Journal of Gerontology: PSYCHOLOGICAL SCIENCES
2002, Vol. 57B, No. 4, P298–P311
Copyright 2002 by The Gerontological Society of America
Working Memory and Online Syntactic Processing
in Alzheimer’s Disease: Studies With
Auditory Moving Window Presentation
Gloria Waters1 and David Caplan2
1Department
of Communication Disorders, Boston University, Massachusetts.
Lab, Massachusetts General Hospital, Boston.
2Neuropsychology
Twenty patients with dementia of the Alzheimer’s type (DAT) and 20 controls were tested on six tests of working
memory and a test of online auditory sentence comprehension in which listening times for each phrase in the sentence, as well as the time required for an end-of-sentence plausibility judgment, were measured. The sentences
differed in syntactic complexity. Patients had lower working memory scores than controls and performed more
poorly on the plausibility judgments. However, patients were not more affected than controls by the syntactic
complexity of a sentence in these judgments, and both groups showed similar effects of syntactic structure in the
listening-time data. The increase in listening times at syntactically capacity-demanding points in complex sentences, compared with comparable points in matched simpler sentences, did not correlate with measures of working memory. The results indicate that early-stage DAT patients are not impaired in their ability to assign syntactic
structure and to use it to determine aspects of sentence meaning, despite their reduced working memories. This
provides evidence for a specialization within working memory for syntactic processing.
T
HE term working memory refers to a limited-capacity
specialized memory system in which small amounts of
information are maintained for short periods of time while
operations are performed on this information to achieve
computational goals (Baddeley & Hitch, 1974). Behavioral
and neurological evidence has suggested that different
working memory systems are engaged by verbal and visual
tasks (Shah & Miyake, 1996; Smith & Jonides, 1997).
Whether there are further specializations within each of
these systems is debated. Just and Carpenter (1992) have argued that there is a single verbal working memory system
that is used in all aspects of language processing—the single resource (SR) theory. In contrast, we have argued that
the verbal working memory system that is used in language
processing can be fractionated, such that the verbal working
memory system that is used to determine the literal, preferred, discourse-appropriate meaning of an utterance (interpretive processing) is distinct from that used in other
aspects of language processing (e.g., Caplan & Waters, 1999a,
1999b; Waters & Caplan, 1996a)—the separate language interpretation resource (SLIR) theory. We have suggested that
a specialized working memory system for interpretive processing may come into existence because the operations that
make up interpretive processing (accessing lexical items,
constructing syntactic structures, intonational contours, thematic roles, focus, topic, etc.) constitute a highly overpracticed set of computations that are always carried out together—
precisely the conditions that would be expected to foster the
development of a degree of independence of a mental skill.
In contrast, postinterpretive processing—using the products
of the comprehension process to accomplish tasks such as
entering information in semantic memory, reasoning, and so
forth—involves much more conscious, controlled, verbally
P298
mediated processing, and different postinterpretive operations occur in association with different utterances rather
than as a uniformly functionally integrated process. We have
argued that postinterpretive processing draws on a set of resources that is at least partially separate from that used for
interpretive processing.
A variety of types of experimental evidence have been
brought to bear on the issue of the nature of the resource
system used in language processing. One approach has been
to determine whether individual differences in verbal working memory capacity (usually measured on a task like the
Daneman & Carpenter [1980] reading or sentence span task)
are related to individual differences in language processing
efficiency. At least moderately high correlations have been
found between reading span and many verbal tasks, such as
verbal Scholastic Aptitude Test scores or scores on the Nelson
Denny reading comprehension test (see Daneman & Merikle,
1996, for a review). We have argued that these correlations
may be largely accounted for by the heavy postinterpretive
demands of these verbal tasks (Waters & Caplan, 1996c).
Far fewer studies have investigated the relationship between interpretive aspects of language processing and verbal
working memory capacity. A number of studies that have
focused on this issue have studied the ability to structure a
sentence syntactically. We have found that participants with
lower working memory capacity do not have increased difficulty in processing syntactically complex sentences, such as
garden path sentences and sentences with object relativization (Waters & Caplan, 1996c; Waters, Rochon, & Caplan,
1998). This is true for several different offline measures including enactment, sentence–picture matching, and sentence acceptability judgment and for online measures such
as word-by-word reading and continuous lexical decision
WORKING MEMORY AND ONLINE SYNTACTIC PROCESSING
(Caplan & Waters, 1999a). Several studies with college students have claimed to find a relationship between online
measures of syntactic processing efficiency and working
memory capacity (e.g., King & Just, 1991; Kluender & Kutas,
1993, MacDonald, Just, & Carpenter, 1992; Pearlmutter &
MacDonald, 1995), but we have argued that the evidence
they present is not convincing. We critically reviewed the
evidential basis for the SR theory and argued on the basis of
detailed analyses of aspects of experimental controls, statistical analyses, and other features of the experiments reported in the literature that the data do not support this theory (see Caplan & Waters, 1999a, 1999b; Waters & Caplan,
1996a, for discussion).
Evidence bearing on the SR and SLIR theories also comes
from neuropsychology. Many studies have shown that patients with neurological disease who have severely reduced
short-term memory due to limited rehearsal or phonological
storage capacities retain the ability to structure sentences
syntactically (see Caplan & Waters, 1990, for a review),
supporting the SLIR model. Patients with Alzheimer’s disease, whose reduced working memory capacity is due to
limited central executive functions (Rochon, Waters, & Caplan, 2000) have also shown normal syntactic processing in
several studies. Rochon, Waters, and Caplan (1994) used a
sentence–picture matching task with 10 sentence types.
They found that Alzheimer’s patients were less accurate
than controls, but were not more affected than controls by
syntactic complexity. Rochon and associates (2000) replicated this finding. Waters and colleagues (1998) extended
these results to a sentence–video verification task and to a
sentence–picture matching task using variable numbers of
foils. Grossman and White-Devine (1998) reported that 22
Alzheimer’s patients performed at the same level in answering questions about thematic roles in semantically reversible
active and passive sentences (see Appendix A, Note 1).
Waters, Caplan, and Rochon (1995) further found that a
concurrent digit memory load equal to participants’ spans
did not cause performance to decline more on syntactically
complex sentences than on syntactically simple sentences in
sentence–picture matching, in both Alzheimer’s patients and
normal controls. Waters and Caplan (1997) reported a similar result, using a timed end-of-sentence measure of judgments about the well formedness of sentences related to reflexive pronoun–antecedent agreement, under no interference
conditions and while engaged in a concurrent task consisting of continuously tracking dots or digits on a computer
screen. All participants showed effects of syntactic complexity, which were not increased in either of the two concurrent task conditions in either patients or controls. All
these results indicate that patients with Alzheimer’s disease
can structure sentences syntactically to derive their meaning
and that loading their already reduced verbal working system does not affect this ability. A few studies have reported
patterns of performance that indicate that there are impairments in syntactic comprehension in Alzheimer’s patients
(Grober & Bang, 1995; Kempler, Almore, Tyler, Andersen,
& MacDonald, 1998). In these cases, it is not clear that the
patients’ impairment in syntactic processing is related to
their reduction in working memory. Grober and Bang
(1995) concluded that the two functional deficits were unre-
P299
lated because the patients’ syntactic deficit in a sentence–
picture matching task was present even when a written sentence and a picture were in view simultaneously. On the
other hand, Kempler and colleagues (1998) argued that
the two deficits were related because performance on a
grammaticality judgment task correlated with a composite
measure of working memory in their patients. Rochon and
associates (2000), however, failed to find a significant correlation between the magnitude of a syntactic comprehension
deficit in sentence–picture matching and any of the measures of working memory in their patients. Overall, the bulk
of the literature has indicated that Alzheimer’s patients do
not have an impairment that affects their ability to construct
complex syntactic structures, as determined by end-ofsentence measures.
Even fewer studies have addressed the possibility that patients with Alzheimer’s disease have online syntactic processing impairments; that is, that they find it harder than
control participants to assign and interpret syntactic structures as they construct such representations while the words
of a sentence are being perceived. Kempler and colleagues
(1998) tested Alzheimer’s patients on a cross-modal naming
task, in which participants heard an introductory sentence
and a sentence fragment and then had to read aloud a visually displayed word that constituted either a grammatical or
an ungrammatical continuation of the fragment. Alzheimer’s patients showed the same increase in reading times
for incorrect continuations as normal controls; moreover,
the difference in reading times for the grammatical and ungrammatical continuations did not correlate with a composite measure of working memory. Almor, Kempler, MacDonald, Andersen, and Tyler (1999) reported that the
pattern of no difference between Alzheimer’s patients and
controls in reading times for grammatical versus ungrammatical sentence-fragment continuations occurred for subject–verb agreement when the verb was separated from the
sentence subject by a long intervening stretch of material.
These results suggest that online syntactic processing is normal in Alzheimer’s patients, at least as far as has been measured to date.
The purpose of this study was to further examine the relationship between working memory capacity and the ability
to structure sentences syntactically online. The study used a
paradigm developed by Ferreira and her colleagues (Ferreira & Anes, 1994; Ferreira, Henderson, Anes, Weeks, &
MacFarlane, 1996) to extend the study of online sentence
processing to the auditory modality—the auditory moving
windows task. This task requires participants to listen to a
sentence one word or phrase at a time by pressing a button
to receive successive words. The time required to call for the
next segment is the dependent variable for each word or
phrase. The assumption underlying this task is that, as in
self-paced reading and eye fixation duration measurements,
the time taken to listen to each segment reflects the time
needed to recognize the words in the segment and to integrate them into the accruing syntactic and semantic representation of the sentence in the discourse. Ferreira and colleagues (1996) have shown that listeners are immediately
sensitive to both lexical frequency and syntactic complexity
during the auditory processing of sentences in this task,
P300
WATERS AND CAPLAN
indicating that it can be used to examine the effect of syntactic variables on processing spoken sentences. We predicted that the auditory moving windows task would be sensitive to local increases in processing load, and, consistent
with SLIR theory, we predicted that this effect would not be
greater in patients with Alzheimer’s disease than in matched
control participants.
