenhaus & Carlson, 1989; MacDonald,
29
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Pearlm utter & Seidenberg, 1993). Consequently, im pairm ents of lexical
semantic ability m ight be expected to disrupt syntactic processing. This
contrasts directly w ith the position taken by researchers w ho assert that
different levels of linguistic representation, especially sem antics and syntax,
are independent m odules, and so either sem antic o r syntactic impairm ent is
predicted to occur w ith no effect on the other type of processing. Most of the
researchers w hose w ork is presented in the follow ing section ascribe to
theories of com positional sem antic representation an d sentence production
similar to th at described here.
2.3 Research that Supports C om positional Sem antic Representations
G arrett's and Levelt's theories explain m any phenom ena associated
with speech production, such as the speech errors of English and Dutch
speakers and the patterns of errors m ade by English and Dutch aphasics.
However, there are several findings that are not easily explained by these
two theories, including: the effects of agram m atism and paragram m atism
on speakers of inflecting languages, im ageability effects on speech
production, the im m unity of closed class w ords to inter-sentential priming,
and frequency effects on lexical access and recognition memory. These
effects seem to be m ore easily accounted for in a theory of compositional
semantics in w hich w ords that overlap in m eaning activate many of the
same features, as described above.
2.3.1 Cross-linguistic studies of aphasia
M uch of the w ork on crosslinguistic studies of aphasia has been done
by Bates and h er colleagues who ascribe to a com positional theory of
30
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
semantics, as described above (Bates & Wulfeck, 1989). They characterize
both open and closed class w ords as being com posed of different num bers
and types of sem antic features, and postulate that a w ord is accessed when
these features are activated by a message from conceptual processes. They
argue that the problem w ith traditional characterizations of agram m atism
and paragram m atism is that they are largely English-specific; aphasics who
are native speakers of languages other than English show quite different
patterns. That is, w hile it is true that English-speaking agram m atics omit
m any closed class item s, agram m atic speakers of highly inflected languages,
such as Hebrew, Italian and German, make more su b stitu tio n s than
omissions of gram m atical forms (Bates et al., 1987; Bates & Wulfeck, 1989;
also see Blumstein, 1988). Furthermore, convergent research has found that
English-speaking paragram m atics do not just substitute closed class words,
as formerly believed, b u t also om it them (Heeschen, 1985; Butterworth &
How ard, 1987; M enn & Obler, 1990). Bates and her colleagues conclude that
the pattern of om ission and substitution errors in agram m atic and
paragram m atic speech depends less on the type of aphasia and more on
language specific differences in the am ount of inform ation carried by closed
class items. For exam ple, these researchers docum ent the finding that
German articles, w hich are alw ays required because they carry necessary case
m arkers for their associated nouns, are rarely om itted b u t frequently
substituted by G erm an-speaking agrammatics. They attribute this different
behavior betw een G erm an-speaking and English-speaking agram m atics to
the fact that G erm an articles have more "m eaning," and consequently more
semantic features, than English articles. According to this hypothesis, words
w ith fewer sem antic features w ould be more subject to om ission than w ords
31
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
with higher num bers of features. This idea is discussed m ore fully in
section 2.3.2 on imageability effects. Because of the sim ilarity in the error
patterns of G erm an-speaking agram m atic and paragram m atic aphasics,
Bates and W ulfeck state that the prim ary difference between the tw o aphasic
types is not in sem antic versus syntactic im pairm ent, but instead can be
attributed to differences in the levels of im pairm ent of fluency, prosody, and
com prehension (Bates & Wulfeck, 1989).
These data are not easily accounted for in m odular theories of speech
production, especially that of Garrett, w h o refers to agram m atism and
paragram m atism as representing a double dissociation betw een im paired
semantics and syntax (Garrett, 1984). In his theory, agram matics w ho speak
any language are supposed to have im paired syntactic processing b u t intact
semantics, w hile paragram m atics show supposedly intact syntax and
impaired sem antics. However, Bates et al. (1987) illustrate that agram m atic
and paragram m atic speakers of the sam e language make sim ilar kinds of
errors, and do not show the predicted double dissociation. That is, across
languages, agram m atics do not system atically om it closed class inform ation,
and paragram m atics do not solely substitute words. Furtherm ore, they
point out that by classifying the om ission of w ord final m orphology as an
omission of syntactic material, these theories ignore the fact that, at least in
English, the om ission of word final m orphology, (i.e., the plural -s or the
past tense -ed) alm ost always results in a real word. Consequently, these
errors m ight just as easily be considered to be the substitutions of related
forms of the sam e root— the most com m on type of error m ade by
agrammatic speakers of highly inflected languages (Bates & W ulfeck, 1989).
32
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
O ther studies of non-English speaking aphasics have also offered
different explanations of difficulties w ith closed class item s, rather than
positing that there are specific syntactic deficits w hich result in the om ission
of closed class words. For exam ple, Hofstede & Kolk (1994) have suggested
that their om ission m ight be an adaptation to the extrem e difficulties w ith
the fluent production of speech w hich are characteristic of agram m atics.
They suggest that a subject w ho has difficulty producing any w ords will
naturally concentrate on the content elements of the m essage.4 These
researchers have show n that G erm an-speaking aphasic subjects who, in one
instance, om itted m any closed class words, were able to produce them in
context w hen requested, though w ith increased num bers of substitutions.
In a related study, Kolk (1995) suggests closed class item s are sem antically
represented and have their ow n m orphology which m u st be m atched or
synchronized w ith the available sentence frame. A ccording to this theory,
syntactic and sem antic inform ation m ight become available to speech
production processes at different tim es after language processing has been
affected by a stroke, especially in the case of paragram m atism . Kolk suggests
that this desynchronization results in a subject having m any alternative
m orphological forms activated at sam e time w ith no w ay to choose betw een
them , since the syntactic inform ation necessary for this choice is not
sim ultaneously available. T hus, subjects make m any m ore substitution
4 In some cases, this is likely a result of subjects trying to adhere to the Gricean Principles of
conversation, and not wanting to prolong their (already overly long) conversational turn by
33
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm ission .
errors. Based on these studies, cross-linguistic research has not supported
either the assum ption that agram m atism and paragram m atism represent a
double dissociation betw een semantic im pairm ent and syntactic
im pairm ent or the assertion that closed class elem ents lack sem antic
representations.
2.3.2 Imageabilitv Effects
Theorists have also explored the dim ension of imageability as a
m eans of explaining differences between w ord types, particularly w ith
respect to the part-of-speech effects found in deep dyslexics. In this
syndrom e patients have m ore difficulty reading w ords that are less
imageable: i.e. DOG is easier to image and read than WENT, which is easier
to im age and read than OUT. This issue is particularly salient to closed class
item s since, for the m ost part, they are the least imageable w ords in a
language. Im ageability effects are usually explained in terms of the types
a n d /o r num bers of features which m ake up a w ord's semantic
representation (W arrington & McCarthy, 1987). Highly imageable w ords are
postulated to have m ore features, because m ore imageable w ords have a
num ber of visual-perceptual features in addition to the functional and
encyclopedic features th at m ost words have (Plaut & Shallice, 1993; Farah &
M cClelland, 1991).
spending a lot of time attempting to articulate non-content items.
34
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Plaut and Shallice (1993) have explored imageability effects by
im plem enting a connectionist net that m aps orthography and phonology to
sem antic representations. In this sim ulation, the effects of im ageability are
instantiated by using different num bers of semantic features for imageable
and abstract w ords; they use a m ean of eighteen semantic features for highly
imageable w ords and only five for less imageable words.5 They find that
they can cause a disproportionate im pairm ent in accessing less imageable
w ords w ith m oderate dam age to the direct pathw ays in the net betw een the
input level, orthography, and sem antics. In effect, this results in the input
to sem antics being "noisy" and underdeterm ined, which causes w ords with
fewer features to be less accessible. These computational results suggest two
im portant things about less imageable w ords (including closed class items):
first, they m ay be successfully described as having fewer sem antic features
than m ore im ageable w ords and, second, they may be disproportionately
affected by dam age that has little effect on more imageable w ords.
A n alternate explanation of im ageability effects is based on the
observation that m ore imageable w ords tend to maintain the sam e m eaning
in different contexts, whereas abstract w ords, such as verbs and closed class
items, tend to change m eaning in different contexts (Saffran, Bogyo,
Schwartz, & M arin, 1980). Variability in m eaning would lead to a smaller
set of features being consistently activated for a given lexical item, in effect
yielding separate features sets for each m eaning which only overlap in a
core set of features. In contrast, constancy of m eaning would lead to a larger
proportion of consistently activated features in the semantic representation.
5 Plaut and Shallice (1993) are careful to state that they are not trying to capture the entire
meaning of the words used, but are more intent on representing the relationships between the
words.
35
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
M artin, Saffran and Dell (1996) assert that theories w hich claim that
imageability effects are the result of differing num bers of features and those
that claim it is due to variability in m eaning betw een contexts lead to the
sam e conclusion: highly imageable w ords tend to share m ore features w ith
related w ords than abstract w ords do with other related abstract words.
Thus, imageability effects are relatively easily accounted for w hen sem antic
representations are defined as sets of semantic features, given the finding in
Plaut and Shallice (1993) that the activation of w ords w ith fewer features is
m ore affected by noisy input than w ords w ith m ore features.
In a noncom positional sem antic system , the prim ary challenge
w ould be to explain w hy m any abstract w ords w ere im paired, and in
particular, w hy w ords that are m ore abstract are im paired m ore frequently
than less abstract w ords. In this case, the usual approach is to assert that
abstract w ords are represented in a different location in the semantic
netw ork than concrete w ords, and a lesion has preferentially affected this
area. Presumably, the m ost abstract words w ould be at the focus of this
hypothetical lesion site, w hereas less abstract w ords w ould be on the
periphery. How ever, deep dyslexia, the syndrom e m ost com m only
associated w ith im ageability effects, typically involves large portions of the
left tem poral and parietal lobes and is initially associated w ith global aphasia
w hich only resolves into deep dyslexia w ith rem ediation (Plaut, 1996).
Thus, the lesions that lead to imageability effects can hardly be called focal.
Furtherm ore, while the m ost abstract words, like closed class items, are
often om itted in reading, the ability to read open class w ords is not fully
preserved, as subjects also m ake m any sem antic substitutions w hen reading
these, and often m ake com bination visual and sem antic errors, such as the
36
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
target SYMPATHY being read as ORCHESTRA, presum ably via m isreading of
the target as SYMPHONY (Caplan, 1992). Thus, it appears that m ore of
language system is affected in deep dyslexia than just an abstract w ord
m odule.
2.3.3 Syntactic prim ing and closed class w ords
Prim ing studies have investigated w hether syntactic structures prime
each other, as w ell as w hether sem antically related words prim e each other.
Research has show n that subjects tend to produce similar syntactic
structures to those they have just heard (Bock, 1986; Bock & Loebell, 1990).
Bock (1989) tests G arrett's hypothesis that closed class w ords are accessed as
p art of the sentence fram e by exam ining w hether there is sentence-to-
sentence p rim in g of either sentence fram es, the closed class w ords in them
or both. She reconfirm s that using a particular sentence structure prim es
subsequent usage of the same structure (i.e. a prepositional dative
construction in sentence #1 increases the likelihood of a prepositional
dative in sentence #2), as she had found in other studies. M ore im portantly
for the cu rren t research, this study also establishes that the particular closed
class w ords in a sentence structure do not prim e themselves to appear in a
following sentence. That is, a dative sentence using TO in sentence #1, (e.g.
"The cheerleader offered the seat to her friend"), is associated w ith a
prepositional dative in sentence #2, b u t that dative form is just as likely to
use FOR as TO in the benefactor's prepositional phrase (e.g. "The cheerleader
saved the seat fo r her friend"). Therefore, because prepositions are not
37
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
prim ed betw een sentences but sentence structure is, Bock concludes that
gram m atical prepositions are not inherent to the sentence frame, thus,
arguing against G arrett's assertion to the contrary. She also observes that
closed class item s seem to be less susceptible to prim ing than open class
items or sentence structures, and suggests that this could be related to the
fact that closed class w ords are accessed so fast. This suggestion is based on
the consistently high frequencies of closed class items, and the fact that
increased lexical frequency is consistently associated w ith faster lexical
access, as described below . The implication of Bock's statem ent in this
regard is that w ords that are accessed very fast due to high frequency of use
also decay very fast in memory.
2.3.4 Frequency effects in recognition m em ory tasks
Frequency differences pertinent to this research have also been
observed in studies of recognition memory. In recognition memory tasks,
subjects are given lists of w ords to study, then, after doing a filler task, they
m ust indicate w hether the w ords presented in a new list had been included
in the original study group. It is a common finding that lower frequency
item s are m ore easily rem em bered in recognition m em ory tasks than
higher frequency w ords are and that this effect is independent of the time
subjects are given to study the words initially (H irshm an & Palij, 1992).
This phenom enon offers converging evidence that the increased speeds of
lexical access associated w ith higher frequency w ords m ight also be
associated w ith faster decay of activation in m em ory. If this were true,
38
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
closed class w ords, as a group, w ould be the m ost severely affected by fast
decay, since they are by far the m ost frequent words in a language, which, in
turn, could m ake these w ords very vulnerable to lexical errors, particularly
om issions.
2.3.5 Frequency effects on lexical access
The effects of frequency on lexical access have been explored by
Bradley and G arrett (1983) and G ordon and Caramazza (1982,1983). Bradley
and G arrett (1983) propose that open and closed class w ords are accessed
differently, based on their findings that only aphasics, but not norm al
speakers, show evidence of frequency sensitivity for closed class words,
w hereas both groups of speakers show frequency effects for open class
words. Based on Garrett's theory of speech production, they postulate that
this effect is due to a special, non-frequency-sensitive access m echanism
dedicated to the processing of closed class w ords that has been im paired in
the aphasics. However, one of the difficulties with this type of research is
that there are very few open class w ords that are even close to the frequency
of most closed class words (Bock, 1989), and the few that do overlap include
closed class w ords that only appear in w ritten text rather than speech (e.g.,
THEREUPON, WHEREAS), problems that are not addressed in the Bradley and
G arrett study. Gordon and Caram azza (1982, 1983) investigate the claims of
Bradley and G arrett further but take into account differences in frequency,
and find sim ilar frequency effects in both open and closed class words. That
is, at frequencies of less than 316 per m illion both types of w ords show
sim ilar effects of frequency on lexical access, but higher frequency open and
closed class w ords show the same lack of frequency sensitivity, especially at a
39
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
frequency of over 399 per m illion (using Kucera a n d Francis word
frequencies). That is, reaction times to the m ost frequent items of both
categories approach a n absolute lower lim it on the tim e needed for lexical
access, even w hen perceptual and m otor response tim es are controlled for.
Thus, the highest frequency words, most of which are closed class, can be
described as being at the "asym ptote of accessibility" (Bock, 1989). Gordon
and Caram azza interpret these findings as signaling a "frequency saturated"
region in w hich increases in w ord frequency cease to have any effect on
lexical access time. These researchers conclude that both w ord types are
accessed via the sam e speech production processes, in contrast to the
assertions of Levelt (1989) and Bradley and G arrett (1983).
A nother effect of frequency relates directly to G arrett's claim (1975,
1982) that only open class w ords participate in sound exchange errors,
because the phonological forms of closed class item s are not postulated to be
available to the speech production system at the tim e that these errors occur.
Dell (1990) explores this assertion by com paring speech errors on open and
closed class hom onym s, such as HYMN and HIM, WEAR and WHERE, and BUY
and BY. He show s th at the higher the frequency of the phonological form of
a w ord, i.e. [him], [wEr] and [bai], the fewer sound exchange errors there are
on either form. A pparently, the phonological patterns of frequently used
w ords are m ore resistant to sound errors than less frequent phonological
patterns. Thus, another difference between open an d closed class speech
errors that had been claim ed as evidence for separate access pathways for
open and closed class w ords can be accounted for by frequency differences,
rather than to requiring that there be differences in lexical access or
representation.
40
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
2.3.6 Sum m ary
In this section, several im portant findings about the processing of
dosed d a ss w ords have been introduced. Evidence has been presented that
dosed d ass w ords are accessed using the sam e process as open class w ords
(Gordon & C aram azza, 1982, 1983.) In these studies, m ore frequent closed
dass w ords are accessed m ore quickly than less frequent ones, though
evidence for a low er lim it on access tim es is found for the most frequent
words of both types. Further, English closed class w ord forms do not prim e
themselves to occur in subsequent sentences (Bock, 1989). Furthermore,
higher frequency w ords, such as dosed class w ords, are also more difficult to
remember (H irshm an & Palij, 1992). These findings suggest that high
frequency, low-sem antic-content w ords (e.g., closed d ass words) m ay be
very difficult to hold in m em ory due to a very fast decay rate which is
related, perhaps, to an exceedingly fast access rate. This fast decay rate may
lead to increased num bers of omissions of these w ords in connected speech,
when extraordinary am ounts of attention are dem anded by other aspects of
the speech production process, such as articulatory movements in the case
of agram m atism , or w hen there are other extraordinary demands on WM
resources. From a different point of view, com putational simulations have
shown that w ords w ith a small num ber of sem antic features are m ore likely
to be om itted th an w ords w ith large num bers of features when there is
damage betw een sem antics and input system s (Farah & McClelland, 1991;
Plaut & Shallice, 1993). This last finding is supported by data from aphasic
speakers of languages in which closed class w ords carry more m eaning (i.e.
have m ore sem antic features), in that, in these languages, closed class
41
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
elem ents are more resistant to om ission errors, but do show increased
vulnerability to substitution errors instead.
2.4 C onclusions and Predictions
In this chapter, I have presented the two dom inant m odels of speech
production, both of which are grounded in assum ptions of encapsulated
linguistic m odules and noncom positional semantics. Subsequently, two
theories of semantic representation are discussed; the first being a non
compositional model sim ilar to th at accepted by Levelt and G arrett and the
second representing an overview of a compositional, feature-based semantic
system. Various research findings are also discussed which su p p o rt this
latter theory. However, several of the im portant issues distinguishing
betw een these theories are still unresolved, in particular, w hether closed
class w ords have semantic representations and w hether syntactic processing
is independent of semantic processing.
In any research, the theory adopted by researchers will necessarily
determ ine the types of data that w ill be considered relevant to their research
(Quine & Ullian, 1970). For exam ple, the m odular speech theories discussed
at the beginning of this chapter specify that closed class w ords are accessed by
syntactic rather than sem antic processes. Therefore, research based on these
theories has not considered the possibility that a sem antic deficit m ight
im pair the use of closed class item s. However, the studies discussed above
suggest that the theoretical assum ptions of Garrett and Levelt regarding
closed class items may be inaccurate. For example, closed class w ords are
apparently not inherent to the sentence frame, as specified by G arrett, they
do not show different frequency effects than open class w ords of sim ilar
42
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
frequency as suggested by Bradley and Garrett (1983), or differ in
phonological error patterns com pared to their open class hom onym s (Dell,
1990). Furthermore, evidence from other languages w ith different
m orphological structures show s that the tendency of agram m atic aphasics to
om it rather than substitute closed class items m ay be language specific.
These findings are m ore easily explained using an alternate theory of lexical
representation which specifies that both open and closed class word are
represented by sets of sem antic features, such as that described above.
W orking in this paradigm , both types of w ords m ight be expected to show
effects of semantic im pairm ent. Thus, the question of w hether closed class
w ords are affected by a sem antic im pairm ent is crucial to theories of speech
production. This is one of the goals of this research, to determ ine whether
the use of closed class w ords m ight be affected in populations without a
syntactic impairment, like healthy elderly people and A lzheim er's Disease
patients.
In sum m ary, based on m odular theories of speech production as
represented by the w orks of G arrett and Levelt, AD subjects' and healthy
elderly speakers' speech will be significantly different only in the num ber of
open class word substitutions, and only these errors will be dependent on
subjects' semantic ability. In addition, certain types of errors are not
expected to occur in significant num bers in the speech of either group; these
unanticipated errors include w ord omissions and closed class word errors of
any type. Moreover, in this type of theory, sem antic im pairm ents should
have Litle effect on syntactic processing or on the use of grammatical
formatives, like closed class w ords. In contrast, a theory of speech
production in which all w ords are semantically represented predicts that the
43
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
use of both open and closed dass w ords will be ham pered when there are
im pairm ents in sem antic ability. The current study explicitly tests these
predictions.
44
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
C hapter 3.
Literature Review:
W orking M em ory, Speech Production and A lzheim er's Disease
At one tim e or another everyone has found them selves com pletely
tongue-tied w hen som eone has interrupted them in the m iddle of a task
involving intense concentration. Some w ords seem com pletely inaccessible
an d other w ords crop up unexpectedly w here they d o n 't belong; completing
a sentence m ay seem alm ost impossible. The usual response of an
intelligent person at tim es like this (besides em barrassm ent), is som ething
along the lines of, "Sorry, I have a lot on m y m ind right now ." This is
actually a very insightful com m ent that captures the speaker's intuition that
processing m any ideas or tasks at once m ay adversely affect the accuracy of
the speech production process. This type of speech production difficulty is
slightly m ore com m on am ong elderly speakers than younger ones (Light,
1993, 1994), and is even m ore profound in A lzheim er's Disease (AD)
patients, whose spontaneous speech is m arked by w ord finding difficulties
an d inaccurate w ord choices (Patel & Satz, 1994; K em pler & Zelinski, 1994).
In fact, the speech of AD patients is often described as "vague, repetitive, and
fairly em pty of content w ords, instead being full of pronouns w ithout
antecedents, em pty phrases, indefinite terms, sem antically related but
incorrect w ords, and circum locutions" (Nebes, 1989, p. 378). Their
difficulties w ith speech production are com m only attributed to their
sem antic im pairm ent (Bayles, 1982; Hier, Hagenlocker, & Shindler, 1985;
Nicholas, Obler, Albert, & Helm-Estabrooks, 1985; Kem pler, Curtiss &
Jackson, 1987; Nebes, 1989; Patel & Satz, 1994). H ow ever, as suggested above,
this is not the only possible explanation. AD patients also suffer from
45
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
significant im pairm ents of w orking m em ory (WM) even at an early stage of
the disease (Parasuranam & Haxby, 1993; M orris, 1994; Patel & Satz, 1994),
and WM has been found to be necessary for m any aspects of language use.
Therefore, it is possible that WM deficits m ay also play a role in their speech
production difficulties.
This chapter discusses various theories of w orking m em ory and its
strong association w ith m any aspects of language ability. In addition, it
includes reviews of the few studies which have exam ined the relationship
betw een working m em ory and speech production, as well as the aspects of
speech production which these researchers predict should be affected by
individual differences in WM. These predictions are then com pared to
various findings about the speech production ability of AD patients. The
last section of this chapter presents several issues w hich have been
unresolved in both the WM and the AD literature and which are addressed
in the current research.
3.1 Theories of W orking M em ory
While there are several different theories about w hat working
m em ory (WM) is and how it works, these theories all share basic
assum ptions. First, all su p p o rt the idea that there are separate memory
system s that store sensory input in verbatim form and that can be
independently im paired in addition to a m ore general w orking m emory
capacity. Second, all ascribe, in varying degrees of explicitness, to the
m etaphor that items in m em ory are "activated" above som e resting value.
Finally, all agree that there are individual differences in WM capacity which
have far-reaching effects on cognition and linguistic abilities.
46
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Baddeley and Hitch (1974) propose a m odel of the w orking m em ory
system w hich includes two short-term , perceptually-specific "slave system s"
for verbatim storage of auditory and visual information (short term
auditory and visual memory) and a central executive system for
coordinating, scheduling and processing inform ation from the slave
systems and long-term m em ory (Baddeley, 1990). An alternate view point is
that WM is n o t a separate cognitive system but consists of the item s in long
term m em ory that are currently active, a hypothesis proposed by Engle and
colleagues (Engle, C antor & Carullo, 1992; C antor & Engle, 1993). A similar
hypothesis conceives of WM as the fuel which is used to activate ongoing
cognitive processes; this theory is particularly concerned w ith the effects of
individual differences in WM on language com prehension (Just &
Carpenter, 1992). Further, the latter theorists also speculate about the effects
of lim ited WM on language use. A fter som e inform ation on WM in
general, these three approaches to WM are presented below, follow ed by a
discussion of the effects of WM on speech production.
3.1.1 Baddeley and the "Central Executive"
The central executive, as described by Baddeley (1990), has several
regulatory cognitive functions, such as retrieving inform ation from long
term m em ory, controlling the transm ission of inform ation betw een
different parts of the cognitive system , and coordinating the activities of the
phonological loop (short-term auditory m em ory) and sketchpad (short-term
visual m em ory) "slave system s" (Gathercole & Baddeley, 1993). In other
w ords, all cognitive functions that require activation of previously
assim ilated inform ation or complex thought processes, such as language
47
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
com prehension and d raw in g inferences, are considered to require central
executive involvem ent.
Baddeley's conceptualization of the central executive has been
inextricably linked w ith attentional processes from early in its theoretical
developm ent, such th a t an action or process requires WM to the extent that
it requires conscious attention (Baddeley, 1990). T hus, as an action becomes
well learned, it requires less attention and less WM. In som e of his
w ritings, it appears th a t the central executive is considered to be only a
general attentional m echanism to coordinate the inform ation gathered
from different subsystem s th at has no processing o r storage capacity of its
ow n (Baddeley, 1990; R ichardson, 1996). Indeed, the m ajority of tests of
central executive function are identical to tests of divided attention in
which subjects m ust carry o u t tw o tasks at once (Parasuranam & Haxby,
1993; Kopelman, 1994; M oskovitch, 1994). In these tests typically one task is
a sim ple span task (i.e. trying to rem ember a sh ort list of w ords or num bers),
w hile the second task m ay involve anything from finger tapping to doing
arithm etic problem s o r reading com prehension. The typical finding is that
perform ance on the second task becomes im paired as the num ber of digits
or w ords in the span task increases. These tasks are considered to require
WM particularly because they involve sim ultaneous storage and processing
of inform ation.
A prim ary am biguity in Baddeley's theory has to do with the nature
of the central executive and W M in general. The central executive is often
described as if it w ere a specific place in the brain to w hich information is
copied that is lim ited b y capacity— like a com puter that has limited RAM.
48
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
However, there has been little evidence for the existence of this type of
processing area. Instead, clinical studies of attention tend to implicate many
areas of the brain and the connections betw een them in attentional
processing (for a review, see LaBerge, 1995).
3.1.2 Engle and the General Capacity M odel
An alternate theory, the G eneral Capacity Model, has been suggested
by Engle, C antor and Carullo (1992). This model asserts that the contents of
WM consists of inform ation in long term m em ory, such as a sem antic
representation, that has been activated above som e critical threshold
(Cantor & Engle, 1993). This m odel also assum es that there is a single WM
capacity that is used to fuel all concurrent cognitive processes, w hether these
processes are verbal, quantitative or spatial. Activation is postulated to be a
limited resource that autom atically spreads am ong related knowledge
structures and that can accumulate slow ly (due to connections from related
items) or quickly (if activated directly by items in attentional focus).
An im portant aspect of this m odel is that an item may be raised
above threshold enough to be "in W M " and, thus, accessible to WM
processes, w ithout necessarily being active enough to be at the level of
consciousness. Thus, it seems that these researchers are suggesting dual
thresholds, the low er one allows a concept to be recognized and used by
cognitive processes, and the higher one allows the subject conscious
awareness of the inform ation (Engle, C antor, & Carullo, 1992). If this were
true, it w ould offer a m echanism to explain how much of the data m anipu
lation involved in com plex cognition and language processing could be
accomplished w ithout conscious aw areness. For example, facilitation in a
49
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
prim ing task only requires that the level of activation of sem antic represen
tations be raised to the first, lower level of activation, while nam ing tasks
require that the sem antic representation be raised to the higher level of
activation so that the w ord form can reach conscious awareness. Failure to
raise the level of activation of the complete sem antic representation to the
conscious level can lead to a circumlocution or to a Tip-of-the-Tongue state
(e.g. Burke, M ackay, W orthley, & Wade, 1991). In this theory, the contents
of short term m em ory consists of those knowledge units which are at the
higher level of activation, and thus within the grasp of im m ediate
consciousness (Engle, 1996).
Another interesting aspect of this work is that these researchers
believe that individual differences in WM are a function of the am ount of
activation resources a person has available to distribute am ong different
knowledge structures. Thus, subjects with high WM spans can hold large
am ounts of inform ation at a level below conscious aw areness while
attending to another task, and then reactivate these items for conscious
retrieval. Engle et al. consider dual-task WM span tests, such as reading
span (Daneman & C arpenter, 1980), to be good m easures of subjects'
activation capacity limits, because they require the sw itching of attention
between tasks, and sw itching and m aintaining attention are just as crucial to
this theory of WM as they are to Baddeley's. In fact, recently Engle has
increased his em phasis on attention to the point of suggesting that WM
tasks m easure "w orking attention" more than w orking m em ory (Engle,
1996). He has concluded that the most im portant difference between high
and low WM span subjects is that, in complex tasks, those w ith higher spans
seem to use controlled, intentional processing (which requires the conscious
50
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
direction of attention). In contrast, low span subjects appear to rely m ore on
the passive, autom atic spread of activation in the same tasks.
3.1.3 lust and C arpenter and the effects of lim ited W orking M em ory
Similar to the view s of Engle and colleagues, discussed above, Just
and C arpenter (1992) and their colleagues ascribe to a view of W M in which
the same quantity of activation resources, m easured by WM capacity, is used
for both storage an d com putation of language information. They, m ore
than any other group, have theorized about the effects of reduced WM
capacity on language use, in particular, on comprehension. A ccording to
their theory, WM capacity limits not only the am ount of inform ation a
person can activate about a sentence or situation, but also the tim e course of
its activation (M iyake, Carpenter & Just, 1995). Thus, when WM capacity is
more lim ited, not only are subjects less able to activate and use all they
know about a given w ord or sentence b u t the knowledge that they do have
is available m ore slowly, resulting in decrem ents in language com pre
hension perform ance. U nfortunately, the tasks that these researchers rely
on to m easure WM require the activation and m anipulation of w ords, and,
consequently, it is not clear w hether these tasks measure a separate
cognitive capacity, as these authors claim, or a characteristic function of the
sem antic system , such as tem porary storage of semantic inform ation
(Altmann, K em pler & Andersen, 1997; M artin & Saffran, 1996; M acD onald
& Christiansen, subm itted).
51
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
3.2 W orking M em ory and Speech Production
Several groups of researchers have exam ined how individual
differences in WM capacity can affect aspects of language use, particularly
com prehension. The usual types of tests used to m easure WM capacity are
complex span tests, such as reading and listening span.1 D anem an and
C arpenter's (1980) reading span test requires the subject to read a series of
long sentences (usually 13 to 17 words) and recall the last w ord of each;
listening span is sim ilar except the stim uli are presented auditorially.
Consequently, success on these tasks requires that subjects be able to activate
sem antic representations and to keep them above activation threshold
despite concurrent linguistic distraction. As m entioned above, it is unclear
w hether these tasks m easure an inherent function of sem antic m em ory or
an independent cognitive capacity (see also Gathercole & Baddeley, 1993).
Hence, these tasks are referred to herein as sem antic WM tasks, in contrast
to those WM tasks which do not require the use of sem antic information,
such as digit span backward.
Scores on m easures of sem antic WM capacity show considerable
variation in the norm al population and have been found to correlate
especially highly w ith m any language abilities. For exam ple, children are
lim ited by w orking m em ory span w hen they are learning to speak and
understand spoken language (Gathercole & Baddeley, 1993), as well as w hen
they learn to read (Oakhill, Yuill, & Parkin, 1986; Yuill, Oakhill, & Parkin,
1989) and spell (Ormrod & Cochran, 1988). Further, the differences between
‘Some researchers, particularly Meredith Daneman, believe that each of these tests measures
the efficiency with which WM is used in accomplishing a particular task, like reading,
listening or speaking, and so reading and speaking span might differ (Daneman, 1991).
52
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
children w ith high and low w orking m em ory capacities are reflected in
differences in vocabulary know ledge and other language abilities from an
early age (Gathercole & Baddeley, 1993). These differences persist into
adulthood, w ith individual differences in WM span being related to various
language abilities, such as: vocabulary size (Daneman & C arpenter, 1980;
Engle, C antor & Carullo, 1992), learning new w ord m eanings from context
(Daneman & Green, 1986), com prehending complex syntax (King & Just,
1991), integrating new inform ation w ith previously presented m aterial
(Daneman & Carpenter, 1983), w riting ability (Benton, Kraft, G lover &
Plake, 1984), and the speed of lexical access (Daneman & Green, 1986). In
addition, subjects w ith higher spans have been show n to read complex
sentences faster and w ith better understanding than lower span subjects
(Just & Carpenter, 1992). Furtherm ore, higher span subjects are able to
m aintain the activation of several alternative interpretations for am biguous
sentences, which leads to longer reading times (MacDonald, C arpenter &
Just, 1992), and are also better at using plausibility and contextual constraints
when trying to understand am biguous sentences (Pearlm utter &
MacDonald, 1995; Stevens, H arm , Schuster, & MacDonald, 1995). In
contrast, lower span subjects have m ore difficulty disam biguating sentences