METHODS
Participants
Twenty patients with dementia of the Alzheimer’s type
(DAT) and 20 age- and education-matched controls participated in the study. The patients were referred by neurologists in a memory disorders unit at a local hospital, and the
controls were obtained from a large pool of elderly participants who had volunteered to participate in studies of memory and language. All participants were required to have English as their mother tongue and at least a high school
education. Controls were required to report that they were
aging normally and to be living independently. All participants were screened using pure-tone audiometry to rule out
a hearing impairment. Exclusionary criteria consisted of a
40-db loss in both ears at either 1000 or 2000 Hz or in their
better ear at 1000 or 2000 Hz.
All participants were pretested on a battery of neuropsychological tests to rule out any evidence of cognitive decline
or dementia in the controls and to document cognitive impairment in the DAT patients. The background measures included the Mini-Mental State Exam (Folstein, Folstein, &
McHugh, 1975); the visual memory, mental control, and
logical memory immediate (I) and delayed (II) subtests of
the Wechsler Memory Scale–Revised (Wechsler, 1987);
the Boston Naming Test (Goodglass & Kaplan, 1972); and
the block design subtest of the Wechsler Adult Intelligence
Scale–Revised (Wechsler, 1981).
Table 1 shows the performance of the participants on the
background measures. The DAT patients had a mean score
of 21.4 on the Mini-Mental State Exam, which indicates that
they were in the mild to moderate stages of the disease, and
the controls obtained a mean score of 28.6, which is in the
normal range. Scaled scores using Mayo’s Older Americans
Table 1. Background Characteristics of the Elderly Control
(EC) and Dementia of the Alzheimer’s Type (DAT) Patients
EC
Characteristic
Agea (years)
Educationa (years)
Mini-Mental State Exama(/30)
Visual Memoryb
Mental Controlb
Logical Memory Ib
Logical Memory IIb
Boston Naming Testb
Block Designb
aMean
bMean
DAT
M
SD
M
SD
75.2
15.8
28.6
10.0
11.7
9.3
9.8
11.8
10.9
9.2
2.8
1.4
3.6
0.79
3.0
3.1
3.1
2.4
75.3
15.4
21.4
7.1
8.5
4.1
2.6
6.4
7.6
9.3
2.7
4.2
3.6
3.6
2.9
1.3
3.8
4.0
scores.
scaled Mayo’s Older Americans Normative Studies scores (2–18).
Normative Studies norms (Ivnick et al., 1992) were in the
average range for the controls on all of the background tests.
Scaled scores for the DAT patients were in the impaired
range on all of the subtests.
DAT patients were tested individually in their homes in a
quiet room. Controls were tested at the university. All participants were tested over several sessions lasting from 1 to
1.5 hr each. The auditory moving windows task was given
in one session, and the background measures and the working memory span tasks were given in additional sessions.
The total number of sessions ranged from 4 to 7.5 for the
patients and 4 to 6.5 for the controls.
Procedures and Materials
Working memory measures.—Working memory capacity was tested using six operation span tasks. These tasks
were taken from a large study investigating working memory and aging (Waters & Caplan, 2001b; see Appendix A,
Note 2) and included Alphabet Span (Craik, 1986), Backward
Digit Span (Botwinick & Storandt, 1975), Subtract 2 Span
(Salthouse, 1988), and Sentence Span (Daneman & Carpenter, 1980) for syntactically simple and complex sentences.
In the Alphabet Span task, participants were required to
repeat a series of words after rearranging them in alphabetical order. The stimuli consisted of monosyllabic words of
moderate frequency. The words presented on each trial were
semantically and phonologically dissimilar, and no words
were repeated within a trial.
In the Backward Digit Span task, participants were required on each trial to repeat a series of digits in reverse
order of presentation. In the Missing Digit Span task, participants were read a string of digits. The experimenter
then reread the string in a different random order with one
item omitted. The participant was required to report the
missing item. In the Subtract 2 Span task, participants
were required to repeat a random sequence of digits after
subtracting 2 from each. The stimuli for the three tasks involving digits were drawn from the digits 1 to 9 and presented randomly.
Finally, Sentence Span was measured using the methods
and a subset of the materials from Waters and Caplan
(1996b). Participants were presented with a series of sentences on the video screen of a computer. They were required
to read each sentence silently and then to make a judgment
about the acceptability of each sentence in the series. When
the participant had made a decision about the last sentence in
the series, an asterisk appeared to indicate recall of the last
word of each of the sentences in the series in the correct serial
order. Participants were instructed not to recall the last word
presented first. In addition, they were told to perform the sentence task very accurately and then to perform as well as they
could on the recall task. Participants were tested with two different sets of stimulus materials that varied in syntactic complexity and number of propositions. The syntactically simple
sentences consisted of one-proposition sentences of the cleft
subject form (e.g., It was the gangsters that broke into the
warehouse), and the complex sentences consisted of subject–
object relatives that contained two propositions (e.g., The
meat that the butcher cut delighted the customer).
WORKING MEMORY AND ONLINE SYNTACTIC PROCESSING
The order of presentation of the tasks was randomized
across participants. For all tasks, testing began at Span Size
2. Controls were tested up to Span Size 8, and DAT patients
were tested up to Span Size 5, other than for the Sentence
Span task. Owing to time limitations, participants were only
tested up to span on the Sentence Span tasks. There were
five trials at each span size. Span was defined as the longest
list length at which participants were correct on 3 of 5 trials.
An additional 0.5 was added if participants were correct on
2 of 5 trials at the next span size. In a few cases, DAT patients correctly recalled the items on 3 of 5 trials at Span
Size 5. In this case, testing was continued until they were
correct on fewer than 2 of 5 trials. Because our previous research had shown that a composite measure based on performance on several different working memory tasks was
more reliable and stable than working memory span scores
based on a single task, performance across the six tasks was
averaged for each participant to yield a single span score
(Waters & Caplan, 2001b).
Online measure of sentence processing efficiency.—As
noted above, online sentence processing efficiency was assessed using the auditory moving windows paradigm (Ferreira et al., 1996). The methods were taken from our previous studies with young and elderly participants (Waters &
Caplan, 2001a, 2002).
The stimuli consisted of 130 semantically acceptable and
130 semantically unacceptable sentences divided equally
among the five sentence types—cleft subject (CS), cleft object (CO), subject–subject (SS), object–subject (OS), and
subject–object (SO) sentences—as shown in Table 2. The
CS and CO sentences were developed in pairs, such that
they had the same words but differed in terms of word order.
The SS, OS, and SO sentences were developed in triplets
and also contained the same words but differed in terms of
word order.
These sentences were selected because models of parsing
agree that object-relativized clauses impose a greater processing load than subject-relativized clauses at the verb of the
relative clause (see, e.g., Gibson, 1998). Empirical studies
have demonstrated this effect in the form of longer reading
times at the verb of object-relativized sentences and, in
some studies with SO and SS sentences, at the verb following the relative clause. We used both OS and SS sentences
because they offer different advantages as the syntactically
simple counterparts to SO sentences. OS sentences can be
constructed to have identical meanings to SO sentences;
verbs and noun phrases can be compared within the relative
and the main clauses across SS and SO sentences.
On half the trials the sentence was acceptable, and on the
other half it was unacceptable. The stimuli were designed so
that acceptability judgments did not require judgments
based on detailed semantic knowledge. Sentences had verbs
that required either animate objects or animate subjects
(e.g., It was the girl that the food nourished). Unacceptable
sentences violated these constraints. Detection of anomaly
could thus be based on fairly accessible, general semantic
features and did not require extensive searches through semantic memory for item-specific information. The animacy
of the nouns not affected by these constraints was varied
across sentences of each type and counterbalanced across
acceptable and unacceptable sentences of each type to prevent the use of animacy order in determining acceptability.
The point of anomaly was varied across clauses in the SS,
OS, and SO sentences to ensure that participants paced their
way through the entire sentence before making a decision
regarding acceptability.
A male speaker read the stimulus sentences in a semi-
Table 2. Sample Stimuli Used in the Auditory Moving Windows Paradigm
Sentence Type
P301
Sample
Cleft subject
Phrase:
Acceptable:
Unacceptable:
Intro
NP1 Pro
V
NP2
Continuation
It was/ the book/ that/ interested/ the teenager/because it was a romance.
It was/ the television/ that/ bought/ the family/ after the sale.
Cleft object
Phrase:
Acceptable:
Unacceptable:
Intro
NP1
Pro NP2
V
Continuation
It was/ the teenager/ that/ the book/ interested/ because it was a romance.
It was/ the family/ that/ the television/ bought/ after the sale.
Subject–subject
Phrase:
Acceptable:
Unacceptable:
NP1 Pro V1
NP2
V2
NP3
The law/ that/ favored/ the millionaire/ frustrated/ the workers.
The mother/ who/ tickled/ the child/ planted/ the father.