which require the access of low frequency word m eanings (Miyake, Just &
Carpenter, 1995), and their reaction times to am biguous sentences are not
affected by the possibility of m ultiple meanings, since it appears that only
the most frequent m eaning is available to them (MacDonald, C arpenter &
Just, 1992). Thus, semantic WM and language abilities are intricately
intertw ined from the beginning of language learning. This m akes the
53
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
relationship betw een sem antic WM an d language abilities even m ore
suspect w ith respect to the independence of these m easures.
3.2.1 D anem an and Green (1986)
M ost of the findings to date regarding the effects of sem antic WM on
language have focused on com prehension. However, a few researchers
have considered the possibility that lim ited WM may also affect speech
production. Danem an and Green (1986) have designed a type of complex
span task, the speaking span, that is sim ilar to the reading span task
described above, which, theoretically, targets an aspect of sem antic WM
which is involved with speech production. The speaking sp an task gives
subjects increasingly long lists of w ords, and they m ust construct a sentence
for each w ord in the list. Scores on the speaking span task are very highly
related to subjects' reading spans. D anem an and Green find that subjects
w ith higher speaking spans are able to complete sentences and produce
synonym s for w ords in sentential contexts faster than subjects w ith lower
WM capacity. Performance on the speaking span task is also show n to
correlate very highly with the ability to abstract the m eaning of new words
from context, as well as w ith the speed of lexical access (D anem an & Green,
1986). So this study asserts that higher (semantic) WM capacity is associated
w ith faster sem antic and lexical access and better vocabulary acquisition.
3.2.2 D anem an (19911
In a m ore recent experiment, D anem an (1991) exam ines subjects' false
start and repetition rate w hen reading aloud, and correlates this w ith
speaking and reading span. D anem an compares subjects' perform ance on
54
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
two tasks, picture description and oral reading, to their perform ance on the
speaking span W M task. She finds that several m easures correlate w ith
speaking span, including: the num ber of w ords produced in one m inute
and the richness of vocabulary used in the picture description, as well as
oral reading speed and num ber of false starts in reading. The second
experim ent uses a speech error prim ing paradigm developed by Motley and
Baars (1976) to show that low span subjects m ake m ore sound exchange
errors than high sp an subjects. This prim ing test attem pts to induce
phonological exchange errors by presenting pairs of w ords that all start w ith
the same tw o phonem es (i.e., [t] and [p]) and then sw itching the order of the
phonem es in the fourth pair (e.g. primes: tab-peck, tin-pane, tune-pole;
target: pun-toe). In the first task, Danem an finds that higher w orking
mem ory capacity is associated with faster lexical access (num ber of words
produced in one m inute) and also w ith "vocabulary richness" in the picture
description task. In the speech error prim ing task, she concludes that
subjects w ith low er spans had more phonological intrusions.
W hile the results of this research seem to support those of Danem an
and Green (1986), this study has several flaws w hich m ake its results
difficult to generalize. First, the cognitive requirem ents of reading aloud are
very different from those of spontaneous speech, since sem antic and
conceptual access are not necessarily required in reading aloud (Caplan,
1992; Raymer & Bem dt, 1996). In addition, the speed of lexical access and
richness of vocabulary are likely a function of vocabulary size and overall
language experience (MacDonald & C hristiansen, subm itted), and the study
included no m easure of these. Furtherm ore, it is not clear w hether the
Motley and Baar experim ental paradigm tests WM effects or merely tests
55
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
prim ing a n d activation perseverance in the phonological loop. Thus, it
could easily be argued that high span subjects should m ake m o re errors on
this task, since they w ould be expected to be able to accurately store previous
stim uli for longer periods of time. In addition to this am biguity, all of the
reported statistics are correlational, and although the researcher reports
m ean differences betw een low, m edium and high capacity subjects, the
differences in scores between groups do not appear to be significant.2
N evertheless, this study does suggest different ways in w hich individual
differences in sem antic WM m ight affect speech production, notably in the
num ber of false starts and richness of vocabulary used.
3.2.3 Pow er (1985)
As presented above, WM span, especially when m easured by tasks
like the reading span, has been found to be highly related to m any aspects of
language know ledge and use: vocabulary size, reading com prehension,
inferring m eaning from context, etc. All of these language skills appear to
be sem antically based; that is, they deal w ith the num ber of w ords people
know, how w ell they know them, an d how well they can use this
know ledge to learn about new w ords. This suggests that sem antic processes
and WM m ight be particularly linked. This conclusion is supported by an
early speech production study by Pow er (1985) that looks at the effects of
divided attention on sentence production latencies, the originality of the
sentences produced, and subjects' recall of their own sentences. However,
in this case, the researcher uses a sim ple digit retention task, rather than
reading or listening span to m easure WM. Subjects w ere given zero, three
2 Low span-17% spoonerisms, medium span-12%, high span-10%. Total N=29, distribution of
subjects among span conditions was not reported.
56
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
or six digits to rem em ber, then w ere asked to produce a sentence including
two nom inal stim uli w hich varied in relatedness. Sentences produced in
the digit load conditions w ere m ore stereotypic and had "less information"
in them than sentences created in the "no digit load" condition, in effect
foreshadow ing the finding in Danem an (1991) that reduced resources can
dim inish the richness of sem antic content. In addition, Pow er's subjects
recalled fewer of the sentences they had produced u n d er digit load
conditions than in the no-digit-load condition w hen asked to recall their
sentences later.
Pow er concludes that WM is necessary for the creative and
m eaningful production of speech, and that w hen W M capacity is reduced by
divided attention (or, by analogy, age or pathology) speakers m ight employ
various sim plification strategies to m aintain fluency. Some of the strategies
which she predicts m ight be adopted are: reducing the num ber of words
used, choosing m ore stereotyped expressions such as cliches and idioms,
and using fewer com plex structures. Also, she speculates that there may be
increases in false starts, as well as semantic errors, since her prim ary finding
is that sentence production does take "semantic w ork."
3.2.4 Valencia-Laver (1992)
As just m entioned, w orking memory lim itations have also been
suggested as being a factor in speech dysfluencies, especially increased
production of sentence fragm ents, false starts involving retracing part of an
utterance, and error correction abilities (Power, 1985; Danem an, 1991). Some
of these hypotheses have recently been tested experim entally. Valencia-
Laver (1992) com pares the speech output of healthy young and elderly
57
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm ission .
subjects on a task in which subjects are required to describe an abstract
design consisting of colored dots connected by lines. Perform ance is
evaluated in term s of the relative num ber of lexical errors, repaired errors,
and retracing versus fresh start dysfluencies, as well as the "depth" of the
repairs and retraced dysfluencies. The "depth" of a repair refers to how
m any w ords are uttered betw een an error and the interruption m ade to
correct it, while the depth of a retracing refers to the num ber of w ords that
are repeated, i.e. how far the speaker "backs up," after speech has been
interrupted.
Am ong other things, Valencia-Laver finds that elderly speakers m ake
m ore errors of all types and correct a lower proportion of lexical errors
com pared to young speakers. Furtherm ore, younger speakers are m ore
likely to repeat or retrace their w ords w hen they encounter difficulties in an
utterance, rather than make a fresh start, but the elderly speakers m ake
equivalent proportions of retracings and fresh starts. Besides com paring
error and dysfluency rates betw een groups, Valencia-Laver also exam ines
the depth of repairs and retracings, and the detection of non-phrase-
bounded errors for effects of age, WM (using digit span backward), and
vocabulary size. How ever, no significant effects of any of these variables are
found. In sum m ary, Valencia-Laver finds that older adults' speech is
characterized by increases in speech dysfluencies and lexical errors as well as
fewer error corrections com pared to younger speakers, but cannot relate any
specific error pattern or speech dysfluency to WM or vocabulary size.
58
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
3.3 Speech Production in A lzheim er's Disease
The studies just presented have identified several aspects of speech
production that m ight be affected by individual differences in WM. These
include: accuracy of semantic access (Pow er, 1985; D anem an & Green, 1986),
speed of lexical access (Daneman, 1991; Just & C arpenter, 1992), num ber of
dysfluencies— especially the relative frequencies of retracings versus fresh
starts (Pow er, 1985; Valencia-Laver, 1992), the proportion of errors corrected
(Valencia-Laver, 1992) and a decline in the specificity of vocabulary used
(Pow er, 1985; Danem an & Green, 1986; D anem an, 1991). As a point of fact,
this list of predicted and docum ented effects of WM on speech production is
rem arkably sim ilar to the effects of A lzheim er's Disease on speech
production. The typical speech production of AD patients contains m any
sem antic substitutions for content w ords in spontaneous speech, and m ay
be m arked by a lack of m eaningful, sem antic content (Bayles, 1982; Nebes,
1989; Kem pler, 1991). In addition, they produce m ore sentence fragm ents, a
type of speech dysfluency, com pared to healthy elderly (Lyons et al., 1994),
and show slow er lexical access in alm ost all linguistic tasks (Nebes, M artin
& H orn, 1984; Nebes, 1989). In addition, there is little controversy over the
fact th at AD patients suffer from both sem antic and m em ory deficits even
early in the disease (Kempler, 1991; Patel & Satz, 1994; Green, 1995).
Furtherm ore, these two abilities decline at different rates early in the
disease, so that subjects can be found w ith either prim arily sem antic or
prim arily m em ory im pairm ents though they exhibit the sam e level of
overall disease severity (Martin, 1990). C onsequently, in the current study
A lzheim er's patients have been chosen as the appropriate experim ental
p opulation in w hich to observe the effects of lim ited WM and im paired
59
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
semantic ability on speech production. Furtherm ore, the effects studied will
be two of those suggested by the previously discussed research: errors in
lexical selection and false starts.
Several studies have exam ined the speech production abilities of AD
patients. For exam ple, Bayles (1982) reports evidence that AD patients do
not correct their lexical or gram m atical errors in spontaneous speech.
Further, Kem pler, Curtiss and Jackson (1987) analyze the spontaneous
speech of AD patients and find that lexical errors are common in their spon
taneous speech, b u t m orphosyntactic errors are very infrequent, although
the range of syntactic structures used is norm al. These findings have led to
the assertion that m orphosyntactic processes are preserved in AD.
M ore recently, Bates, M archman, H arris, and W ulfeck (1994) have
reported that AD patients m ake significantly m ore lexical errors in scene
descriptions than elderly norm al controls or college students. In addition,
they find that AD patients are quite im paired in choosing the correct
syntactic form for a sentence w hen their productions are constrained to
certain points of view w hich require less fam iliar w ord order patterns.
Their study concludes that, just as AD patients have difficulty choosing the
correct w ord to nam e a target picture, they are also im paired at choosing the
correct sentence structure to express a given concept or relationship,
particularly w hen the required structure is non-canonical.
In a study of the spontaneous speech of AD patients, Lyons, Kemper,
LaBarge, Ferraro, Balota, and Storandt (1994) find that, as disease severity
increases, there is a significant increase in the proportion of utterances
60
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
consisting of sentence fragments, som e of which result in a type of speech
dysfluency know n as a "fresh start" (Levelt, 1983; Valencia-Laver, 1992).
In 1987, Ripich and Terrell report that high num bers of pronoun
errors in AD speech are a significant contributing factor to their
incoherence. Furtherm ore, Kempler and colleagues (Kempler, 1984;
Kempler et al., 1987) report that AD subjects make m ore pronoun errors
than elderly control subjects.
It is necessary to point out, how ever, that none of these studies have
found any links betw een these phenom ena and limited WM or sem antic
deficits. K em pler et al., Bates et al. and Ripich and Terrell do not examine
their data for specific correlates of either semantic or WM im pairm ent. On
the other hand, Lyons et al. look for effects of WM, but not semantics, and
they find no correlation between the proportion of sentence fragm ents and
their only WM m easures, digit span forw ards and backwards. They propose
that this lack of correlation m ight be due to the limited num ber and type of
WM m easures used and suggest that m ore complex WM m easures m ight
show different results (though they do not suggest which WM tasks might
have been useful). O ur pilot studies explored the spontaneous speech of AD
patients and found that it contained both substitutions and omissions of
pronouns, m orphosyntactic inflections, and open and closed class words.
Moreover, in a longitudinal study, the frequency of errors and proportion of
omissions w ere related to WM deficits rather than to severity of cognitive
im pairm ent (A ltm ann, Andersen & Kempler, 1993, 1995).
Thus, the few relevant studies on speech production in AD have
found that, in com parison to elderly controls, AD patients: m ake more
61
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
lexical errors, produce m ore sentence fragm ents, have m ore difficulty with
pronoun use, b u t show few m orphosyntactic errors, and have difficulty
producing sentences using a noncanonical point of view . However, to date
there is little evidence, besides our ow n pilot studies, that there m ight be a
direct connection betw een either semantic or WM im pairm ent and any of
these types of errors in the speech of AD patients.
3.4 Conclusion and Predictions
In sum , the volum e of research supporting a relationship between
semantic WM capacity and various language abilities suggests that WM may
affect perform ance on all types of language tasks. H ow ever, much of this
evidence is suspect since the WM tasks that show such strong relationships
with language abilities them selves have a strong sem antic component. To
avoid this difficulty, all WM and semantic m easures used in the following
study will be tested for collinearity, to m ake sure that they m easure distinct
abilities.
A prim ary goal of the current research is to investigate the frequency
of speech dysfluencies to determ ine, if possible, their cause in healthy
elderly and AD patients. Levelt (1989) postulates th at WM resources are
crucial for producing a certain types of dysfluency w hich requires that the
ongoing utterance be interrupted, in order to backtrack and "retrace" the last
few w ords said. Retracing is often used as a form of error correction, and it
should be particularly dependent on m em ory, because it requires that part of
the already articulated utterance be repeated verbatim , w hich clearly
requires som e type of short term recall, though it is not clear whether this is
a WM function or a function of a verbatim articulatory store. Nonetheless,
62
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
as discussed above, Valencia-Laver (1992) found no influence of WM on the
depth of retracings in young or elderly subjects as had been theoretically
predicted by Levelt (1983,1989), nor did Lyons et al. (1994) find any effect of
WM on the n u m b er of sentence fragm ents in AD subjects' speech.
However, both of these studies only used backw ards digit span as their WM
m easure, and this m ay have limited their findings. To avoid this difficulty,
this dissertation em ploys several WM m easures to gauge the contribution
of WM to different types of false starts and errors. On the other hand, it is
necessary to be aw are that in a population such as AD patients, sentence
fragm ents and fresh starts may also arise due to semantic im pairm ent if the
appropriate w o rd s to continue an utterance are unavailable to the speaker.
Thus, this dissertation also attem pts to discern w hether sem antic ability
contributes to the frequency of speech dysfluencies of AD patients and
healthy elderly subjects.
A nother goal of this dissertation is to exam ine the speech errors of
both healthy elderly and AD subjects to determ ine whether WM contributes
to the rates at w hich lexical errors occur. This is based on the suggestion in
Power (1985) th at lim ited WM may increase lexical errors and dysfluency
rates, as well as restrict the semantic content of utterances. As pointed out
above, all of these effects are present in the speech of Alzheimer's Disease
patients. H ow ever, AD patients suffer from both semantic and WM
im pairm ents, an d either deficit m ight be responsible for these phenom ena.
Furtherm ore, the effects of semantic im pairm ent on AD patients' speech
have not been experim entally confirmed; that is, while several character
istics of their speech have been docum ented, these have not been tested for
a relationship to sem antic impairm ent. Thus, the issue of the etiology of
63
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
speech production difficulties in AD is still unresolved. The current study
explicitly compares the frequency of lexical errors and the dysfluency rates of
AD patients and healthy elderly speakers to several m easures of WM and
sem antic im pairm ent in o rder to address this question.
64
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Chapter 4.
M ethods
This chapter describes the twenty-six subjects w ho participated in
the study, as well as the sem antic and working m em ory tests that they
w ere given and how these tests were scored. In addition, a short overview
of the speech production tasks used is included, although more in depth
discussions of these tasks are found in Chapters 5 an d 6.
4.1 Subjects
Ten subjects diagnosed w ith Probable A lzheim er's Disease (AD)
using NINCDS-ADRDA criteria (McKhann, Drachm an, Folstein,
Katzman, Price, & Stadlan, 1984) participated in the study. These AD
patients had been enrolled in a longitudinal study of the effects of AD on
language abilities a t the University of Southern California for between one
and three years. In addition, sixteen healthy control subjects were
recruited by advertising in alum ni newsletters, local new spapers and
senior centers in the Los Angeles area. All subjects had normal, or
corrected to norm al, vision and hearing and no history of stroke or other
neurological im pairm ent. G roup means for the M ini-M ental Status Exam
(MMSE) (Folstein, Folstein & McHugh, 1975), age, education, and
proportion of female are show n in Table l.1 W hile the two groups do not
overlap in MMSE, there is no difference between the tw o groups in any of
the other m easures (age: (t(24) = 1.429, p > .15); education: (t(24) = -.801, p >
.40).
1 Demographic information on all subjects is shown in Appendix A.
65
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
Table 1. C om parison of AD and C ontrol G roups
AD subjects
i
Control subjects j
MMSE 20.45 29.00
(s.d.)
(1-4)
(1.32) ;
A ge 80.40
|
76.44
(s.d.) (6.98) (6.82)
E ducation 15.0 15.7
(s.d.) (2-4) (1-9)
Proportion Fem ale .60
*
4.2 Procedures
All AD subjects were tested at their hom es or day care facilities by
trained, psycholinguistic graduate students. There were five sessions to
complete the battery, each lasting less than 90 minutes. Each session
contained at least two of the three types of tasks (working m em ory,
semantic and speech production) along w ith other tests unrelated to the
current study. NC subjects com pleted the battery in one test session either
at USC or at the subjects' home. A single NC session was possible because
the only sem antic m easure in the NC battery w as the Peabody Picture
Verification Test. Semantic im pairm ent tests were not adm inistered to
NC subjects, since they perform at ceiling on these tasks (e.g., K em pler &
Zelinski, 1994; Balota & Duchek, 1988; Salthouse, 1988).
4.2.1 W orking M em ory and Cognitive Assessm ent
All subjects completed several w orking memory (WM) tests as a
part of the experim ent. These tasks have been combined into three
66
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
m easures that are hypothesized to test different aspects of WM function:
Nonsemantic-W M , Backward-Recall, and O rdering ability. These tasks are
described below, and subject's scores are show n in Table 2.
To test N onsem antic-W M ability, the scores on two common tests,
digit span backw ards and counting by 3s from 1 to 40 (Wechsler, 1987),
w ere averaged. In this task, subjects are presented w ith increasingly long
lists of digits w hich they m ust repeat back in the reverse order of
presentation. Digit span backw ard is often used as a m easure of WM in
studies of pathological populations (e.g., Cherry, Buckwalter, &
H enderson, 1996; Lyons et al., 1994), since it requires both storage and
processing functions but does not rely on the m eanings of the elements
(num bers) involved. In the C ounting from One to Forty by Three's task,
subjects were asked to do just that, to count by three's from one to forty. If
necessary, the response w as m odeled, using the follow ing formula
"continue the sequence {1, 4 . . . }." Scores on these tw o tests were highly
related, and so w ere com bined into one m easure w hich has been called
"N onsem antic-W M ."
The Backward-Recall test had two parts, both of which required the
generation of a w ell-know n sequence in reverse order. In the first portion
of the test, subjects had to count from tw enty to one backw ards, while in
the second portion they recited the m onths of the year backwards. In
term s of task dem ands, this test seem ed to require m ore concentration and
attention than the other nonsem antic WM tasks, since subjects had to be
able to direct their attention to a familiar, self-generated sequence from a
(m ore or less) unfam iliar view point.
67
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
The O rdering task presented subjects w ith lists of num bers and lists
of digits, w hich then had to be put into canonical order. The first level of
the test used tw o item lists; if subjects m et criteria on this level, (i.e. three
out of four lists recalled and ordered properly), they w ere tested at the next
highest level also. The largest set of stim uli tested had six items to be
ordered. This task is hypothesized to test the sim ultaneous storage and
m anipulation of m eaningful information since subjects had to use w hat
they know about m onths and digits, i.e. their place in the sequence, in
order to succeed at this task.
The scores for all WM tasks were rendered as the proportion correct
of the total possible. Thus, if a subject got six trials correct on each of the
digit and m onth O rdering subtests, her O rdering score would be 12/40 or
30%. Similarly, w hen she successfully reversed a list of four num bers on
the digit span backw ard, her score would be 4 /7 or 57%, since seven-digit
lists were the longest tested.
The M ini-M ental Status Exam (MMSE) w as given as a m easure of
overall cognitive functioning to AD subjects in both the first and last
testing sessions an d to the NC subjects once. This task was used prim arily
as a diagnostic tool to establish the relative severity of cognitive
im pairm ent, but w as not used in speech production analyses.
4.2.2 Semantic Ability Assessment
The sem antic ability of the AD and N C groups was assessed using
different m easures. Because healthy elderly subjects do not suffer from a
semantic im pairm ent and usually score at o r near ceiling on nam ing tasks,
68
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Table 2: Scores on th e three WM m easures (proportion correct).
Subject2 N on-S em antic Backward-Recall O rdering
W M
106 .7857 .9583 .6250
120 .2857 .9500 .1750
126 .8571 .8333 .4500
131 .3571 .3667 .1000
132 1.0000 1.0000 .5750
136 .7500 1.0000 .5000
138 .7143 .9750 .1750
141 .1429 .9167 .2250
146 .2143 .5750 .1750
147 .7857 1.0000 .6250
301 .9286 1.0000 .8750
302 .9286 1.0000 .8250
310 .8571 1.0000 .8750
320 .7857 1.0000 .7750
321 .6429 1.0000 .8000
329 .8571 .9167 .6500
330 .7857 1.0000 .9000
331 .7143 1.0000 .5500
332 .9286 1.0000 .6000
334 .9286 1.0000 .7500
337 .9286 1.0000 .7000
340 .9643 1.0000 .8000
341 .9286 1.0000 .8500
342 .7143 1.0000 .4500
343 .8571 1.0000 .8500
345 .8571 1.0000 .7750
Subject numbers less than 199 refer to AD subjects; numbers over 300 are MC subjects.
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
their semantic ability w as assessed using the Peabody Picture Vocabulary
Test (PPVT), described below . However, since AD perform ance on such a
task would necessarily reflect both prem orbid vocabulary skills as well as
sem antic im pairm ent, this task has not been used as a sem antic m easure
for the experim ental group. Instead, subjects com pleted a 96 item naming
test, and sem antic ability (or im pairm ent) m easures w ere based on this.
The Peabody Picture Vocabulary Test (PPVT) w hich w as used to
assess the sem antic ability of the NC subjects3 is a w ord-picture matching
test in w hich subjects m ust choose the correct picture d enoting the target
w ord from a tw o by tw o array of pictures w ith com parable complexity.
This task tests com prehension of verbs, adjectives, and nouns of varying
frequency and abstractness. The difficulty of test item s increases
throughout the task, w ith the first twenty items being ap p ro p riate for
young children, and the last tw enty targeted at college-educated subjects.
As an exam ple of the overall difficulty, five of the tw enty m ost difficult
items were: CONVEX, CUPOLA, HOMUNCULUS, REPOSE, an d JUBILANT.
Testing was initiated w ith m id-level items (e.g. item 85 of 175), and
continued until subjects m ade six errors on eight consecutive items.4 To
determ ine scores on this task, the experim enter term inates the test w hen a
subject m akes errors on six of eight consecutive items, calculates the
num ber of errors betw een the ceiling item (the last item tested) and the
last run of eight consecutive correct items, then subtracts the num ber of
errors m ade betw een these tw o item s from the num ber of the ceiling item.
3ppvr results were available from only 13 of the 16 NC subjects.
4 If subjects had made six errors immediately, items would have been given in reverse order until
8 items were answered correctly. However, this was not necessary with any of these subjects.
70
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
For the purposes of this study, scores w ere converted to the percentage
correct of the possible 175 items.
The nam ing task used to assess AD subjects' semantic ability consisted
of 96 black and w hite line draw ings presented one at a time on the screen of
a M acintosh Classic II or sim ilar com puter. The test contained tw elve items
from each of eight sem antic categories: vehicles, tools, furniture, clothing,
anim als, fruits/vegetables, body parts and birds. Items in each category
varied in prototypicality, frequency and familiarity. Scores w ere rendered as
the proportion of all nam ing responses w hich fell into each of the categories
below. A com plete list of the nam ing errors m ade by AD patients and how
these errors w ere categorized is located in A ppendix B.
Responses in the nam ing task w ere scored as follows. R ather than
accepting the first attem pt to nam e a picture as the subjects' response (e.g.
Nicholas, Obler, Au, & Albert, 1996), the whole answer given by the subject
w as considered, and the best, usually the final, nam ing attem pt w as used as
the response to be graded. This scoring strategy was adopted because subjects
w ould often self-correct an initially w rong response, indicating that they
w ere aw are th at they had produced an incorrect nam e for the object
pictured. This aw areness appeared to signal existing semantic know ledge,
and thus, could not be ignored w hen the goal here was the assessm ent of
this know ledge. Furtherm ore, since NC subjects also are know n to
occasionally produce an erroneous response and then correct them selves,
this type of response does not necessarily indicate semantic im pairm ent.
There is a trend in recent nam ing studies to analyze different types of
nam ing errors according to the am ount of target-specific sem antic
71
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
know ledge show n by the nam ing response (Nicholas et al., 1996; Hodges,
Salm on & Butters, 1991). H ow ever, Nicholas et al., in particular, used
subjective ratings on a scale that had five different levels of relatedness, and
excluded those responses w hich show ed no semantic know ledge of the
target. A less arbitrary m ethod w as preferred here in w hich all naming
errors which consisted of the substitution of a member of the sam e semantic
category (i.e., contrast coordinates) constituted the first type of errors,
W ithin-Category errors. This type of error has been considered a hallmark
of m ild semantic im pairm ent (e.g., Hodges, Patterson, O xbury, & Funnell,
1992; Nicholas et al., 1996), and considering the relatively low level of
overall im pairm ent of the experim ental group (i.e., m ean MMSE 20.4), it
w as expected that most of the nam ing errors from AD subjects' would be of
this type. This prediction w as correct; W ithin-Category errors comprised
54% of all nam ing errors (88 of 163). Attempts to develop objective criteria
to further subdivide this category failed largely due to differences between
categories; for example, criteria based on function for tools or furniture were
not applicable to the fruits and vegetables or vehicles categories, since all
m em bers of these two categories share the same functions (i.e. TO BE EATEN
or TO MOVE PEOPLE/THINGS, respectively). The W ithin-Category errors score
is interpreted here as a m etric of the proportion of the sem antic system
affected by mild damage.
As the severity of sem antic im pairm ent worsens in AD and other
diseases affecting semantic know ledge, such as prim ary progressive aphasia
and sem antic dem entia, there seem s to be an increase in superordinate and
circum locution errors hypothesized to be the result of a breakdow n in the
aw areness of subordinate attributes of an item with retention of category
72
R ep ro d u ced with p erm issio n o f th e cop yright ow n er. Further reproduction prohibited w ithout p erm issio n .
knowledge (H odges et al., 1992). In this sam ple of AD subjects these types of
response frequently co-occurred (e.g. subject #120, for target STRAWBERRY, "a
fruit, you eat it.") and so have been com bined into a single category. This
error category, S uperordinate/ Circum locution errors, com prised 23% of all
nam ing errors.*
W hen sem antic im pairm ent becomes severe, subjects often produce
nam ing responses w hich reveal no sem antic know ledge of the target
(Hodges et al., 1992). Errors falling into this category included "I don't
know" responses, affective comments ("I like them "), visual erro rs,* and
erroneous responses from outside the sem antic category of the target. These
responses com prised the rem aining 23% of nam ing errors by these AD
patients. In contrast to Nicholas et al. who argue that these errors are
uninterpretable, the current study argues th at No-Response errors signal the
proportion of the sem antic system that has sustained serious damage.
As described above, W ithin-Category an d No-Response7 error rates
have been interpreted as indicating the proportion of the semantic system
affected by m ild and severe semantic im pairm ent, respectively. This
assum ption is necessary as they will be used as metrics of different degrees
of im pairm ent in the analyses in Chapters 5 an d 6. This analysis of errors
is based on a theory involving a com positional sem antic system in which
individual features m ay participate in the sem antic representations of
5 This category also included errors with a Part-Whole relationship to the target, e.g. "buckle"
for "belt" and "library books" for "bookcase."
6 There was no correlational relationship between visual errors and visual-perceptual ability,
so these errors were included with other errors that showed no semantic knowledge of the
target.
7 Superordinate/Circumlocution errors were excluded from further analyses due to collinearity
with the other two types of naming errors. See section 4.2.3 below.
73
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
m any different item s. Thus, the effect of dam age to a particular features
on nam ing w ill depend on the im portance of that feature in a given
sem antic representation. For exam ple, dam age to a feature denoting HAS
STRIPES should have little effect on nam ing CAT but m ay play a crucial role
in nam ing TIGER. Furtherm ore, because features m ay participate in m any
different sem antic representations, dam age m ay affect the nam ing of
several item s. Consequently, nam ing errors are seen more as indicators of
the extent of dam age to the sem antic system in this approach, rather than
as an indicator of specifically dam aged items. This approach to nam ing
errors contrasts im portantly w ith studies in which nam ing errors of any
type are assum ed to designate the particular lexical items that are dam aged
or inaccessible (e.g., Chertkow & Bub, 1990). In studies of that type,
sem antic representations are usually seen as noncom positional (see
C hapter 2) a n d the crucial variable is w hether errors on a particular w ord
are m ade consistently or inconsistently across tasks and time. However,
considering th at dam age to the sem antic system in AD is relatively
random , one could not generalize from a subject's nam ing score on any
particular test to the overall degree of sem antic im pairm ent, since there
w ould be no guarantee that the particular sam ple tested was in any way
representative of the entire vocabulary.
4.2.3 Independence of M easures
Because m any of the analyses planned for the speech production
task included stepw ise m ultiple regressions, all WM and sem antic
m easures w ere exam ined for m ultiple collinearity by regressing each
potential variable on all other variables (Lewis-Beck, 1980). Separate
74
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
sequential stepw ise regressions were completed using each of the three
WM and three sem antic m easures in turn as dependent variables, with
the other five m easures as factors, treating each group separately.
According to Lewis-Beck, variables that share m ore than 90% of their
variance should be considered collinear and unsuitable for concurrent use
in regression equations.
The analyses of AD subjects' WM scores revealed that Backward-
Recall scores shared about 37% of their variance w ith O rdering scores, and
about 26% w ith Nonsem antic-W M scores. Nonsem antic-W M and
O rdering shared 51% of their variance. However, in the stepwise
regression of O rdering scores m ore than one factor entered the equation.
In that analysis, N onsem antic-W M entered the equation first, with an r2
of .51, as m entioned above, but the second factor to enter the equation was
W ithin-Category nam ing errors, one of the sem antic ability measures,
which accounted for an additional .20 of the variance for a total r2 of .71, (p
< .02). Thus, success on the O rdering task for AD subjects required both
WM and sem antic ability. However, considering the definition of
collinearity in Lewis-Beck (1980) (i.e., shared variance m ust approach or
exceed .90), these three WM measures were not collinear and, hence,
qualified to be used as independent WM factors for AD subjects in the
following study.
The analyses of AD subjects' semantic scores revealed that Within-
Category error rates w ere not predictable from any of the other five factors
here. No-Response error rates shared 37% of variance in scores with
O rdering, but did not correlate with any other m easure. However, the
75
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
other sem antic ability m easure, Superordinate/C ircum locution nam ing
responses, proved to be collinear w ith W ithin-Category and No-Response
errors, w ith these tw o factors accounting for 92.7% of its variance.
Consequently, the Superordinate/C ircum locution nam ing errors were
excluded from further analyses, and W ithin-Category and No-Response
error rates becam e the semantic measures for the AD group.
Regarding WM and semantic m easures for the NC group, only
three were tested: O rdering, Nonsemantic-W M, and PPVT. Since fifteen
of sixteen NC subjects scored at ceiling on Backward-Recall, this m easure
was elim inated from NC analyses. Nonsem antic-W M and PPVT both
shared variance w ith O rdering scores, (r2 = .44 and .40 respectively, both p's
< .02), and together accounted for an r2of .581. However, these two factors
did not account for a large enough portion of the variance in O rdering
scores to qualify as collinear measures. Based on these findings, three
measures have been retained as factors for the analyses of NC speech
production: O rdering, Nonsemantic-W M, and PPVT.
Thus, in both populations success on O rdering depended on both
semantic and W M abilities, but in neither case w as it completely collinear
with any com bination of the measures here. Consequently, it will be
included as a factor in the analyses of speech as a measure, perhaps, of the
ability to coordinate sem antic and WM resources. Further, Nonsem antic-
WM represented a relatively independent m easure for both groups, and
appeared to m easure WM ability independent of semantic effects. O n the
other hand, Backward-Recall scores only varied am ong AD subjects, and
thus seemed to m easure a type of WM that w as not affected in norm al
76
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
aging, b u t w as vulnerable in different degrees even at the m ild to
m oderate stage of A lzheim er's Disease, (e.g. seven of the ten AD subjects
scored below 100% on this task). W ith respect to sem antic ability
m easures, am ong AD subjects, the num ber of S uperordinate/
Circum locution errors proved to be almost completely predicted by a
com bination of W ithin-C ategory and No-Response errors, and so w as not
included as a sem antic m easure in the following analyses. This left
W ithin-Category and N o-R esponse error scores to constitute the semantic
m easures used for AD speech in the following study, w hile PPVT was
used as the m easure of sem antic ability in analyses of NC speech.
4.2.4 Speech Production A ssessm ent
Chapters 5 and 6 exam ine the speech of AD and NC subjects under
different conditions. C hapter 5 analyzes spontaneous speech, while
C hapter 6 examines perform ance on a constrained production task. The
spontaneous speech sam ple consisted of the text of a conversation between
the subject and experim enter w hich covered a set of fam iliar topics: the
subject's career before retirem ent, meeting their spouses, their family life
and children, and any hobbies they enjoyed. In the constrained
production task, stim uli consisted of a set of a verb and tw o nouns, and
the task required subjects to create a sentence using all three w ords. Both
production tasks w ere audio taped, transcribed verbatim , an d checked by
the author for accuracy. The author then scored the transcripts for lexical
and gram m atical errors. The frequency of errors w as converted to the
num ber of errors per h u n d red w ords to control for differences in verbosity
and to facilitate com parison betw een tasks.
77
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Four kinds of errors were analyzed: pronouns, m orphosyntax, open
class w ord, and closed class w ord errors. N ote that pronouns are not
considered to be closed class items in this stu d y , although in linguistic
theory they are, because they also come from a sm all set that is resistant to
historical change. H ow ever, pronouns perform an entirely different kind of
function than other closed class w ords and have a m eaning that is prim arily
referential rather than grammatical; consequently, they have been
considered a separate w ord type in this research. In addition, all errors were
categorized as om issions or substitutions w here applicable. Pronoun, open
and closed class errors counted as om issions w hen a w ord was not
pronounced and a substitution when an incorrect w ord was used. Pronoun
errors also had an extra category, am biguous reference errors, that included
all instances in w hich the referent for a pro n o u n w as unrecoverable from
the context. N oun phrase omissions w ere only counted as pronoun
omissions if the only possible missing w ord w as a pronoun, (i.e., WE, US, ME,
YOU). M orphosyntactic errors included incorrect m orphological inflections,
m issing m atrix or required subordinate clauses, verb subcategorization
errors, and abnorm al w ord order. These errors w ere not divided into
om ission or substitution error types, since it w as not clear that this
distinction could be system atically applied to these errors without invoking
a particular gram m atical theory. For exam ple, using CALL instead of
CALLING in a sentence w ould be categorized as the om ission of an inflection
in one theoretical paradigm but the substitution of an inflection (the zero-
grade form) in another. Appendix C show s the num bers of raw errors of
each type that w ere m ade by the two groups in each task.
78
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Besides errors, this research also examined two types of speech
dysfluencies, "retracings" and "fresh starts." Retracings occurred w hen a
speaker began an utterance, but broke it off in order to backtrack for two or
m ore w ords, and then continued the utterance, repeating several words,
(e.g. "Paul show ed, Paul show ed the movie in the evening," subject # 132).
The second form of dysfluency consisted of those instances in w hich the
speaker abandoned an utterance that had been started, and m ade a "fresh
start" using completely different w ords and syntax, (e.g. "The pill, Diane
took the pill," subject #146). Single w ord repetitions are not included as
false starts for the purpose of this research, because these m ay be quite
different from other retracings in th at they do not have a disruptive effect
on discourse processing (Tree, 1995).
The tw o production tasks, the rationale behind them , and their
results are described in more detail in Chapters 5 and 6.
79
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Chapter 5.
A nalyses of Spontaneous Speech
This chapter com pares the spontaneous speech production of
Alzheim er's Disease (AD) patients and healthy elderly speakers w ith respect
to the num ber and distribution of lexical and gram m atical errors, as well as
the types of false starts they m ake. In addition, each type of error is
exam ined for the influence o f working m em ory (WM) and sem antic
im pairm ent. Regarding dysfluencies, this chapter explores the assertion by
Levelt (1989) that certain types of speech dysfluencies (restarts) require the
use of WM resources, and as such, should decrease in frequency w hen WM
is limited. W ith respect to lexical errors, the analyses in this chapter test
several predictions based on the two dom inant theories of speech
production (Garrett, 1975,1984; Levelt, 1989) as described in Chapter 2:
1) only lexical choice errors which involve open class w ords should
show a correlational relationship w ith sem antic im pairm ent
m easures;
2) om ission errors should be relatively absent; and
3) individual differences in WM should have no im pact on speech
production.
On the other hand, spontaneous conversation m ay present
significant challenges to subjects w ith language a n d /o r WM im pairm ents
because of the m any concurrent dem ands for com putation and storage at
several linguistic levels (D anem an & Carpenter, 1986; H ofstede & Kolk,
1994; Kempler & Zelinski, 1994). That is, carrying on a conversation
requires subjects to sim u ltan eo u sly understand the interlocutor, form ulate
80
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
coherent responses, access all the open class w ords necessary to express one's
ow n thoughts, m aintain pronoun reference throughout the discourse, and
use correct gram m ar (e.g. m aintain subject-verb agreem ent). From this
point of view, spontaneous speech m ight be expected to cause difficulty w ith
the use of open class w ords, m orphological inflections, and pronouns w hen
sem antic and WM ability w ere im paired. N evertheless, spontaneous
conversation is a w ell-practiced skill th at allow s subjects to choose their
ow n vocabulary an d syntactic constructions. Thus, the flexibility inherent
in spontaneous conversation in term s of the topics covered and w ord choice
m ight well m itigate any effects of the difficulties m entioned above. These
three conflicting hypotheses are exam ined in the following analyses.
This chapter is organized into four parts. The first section describes
the speech production task and the scoring technique used. The following
two sections present analyses of speech dysfluencies and lexical choice
errors, respectively, discussing both their distribution and etiology. The
final section of this chapter sum m arizes these findings.
5.1 M ethods
The spontaneous production task consisted of a personal narrative
(PN) elicited in an inform al conversation w ith the experim enter. The
purpose of the PN w as to obtain sam ples of subjects' speech that w ere
representative of their norm al production abilities. The experim enter
engaged subjects in conversation for five to ten m inutes by asking questions
about their lives, w here they were bom , w h at their childhood w as like,
w hat type of career or jobs they had had, favorite hobbies, m em orable
holidays, and how they m et their spouses. The intention of this interview
81
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
w as both to obtain data for this study and to acquaint the subject and
experim enter w ith each other; thus, conversation ranged freely at the
experim enter's discretion, since it was only errors, not content, that w ere the
concern of this study. The speech sam ples used for analysis consisted of
approxim ately 500 w ords of consecutive dialogue from each subject's
narrative.
All personal narratives w ere audio taped, transcribed verbatim by a
trained research assistant, and checked by the author for accuracy. The
author then scored the transcripts for false starts and unrepaired errors. The
num bers of errors and false starts were converted to a m easure of errors or
false starts p er hundred w ords in order to control for effects of verbosity.
As described in Chapter 4, there were four categories of errors that
were analyzed: pronoun, open class word, closed class word, and
m orphosyntax. A further distinction was m ade betw een omissions and
substitutions in all categories save m orphosyntax. In addition, the num ber
of dysfluencies, or false starts, were tabulated for both tasks. Two types of
dysfluency w ere tracked: 1) "retracings" consisted of those instances in
which a speaker began an utterance, but broke it off to backtrack for two or
m ore w ords an d then continued the utterance, and 2) "fresh starts"
consisted of those instances in which the speaker broke off an utterance, and
then continued the turn using different w ords and syntax (Levelt, 1989;
Valencia-Laver, 1992).
82
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
5.2 Dysfluencies in Spontaneous Speech
This section presents analyses of the rate at w hich subjects produced
speech dysfluencies. N orm al speakers may produce dysfluencies as the
result of various transitory conditions, such as lapses of attention, word
finding difficulties, a poor initial choice of sentence structure or just
thinking of a better w ay to say something. It can even be a habitual m ode of
speech; som e people, especially when speaking to people they know well,
frequently begin a sentence and stop, knowing by facial expressions, nods or
other cues that the other people in the group follow their train of thought.
In a population w ith both sem antic and WM deficits such as AD either of
these im pairm ents m ight increase the frequency of dysfluencies.
To com pare the frequency and distribution of the tw o types of
dysfluencies in the spontaneous speech of the two groups, their false start
rates were entered into a tw o by two (Group by Dysfluency Type) ANOVA
using false starts per hundred w ords as the dependent variable. This
analysis revealed that in overall dysfluency rates there w as little difference
in how often the tw o groups interrupted their ow n speech; the main effect
of group did not approach significance, (F(l, 48) = .197, p > .6). This confirms
reports in the literature that the speech of AD subjects is as fluent as that of
healthy elderly speakers. In addition, both groups tended to make more
fresh starts than retracings, (F(l, 48) = 3.431, p = .07). H ow ever, the
interaction betw een group and dysfluency type was only m arginally
significant, (F(l, 48) = 2.257, p < .14). This finding w as explored with
Bonferroni analyses w hich show ed that NC subjects w ere equally likely to
retrace an utterance as to m ake a fresh start, (f(15) = .387, p > .7), whereas, AD
83
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
subjects tended to m ake a higher proportion of fresh starts in their
spontaneous speech, (f(9) = 1.767, p = .11), though the difference w as not
significant, as show n below in Figure 2.
Figure 2. False Start Rates in spontaneous speech.
1.8
1.6 -
-a 1.4 -
U t
0
i , j i
c 1 -
w
Q J
1 .8 1
v
3
.6 -
.4 -
.2 '
□ Fresh Starts
Hi Restarts
1
AD NC
In order to determ ine w hether the dysfluency rates of either group
related to sem antic or WM ability, their fresh start and retracing rates were
subm itted to a sequential stepwise regression (F to enter = 3). First, NC
dysfluency rates were analyzed using the factors described in C hapter 4, two
WM m easures (O rdering and Nonsemantic-W M ) and the sem antic
measure, PPVT. However, the frequency of both types of dysfluency proved
to be independent of these factors. This parallels findings in recent research
which found no relationship betw een WM and other dysfluency m easures
in the speech of healthy elderly speakers (Valencia-Laver, 1992), while
extending the statem ent to show no relationship betw een dysfluency rates
and sem antic ability either. From this it appears that dysfluency rates of
84
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
healthy elderly subjects are likely the result of norm al variation in the
population due to differences in speech style.
Secondly, the rates of both types of dysfluencies by the AD subjects
were analyzed for effects of WM and sem antic im pairm ents, using the two
sem antic m easures (W ithin-Category and N o-R esponse nam ing errors) and
three WM m easures (O rdering, Backward-Recall and Nonsemantic-W M )
described in C hapter 4. The num ber of fresh starts w as m arginally predicted
by Nonsem antic-W M scores, (r2 = ".34, p <.08), w ith better WM being
associated w ith few er fresh starts. In the analysis of retracings, only
Backward-Recall entered the equation for an r2 of .48, (p < .05), again with
the effect that m ore im paired WM predicted increased restarts. This latter
finding contradicted the prediction based on the assertion in Levelt (1989)
that the production of retracings in particular should consum e WM, and,
consequently, that their frequency should decrease as WM ability decreased.
Instead, these results reflected the opposite p attern — increased dysfluencies
of both types w ere associated w ith m ore im paired WM.
The analysis of the frequency of speech dysfluencies presented here
confirms reports that fluency of spontaneous speech is preserved in AD
patients. How ever, it also hints that the types of dysfluencies m ade by AD
subjects maybe qualitatively different than those m ade by NC subjects, in
that AD speakers tend to m ake a higher proportion of fresh starts. The
results concerning the etiology of AD subjects' speech dysfluencies suggest
that decreases in W M ability m ay be accom panied by increased dysfluencies,
especially fresh starts. Thus, w hen sufficient W M resources are not
available, fluency of production m ay be som ew hat com prom ised, though
85
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
not necessarily significantly. H ow ever, differences in WM did n o t affect the
dysfluency rates of the NC subjects. From this, it appears that the NC
dysfluency rates reported here represent a baseline dysfluency rate w ith no
WM com ponent, at least as m easured by the two WM tasks used here.
W hat is m ore difficult to explain is how the AD dysfluency rate can be so
sim ilar to the N C rate, and predicted b y WM im pairm ents, w hen these have
little effect on NC scores. It could be th at there is a threshold effect, such
that im pairm ents in WM only affect dysfluency rates when they are of a
certain severity; in this case, NC subjects w ould have sufficient resources to
be im m une to WM effects, while the frequency of dysfluencies am ong the
AD patients m ay reflect the fact that their WM ability is just below the
crucial threshold. This hypothesis im plies that, although the AD speakers
do not show significant increases in the overall frequency of dysfluencies at
this tim e, they m ight as WM becom es m ore impaired.
5.3 The Distribution of Lexical Choice Errors
This section examines the distribution of speech production errors in
the spontaneous speech of NC and AD subjects in terms of error type
(omission or substitution) and the w ord or m orphem e type involved. As
previously stated, only open class, closed class, and pronoun errors are easily
classified as om issions and substitutions, since it is not clear that
m orphosyntactic errors can be separated into these categories using
atheoretical criteria. Consequently, m orphosyntactic error rates will be
analyzed separately. As presented in C hapter 2, extrapolations based on
traditional speech production theories predict that, since neither group has a
syntactic im pairm ent, neither group w ill m ake a significant num ber of
86
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
lexical om issions or closed class errors. Similarly, since the prim ary
linguistic deficit in AD is semantic, these theories predict only open class
substitution errors for this group. In contrast, interactive theories of
language representation predict that AD subjects m ay make errors on all
word types, and that the frequency of these errors will be related to semantic
im p airm en t.
First, it m ust be m ade clear that the num ber of errors in both groups
was quite low. The m ean num ber of total errors by AD subjects was 1.45 per
hundred w ords, or approxim ately 7.25 errors in a five hundred word sam ple
of speech; however, all AD subjects m ade at least three errors and the most
error-prone m ade nineteen. In contrast, the NC subjects erred at a rate of
only .44 errors per hundred w ords, or 2.2 errors in the five hundred word
sample. O nly one NC subject produced error-free speech, five more m ade
only one error, and the m axim um num ber of errors in a sample of NC
speech w as six. The distribution of these errors across word types is pictured
in Figure 3.
As can be seen in Figure 3, AD subjects m ade more errors on all w ord
types com pared to NC subjects. Error rates on open and closed class w ords
and on pronouns were analyzed using a two by two by three (Group x Error
Type x W ord Category) ANOVA with the num ber of errors per hundred
w ords as the dependent variable. The m ain effect of group was highly
significant, since AD subjects m ade m any m ore errors overall than NC
subjects, (F(l, 144) = 22.444, p < .0001). O n the other hand, both the w ord
category m ain effect and group by word category interactions were
insignificant, (F(2,144) = 1.225, p > .30, and F(l,144) = .232, p > .75,
87
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
respectively). Thus, for each group, the proportion of errors m ade in these
three w ord classes d id not significantly differ.
ngure 3. The frequency of errors in spontaneous speech by w ord type.
Error Category
□ Closed Class
I Open Class
M orphosyntax
Pronoun
■ S .40
< u .30
Sim ilarly, neither the main effect of error type (omission or
substitution) o r the group by error type interaction were significant, (F(l,
144) = .053, p > .8, and F(l, 144) = 1,382, p > .20, respectively), revealing that
the proportions of substitution and om ission errors did not vary
significantly w ithin or between groups. O n the other hand, closed class
words w ere m ore likely to be om itted than substituted, while open class
words and pronouns show ed the opposite pattern, as reflected in a
significant w ord type by error type interaction, (F(2,144) = 7.617, p < .001), as
shown in Figure 4. Im portantly, this pattern held true for both groups, as
shown by the lack of a three way interaction, (F (2,144) = 1.298, p > .30); thus,
in the spontaneous speech of both populations, closed class w ords w ere
88
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
m ore likely to be om itted than substituted, w hile open class w ords and
pronouns show ed the opposite tendency.
Figure 4. M ean om ission and substitution rates, com bined across groups.
.60
Omissions
Substitutions
c: -30
C O
o
£ .20
.10
Open Closed
Pronoun
Class Class Errors
Errors Errors
The final error analysis for spontaneous speech exam ined the
frequency of m orphosyntactic errors by the tw o groups. A two-tailed, two
groups f-test revealed that AD subjects m ade significantly m ore
m orphosyntactic errors than NC subjects, (f(24) = 2.170, p < .05), with means
of .310 errors per h u n d red w ords compared to only .122 for NC subjects.
These results have illustrated several characteristics of the
spontaneous speech of both AD and NC subjects. First, they show ed that
AD patients m ade m ore errors compared to NC subjects on pronouns, open
and closed class w ords, revealing that AD subjects' errors w ere not limited
to open class substitutions. Second, they revealed that om ission errors
constituted a considerable proportion of the errors of both groups, in that
there was no effect of error type in either group. T hird, these findings point
out that both groups m ade a considerable proportion of their errors on
89
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
closed class item s, and that closed class w ords w ere more likely to be om itted
than substituted. Lastly, they showed that AD subjects also m ade m ore
m orphosyntactic errors than NC subjects. Thus, AD subjects m ade m ore
errors of all types than the control group.
5.4 The Etiology of Lexical Errors
This section addresses the central question of this research— w hether
limitations in W M a n d /o r semantic ability affect the frequency of speech
errors in healthy elderly speakers and subjects w ith AD. The following
analyses exam ine the relationships betw een individual differences in WM
capacity and sem antic ability and the frequency of different types of speech
errors, first for N C subjects, then for AD patients.
5.4.1 Analyses of the etiology NC speech errors
This section examines the errors of the NC group for effects of WM
and semantic ability using stepwise regression analyses in which all factors
with an F value greater than 3.0 were allow ed to enter the equation. In
order to be eligible for a regression analysis, the m ean frequency of the error
type had to be at least .20 per hundred w ords and at least half the subjects in
the group had to have m ade errors of the given type. Since the error rates of
the NC subjects w ere extremely low, only tw o error types qualified for
analysis: total errors and total omissions. Factors in these analyses included
the NC sem antic ability measure (PPVT) , and the WM factors, O rdering and
nonsem antic-W M .1
lPPVT scores were available for only 13 of the 16 NC subjects.
90
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
The total num ber of errors in NC speech proved to predictable from
PPVT scores, (r2 = .327; F (l,ll) = 5.344, p < .05). In this analysis, subjects with
higher overall sem antic ability m ade fewer total errors in spontaneous
speech. In contrast, O rdering, a WM measure, was th e only factor to enter
the equations for total omissions, accounting for 33% of the variance in
scores, (F (l,ll) = 4.674, p < .05), reflecting the fact th at NC subjects with lower
O rdering scores om itted m ore words in spontaneous speech. It should also
be pointed out that about two thirds of the words om itted by NC subjects
were closed class words.
These analyses show that both WM and sem antic ability may affect
the accuracy and gram m aticality of speech in healthy elderly subjects.
Semantic ability, as m easured by vocabulary size, w as associated with the
overall accuracy of lexical choice. On the other hand, lower WM predicted
an increased frequency of omissions, the majority of w hich consisted of
omissions of closed class words. Thus, both sem antic ability and WM may
play a role in ensuring the accuracy and gram m aticality of spontaneous
speech in norm al speakers.
5.4.2 Analyses of AD patients' speech errors
The follow ing stepw ise regression analyses exam ine the relative
contribution of sem antic ability and working m em ory capacity to different
types of speech errors m ade by AD patients. As above, the frequency of
errors needed to be at least .20 errors per hundred w ords, and the particular
error type had to occur in at least 50% of the transcripts for the error type to
be analyzed. Consequently, four categories of errors w ere excluded from
analysis as a result of these criteria: pronoun and closed class word
91
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
substitutions, and pronoun and open class w ord omissions. The
independent factors used in the following analyses consisted of three WM
m easures (Backward-Recall, Nonsem antic-W M and O rdering) and tw o
sem antic m easures (W ithin-Category and N o-Response nam ing errors); the
criteria for entering the stepw ise regression w as m aintained at F 3.0. To
foreshadow the results of these analyses, all of the AD error rates exam ined
proved to be related to either semantic im pairm ent, or a com bination of
WM and sem antic im pairm ents, except the frequency of m orphosyntactic
errors w hich w as n o t predictable from these m easures.
The only types of pronoun errors to fulfill the above inclusion
criteria w ere am biguous reference errors, pronouns whose referent w as
unrecoverable from the context, and the total num ber of pronoun errors,
which included the am biguous reference errors as well as substitutions and
omissions of pronouns. These types of difficulties w ith pronoun use have
been well docum ented in studies of AD subjects' speech but the etiology of
these errors has n o t been experim entally established, although it has been
suggested that they result from failures in m em ory (Ulatowska et al., 1986)
or a pragm atic deficit (Ripich & Terrell, 1987). However, the results
obtained here identified sem antic im pairm ent as being a prim ary factor in
both categories of pronoun errors. In the analysis using total pronoun
errors as the dep en d en t variable, W ithin-Category nam ing errors was the
only factor to enter the equation, w ith an r2 of .554, (p < .01), such that
increased num bers of W ithin-Category errors in nam ing predicted increased
pronoun errors. Sim ilarly, the frequency of pronouns with am biguous
reference also proved to be related to sem antic but not WM capacities. In
this case, W ithin-C ategory nam ing errors predicted the frequency of
92
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
am biguous reference errors, (r2 = .42, p < .04), w ith the sam e effect— increased
difficulty w ith nam ing w as associated w ith higher num bers of pronouns
w ith am biguous reference.
The only error type for w hich m odular theories of speech production
clearly predicted effects of sem antic im pairm ent was open class
substitutions. In the regression equation addressing these errors, W ithin-
Category nam ing errors again entered first, such that increased W ithin-
Category errors in nam ing predicted increased num bers of open class
substitutions in the narrative, (r2 = 42%). Subsequently, Nonsem antic-W M
entered the equation w ith the effect th at subjects with higher W M m ade
m ore errors. The result after this factor entered the equation w as r2 = .638,
(p < .03). Thus, it w as the conjunction of high num bers of W ithin-Category
errors plus relatively intact WM that resulted in the highest rates of open
class substitutions. This pattern of factor usage is not easy to understand
and is addressed at length in C hapter 7.
In the analysis of closed class om issions, the only factor to enter the
equation w as W ithin-Category nam ing errors, which accounted for 55% of
the variance in scores, (p < .02). As w ith pronoun and open class errors,
higher num bers of W ithin-Category nam ing errors were associated with
increased num bers of closed class om issions.
A lthough AD subjects did m ake m ore m orphosyntactic errors than
NC subjects overall, the frequency of these errors was not predictable from
sem antic or W M m easures.
93
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
5.5. Discussion of Results
The analyses in this chapter com pared the fluency and frequency of
different types of lexical and gram m atical errors in the spontaneous speech
of AD and NC subjects. Some of these findings confirm long-standing
assum ptions of the AD literature; however, m ost of the results detailed here
contradict the m ore explicit expectations of m odular speech production
theories. The m ost im portant of these include:
1) AD patients produced the sam e num ber of overall speech
dysfluencies as NC subjects in the PN, although they m ade m arginally more
fresh starts, a subtype of dysfluency, com pared to the NC group. This
reconfirms the usual description of AD speech production as being fluent.
2) AD subjects w ith lim ited WM tended to produced more
dysfluencies of both types than AD subjects w ith m ore intact WM.
However, the dysfluency rates of NC subjects w ere not affected by
individual differences in WM.
3) C om pared to NC subjects, AD subjects m ade more errors in all
error categories, including closed class w ords and morphosyntax. This does
not support the prediction of m odular speech production theories that only
open class w ord errors w ould increase in AD patients' speech.
4) A high proportion of the errors of both groups of subjects consisted
of om itted w ords. Thus, om itted w ords are not necessarily an abnorm al
phenom enon that is the result of syntactic deficits.
94
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
5) AD an d N C groups did not differ in the distribution of errors
among w ord types or betw een omissions and substitutions. Therefore,
errors w ere sim ilarly distributed across w ord types and between error types
in the speech of both healthy elderly and AD patients.
6) Total errors by the NC group were predicted by their sem antic
ability as m easured by vocabulary size. Thus, better semantic ability in the
form of w ord know ledge w as associated w ith m ore accurate speech.2
7) The frequency of whole w ord om issions by NC subjects proved
predictable using a m easure of WM, O rdering. Further, 75% of these errors
consisted of the om ission of closed class words. Therefore, the tendency to
omit w ords in spontaneous speech particularly closed class words, m ay be a
relatively com m on phenom enon that m ay be exacerbated when WM is
lim ited.
8) The m ost pow erful predicter of AD speech errors was the num ber
of W ithin-Category nam ing errors on a 96 item nam ing test. Based on this,
there appears to be a direct relationship betw een the ability of subjects to
accurately nam e pictures and the accuracy with which they can produce
spontaneous speech. This comm on elem ent has been interpreted here as
the degree of sem antic im pairm ent. Interestingly, the measure of severe
semantic im pairm ent, No-Response errors, did not contribute to any error
rates in spontaneous speech.
9) Pronoun errors, and in particular pronoun errors w ith am biguous
reference, w ere highly related to the sem antic m easure, W ithin-Category
interestingly, there was no relationship between vocabulary size and education, or education
and production errors in this group of elderly control subjects
95
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
errors. This is theoretically im portant, because the role of sem antic
activation of the referent for later use in defining a pronoun is not usually
considered in discussions of pronoun production. How ever, it is clear that a
failure to fully activate the sem antic representation of a referent initially.
and to m aintain this activation over time, could im pair subjects' pronoun
use later in a conversation (Almor, Kem pler, M acDonald & A ndersen,
1997).
10) In the speech of AD subjects, sem antic im pairm ent an d WM
abilities may interact in com plex ways, exemplified by the finding that more
intact WM m ay lead to increased open class errors w hen sem antics are
im paired.
The m ost theoretically interesting findings here are those dealing
w ith closed class w ord errors. Since closed class w ords are crucial to speech
production theories (see C hapter 2), their vulnerability u nder conditions of
both lim ited WM and sem antic im pairm ent is theoretically quite
im portant. The finding that the frequency of nam ing errors, a m easure of
sem antic im pairm ent, w as associated w ith the frequency of closed class
om issions cannot be easily explained w ithin either G arrett's or Levelt's
theories. For instance, one w ould have to posit that the closed class words
affected in spontaneous speech w ere those which, according to Levelt, have
sem antic representations, (e.g. m odals and locational prepositions).
H ow ever, in an uncontrolled setting such as spontaneous speech, it is
difficult to directly com pare the frequency of errors on only "gram m atical,"
as opposed to "sem antically represented," closed class w ords d u e to inter
subject variability of topic and w ord choice. That is, subjects m ay not
96
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
produce enough of the tw o types of closed class w ords to directly compare
their error rates. Furtherm ore, even from the point of view of an
interactive theory of speech production, these closed class errors may have
been either a direct result of semantic im pairm ent or a "side effect" of the
sem antic dem ands of spontaneous conversation (e.g. m aintaining pronoun
reference and accessing lexical items). To address this issue, an experiment
w as devised w hich both m inim ized lexical access requirem ents and
required subjects to provide only gram m atical closed class words in a
sentential context. This task and its results are presented in the following
chapter.
97
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Chapter 6.
Analyses of Constrained Production
This chapter addresses the question of w hether Alzheim er's Disease
(AD) patients are able to produce the required closed class items for a
gram m atical sentence w hen the dem ands of open class word lexical access
are m inim ized. This question was posed as a direct result of the findings in
the previous chapter in which AD patients show ed relatively high rates of
closed class w ord errors in spontaneous speech. To recap this argum ent, AD
patients are often described as suffering from a deficit in lexical access,
therefore, it is possible that the high error rates on closed class w ords were
actually a side effect of the intense dem ands that spontaneous speech put on
lexical access m echanism s, which, in turn, increased errors in production
across the board at a rate proportional to sem antic impairment. To
minimize these effects, the current experim ent constrained production of
sentences by providing all the open class w ords necessary for a sentence to
subjects, thereby, em phasizing sentence construction abilities rather than
lexical access.
Because of the characterization of AD as a lexical access deficit and the
nature of the task, the original concern w as that this task m ight be trivially
easy for AD patients, as it proved to be for the elderly control subjects. In
fact, AD subjects had a variety of difficulties w ith the task which were not
limited to closed class w ord usage. Exam ination of the data revealed five
types of difficulties: m ultiple dysfluencies, whole sentence productions that
subjects rejected as unacceptable (e.g. saying "Charlie beaten Don, no it has
to be Don beaten Charlie."), lexical or gram m atical errors (i.e. om itted closed
98
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
class w ords or subject-verb agreem ent errors), m isused stim uli (responding
"M ary grew roses." to the stim ulus GROWN ROSES MARY), and inability to
m ake a sentence (e.g. "I d o n 't know " responses). These difficulties are first
analyzed as a whole, then the distribution and etiology of the three most
comm on error types are considered separately.
To foreshadow o u r findings, AD patients had considerable difficulty
w ith different aspects of sentence construction in this task, as show n by
increases in dysfluency rates, closed class w ord om issions, and syntax errors.
Moreover, m any of the difficulties that subjects experienced w ith this task
w ere concentrated in response to stim uli from one particular verb condition,
irregular past participles. To explore this finding further, item analyses were
com puted to determ ine the relative impact of w ord frequency, num ber of
m eanings and noncanonical m orphosyntactic form. The results of these
analyses comprise the last section of this chapter.
6.1 Method
The Constrained Production (CP) task was designed to elicit purely
gram m atical closed class w ords, (i.e. auxiliary verbs, agentive prepositions,
and determ iners), in a sentential context. To accom plish this, the task
encouraged the production of passive sentences by using a paradigm similar
to that of an experim ent designed by F. Ferreira (1994) which examined the
influence of verb type on the frequency of production of passive sentences.
She presented subjects w ith a com puter screen show ing tw o nouns, which
consisted of either two anim ate nouns or one anim ate and one inanimate
noun (in either order), follow ed by a verb. Ferreira reported that subjects
produced significantly m ore passive sentences w hen using Theme-
99
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Experiencer, "psyche" verbs (e.g., THRILLED, SURPRISED) than w hen using
"regular," A gent-Them e, transitive verbs (e.g., CALLED, POURED). M oreover,
she found that this tendency w as m ore pronounced when there w as a
m ism atch in anim acy betw een the tw o nouns provided for use in the
sentence, that is, w hen the stim ulus contained an inanimate and an
anim ate noun. C onsequently, Them e-Experiencer verbs and noun stim uli
m ism atched in anim acy com prise several of the conditions used in this
experim ent.
The design of the present experiment varied slightly from that of
Ferreira's study. First, the dependent variables of interest were error and
dysfluency rates rather than reaction times, and so an off-line presentation
of stimuli was adopted. Second, there was concern with the linear
arrangement of stim ulus words that Ferreira had used, since she reported a
small, marginally significant trend for subjects to use the left-most "first"
noun as the sentence subject. To avoid this, we chose not to present the
stimuli in a linear form and positioned the verb in the prominent, left-most
position. Thus, the stimuli for this experiment appeared as shown below
in Figure 5, presented in 24 point Helvetica type on 5" x 8" index cards. The
word on the left of the card was always the past participle of a verb from one
of the three verb types described below, and the other two words were either
common names (e.g. SUSAN or MIKE) or inanimate nouns (e.g. MOVIE, TOY,
MEETING). The names used were chosen to be relatively frequent and
unambiguous as to gender, such as TOM, SUSAN, FRED, and LAURA. The
inanimate nouns were chosen specifically to be highly appropriate
arguments for particular verbs so that, semantically, a connection would be
relatively easy to make between them, i.e. POURED/MILK, THRILLED/MUSIC,
100
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
INSPIRED/SERMON, GROWN/ROSES. There were three different noun
com binations used: Nam e above1 N am e (2-Names), N am e above
Inanim ate N oun (Name-first), and Inanim ate N oun above N am e
(Inanim ate-first). The fourth possibility, two inanim ate nouns, was not
used because one of the verb types, (Theme-Experiencer verbs), requires at
least one anim ate argum ent.
Figure 5. Sample of the stim uli used in the CP task.
ball
thrown
Tommy
The first of the three verb types, regular (REG) verbs, has identical
past participle and sim ple past tense forms, and, in a sim ple active sentence,
assigns the thematic roles agent and theme to its subject and object,
respectively. Some examples of the REG verbs used were BUMPED, c a r r ie d ,
and POURED. The second type of verb, also known as "psyche verbs" or
Theme-Experiencer verbs, assigns the thematic roles them e and experiencer
to its subject and object in a sim ple active sentence (T-E verbs); thus, active
sentences will be of the form "inanim ate noun-verb-anim ate noun" (e.g.,
THE MOVIE SHOCKED S a ra h .) All of these verbs have equivalent
'Refers to placement of noun stimuli on presentation card.
101
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
phonological forms for the sim ple past and the past participle and come
from the category of "am use" verbs (Levin, 1993), for example, SURPRISED,
SHOCKED, STARTLED and ANNOYED. The last verb type consists of transitive
verbs with "irregular" past participles (IRREG verbs), i.e. the simple past and
past participles have distinct phonological forms. Some examples of these
are TORN, THROWN, and FORGOTTEN. All IRREG verbs assigned thematic
roles in the sam e w ay as REG verbs did: in a sim ple active sentence the
subject was the agent and the direct object w as the theme. When stimuli
included IRREG verbs (e.g. THROWN PLATE LISA), the morphological form
forced subjects to use complex syntactic structures, such as an active perfect
(i.e. HAD THROWN), a passive (WAS THROWN), or a deverbal adjective (i.e.
THE THROWN PLATE). A full list of CP stimuli is located in Appendix D. In
fact, this task proved to be very successful at eliciting passive sentences, w ith
control subjects producing 62.5% of their responses to T-E verbs in the
passive voice (as com pared to passive rates of 19-35% in Ferreira's original
study), and 29.5% of the responses to IRREG verbs in the passive.
Responses on the CP were scored using the same criteria as used in
the analyses of spontaneous speech in the preceding chapter. Dysfluencies
were categorized as either fresh starts, if an utterance was abandoned in
favor of a different attem pt, or retracings, w hen an utterance in progress was
interrupted so that the speaker could back up and repeat two or more of the
w ords already articulated and then continue the same sentence. Lexical
errors consisted of om issions or substitutions of open and closed class w ords
or pronouns and errors in morphosyntax. These error types have been
described m ore fully in Chapter 4. In addition, there were three response
types unique to this task, m isused stimuli, rejected responses and no
102
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
response. M isused stim uli refer to those instances in which one of the three
stim ulus w o rd s w as either omitted from the response (e.g., stim ulus:
GROWN ROSES MARY; response, "M ary liked roses), m orphologically altered
(e.g., "M ary grew roses."), or sem antically substituted (for exam ple, "M ary
raised roses."). This error type is described and analyzed at length in section
6.2.4 of this chapter. Rejected responses are actually a subtype of dysfluency
and include those responses in which subjects offered a full sentence
response to the stim uli, but rejected it for one reason or another (e.g.
"Charlie beaten Don, no, It has to be 'D on beaten Charlie.'"). Two patients
were occasionally unable to combine the stim ulus words into a sentence
after reading the stimuli, (e.g., "driven M aggie Frank. I d o n 't get anything
from that one." Subject 126.); these have been categorized as no response
errors.
6.2 Subjects' Analyses
6.2.1 Difficulties w ith the Task
A lthough NC subjects found this C onstrained Production task
trivially easy, A D subjects varied w idely in the range of the num ber and
types of difficulty they experienced w ith it. Further, even though the
majority of certain errors, such as lexical errors and m isused stim uli, were
limited to only a few subjects, all AD patients experienced som e degree of
difficulty w ith the task, and the severity of the difficulty varied in
proportion to b o th sem antic and WM deficits. For example, even though
the corpus of speech was about half as large in this constrained production
task com pared to the spontaneous speech sam ple (2350 w ords com pared to
4711 w ords), the num ber of closed class om issions jum ped from 23 to 56,
103
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
and fresh start dysfluency rates increased by 20%, from 69 to 84. In addition,
there w ere 56 m isused stim uli, 31 rejected responses and four "I don't
know" responses. The frequency of each difficulty type in the responses of
individual AD subjects is show n in Figure 6. Only tw o subjects (#120 and
#126) m ade "N o-R esponse" errors, but all subjects tended to exhibit the four
other types of difficulty, although each individual show ed a unique profile
of problems. Specifically, three subjects m ade prim arily lexical or
gram m atical errors (subjects #126, 141, and 132), tw o subjects misused m any