Object–subject
Phrase:
Acceptable:
Unacceptable:
NP1
V1
NP2 Pro
V2
NP3
The millionaire/ favored/ the law/ that/ frustrated/ the workers.
The child/ tickled/ the mother/ who/ planted/ the father.
Subject–object
Phrase:
Acceptable:
Unacceptable:
NP1 Pro
NP2
V1
V2
NP3
The law/ that/ the millionaire/ favored/ frustrated/ the workers.
The mother/ who/ the child/ tickled/ planted/ the father.
Note: Intro 5 introductory phrase; NP1 5 first noun phrase; Pro 5 pronoun; V 5 verb; NP2 5 second noun phrase.
P302
WATERS AND CAPLAN
anechoic chamber. The stimuli were recorded and digitized
using SoundEdit (Dunn, 1994) with a sampling rate of 44.1
kHz and 16-bit quantization, stored as a waveform file, and
edited using SoundEdit. A marker (referred to as a tag) was
placed in the waveform at the locations defining the boundaries of presentation of segments. Table 2 shows the location of tags (indicated by / ) for the five sentence types used
in the experiment. To make segment-to-segment transitions
smooth, tags were placed in the waveform at areas of low
signal amplitude, as indicated by auditory and visual inspection, whenever possible. When word boundaries did not coincide with areas of low signal amplitude, the tags were
placed so as to maximize the intelligibility of the words. Following Ferreira and associates (1996), a tone was appended
to the waveform of every sentence immediately following
the offset of visible and auditory activity associated with the
sentence-final word to signal the end of the sentence.
The resulting waveform files were then entered into PsyScope (Cohen, MacWhinney, Flatt, & Provost, 1993). The
experiment was controlled by a Macintosh PowerPC computer equipped with an additional digital I/O board and button
box for gathering interresponse times. The participants’ task
was to pace their way through the sentence as quickly as
possible by pressing a button on a box interfaced with the
computer for the successive presentation of each phrase, and
then to make an acceptability judgment about the sentence
they had just heard by responding yes or no into a voice key.
To discourage participants from pressing the button before
they had heard and processed each segment, the segment
was truncated at the point of the button press (see Appendix
A, Note 3) if a participant pressed the button before the end
of a segment. Reaction times for each button press, as well
as latency to initiate vocalization and accuracy on the acceptability judgment (as measured from the last button
press), were recorded in PsyScope (Cohen et al., 1993).
Predictions
The critical prediction of the SR theory is that Alzheimer’s patients will be more affected than control participants by the processing load imposed by syntactically complex constructions. In terms of this experiment, SR theory
predicts that there will be a three-way interaction between
group, phrase, and sentence type. In contrast, SLIR theory
predicts that this three-way interaction will not occur. SLIR
theory predicts a two-way interaction between phrase and
sentence type (that there will be longer listening times at the
most capacity-demanding phrases of syntactically complex
sentences compared with syntactically simple sentences),
and it predicts effects of group (that Alzheimer’s patients
will respond more slowly than controls). However, it predicts that Alzheimer’s patients and control participants will
be equally affected by syntactic load; that is, that there
will be no interactions involving the group factor (see Caplan & Waters, 1999b, for more detailed discussion).
RESULTS
Working Memory Measures
Table 3 shows the mean span scores for the elderly control participants and the DAT patients on the six operation
Table 3. Performance of the Elderly Control (EC) and Dementia
of the Alzheimer’s Type (DAT) Patients on the Working
Memory Measures (Span Scores)
EC
Measure
Alphabet span
Backward digit span
Subtract 2 span
Running item span
Sentence span simple (cleft subject)
Sentence span complex (subject–object)
DAT
M
SD
M
SD
3.88
5.43
4.82
3.79
3.50
2.65
0.76
1.6
1.0
1.0
1.4
1.1
2.67
3.98
3.00
3.00
1.28
1.10
1.4
1.8
2.1
1.7
0.82
1.1
span measures. Unpaired t tests for each task showed that
the DAT patients had significantly lower spans than the controls on all but the Running Item task: Alphabet Span, t(38) 5
3.37, p , .001; Backward Digit Span, t(38) 5 2.70, p ,
.001; Subtract 2 Span, t(38) 5 3.36, p , .001; Running Item
Span, t(38) 5 1.74, ns; Sentence Span Simple, t(38) 5 6.27,
p , .001; and Sentence Span Complex, t(38) 5 4.43, p ,
.001. A composite working memory span was calculated for
each participant by obtaining the average of the six working
memory measures. The elderly control participants had a
composite score of 4.0, and the DAT patients had a score of
2.5. Unpaired t tests showed that the two groups also differed on this measure, t(138) 5 4.2, p , .001. This composite measure was used in the correlational analyses reported
below, investigating the relationship between working memory capacity and online sentence processing efficiency.
End-of-Sentence Acceptability Judgment
We calculated mean reaction times for correct acceptability judgments at the end of sentences in the auditory moving
windows task and judgment accuracy as measured by A9.
The reaction time data in this and all other analyses were
subjected to standard data-trimming procedures. Before
analysis, any reaction times that were 63 standard deviations from the mean for an individual participant in a particular condition were replaced by the value that was 63 standard deviations as appropriate. A9s were used as the measure
of accuracy, as they control for a strategy of simply responding yes or acceptable on each trial and as such are considered to be a more sensitive measure of accuracy (Pollack &
Norman, 1964).
Figure 1 shows the mean reaction time for the acceptability judgments for acceptable sentences of each type for
which participants had made a correct acceptability judgment. The data for each pair of matched sentences (CS vs
CO, SS vs SO, OS vs SO) were analyzed in 2 (group) 3 2
(sentence type) analyses of variance (ANOVAs) using both
participant and item means as units.
As can be seen in Figure 1 there was an effect of group in
the analysis of reaction times to the acceptability judgments
in all three sets of data, with the judgment times of the DAT
patients being significantly longer than those of the controls,
CS versus CO, F1(1,38) 5 6.88, MSE 5 1,439,889.3, p ,
.05, F2(1,50) 5 75.2, MSE 5 170,821.5, p , .001; SS versus SO, F1(1,38) 5 4.03, MSE 5 3,625,082.2, p , .05,
WORKING MEMORY AND ONLINE SYNTACTIC PROCESSING
P303
SO, F1(1,38) 5 0.04, MSE 5 .002, ns (see Appendix A,
Note 4).
Online Measures of Sentence Processing Efficiency
Figure 1. Mean reaction time in milliseconds on the sentence acceptability judgment component of the auditory moving windows
task for cleft subject (CS), cleft object (CO), subject–subject (SS), object–subject (OS), and subject–object (SO) sentences. EC 5 elderly
control; DAT 5 dementia of the Alzheimer’s type.
F2(1,50) 5 88.4, MSE 5 208,533.3, p , .001; OS versus SO,
F1(1,38) 5 3.82, MSE 5 3,425,690.6, p , .05, F2(1,50) 5
79.8, MSE 5 201,695.3, p , .001. The effect of sentence
type was only significant in the CS versus CO comparison,
CS versus CO, F1(1,38) 5 5.54, MSE 5 298,082.5, p , .05,
F2(1,50) 5 5.1, MSE 5 377,954.0, p , .05; SS versus SO,
F1(1,38) 5 0.92, MSE 5 169,872.2, ns, F2(1,50) 5 0.39,
MSE 5 668,509.8, ns; OS versus SO, F1(1,38) 5 0.08, MSE 5
369,247.6, p , .05, F2(1,50) 5 0.02, MSE 5 201,695.3, ns.
The interaction between group and sentence type was not significant in any of the analyses, CS versus CO, F1(1,38) 5 1.24,
MSE 5 298,082.5, ns, F2(1,50) 5 1.6, MSE 5 170,821.5, ns;
SS versus SO, F1(1,38) 5 0.51, MSE 5 169,872.2, ns,
F2(1,50) 5 0, MSE 5 208,533.8, ns; OS versus SO,
F1(1,38) 5 0.67, MSE 5 369,247.6, ns, F2(1,50) 5 0.32,
MSE 5 201,695.3, ns.
Mean A9s were .98, .92, .93, .91, and .91 for the elderly
control participants and .89, .85, .86, .83, and .83 for the
DAT patients for the CS, CO, SS, OS, and SO sentences, respectively. These data were analyzed in the same manner as
the reaction time data, other than the fact that item statistics
were not calculated as the A9 measure is not amenable to
item statistics. Analysis of these data also resulted in a significant effect of group for all three comparisons, with the
judgments of the DAT patients being less accurate than
those of the controls, CS versus CO, F1(1,38) 5 11.9, MSE 5
.011, p , .05; SS versus SO, F1(1,38) 5 17.2, MSE 5 .007,
p , .05; OS versus SO, F1(1,38) 5 17.7, MSE 5 .008, p ,
.05. In addition, there was an effect of sentence type in the
CS versus CO and SS versus SO comparisons, CS versus
CO, F1(1,38) 5 40.6, MSE 5 .001, p , .05; SS versus SO,
F1(1,38) 5 7.7, MSE 5 .001, p , .05; OS versus SO,
F1(1,38) 5 0.26, MSE 5 .002, ns. The interaction between
group and sentence type was not significant in any of the
analyses, CS versus CO, F1(1,38) 5 0.74, MSE 5 .001, ns;
SS versus SO, F1(1,38) 5 0.29, MSE 5 .001, ns; OS versus
Phrase-by-phrase listening time.—The online dependent
measure for the auditory moving windows task consisted of
the response time for each phrase. Corrected response times
(listening times) were calculated by subtracting the segment’s tag-to-tag duration from the response time. The data
were trimmed as outlined above for the end-of-sentence acceptability judgment measure. Inspection of the dataset as a
whole showed that the degree of skewness and kurtosis did
not warrant transformation. Moreover, inspection of the normal probability plots for each pair of sentences showed that
the fit was not improved by transformation.