stim uli (subjects #131 and 146), and four produced a high proportion of
dysfluent responses (subjects #120, 106,138, and 147). The m ildest subject
w ith respect to both sem antic and WM im pairm ents, subject #136, made
one lexical error and one dysfluency.
As stated above, the sixteen healthy control subjects found this task
very easy, as show n by the raw error rates listed in A ppendix C. In fact they
m ade only 15 errors and 33 dysfluencies in 864 responses (54 responses for
each of 16 subjects). T here w ere also four m isused stim uli (all
m orphological substitutions), b u t no "Rejected" responses or "I don't know"
type responses, as described below. Using the sam e criteria for analysis as in
Chapter 5,2 none of these types of difficulty were frequent enough to qualify
for further analysis.
As illustrated in Figure 7 below, AD subjects displayed quite variable
perform ance on the CP task, w ith the num ber of difficulties ranging from
tw o to 44. C onsidering that subjects were relatively hom ogenous in their
2 To qualify for analysis, the frequency of errors had to be at least .20 per hundred words and at
least half of the subjects in the group had to have made that type of error.
104
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Insert Figure 6 here,
Page 105
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
MMSE scores (s.d. of MMSE = 1.4), disease severity cannot be the cause of
these problems. Consequently, a stepw ise regression w as conducted (F to
enter = 4) using the total num ber o f "problem s" on the CP, (sum of the total
num ber of raw: dysfluencies, rejected responses, lexical/gram m atical errors,
m isused stimuli, and "I d o n 't know " responses), as the dependent variable.
Factors included the three WM m easures (Backward-Recall, Nonsemantic-
WM, and Ordering) and two sem antic measures (W ithin-Category and No-
Response nam ing errors) w hich w ere used in the previous chapter in
sim ilar analyses. Backward-Recall was the first factor to enter the equation,
w ith the effect that the lower subjects scored on Backward-Recall, the more
overall problem s they had w ith the CP task. This variable accounted for
57% of the variance in total CP problem s, (p < .01). W ithin-Category
nam ing errors, a sem antic m easure, entered the regression equation next,
with the effect that as these nam ing errors increased, the num ber of
problem s on the CP task also increased. These two variables accounted for
84% of the total variance, (p <.002). Thus, im pairm ents in either semantics
or WM (or both) led to increased difficulty with this task.
In sum m ary, this constrained speech task affected the speech
production of individual AD subjects in very different ways; for instance,
some subjects showed increased num bers of lexical errors, or m isused
stim uli, while others had problem s form ulating sentences and,
consequently, produced m any fresh starts. However, the degree of difficulty
that a subject experienced w ith th e task overall was highly predictable using
a com bination of WM and sem antic im pairm ent m easures. The following
sections look at the four m ost prevalent types of difficulty found in this task
106
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
(dysfluencies, rejected responses, errors, and m isused stimuli) in ord er to
determ ine the etiology of each, and to ascertain the stim ulus conditions that
caused the m ost difficulty.
6.2.2 Dysfluency rates and rejected responses
This section discusses the distribution an d etiology of dysfluencies
and rejected responses in the constrained production task. Initially, it
compares the types of dysfluencies m ade in the CP to those m ade in the PN
by both groups. Subsequently, the etiology of the dysfluencies m ade by the
AD group are analyzed, along with their distribution across stim ulus
conditions. These analyses reveal that the frequency of AD subjects' fresh
start dysfluencies increased dram atically in this task com pared to baseline
rates in spontaneous speech, and this frequency w as highly related to the
individual's sem antic im pairm ent. Further, rejected responses proved to be
related to both W M and sem antic im pairm ent, w ith deficits in either ability
leading to an increased num ber of rejections. Both types of difficulty are
found largely in response to IRREG verb stim uli.
Recall that, in spontaneous speech, the tw o groups, AD and NC
subjects, did not differ w ith respect to num ber of dysfluencies, and both
groups m ade slightly m ore fresh starts than retracings (although this
difference only approached significance am ong AD patients). In order to
determ ine the differential effects of the CP task on dysfluency rates, a group
by task by dysfluency type (2x2x2) ANOVA w as com puted using as the
dependent variable the num ber of dysfluencies of each type (retracings and
fresh starts, as described in chapter 5) per hundred w ords. The results of this
analysis, show n below in Figure 7, revealed that the distribution of
107
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
dysfluencies in the CP task differed significantly from that in spontaneous
speech, as revealed by a highly significant three way interaction, (F(l, 96) =
16.740, p < .0001). This interaction reflected the fact that AD subjects made
m any m ore fresh starts in the CP than they had in spontaneous speech,
while NC subjects virtually did not m ake fresh starts in the CP task. In
contrast, the num ber of retracings by both groups rem ained stable across
tasks. All two way interactions also attained significance, d u e to the
strength of the three-way interaction. This pattern of results also
m anifested in m ain effects of both group, (F(l, 96) = 27.613, p < .0001) and
dysfluency type, (F(l, 96) = 20.037, p < .0001). However, the task effect was
only m arginally significant, (F(l, 96) = 2.175, p < .15), since the increase in
AD fresh start dysfluencies com pensated for the lack of dysfluencies of this
type in NC speech.
Figure 7. Frequency of two types of speech dysfluency in tw o tasks.
4.5
□ AD
□ NC
C/3
- § 2.5
C/3
1.5
Fresh Starts Retracings Fresh Starts Retracings
Constrained Speech Spontaneous Speech
108
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
These results revealed that the constrained production task
significantly disrupted the fluency of the AD patients, as show n in the
sam ple responses in A ppendix E. Moreover, increases in the frequency of
fresh starts in the C P could not be blam ed for difficulties w ith open class
lexical access, since all the open class w ords necessary for a grammatical
sentence w ere provided to subjects. In order to determ ine the etiology of
both types of dysfluency, tw o stepwise regressions were conducted (F to
enter = 3) using as dependent variables the num ber of fresh starts and
retracings m ade by AD subjects in the CP. The two semantic and three WM
measures used in previous analyses were included as factors. The results,
show n in Figure 8 below, revealed that the num ber of fresh starts increased
in proportion to the num ber of W ithin-Category nam ing errors, which
accounted for 62.8% of the variance in scores, (p < .007). The num ber of
retracings by AD patients in the CP task (which did not differ from the
num ber of sim ilar dysfluencies by NCs) was not predictable using either
semantic or WM factors, and thus may represent a baseline rate for these
dysfluencies. These findings suggest that the sam e type of im pairm ent that
led to W ithin C ategory nam ing errors also dim inished the ability of subjects
to im m ediately activate the semantics associated w ith the stimuli. This led
to a period of tim e in w hich word forms were available to subjects but
apparently not enough of the semantics of the stim uli to allow subjects to
combine them into a sentence. For example, subject #120 produced this
response to the stim ulus WOVEN CLOTH G a il, "um , woven Gail? Woven
cloth? I d o n 't know. Cloth, Gail. Cloth w as w oven by Gail." As can be seen
here, this subject repeatedly tried to combine these w ords in a meaningful
way and finally produced a completely gram m atical sentence.
109
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Figure 8. The num ber of AD subjects' fresh starts in the CP task w as highly
related to the n um ber of their W ithin-C ategory nam ing errors.
3.5
2.5
1.5
14
W ithin-Category N am ing Errors
In order to explore these findings further and to determ ine if any
particular aspects of the CP task stim uli w ere associated w ith increased
dysfluencies for AD patients, we analyzed the distribution of dysfluencies
across the three noun and verb conditions. This question w as explored with
a three by three (Verb Type x N oun O rder) ANOVA which used as the
dependent variable the num ber of dysfluencies m ade by AD subjects in each
condition of the C P task. Neither the m ain effect of noun ord er nor the
interaction betw een noun order and verb type were significant. How ever,
the main effect of verb type proved highly reliable, (F(2, 81) = 7.288, p < .002),
since there w ere m any more dysfluencies in response to IRREG verbs (mean
= 2.37 dysfluencies per hundred w ords) than to either of the other tw o verb
types (means: REG = .83, T-E = 1.00).
110
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
A kind of difficulty that w as closely related in form to dysfluencies
w as rejected responses in w hich subjects offered a response to a CP stim ulus
and then discarded it in favor of an alternate response, a type of response
not found in spontaneous speech. Examples of this type of difficulty, found
in A ppendix F, consist prim arily of responses in w hich either closed class
w ords w ere m issing or stim ulus w ords were altered or missing. In a
stepw ise regression exam ining the etiology of these difficulties, Backward-
Recall entered the equation first to account for 57% of the variance in scores,
followed by the n u m ber of No-Response nam ing errors, yielding a total r2 of
.73, (p < .01). In this analysis, low er Backward-Recall scores and higher
num bers of N o-Response nam ing errors predicted higher num bers of
rejected responses. T hus, better WM and m ore sem antic im pairm ent (as
m easured by N o-Response nam ing errors) led to better self-monitoring of
responses. The apparently anom alous contribution of No-Response errors
is discussed in C hapter 7. Due to the num ber of em pty cells, rejected
responses w ere not analyzed w ith an ANOVA; how ever, examination of
the data revealed that they w ere found prim arily in tw o conditions, the
IRREG verb condition (57% of rejected responses) an d the 2-Names noun
condition (49% of rejected responses).3
To recapitulate the findings presented in this section, the frequency of
dysfluencies in this constrained production task differed substantially from
the results of the analyses of dysfluencies in spontaneous speech. AD
patients m ade m any m ore fresh starts in the CP com pared to NC subjects,
3 I suspect the 2-Names condition was difficult because it offered subjects no clues as to which
person should be the agent of the verb, leading perhaps to rejected responses based on people
subjects had known with that name, (e.g. Greg, Greg stopped Bob, no that's not any good. Bob
stopped Greg.)
I l l
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
but did not differ from them in the num ber of fresh starts in the PN or in
the num ber of retracings they m ade on either task. Further, the frequency
of AD patients7 fresh starts on the constrained production tasks w as highly
predicted by their sem antic im pairm ent, w hile increased dysfluencies in
spontaneous speech w ere associated w ith lower WM scores. This is
particularly interesting since fluency of sentence production w as vulnerable
to semantic im pairm ent in the task in w hich w ords were provided to the
subjects, but n o t in spontaneous conversation, in which w ord choice was at
the discretion of the speaker. In contrast to these findings on the dysfluency
rates of AD subjects, and despite their surface similarity to dysfluencies, the
frequency of rejected responses proved to be related to both WM and
semantic im pairm ent, suggesting that deficits in either dom ain m ay affect
subjects' ability to m onitor their speech output on-line, leading to increases
in post-articulation editing. Furtherm ore, the IRREG verb condition was
marked by increases in both dysfluencies and rejected responses, while
rejected responses also occurred m ore frequently when noun stim uli were
both anim ate.
6.2.3 The D istribution and Etiology of Lexical Errors
This section identifies the types of lexical and gram m atical errors
which predom inated in the CP task, and determ ines their etiology, as well
as the stim ulus conditions associated w ith increased errors. Due to the
nature of the task, (e.g. open class w ords provided, no continuity betw een
responses, an d prespecification of the morphological form of the verb), the
expectation w as that open class errors, pronoun errors and m orphological
errors should be m inim ized. This prediction was proven to be true, but
112
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
only for open class and pronoun errors; the frequency of m orphosyntactic
errors did not change betw een tasks. On the other hand, since providing
closed class words was the prim ary requirem ent of the task, these types of
difficulties were predicted to increase; this prediction was also borne out.
The following analyses also show ed that the frequency of closed class errors
w as highly related to sem antic impairment. Finally, stim uli containing
IRREG verbs or Inanim ate-first noun orders contained increased errors.
To test the prediction that open class, m orphosyntactic and pronoun
errors w ould be m inim ized in the CP task, a three way, 2 x 3x2 (Task by
W ord Class by Group) ANOVA was com puted using the num ber of errors
per 100 w ords as the dependent variable. The results of this analysis appear
in Figure 9. As shown, rate and distribution of NC errors did not change
betw een tasks, and, for the m ost part, the perform ance of the AD patients in
the CP task was identical to that of the NC subjects. Only the performance of
the AD subjects in the PN differed, showing large increases in open class
and pronoun errors, a pattern which led to a significant interaction between
group and task, (F(l, 300) = 7.694, p < .006). This large increase in errors in
the PN by AD subjects also resulted in a significant m ain effect of task, (F(l,
300) = 5.3767, p < .01), and group, (F(l, 300) = 10.035, p < .002). However, the
m ain effect of error category w as only m arginally significant, (F(2, 300) =
2.179, p < .12), due to the sim ilar error distributions across w ord classes and
to excessive variance in scores. Post-hoc analyses revealed that, compared to
the PN, the CP task did m inim ize open class and pronoun errors in the
productions of AD subjects, (f(9) = 3.202, p < .01 and f(9) = 1.99, p < .08,
respectively), but not m orphosyntactic errors, (f < 0.2).
113
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
Figure 9. C om parison of AD and NC error frequencies betw een tasks.
Error Type
O M orphosyntactic
I~1 O pen Class
■ Pronoun
Lj YJ
CP PN
AD
NC
Since providing closed class items was one of the prim ary
requirem ents of the CP task, closed class errors have been analyzed
separately in a (2 x 2 x 2) group by task by error type (omission or
substitution) ANOVA. As can be seen in Figure 10 below, there was a
significant three w ay interaction as a result of the large increase in closed
class omissions by the AD group in the CP, (F(l, 96) = 4.930, p < .03). All
m ain effects and other interactions w ere also significant, b u t these were
attributable to the high num ber of closed class om issions by the AD group in
the CP. Exam ples of CP responses in which subjects om itted closed class
w ords are found in A ppendix G.
Post-hoc analyses revealed that the frequency of AD subjects' closed
class su b stitu tio n s did not differ by task, nor did the frequency of either type
of closed class errors by NC subjects. Surprisingly, the increase in AD
114
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
subjects' closed class om issions between tasks w as only m arginally
significant, (f(9) = 1.879, p = .09). Examination of the data revealed that this
failure to reach significance resulted from a high variance in error rates;
three of the AD subjects produced the majority of the closed class omission
errors, in som e cases producing alm ost agram m atic speech, (e.g. subject 141,
"Diane ( ) taken ( ) pill;" subject 126, "Lisa ( ) throw n () plate.")4 In fact,
subjects #126 and 141 om itted 40% of the closed class item s required by
sentence context, w hile subject #132 om itted 20%. Based on these figures, it
cannot be argued that these subjects did not understand the task since they
produced m ore than half of the closed class items required, an d all three of
these subjects did produce some completely gram m atical responses
interspersed w ith agram m atic ones.
Figure 10. AD subjects m ade a disproportionate num ber of closed class
om issions in the CP task.
4 :
3.5
7 3
I 2.5 -j
§
S *
O J
Q *
t n 1.5 1
M
C
i*
u 1 :
.5
0
s b i
CP PN
□ Omissions
Substitutions
a -L
CP PN
AD NC
4 As in Menn, Reilly, Hayashi, Kamio, Fujita, & Sasanuma (1998), word order was used to
determine the intended sentence structure of response in these cases, since it is almost
universally preserved in AD and in most aphasias.
115
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
The difference betw een the three "high om itters" and seven "low
om itters" w as explored extensively, b u t analyses were ham pered by the
small num ber of subjects; however, the tw o groups did not differ in age,
education, MMSE, sem antic im pairm ent, o r WM scores. Further, the tw o
AD groups did not differ from each other or from NC subjects w ith respect
to the num ber of o th e r errors made on the CP task, not including closed
class omissions. These findings suggested that these three AD subjects
m ight have a specific difficulty with closed class w ord production in general.
To test this hypothesis, I re-examined the spontaneous speech error patterns
w ith a G roup x Error Category ANOVA that used the num ber of errors per
100 w ords on the personal narrative as the dependent variable and treated
high om itters, low om itters and NC subjects as three separate groups. The
results of this analysis are shown in Figure 11. As illustrated, the prediction
of a specific closed class w ord im pairm ent w as not supported, as the high
om itters w ho m ade m ore errors of all types com pared to low om itters w ho
m ade an interm ediate num ber of errors in all categories, and the NC
subjects w ho m ade very few errors. Reflecting this, there was a significant
main effect of group, (F(2,196) = 20.798, p < .0001), and post-hoc Scheffe tests
verified that the total error rates of all three groups differed significantly
from each other. N either the effect of error category nor the interaction was
significant, reflecting the previous finding from Chapter 5, that errors w ere
distributed relatively evenly across all error categories. Post-hoc
comparisons revealed that the low om itter group only differed significantly
from the NC subjects in the frequency of overall errors and open class
errors. However, high om itters differed from NC subjects in all error types
but only significantly differed from low om itters in the frequency of their
116
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
pronoun errors. N evertheless, the pattern is clear— despite sim ilar levels of
semantic and WM im pairm ents, the three subjects w ho om itted many
closed class item s in the CP task also differed considerably from the rest of
the AD group in their accuracy of production in spontaneous speech. The
existence of this high om itter group is difficult to explain and is discussed
more fully in C hapter 7.
Figure 11. The relative error rates in the PN of NC subjects and AD subjects
who were high or low om itters of closed class w ords in the CP task.
G roup
□ High Omitters (AD)
Low Omitters (AD)
Normal Controls
m
Closed Class M orphosyntactic O pen Class Pronouns
Error Categories
Returning to the discussion of the CP task, the next question
addresses the etiology of errors m ade by the AD subjects. Two particular
types of errors attracted attention in previous analyses of this task, the large
num ber of closed class om issions and the frequency of m orphosyntactic
errors which did not decrease in the CP as predicted; these two error types
were exam ined using stepw ise regression analyses (F = 3.0 to enter). As in
previous analyses of this type, the factors used included the two semantic
117
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
m easures (W ithin-Category and N o-Response nam ing errors) and the three
WM m easures (Backward-Recall, N onsem antic-W M , and O rdering) w hich
w ere described in C hapter 4.
Figure 12. Closed class om issions as a function of semantic im pairm ent in
the high and low om itter groups.
( A
" O
£
o
o
C A
C
o
< A
( A
( 0
u
T3
0 1
t A
o
U
10
9
O H igh Om itters
♦ Low Om itters
8
7
6
5
4
3
2
1
0
12 14 0 2 6 8 10 18 4 16
W ithin-Category nam ing errors
CP omitted closed class = -.2852 + .0648 *W/in-Cat; RA 2 = .778 (Low omitters)
CP omitted closed class = -1.5508 + .7538*W/in-Cat, RA 2 = .775 (High omitters)
Considering all ten AD subjects together, the frequency of om itted
closed class w ords w as predicted by W ithin-Category naming errors, (r2 = .46,
p < .02), w ith the effect that subjects w ho m ade m ore W ithin-Category
nam ing errors also om itted m ore closed class w ords in the CP. H ow ever, a
better analysis resulted from treating separately the two groups identified
above (high versus low om itters), as show n in Figure 12. A lthough there
w ere too few subjects for these analyses to reach significance, W ithin-
Category nam ing errors w as a good predicter of the num ber of closed class
omissions in both groups. Further, it is obvious from this graph that closed
118
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
class production in the high om itter group was m uch m ore strongly affected
by individual differences in semantic ability, as m easured by W ithin-
Category nam ing errors, than it w as in the low om itter group.
In exam ining the data on m orphosyntactic errors, it became evident
that all of AD subjects' errors of this type w ere syntactic rather than
inflectional. Indeed, the majority of these errors involved the use of
incorrect subcategorization frames for T-E verbs, for example, responding
"Robert terrified the fire," (Subject #146) w hen the stim uli consisted of
TERRIFIED ROBERT FIRE. The stepwise regression analyzing syntactic errors in
the CP revealed three factors which contributed to the frequency of this
error type.5 W ithin-C ategory nam ing errors entered first to account for 59%
of the variance in scores, followed by No-Response errors and Ordering,
which each accounted for an additional 17% of the variance, for a total r2 of
93%, (p < .001). The effect of W ithin-Category errors w as the same as in
m ost of the other analyses, such that increased nam ing errors predicted
increased syntactic errors. The effect of O rdering w as also as predicted-
lower WM scores w ere associated w ith more errors in syntax. No-Response
nam ing errors had the opposite effect here— subjects w ho m ade more No-
Response nam ing errors m ade fewer syntactic errors in the CP task.
The analyses in this section have show n th at closed class om issions
and syntactic errors on the CP task w ere associated, in different degrees, w ith
sem antic im pairm ent, such that as W ithin-Category nam ing errors
increased, so d id errors in the constrained production task. W hat these data
do not reveal is w hether it was just the nature of the task itself that caused
5 "High omitters" and "low omitters" did not differ in the frequency of syntactic errors here.
119
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
difficulties (in w hich case, errors w ould be spread random ly across stim ulus
conditions) or w hether som e aspects of the stim uli w ere especially difficult
for the AD subjects (in w hich case the errors w ould be concentrated in
specific stim ulus conditions.) The following analysis exam ines the
distribution of AD subjects' errors w ith respect to variations in the CP
stim uli.
To determ ine w hether AD subjects' errors w ere random ly spread
across conditions or concentrated in a few conditions, a three by three (Verb
Type x N oun O rder) ANOVA was com puted using the num ber of errors per
condition m ade by AD patients as the dependent variable. The results of
this analysis are show n in Figure 13 below. This analysis revealed a strong
main effect of noun order, (F(2, 81) = 5.219, p < .01), w ith the highest num ber
of errors occurring in the Inanim ate-first noun condition (mean = 4.20
errors per condition p er subject), fewer in the N am e-first condition (mean =
2.31), and the fewest errors in the 2-Names condition (m ean = 1.11). Post-
hoc Scheffe analyses revealed that the error frequency in the Inanim ate-first
condition differed significantly from both the 2-Name (p < .002) and the
Name-First conditions (p = .053), but the latter two conditions did not differ
from each other (p > .25). Furthermore, the m ain effect of verb type was
marginally significant, (F(2, 81) = 2.613, p < .08), due to the fact that there
were m ore errors in the IRREG verb conditions in all noun order
conditions, though it only differed significantly from the REG verb
condition. In addition, the num ber of errors in the T-E verb condition did
not differ significantly from those in the REG condition w hen paired w ith 2-
Name and N am e-first stim uli, but increased w hen the noun stimuli
appeared in the Inanim ate-first order. The interaction betw een verb type
120
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
and noun order w as insignificant. In sum m ary, for AD subjects the prim ary
stim ulus m anipulation w hich led to increased errors was the order of noun
stim uli, w ith m ore errors occurring in stim uli w ith the inanim ate no u n
first than in the other tw o noun conditions. Further, IRREG verbs w ere
also associated w ith consistent, but m arginal increases in error rates across
all noun conditions.
Figure 13. D istribution of AD subjects' errors across CP stim ulus conditions.
c
,o
2.5
c
o
O
2
o j ^
O -
2
o 1.5
fc*
U
V
□ REG verbs
H T-E verbs
■ IRREG verbs
V
X I
6
3
2
2-Names Name-
first
Inanimate-
first
Noun O rder
Inanim ate-first stim uli m ay have been m ore difficult for at least three
reasons: 1) they encouraged subjects to produce a sentence using the
inanim ate noun first, a w ord order which is dispreferred in English
(McDonald et al., 1993), 2) if the noun stim uli were used in the given order,
a determ iner had to be supplied imm ediately at the beginning of the
response, or 3) reordering Inanim ate-first stim uli m ay consum e m ore
processing resources, leading to more errors. Examination of the data
revealed that errors w ere not limited to the om ission of sentence-initial
121
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
determ iners, although these errors did constitute 30% of all closed class
om issions. Further, this type of error never com prised m ore than 45% of an
individual subjects' total closed class omissions. In actuality, the largest
group of om issions consisted of m issing auxiliary verbs (42%).6 The
rem aining om issions consisted of m issing prepositions (8%) and m issing
sentence-internal determ iners (20%). Thus, although m issing sentence-
initial determ iners com prised a large proportion of closed class omissions,
they w ere not the prim ary source of these errors. To explore the possibility
that reordering noun stim uli caused m ore errors in AD subjects' responses,
the frequencies w ith w hich AD and NC subjects reordered Inanim ate-first
stimuli w ere com pared. In fact, the AD group reordered about the same
proportion of noun stim uli in the Inanim ate-first condition as NC subjects
did, (AD: .678, NC: .750; f(25) < 1.0). Subsequently, a correlation was
com puted betw een the proportion of Inanim ate-first n o u n pairs reordered
and the num ber of errors m ade in this condition to determ ine w hether
subjects w ho reordered m ore Inanim ate-first noun pairs m ade m ore errors
due to increased use of lim ited resources.7 This analysis revealed that these
two variables were m arginally related (r = -.339, p = .067); how ever, the
direction of the correlation indicated that increased errors w ere associated
w ith a decrease in noun reordering. In other w ords, AD subjects who did
n o t consistently reorder the Inanim ate-first stim uli tended to m ake more
errors in this condition. This difficulty in creating gram m atical sentences
w hen the stim uli encouraged a less frequent, inanim ate-first w ord order
was lim ited to the AD group. Therefore, it appears that the accuracy of
6 This a rather bizarre finding because more errors occurred in the Inanimate-first condition,
rather than in IRREG verb constructions which almost always required an auxiliary.
Correlations between error rates and proportion of noun-pair reorderings in all 3 noun conditions
were computed. Only those in the Inanimate-first condition were significant.
122
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
production of AD subjects suffered w hen trying to use the noun stim uli in
the non-canonical, dispreferred order, and this effect w as exacerbated w hen
the verb form was also noncanonical (i.e., the IRREG verb condition).
This section has reported several im portant findings with respect to
the errors m ade by AD subjects in the CP task. First, open class and pronoun
errors w ere m inim ized by providing subjects w ith open class w ords to use.
However, com pared to error rates in spontaneous speech, the frequency of
m orphosyntactic errors w as not affected by the reduction in discourse
dem ands, and the rate of closed class om issions increased dramatically.
Both total errors on this task and closed class om issions proved to be
predictable from the num ber of subjects' W ithin-Category nam ing errors,
although the m ajority of these errors was attributable to three AD subjects
w ho om itted 20 to 40% of the necessary closed class words. However, the
frequency of open class, pronoun and m orphosyntactic errors in the CP did
not differ betw een these two AD groups. These three "high om itters" also
proved to have m ade m ore errors of all types in spontaneous speech than
those AD subjects w ho om itted a norm al num ber of closed class w ords in
the CP task. Exam ination of syntactic error from the CP task revealed that
there were no inflectional errors by AD patients, but instead, errors of
argum ent structure prevailed, especially in T-E verb conditions. These
errors were also highly predictable from sem antic m easures, w ith the WM
m easure, O rdering, also contributing to the variance in scores. The highest
num bers of errors occurred in Inanim ate-first noun conditions, and subjects
m ade more errors here w hen they did n o t reorder these stimuli into the
canonical anim ate-first w ord order. IRREG verb conditions also contained
123
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
consistently m ore errors than other verb conditions, but this effect w as only
m arginally significant.
6.2.4 The D istribution and Etiology of M isused Stimuli
Besides gram m atical and lexical errors and dysfluencies, som e AD
subjects8 had another type of difficulty w ith the CP task-using all three
stim ulus w ords as written. These errors included instances in w hich one or
more of the three stim uli were om itted from the response or w ere
obviously substituted.9 Examples of responses w ith misused stim uli are
found in A ppendix H. Thus, these errors are distinct from and do not
overlap w ith the gram m atical and lexical errors analyzed previously. It is
also im portant to point out that, just as three subjects m ade the m ajority of
closed class om ission errors, about 75% of m isused stimuli w ere attributable
to two subjects, although eight of the ten AD subjects did make this type of
error. This section examines the ways in w hich subjects m isused CP
stimuli, the conditions in which stim uli w ere m isused, and w hether
increased incidence of misused stim uli resulted from WM or sem antic
im pairm ent, or both. To preview the findings below, these analyses
revealed that the m ost common ways of m isusing stimuli were to om it
them or to change the morphological form of an irregular verb. In addition,
the frequency w ith which these processes occurred was highly related to
WM im pairm ent, and m ost m isused stim uli occurred in response to
IRREG verb stim uli.
8 There were only four Stimulus Misuse errors by the NC group, three of which were by the same
subject, #331. All four consisted of morphological substitutions of the simple past form for an
irregular past participle.
124
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
Exam ination of the data identified four different w ays in which
subjects m isused stim uli: 1) Visual errors, such as using "crow n" for
GROWN, 2) Semantic Substitutions of the stim uli, e.g. saying "Pam
raised roses," in response to the stim uli GROWN PAM ROSES, 3)
O m ission of one or m ore of the three stim ulus w ords from the
response offered, and 4) M orphological Substitution of the sim ple past
tense form for the past participle form, i.e. "grew " for GROWN or "beat"
for BEATEN, a type of error only possible w ith IRREG verb stimuli.
To distinguish betw een sem antic substitutions and omissions, errors
w ere only classified as substitutions if a response included a w ord that was
sem antically related to one of the stim ulus w ords w hich was m issing from
the response. As an exam ple of a semantic substitution, subject # 131
responded, "Jeff w as totally baffled by the T.V. m onitor." in response to
BAFFLED MACHINE JEFF, which w as interpreted as the substitution of "T.V.
m onitor" for "m achine." A n om ission was counted w hen one of the
stim uli, or a close sem antic associate, was not included in the utterance
produced (e.g., from subject # 131, "A rthur was shaken w hen he fell down
on the basketball court" in response to SHAKEN ARTHUR JOAN.) Figure 14
show s the relative distribution of these four error types. As illustrated,
visual and sem antic substitution errors com prised 11 and 12%, respectively,
of the m isused stim uli in the corpus. The other two error types, omissions
and m orphological substitutions, each com prised 38% of total stimulus
m isuse errors.
125
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Figure 14. The relative frequency of the four types of stim ulus m isuse
errors in the CP by AD subjects.
■ Visual Substitutions
Semantic Substitutions
■ Omitted Stimuli
□ Morphological Substitutions
To investigate w hether m isused stim uli resulted from sem antic
im pairm ents, WM deficits or both, a stepw ise regression was conducted (F
to enter = 4) using the total num ber of m isused stim uli as the dependent
variable. The three W M and two sem antic m easures that have been
em ployed in all other analyses of AD difficulties w ere used as factors.1 0 The
first variable that entered the equation w as the WM measure Backward-
Recall, w hich accounted for 85.6% of the variance in scores. The only other
variable to enter the regression equation w as the num ber of W ithin-
Category nam ing errors, r2 = .934, (F(2,7) = 49.167, p < .0001).1 1 H ow ever, the
effect of the sem antic variable was anom alous— the m ore W ithin-Category
errors subjects m ade, the fewer stimuli they m isused. This counter
intuitive finding w as m ost likely attributable to the fact that, of the tw o
subjects w ho had the m ost difficulty correctly using all three stim ulus
'“ These factors consisted of the WM scores on Nonsemantic-WM, Backward-Recall, and
Ordering; and the semantic measures Within-Category and No-Response naming errors.
“Separate analyses of omitted stimuli and substituted stimuli showed similar patterns; both
were primarily predicted by Backward-Recall with r2 values of .80-.88.
126
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
w ords, the one w ith more preserved sem antics m ade m ore stim ulus misuse
errors. H ow ever, this subject also consistently attem pted responses that
were m uch m ore complex both syntactically and sem antically than those of
other subjects, a pattern which likely had a deleterious effect on his
perform ance (see Appendix E, subject #131). Thus, the num ber of misused
stim uli w as highly associated w ith the severity of attentional or WM
deficits, b u t sem antic im pairm ent did not increase the rate of these errors.
In order to determ ine w hether m isused stimuli clustered w ithin a
particular stim ulus condition or occurred random ly across all stim uli, the
num ber of stim ulus-m isuse errors w as used as the dependent variable in a
three by three (Verb Type x N oun Order) ANOVA, as has been done with
both errors and speech dysfluencies. In this analysis, the only significant
finding w as a m ain effect of verb type, (F(2,81) = 4.222, p < .02), as a result of
large num bers of m isused stim uli occurring in response to IRREG verb
stimuli (m ean = 3.51 per condition per person), and significantly fewer in
the T-E verb (1.41) and REG (.69) verb conditions.
In this section, misused stim uli were found to consist m ainly of two
types, substituted morphological form s and omissions of one of the
stim ulus w ords. There were also sm all num bers of sem antic substitutions
of stim uli and visual errors. The frequency of stimulus m isuse errors was
very accurately predicted by WM im pairm ent, as m easured by Backward-
Recall scores. Furtherm ore, m ost m isused stimuli occurred in the IRREG
verb conditions.
127
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
6.2.5 The Syntactic Com plexity of Responses
The IRREG verb stim uli were associated w ith all four types of
difficulty that subjects had w ith this task, (i.e., lexical errors, dysfluencies,
rejected responses, and m isused stimuli). H ow ever, it is not clear just w hat
characteristic of these irregular past participles m ade them so difficult. One
possibility was that IRREG verbs caused m ore difficulty for AD patients
because their form prohibited the use of a canonical, sim ple active sentence
as a response, but instead forced the use of a complex tense (e.g., a passive or
a perfect), whereas the other two verb types allow ed the use of the canonical
syntactic form. If com plexity of response w ere the reason behind the
difficulties w ith this task, then only the difficult IRREG verb condition
w ould show large num bers of passives or other complex tenses. This was
tested here by com paring the frequency of passive production across
stim ulus conditions and betw een groups, in order to establish a "norm al"
baseline frequency of passive use. It was im portant to also include the effect
of noun order in this analysis, since verb type an d nouns order were
predicted to influence the structure of the sentence produced. For example,
due to differences in argum ent structure, Inanim ate-first noun stim uli
paired with T-E verbs w ould be expected to encourage an active usage (e.g.
SURPRISED GIFT MARY: "The gift surprised M ary"), whereas the sam e noun
stimuli with an IRREG or REG verb would be expected to lead to a passive
(i.e. CHOSEN GIFT MARY: "The gift was chosen b y Mary.").
To explore w hether complexity of response was limited to the IRREG
verb condition, the sentence structure of subjects7 responses were com pared
using a two by three by three (Group x Verb Type x N oun Order) ANOVA
128
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
w ith the proportion of passives produced in each condition as the
dependent variable. The main effect of group was insignificant, (AD = 26%
passives, NC = 33%; F(l, 216) = .398, p > .5), show ing that the two groups of
subjects did not differ in the overall frequency w ith which they produced
passive sentences. As predicted, the m ain effect of noun order w as also
insignificant (F(2, 216) = .639, p > .5), show ing that none of the three noun
orders consistently resulted in significantly higher proportions of passives.
The m ain effect of verb type, however, w as highly significant, (F(2, 216) =
68.786, p < .0001), because neither group produced a significant num ber of
passives using REG verbs (mean = .056), (thus replicating findings in
Ferreira (1994)), and also because the proportion of passives produced by the
com bined groups in the T-E verb condition w as higher than in the IRREG
verb condition (M eans = .57 versus .34, respectively). Interestingly, the
frequency of passive production w ith T-E verb stimuli in the current study
considerably exceeds the passive production rates of between 19 and 35
percent for T-E verbs reported by Ferreira (1994), a difference that m ay be due
to the different spatial arrangem ent of the stim uli. Ferreira presented
stim uli in a linear arrangem ent w ith n o u n stim uli first and so m ay have
triggered w ell-entrenched production strategies of using the first activated
noun as the subject. The current task attem pted to minimize the effects of
stim ulus order by using a non-linear w ord arrangem ent (see Figure 6 on
page 94) and placing the verb in the left-m ost position, thereby giving visual
and probably tem poral precedence to characteristics of the verb.
In addition to the verb type m ain effect, there was also a significant
group by verb type interaction, as illustrated in Figure 15, because the two
groups varied in their patterns of passive production w hen using T-E and
129