Separate analyses were carried out on listening times for
CS versus CO sentences, SS versus SO sentences, and OS
versus SO sentences. Listening times for acceptable sentences of each type for which participants had made a correct acceptability judgment were analyzed in 2 (group) 3 2
(sentence type) 3 4 (phrase) ANOVAs using both participant
and item means as units.
Figure 2 shows the mean listening time for CS and CO
sentences at each phrase for the two participant groups.
ANOVAs by participants and items showed that there were
significant main effects of group, F1(1,38) 5 6.5, MSE 5
1,457,473.8, p , .05, F2(1,200) 5 299.1, MSE 5 3,519.2,
p , .001; sentence type, F1(1,38) 5 30.8, MSE 5 32,808.2,
p , .001; F2(1,200) 5 8.7, MSE 5 182,517.9, p , .01; and
phrase, F1(3,114) 5 28.9, MSE 5 64,351.9, p , .001;
F2(3,200) 5 14.0, MSE 5 182,517.9, p , .001; a significant
Sentence Type 3 Phrase interaction, F1(3,114) 5 59.6,
MSE 5 45,438.6, p , .001, F2(3,200) 5 19.7, MSE 5
182,517.9, p , .001; and a Group 3 Phrase interaction,
F1(3,114) 5 59.6, MSE 5 64,351.9, p , .001; F2(3,200) 5
4.4, MSE 5 35,919.2, p , .01; and a Group 3 Sentence
Type 3 Phrase interaction, F1(3,114) 5 10.2, MSE 5
Figure 2. Mean listening time in milliseconds at each phrase for the
dementia of the Alzheimer’s type (DAT) patients and elderly controls
(EC) on cleft subject (CS) and cleft object (CO) sentences. Intro 5
introductory phrase; NP 5 noun phrase; V 5 verb.
P304
WATERS AND CAPLAN
45,438.6, p , .001, F2(3,200) 5 11.4, MSE 5 35,919.2,
p , .001.
As can be seen in Figure 2, the effect of group was due to
significantly longer listening times for the patients than for
the controls, and the effect of sentence type was due to
longer listening times for CO than for CS sentences. Analysis of the phrase effect using Tukey post hoc tests showed
that listening times for the verb (V) were longer than for all
other sentence positions and for the second noun phrase
(NP2) than for the first noun phrase (NP1) and for the introductory phase (Intro) than for NP1. Analysis of the Sentence
Type 3 Phrase interaction showed that listening times were
longer on NP2 in CS sentences than in CO sentences and on
V in CO sentences than in CS sentences.
As can be seen in Figure 2, both groups of participants
showed the pattern outlined above of longer listening times
at V of CO sentences and NP2 of CS sentences. However,
there was a significant Group 3 Sentence Type 3 Phrase interaction. To investigate this further, we calculated difference scores (listening times for CO–CS) at each phrase, for
each participant group. These scores served as an index of
the magnitude of the effect of increased complexity in the
more complex sentence type at each phrase. The difference
scores were analyzed in a 2 (group) 3 5 (phrase) ANOVA.
The Phrase 3 Group interaction was significant, F(4,152) 5
8.0, MSE 5 88,290.8, p , .001. Tukey’s post hoc tests
showed that the magnitude of the difference between listening times for V in CO and CS sentences was greater in
the Alzheimer’s patients than in the controls. However, the
magnitude of the difference between listening times for NP2
in CS sentences and for V in CO sentences between the
Alzheimer’s patients and controls was not significant.
Figure 3 shows the data from the SS and SO sentences.
The ANOVA showed that there were significant main effects
of sentence type, F1(1,38) 5 14.7, MSE 5 32,400.2, p ,
.001, F2(1,200) 5 5.5, MSE 5 90,120.8, p , .001; and
phrase, F1(3,114) 5 4.9, MSE 5 46,868.4, p , .001;
F2(3,200) 5 2.9, MSE 5 90,120.8, p , .001; and a significant Sentence Type 3 Phrase interaction, F1(3,114) 5 11.4,
MSE 5 41,639.1, p , .001; F2(3,200) 5 6.9, MSE 5
90,120.8, p , .001. The group effect was significant in the
item but not in the participant analyses, F1(1,38) 5 3.5,
MSE 5 1,716,334.2, ns; F2(1,200) 5 190.7, MSE 5
36,691.7, p , .001. As can be seen in Figure 3, the effect of
sentence type was due to longer listening times for SO than
for SS sentences. Post hoc tests showed that the effect of
phrase was due to longer listening times at V1 than at either
NP1 or NP2. The Sentence Type 3 Phrase interaction was
due to significantly longer listening times at V1 in SO sentences compared with SS sentences. Furthermore, listening
times at V1 in SO sentences were significantly longer than
those at NP2 in SS sentences, showing that the effect at V1
was not simply due to its position at the point of possible
thematic closure. There were no interactions involving the
group factor.
Figure 4 shows the data from the OS and SO sentences.
The ANOVA showed that there was a significant main effect
of phrase, F1(3,114) 5 7.9, MSE 5 47,084.5, p , .001,
F2(3,200) 5 4.8, MSE 5 77,062.2, p , .001; and a significant Sentence Type 3 Phrase Interaction, F1(3,114) 5 14.5,
Figure 3. Mean listening time in milliseconds at each phrase for the
dementia of the Alzheimer’s type (DAT) patients and elderly controls (EC) on subject–subject (SS) and subject–object (SO) sentences.
NP 5 noun phrase; V 5 verb.
MSE 5 39,096.3, p , .001, F2(3,200) 5 10.8, MSE 5
77,062.2, p , .001. The group effect was significant in the
item but not the participant analyses, F1(1,38) 5 3.3, MSE 5
1,597,117.3, ns; F2(1,200) 5 171.4, MSE 5 35,376.3, p ,
.001. The phrase effect was due to longer listening times at
V1 than at NP1 and at NP2 than at NP1. The interaction was
due to longer listening times at V1 in SO sentences than
in OS sentences and at NP2 in OS sentences than in SO
sentences. Listening times at V1 in SO sentences and NP2
Figure 4. Mean listening time in milliseconds at each phrase for the
dementia of the Alzheimer’s type (DAT) patients and elderly controls (EC) on object–subject (OS) and subject–object (SO) sentences.
NP 5 noun phrase; V 5 verb.
WORKING MEMORY AND ONLINE SYNTACTIC PROCESSING
in OS sentences were compared in order to determine the
extent to which the increased listening time at V1 was due
to its position at the end of a clause. This analysis
showed that listening times at V1 in SO sentences were
longer than those at NP2 in OS sentences, suggesting that
the increase in listening time was not simply due to its position in the phrase. There were no interactions involving the
group factor.
Given the failure to find a Group 3 Sentence Type 3
Phrase interaction in the analysis of both the SS–SO and
OS–SO data and the theoretical importance of this finding,
we reanalyzed the data for the SS–SO and the OS–SO sentences using tests of simple effects to ensure that the effect
of sentence type did not vary by group. The results of this
analysis were consistent with those presented above and
showed that both the DAT patients and the elderly controls
showed a significant Sentence Type 3 Phrase interaction for
both the SS–SO data set—elderly controls: F1(3,57) 5 5.4,
p , .01; F2(3,200) 5 5.25, p , .001; DAT: F1(3,57) 5 6.2,
p , .001; F2(3,200) 5 5.2, p , .01—and the OS–SO data
sets—elderly controls: F1(3,57) 5 5.8, p , .01; F2(3,200) 5
6.9, p , .001; DAT: F1(3,57) 5 8.8, p , .001; F2(3,200) 5
8.4, p , .01—and that this interaction was due to increased
listening time at V1 in the more complex SO sentences than
in the less complex SS or OS sentences.
Correlation of working memory measures and online
sentence processing measures.—Table 4 shows the correlations between measures of working memory and difference scores for phrases in the syntactically more complex
sentences (CO, SO) and syntactically less complex sentences (CS, SS, OS). These correlations were calculated
separately for each participant group (elderly control and
DAT) and for the group as a whole (all). For the SO–SS and
SO–OS sentence pairs, virtually none of the correlations
were significant. For the CO–CS sentence pair, the correlation between the composite working memory measure for
all participants and the difference in listening times on the
verb in CS and CO sentences was significant.
Table 4. Correlations Between the Composite Working
Memory Measure and Online Effects
Sentences
EC
DAT
All
CS vs CO
COV–CSV
COV–CSNP2
2.28
2.17
2.21
.11
2.44*
2.10
SS vs SO
SOV1–SSV1
SOV2–SSV2
SOV1–SSNP2
.01
2.29
2.02
2.06
.14
.16
2.07
.05
.13
OS vs SO
SOV1–OSV1
SOV2–OSV2
SOV1–OSNP2
2.03
2.24
2.12
.01
2.04
.19
2.14
2.08
.15
Note: EC 5 elderly control; DAT 5 dementia of the Alzheimer’s type; CS 5
cleft subject; CO 5 cleft object; SS 5 subject–subject; SO 5 subject–object;
OS 5 object–subject; V 5 verb; NP2 5 second noun phrase.
*p , .05.
P305
Table 5. Correlations Between the Composite Working Memory
Measure and Reaction Times (RT) and Accuracy (A9)
on the Acceptability Judgment Task
Measure
EC
DAT
All
CO–CS
SO–OS
SO–SS
.48*
2.01
2.05
.24
2.21
2.25
.11
2.17
2.21
Accuracy (A9)
CO–CS
SO–OS
SO–SS
.28
2.49*
2.35
2.24
2.06
2.09
2.12
2.15
2.09
RT
Note: EC 5 elderly control; DAT 5 dementia of the Alzheimer’s type; CO 5
cleft object; CS 5 cleft subject; SO 5 subject–object; OS 5 object–subject, SS 5
subject–subject.
*p , .05.
Table 5 shows the correlation between the working memory measures and the end-of-sentence acceptability judgment reaction time and accuracy data. Only two of these
correlations were significant. For elderly control participants, the composite working memory measure was related
to the increase in time to make a plausibility judgment for
CO compared with CS sentences and to the decrease in accuracy when making a judgment about SO compared with
OS sentences.
DISCUSSION
This study provides support for the view that there is a
specialization within verbal working memory for syntactic
processing. The Alzheimer’s patients tested here had consistently low working memory scores. Nonetheless, they
showed the same sensitivity to syntactic structure that normal control participants did.
With respect to end-of-sentence measures, the patients
with Alzheimer’s disease were slower and made more errors
than the matched controls in end-of-sentence acceptability
judgments, but were affected to the same degree by the syntactic complexity of the sentences. As noted in our review of
the literature in the introduction to this article, these end-ofsentence data replicate the majority of findings in the literature. Patients with Alzheimer’s disease typically do not do
as well as control participants on tasks such as sentence–
picture matching, video verification, and so forth, but are
not more affected than controls by syntactic complexity.
The overall poorer performance of Alzheimer’s patients
could be due to any of a number of factors; the lack of a
greater effect of syntactic complexity is evidence against the
SR model of working memory.
This study provides new data regarding online processing
of syntactic structure in Alzheimer’s patients. The selfpaced listening technique is sensitive to local increases in
processing load. In the present study, this sensitivity was
tested by constructing three matched pairs of sentences with
object- and subject-relativized clauses: CO and CS sentences, SO and SS sentences; and SO and OS sentences. In
the present study, the effects of increases in processing load
were discernible in all three sentence-type comparisons in
the form of longer listening times at the verb of the object-
P306
WATERS AND CAPLAN
relativized clause than at either the verb or the object noun
of the subject-relativized clause.
With one exception among six comparisons, which we
discuss below, these differences in processing times were
not greater in patients with Alzheimer’s disease than in control participants. This pattern strongly suggests that these
patients’ end-of-sentence results do not mask greater than
normal sentence-internal effects of processing load that are
only seen locally and are gone by the time a sentence has
ended. This in turn suggests that a reduction in working
memory capacity, as measured by standard tests, is not associated with a decreased ability to devote processing resources to syntactic analysis and to the use of syntax to determine sentence meaning. This conclusion is bolstered by
the fact that all but 1 of 24 correlations between a composite
measure of working memory and difference scores in listening times for critical phrases in the matched sentence pairs
were not significant.
One caveat to this conclusion concerns the fact that there
were only 20 participants per group in the present study. It is
possible that with a greater sample size and increased statistical power a different pattern of results might emerge. One
argument against this is that the basic pattern of results
found here has been found in several other studies of nondemented participants who nonetheless had reduced working memory capacity. These studies have included both elderly participants and college students with low working
memory spans (Waters & Caplan, 2001a; 2002).
In addition, one online finding could be taken to suggest
that the Alzheimer’s patients were more affected by syntactic complexity than the controls. This finding was the
greater increase in listening times at the verb of CO compared with CS sentences in Alzheimer’s patients than in
controls. However, this finding is likely due to nonsyntactic processing that occurs when the thematic roles required
by a verb are assigned, not syntactic processing per se.
This conclusion is based on considering both the difference in listening times for V in CO sentences compared with
V in CS sentences and the difference in listening times for
V in CO sentences compared with NP2 in CS sentences.
Both these comparisons showed longer listening times for
V in CO sentences, but only the former (V in CO sentences compared with V in CS sentences) was greater in
Alzheimer’s patients than in controls. However, adopting
Gibson’s (1998) model of syntactic processing, the difference in local syntactic processing load between V in CO
sentences and V in CS sentences is the same as that between V in CO sentences and NP2 in CS sentences (see
Appendix A, Note 5). Accordingly, assuming that Gibson’s
model provides an adequately accurate means of assessing
local syntactic processing load, the greater increase in listening time in Alzheimer’s patients at V of CO sentences
than at V of CS sentences cannot be due to the increased
syntactic processing load at these two points, because an
identical increase in syntactic processing load at V of CO
sentences compared with NP2 of CS sentences did not
lead to disproportionate increase in listening times in the
Alzheimer’s patients. This explanation is consistent with
the finding that Alzheimer’s patients did not show a larger
effect of increased processing load compared with controls
at the capacity-demanding region of SO sentences when
compared with either SS or OS sentences.
One source of increased load at V in CO sentences is related to discourse-level representations. In addition to constructing sentence-level semantic properties, such as thematic roles, listeners also construct discourse-level semantic
properties, such as an inventory of entities in the universe of
discourse, the entity in the focus of the discourse, and so
forth. From the perspective of creating a discourse structure,
sentences presented in isolation act as discourse-initial sentences (Altmann & Steedman, 1988). Clefting and other
markers of discourse focus are usually inappropriate in
discourse-initial sentences, and therefore in sentences presented in isolation. CO sentences presented in isolation
make additional demands on the construction of discourse
focus. The focus of a discourse is usually the subject of a
sentence and the agent of a verb (Kintsch & VanDijk, 1978).
In a CO sentence, this usual co-occurrence of agent and
focus is violated, because the theme of the verb is put into
the focused, clefted position. The verb of CO sentences is
the first point at which the dissociation of the usual combination of agent and focus can be recognized. All these factors increase the processing load at the V of CO sentences.
Because of the unnaturalness of the clefted construction in
discourse-initial sentences, constructing the discourse representations associated with clefted structures in singlesentence presentations may lie outside the usual interpretive
process and involve the putative “general purpose” pool of
processing resources; the demands on this pool would be
particularly high at the V of CO sentences. Increases in the
processing load on the general pool of processing resources
are expected to affect Alzheimer’s patients more than controls, due to their reduced general pool of such resources
(see Appendix A, Note 6).
This account is admittedly post hoc. However, it provides
a possible explanation of the entire pattern of data, something that an account based on syntactic processing load
cannot do, for the reasons we presented above. Furthermore,
we note that there is much evidence that CO sentences are
surprisingly difficult to process—more difficult than would
be expected on the basis of their syntactic structure alone—
an explanation of the particular difficulty associated with
these sentences in terms of their discourse properties is attractive. The more speculative key claim that underlies the
implication of this account for the SR and SLIR theories is
that computing these discourse properties in the case of CO
sentences presented in isolation involves the general pool of
working memory resources. This claim is speculative, but is
not a priori unreasonable, and it is testable. If the processing load associated with constructing a discourse focus uses
a general pool of processing resources when these sentences
are presented in isolation, reliance on this pool should disappear when these sentences are preceded by a context establishing the appropriateness of a CO structure. In that
case, listening times should not be more elevated at V of CO
sentences compared with V of CS sentences in Alzheimer’s
patients compared with controls.
The results of this experiment are consistent with online
studies comparing processing of subject- and object-relativized
sentences using other techniques, such as self-paced read-
WORKING MEMORY AND ONLINE SYNTACTIC PROCESSING
ing, in demonstrating that there is an increase in processing
time at the most capacity-demanding portion of the syntactically more complex sentence (Berwick & Weinberg,
1984; Haarman, Just, & Carpenter, 1997; Caplan, Hildebrandt, & Waters, 1994; King & Just, 1991; Stine-Morrow,
Ryan, & Leonard, 2000). They are also consistent with our
previous studies with normal participants in finding that the
increase in listening time at the capacity-demanding portion
of the more complex sentences was not larger in participants
with lower working memory capacities or in older participants. Aside from our studies, two other papers in the literature have examined individual differences in processing
these sentences using self-paced reading.
King and Just (1991) claimed to have found a greater increase for high- than low-span college students in reading
times for the main verb of SO sentences compared with SS
sentences. Stine-Morrow and colleagues (2000) reported a
complex pattern of results with these sentences: Low span
(older) participants made more errors in answering questions about thematic roles in object-relativized than in subjectrelativized sentences but showed smaller online effects of
complexity. As noted in the introduction to this article, we
have argued that the data in the King and Just article do not
support their conclusions (Caplan & Waters, 1999a, 1999b;
2002; Waters & Caplan, 1996a). The results reported by
Stine-Morrow and colleagues (2000) deserve comment.
First, Stine-Morrow and her colleagues did not report
Group 3 Sentence Type interaction in the question accuracy
data, or a three-way interaction of group, sentence type, and
region in the reading time data, leaving some uncertainty regarding whether the effects they describe truly differed in
the older and younger participants. This aside, the reading
time data are the opposite of those said to have been found
by King and Just for the low-span group as a whole, where
low-span participants showed a greater effect of object relativization at the verb. Stine-Morrow and her colleagues suggested that the low-span, older participants in their
study stopped processing the complex portion of the objectrelativized sentences earlier than the high-span, young participants. This would be consistent with King and Just’s
finding that low-comprehending, low-span participants
showed less of an effect of object relativization than the
better comprehending low-span participants. These data call
into question the near universal assumption in psycholinguistic research that increased difficulty integrating a phrase
into the accruing representation of a sentence leads to
longer, not shorter, reaction times. In addition, they point to
the need for more complex models of the effect of resource
limitations on language processing.