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
IRREG verbs, especially in the m ixed anim acy conditions. As pictured, AD
subjects show ed a distinct flattening of the curves show n by the NC group.
In fact, post-hoc t-tests collapsing across all noun conditions revealed no
difference in the proportion of passives produced by AD subjects when
using T-E verbs and IRREG verbs, (Means: .468 and .402, respectively; t(58) =
.4125, p > .40). In contrast, NC subjects exhibited a substantial difference in
the proportion of passives produced using these two verb types (Means: T-E
= .625, IRREG = .295; f(82) = 4.384, p < .0001). The only other significant
interaction in this analysis w as betw een verb type and noun order, (F(4, 216)
= 3.485, p < .01), which is a result of the total lack of noun order effect with
REG verb stim uli and the opposing effects that the tw o mixed animacy
conditions had on the production of passives in T-E and IRREG verb
conditions.
Figure 15. The proportions of passives produced by both groups divided by
verb type and noun order.
.9
Verb type
0
2-Name Name- Inanimate- 2-Name Name- Inanimate-
first first first first first first
AD NC
130
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
In addition, a separate analysis revealed that, com pared to N C
subjects, AD subjects produced m ore active sentences throughout the CP
task, (AD = 59%, NC = 44%;f(24) = 2.087, p < .05), but produced about the
sam e proportion of passives, as show n above. Thus, AD speakers produced
approxim ately 85% of their responses in either the passive or the active
voice, w hile NC subjects produced only 77% of their responses using these
two structures, a difference that approaches significance (f(24) = 1.855, p<
.08). The rem aining 15% of AD responses and 23% of NC responses
consisted largely of responding to IRREG verbs w ith the perfect aspect or a
deverbal adjective. These findings reveal, that although AD patients relied
m ore on the canonical English active structure than NC subjects, they had
no difficulty w ith using passive structures but m ay have been less able to
utilize alternate structures.
These results show that increased sentence complexity was not
lim ited to the IRREG verb conditions but w as also found in responses to T-E
verbs in the responses of both groups, w hich did not show significant
increases in errors, dysfluencies or m isused stim uli in previous analyses.
C onsequently, increased response com plexity could not have been a
proxim al cause of the various difficulties that AD subjects had in the IRREG
verb condition. Furthermore, healthy elderly speakers dem onstrated clear
effects of verb type in determining the proportion of passives that w ould be
produced in m ixed animacy conditions, producing a high proportion of
passives in response to T-E verb stim uli, few er in response to IRREG verbs,
and alm ost none in response to REG verbs. However, compared to the
norm al elderly, the AD group used passives less frequently in response to T-
E verbs and m o re frequently in response to IRREG verbs. In fact, AD
131
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
patients only responded using a norm al pattern of passive (non)production
w ith REG verb stimuli, to w hich all subjects generally responded w ith a
sim ple active sentence. In addition, AD patients used m ore active sentences
overall and fewer alternate structures besides active and passive sentences.
Based on these findings, it seem s that these AD patients had a dim inished
aw areness of the custom ary sentence forms associated w ith verbs that differ
from canonical, regularly inflected, agent-patient verbs, especially verbs
w ith different argum ent structures (T-E verbs) and verbs w ith irregular
m orphological forms (IRREG verbs).
6.2.6 Sum m ary of Subjects' A nalyses
The Constrained Production task required subjects to produce a
sentence using a verb and two nouns w hich were provided by the
experim enter on a card. Stim uli w ere chosen specifically to elicit passive or
perfect sentences, and the p rim ary dem ands of the task involved supplying
the sentence form and the closed class w ords necessary for a grammatical
sentence. Analysis of AD subjects' perform ance on the task revealed that
the task was quite difficult for m any of the AD subjects, resulting in large
num bers of fresh start dysfluencies, rejected responses, closed class
om issions, syntactic errors, a n d m isused stimuli. Further, the level of
difficulty experienced by a given subject on the task proved to be related to
his or her WM and sem antic im pairm ents, such that both im pairm ents
resulted in increased difficulty w ith the task. The specific results relating to
these phenom ena are enum erated below.
1) The AD group m ade m any m ore fresh start dysfluencies in the
constrained production task th a n they had in spontaneous speech, while
132
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
NC subjects did n o t produce this type of dysfluency in this task. Thus,
constraining the speech of AD subjects to include particular content w ords
significantly im paired their fluency.
2) The frequency of AD subjects' fresh sta rt dysfluencies in the CP
proved to be significantly related to the degree of their sem antic
im pairm ent. T hus, although subjects w ere able to read and repeat the
w ords of the stim uli, they could not alw ays im m ediately combine these
w ords into sentences, and the frequency w ith w hich this occurred was
proportional to th eir sem antic im pairm ent.
3) AD patients m ade a disproportionate num ber of closed class errors
in the CP task, an d these errors were predictable from the semantic
im pairm ent m easure, W ithin-Category nam ing errors. Further, a subset of
AD patients show ed m uch stronger effects of sem antic im pairm ent and
om itted m any m ore closed class w ords than the m ajority of the group.
These sam e patients also show ed a sim ilar p attern of errors in spontaneous
speech in that they consistently m ade m ore errors of all types than the
majority of AD subjects. Based on these findings, there appear to be two
groups of AD patients in which sim ilar levels of sem antic im pairm ent have
quite different effects on the production of certain types of words,
particularly closed class items. This finding is addressed m ore fully in
Chapter 7.
4) The syntactic errors of AD patients in the CP task were highly
related to sem antic im pairm ent, and were alm ost entirely accounted for
using a com bination of WM and semantic m easures. These errors occurred
alm ost exclusively in response to T-E verbs, w hich have a noncanonical
133
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
argum ent structure. This suggests that sem antic im pairm ent m ay dim inish
a person's ability to use his or her know ledge about the syntactic
requirem ents of specific verbs, especially those with noncanonical argum ent
structures.
5) G ram m atical and lexical errors w ere not random ly distributed
across all conditions in the CP, b u t w ere found prim arily in responses to
those stim uli th at positioned the inanim ate noun in the prom inent, top
position on the stim ulus card, the Inanim ate-first condition.
6) Both NC and AD subjects reordered a majority of Inanim ate-first
stimuli, but the less frequently AD subjects reordered these, the m ore errors
they m ade in this condition. Therefore, lexical errors in the C P occurred
largely in the Inanim ate-first noun order condition which encouraged a
non-canonical w ord order for English speakers, and subjects w ho tried to
use the stim uli in the non-canonical w ord order were likely to m ake more
errors in this condition.
7) The m ost comm on ways that subjects m isused CP stim uli
consisted of om itting one of the three w ords from a response o r changing
the m orphological form of an ERREG past participle to the sim ple past.
8) The frequency of misused CP stim uli was predicted by WM
im pairm ent as m easured by Backward-Recall.
9) The IRREG verb condition w as consistently the m ost difficult for
AD subjects, in that responses to this verb type were associated w ith
increases in all of the types of difficulties exam ined here: dysfluencies,
rejected responses, lexical errors, and m isused stimuli. In addition, two of
134
R ep ro d u ced with p erm issio n o f th e cop yright ow n er. Further reproduction prohibited w ithout p erm issio n .
the three instances in w hich subjects were unable to provide a response to
the stim uli occurred in response to IRREG verb stim uli. Therefore, despite
the fact that different subjects experienced different types of difficulty with
this task, the IRREG verb condition caused difficulties for all of them.1 2
10) Difficulties w ith ERREG verbs cannot be attributed to differences
in the syntactic com plexity of the responses given, since both groups
produced at least as m any passive structures w ith T-E verbs as with IRREG
verbs, while the num ber of difficulties in response to T-E verb stimuli was
m uch less than that found in response to IRREG verbs.
These last tw o findings are problematic in th at it is not obvious w hat
it w as about the IRREG verb stim uli that caused AD subjects' difficulty. The
difference cannot be argum ent structure, since it w as the T-E verbs that
differed from the other tw o verb types in that respect, and for the most part,
the num ber of difficulties associated with REG and T-E verbs were similar.
Further, the difference cannot be in the presence of m ore complicated
syntactic structures in responses to IRREG verbs, since AD subjects produced
sim ilar num bers of passives w ith T-E and IRREG verbs, but only m ade large
num bers of errors, dysfluencies and m isused stim uli in response to IRREG
verbs. The m ost obvious w ay that IRREG verbs differed from the other two
verbs types was that they used separate, phonologically distinct forms for
past participles and sim ple past tense. However, there m ay be other
differences betw een verb types which may also have contributed to their
difficulty. For exam ple, both frequency and num ber of m eanings have been
1 2 This was confirmed with a Verb Type x Noun Order ANOVA using the total difficulties per
condition as the dependent variable, in which only a main effect of Verb Type was found
(F(2,81) = 12.660, p < .0001).
135
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
found to affect lexical accessibility, and as such may have increased the
difficulty of a specific verb type. These issues are addressed below in a series
of item analyses.
6.3 Item Analyses on CP Stimuli
This section of the dissertation explores the possibility, presented
above, that the frequency and num ber of m eanings of the verb stim uli may
have increased the difficulty of the task for AD subjects, apart from
canonidty. Frequency has been established as a variable that has w idespread
effects on linguistic processing. For exam ple, lower frequency has been
assodated experim entally with slower access times (e.g. G ordon, 1982,1983);
therefore, creating sentences with less frequent verbs m ight take m ore time,
leading especially to increased num bers of dysfluencies. O n the other hand,
h igh lexical frequency m ay be assodated w ith a faster than average
activation decay rate, which would m ake higher frequency item s m ore
difficult to m aintain in m em ory (H irschm an & Palij, 1992), leading perhaps
to m isused stim uli. Alternatively, the theoretical im portance of having
m ore m eanings stem s from the hypothesis that, when a w ord has m any
different usages, the set of semantic features shared by all m eanings will be
sm aller (Martin et al., 1996). Based on this hypothesis, seeing o r hearing a
polysem ous w ord o u t of context will activate a less com plete sem antic
representation than hearing a w ord w ith only few m eanings w hich all share
a majority of features. Thus, polysem y could conceivably m ake it m ore
difficult to fully activate one specific m eaning of a word. In o rd er to explore
these possibilities a series of item analyses w ere com puted to explore the
1 3 6
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
relative contribution of frequency, num ber of m eanings1 3 and canonicity to
the error, dysfluency an d m isused stim ulus rates of each verb type used in
this study.
Canonicity, w ith respect to the verbs used in the current study, could
be defined in either m orphological or syntactic term s. A morphological
definition w ould be based sim ply on w hether the sim ple past tense of the
verb and the past participle form were identical. This m ethod would yield a
sim ple binary variable th at in effect isolated IRREG verbs from the other
tw o verb types. A lternatively, a syntactic definition of canonicity could be
calculated based on, the frequency w ith which healthy, adult subjects used a
verb form in a sim ple active sentence, which is w idely accepted as the
canonical English sentence form. The advantage of this approach w ould be
th at individual differences in verbs within categories m ight be captured,
leading to m ore of a continuum of scores. Both types of canonicity were
calculated. Syntactic canonicity was calculated by determ ining the
proportion of norm al control subjects w ho responded to each verb with a
gram m atical, sim ple active sentence. Using this m ethod, the average
canonicity of IRREG verbs w as 0 since this verb type could not be
gram m atically used in a sim ple active sentence; the m ean canonicity of T-E
verbs was .34 reflecting NC subjects' preference for using passive structures
w ith these verbs; and the canonicity score of REG verbs w as .92 show ing the
alm ost universal preference for using these verbs in an active construction.
M orphological canonicity w as captured by assigning all REG and T-E verbs
the value "1" and IRREG verbs "0." In fact, the two types of canonicity
uNumber of meanings was established by a count of the number of different senses listed for the
verb and its various morphological forms (including idioms) in Webster's Nexu World
Thesaurus, New York: Prentice -Hall, 1985.
137
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
determ ined in this w ay were highly correlated, (r 2 = .56, p<.0001).
Consequently, the m ean of the two canonicity scores was used in the
follow ing analyses.1 4 Only the canonicity scores of the T-E verbs w ere
appreciably affected by this m ethod, since their morphological canonicity
was assigned to be "1," but their m ean syntactic canonicity score was only
.34. Thus, this m ethod categorizes T-E verbs as interm ediate in relative
canonicity betw een IRREG and REG verbs.
In the follow ing stepwise regression analyses, the variance associated
w ith frequency and num ber of m eanings w as partialled out of the following
analyses first before the canonicity variable was allowed to enter, because
these variables m ight be confounded w ith canonicity. In the regression
analysis using AD subjects' errors per item as the dependent variable,
frequency and num ber of meanings accounted for 7.5% (p < .07) of the
variance across all 54 verbs, with 5% being contributed by num ber of
m eanings and only 2.5% attributed to frequency differences. Canonicity
contributed an additional 20% for a total r2 of .276, (p < .001). In the analysis
of the num ber of dysfluencies per verb, the num ber of m eanings accounted
for 4.2% of the variance, and frequency only accounted for an additional
.3%. C anonicity independently contributed 17.1% to the variance, yielding a
total r2 of .216, (p < .01). In both of these analyses, frequency had little effect;
increased num bers of meanings w ere m arginally associated w ith higher
error rates, and increasingly lower canonicity scores consistently predicted
m ore difficulties. Considering that both errors and dysfluencies were
predicted by sem antic impairm ent in the subject analyses, certain
1 4 Using either morphological or syntactic canonicity leads to the same results: less canonical
stimuli are significantly associated with more errors, dysfluencies and misused stimuli.
138
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
generalizations can be inferred regarding the relative difficulty of
constrained production stim ulus items for subjects w ith semantic
im pairm ent: 1) creating sentences using less canonical past participles was
very difficult, 2) using verbs w ith more m eanings caused slight difficulties,
and 3) using verbs of differing frequencies had very little effect on errors or
dysfluency rates per item.
The findings w ith respect to the num ber of m isused stim uli per verb
item differed in several w ays from those above, w hich is not surprising
since these difficulties w ere associated prim arily w ith WM deficits rather
than sem antic im pairm ents in the subject analysis. In the regression
analysis in which the dependent variable was the num ber of misused
stim uli per verb, frequency played a significantly larger role, (r2 = .121, p<
.01), and the num ber of m eanings also contributed significantly, (r2= .06,
p<.05), in predicting w hich verbs would be associated w ith more misused
stimuli. Even so, canonicity again contributed the m ajor portion of the
variance, 21.3%, for a total r2 in this analysis of .394, (p < .0001), again with
the effect that subjects had m ore difficulty w ith non-canonical stimuli. In
contrast to the pattern show n in the above analyses, increased numbers of
m isused stim uli w ere associated w ith higher frequency verbs and fewer
m eanings.
These item analyses w ere based on the assum ptions that the overall
canonicity of a particular verb was a combination of the relative frequency
w ith w hich it was used in canonical, simple active sentence structures and
w hether it had a regular past participle form. In addition, it was assumed
that canonicity m ight be a relative value, that is, that som e T-E verbs, for
139
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
exam ple, m ight be used in active sentences m ore often than others, despite
sharing the sam e argum ent structure. These analyses have show n that,
conceived in this w ay, the canonicity of verbs significantly predicted how
difficult stim uli using that verb would be for AD patients, w hether their
prim ary deficit w as in semantics or WM. Further, they have show n that, in
determ ining w hich stim uli would be difficult for AD patients, the im pact of
frequency and n u m b er of m eanings of the verb stim ulus was m uch less
im portant than its relative canonicity.
140
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
C hapter 7.
C onclusion
This dissertation evaluates the efficacy of m o d u lar and interactive
theories of speech production for predicting the patterns of speech errors
and dysfluencies in the speech of subjects with A lzheim er's Disease (AD).
In its early stages, AD im pairs both semantic and m em ory abilities but does
not obviously dim inish syntactic competence. As a result, these patients
have previously been used as support for theories advocating the
m odularity of syntactic and sem antic processing. This stu d y directly
addressed this question by exam ining the speech of AD patients for errors in
dom ains usually considered syntactic. My results provide no support for
the m odular theory. AD patients exhibited significant im pairm ents
com pared to healthy elderly in several aspects of speech production, such as
sentence construction ability and the use of closed class w ords.
Furtherm ore, there w ere stro n g correlational relationships betw een
subjects' sem antic im pairm ent and the frequency w ith w hich they made
both sem antic and syntactic errors. In fact, sem antic im pairm ent predicted
these latter errors as well as it did the semantic ones.
Besides m odularity w ith in the language system , language in general
is often hypothesized to be an independent processing m odule within the
cognitive system and thus, encapsulated from individual differences in
extra-linguistic resources, such as w orking m em ory (WM). The
concom itant m em ory and sem antic deficits in AD allow ed the testing of this
hypothesis in that all errors w ere analyzed for the effects of both limited
WM and sem antic im pairm ent. The results of the stu d y provide little
141
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
evidence to support a theory in w hich the speech production process is
encapsulated from lim itations in WM. In fact, individual differences in
WM account for significant portions of the variance in several types of
errors by both AD patients and healthy elderly people.
While providing no support for m odular theories of language
representation and use, the findings herein are completely consistent with
theories in w hich gram m atical speech production requires interaction
between several levels of linguistic and cognitive processing. For example,
the model described later in this chapter adopts the idea that closed class
w ords are sem antically represented (cf. Bates & Wulfeck, 1989) and posits
that the syntactic inform ation associated w ith w ords is either p art of their
semantic representations or is activated from their sem antic
representations. In this kind of a language system, sem antic im pairm ent
w ould be expected to be associated w ith both closed class w ord errors and
difficulties w ith sentence form ulation. M oreover, in this m odel WM is
hypothesized to have a crucial role in the speech production process,
providing the fuel for the activation of sem antic representations. Besides
accounting for the various error patterns docum ented here, the proposed
m odel also predicts the interaction found betw een accuracy of lexical choice
and fluency in AD subjects' speech across tasks, while this phenom enon
exceeds the scope of other speech production models.
Below, I review the m ost salient findings of the current stu d y that
cannot be explained w ithin current theories of speech production.
Subsequently, I describe an interactive m odel of speech production which:
1) specifies that the sem antic representations of all w ords are com positional,
142
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
2) suggests that WM provides the fuel for sem antic activation, and 3) uses
connectionist principles to account for m any of the findings in this study.
7.1 Discussion of Results
Some of the m ost im portant results of this stu d y concern errors on
closed class w ords because of their theoretical prom inence in m odular
speech production theories. Extrapolating from traditional, m odular
theories, the sem antic im pairm ent of AD patients should only have
resulted in increases in e rro r rates on open class w ords, com pared to the
error rates of NC subjects. H ow ever, the analyses of spontaneous speech
dem onstrated that AD patients m ade considerably m ore errors of all types,
including errors on pronouns and m orphosyntax, as well as on open and
closed class words. In fact, AD patients m ade approxim ately the same
num ber of closed class erro rs as open class errors. Furtherm ore, like the
frequency of AD subjects' open class substitution errors in spontaneous
speech, the frequency of closed class omissions was predictable from the
degree of subjects' sem antic im pairm ent. The relationship betw een open
class word substitutions a n d sem antic im pairm ent w as fully expected from
the point of view of either m odular or interactive speech production
theories since in both they are accessed via sem antic processes. However,
m odular theories have m uch m ore of a challenge explaining the
concomitant relationship betw een closed class errors and sem antic
impairment, since these theories usually postulate th at closed class w ords
are accessed using syntactic mechanisms. It is possible of course, that these
findings m ay m erely reflect concurrent, coincidentally proportional deficits
in syntax and semantics, b u t the lack of other su p p o rt for a syntactic deficit
143
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
in the spontaneous speech of AD patients argues against this (Kempler, et
al., 1987). In contrast the interactive theory of speech production described
below specifically predicts that sem antic im pairm ents should result in
difficulties producing both open and closed class words, and predicts that the
frequency of these errors should be related to the degree of subjects'
sem antic im pairm ent.
Along sim ilar lines, m odular theories of speech production assign
sentence construction abilities to a syntactic processing m odule that has no
input from or interaction w ith sem antic processes. The current study
revealed tw o effects w hich cannot be explained w ithin that theoretical
paradigm , but w hich are m uch m ore readily explained in a interactive
theory of language in w hich certain syntactic inform ation, such as the
argum ent structure of verbs, is encoded sem antically (e.g., T anenhaus &
Carlson, 1989; M acD onald, Pearlm utter, & Seidenberg, 1994). For example,
the results revealed that the frequency of syntactic errors in the constrained
production task w as highly predictable from AD subjects' sem antic
im pairm ent. These syntactic errors consisted largely of argum ent structure
errors sw itching the them e and experiencer w hen using a T-E verb (e.g.,
^Ro b e r t TERRIFIED t h e FIRE). These verbs im pose an unusual argum ent
structure assignm ent on active sentences in that they position the
inanim ate them e as the subject and the anim ate experiencer in the direct
object place, thereby requiring speakers to overcome their tendency to use
the canonical English, anim ate-first/active sentence type (M cD onald et al.,
1992). Judging from the errors they m ade, these AD patients show ed a
tendency to rely m ore on canonical sentence form s (e.g., A nim ate N oun-
Verb-Inanim ate N oun) as their sem antic im pairm ent becam e m ore severe,
144
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
even w hen the argum ent structure of the verb prohibited this. As suggested
above, if argum ent structure inform ation is semantically encoded, as
suggested by interactive, constraint-based m odels of language use, then this
type of error is relatively easily explained as a sym ptom of sem antic
breakdow n. O n the other h and, these results present a challenge w ithin a
m odular theory of language representation since there should be little effect
of sem antic im pairm ent on the choice of sentence frame. In fact, within
G arrett's theory, (1982), one w ould have difficulty even explaining how a
syntactic deficit could be related to a particular verb type, since w ords are
chosen independently of sentence frames. W ithin Levelt's theory, it w ould
be necessary to postulate a partial im pairm ent of verb-type-specific, syntactic
building procedures that occasionally prevented the use of argum ent
structure inform ation associated w ith them e-experiencer verbs, while
specifying that this dam age w as only coincidentally related to semantic
im pairm ent. Thus, while this pattern of results can be explained within
these theories, the explanation is not straightforw ard and requires the
assum ption that errors in argum ent structure are only coincidentally related
to sem antic im pairm ent
The second finding th at points to an im portant role for sem antics in
sentence construction ability concerns the fluency of AD subjects' speech in
the constrained production task. In m odular theories of language, the
sem antics of a w ord are activated in parallel with its syntax, and both are
assum ed to consistently precede articulation (Garrett, 1980, 1982; Levelt,
1989; Levelt et al., 1991). Thus, a sem antic im pairm ent m ight be expected to
lead to w ord finding problem s or difficulties in word reading, but not to
difficulties w ith sentence production once the w ords for the sentence had
145
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
been read aloud since successful articulation presum es the previous access
of sem antic and syntactic information. H ow ever, in the constrained
production task, AD patients attem pting to use the stim ulus w ords to create
sentences m ade m any m ore fresh start dysfluencies, (i.e., aborted attem pts at
sentence creation), than healthy subjects in the sam e task, and many m ore
than either group h ad m ade in spontaneous speech. Examination of the
data (see A ppendix E) suggests that there w as a tim e period in which
subjects w ere able to pronounce the stim uli b u t not to combine them into
sentences, suggesting, perhaps, that subjects suffered a partial failure of the
processes that quickly activate syntactic or sem antic representations from
reading. H ow ever, since the sentences ultim ately produced were usually
completely gram m atical, and there were few syntactic difficulties in
spontaneous speech, it is difficult to blame these dysfluencies on a syntactic
deficit. A possible explanation stems from the fact that the frequency of AD
subjects' fresh starts on the CP was highly related to their semantic
im pairm ent. If these AD subjects were able to read the stim ulus words
aloud w ithout fully activating their sem antic content, as has been recently
postulated for another group of AD subjects (Rayner & Bemdt, 1996), this
pattern of results m ight obtain. I postulate that severity of semantic
im pairm ent m ay directly affect the speed of activation of semantic features
of visually presented words. A necessary corollary to this hypothesis is that,
in order to use a particular word in a sentence, a m inimal proportion of its
semantic features m ust be activated. Given these assum ptions, AD subjects
were incapable of producing a sentence in response to CP stimuli initially,
which m ay have resulted from the inability to quickly activate a sufficient
146
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
set of sem antic features of the stim ulus w ords in order to determ ine how to
combine them into a sentence.
The phenom enon just described, in w hich the frequency of fresh
starts w ere strongly related to sem antic im pairm ent, occurred only in the CP
task in w hich subjects had no choice of the w ords to use. In spontaneous
conversation AD patients spoke just as fluently as their healthy elderly
counterparts, despite making m ore lexical errors. M oreover, sem antic
im pairm ent predicted both dysfluency rates in the constrained production
task and error rates in spontaneous speech. This set of findings supports
one of the hypotheses suggested in the introduction: that sem antic
im pairm ent m ay lead to a trading relationship between fluency and lexical
accuracy in the speech of AD patients. A pparently, this trade-off was
exacerbated by task demands. That is, w hen AD patients w ere constrained to
produce a sentence using the specific w ords in the CP stim uli, their speech
became m uch less fluent, but the frequency of open class and pronoun
errors w as m inim ized. In contrast, w hen free to choose their ow n w ords in
spontaneous speech, AD patients were as fluent as healthy elderly speakers,
although their speech contained m any m ore lexical errors. This pattern of
results cannot be explained w ithin traditional theories of speech production,
because they cannot explain the finding that AD subjects' could read w ords
but not combine them into a sentence, unless they posit a syntactic
im pairm ent, a hypothesis not supported by data here or elsew here. Further,
m odular theories prohibit sem antic im pairm ent from interacting w ith
sentence construction processes, so again they w ould have to label this
correlational relationship as coincidental.
147
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
In general, Levelt (1989, Figure 1.1, p. 9) m ight attem pt to explain
m any of these results by suggesting that the dam age in AD w as to either the
conceptualizer or the organizing lemma node, rath er than to the semantic
representation itself. W hile Levelt does attribute sem antic substitution
errors to slips at the conceptual level, it seems unlikely that these would be
the only results of dam age at this fundam ental level of representation. One
m ight also expect bizarre, alm ost schizophrenic speech, or incoherent,
indecipherable and m alform ed utterances from dam age at this level, since it
is the prim ary in p u t to all other linguistics levels. H ow ever, since the
majority of the AD subjects' utterances were w ell-form ed and coherent,
especially in spontaneous speech, it seems unlikely th at the damage causing
problem s w ith closed class w ord usage and sentence production is at the
conceptual level. O n the other hand, dam age at the lem m a level offers a
better explanation in th a t it could lead to problem s understanding and
producing w ords d ue to difficulty activating the lem m a node. However,
considering the unitary n atu re of the lemma node, one w ould expect
equivalent rates of difficulty for all tasks that require the use of that lemma,
so one w ould never expect, for example, that picture nam ing ability m ight
be m ore im paired than w ord picture m atching tasks, a com m on finding in
all populations w ith sem antic im pairm ents (Hodges et al., 1992). Moreover,
dam age at the lem m a level w ould also lead to difficulties w ith access to all
lem m a com ponents including phonology, which is com pletely preserved in
AD. Based on these arguments, it does not appear th at dam age at either the
lem m a or conceptualizer level can account for the p attern of results found
here. As previously stated, though, these results can be easily accounted for
w ith a m odel that specifies that all w ords are sem antically represented, and
148
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
the syntactic inform ation associated w ith a w ord is activated as a p art of or
from its sem antic representation, as described below.
M odular theories of speech production have also asserted that
linguistic processes are encapsulated from the effects of w orking memory
limitations; researchers in that paradigm , then, would have expected no
relationships betw een WM ability and language difficulties. In contrast,
both interactive theories of language use and WM research suggest that
WM plays an im portant role in speech production. Several of the results
from the p resent stu d y support the latter theory, suggesting that WM plays
an crucial role in the initial activation of the semantic representations of the
w ords to be used in a sentence. For exam ple, the data revealed that normal
elderly subjects w ith lower WM scores om itted more w ords than others in
their cohort w ith better WM scores. Furtherm ore, three quarters of these
omitted w ords belonged to closed class categories. This finding is consistent
with the prediction m ade in Chapter 2 that, due to having very sm all sets of
semantic features and very high frequencies, closed class items m ight be
particularly vulnerable w hen WM is lim ited.
O ther findings that also argue for the importance of WM in the
activation of the w ords chosen for an utterance include the fact that the AD
subjects w ith the m ost im paired WM m isused m ore of the stim uli in the
constrained production task, either using morphological substitutions of the
verbs in the stim uli or om itting one of the three stim ulus w ords altogether.
Given that the three stim ulus w ords rem ained in clear view to all subjects
throughout each response, it is difficult to conceive of an explanation for
these errors a n d their strong relationship w ith WM if there is no interaction
149
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
between WM abilities an d language processing. In the m odel presented
below, I suggest that low ered WM resources left these AD patients unable to
completely activate the sem antic features of all three stim ulus w ords at one
time. This led, in the m ost severe cases, to the om ission of one of the
required stim ulus w ords, or in less severe cases, to the partial activation of
the past participle verb form , leading to the substitution of a
morphologically less com plex form, the simple past.
In addition to the m isused stimuli, there were also complicated
interactions betw een sem antic and WM impairm ents in the speech of AD
patients. In particular, the results showed that, in the spontaneous speech
of AD subjects w ith sim ilar degrees of semantic im pairm ent, subjects with
m ore intact WM m ade m o re open class substitution errors than subjects
w ith lower W M.1 This puzzling phenom enon contradicts both the
predictions of m odular speech production theorists and also those of the
proponents of lim ited W M as a major cause of lexical errors. However,
Power (1985) and the findings of Daneman (1991) suggest that WM may
have its greatest effect a t the semantic level, acting as the activating force for
semantic representations, w ith m ore semantically com plex concepts
requiring m ore WM for activation. The model described below adopts this
idea, in com bination w ith other findings, to account for these apparently
anom alous findings.
Since the traditional, m odular theories of speech production adm it
no influence of sem antics or WM on closed class w ord production, one of
‘In this analysis, semantic impairment, as measured by Within-category errors, entered the
equation first and so was the primary predictor of the number of open class substitution errors,
(r^ = .42), but subsequently, nonsemantic-WM accounted for an additional 21% of the variance.
150
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ith out p erm issio n .
the m ost interesting and theoretically im portant findings of this stu d y was
the vulnerability of closed class w ords to individual differences in WM and
sem antic ability. Indeed, as stated above, in the spontaneous speech of
control subjects closed class omissions com prised the largest single category
of errors, and the frequency of closed class omissions (indeed, all w ord
omissions) w as associated with individual differences in WM. Sim ilarly, in
the spontaneous speech of AD subjects closed class w ord om issions
significantly outnum bered closed class substitutions and were alm ost as
num erous as open class substitutions. H ow ever, the much higher om ission
rates of closed class w ords among AD subjects compared to controls w ere
predicted by sem antic im pairm ent. Based on these findings, it appears that
either WM lim itations or semantic im pairm ent can lead to increased closed
class w ord om issions, b u t where both im pairm ents are present, as in AD
patients, the effect of sem antic im pairm ent is likely to be larger and m ore
obvious. This finding m ay account for som e of the difficulties that W M-
a n d /o r sem antically-im paired aphasic patients have with closed class items.
O ne of the m ost consistent findings of this study was the deleterious
effects that various kinds of noncanonical stim uli in the CP task had on the
speech of AD patients. Two types of stim uli which encouraged
noncanonical responses were Them e-Experiencer verbs and Inanim ate-first
stimuli. Both of these stim uli challenged the sam e conceptual constraint,
the tendency of English speakers to position an anim ate noun before an
inanim ate one in a sentence, a tendency called here the "A nim ate-first
constraint." T he Anim ate-first constraint is quite strong in English an d
constructions th at violate it are usually avoided (McDonald et al., 1992).
Them e-experiencer verb stimuli (e.g., SHOCKED, TERRIFIED) required subjects
151
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
to choose betw een two noncanonical alternatives, either producing a
passive sentence w ith noun stim uli in the preferred English order (e.g.,
SU SAN w a s SHOCKED BY THE BOOK) or producing an active sentence that
placed the inanim ate noun stim ulus first (such as, THE BOOK SHOCKED
SU SA N ). In contrast to control subjects who consistently chose a passive
sentence form with stim uli of this type, thereby accom m odating the
Anim ate-first constraint, the AD group was basically at chance w ith respect
to w hich constraint they w ould violate. In addition, AD patients produced
higher rates of syntactic errors specifically related to argum ent structure
w ith these verbs. These errors (i.e., *SUSAN SHOCKED THE BOOK) neither
used noncanonical passive sentence structures nor violated the animate-
first constraint, but did violate the argum ent structure requirem ents of
these verbs. The second type of stim ulus which tested the Anim ate-first
constraint involved those stim uli in which inanim ate n ouns w ere
prom inently positioned over anim ate nouns, thus encouraging the
production of sentences w ith inanim ate nouns first. In fact, both groups
reordered m ost of the Inanim ate-first stimuli in order to use the anim ate
noun first, and those subjects w ho did n o t m ade the m ost errors in this
condition. Further, a subset of AD subjects produced alm ost agram matic
responses to Inanim ate-first stim uli, om itting m any necessary closed class
w ords. Therefore, violating (or even being encouraged to violate) this
"anim ate-first" constraint on English sentence production led difficulties for
AD subjects; however, the instantiation of this difficulty varied depending
on w hich aspect of the stim ulus encouraged a noncanonical response.
W hen the verb type encouraged a noncanonical response, there w ere errors
affecting the verb's argum ent structure, but w hen noun order encouraged
152
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
the noncanonical response, the result was an overall increase in difficulty in
assembling the necessary elements for a gram m atical sentence.
The last and m ost pervasive set of noncanonicity effects in the
constrained production task occurred in response to irregular past participle
stimuli, (e.g., SHOW N, CHOSEN). Responses to these verbs contained
significantly m ore dysfluencies and m isused stim uli, and also m ore lexical