An issue that needs to be addressed regarding the present
study regards the auditory moving windows paradigm. A
concern that has been raised about this technique is that
splicing the speech signal into chunks for self-paced presentation may affect low-level word recognition processes and
the use of prosodic information. However, Ferreira and colleagues (1996) reported that participants who rated the naturalness and intelligibility of sentences presented in their
study found that the stimuli sounded reasonably natural and
were judged to be intelligible. The same was true of informal judgments of the naturalness of the stimuli in our exper-
P307
iment made by lab personnel (both naïve and non-naïve to
the goals of the experiment) at the materials development
stage. In addition, although not a response to concern about
the naturalness of auditory stimuli presented phrase by
phrase, we note that the auditory moving windows task is
not obviously less natural than many other tasks used in
psycholinguistic research, in particular, self-paced reading,
in which participants are denied parafoveal preview and are
unable to make regressive eye movements (Ferreira et al.,
1996). Finally, in defense of the auditory moving windows
task, we note that it has the advantage of presenting the
stimuli in the primary modality of language reception,
which is particularly important when studying pathological
populations, such as DAT patients, who may strongly prefer
this input modality and who may have impairments in reading. Moreover, this paradigm has proven a useful method of
investigating age differences in auditory processing time
(Titone, Prentice, & Wingfield, 2000).
In summary, the results of this study provide evidence
that patients with early Alzheimer’s disease are able to construct the syntactic structure of a sentence and use that
structure to determine the meaning of a sentence, despite
stable reductions in verbal working memory. This result
supports the view that the verbal working memory system is
fractionated, with some operations relying on a working
memory system not well measured by standard tests of this
function. The results are consistent with the hypothesis that
one such specialization involves computing the syntactic
structure of a sentence.
An alternative interpretation of these results is that syntactic processing uses the same working memory system
that is measured on standard tests but is so automatic that it
makes very few demands on this system, and even the reduced working memory capacity of the DAT patients tested
here is sufficient to assign syntactic structure. There are several arguments against this account. First, the complex sentences used in this study are so resource demanding as to be
at the limit of most individuals’ ability to structure. Adding
one more embedding to the SO sentences produces doubly
center-embedded sentences (e.g., The woman who the bodyguard who the trainer encouraged attracted closed the
drapes) that most people cannot comprehend unless they rehearse them or see them in written form. Arguably, it requires all, or almost all, of the working memory that a normal individual has available for syntactic processing to
structure the sentences presented here. It therefore seems
unlikely that working memory could be reduced to the extent seen in the DAT partients tested in this study (an average reduction of 38% across all span measures) without any
effect on a measure of online syntactic processing, if such
processing used the working memory system being measured. Second, working memory was not correlated with online measures of syntactic processing. This suggests that the
availability of working memory resources for syntactic processing does not depend on the size of the general purpose
working memory resource pool. Admittedly, such arguments cannot rule out this proposed analysis. Further empirical research, in which the processing load exerted by parsing is measured and compared with that exerted by working
memory tasks, could address the question of how much
P308
WATERS AND CAPLAN
resource is required by parsing. Comparing the effects of
parsing and other working memory tasks on a common third
task is one way to investigate this question (see Waters, Komoda, & Arbuckle, 1985, for an example).
This discussion raises two related questions: How many
resource pools are there, and can we delineate those aspects
of comprehension that use the hypothesized special pool(s)
and those that use the general pool? The discrepancy between performance on working memory tests and syntactic
processing points to a specialized working memory pool for
the latter. We have suggested that the processes that use this
specialized system are online, first pass, obligatory, and unconscious. In the domain of syntactic processing, this includes the assignment of syntactic structure on the basis of
grammatical categorical information (Frazier & Clifton, 1996)
as well as the utilization of lexical information (frequency
of co-occurrence data as well as pragmatic information) to
determine preferred syntactic structures (MacDonald, Pearlmutter, & Seidenberg, 1994; Pearlmutter & MacDonald,
1995). Conscious revision of syntactic analyses, as occurs in
deliberate recovery from garden path situations, is not part
of the syntactic processing we associate with this specialized resource pool. As noted in the introduction, we have
suggested that this resource system supports all the online,
first-pass, obligatory, unconscious processes that are devoted to assignment of the literal, preferred, discoursecongruent meaning of utterances, not just syntactic operations (Caplan & Waters, 1999a, 2002; Waters & Caplan,
1996c, 1997). We therefore hypothesize that there are two
verbal working memory systems: one devoted to online,
first-pass interpretive processing and one devoted to postinterpretive processing. Further studies are needed to determine whether this is the best way to characterize the differences in the types of processes that each of these two
hypothesized working memory systems supports, or whether
some other characterization of these domains, such as that
between unconscious and conscious verbally mediated processes, is empirically supported.
Acknowledgments
This work was supported by Grant AG0096610 from the National Institute on Aging.
Address correspondence to Gloria Waters, 635 Commonwealth Ave.,
Boston, MA 02215. E-mail: [email protected]
References
Almor, A., Kempler, D., MacDonald, M. C., Andersen, E. S., & Tyler, L. K.
(1999). Why do Alzheimer’s patients have difficulty with pronouns?
Working memory, semantics, and reference in comprehension and production in Alzheimer’s disease. Brain and Language, 67, 202–228.
Altmann, G., & Steedman, M. (1988). Interaction with context during
human sentence processing. Cognition, 30, 191–238.
Baddeley, A. D., & Hitch, G. (1974). Working memory. In G. A. Bower
(Ed.), The Psychology of learning and motivation (Vol. 8, pp. 47–89).
New York: Academic Press.
Balogh, J., Zurif, E. B., Prather, P., Swinney, D., & Finkel, L. (1998). Gap
filling and end of sentence effects in real-time language processing: Implications for modeling sentence comprehension in aphasia. Brain and
Language, 61, 169–182.
Berwick, R. C., & Weinberg, A. (1984). The grammatical basis of linguistic
performance: Language use and acquisition. Cambridge, MA: MIT Press.
Botwinick, J., & Storandt, M. (1975). Memory, related functions and age.
Springfield, IL: Charles C Thomas.
Caplan, D., Hildebrandt, N., & Waters, G. S. (1994). Interaction of verb selectional restrictions, noun animacy and syntactic form in sentence processing. Language and Cognitive Processes, 9, 549–585.
Caplan, D., & Waters, G. S. (1990). Short-term memory and language comprehension: A critical review of the neuropsychological literature. In G.
Vallar & T. Shallice (Eds.), Neuropsychological impairments of shortterm memory (pp. 337–389). Cambridge, England: Cambridge University Press.
Caplan, D., & Waters, G. S. (1999a). Issues regarding general and domainspecific resources. Behavioral and Brain Sciences, 22, 77–94.
Caplan, D., & Waters, G. S. (1999b). Verbal working memory capacity
and language comprehension. Behavioral and Brain Science, 22,
114–126.
Caplan, D., & Waters, G. S. (2002). Working memory and connectionist
models of parsing: A response to MacDonald and Christiansen. Psychological Review, 109, 66–74.
Cohen, J. D., MacWhinney, B., Flatt, M., & Provost, J. (1993). PsyScope:
An interactive graphic system for designing and controlling experiments in the psychology laboratory using Macintosh computers. Behavior Research: Methods, Instruments and Computers, 25, 257–271.
Craik, F. I. M. (1986). A functional account of age differences in memory.
In F. Klix & H. Hagendorf (Eds.), Human memory and cognitive capabilities (pp. 409–421). North Holland: Elsevier.
Daneman, M., & Carpenter, P. (1980). Individual differences in working
memory and reading. Journal of Verbal Learning and Verbal Behavior,
19, 450–466.
Daneman, M., & Merikle, P. M. (1996). Working memory and language
comprehension: A meta-analysis. Psychonomic Bulletin and Review, 3,
422–433.
Dunn, J. (1994). SoundEdit (Version 2) [computer software]. San Francisco, CA: Macromedia Inc, Apple Inc.
Ferreira, F., & Anes, M. (1994). Why study spoken language? In M. A.
Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 33–56). San Diego,
CA: Academic Press.
Ferreira, F., Henderson, J. M., Anes, M. D., Weeks, P. A., Jr., & McFarlane,
D. K. (1996). Effects of lexical frequency and syntactic complexity in
spoken language comprehension: Evidence from the auditory moving
window technique. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 22, 324–335.
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 of Psychiatric Research, 12, 189–198.
Frazier, L., & Clifton, C. (1996). Construal. Cambridge, MA: MIT Press.
Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76.
Goodglass, H., & Kaplan, E. (1972). Assessment of aphasia and related
disorders. Philadelphia: Lea & Febiger.
Grober, E., & Bang, S. (1995). Sentence comprehension in Alzheimer’s
disease. Developmental Neuropsychology, 11, 95–107.
Grossman, M., & White-Devine, T. (1998). Sentence comprehension in
Alzheimer’s disease. Brain and Language, 62, 186–201.
Haarman, H. J., Just, M. A., & Carpenter, P. A. (1997). Aphasic sentence
comprehension as a resource deficit: A computational approach. Brain
and Language, 59, 76–120.