errors (syntactic errors and om itted closed class items) than responses to the
other two verb types. Furthermore, these items caused difficulties for
subjects w ith prim arily sem antic impairm ents as well as those w ith severe
WM im pairm ents, though these difficulties m anifested in different ways.
Item analyses su p p o rt the assertion that difficulties w ith IRREG verbs did
not stem from w ord frequency or the num ber of m eanings of the individual
verbs, but could be predicted by a measure derived from morphological
canonicity (w hether the past participle form w as the sam e as the sim ple past
tense) and syntactic canonicity (the frequency w ith which NC subjects used a
stim ulus in the canonical active form). Therefore, we have been forced to
conclude that these irregular morphological form s are inherently m ore
difficult than regular verb forms, both because they require subjects to use
m ore complex sentence forms and because they have non-standard
morphological forms. Apparently, the conjunction of these two types of
noncanonicity necessitated a more intense usage of both semantic and WM
resources, leading to increased difficulties by subjects w ith im pairm ents to
either faculty. The pervasive and deleterious effects of noncanonical
stimuli on AD speech production may contribute to the dependence of AD
subjects' on all types of canonical forms (Bates et al., 1994) to preserve fluent,
gram m atical speech. These findings exceed the scope of current theories of
153
R ep r o d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
semantic representation and speech production and are discussed again
below.
In this section I have presented th e principle results of this study and
have explained how these findings contradict w hat might be predicted based
on the traditional, m odular theories of speech production. These results
have show n that there is a considerable am ount of interaction betw een
semantic ability an d the generation of syntactic form, and that individual
differences in W M ability influence the com pleteness of sentences produced
as well as the accuracy of lexical selection. Secondly, these findings support
the hypothesis that the speech of AD subjects is characterized by a trade-off
between lexical accuracy and fluency, such that w hen fluency is required, the
accuracy of lexical choice may suffer, an d w hen the accuracy of production is
at a prem ium , fluency m ay suffer. Further, these results have dem onstrated
the reliance of AD speech on canonical form s at conceptual, m orphological,
and syntactic levels, since employing noncanonical forms led to consistent
difficulties in sentence production at all these levels. To account for these
findings, a new m odel of speech production processes is described below.
7.2 An Interactive, Distributed M odel of Speech Production
The results of the current study have show n that the traditional,
m odular theories of speech production are inadequate for describing the
patterns of speech errors and dysfluencies of Alzheimer's Disease patients in
either spontaneous or constrained speech. As such, I have developed a
model that can better account for these findings by em phasizing the
interaction betw een WM, semantic ability, and syntactic processing. M any
of the ideas presented below are based o n Parallel Distributed Processing
154
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
(PDP) m odels of com plex behaviors (Rum elhart, McClelland, & the PDP
Research Group, 1989; McClelland, R um elhart & the PDP Research G roup,
1989). For example, a recent PDP m odel of the sem antic system postulated
that w ord m eanings w ere com prised of sets of overlapping sem antic
features which are them selves defined as distributed patterns of activation
w ithin the sem antic system (Plaut & Shallice, 1993). Typically, in these
m odels, the units that com prise a representation are highly interconnected,
and the connections betw een them have different strengths that reflect the
frequency w ith w hich two units have been activated sim ultaneously. These
changes in connection strength betw een units reflect the learning that has
taken place in the netw ork (McClelland, R um elhart & Hinton, 1989; Elman,
Bates, Johnson, Karmiloff-Smith, Parisi, & Plunkett, 1996). O ther properties
derived from various PDP m odels are discussed below in m ore detail, such
as the principle of graceful degradation (M cClelland et al., 1989) and various
phenom ena associated w ith intercorrelated and distinguishing features
(Gonnerm an, A ndersen, Devlin, Kempler & Seidenberg, 1997).
At the heart of this m odel is the idea th at attention or WM provides
the fuel or im pulse to activate the sem antic representations necessary for
speech. Specifically, the speech production process is envisioned as
involving, first, a decision to articulate a m essage originating at a (more-or-
less2) nonlinguistic, conceptual level; this decision focuses attention on the
m essage which then acts as input to the sem antic level of representation for
linguistic encoding. Thus, the am ount of attention allotted to this
"transfer" of inform ation will determ ine the clarity, accuracy, and
2 In a highly interactive system such as that conceived here, it is entirely possible that aspects
of semantic representations may interact with and influence the development of, and
relationships among, conceptual structures at this "nonlinguistic" level.
155
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
com pleteness of th e resulting utterance. If insufficient attention or WM
resources are available for the task, subjects m ay om it w ords, particularly
closed class w ords, as d id healthy elderly control subjects in this study. In
cases in w hich W M is pathologically limited, such as in AD subjects, there
also may be increases in speech dysfluencies in spontaneous conversation;
that is, subjects m ay have to backtrack slightly to refresh the m em ory trace
of the planned utterance, or abandon an utterance entirely if the final
portion of the m essage has not been activated strongly enough to be
sem antically encoded.
Along sim ilar lines, this study has presented evidence that the
am ount of available WM m ay limit the am ount of inform ation that a
subject can sem antically encode and incorporate into ongoing speech. This
w as show n in the strong relationship betw een W M and m isused stim uli in
the constrained production task, in which the AD subjects with the m ost
im paired WM om itted and substituted w ords from the CP stimuli in
proportion to their WM deficit. I suggest that, in the case of omissions, the
extent of WM lim itation allow ed only two of the three stim uli to be fully
activated for use in a sentence. Similarly, m orphological substitutions of
irregular past participles verbs may have occurred w hen limited WM
allowed only a subset of the verb's semantic features to be activated,
resulting in the use of the sim ple past tense instead of the past participle
show n on the stim ulus card. The model described here explains this last
phenom enon by ad o p tin g the suggestion of Bates and Wulfeck (1989) that
m orphological form s of the sam e verb share m ost of their semantic
features. Based on this, I posit that simple past an d past participle forms of
irregular verbs share a core set of semantic features, w ith the irregular past
156
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
participle form requiring the activation of additional, less frequent features
to distinguish it from the sim ple past. Thus, the sim ple past and past
participle forms of irregular verbs represent two phonologically distinct
w ords w ith highly overlapping m eanings which are associated w ith distinct
syntactic structures. Im portantly, it is the activation of the distinguishing
features that signals the appropriate syntactic forms, so failure to activate
these will lead to the use of the com m on form, the sim ple past. When WM
is lim ited, these less frequent, distinguishing features m ay not reach full
activation,3 although the m ore frequent shared features w ill, and the simple
past form is produced. This difference in representation w ould m ake
irregular past participles m ore difficult to utilize in building sentences since
subjects had to be able to activate those crucial distinguishing features that
differentiate the past participle from the simple past form in order to use the
irregular form in a sentence. This hypothesis entails that irregular past
participles are fundam entally different from past participles form ed
regularly; in the latter case, the sim ple past and past participle forms are
identical and so represent o n e w ord w ith a single m eaning w hich is linked
to at least three syntactic structures (active, passive, and perfect). Thus,
activation of this form allow s the use of any of these structures w ith little
consum ption of extra resources. Furtherm ore, it is likely that the deficits in
the attentional aspect of WM also contributed to the frequency of these
errors, since stim uli w ere continuously visible throughout the response and
these subjects still did not notice their errors.
In PDP models, sets of features that frequently co-occur in different
representations develop strong interconnections. One effect of this is that
^The vulnerability of distinguishing features is discussed below.
157
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
w hen a pattern is inadvertently only partially activated this may be
com pensated for, to a certain extent, by strong connections w ith other co
occurring features w hich h a v e been fully activated. In other words, given a
semantic representation w ith a num ber of highly related features (e.g. TAIL,
FUR, 4 LEGS, LIVE BABIES), w hen only a subset of these are activated from
conceptual processes, the strong connections to the other frequently co
occurring, or "intercorrelated," features can subsequently activate the rest of
the set (G onnerm an et al., 1997; Devlin, G onnerm an, A ndersen &
Seidenberg, 1998). This ability, which leads to a phenom enon called
"graceful degradation," is crucial to this m odel's explanation of lexical
substitution errors. Graceful degradation allows the system to make its "best
guess" for a w ord based on the set of currently activated units, even if only a
partial set of features w ere originally activated (Rum elhart et al., 1989). As
long as the set of units activated is sufficient to uniquely designate one item
m ore strongly than any other, a word can be produced. However, graceful
degradation does not guarantee that the o u tput from this process will be
correct; that depends on w hether the resulting pattern of activation is m ost
sim ilar to that of the target or if the pattern activated has more in com m on
w ith another word.
D istinguishing features are those features w hich only occur in a few
m embers of a sem antic category and serve to distinguish betw een sim ilar
items (Gonnerm an e t al., 1997; Devlin et al., 1998). For example, the feature
HAS STRIPES distinguishes betw een TIGERS and LIONS, as well as betw een
ZEBRAS and HORSES, bu t occurs in very few other sem antic representations
of animals. Therefore, by their nature, distinguishing features are less
frequent than the intercorrelated features that are shared by all m em bers of
158
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
a sem antic category. Furtherm ore, distinguishing features will be m ore
vulnerable th an shared core features w h en there is dam age to the system
because, by definition, they lack the strong connections due to frequency of
co-occurrence w hich are found betw een intercorrelated features.
Consequently, distinguishing features m u st rely on the initial activation
im pulse from conceptual processes to reach threshold activation, and if this
impulse is n o t strong or clear enough, a n incorrect w ord m ay be produced.
In contrast, as stated above, failures to initially activate the com plete set of
intercorrelated features are com pensated for because the strong connections
from other intercorrelated features th at h a v e been activated will bring those
features not initially activated up to threshold. Hence, the precision of the
initial activation is particularly im portant for distinguishing features,
because the netw ork can only com pensate for partial activation of
intercorrelated features.
H ow m ight this m odel account for the findings of this study
regarding closed class errors? The tw o m ost im portant findings to explain
are that 1) om issions of closed class w ords w ere m ore prevalent than
substitutions in both groups, and 2) both lim ited WM and sem antic
im pairm ent w ere associated w ith the om ission of closed class w ords. In this
model, all w ords are sem antically represented b u t do not have the sam e
num ber of features. Following Bates an d W ulfeck (1989) and Bock (1989),
closed class w ords are posited to have very sm all feature sets that, as a result
of their exceedingly high frequency, are very quickly activated.
Furtherm ore, there is little overlap in m eaning betw een m em bers of the
closed classes and so any set of intercorrelated features is very lim ited,
leading to sem antic representations th at consist, in essence, of sm all sets of
159
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
distinguishing features. W ith no core, intercorrelated features to boost
activation levels, any type of partial activation, d u e to lim ited WM or a
noisy sem antic system, m ay lead to n£> w ord being clearly activated— leading
to an om ission error rather than the substitution of a related w ord. Thus,
un d er conditions of either lim ited W M or sem antic im pairm ent the model
predicts that closed class w ords should be extraordinarily vulnerable to
om ission errors, while w ords w ith richer sem antic representations should
participate in m ore substitution errors, as described below.
These concepts of intercorrelated and distinguishing features are also
fundam ental to the explanation of the open class errors m ade by
A lzheim er's patients. In AD, dam age to the sem antic system occurs quite
early and, in m any mildly affected patients, m ay be the m ost noticeable
sym ptom of the disease (M artin, 1990; Kempler, 1991). One of the primary
neurological effects of AD is the m alform ation and loss of synapses (Terry,
Peck, DeTeresa, Schechter, & H oroupian, 1981; M asliah, H ansen, Albright,
M allory, & Terry, 1991; Terry, M asliah, Salmon, Butters, DeTeresa, Hill,
H ansen, & Katzman, 1991). In a connectionist m odel, these effects could be
realized as either the distortion of weights on connections or the loss of
connections betw een units in the system (Devlin et al., 1998). These changes
in the sem antic netw ork m ake the patterns w hich are activated "noisy" and
am biguous, and consequently, it is m ore difficult to accurately activate
sem antic representations. The increasing noisiness w ould have its largest
initial effect on distinguishing features, due to their lack of compensatory
intercorrelations, leading to increased substitutions of related w ords from
the sam e semantic category.
160
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
As the loss of connections in the sem antic system becomes more
w idespread in AD, speech production becomes m ore difficult, particularly
when the production task dem ands lexical accuracy, such as in naming.
This task is especially difficult because strong connections between co
occurring features will activate the complete set of intercorrelated features
associated w ith bo th the target and other sim ilar w ords from the same
category, b u t the distinguishing features of the pictured item may be m ore
difficult to activate, as m entioned above. Thus, if the necessary
distinguishing features are not initially activated to the necessary level, the
correct w ord m ay not be available, and a related w ord w ith a very similar
pattern of activation m ay be produced. In the early stages of semantic
im pairm ent, the m odel predicts that nam ing errors will involve the
substitution of sim ilar item s from the sam e sem antic category, (e.g., W ithin-
Category nam ing errors). As it turns out, this is the m ost prevalent type of
error m ade by m ild Alzheim er's patients both in spontaneous speech and in
naming tests (Nebes, 1989). As discussed in C hapter 4, within this model
the frequency of these errors represents an overall m etric of the difficulty
associated w ith achieving the accurate activation of specific items within the
semantic system . The patterns of activation becom e progressively more
am biguous as the sem antic network becomes m ore degraded, resulting in
only the intercorrelated features being activated and the subsequent the
production of a superordinate category name (Tippett, McAuliffe & Farah,
1995; Done & Gale, 1997). Alternatively, only a feature of the item may
become active, leading to a description of this sem antic feature of the target,
(e.g., Superordinate-Circum locution errors). Eventually, as more
connections are lost due to disease and the areas of im pairm ent begin to
161
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
overlap, the p attern activated in th e sem antic system will have little
resem blance to a real w ord representation, and the subject's nam ing
response w ill show no semantic know ledge of the target w ord, (e.g., No-
Response errors). Thus, in AD N o-Response errors m ay represent the size
or frequency of large "patches" of diseased or inactive sem antic units. This
progression of errors recapitulates that described in H odges et al. (1992).
Paradoxically, having m any inactive sem antic units, as represented by high
num bers of No-Response errors, m ay actually lower overall noise in the
sem antic system and allow for better perform ance on som e tasks or for
som e patients. This is a possible explanation for two anom alous findings
concerning sentence construction in the current study: 1) subjects w ith
m ore N o-Response errors apparently m ade fewer syntactic errors on the CP
task, and 2) subjects with m ore N o-Response errors rejected m ore whole
sentence responses in the CP task (m any, b u t not all, of w hich contained
errors). Thus, losing already dam aged units and their anom alous
contributions to sem antic com putations m ay actually provide less noisy
input to syntactic processing, allow ing, in this case, m ore accurate sentence
construction and better m onitoring of speech output.
In conversational speech, m any types of errors by the AD subjects
were predictable using the W ithin-C ategory nam ing errors m easure, which
has just been analyzed as a m easure of subjects' difficulty in accurately
activating sem antic representations. H ow ever, open class errors show ed an
odd pattern of predictive factors in that, while increases in sem antic
im pairm ent led to increases in sem antic substitution errors as predicted,
subjects w ith h ig h er WM m ade even m ore errors of this type. This model
suggests that, in subjects w ith higher WM capacity, the sem antic
162
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
representations activated w ere more complete, consisting of both
intercorrelated and distinguishing features, b u t there w ere not enough
distinguishing features activated to unam biguously allow the correct w ord
to be produced. O n the other hand, subjects w ith low er WM capacity
probably activated only the m ore frequent, intercorrelated features of the
sem antic representation, thus producing an appropriate, but less specific
w ord. The key to this explanation is that a language contains a variety
w ords with a range of specificity, any of which can be used correctly in a
given situation. For exam ple, the words DO < MAKE < COOK < {BAKE, FRY,
GRILL} represent a range of specificity in m eaning, and any of the w ords
could be used in the sentence, "I'm going to X the chicken now," and no
error w ould be noticed. This phenom enon can be accounted for in the
current m odel because it specifies that the features that comprise m ore
general, superordinate term s actually consist of the core features that
underlie several m ore specific subordinate term s. Hence, since these
features participate as a group in all subordinate w ords as well as the
superordinate term , the superordinate feature com plex is more frequent
and has m ore intercorrelations than those of subordinate terms, and thus,
reaches activation m ore easily than more specific w ords (Tippett et al., 1995;
Done & Gale, 1997). This sam e principle probably contributes to AD
patients' ability to m aintain fluency in spontaneous speech despite
increasing sem antic im pairm ent. That is, if com plete and accurate sem antic
representations of the specific w ords for a m essage are not readily available
to speakers w ith AD, they can use other less specific, "em pty" w ords w ithout
being in error and w ith o u t having to interrupt their speech with frequent
bouts of w ord finding. As illustrated above, these less specific w ords will
163
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
not necessarily be inappropriate in conversational speech, though a more
precise w ord m ight be preferred by both the speaker and the listener.
The im portance for AD patients of being able to choose one's ow n
w ords in speech is show n graphically by this study in that constraining their
speech to include very specific w ords in the CP task m ade it exceedingly
difficult for subjects to produce sentences, which resulted in increased errors
for a few patients and decreased fluency for m ost of them. Indeed, for
several patients the m ain effect associated with semantic im pairm ent was
that it seem ed to take them an extraordinarily long tim e to construct a
sentence using the three stim ulus w ords. While apparently struggling for
semantic access, these subjects often abandoned a num ber of responses,
before eventually being able to produce a grammatical sentence. This
pattern w as particularly evident w hen subjects were required to use
irregular past participles. This supports the earlier suggestion that these
verb forms are fundam entally m ore difficult than other verb types, having
perhaps m ore distinguishing features that m ay take a longer tim e to
activate, and requiring more precise patterns of activation for correct usage.
From this, w e can infer that sem antic im pairm ent affected subjects' ability to
quickly and accurately activate the sem antic representations of the irregular
verb stim uli, which, in turn, im paired subjects' ability to construct a
response to the stimuli.
The actual form ulation of sentences from a set of activated sem antic
representations is hypothesized in this m odel to involve satisfying a
m ultidim ensional set of constraints on sentence structure (Tanenhaus &
Carlson, 1989; MacDonald, 1993). Thus, each sentence produced is composed
164
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
or built to satisfy the sem antic requirem ents of the message, the syntactic
constraints associated w ith the words chosen for use, and any other
language-specific or pragm atic constraints that m ight be operable at the time
(e.g., MacDonald, Pearlm utter, & Seidenberg, 1994), rather than being
accessed from a store of previously compiled structures (Garrett, 1982,1984).
One of the m ost lim iting constraints w ould be the verb chosen for use.
Since the syntactic form s associated w ith a verb are contingent upon its
m eaning (Levin, 1993), sem antic im pairm ent m ight very well lead to
difficulties in sentence construction, syntactic errors, or argum ent structure
errors. However, there is nothing in this m odel that discrim inates betw een
models in which the syntactic requirem ents of a verb are semantically
represented, and so are activated sim ultaneously w ith their semantics
(MacDonald et al., 1994), or m odels in which the syntactic representations of
verbs are em ergent properties of their sem antics and thus are activated
secondarily from sem antics, as seems to be the suggestion in Levin (1993).
A nother constraint that proved to be very pow erful in predicting
difficulties for the AD subjects was the anim ate-first constraint which
seemed to operate at a m ore conceptual than linguistic level. This
constraint was credited w ith the high proportion of passives used by norm al
control subjects w ith T-E verbs (Ferreira, 1994), as well as with the high error
rates on items in w hich AD subjects did not reorder inanim ate first noun
stimuli. Using noun stim uli presented in this ord er increased the difficulty
of the task for AD patients, leading to exceptionally high closed class
omissions by a subset of patients. I suggest that supralinguistic, conceptual
constraints like this one helped AD patients structure their speech w hen the
semantic activation of specific words was com prom ised, leading to a
165
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
reliance on the canonical conceptual o rd er "anim ate noun-verb-inanim ate
noun." In support of this, AD subjects had considerably less difficulty
producing speech w hen using canonical linguistic structures, such as sim ple
active sentences and regular m orphological forms. This pervasive pattern
of preferring canonical sentence structures w as likely not a conscious
strategy, but rather w as an alm ost autom atic effect of the im paired linguistic
system which tended to settle into general, high frequency linguistic forms
w hen sem antic inform ation w as unavailable to counteract this tendency.
Nevertheless, these constraints can be overridden, as show n in the speech
of the NC subjects w ho had no difficulty in producing noncanonical passive
and active sentences. How ever, the results of this study suggest that
avoiding canonical linguistic pathw ays m ay require more extensive use of
semantic resources. Thus, as the sem antic im pairm ent of AD patients
became m ore severe, their speech production tended to follow m ore
canonical syntactic patterns, especially in tasks that stressed the sem antic
system such as the constrained production task used in this study.
7.3 Unaccounted For Results
U nfortunately, as presently conceived this model cannot account for
all of the findings in this study. Most notably, I can only speculate about the
significance of the existence of the two groups of AD patients w ith different
patterns of errors w ith respect to closed class w ords and pronouns. That is,
the current study found that three of the ten AD subjects show ed extrem e
effects of sem antic im pairm ent on their closed class word production,
leading to m any agram m atic-like responses in the CP task, as listed in
Appendix G. These three subjects also om itted m ore closed class w ords and
166
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
m ade m ore pronoun errors (and m arginally m ore open class and
m orphological errors) in the personal narrative than the rest of the AD
group. Since there w ere no dem ographic, sem antic ability or WM
differences betw een subjects w ho om itted large num bers of closed class
items and those w ho did not, w e m ust look for an explanation for this
difference elsewhere. Three possible explanations are offered here.
First, Plaut and Shallice (1993) found that in a highly connected
semantic netw ork, the production of w ords w ith few er semantic features
was affected to a m uch higher degree by "noise" in the connections in the
network. Based on this, the three subjects w ho have been called "high
om itters" m ay have m ore "noise" in the connections in their semantic
networks due to slight differences in disease processes, making it more
difficult to produce w ords w ith very few sem antic features, such as closed
class words.4 This implies that the effects of AD on the semantic system
may differ betw een patients, in som e people causing an increase in "noise"
in addition to dam age in the system , and in other people causing similar
dam age in the sem antic system but w ithout excessive noise. Since the
effects of AD on different patients does vary considerably, this is a relatively
plausible explanation.
An alternative explanation requires th at a substantial subset of
English speakers organize their language representations in a qualitatively
different m anner from the m ajority of English speakers. In this alternate
system, closed class w ords w ould be activated, not directly via their own
semantics, b u t indirectly via the heads of the relevant phrases in which they
4 It is likely that pronouns also have few semantic features and so would be expected to show
similar effects.
167
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
occur as in Levelt (1989). For example, the activation of a determ iner would
depend upon the previous activation of the n o u n to which the determ iner
refers. In this scenario, failure to sufficiently activate the head of the phrase
m ight then lead to the om ission of the required closed class w ord. This
m anner of organizing a linguistic system w ould lead closed class w ord
production to be doubly dependent on the ability to fully activate lemmas,
because both the initial activation of an open class w ord lemma (and its
syntactic representation, according to Levelt), a n d the subsequent activation
of the lemma denoting the closed class w ord m ust be successful in order for
a closed class w ord to be produced. Acceptance of this alternative leads to
the inference that there m ay be at least two different ways in which English
speakers m ay organize their linguistic system s— one in which closed class
w ord usage depends on the prior activation of an open class w ord or
syntactic structure, and the other using, perhaps, direct activation of closed
class w ords via their ow n semantic features.
The third a n d final speculative explanation for this phenom enon
springs from a recent theory proposed by Deacon (1997), in which closed
class items m ight be analyzed during the acquisition process as either
semantic elem ents or purely grammatical elem ents. In the former case,
closed class w ords w ould be completely sem antically encoded (probably in
tem poral cortex), and, as such, subject to errors w hen there is sem antic
impairm ent. If closed class words were analyzed as purely gram m atical
elements signaling clause and phrase boundaries, Deacon predicts they
would be encoded in frontal cortical areas and, as such, resistant to semantic
im pairm ent b u t vulnerable to dam age w ithin prefrontal cortex or to
connections betw een tem poral and prefrontal cortex. However, Deacon
168
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
does not acknowledge that b o th patterns m ight occur in a group of speakers
of the sam e language. H e predicts specifically that languages using
morphological inflections as the prim ary indicators of syntactic information
w ould be m ore likely to encode this inform ation sem antically, in temporal
cortex, while languages that use w ord order and separate particles to carry
syntactic inform ation, such as English, w ould m ost likely encode this in a
part of the brain specialized for ordering relationships, prefrontal cortex.
Acceptance of this hypothesis w ould again entail that different speakers of a
given language m ight encode crucial parts of the language in very different
ways, that is, the high om itter group m ight encode closed class words
semantically, in tem poro-parietal cortex, while the low om itters m ight
encode closed class vocabulary as gram m atical inform ation in prefrontal
cortex. This could perhaps b e tested using cortical im aging technology to
study closed class w ord production.
In this section, three hypothetical explanations have been offered to
account for the fact that som e AD patients were consistently im paired in the
used of closed class vocabulary, while the majority were not. At this stage,
although I lean tow ard the first explanation for its simplicity, there is no
w ay to choose betw een these three hypotheses. Further exam inations of
error patterns in aphasics an d other language-im paired subjects for
relationships betw een closed class error use and sem antic im pairm ents may
eventually help to clarify this finding.
7.4 Sum m ary
This research has explored new territory in speech production
research. From a m ethodological view point, it has suggested that the
169
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
num ber of w ithin-sem antic-category errors in a large nam ing test
(approxim ately 100 item s) m ight act as a particularly good m easure of
sem antic im pairm ent, o r at least as a m easure of an underlying im pairm ent
that proportionately affects other speech production processes. In addition,
the constrained production task described in C hapter 6 was not only very
good at eliciting passive sentences, but also was extraordinarily difficult for
A lzheim er's patients w ho had prim ary deficits in either sem antics o r WM.
As such, this task m ight be very useful for testing the speech production
abilities of other im paired populations.
In addition to these m ethodological innovations, the results of this
study have also advanced our know ledge about the pervasive effects of
sem antic im pairm ent on the speech production abilities of A lzheim er's
patients. For several years AD speech production has been described as
being "preserved." H ow ever, this study has show n clearly that the integrity
of AD speech is strongly dependent on the speech task involved and on
how "preserved" is defined. If "preservation" refers only to norm al fluency
in spontaneous speech, then the results here su p p o rt this assertion— w hen
AD speakers w ere allow ed to choose their ow n w ords for an utterance,
fluency w as sim ilar to that of norm al control speakers. How ever, w hen
subjects' utterances w ere required to include specific w ords, as in the
Constrained P roduction task or even picture nam ing, fluency w as disrupted
in proportion to subjects' sem antic im pairm ent. O n the other hand, if
preservation is interpreted to m ean "nearly norm al," then the results of this
study again contradict this assertion, since it found that the num ber of AD
patients' lexical and gram m atical errors in spontaneous speech w as m uch
higher than norm al— effects that w ere also linked to sem antic im pairm ent.
170
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
M oreover, the accuracy and fluency of speech by these AD patients suffered
w hen they w ere required to include anything in their speech that
encouraged noncanonical sentence structures or w ord orders, w hich was
again an abnorm al response com pared to healthy controls. Interpreting
these results using the m odel presented above, it appears that the effects of
sem antic im pairm ent on the speech of Alzheimer's D isease patients are
profound and cut across m any of the traditional linguistic levels of
representation.
From a m ore theoretical vantage, the results of this study have
challenged the tw o m ost w ell-know n m odular theories of speech
production by docum enting significant relationships betw een sem antic
im pairm ent and the frequency of errors on aspects of language considered
by them to be purely syntactic. In particular, closed class w ords and syntactic
errors proved to be vulnerable w hen semantic ability w as im paired. To
account for these data w ithin m odular theories of speech production, one
w ould have to posit concurrent deficits in both sem antic and syntactic
processing, and attribute the correlational relationship betw een closed class
errors and sem antic im pairm ent to spurious correlations, w hile accepting
the relationship betw een open class errors and sem antic im pairm ent as
true. In order to avoid this type of inconsistent argum ent, the m odel
conceived to account for the current findings suggests th at closed class items
are semantically encoded (see C hapter 2), and that sentence production may
be more interactive than previously assumed. This m odel of speech
production incorporates W M as the fuel for activating sem antic
representations, and em ploys connectionist, Parallel D istributed Processing
principles to account for the structure and behavior of the sem antic
171
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
netw ork, w hile using constraint satisfaction as the guiding principle behind
sentence construction.
The novelty of m any of these findings suggests a wide range of
further studies. First, these findings need to be replicated both w ithin a set
of AD patients and hopefully also in aphasic patients with different types of
im pairm ents. For exam ple, different theories about the prim ary
im pairm ent in Broca's aphasia lead to very different predictions about how
this group m ight perform on the constrained production task. If their
im pairm ent lies prim arily in keeping an utterance in memory w hile
struggling to articulate it, this task m ight very well prove much easier for
them than spontaneous speech, although there m ight be problem s w ith the
irregular verb stim uli. However, if their im pairm ent lies in sentence
construction abilities and sentence planning, then they w ould have
considerable difficulty w ith this task across all verb types. On the other
hand, the results of this stu dy w ould predict that mild W ernicke's aphasics
should show sim ilar types of difficulties w ith the constrained production
task as the m ore sem antically im paired A lzheim er's patients, including
m any argum ent structure errors and closed class w ord omissions.
A nother interesting finding that should be pursued is the
relationship betw een W M and errors in spontaneous speech by healthy
speakers. This finding should be exam ined using a larger sam ple of healthy
elderly subjects, as w ell as younger subjects of varying WM ability, and,
perhaps even pathological populations w ith little semantic im pairm ent but
m ore severe WM im pairm ent. In these studies, particular im portance
should be placed on the use of closed class item s, since the current study has
172
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
suggested that dosed d a ss om issions m ay be espedally vulnerable to
om ission u n d e r conditions of low WM.
A nother set of findings from this study that m erits further
investigation is the different levels of difficulty that AD subjects had with
various verb types, espedally those verbs w ith distinct m orphological forms
for the sim ple past and past participle. The pervasiveness of difficulties
with this verb type was unexpected and not easily explained, and as such,
should be the topic of further research using other im paired populations as
well as AD patients of differing severity. In addition, it w ould be interesting
to exam ine the reaction time data of young and older norm al control
subjects in response to these different verb types to see w hether the patterns
revealed here in errors and dysfluendes persists. The results of studies like
these could shed considerable light on the underlying relationships between
differing inflectional forms of the sam e verb, and by extension, perhaps also
betw een derivationally related m orphological forms.
In sum m ary, the results of the current study have im portant
consequences for theories about sem antic representation and the effects of
sem antic im pairm ent on speech production. Furtherm ore, it m akes testable
predictions about speech production in other language-im paired
populations w hose investigation m ay eventually lead to a m ore
com prehensive understanding of speech production in general.
173
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithou t p erm issio n .
Appendix A
D em ographic inform ation on subjects w ho participated in the study.
Subject Education
Subject Sex Age
(XJSsl MMSE
M H AD f 79 16 21.5
MG AD f 78 12 22.0
BS AD f 83 16 20.0
MK AD m 64 16 19.0
AS AD m 87 20 21.0
HF AD m 82 12 22.5
MD AD f 89 14 20.0
H W AD m 83 13 19.5
KB AD f 84 16 18.0
MF AD f 76 16 21.0
PH N C m 66 16 30.0
RV N C m 77 17 29.0
BZ N C f 65 16 30.0
GE N C f 78 16 30.0
HM N C f 81 13 27.0
ES N C f 84 12 27.0
IS N C m 68 18 30.0
LS N C f 85 13 29.0
DR N C f 87 15 28.0
NS N C f 76 16 30.0
LG N C m 85 16 30.0
PH N C f 73 20 29.0
MM N C f 73 15 29.0
PM N C m 78 16 26.0
AM N C f 75 16 30.0
JO N C m 72 16 30.0
R ep r o d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
A ppendix B
Nam ing Errors of AD patients
Subject Target ResDonse Error tvpe
106 deer elk WITHIN CATEGORY
106 elbow knee WITHIN CATEGORY
106 flamingo crane WITHIN CATEGORY
106 rope a cord, no heavier WITHIN CATEGORY
106 sailboat ship WITHIN CATEGORY
106 screw hom No Response, visual
106 skunk not a racoon SUPERORDINATE/
CIRCUMLOCUTION
106 table stool or bench WITHIN CATEGORY
106 turkey peacock WITHIN CATEGORY
120 canoe boat SUPERORDINATE/
CIRCUMLOCUTION
120 celery NR No Response
120 deer EDK No Response
120 duck goose WITHIN CATEGORY
120 file cabinet a file of papers SUPERORDINATE/
CIRCUMLOCUTION
120 flamingo pelican? no. DDK. WITHIN CATEGORY
120 giraffe kangaroo WITHIN CATEGORY
120 goat EDK No Response
120 helicopter airplane WITHIN CATEGORY
120 hoe NR No Response
120 hummingbird bird SUPERORDINATE/
CIRCUMLOCUTION
120 jeep auto SUPERORDINATE/
CIRCUMLOCUTION
120 kangaroo DD K No Response
120 mitten glove WITHIN CATEGORY
120 mushroom lamp, candy dish No Response, visual
120 ostrich NR No Response
120 parakeet parrot WITHIN CATEGORY
120 sailboat ship, boat WITHIN CATEGORY
120 screw hom, you blow it No Response, visual
120 skunk that awful old thing No Response
120 squirrel animal up in a tree,
monkey?
WITHIN CATEGORY
120 strawberry fruit, you eat it SUPERORDINATE/
CIRCUMLOCUTION
120 tape measure what you measure with SUPERORDINATE/
CIRCUMLOCUTION
120 toe hand? no. WITHIN CATEGORY
120 trunk you put clothes in it for
travelling
SUPERORDINATE/
CIRCUMLOCUTION
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
120 wheelbarrow what you push in the
garden
SUPERORDINATE/
CIRCUMLOCUTION
120 wheelchair for someone who can’t
walk
SUPERORDINATE/
CIRCUMLOCUTION
120 woodpecker bird nest No Response
126 bus streetcar WITHIN CATEGORY
126 canoe boat SUPERORDINATE/
CIRCUMLOCUTION
126 crow robin WITHIN CATEGORY
126 dresser cabinet with 4 drawers WITHIN CATEGORY
126 file cabinet drawers SUPERORDINATE/
CIRCUMLOCUTION,
part/whole
126 flamingo ostrich WITHIN CATEGORY
126 giraffe zebra WITHIN CATEGORY
126 goat bull calf WITHIN CATEGORY
126 hand 5 fingers SUPERORDINATE/
CIRCUMLOCUTION.
part/whole
126 helicopter airplane WITHIN CATEGORY
126 jeep automobile, coupe WITHIN CATEGORY
126 kangaroo antelope WITHIN CATEGORY
126 motorcycle bicycle WITHIN CATEGORY
126 parakeet canary, parrot WITHIN CATEGORY
126 pelican related to a duck, not a
goose
SUPERORDINATE/
CIRCUMLOCUTION
126 penguin not an owl, not a
partridge
SUPERORDINATE/
CIRCUMLOCUTION
126 ship sailboat WITHIN CATEGORY
126 skunk smells bad SUPERORDINATE/
CIRCUMLOCUTION
126 swan goose or duck
swimming
WITHIN CATEGORY
126 turkey gander or goose WITHIN CATEGORY
126 vest jacket WITHIN CATEGORY
126 woodpecker it picks holes in trees SUPERORDINATE/
CIRCUMLOCUTION
131 arm leg WITHIN CATEGORY
131 file cabinet a file SUPERORDINATE/
CIRCUMLOCUTION
131 finger safety pin No Response, visual
131 flamingo saw a whole bunch
down at the Animal
Park in San Diego
SUPERORDINATE/
CIRCUMLOCUTION
131 jeep off-road vehicle SUPERORDINATE/
CIRCUMLOCUTION
131 knee NR No Response
131 leg heel WITHIN CATEGORY
131 lemon potato bug No Response, visual
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
131 mitten boxing glove WITHIN CATEGORY
131 nose ear WITHIN CATEGORY
131 shirt jacket WITHIN CATEGORY
131 toe water beetle No Response, visual
132 airplane NR No Response
132 arm foot WITHIN CATEGORY
132 asparagus branch No Response, visual
132 celery EDK No Response
132 deer a hom...what is it? No Response, visual
132 elbow hand WITHIN CATEGORY
132 flamingo giraffe, lamp No Response, visual
132 goat ram WITHIN CATEGORY
132 helicopter airplane WITHIN CATEGORY
132 hummingird bird SUPERORDINATE/
CIRCUMLOCUTION
132 jeep car SUPERORDINATE/
CIRCUMLOCUTION
132 leg foot WITHIN CATEGORY
132 parakeet bird SUPERORDINATE/
CIRCUMLOCUTION
132 sailboat a boat SUPERORDINATE/
CIRCUMLOCUTION
132 shirt jacket WITHIN CATEGORY
132 squirrel skwir (phon)
132 swan goose WITHIN CATEGORY
132 toe foot WITHIN CATEGORY
136 dresser drawers SUPERORDINATE/
CIRCUMLOCUTION.
part/whole
136 pelican penguin WITHIN CATEGORY
136 wheelchair invalid chair SUPERORDINATE/
CIRCUMLOCUTION
138 bench fence, sit on it in the
park
SUPERORDINATE/
CIRCUMLOCUTION
138 bookcase library books SUPERORDINATE/
CIRCUMLOCUTION,
part/whole
138 camel elephant WITHIN CATEGORY
138 canoe boat SUPERORDINATE/
CIRCUMLOCUTION
138 elbow hand WITHIN CATEGORY
138 file cabinet chest of drawers WITHIN CATEGORY
138 flamingo duck or swan WITHIN CATEGORY
138 giraffe elephant WITHIN CATEGORY
138 goat calf WITHIN CATEGORY
138 hoe spade WITHIN CATEGORY
138 hummingbird IDK what kind of bird
it is
SUPERORDINATE/
CIRCUMLOCUTION
138 kangaroo long tail fox or bear WITHIN CATEGORY
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
138 knee IDK No Response
138 lemon fish No Response, visual
138 lips lampshade, pillow No Response, visual
138 mushroom lamp No Response, visual
138 nose IDK No Response
138 ostrich chicken or turkey WITHIN CATEGORY
138 parakeet bird SUPERORDINATE/
CIRCUMLOCUTION
138 pelican duck WITHIN CATEGORY
138 penguin bird standing up SUPERORDINATE/
CIRCUMLOCUTION
138 rope wreath No Response, visual
138 screw hom No Response, visual
138 ship sailboat WITHIN CATEGORY
138 skunk squirrel WITHIN CATEGORY
138 toe foot, toenail WITHIN CATEGORY
138 truck van WITHIN CATEGORY
138 trunk to travel with SUPERORDINATE/
CIRCUMLOCUTION,
138 wheelbarrow shovel SUPERORDINATE/
CIRCUMLOCUTION
138 woodpecker bird SUPERORDINATE/
CIRCUMLOCUTION
141 bus streetcar ^ WITHIN CATEGORY
141 celery vase of flowers No Response, visual
141 chicken turkey WITHIN CATEGORY
141 crow sparrow 1 WITHIN CATEGORY
141 duck pelican WITHIN CATEGORY
141 ear womb No Response, visual
141 file cabinet drawers SUPERORDINATE/
CIRCUMLOCUTION,
part/whole
141 flamingo ostrich WITHIN CATEGORY
141 hummingbird sparrow WITHIN CATEGORY
141 knee person’s leg WITHIN CATEGORY
141 lips DDK No Response
141 mitten slipper WITHIN CATEGORY
141 parakeet sparrow WITHIN CATEGORY
141 rake hoe WITHIN CATEGORY
146 asparagus carrot WITHIN CATEGORY
146 belt buckle SUPERORDINATE/
CIRCUMLOCUTION,
part/whole
146 desk dresser WITHIN CATEGORY
146 file cabinet drawers SUPERORDINATE/
CIRCUMLOCUTION,
part/whole
146 finger N/R No Response
146 flamingo stork WITHIN CATEGORY
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
146 hand 5 fingers WITHIN CATEGORY
146 helicopter plane WITHIN CATEGORY
146 hummingbird blackbird WITHIN CATEGORY
146 mushroom toad No Response (visual?)
146 parakeet bird SUPERORDINATE/
CIRCUMLOCUTION
146 pelican bird SUPERORDINATE/
CIRCUMLOCUTION
146 penguin parrot WITHIN CATEGORY
146 sailboat ship WITHIN CATEGORY
146 screw hom No Response, visual
146 strawberry pineapple WITHIN CATEGORY
146 swan duck WITHIN CATEGORY
146 tape measure measuring stick WITHIN CATEGORY
146 toe fingers WITHIN CATEGORY
146 vest jacket WITHIN CATEGORY
146 woodpecker bird SUPERORDINATE/
CIRCUMLOCUTION
147 canoe boat SUPERORDINATE/
CIRCUMLOCUTION
147 jeep truck without a top WITHIN CATEGORY
147 toe toenail, foot WITHIN CATEGORY
147 truck big van or bus WITHIN CATEGORY
147 woodpecker crow WITHIN CATEGORY
KEY
WITHIN CATEGORY: Closely related substitution (within sem antic category
substitution)
SUPERORDINATE/CIRCUMLOCUTION: Some semantic know ledge
(superordinate term , circum locution, p a rt/w h o le
substitution)
N o Response: No Response semantic know ledge (IDK, NR, visual errors,
affective comments)
NR: No Response response.
IDK: I d on't know.
179
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Appendix C
The num ber of raw errors by group, task, error category and type.
A D (n=10)i Closed Class O pen Class P ro n o u n s1 T otal
M orpho- |
Syntax
PN:
O m issions 16 9 3 28
PN: i
Substitutions 7 11
(9 reference)
4
(9)
22
15
CP:
O m issions 49 0 0 49
CP: |
Substitutions | 7 3 1 11
6
TOTALi 79 23 17 119 21
N C (n=16)j Closed Class
O pen Class
- |
|
P ronouns j Total
M orpho- |
Syntax
PN:
O m issions 11 4
I
0 !
15
PN:
.
Substitutions : 2
.
(2 reference) |
i i
(2)
7
10
CP:
O m issions 5 1
i
0 !
6
I
CP:
Substitutions 4
4
i
i
9
8
TOTAL 22 13 4 !
i
39 18
Subject S PN PN CP CP i
Type | Fresh Starts R estarts Fresh Starts Restarts i
AD ! 69 31 84 27
NC 74 67 1 32
CP Misused Stim: AD-56, NC-5
lThe numbers of pronouns with no discernible referent are shown in parentheses.
180
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm ission .
A ppendix D
C onstrained P roduction Stim uli
REG (regular! Verb Stimuli Tod N oun Bottom N oun
bum ped Liz M ary
repeated Jim Gary
avoided Jane R u th
carried Becky Jill
found M ike Susie
stopped Greg Bob
taught history S usan
delayed m eetin g Frank
loved cookies Lynn
caused u p ro ar David
poured m ilk Betty
typed letter Jo h n
clim bed Debbie ladder
spilled T om juice
lost Jim gam e
w arm ed Sally bread
studied Rick notes
explored Sam cave
IRREG (irregular! Verb Stimuli Tod N o u n Bottom N oun
d riv en M aggie Frank
shaken A lbert Joan
beaten D on C harlie
draw n A m y Lisa
forgotten Cathy Joe
seen Larry M ike
eaten Bill cake
taken
D iane pill
broken* A lbert toy
chosen Gary restau ran t
h idd en F ran candy
stolen H enry stereo
to rn
pants Jack
w oven cloth Gail
w ritten book Peter
show n m o v ie Paul
th ro w n plate Lisa
grow n roses Pam
181
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
T-E (them e-experiencer) Stimuli Top N oun_________Bottom N oun
depressed
am used
annoyed*
bored
tem pted
confused
shocked
excited*
thrilled
terrified
inspired
frightened
disgusted
baffled
enraged
delighted
surprised
startled
Practice:
h it
played
Interested
given
182
Ken
N ancy
N ed
Paula
Dan
Laura
Sarah
Randy
Betsy
Robert
K aren
Grace
n o v e l
m achin e
program
gift
new s
noise
Lois
K evin
Kelly
Scott
Terry
M ark
m o v ie
trip
m usic
fire
serm o n
th u n d e r
Linda
Jeff
Ben
A n n
D onna
R on
ball
Leah
history
book
T om m y
cards
Ellen
M olly
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Appendix E
CP responses with Dysfluendes (underlined). 2
Sub.# S tim u lu s Response
120 shaken A rth u r Joan Hm m , loan. O h gosh. I d o n 't know. Jo a n . . . I
can't seem to m ake anything out of it. A rthur
and Joan w ere shaken.
120 show n m ovie Paul Paul show n the m ovie rr, no. Paul liked to . . .
Paul. Paul liked the m ovie that was shown.
120 throw n plate Lisa
120 repeated Jim Gary
120 w oven cloth Gail
106 grow n roses Pam
Lisa plate. I d o n 't know w h a t's... Lisa, um.
Lisa throw n the plate rr, that doesn't make
sense. Lisa . . . TTie plate was thrown to Lisa.
Um, that isn't repeated, is it? Yeah, that is
repeated. Tim and G ary repeated— no. Tim
helped, no. Gary repeated for Jim rr. Gary
repeated a- I d o n 't know.
um, woven Gail a? W oven cloth a? I don't
know. Cloth. G ail. Cloth was woven by Gail.
Pam grew . . . The roses th at Pam grew were -
no that's wrong...Pam grew beau- no that's
also, not right. The roses that were grown by
Pam were beautiful.
106 explored Sam cave They found m uch to ex a— no ? Sam found
very — no, that's n o t good either. Sam explored
the cave. Stupid sentence.
138 lost Jim gam e Tim . . . the gam e. .. had lo s t. .. Jim was upset
because he lost the game.
141 thrilled Betsy m usic Betsy thrilled^ .. er, they. Music thrilled Betty,
uh BETSY!
146 terrified Robert fire Robert was terrified— Rob terrified the fire rr,
Rob was terrified by the fire.
2 Also includes Rejected Responses marked rr).
183
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
Appendix F
CP responses show ing Rejected Responses (underlined)
Sub.# S tim u lu s Response
120 forgotten Cathy Joe Cathy, C athy forgotten about Toe. No. Joe had
forgotten Cathy.
120 stopped G reg Bob
120 chosen G ary
restau ran t
120 disgusted novel
Linda
146 shaken A rth u r Joan
146 terrified Robert fire
131 grow n roses Pam
131 taken D iane pill
131 thrilled Betsy music
132 annoyed N ed Kelly
Greg, G reg stopped Bob, no that's not any good.
Bob stopped Greg.
Gary um. I don't know. Gary chosen the
restaurant, no. I don't know! The restaurant
was chosen by Gary.
Linda, let's see, Linda disgusted novel, um . I'm
trying to. . . I guess I'm mixed u p on this one.
Linda w as disgusted with the novel.
A rthur um . . . A rthur claim ed lo an . Joanne,
A rthur had shaken Joanne.
Robert w as terrified — Rob terrified the fire. Rob
was terrified by the fire.
The roses w ere prickelv once Ben w alked
through the garden . . . Pam grew roses this past
year.
Diane, uh, Diane played music on the piano at
W illow Lake last week. .. OK, Diane ate so
m any pills that she rattled as she walked.
Betsy m ade some brownies. .. Oh. Betsy was
thrilled by the Sousa band music.
Either N ed or Kelly annoyed each other. Kelly
annoyed N ed, and Ned annoyed Kelly.
184
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Appendix G
CP responses showing difficulty including closed class words.
Subr# Stimulus Response
126 grown roses Pam Pam , no. was...Roses grow n by Pam.
126 explored Sam cave Cave explored by Sam.
126 taken Diane pill Pill was taken by Diane.
126 delayed m eeting M eeting was de- delayed by Frank.
Frank
126 throw n plate Lisa Lisa throw n plate, I guess.
126 eaten Bill cake I can't make sense o u t of it. Cake. Bill eaten
cake
126 delighted gift A nn A nn delighted gift. T hat doesn't make sense
either. Gift delighted Ann.
126 excited Randy trip Trip excited Randy.
132 shown m ovie P aul Paul show- show n the m ovie.
132 broken Albert toy Albert's toy broken
132 beaten Don C harlie Don beaten by Charlie, or Charlie - no, it has to
be Don beaten by Charlie.
132 stolen H enry stereo OK, Stereo was stolen by Henry.
141 explored Sam cave Sam explored cave.
141 taken Diane pill Diane .. taken pill.
141 delayed m eeting M eeting delayed Frank.
Frank
141 show n m ovie P aul Paul sh- show n m ovie
141 chosen Gary G ary had chosen restaurant,
restaurant
141 thrown plate Lisa Plate was throw n at Lisa.
185
R ep ro d u ced with p erm issio n o f th e copyright ow ner. Further reproduction prohibited w ithout p erm issio n .
Appendix H
CP responses that have stimulus substitutions or omissions
Sub.# Stimulus
126 hidden Fran candy
141 beaten Don Charlie
146 enraged program Ben
146 grown roses Pam
146 taken Diane oill
Rgpgnse
Candy is, candy in Fra- Fran.
Charlie, um. Don beat Charlie in the race.
Ben made a program.
Pam grew roses.
The pill, Diane took the pill.
146 chosen Gary restaurant The restaurant, Gary chold- chose the restaurant.
146 hidden Fran candy Frank ate the candy.
Lisa threw the plate.
Peter wrote a book.
Henry stole the stereo.
146 thrown plate Lisa
146 written book Peter
146 stolen Henry stereo
131 forgotten Cathy Joe
131 shaken Arthur Joan
131 amused Nancy Kevin
131 frightened Grace
thunder
131 thrown plate Lisa
131 written book Peter
131 caused uproar David
1 3 1 stolen Henry stereo
131 repeated Jim Gary
Cathy uh forgot that she had a cake in the oven that she
was cooking, baking for Joe.
Arthur was shaken when he fell down on the basketball
court.
Nancy and Kevin got all excited when they saw the
parachuters fall off the bridge.
Grace was exhausted after she ran away from the
tornado, and there was lots of loud thunder, it
frightened her.
A plate came smashing through the window of Lisa's
car.
The book submitted by Peter was deemed to be
plagiarized.
David said that the peanuts were too salty. David
created havoc. (I don’t believe I can make short
sentences.)
Henry listened to his stir- stereo.
Jim repeated the bouncing ball trick.
186
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
187
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
REFERENCES
Albert, M. & Milburg, W. (1989). Semantic processing in patients with Alzheimer’s
Disease. Brain and Language, 37, 163-171.
Allen, J. (1995). Argument structures and lexical entries: A connectionist model of
verb class acquisition. Unpublished manuscript, University of Southern
California.
Almor, A., Kempler, D., MacDonald, M. C. & Andersen, E. S. (1997). Discourse
comprehension in Alzheimer’s Disease. Brain and Language, 60(I), 185-188.
Altmann, L. P., Andersen, E. S. & Kempler, D. (1993). Re-evaluating syntactic
preservation in Alzheimer’s Disease. A study presented at the 1993 Academy
of Aphasia, Tucson, Arizona, October (1993).
Altmann, L. P., Andersen, E. S. & Kempler, D. (1995). Open and Closed Class Speech
Errors in Alzheimer’s Disease. A poster presented at 1995 annual meeting of the
International Neuropsychological Society, Seattle, WA. February 9, (1995).
JINS, 1(2), 148.
Altmann, L. P., Kempler, D. & Andersen, E. S., (1997). Working Memory tasks and
offline semantic tasks: Are they measuring the same thing? JINS, J( 1).
Anderson, J. R. (1983). The architecture o f cognition. Cambridge, MA: Harvard
University Press.
Baddeley, A. D. (1990). Human Memory: Theory and Practice. Boston. MA: Allyn
and Bacon.
Baddeley, A. D. & Hitch, G. J. (1974). Working memory, in G. Bower (Ed.), The
psychology o f learning and motivation (Vol.8). (pp. 47-90). New York,:
Academic Press.
Balota, D. A. & Duchek, J. M. (1988). Age-related differences in lexical access.
spreading activation, and simple pronunciation. Psychology and Aging, 4, 84-93.
Bates, E., Friederici, A. & Wulfeck, B. (1987). Comprehension in Aphasia: A cross-
linguistic study. Brain and Language, 32, 19-67.
Bates, E., Marchman, V., Harris, C. & Wulfeck, B. (1994). Production of complex
syntax in normal aging and Alzheimer’s disease. Language and Cognitive
Processes.
Bates, E. & Wulfeck, B. (1989). Crosslinguistic studies of aphasia, in B. MacWhinney
and E. Bates (Eds.), The crosslinguistic study o f sentence processing, (pp. 328-
371). Cambridge, UK: Cambridge University Press.
Bayles, K. A. (1982). Language function in senile dementia. Brain and Language, 16.
265-280.
187
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Benton, S. L., Kraft, R. G., Glover, J. A. & Plake, B. S. (1984). Cognitive capacity
differences among writers. Journal o f Educational Psychology, 76, 820-834.
Bock, J. K. (1986). Syntactic persistence in language production. Cognitive
Psychology, 18, 355-387.
Bock, J. K. (1989). Closed class immanence in sentence production. Cognition, 31.
163-186.
Bock, J.K. & Loebell, H. (1990). Framing sentences. Cognition, 35, 1-39.
Bock, K., Loebell, H. & Morey, R. (1992). From conceptual roles to structural relations:
Bridging the syntactic cleft. Psychological Review, 99 (I), 150-171.
Blumstein, Sheila E. (1988). Chapter 8: Neurolinguistics: An overview of language-
brain relations in aphasia, in Newmeyer, F.J. (Ed.), Linguistics: The Cambridge
Survey, vol. 3, Language: Psychological and Biological Aspects. Cambridge,
UK: Cambridge University Press, pp. 210-236.
Bradley, D. C., & Garrett, M. F. (1983). Hemisphere differences in the recognition of
open class words. Neuropsychologica,2l(2), 155-159.
Burke, D.M., MacKay, D.G., Worthley, J.S., and Wade, E. (1991). On the tip of the
tongue: What causes word finding failures in young and older adults? Journal
o f Memory and Language, 30, 542-579.
Butterworth, B. (1989). Lexical access in speech production, in W. Marslen-Wilson
(Ed.), Lexical representation and process, (pp. 108-135). Cambridge, MA: The
MIT Press.
Butterworth, B. (1992). Disorders of phonological encoding, in W.J.M. Levelt (ed.).
Lexical access in speech production, (pp. 261-286). Cambridge, MA: Basil
Blackwell.
Butterworth, B. & Howard, D. (1987). Paragrammatisms. Cognition, 26, 1-37.
Cantor, J., & Engle, R. W. (1993). Working memory capacity as long-term memory
activation: An individual differences approach. Journal o f Experimental
Psychology: Learning, Memory and Cognition, 19 (5), 1101-1114.
Caplan, D. (1992). Language: Structure, processing and disorders. Cambridge, MA:
The MIT Press.
Caplan, D., Baker, C., & Dehaut, F. (1985). Syntactic determinants of sentence
comprehension in aphasia. Cognition, 21. 117-175.
Caplan, D. & Hildebrandt, N. (1988). Disorders o f syntactic comprehension.
Cambridge, MA: The MIT Press.
Carramazza, A. & Zurif, E. (1976). Dissociation of algorithmic and heuristic processes
in language comprehension: Evidence from aphasia. Brain and Language, 3,
572-582.
188
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Cazden, C. B. (1968). The acquisition of noun and verb inflections. Child
Development, 39, 433-448.
Cherry, B. J., Buckwalter, J. G. & Henderson, V .W. (1996). Memory span procedures
in Alzheimer’s disease. Neuropsychology, 10(2), 286-293.
Chertkow, H. & Bub, D. (1990). Semantic memory loss in Alzheimer-type dementia,
in Schwartz, M. (ed.), Modular deficits in Alzheimer-type dementia, (pp. 207-
244). Cambridge, MA: The M TT Press.
Daneman, M. (1991). Working memory as a predictor o f verbal fluency. Journal of
Psycholinguistic Research, 20 (6), 445-464.
Daneman, M. & Carpenter, P. A. (1980). Individual differences in working memory
and reading. Journal o f Verbal Learning and Verbal Behavior, 19, 450-466.
Daneman, M. and Carpenter, P.A. (1983). Individual differences in integrating
information between and within sentences. Journal o f Experimental Psychology:
Learning, Memory and Cognition, 9, 561-583.
Daneman, M. & Green, I. (1986). Individual differences in comprehending and
producing words in context. Journal of Memory and Language, 25, 1-18.
Davis, G. A. (1983). A survey of adult aphasia. Englewood Cliffs, NJ: Prentice-Hall.
Deacon, T. W. (1997). The Symbolic Species: The co-evolution of language and the
brain. New York: Norton.
Dell, G. S. (1990). Effects of frequency and vocabulary type on phonological speech
errors. Language and Cognitive Processes, 5, 313-349.
Dell, G.S., Juliano, C., and Govindjee, A. (1993). Structure and content in language
production: A theory of frame constraints in phonological speech errors.
Cognitive Science, 17. 149-195.
Dell, G. S. & O’Seaghdha, P. G. (1992). Stages of lexical access in language
production, in W.M.J. Levelt (ed.), Lexical access in speech production, (pp.
287-314). Cambridge, MA: Blackwell.
Devlin, J.T., Gonnerman, L.M., Andersen, E.S. & Seidenberg, M.S. (1998) Category
specific semantic deficits in focal and widespread brain damage: A
computational account. Journal o f Cognitive Neuroscience.
Done, D.J. & Gale, T.M. (1997). Attribute verification in dementia of Alzheimer’s
type: Evidence for the preservation of distributed concept knowledge. Cognitive
Neuropsychology, 14(4). 547-571.
Duchek, J. M., & Balota, D. A. (1993). Sparing activation processes in older adults, in
J. Cerella, J. Rybash, W. Hoyer, & M. L. Commons (Eds.), Adult Information
Processing: Limits on Loss. (pp. 383-405). New York: Academic Press.
189
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Elman, J. L., Bates, E. A., Johnson, M. H., Karmiloff-Smith, A., Parisi, D. & Plunkett,
K. (1996). Rethinking innateness: A connectionist perspective on development.
Cambridge, MA: MIT Press.
Engle, R. W. (1996). Working memory and retrieval: An inhibition-resource approach,
in Richardson, J. T. E., Engle, R. W., Hasher, L., Logie, R. H., Stoltzfus, E. R. &
Zacks, R. T., Working memory and human cognition, (pp. 89-119). New York:
Oxford University Press.
Engle, R. W„ Cantor, J. & Carullo, J. J. (1992). Individual differences in working
memory and comprehension: A test of four hypotheses. Journal of
Experimental Psychology: Learning, Memory and Cognition, 18 (5), 972-992.
Ervin, S. M. (1964). Imitation and structural change in children’s language. In E. H.
Lenneberg (Ed.), New directions in the study of language. Cambridge: MIT
Press.
Farah, M. J. & McClelland, J. (1991). A computational model of semantic memory
impairment: Modality specificity and emergent category specificity. Journal of
Experimental Psychology: General, 120, 339-357.
Ferreira, F. (1994). Choice of passive voice is affected by verb type and animacy.
Journal o f Memory and Language, 33, 715-736.
Folstein, M. F., Folstein, S. E., & McHugh, P. R. (1975). Mini-Mental State: A
practical method for grading the cognitive state of patients for the clinician.
Journal o f Psychiatric Research, 12, 189-198.
Francis, W. N. & Kucera, H. (1982). Frequency analysis of English usage: Lexicon
and grammar. Boston: Houghton-Mifflin.
Garrett, M. F. (1975). The analysis of sentence production, in G. Bower (Ed.).
Psychology o f learning and motivation: vol. 9. (pp. 137-177). Academic Press:
New York.
Garrett, M. F. (1980). Levels of processing in sentence production, in Brian
Butterworth (Ed.). Language production. Volume 1: Speech and talk. (pp. 177-
220). London: Academic Press.
Garrett, M. F. (1982). Production of speech: Observations from normal and
pathological language use. in A.W. Ellis (Ed.). Normality and pathology in
cognitive junction. (pp. 19-75). London: Academic Press.
Garrett, M. F. (1984). The organization of processing structure for language
production: Applications to Aphasia Speech, in D.Z Caplan, A. R. Lecours and
A. Smith (Eds.). Biological perspectives o f language, (pp. 172-193).
Cambridge, MA: MIT Press.
Garrett, M. F. (1992). Disorders of lexical selection, in W.M.J. Levelt (Ed.), Lexical
Access in speech production, (pp. 143-180). Cambridge, MA: Blackwell.
Gathercole, S. E. & Baddeley, A. D. (1993). Working memory and language.
Hillsdale, NJ: Erlbaum.
190
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .
Gonnerman, L. M., Andersen, E. S., Devlin, J. T., Kempler, D., & Seidenberg, M. S.
(1997). Double dissociation of semantic categories in Alzheimer’s disease.
Brain and Language, 57, 254-279_
Goodglass, H. & Menn, L. (1985). Is agrammatism a unitary phenomenon? In M. L.
Kean (Ed.), Agrammatism. New York: Academic Press, p. 1-26.
Gordon, B. & Caramazza, A. (1982). Lexical decision for open-and closed-class
words: Failure to replicate differential frequency sensitivity. Brain and
Language, 15, 143-160.
Gordon, B. & Caramazza, A. (1983). Lexical decision for open- and closed-class
words: Failure to replicate frequency sensitivity. Brain and Language, 15, 143-
160.
Green, R. (1995). Alzheimer’s Disease and other dementing disorders in adults, in R. J.
Joynt (Ed.), Clinical Neurology. Philadelphia: Lippincott-Raven.
Green, S. B., McKoon, G., & Ratcliff, R. (1992). Pronoun resolution and discourse
models. Journal of Experimental Psychology: Learning, Memory and
Cognition, 18(2), 266-283.
Heeschen, C. (1985). Agrammatism vs. paragrammatism: A fictitious opposition. In
M.L. Kean (Ed.), Agrammatism.. London: Academic Press.
Henderson, V. W„ Kratz, K. E., Smith, C. A., & Buckwalter, J. G. (1995). Working
Memory and Language Comprehension in Alzheimer’s Disease. Brain and
Language, 51(1), 142-144.
Hier, D.B., Hagenlocker, K., & Shindler, A.G. (1985). Language disintegration in
dementia: Effects of etiology and severity. Brain and Language, 25, 117-133.
Hirshman, E. & Palij, M. (1992). Rehearsal and the word frequency effect in
recognition memory. Journal o f Memory and Language, 31, 477-484.
Hodges, J. R., Salmon, D., & Butters, N. (1991). The nature of the naming deficit in
Alzheimer’s and Huntington’s disease: The role of attention and memory.
Brain, 114, 1547-1558.
Hodges, J. R„ Patterson, K., Oxbuiy, S., & Funnell, E. (1992). Semantic dementia:
Prgressive fluent aphasia with temporal lobe atrophy. Brain, 115, 1783-1806.
Hofstede, B. T. M. & Kolk, H. H. J. (1994). The effects of task variation on the
production of grammatical morphology in Broca’s aphasia: A multiple case
study. Brain and Language, 46, 278-328.
Huff, F. J., Corkin, S. & Growdon, J. H. (1986). Semantic impairment and anomia in
Alzhiemer’s disease. Brain and Language, 28, 235-249.
Jorm, A. F. (1986). Controlled and automatic information processing in senile
dementia: A review. Psychological Medicine, 16, 77-88.
191
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Just, M. A. & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual
differences in working memory. Psychological Review, 99, 122-149.
Kempler, D. (1984). Syntactic and symbolic functions in dementia o f the Alzheimer’ s
type, unpublished doctoral dissertation. University of Califoria. Los Angeles.
Kempler, D. (1991). Language changes in dementia of the Alzheimer type. Lubinski,
(Ed.). Dementia and Communication: Research and Clinical Implications.
New York: Decker Publishing.
Kempler, D., Curtiss, D., & Jackson, C. (1987). Syntactic preservation in Alzheimer’ s
Disease. Journal o f Speech and Hearing Research, 30, 343-350.
Kempler, D. & Zelinski, E. M. (1994). Language in dementia and normal aging, in
F.a. Huppert, C. Brayne, & D.W. O’ Connor, Dementia and normal aging, (pp.
331-365). Cambridge: Cambridge University Press.
King, J. & Just, M. A. (1991). Individual differences in syntactic processing: The role
of working memory. Journal o f Memory and Language, 30, 580-602.
Kolk, H. (1995). A time-based approach to agrammatic production. Brain and
Language, 50, 282-303.
Kopelman, M.D. (1994). Working memory in the amnesic syndrome and degenerative
dementia. Neuropsychology, 8 (4), 555-562.
LaBerge, David. (1995). Attentional Processing: The Brain’ s Art of Mindfulness.
Cambridge, MA: Harvard University Press.
LaVoie, D., Andersen. E.S., Kempler, D., & Seidenberg, M. (1996) Maintenance of
Semantic Featural Knowledge in Alzheimer’ s Disease. Presented at the
International Neuropsychological Society, February, Chicago, 1996. Journal of
the International Neuropsychological Society, Vol. 2., No. 1, p 46.
Levelt, W. J. M. (1983). Monitoring and self-repair in speech. Cognition, 74(1), 41-104.
Levelt, Willem J. M. (1989). Speaking: From Intention to Articulation. Cambridge.
MA.: The MIT Press.
Levelt,W. J. M., Schriefers, H., Vorberg, D., Meyer, A. S., Pechmann, T. & Havinga, J.
(1991). The time course of lexical access in speech production: A study of
picture naming. Psychological Review, 98(1), 122-142.
Levin, B. (1993). English Verb Classes and Alternations: A Preliminary Investigation.
Chicago, E L : The University of Chicago Press.
Levin, B. & Pinker, S. (1991). Introduction, in B. Levin & S. Pinker (Eds.). Lexical
and Conceptual Semantics, (pp. 1-8). Cambridge, MA: Blackwell.
Lewis-Beck, M. S. (1980). Applied Regression: An Introduction. Newbury Park:
Sage Publications.
192
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Light, L. L. (1993). Language changes in old age. in G. Blanken, J. Dittmann, H.
Grimm, J. C. Marshall, & C-W. Wallesch (Eds.), Linguistic Disorders and
Pathologies: An international handbook. Berlin: Walter de Gruyter.
Light, L. L. (1994). Language and aging: Competence versus performance. inJ. F.
Birren & V. L. Bengtson (Eds.). Handbook of Theories o f Aging. New York:
Springer.
Lyons, K., Kemper, S., LaBarge, E., Ferraro, F. R., Balota, D., & Storand, M. (1994).
Oral language and Alzheimer’s Disease: A reduction in syntactic complexity.
Aging and Cognition, 1(4), 271-281.
MacRae, K., Spivey-Knowlton, M. J., & Tanenhaus, M. K. (1998). Modeling the
influence of thematic fit (and other constraints) in on-line sentence
comprehension. Journal o f Memory and Language, 38, 283-312.
MacDonald, M. C. (1989). Priming effects from gaps to antecedents. Language and
Cognitive Processes, 4, 35-56.
MacDonald, M. C. (1993). The interaction of lexical and syntactic ambiguity. Journal
o f Memory and Language, 32, 692-715.
MacDonald, M. C. & Christiansen, M. (submitted). Language comprehension without
working memory: A reply to Just 8c Carpenter and Waters & Caplan.
MacDonald, M. C., Just, M. A., & Carpenter, P. A. (1992). Working memory
constraints on the processing of syntactic ambiguity. Cognitive Psvchologv. 24,
56-98.
MacDonald, M. C. & MacWhinney, B. (1990). Measuring inhibition and facilitation
from pronouns. Journal o f Memory and Language, 29(6), 469-492.
MacDonald, M. C., Pearlmutter, N. J. & Seidenberg, M. S. (1994). The lexical nature
of syntactic ambiguity resolution. Psychological Review, 101, 676-703.
MacKay, D.G. 1987. The organization o f perception and action: A theory for
language and other cognitve skills. New York: Springer-Verlag.
MacWhinney, B. & Bates, E. (eds.). (1989). The crosslinguistic study o f sentence
processing. Cambridge, UK: Cambridge University Press.
Martin, A. (1990). Neuopsychology of Alzheimer’s Disease: The case for sub-groups,
in Myma F. Schwartz (Ed.), Modular deficits in Alzheimer-type dementia, (pp.
143-176). Cambridge, MA: The MIT Press.
Martin, N. & Saffran, E. M. (1996). Evidence for a common impairment affecting
lexical processing and short-term memory in aphasia. Brain and Language, 55
(1), 191-194.
Martin, N., Saffran, E. M. & Dell, G. S. (1996). Recovery in deep dysphasia: Evidence
for a relation between auditory-verbal STM and lexical errors in repetition.
Brain and Language, 52, 82-113.
193
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Martin, R. C. & Feher, E. (1990). The consequences of reduced memory span for the
comprehension of semantic versus syntactic information. Brain and Language,
38, 1-20.
Masliah, E., Hansen, L., Albright, T., Mallory, M. & Terry, R. D. (1991).
Immunoelectron microscopic study of synaptic pathology in Alzheimer’s
disease. Acta Neuropathologica, 8 1 ,428-433.
McClelland, J. L., Rumelhart, D. E„ & Hinton, G. E. (1989). The appeal of parallel
distributed processing, in D. E. Rumelhart, J. L. McClelland and the PDP
Research Group, Parallel distributed processing: Explorations in the
microstructure o f cognition, vol. 1: Foundations, (pp. 3-44). Cambridge, MA:
The MIT Press.
McClelland, J. L., Rumelhart, D. E., and the PDP Research Group. (1989). Parallel
distributed processing: Explorations in the microstructure of cognition, vol. I:
Foundations, Volume 2. Cambridge, MA: The MIT Press.
McDonald, J. L., Bock, K. & Kelly, M. H. (1993). Word and world order: Semantic,
phonological, & metrical determinants of serial position. Cognitive Psychologv,
25, 188-230.
McKhann, G., Drachman, D., Folstein, M., Katz man, R„ Price, D., Stadlan, E. M.
(1984). Clinical diagnosis of Alzheimer’s Disease: Report of the NINCDS-
ADRDA work group under the auspices of Health and Human Services Task
Gorce on Alzheimer’s Disease. Neurology, 34: 939-944.
McNamara, P., Obler, L. K., Au, R., Durso, R. & Albert, M. L. (1992). Speech
monitoring skills in Alzheimer’s Disease, Parkinson’s Disease, & normal aging.
Brain and Language, 42, 38-41.
Menn, L. & Obler, L. (1990). Cross-language data and theories of agrammatism. In L.
Menn and L. K. Obler (Eds.), Agrammatic aphasia: A cross-language narrative
sourcebook. Philadelphia: John Benjamins. 369-390.
Menn, L., Reilly, K.F., Hayashi, M., Kamio, A., Fujita, I. & Sasanuma. S. (1998). The
interaction of preserved pragmatics and impaired syntax in Japanese and English
aphasic speech. Brain and Language, 61 (2), 183-225.
Miyake, A., Carpenter, P. A. & Just, M. A. (1995). A capacity approach to syntactic
comprehension disorders: Making normal adults perform like aphasic patients.
Cognitive Neuropsychology, 11 (6), 671-717.
Moskovitch, M. (1994). Cognitive resources and dual-task interference effects at
retrieval in normal people: The role of the frontal lobes and medial temporal
cortex. Psychological Bulletin, 106(3), 524-534.
Motley, M. T. & Baars, B. J. (1976). Laboratory induction of verbal slips: Anew
method for psycholinguistic research. Communication Quarterly, 24, 28-34.
Nebes, R. D. (1989). Semantic memory in Alzheimer’s Disease. Psychological
Bulletin, 106(3), 377-394.
194
R ep ro d u ced with p erm issio n o f th e copyrigh t ow ner. Further reproduction prohibited w ithout p erm issio n .
Nebes, R.D. (1992). Cognitive dysfunction in Alzheimer’s Disease. In F.I.M. Craik
and T.A. Salthouse (Eds.), The handbook of aging and cognition. Hillsdale,
N.J.: Erlbaum.
Nebes, R. D. & Brady, C. B. (1989). Integrity of semantic fields in Alzheimer’s
Disease. Cortex, 24, 291-299.
Nebes, R.D. & Madden, D.J. (1988). Different patterns of cognitive slowing produced
by Alzheimer’s Disease and normal aging. Psychology and Aging, 3. 102-104.
Nebes, R. D., Martin, D. C. & Horn, L. C. (1984). Sparing of semantic memory in
Alzheimer’s disease. Journal o f Abnormal Psychology, 93, 321-330.
Nicholas, M., Obler, L., Albert, M., & Helm-Estabrooks, N. (1985). Empty speech in
Alzheimer’s disease and fluent aphasia. Journal o f Speech and Hearing
Research, 28, 405-410.
Nicholas, M., Obler, L. A., Au, R. & Albert, M. L. (1996). On the nature of naming
errors in aging and dementia: A study of semantic relatedness. Brain and
Language, 54 (2), 184-195.
Oakhill, J., Yuill, N. & Parkin, A. (1989). On the nature of the difference between
skilled and less-skilled comprehenders. Jounal of Research in Reading, 9, 80-
91.
Ober, B. A. & Shenaut, G. K. (1995). Semantic priming in Alzheimer’s disease: Meta
analysis and theoretical evaluation, in P. Allen and T.R. Bashore (Eds.). Age
differences in word and language processing, (pp. 247-271). Amsterdam:
Elsevier.
Ormrod, J. E. & Cochran, K. F. (1988). Relationship of verbal ability and working
memory to spelling achievement and learning to spell. Reading Research and
Instruction, 28, 33-48.
Parasuranam, R. & Haxby, J. V. (1993). Attention and brain function in Alzheimer’s
Disease: A review. Neuropsychology, 7(3), 242-272.
Patel, P. G. & Satz, P. (1994). The language production system and senile dementia of
the Alzheimer’s type: Neuropathological implications. Aphasiology, 8(1), 1-18.
Pearlmutter, N. J. & MacDonald, M. C. (1995). Individual differences and probabilistic
constraints in syntactic ambiguity resolution. Jounal o f Memory and Language,
34,(4), 521-542.
Pinker, Steven. (1988). Language Acquisition. In Lila R. Gleitman & Mark Liberman
(Eds.), An invitation to Cognitive Science, Vol. I Language. Cambridge, Ma:
MIT Press, pp. 135-182.
Pinker, Steven. (1994). The Language Instinct: How the mind creates language. New
York: William Morrow and Company.
Plaut, D. C. (1996). Relearning after damage in connectionist networks: Toward a
theory of rehabilitation. Brain & Language, 52(1), 25-82.
195
R ep ro d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Plaut, D. C. & Shallice, T. (1993). Deep Dyslexia: A case study of connectionist
neuropsychology. Cognitive Neuropsychology, 10(3), 377-500.
Power, M. J. (1985). Sentence production and working memory. The Quarterly
Journal o f Experimental Psychology, 37a, 367-385.
Pratt, M.W., Boyes, C„ Robins, S. & Manchester, J. 1989. Telling tales: Aging,
working memory, and narrative cohesion of story retellings. Developmental
Psychology, 25 (4), 628-635.
Quine, W. V. & Ullian, J. S. (1970). The Web o f Belief New York: McGraw-Hill.
Rapp. B. & Caramazza, A.. (1993). On the distinction between deficits of access and
deficits of storage: A question of theory. Cognitive Neuropsychology, 10(2),
113-142.
Raymer, A. M.. & Bemdt, R. S. (1996). Reading lexically without semantics:
Evidence from patients with probable Alzheimer’s disease. JINS, 2(4), 340-349.
Richardson, J. T. E., Engle, R. W., Hasher, L., Logie, R. H., Stoltzfus, E. R. & Zacks, R.
T. (1996). Working Memory and Human Cognition. New York: Oxford
University Press.
Richardson, J. T. E. (1996). Evolving concepts of working memory, in Richardson, J.
T. E., Engle, R. W„ Hasher, L., Logie, R. H., Stoltzfus, E. R. & Zacks, R. T„
Working Memory and Human Cognition, (pp. 3-30). New York: Oxford
University Press.
Ripich, D. N. & Terrell, B. Y. (1988). Patterns of discourse cohesion and coherence in
Alzheimer's Disease. Journal of Speech and Hearing Disorders, 53. 8-15.
Rumelhart, D. E., Hinton, G. E. & McClelland, J. L. (1989). A general framework for
parallel distributed processing, in D. E. Rumelhart, J. L. McClelland and the
PDP Research Group, Parallel distributed processing: Explorations in the
microstructure o f cognition, vol. I: Foundations. Cambridge, MA: The MIT
Press. 45-76.
Rumelhart, D. E., McClelland, J. L., and the PDP Research Group. (1989). Parallel
distributed processing: Explorations in the microstructure of cognition, vol. I:
Foundations, Volume I. Cambridge, MA: The MIT Press.
Saffran, E. M., Bogyo, L. C., Schwartz, M. F. & Marin, O. S. M. (1980). Does deep
dyslexia reflect right hemishpere reading? In M. Coltheart, K. Patterson, J.C.
Marshall (Eds.). Deep dyslexia (pp. 381-406). London: Routledge.
Salthouse, T. A. (1988). Effects of aging on verbal abilities: Examination of the
psychometric literature, in L.L. Light & D. M. Burke (Eds.). Language,
Memory, and Aging, (pp. 17-35). Cambridge: Cambridge University Press.
Salthouse, T. A. (1994). The aging of working memory. Neuropsychology, 8 (4), 535-
543.
196
R ep r o d u ced with p erm issio n o f th e copyright ow n er. Further reproduction prohibited w ithout p erm issio n .
Shallice, T. (1988). From neuropsychology to mental structure. Cambridge:
Cambridge University Press.
Small, J. A. (1994). On- and off-line sentence comprehension abilities in Alzheimer’ s
Disease. Unpublished doctoral dissertation, University of Southern California,
Los Angeles.
Smith, S., Faust, M., Beeman, M., Kennedy, L. & Perry, D. (1996). A property level
analysis of lexical semantic representation in Alzheimer’s disease. Brain and
Language, 49(3), 263-280.
Smolensky, P. & Prince, (1996) Optimality theory, of linguistic processing.
Cambridge, MA: The MIT Press.
Stevens, K. L., Harm, M., Schuster, S. P., & MacDonald, M. C. (1995). Aging and the
use of context and frequency information during ambiguity resolution. Presented
at CUNY Conference on Human Sentence Processing, Tucson, AZ. March
1995.
Tanenhaus, Michael K. & Carlson, Greg N. (1989). Lexical Structure and Language
Comprehension, in William Marslen-Wilson (Ed.), Lexical Representation and
Process. Cambridge, MA: MIT Press, pp. 529-561.
Terry, R. D., Masliah, E., Salmon, D. P., Butters, N., DeTeresa, R., Hill, R., Hansen, L.
A., Katzman, R. (1991). Physical basis of cognitive alternations in Alzheimer’s
Disease: Synapse loss is the major correlate of cognitive impairment. Annals of
Neurology, 30, 572-580.
Terry, R. D., Peck, A., DeTeresa, R., Schechter, R„ Horoupian. D. S. (1981). Some
morphometric aspects of the brain in senile dementia of the Alzheimer type.
Annals o f Neurology, 10, 184-192.
Tippett, L. J., McAuliffe, S. & Farah, M. J. (1995). Preservation of categorical
knowledge in Alzheimer’s Disease: A computational account. Memory, 3(3/4),
519-533.
Tree, J. E. F. (1995). The effects of false starts and repetitions on the processing of
subsequent words in spontaneous speech. Journal of Memory and Language,
34(6), 709-738.
Ulatowska, H. K., Hayashi, M. M., Cannito, M. P., & Fleming, S. G. (1986).
Disruption of reference in aging. Brain and Language, 28, 24-41.
Valencia-Laver, D. L. (1992). Adult age differences in the production, detection, &
repair o f speech errors. Unpublished doctoral dissertation, Claremont Graduate
School, Claremont, CA.
Warrington, E. K. & McCarthy, R. (1987). Categories of knowledge: Further
fractionation and an attempted integration. Brain, 110, 1273-1296.
Warrington, E.K. & Shallice, T. (1979). Semantic access in dyslexia. Brain, 102,42-
63.
197
R ep ro d u ced with p erm issio n o f th e cop yright ow ner. Further reproduction prohibited w ithout p erm issio n .
Wechsler, D. (1987). Wechsler Memory Scale Revised. New York: Psychological
Corp.
Yuill, N., Oakhill, J. & Parkin, A. (1986). Working memory, comprehension, ability
and the resolution of text anomaly. British Journal of Psychology, 80, 351-361.
198
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm ission .
V *
IMAGE EVALUATION
TEST TARGET (Q A -3 )
A
✓
A
a
1 . 0
l.l
1.25
U ±
2.2
1 .4
1 M
I 1 .8
1 . 6
150mm
IM /1 G E . I n c
1653 East Main Street
Rochester. NY 14609 USA
Phone: 716/462-0300
Fax: 716/288-5989
R ep ro d u ced with p erm issio n o f th e copyrigh t ow n er. Further reproduction prohibited w ithout p erm issio n .