Ivnick, R. J., Malec, J. F., Smith, G. E., Tangalos, E. G., Petersen, R. C.,
Korkment, E., et al. (1992). Mayo’s Older Americans Normative
Studies: WAIS-R, WMS-R and AVLT norms for age 56 through 97. The
Clinical Neurologist, 6 (Suppl.), 49–82.
Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review,
99, 122–149.
Kempler, D., Almor, A., Tyler, L. K., Andersen, E. S., & MacDonald, M. C.
(1998). Sentence comprehension deficits in Alzheimer’s disease: A
comparison of off-line vs. on-line sentence processing. Brain and Language, 64, 297–326
King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language,
30, 580–602.
Kintsch, W., & VanDijk, T. A. (1978). Toward a model of text comprehension and production. Psychological Review, 85, 363–394.
Kluender, R., & Kutas, M. (1993). Bridging the gap: Evidence from ERPs
on the processing of unbounded dependencies. Journal of Cognitive
Neuroscience, 5, 196–214.
WORKING MEMORY AND ONLINE SYNTACTIC PROCESSING
MacDonald, M. C., Just, M. A., & Carpenter, P. A. (1992). Working memory constraints on the processing of syntactic ambiguity. Cognitive Psychology, 24, 56–98.
MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). The
lexical nature of syntactic ambiguity resolution. Pschological Review,
101, 676–703.
Pearlmutter, N. J., & MacDonald, M. C. (1995). Individual differences and
probabalistic constraints in syntactic ambiguity resolution. Journal of
Memory and Language, 34, 521–542.
Pollack, I., & Norman, D. A. (1964). A non-parametric analysis of recognition experiments. Psychonomic Science, 1, 125–126.
Rochon, E., Waters, G. S., & Caplan, D. (1994). Sentence comprehension
in patients with Alzheimer’s disease. Brain and Language, 46, 329–
349.
Rochon, E., Waters, G. S., & Caplan, D. (2000). The relationship between
measures of working memory and sentence comprehension in patients
with Alzheimer’s disease. Journal of Speech, Language and Hearing
Research, 43(2), 395–413.
Salthouse, T. A. (1988). The role of processing resources in cognitive aging. In M. L. Howe & C. J. Brainerd (Eds.), Cognitive development in
adulthood (pp. 185–239). New York: Springer-Verlag.
Shah, P., & Miyake, A. (1996). The separability of working memory resources
for spatial thinking and language processing: An individual differences
approach. Journal of Experimental Psychology: General, 125, 4–27.
Smith, E. E., & Jonides, J. (1997). Working memory: A view from neuroimaging. Cognitive Psychology, 33, 5–42.
Stine-Morrow, E. A. L., Ryan, S., & Leonard, J. S. (2000). Age differences in
on-line syntactic processing. Experimental Aging Research, 26, 315–322.
Talland, G. A. (1965). Three estimates of word span and their stability over
the adult years. Quarterly Journal of Experimental Psychology, 17,
301–307.
Titone, D., Prentice, K. J., & Wingfield, A. (2000). Resource allocation during spoken discourse processing: Effects of age and passage difficulty as
revealed by self-paced listening. Memory and Cognition, 28, 1029–1040.
Waters, G., & Caplan, D. (1996a). The capacity theory of sentence comprehension: A reply to Just and Carpenter (1992). Psychological Review,
103, 761–772.
Waters, G. S., & Caplan, D. (1996b). The measurement of verbal working
memory capacity and its relation to reading comprehension. Quarterly
Journal of Experimental Psychology, 49A, 51–74.
Waters, G., & Caplan, D. (1996c). Processing resource capacity and the
comprehension of garden path sentences. Memory and Cognition, 24,
342–355.
Waters, G. S., & Caplan, D. (1997). Working memory and on-line sentence
comprehension in patients with Alzheimer’s disease. Journal of Psycholinguistic Research, 26, 377–400.
Waters, G. S., & Caplan, D. (2001a). Age, working memory, and on-line
syntactic processing in sentence comprehension. Psychology and Aging, 16, 128–144.
Waters, G. S., & Caplan, D. (2001b). The reliability and stability of working memory measures. Manuscript submitted for publication.
Waters, G. S., & Caplan, D. (2002). Individual differences in verbal working memory capacity and on-line sentence processing efficiency in college students. Manuscript submitted for publication.
Waters, G. S., Caplan, D., & Rochon, E. (1995). Processing capacity and
sentence comprehension in patients with Alzheimer’s disease. Cognitive Neuropsychology, 12, 1–30.
Waters, G. S., Komoda, M. K., & Arbuckle, T. Y. (1985). The effects of
concurrent tasks on reading: Implications for pharmacological recoding. Journal of Memory and Language, 24, 27–45.
Waters, G. S., Rochon, E., & Caplan, D. (1998). Task demands and sentence comprehension in patients with dementia of the Alzheimer’s type.
Brain and Language, 62, 361–397.
Wechsler, D. (1981). The Wechsler Adult Intelligence Scale–Revised. San
Antonio, TX: Psychological Corp.
Wechsler, D. (1987). The Wechsler Memory Scale–Revised. San Antonio,
TX: Psychological Corp.
Received June 26, 2000
Accepted August 7, 2001
Decision Editor: Margie E. Lachman, PhD
P309
Appendix A
Notes
1. Grossman and White-Devine (1998) also reported that Alzheimer’s patients were disproportionally impaired on “periphrastic” sentences with transitive verbs (e.g., The boy made
the girl kiss) than on such sentences with causative verbs (e.g.,
The mother made the baby awaken). However, this result is
hard to interpret for two reasons: (a) The periphrastic sentences
with transitive verbs were all semantically unconstrained, and
those with causative verbs were both constrained and unconstrained (e.g., The boy made the vase break), and (b) at least
some of the periphrastic sentences with transitive verbs seem to
be ungrammatical, as with the example The boy made the girl
kiss, taken from Grossman and White-Devine’s illustration of
their stimuli (kiss cannot be an intransitive verb).
2. Participants were also tested on the Missing Digit Span task
(Talland, 1965). However, the data from our previous study of
aging using these tasks showed that this task did not share a
significant amount of variance with the other variables. In addition, the assessment of reliability, stability, and internal consistency showed that this measure had extremely low test–retest
reliability, resulted in unstable classification of participants,
and had low internal consistency.
3. Participants pressed the button before the end of a segment and
thus truncated the segment before the end of the sentence on a
very small proportion of trials. Moreover, the percentage of trials on which this happened did not differ significantly across
the two groups. For CS and CO sentences, this occurred on 5%
of trials for the patients and 6% of trials for controls; for OS
and SO sentences, on 6% of trials for both groups; and for SS
and SO sentences, on 6% of trials for the patients and 7% of trials for the controls.
4. Analysis of the percentage of plausibility judgments that were
correct resulted in a pattern of results that was extremely similar to that found for the A9 data. The percentage of correct plausibility judgments for the DAT patients were 83%, 78%, 77%,
75%, and 73%, respectively, for the CS, CO, SS, OS, and SO
sentence types and for the elderly control participants were
96%, 86%, 87%, 84%, and 83%, respectively.
5. Gibson’s (1998) model assigns local processing load costs as a
result of integration and storage demands. Integration costs are
due to the distance (in number of words) between a word and
the item that it is related to (e.g., a verb and a noun playing a
thematic role around that verb), the number of new discourse
referents introduced between the word and the item that it is related to, and the number of integrations that are performed at
the word. Storage costs are the number of constituents that are
necessary to complete any obligatory structures associated with
a word that has occurred in a sentence (these constituents are
said to be predicted and thus maintained in storage until they
are recognized). Applying these measures, integration costs are
6 at V of CO sentences and 2 at V and NP2 of CS sentences.
Storage costs are 1 at V of CS sentences and NP2 of CO sentences and 0 at NP2 of CS sentences and V of CO sentences.
6. An argument worth developing is based on the idea that participants reactivate all nouns at the end of each sentence and assign
each to all thematic roles that are pragmatically possible, which
are checked against the thematic roles determined by the syntactic structure of a sentence (Balogh, Zurif, Prather, Swinney,
& Finkel, 1998). Though Balogh and associates suggested that
this process occurs at the end of a sentence, their data are compatible with the equally (if not more) plausible view that it occurs when the argument structure of a verb is satisfied, not (or
not only) at the end of a clause or a sentence. A process such as
P310
WATERS AND CAPLAN
this could underlie the additional listening time at V of CO sentences compared with V of CS sentences in Alzheimer’s patients. However, such a checking process would also be expected to operate in comparing V1 in SO sentences against V1
in both SS and OS sentences, and in neither case did the Alzheimer’s patients show disproportionately longer listening times.
Therefore, we reject this possible explanation of the greater increase in listening times for Alzheimer’s patients than for controls at V of CO sentences compared with CS sentences.
Appendix B
CS Acceptable Sentences
1. It was/ the movie/ that/ terrified/ the child/ because it
showed a monster.
2. It was/ the water/ that/ appealed to/ the boy/ because it was
hot.
3. It was/ the toy/ that/ interested/ the man/ because it was red.
4. It was/ the rain/ that/ woke up/ the father/ because of the
leak.
5. It was/ the fire/ that/ injured/ the policeman/ on the
highway.
6. It was/ the store/ that/ delighted/ the secretary/ yet it was
closed.
7. It was/ the box/ that/ distracted/ the dog/ for it held a cat.
8. It was/ the food/ that/ nourished/ the child/ because it was
plentiful.
9. It was/ the coffee/ that/ pleased/ the secretary/ because it
was caffeinated.
10. It was/ the camera/ that/ interested/ the girl/ for it was
automatic.
11. It was/ the car/ that/ amazed/ the boy/ although it was old.
12. It was/ the playground/ that/ pleased/ the mother/ for it was
safe.
13. It was/ the painting/ that/ delighted/ the woman/ because it
was beautiful.
14. It was/ the food/ that/ distracted/ the dog/ for he was hungry.
15. It was/ the book/ that/ interested/ the teenager/ because it
was a romance.
16. It was/ the radio/ that/ scared/ the child/ because it was too loud.
17. It was/ the sweater/ that/ warmed/ the child/ for she was
cold.
18. It was/ the girl/ that/ wrote/ the letter/ that was sent to the
wrong address.
19. It was/ the girl scout/ that/ sold/ the cookies/ for charity.
20. It was/ the janitor/ that/ mopped up/ the spill/ although it
was sticky.
21. It was/ the child/ that/ carved/ the pumpkin/ even though it
wasn’t Halloween.
22. It was/ the boyfriend/ that/ bought/ the flowers/ that cheered
the girlfriend.
23. It was/ the announcer/ that/ read/ the message/ as a warning
to all.
24. It was/ the swimmer/ that/ heard/ the whistle/ before he
started.
25. It was/ the monkey/ that/ ate/ the banana/ before the show.
26. It was/ the man/ that/ read/ the map/ but still got lost.
CO Acceptable Sentences
1. It was/ the child/ that/ the movie/ terrified/ because it
showed a monster.
2. It was/ the boy/ that/ the water/ appealed to/ because it was
hot.
3. It was/ the man/ that/ the toy/ interested/ because it was red.
4. It was/ the father/ that/ the rain/ woke up/ because of the
leak.
5. It was/ the policeman/ that/ the fire/ injured/ on the highway.
6. It was/ the secretary/ that/ the store/ delighted/ yet it was
closed.
7. It was/ the dog/ that/ the box/ distracted/ for it held a cat.
8. It was/ the child/ that/ the food/ nourished/ because it was
plentiful.
9. It was/ the secretary/ that/ the coffee/ pleased/ because it
was caffeinated.
10. It was/ the girl/ that/ the camera/ interested/ for it was
automatic.
11. It was/ the boy/ that/ the car/ amazed/ although it was old.
12. It was/ the mother/ that/ the playground/ pleased/ because it
was safe.
13. It was/ the woman/ that/ the painting/ delighted/ because it
was beautiful.
14. It was/ the dog/ that/ the food/ distracted/ for he was hungry.
15. It was/ the teenager/ that/ the book/ interested/ because it
was a romance.
16. It was/ the child/ that/ the radio/ scared/ because it was too
loud.
17. It was/ the child/ that/ the sweater/ warmed/ for she was
cold.
18. It was/ the letter/ that/ the girl/ wrote/ that was sent to the
wrong address.
19. It was/ the cookies/ that/ the girl scout/ sold/ for charity.
20. It was/ the spill/ that/ the janitor/ mopped up/although it
was sticky.
21. It was/ the pumpkin/ that/ the child/ carved/ even though it
wasn’t Halloween.
22. It was/ the flowers/ that/ the boyfriend/ bought/ that cheered
the girlfriend.
23. It was/ the message/ that/ the announcer/ read/ as a warning
to all.
24. It was/ the whistle/ that/ the swimmer/ heard/ before he
started.
25. It was/ the banana/ that/ the monkey/ ate/ before the show.
26. It was/ the map/ that/ the man/ read but/ still got lost.
OS Acceptable Sentences
1. The advertisement/ included/ the picture/ that/ depicted/ the
product.
2. The mechanic/ transported/ the car/ that/ passed/ the inspection.
3. The activist/ protected/ the tree/ that/ sheltered/ the squirrel.
4. The report/ omitted/ the biographer/ who/ insulted/ the publisher.
5. The bodyguard/ attracted/ the woman/ who/ closed/ the drapes.
6. The audience/ bored/ the child/ who/ embarrassed/ the
mother.
7. The coach/ selected/ the jockey/ who/ rode/ the horse.
8. The evidence/ defended/ the man/ who/ broke/ the window.
9. The mallet/ hit/ the man/ who/ pierced/ the pipe.
10. The soldier/ protected/ the sailor/ who/ piloted/ the boat.
11. The chemical/ burned/ the man/ who/ instructed/ the student.
12. The neighbors/ harassed/ the tenant/ who/ irritated/ the
landlord.
13. The secret/ identified/ the aide/ who/ infuriated/ the senator.
14. The jeweler/ liked/ the customer/ who/ bought/ the bracelet.
15. The millionaire/ favored/ the law/ that/ frustrated/ the
workers.
16. The yard/ surrounded/ the fence/ that/ pleased/ the carpenter.
17. The rocks/ encircled/ the wolf/ that/ watched/ the campfire.
18. The procedure/ used/ the computer/ that/ recorded/ the data.
19. The soldiers/ protected/ the fort/ that/ stored/ the ammunition.
WORKING MEMORY AND ONLINE SYNTACTIC PROCESSING
20.
21.
22.
23.
The writer/ reviewed/ the article/ that/ satisfied/ the editor.
The captain/ towed/ the boat/ that/ won/ the race.
The scout/ warmed/ the cabin/ that/ contained/ the firewood.
The stocks/ devalued/ the investment/ that/ displeased/ the
banker.
24. The performance/ disturbed/ the man/ who/ demanded/ the
refund.
25. The table/ scratched/ the butcher/ who/ summoned/ the doctor.
26. The inflation/ slowed/ the recession/ that/ disturbed/ the
consumers.
SS Acceptable Sentences
1. The picture/ that/ included/ the advertisement/ depicted/ the
product.
2. The car/ that/ transported/ the mechanic/ passed/ the inspection.
3. The tree/ that/ protected/ the activist/ sheltered/ the squirrel.
4. The biographer/ who/ omitted/ the report/ insulted/ the
publisher.
5. The woman/ who/ attracted/ the bodyguard/ closed/ the drapes.
6. The child/ who/ bored/ the audience/ embarrassed/ the
mother.
7. The jockey/ who/ selected/ the coach/ rode/ the horse.
8. The man/ who/ defended/ the evidence/ broke/ the window.
9. The man/ who/ hit/ the mallet/ pierced/ the pipe.
10. The sailor/ who/ protected/ the soldier/ piloted/ the boat.
11. The man/ who/ burned/ the chemical/ instructed/ the student.
12. The tenant/ who/ harassed/ the neighbors/ irritated/ the
landlord.
13. The aide/ who/ identified/ the secret/ infuriated/ the senator.
14. The customer/ who/ liked/ the jeweler/ bought/ the bracelet.
15. The law/ that/ favored/ the millionaire/ frustrated/ the
workers.
16. The fence/ that/ surrounded/ the yard/ pleased/ the carpenter.
17. The wolf/ that/ encircled/ the rocks/ watched/ the campfire.
18. The computer/ that/ used/ the procedure/ recorded/ the data.
19. The fort/ that/ protected/ the soldiers/ stored/ the ammunition.
20. The article/ that/ reviewed/ the writer/ satisfied/ the editor.
21. The boat/ that/ towed/ the captain/ won/ the race.
22. The cabin/ that/ warmed/ the scout/ contained/ the firewood.
P311
23. The stocks/ that/ devalued/ the investment/ displeased/ the
banker.
24. The man/ who/ disturbed/ the performance/ demanded/ the
refund.
25. The butcher/ who/ scratched/ the table/ summoned/ the doctor.
26. The recession/ that/ slowed/ the inflation/ disturbed/ the
consumers.
SO Acceptable Sentences
1. The picture/ that/ the advertisement/ included/ depicted/ the
product.
2. The car/ that/ the mechanic/ transported/ passed/ the inspection.
3. The tree/ that/ the activist/ protected/ sheltered/ the squirrel.
4. The biographer/ who/ the report/ omitted/ insulted/ the
publisher.
5. The woman/ who/ the bodyguard/ attracted/ closed/ the drapes.
6. The child/ who/ the audience/ bored/ embarrassed/ the mother.
7. The jockey/ who/ the coach/ selected/ rode/ the horse.
8. The man/ who/ the evidence/ defended/ broke/ the window.
9. The man/ who/ the mallet/ hit/ pierced/ the pipe.
10. The sailor/ who/ the soldier/ protected/ piloted/ the boat.
11. The man/ who/ the chemical/ burned/ instructed/ the student.
12. The tenant/ who/ the neighbors/ harassed/ irritated/ the landlord.
13. The aide/ who/ the secret/ identified/ infuriated/ the senator.
14. The customer/ who/ the jeweler/ liked/ bought/ the bracelet.
15. The law/ that/ the millionaire/ favored/ frustrated/ the workers.
16. The fence/ that/ the yard/ surrounded/ pleased/ the carpenter.
17. The wolf/ that/ the rocks/ encircled/ watched/ the campfire.
18. The computer/ that/ the procedure/ used/ recorded/ the data.
19. The fort/ that/ the soldiers/ protected/ stored/ the ammunition.
20. The article/ that/ the writer/ reviewed/ satisfied/ the editor.
21. The boat/ that/ the captain/ towed/ won/ the race.
22. The cabin/ that/ the scout/ warmed/ contained/ the firewood.
23. The investment/ that/ the stocks/ devalued/ displeased/ the
banker.
24. The man/ who/ the performance/ disturbed/ demanded/ the
refund.
25. The butcher/ who/ the table/ scratched/ summoned/ the doctor.
26. The recession/ that/ the inflation/ slowed/ disturbed/ the
consumers.