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Anatomic and functional specificity of relationships between explicit memory dysfunction and atrophy of medial temporal structures in Alzheimer's disease

PDF

Morphology and the lexicon: Exploring the semantics-phonology interface

PDF

From "almost impossible" to "almost certain": Novice and experienced psychologists' interpretations of statements of implied probability

PDF

Feeling and doing: Health and mind-body-context interactions during daily occupations of women with HIV/AIDS

PDF

Genetic and environmental influences on osteoarthritis

PDF

Definiteness types in Spanish: A study of natural discourse

PDF

The role of dynamic representations in understanding semantic impairments: Investigations from connectionist neuropsychology.

PDF

Effects of echoic, short-term, and long-term memory on validity and reliability of self-identification of authentic and simulated stuttering

PDF

Evaluating heaviness: Relative weight in the spoken production of heavy-NP shift

PDF

The concreteness effects of bilingual memory in free association tasks

PDF

THE EFFECT OF TIME-OUT FROM SPEAKING AND PARENTAL SOCIAL REINFORCEMENT UPON A PREADOLESCENT'S DYSFLUENT SPEECH.

PDF

Intra-lexical noun-verb dissociations: Evidence from Chinese aphasia

PDF

The influence of cognitive processing and fluid ability on problem-solving expertise

PDF

RCAN1 -1Long is overexpressed in Alzheimer disease

PDF

The president, the press, and the national agenda: White House communication strategies during the Kennedy, Nixon, Reagan, and Clinton administrations.

PDF

Risk-related differences in neurocognitive responses to an ethanol challenge

PDF

Prevention of adolescent substance use: An investigation of cross -level interaction effects of self, peer, family and school level risk and protective factors using multilevel modeling.

PDF

Prenatal Viral Infection, Developmental Abnormalities, and Schizophrenia

PDF

SPACE PERCEPTION AND VISUALIZATION IN CEREBRAL DYSFUNCTION.

PDF

Problem-based learning versus lecture: Effects on multiple choice test scores in associate degree nursing students

Asset Metadata

Creator Altmann, Lori Jean (author)

Core Title Effects of working memory and semantic impairments on speech in Alzheimer's disease.

Degree Doctor of Philosophy

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag health sciences, speech pathology,language, linguistics,OAI-PMH Harvest,Psychology, clinical,psychology, cognitive

Language English

Contributor Digitized by ProQuest (provenance)

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c17-373132

Unique identifier UC11350275

Identifier 9919007.pdf (filename),usctheses-c17-373132 (legacy record id)

Legacy Identifier 9919007.pdf

Dmrecord 373132

Document Type Dissertation

Rights Altmann, Lori Jean

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the au...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA