JOURNAL OF MEMORY AND LANGUAGE 30,580-602 (1991)

Individual Differences in Syntactic Processing: The Role of

Working Memory

JONATHAN KING AND MARCELADAMJUST

Carnegie Mellon University
The results of two experiments indicate that individual differences in syntactic processing
are governed in part by the amount of working memory capacity available for language
comprehension processes. Reading the verbs of an object relative sentence, such as The
reporter that the senator attacked admitted the error, takes more time for readers with less
working memory capacity for language, and their resulting comprehension is less accurate.
Experiment 1 investigated the effects of a concurrent working memory load and found that
with no load or a small memory load many Low Span readers comprehended object relative
sentences very poorly although their reading times in the critical area of these sentences
were greater than those of High Span subjects. Experiment 2 replicated the reading time
effects of Experiment 1 for object relative sentences and showed that pragmatic information
improved the comprehension of the lower capacity readers, although their use of this in-
formation was limited to the clause in which it was presented. o IW Academic press, IIIC.

In this paper, we develop the hypothesis diate and partial products (Daneman & Car-
that individual differences in syntactic pro- penter, 1980). Individual differences in
cessing can be observed in normal readers, working memory capacity may be the result
and that these differences are influenced by of differences in either the size of the pool,
differences in working memory capacity for the efficiency of the processes that perform
language. Although the processing of syn- symbolic computations, or both.
tax has been an active topic of language There are several reasons for hypothesiz-
research for several decades, only a few re- ing a relation between syntactic processing
searchers have examined individual differ- and working memory. Syntactic processing
ences in the processing of syntactic infor- transforms a linear sequence of words into
mation (e.g., Gleitman & Gleitman, 1970; a nonlinear (hierarchical) syntactic struc-
Holmes, 1987). The present research inves- ture, and this transformation requires the
tigates how syntactic processing and the re- temporary storage of word representations
sulting comprehension of sentences may be during the left-to-right processing of a sen-
influenced by a reader’s working memory tence. Moreover, linguistic parsing models
capacity for language. Working memory ca- have appealed to working memory storage
pacity includes not only storage functions limitations to account for the preferences
for intermediate or partial products, but a among alternative interpretations in tempo-
flexibly deployable pool of operational re- rarily ambiguous sentences (e.g., Kimball,
sources needed to perform the symbolic 1973; Frazier 8z Foder, 1978; Gibson,
computations that generate those interme- 1990). Limitations on the processing func-
tions of working memory are similarly im-
This work was supported in part by Grant MH- plicated in syntactic processing by evi-
29617 from NIMH, and Research Scientist Develop- dence that certain syntactic gaps are diffi-
ment Award MH-00662 from NIMH. We thank Patri- cult to find and process (e.g., MacDonald,
cia Carpenter and Brian MacWhinney for comments 1989).
on the manuscript. Requests for reprints should be The classic example of a syntactic struc-
addressed to Marcel Adam Just, Psychology Depart-
ment, Carnegie Mellon University, Pittsburgh, PA ture that makes large demands on working
15213. memory capacity is a center-embedded rel-

580
0749-5%x/91 $3.00
Copy&ht 0 1991 by Academic Press, Inc.
AU rights of reproduction in any form reserved.
 DIFFERENCES IN SYNTACTIC PROCESSING 581

ative clause, such as the one in sentence (1) tive sentence, like (2) below, the main
below. It is also called an object relative for clause is interrupted, but role assignments
brevity, a name that reflects the role that can be made one at a time, and the constit-
the head noun plays as the object in the uents have parallel roles in the two clauses.
relative clause. (2) The reporter that attacked the sena-
(1) The reporter that the senator at- tor admitted the error.
tacked admitted the error. These subject relative sentences are both
Subjects who hear a sentence like (1) and easier to comprehend and easier to process
then try to paraphrase it make errors in on-line than object relative sentences like
matching verbs with their agents approxi- (1) (Holmes & O’Regan, 1981). In the read-
mately 15% of the time (Larkin & Burns, ing of the easier subject relative sentences
1977). There are three kinds of demands like (2), the head of the relative clause
such sentences make on working memory needs to be maintained over only a short
that combine to make them difficult to com- distance, since the agent role can be as-
prehend. First, the embedded clause in signed to the reporter as soon as the verb
such sentences interrupts the main clause, attacked is processed; this leaves only a
drawing on the storage resources of work- single role to assign when the senator is
ing memory. The representation of the reached. This contrasts with the reading of
clause segment that precedes the interrup- an object relative sentence like (1) where
tion must either be retained in working the reader cannot assign a thematic role to
memory during the processing of the em- either the reporter or the senator until the
bedded clause or be reactivated at the con- verb attacked is encountered. One would
clusion of the embedded clause (Miller & also expect a subject relative sentence like
Chomsky, 1963; Wanner & Maratsos, (2) to be easier, because the agent in the
1978). Second, the assigning of the proper relative clause is also the agent in the main
thematic roles to the two noun phrases im- clause, and thus requires no assignment of
poses some difficulty. In particular, know- conflicting roles to the actor and no per-
ing whether the head of the relative clause spective shift. Thus, the three aspects that
is the agent or the patient of the relative make the processing of an object relative
clause verb causes comprehension of the sentence demanding of working memory
relative clause to be less accurate than are all mitigated in a subject relative sen-
comprehension of the main clause in this tence. Our main hypothesis is that the dif-
kind of sentence (e.g., Holmes & O’Regan, ference between these two sentence types
1981). Third, the assignment of two differ- in their working memory requirements will
ent roles to a single syntactic constituent interact with the differences in working
also taxes working memory capacity. In the memory capacity we can observe in a col-
example above, reporter is the agent in one lege level population.
clause and the recipient of the action in the The relation between working memory
other. Associating a single concept with capacity and language processing has been
two different roles simultaneously seems to addressed by the work of Daneman and
be a source of difficulty in language com- Carpenter (1980, 1983) at the referential
prehension (Bever, 1970) and the switch- level of processing, but not at the syntactic
ing of perspective in the construction of level. Daneman and Carpenter have shown
such a concept can also tax cognitive re- that differences in a measure of working
sources (MacWhinney & Pleh, 1988). memory capacity predict substantial differ-
The difficulty of processing an object rel- ences in the ability to compute the referents
ative sentence contrasts with the relative of pronouns and the ability to integrate se-
ease of processing a related construction, mantic information between and within
called a subject relative. In a subject rela- sentences. The measure of working mem-
582 KING AND JUST

ory capacity developed by Daneman and tation Hypothesis (Just & Carpenter, 1980,
Carpenter, called the Reading Span Test, 1987). In the case of the object relative sen-
determines the longest set of sentences (of tence such as (1) the increased processing
approximately 15 words each) that subjects demands first manifest themselves when
can orally read and from which they can the reader reaches the first verb (attacked).
then recall all the sentence-final words of At that point, the head noun (reporter) must
the set. This task requires subjects to main- be assigned to the thematic role of patient
tain the set of unrelated sentence-final and senator must be assigned to the the-
words from preceding sentences while they matic role of agent. Next, the reader en-
process each successive sentence in the counters the verb (admitted) and must find
set. Performance on the Reading Span Test an agent for this verb, but this agent is not
accounts for almost 35% of the variance on the same as the agent of attacked. Thus the
verbal SAT scores and 80% of the variance two adjacent verbs should be the locus of
in performance on a task which measures the extra processing load. This result was
the ability to associate pronouns with their found by Ford (1983) with a continuous lex-
antecedents over varying distances (Dane- ical decision task in which readers judged
man & Carpenter, 1980; see also Turner & each successive word of a sentence to de-
Engle, 1989). By contrast, simple word or termine whether it was a real word or a
digit span tests, which make little demand nonword. Subjects took 25 ms longer to
on any operational resources related to make a lexical decision on the verbs at-
comprehension, typically are not substan- tacked and admitted in an object relative
tially correlated with reading ability or sentence like (1) than in a subject relative
other interesting language tasks (Perfetti 8z sentence like (2). Similar results were ob-
Lesgold, 1977). The advantage of the Read- tained by Holmes and O’Regan (1981) in a
ing Span measure over the simpler mea- reading task which measured the duration
sures may derive from a similarity between of subjects’ eye fixations using sentences in
the reading span task and more interesting French, which has a similar subject/object
linguistic tasks in their demands on the relative clause distinction. Consequently, it
storage and processing functions of work- is reasonable to expect that word-by-word
ing memory. Our main hypotheses are that reading times will be sensitive to syntactic
while all readers should have greater diffi- processing load and to working memory ca-
culty with object relative sentences than pacity differences among individuals. It is
with subject relative sentences, it is the also known that the process of role assign-
readers with less working memory capacity ment in object relative sentences is difficult
for language as assessed by the Reading enough to cause a substantial number of
Span test who should have the most difli- comprehension errors, which suggests that
culty. comprehension scores could also be sensi-
The distribution of processing load dur- tive to working memory capacity differ-
ing comprehension should be reflected in ences.
the word-by-word reading times for object The current studies also experimentally
relative and subject relative sentences like manipulated the subjects’ effective working
(1) and (2): memory capacity by either imposing an ex-
traneous memory load during the process-
(1) The reporter that the senator at- ing of the sentences (Experiment l), or by
tacked admitted the error. supplying pragmatic information that would
(2) The reporter that attacked the sena- make it easier to comprehend object rela-
tor admitted the error. tive sentences (Experiment 2). In previous
We assume that readers try to interpret studies that used a memory load manipula-
each word as soon as they encounter it, in tion, Baddeley (1986) examined sentence
agreement with the Immediacy of Interpre- comprehension that was concurrent with
 DIFFERENCES IN SYNTACTIC PROCESSING 583

retaining a string of digits or articulating liability check on our use of the Reading
nonsense syllables. The concurrent storage Span Test, sentence comprehension accu-
task imposes an additional load which con- racy and reading time results will be crucial
sumes some of the resources of working to our hypothesis. This is particularly true
memory; consequently, it is expected to de- for one- and two-sentence trials, which
grade some facet of performance in a de- should be within the capacity limitations of
manding comprehension task. By contrast, all subjects.
when additional pragmatic information is Our hypothesis is that when the combi-
supplied, comprehension performance nation of storage and processing require-
should improve, because the pragmatic in- ments of a comprehension task exceeds
formation could obviate some of the syn- working memory capacity, performance on
tactic processing usually required by object one or more aspects of the task will deteri-
relative sentences. orate. Thus, when comparing performance
between groups of subjects differing in
EXPERIMENT 1 working memory capacity for language, we
This experiment investigated how differ- would expect the lower capacity subjects to
ences in the working memory capacity of show lower comprehension accuracy for
readers affected the processing of syntactic syntactically more complex object relative
structure. The experimental task itself var- sentences and increased reading times at
ied two factors: the syntactic structure of the sentence locations in object relative
the sentences read and whether or not a sentences where substantial syntactic com-
memory load was carried in working mem- putations must be performed. This pattern
ory concurrently while reading. The third of results can be expected from the in-
and central factor studied was the working creased processing demands of the object
memory capacity of the subjects, which relative sentences compared to subject rel-
was assessed by the Reading Span task. ative sentences, as described above. The
The structure of the reading task was simi- effect of a working memory load should de-
lar to that of the Reading Span task de- pend on the working memory capacity of
scribed in the introduction. Subjects read the subjects themselves. The performance
sets of one, two, or three sentences and of high capacity subjects maintaining a
were asked to recall the sentence final working memory load should more closely
words at the end of the set. The syntactic resemble the performance of low capacity
factor was manipulated by varying the subjects without any load. Similarly, the
structure of the final, target sentence. Half performance of low capacity subjects main-
of these targets included subject relative taining a working memory load should be-
clauses, while the other half contained come worse up until the point when their
more difficult object relative clauses. The capacity is entirely exhausted, at which
external memory load factor was manipu- point patterns of performance may cease to
lated by varying the number of sentences be easily interpretable.
preceding the target. In addition, subjects
were required to respond to a true-false Method
probe item testing their comprehension of Materials. Each target sentence occurred
the final target sentence in the set. This de- as the final sentence of a set of either 1, 2,
sign allowed us to measure the proportion or 3 sentences. In a set with one sentence,
of sentence-final words recalled, the pro- the subject read only the target sentence;
portion of comprehension probes answered then he or she recalled the last word of the
correctly, and the reading times on each sentence and then answered a question
word of the target sentence. While recall of about the sentence. In a set of two sen-
sentence-final words will primarily be a re- tences, there was one sentence prior to the
584 KING AND JUST

target sentence. The subject read each of adverbial phrase that followed the direct
the two sentences, recalled the two sen- object of the main verb. This was done so
tence-final words, and then answered a that the reading time on the direct object
question about the target sentence. Three was not contaminated by any effect due to
sentence trials were the same, except that encountering the end of the sentence,
all three sentence-final words were re- where subjects sometimes pause briefly
called. For convenience, we will refer to (Just & Carpenter, 1980). The filler sen-
these conditions as sets of 1, 2, or 3 words. tences that preceded the last sentence in
However, it is important to remember that either the experimental or filler sets were
when the subject was reading the target sentences that had previously been found
sentence, she or he was retaining only 0, 1, to be fairly easy to process, as measured by
or 2 words, respectively, from the preced- per-character reading times. The final sen-
ing sentences. Thus, the number of words tence of each filler set was chosen from an-
retained during the reading of the critical other set of sentences which had previously
sentence was one less than the number of been found to be rather difficult, in order to
sentences in the set. reduce the likelihood of subjects noticing
There were 60 trials in the experiment, the object relative sentences as being more
evenly divided among sets consisting of 1, difficult than the tiller trials. An example of
2, or 3 sentences. There were 36 experi- this type of sentence is:
mental sets and 24 tiller sets. Half of the
To equivocate on issues that are mal-
experimental sets contained an object rela-
adaptive for mankind is a futile objective.
tive sentence as the last sentence in the
trial, while the other half contained a sub- The comprehension test probe for a tar-
ject relative sentence as the last sentence. get sentence was constructed by combining
The experimental and filler trials were pre- one of the two verbs in the sentence with
sented in a different random order for each two of the three nouns. All pragmatically
subject. All target sentences were con- possible orders were used in equal numbers
structed so that the relative clauses were to defeat guessing strategies. Possible com-
completely reversible and both the correct prehension probes for the object relative
and reversed versions were equally plausi- sentence (1) would include any of the fol-
ble and appropriate in the context of the lowing sentences:
sentence. All the sentences were between
12 and 17 words in length. An example of an (1) The reporter that the senator at-
tacked admitted the error . . .
object relative sentence is:
The reporter attacked the senator.
The reporter that the senator attacked (False)
admitted the error publicly after the hear- The senator attacked the reporter.
ing. W-N
The reporter admitted the error.
Its subject relative counterpart would be:
(True)
The reporter that attacked the senator The senator admitted the error.
admitted the error publicly after the hear- (False)
ing. The same four statements served as the
Eighteen familiar transitive verbs that take comprehension test items for the subject
animate subjects and objects were each relative sentence, except that the two state-
used twice (with different grammatical sub- ments probing the relative clause would
jects and objects) in the embedded clauses have opposite truth values. The number of
of the target sentences. Each of the sen- true and false comprehension test items
tences ended with an extra prepositional or was balanced across clause type and mem-
 DIFFERENCES IN SYNTACTIC PROCESSING 585

ory load. To better assess the comprehen- 1

sion of the embedded clauses of object rel- The
--- first
I I
--- ----- aentcmce
ative sentences, which have proved to be ---
--- ------
-__ ____- ----
--- __---- ---_--_
more difficult to answer in previous re- -_-
--- ------
--- _----- ---
--- _------- ----
search, these were interrogated twice as of- ---.
-----_
--- -------- -
___-__
--- ---_--
ten as the main clauses; to compensate for -- - -__ - -_
--- ------- ------
--- -----_- --
this frequency bias, the main clauses of
subject relatives were interrogated twice as
often as the embedded clauses. (All com-
--- -__--
prehension analyses below are therefore The
--- ----- ---_--__ - -_- ----- --
--- second
based on weighted means to correct for the --- --- ------ *sntencs -__ ---__- -
--- -_- --_.
unequal frequencies of main clause and rel- ---.
-----_
--- ___----_ -___ __
-_- --------

ative clause probes in the two syntactic ---

--- -_---_.
------
-_----

conditions.) --- ------- -_--__--

--- -------

Procedure. Each set of sentences was C

presented word-by-word, using a subject- I --- --- ----_-- --- ----- -------
1 I
paced moving window paradigm on an 1 4-
IBM-PC/XT (Just, Carpenter, & Woolley,
1982). The procedure is schematically illus-
trated in Fig. 1. When subjects clicked the
Da I:__: i!!f
handheld microswitch to advance the dis-
play to the next word in a sentence, the
letters of that word would replace the
dashes, and the letters of the previous word
would revert to dashes. Each sentence in a
set began on a new line, and the final word
~~~
of each sentence was padded out to a length
of 10 characters using underscores, so that
the number of dashes in the display would FIG. 1. A diagram showing the course of a three-
not be a clue to the actual length of the sentence trial in Experiment 1: (A and B) The subject
sentence-final word. The critical sentence reads the first two sentences one word at a time and
was followed by an instruction to recall the encodes the two sentence-final words. (C) The subject
reads the experimental sentence one word at a time
sentence-final words of the 1, 2, or 3 sen-
and then advances the display one more time to the
tences in the set. Subjects were allowed to recall prompt (D), where he recalls as many of the
recall these words in any order, provided sentence final words as possible. After recall is com-
that the sentence-final word of the last sen- plete the subject pushes a button to expose the com-
tence not be recalled first, unless there had prehension probe, and answers “True” or “False” by
been only one sentence in the set. After pressing the appropriate button on the response box.
recalling all the sentence-final words that
they could, subjects advanced the display whose reading spans had not been previ-
to a true/false comprehension probe, which ously measured in other unrelated experi-
always interrogated the last sentence of the ments. Subjects read sets of sentences
set. Subjects were told to favor accuracy aloud, one sentence at a time, and then re-
over speed in their responses to the probe. called each of the sentence-final words
After the probe was answered, the next set from that set. Subjects were given sets of an
started. increasing number of sentences until they
Reading Span Test. This test was admin- failed to recall all final words for every
istered last, and only to those subjects group of a given set size. Reading span was
586 KING AND JUST

defined as the size of the largest set for Data analysis. In each of the experimen-
which the subject had perfect recall for final tal conditions, the mean reading time per
words in three of five sets of that size. If a word was calculated for four areas, as indi-
subject showed perfect recall in two of five cated by the square brackets:
groups of a given size, then half credit for
The [reporter that the senator]
that size was given. For example, a subject
1
who recalled final words correctly in all 2-
sentence groups, in four of the 3-sentence [attacked] [admittea [the error] . . . .
groups, and only two of the 4-sentence sets 2 3 4
received a span score of 3.5. The test sen- The four areas will be referred to below as
tences in each group were 13 to 16 words the initial area, the relative clause ending,
long and were unrelated to each other. (See the main verb, and the final areas, respec-
Daneman & Carpenter, 1980, 1983, for a tively. The time on the first word of each
further description of the Reading Span sentence (the determiner The) and on the
Test.) sentence final phrase was not analyzed be-
All subjects were classified according to cause these data were not crucial to any
their performance as having either being hypothesis. For the object relative sen-
High Span, Medium Span, or Low Span. tences, the two areas of primary interest
High Span subjects were those whose read- were the relative clause ending and main
ing spans were 4.0 or higher, while Low verb areas, which correspond to the loca-
Span subjects were those whose reading tions of the two verbs (attacked and admit-
spans were 2.5 or lower. All subjects whose ted); these are expected to be the loci of
reading spans were either 3.0 or 3.5 were difficulty when compared to the initial and
classified as Medium Span subjects. Given final areas. The corresponding segmenta-
that the aim of this work was to seek out tion of the subject relatives was identical,
differences between readers with large and except that the relative clause ending area
small working memory capacities, only the falls on the object, rather than on the verb,
data from High and Low Span subjects bear of the embedded relative clause.
directly on the hypotheses we present, and
only these data are analyzed below. In this The [reporter that attacked the]
and several other studies undertaken by our 1
research group, the performance of Me- [senator] [admitted] [the error] . . . .
dium span subjects has been generally in- 2 3 4
termediate between that of Low and High
Span subjects. For the reading time analysis, all obser-
Subjects. A total of 46 Carnegie Mellon vations that exceeded 1800 ms were
students participated in the experiment ei- trimmed to that value (less than 4% of the
ther to fulfill a course requirement or to re- times collected). The reading time results
ceive a $5.00 payment. The data from one reported below and following Experiment 2
subject, whose word-by-word reading include those trials in which the compre-
times were almost all greater than 2 s, and hension question interrogating the target
the data from two other subjects, whose sentence was answered incorrectly; how-
misunderstanding of the comprehension ever, the results are not substantially
task led them to respond “true” to every changed when such trials are excluded.
probe in the experiment, were discarded. Results and Discussion
The data considered below come from 24
High Span subjects and 22 Low Span sub- Recall Accuracy
jects. As expected, recall accuracy of sen-
 DIFFERENCES IN SYNTACTIC PROCESSING 587

tence-final words in the experimental task that object and subject relative sentences
was similar to that in the task that measures differ in the demands they make on working
span. Recall was worse for Low Span sub- memory, leaving differential amounts of re-
jects (7%) than for High Span subjects sidual capacity to retain a memory load.
(90%). These results provide a simple reli-
ability check. Moreover, recall was worse Comprehension Accuracy
for large memory loads (68%) than for small High Span readers were hypothesized to
memory loads (100% recall for one-word have sufficient computational resources to
sets). Also, because both Low and High comprehend both object and subject rela-
Span subjects could perfectly recall one- tive sentences at an equally high rate with-
word sets, whereas Low Span readers were out a memory load, whereas Low Span
much worse than High Span readers at re- subjects should be less accurate in their
calling three-word sets, there was a signifi- comprehension of the more difftcult object
cant interaction of Memory Load and Span. relatives, but not necessarily less accurate
An analysis of variance found all these ef- than High Span subjects on the easier sub-
fects to be quite reliable (with all p < .OOl). ject relatives. Figure 2 shows the compre-
As predicted, the recall of the largest hension data for both sentence types with
working memory loads was affected by the the level of working memory load increas-
syntactic complexity of the final, target ing from left to right in each panel. The data
sentence. The recall rate on three sentence for one-sentence trials show both a main
trials was 7% lower if the final sentence was effect of working memory capacity and an
an object rather than a subject relative. interaction between working memory ca-
This interaction of sentence type and load pacity and sentence type. Without any
was significant, with F(2,88) = 4.86, p < memory load, High Span subjects have a
.Ol. Similar results have been demonstrated better overall comprehension rate than do
using different paradigms, (e.g., Foss & Low Span subjects, (F(1,44) = 10.34, p <
Cairns, 1970; Wanner & Maratsos, 1977) .Ol), but this is particularly true for object
but not to our knowledge in an uninter- relative sentences, where the comprehen-
rupted reading task. This finding suggests sion difference between High and Low

NUMBER OF WORDS IN MEMORY LOAD SET

FIG. 2. Mean percentage of comprehension items answered correctly as a function of the number of
sentence-final words preceding the tested sentence. The parameter on the curve, “H” or “L,”
indicates whether comprehension was given by High or Low Span subjects, respectively.
588 KING AND JUST

Span readers is 22%, compared to a much Responses to relative clause probes were
smaller 6% difference on subject relative less accurate than responses to main clause
sentences. This predicted interaction be- probes (F(1,44) = 6.75, p < .02). The effect
tween working memory capacity and sen- of working memory load was in general ac-
tence type was reliable for one-sentence tri- cord with the pattern predicted, although
als (F(1,44) = 6.50, p < .02). the highly variable performance of Low
Figure 2 shows that the comprehension Span subjects noted above prevented the
rate for Low Span subjects on object rela- interaction of working memory capacity
tive sentences even without a memory load and other factors from being reliable. High
is so low that there is very little room for Span subjects show a 10% decline in their
their performance to deteriorate under a comprehension of object relative sentences
memory load. Indeed, further inspection of as the load increases from zero to two
the data indicated that 11 of the 22 Low words, but no decline in their comprehen-
Span subjects had overall comprehension sion of subject relative sentences. Low
rates that fell within a 95% confidence in- span subjects, who as a group show near
terval of chance, while only 2 of the 24 High chance-level performance on object rela-
Span subjects did so. If we label those sub- tives with no memory load, show no de-
jects whose overall comprehension was cline in their comprehension of object rela-
above chance the comprehending subjects tive sentences as the memory load in-
and those whose overall comprehension creases, but do show a decline in their
was at or near chance the noncomprehend- comprehension of the easier subject rela-
ing subjects, an interesting interaction be- tive sentences as their processing capacity
comes evident. With no memory load, non- is increasingly consumed by the demands
comprehending Low Span subjects gave of storing the memory load.
correct responses to object relative sen-
tences only 51% of the time, while compre- Reading Times
hending Low Span subjects averaged 81% The reading times of all comprehending
correct on these trials. This contrasts with subjects (as defined above) are shown in
the much smaller difference on computa- Fig. 3. (The performance of the non-
tionally easier subject relatives, where the comprehending subjects is reported below.)
noncomprehending subjects responded cor- These reading times provide striking sup-
rectly over 70% of the time while compre- port for the hypothesis that differences in
hending subjects responded correctly ap- working memory capacity affect the on-line
proximately 80% of the time. Thus, the processing of syntactic information. In gen-
comprehension of the noncomprehending eral, the difficulty of processing the critical
Low Span subjects was not just globally area was magnified by syntactic complex-
less accurate, but, as one would expect ity, and the amount of magnification was
from our theory, it was particularly poor on dependent on working memory capacity.
the more computationally demanding ob- The reading time results focus on the one-
ject relative sentences. The reading time sentence and two-sentence sets, because it
data will provide further evidence that non- was only in these cases that the comprehen-
comprehending and comprehending Low sion and recall performance were good
Span subjects differ primarily in their pro- enough to indicate that subjects understood
cessing of object relative sentences. the sentences and maintained the extrane-
Overall, as Fig. 2 indicates, High Span ous memory load. Reading times for the
subjects show better comprehension than three-sentence sets, while similar in general
Low Span subjects (F(1,44) = 17.77, p < appearance to the data presented here,
.OOl). The effect of sentence type was mar- were somewhat faster, particularly for low
ginally reliable (F(1,44) = 3.31, p < .08). span subjects. Subject reports indicated
 DIFFERENCES IN SYNTACTIC PROCESSING 589

looo- -1000
subjwt Ftelatives o@ctReliltives
setsL?es1md2 set.stzesland2

.L
SOO- ,/ \ -900
,..:' 3,
T ,..' \
...'
...'
E ./'
0 SOD- J' -WI
\
B ',
/
E ,/~ L....,,
700- ,,_,' "... \ -700
f H-" \
p '...
,.,,L./ l. /
‘L
8 /'I ," / '?
2 m- ,./ 1 -m
L ,...'
2
\-
f n i’
s SW- Hi tica
-7 \-
H
I
(The) senabr admnmd me t-W an&ted amtted me
rape- *rlol repoiler ermr
mat mat
albed me
me senamr
FIG. 3. Mean reading time per word for successive areas of the sentence. (The reading times for
parenthesized words are not included in the plotted points.) “H” and “L” are used to plot reading
times for High and Low Span subjects, respectively. The left panel shows reading times for subject
relatives, and the right panel shows the reading times for object relatives; these reading times are
combined from the no working memory load and one-word working memory load conditions.

that the faster reading times in the three- tactic complexity. Although the reading
sentence sets was part of an unsuccessful times increase in the computationally de-
strategy to rush through the sentence in or- manding areas for both sentence types, the
der to recall sooner the memory load mate- increase is much larger (by about 150 ms)
rials, so it is diffkult to interpret these for the object relatives in the right panel.
times, which are given in the Appendix. To Specifically, the reading time on the clause-
examine the reading time results, an ending word is reliably longer in object-
ANOVA contrasted the two set size condi- relative sentences than in subject relative
tions, the two syntactic constructions, and sentences, F(1,32) = 10.78, p c .Ol. Simi-
various words (areas) in the sentences as larly, the reading time is reliably longer on
within-subjects factors, and the span group the second word in the difficult region, the
as a between-subjects factor. main clause verb, in the case of object rel-
The left panel of Fig. 3 depicts the read- atives, F(1,32) = 23.99, p < .OOl. While
ing times for subject-relative sentences. High Span subjects spend significantly less
There is approximately a 60-ms difference time reading the main clause verb of both
in the reading times of High and Low Span sentence types (F(1,32) = 4.99, p < .05),
subjects, and both groups of subjects show the much sharper divergence of the curves
a similar selective increase in reading time for High and Low Span subjects in the right
at the relative clause-ending sector (senator panel of Fig. 3 suggeststhe predicted inter-
or attacked) and at the main verb (admit- action between reading span and sentence
ted). The contrast between the left panel type, which is indeed reliable (F(1,32) =
and the right panel, which depicts reading 4.26, p < .05). Although the High Span sub-
times for the harder object-relative sen- jects read more quickly in all four sectors of
tences, reveals the substantial effect of syn- the object relative sentences, their reading
590 KING AND JUST

time advantage over the comprehending the non-comprehending Low Span subjects
Low Span subjects is, as predicted, larger show some sensible pattern of reading
in the critical areas and smaller where less times? Because these subjects show little
processing is required; only in the main evidence of having comprehended the more
verb sector of the object relative sentences difftcult object relative sentences, it is pos-
is there a significant difference in reading sible they simply did not spend the time or
times between High and Low Span readers. resources necessary to construct the cor-
Low Span subjects are also the only group rect parse. Thus their reading times should
that shows any evidence of an effect of be similar to those of the comprehending
memory load on their reading times. As Low Spans except that they would be
predicted, their reading times are slower on spending significantly less time at the criti-
both subject and object relative sentences cal area of object relative sentences. Figure
under a one-word load, but the approxi- 4 shows that this is exactly what is happen-
mately 100 ms difference did not reach sig- ing. While reading times for both compre-
nificance due to the small number of com- hending and noncomprehending Low Span
prehending Low Span subjects. subjects were similar outside the critical
The reading times of the 11 Low Span area of the two sentence types, they were
subjects who comprehended at above- very different for the main verb. Here, the
chance levels provide evidence that the ef- Low Span comprehenders took reliably
fects of syntactic structure do interact with longer than the noncomprehenders , F( 1,20)
working memory capacity where the pro- = 6.63, p < .02, and there was an interac-
cessing difficulty is greatest, but this group tion between this factor and sentence type,
of subjects represents only half of the Low F(1,20) = 4.20, p = .05. This pattern of
Span subjects tested in this experiment. Do reading times suggests that the Low Span

t
1000
objecl Relative

FIG. 4. Mean reading time per word for successive areas of the sentence for Low Span readers only.
“C” and “N” are used to plot reading times for Low Span subjects who comprehended at above
chance and for those who did not comprehend at above chance, respectively. Reading times for
subject relatives are on the left, and reading times for object relatives are on the right. As in Fig. 3,
these reading times are combined from the no working memory load and one-word working memory
load conditions.
 DIFFERENCES IN SYNTACTIC PROCESSING 591

noncomprehenders did not spend the time greater use of other kinds of information in
necessary to associate actors with their ap- their processing of complex syntactic struc-
propriate thematic roles, particularly in ob- tures to compensate for their lack of capac-
jective relative sentences where such as- ity. Experiment 2 was designed to replicate
signments were most demanding. the basic findings of Experiment 1 and in-
These reading time results when com- vestigate how these Low Span subjects
bined with the comprehension results dis- might differ from High Span subjects in
cussed above provide converging sources their use of pragmatic information.
of evidence for the hypothesis that individ-
uaI differences in working memory capacity EXPERIMENT 2
can be an important factor in syntactic pro- One surprising result of Experiment 1
cessing. On the less computationally de- was the very low level of comprehension
manding subject relative sentences with no that Low Span subjects showed for the ob-
memory load, Low Span readers performed ject relative target sentences. One half of all
much like High Span subjects both in their Low Span subjects comprehended these
comprehension and in their on-line reading sentences at chance level. These readers
times. On the more computationally de- may normally rely on extra-syntactic infor-
manding object relatives, however, Low mation in order to build correct representa-
Span subjects were much poorer in their tions, and their efforts may fail in the ab-
comprehension, and even those Low Spans sence of such information. A Low Span
who showed adequate comprehension reader could make use of the pragmatic as-
spent much longer reading on-line in the sociations between each of the nouns and
critical sector of these sentences than did the verb in order to help decide which of
High Span subjects. Furthermore, those two possible nouns is the subject and agent
Low Span subjects who showed some evi- of the relative clause. For example, if the
dence of having comprehended object rela- two possible agents of the verb rescued
tive sentences spent more time in the criti- were the fireman and the robber, the prag-
cal area of the sentence than those who matic association between rescued andfire-
showed little evidence of having compre- man might lead a Low Span subject to rep-
hended object relative sentences. The pri- resent fireman as the agent of rescued. Al-
mary effect of a small memory load was to though this kind of pragmatic information is
reduce the comprehension of High Span potentially misleading, it could be an attrac-
subjects on more difficult object relative tive option for Low Span subjects, whose
sentences; this effect did not occur for Low lack of extra processing and storage re-
Span subjects apparently because their sources limit their use of other syntactic
comprehension of object relatives was cues. Because pragmatic associations be-
nearly at chance in the first place. tween nouns and verbs are independent of
The poor performance of many Low syntactic relationships, their use requires
Span subjects on object relative sentences only the retrieval of familiar stored infor-
in this experiment suggests that they were mation rather than computationally expen-
not processing some critical syntactic cues sive syntactic processing. Further, the
effectively, apparently because they lacked pragmatic bias of a verb towards one of two
the working memory capacity to do so. possible agents occurs at precisely the
Fortunately for these subjects, pragmati- point where the information would be use-
cally unbiased and contextually isolated ob- ful, and thus does not tax the storage func-
ject relative sentences, such as the critical tions of working memory. Indeed, an early
sentences in Experiment 1, are rather rare experiment by Stolz (l%i’) demonstrates
in normal reading material. It seems possi- that people can use pragmatic cues to de-
ble that Low Span readers may make code doubly embedded object relative
592 KING AND JUST

clauses whose syntactic structure they may Working memory capacity should deter-
not be able to evaluate. Experiment 2 in- mine not only how much a subject will rely
vestigates how pragmatic information af- on pragmatic information to comprehend
fects the processing of object relative sen- object relative sentences, but also at which
tences by readers who differ in working point in processing this information will be
memory capacity. most beneficial. Low span subjects, who
The target sentences of Experiment 2 are able to retain less information in work-
were all object relative sentences contain- ing memory while reading, should benefit
ing verbs that either did or did not provide most when pragmatic information is pro-
strong pragmatic cues as to which of the vided at the point in processing where it is
two potential actors in the sentence was the needed. By contrast, pragmatic information
agent of the given verb. There were four which occurs after the point where it is
different biasing conditions: both verbs needed should be more difftcult for Low
were pragmatically biased, only the relative span readers to use effectively. Specifi-
clause verb was biased, only the main cally, pragmatic information provided in
clause verb was biased, or neither verb was the main clause verb phrase, which occurs
biased, as shown in the examples below: after the end of the relative clause, might be
of no help to Low Span readers in compre-
The robber that the fireman rescued stole
hending the relative clause, Readers with
the jewelry. (both)
higher working memory capacities, whose
The robber that the fireman rescued
capacity for syntactic processing makes
watched the program. (relative)
them less dependent on pragmatic informa-
The robber that the fireman detested
tion, should comprehend the relative clause
stole the jewelry. (main)
similarly regardless of whether pragmatic
The robber that the fireman detested
information was provided during the pro-
watched the program. (neither)
cessing of that clause. In general, the dif-
The last sentence type, with no pragmatic ferences between High and Low Span sub-
cues associated with either verb, resembles jects should be much smaller when the
the type of sentences used in Experiment 1. main clause is queried than when the rela-
If Low Span readers make substantial tive clause is queried, primarily because the
use of pragmatic information to construct main clause is easier to process, and all the
referential representations, then these prag- pragmatic information available will occur
matic cues should at least partially relieve before or at the point in the sentence where
them of the need to process syntactic cues it is most needed.
based on nonlocal information such as In contrast to its large effect on compre-
word order, which they were poor at using hension, the immediate effects of pragmatic
in Experiment 1. Thus, if both the main information on syntactic processing mea-
clause and the relative clause verb are prag- sured with on-line techniques have proven
matically biased towards their agents, Low difficult to find in the past (e.g., Clifton &
Span readers should perform comparably Ferreira, 1987). Given that our materials
to readers who have a greater working are manipulating a pragmatic variable that
memory capacity. As the amount of prag- is only heuristic in value and does not di-
matic information decreases, the compre- rectly affect syntactic well-formedness, it is
hension performance of the Low Span sub- likely that considerable syntactic process-
jects should diverge from that of the other ing will occur before this information af-
subjects. Indeed, overall performance in fects processing. This does not mean that
the condition where neither verb is biased the pragmatic cueing is unimportant, (it
should resemble that observed under no may be quite important for Low Span sub-
memory load in Experiment 1. jects) but that it may not have as large an
 DIFFERENCES IN SYNTACTIC PROCESSING 593

effect on reading times as does working while the third line contained the compre-
memory capacity. Indeed, without any ad- hension probe. There were live practice tri-
vance context, this pragmatic information als followed by the 32 experimental and 68
may not become available rapidly enough filler sentences presented in a different ran-
to guide a reader’s initial parsing efforts, so dom sequence for each subject.
that the factors of working memory and Subjects. The subjects were 48 students
pragmatic bias may not interact in the read- of Carnegie Mellon or the University of
ing time data. Pittsburgh who were native monolingual
speakers of English and served either for
Method course credit or a $5.00 payment. The data
Materials. The 32 experimental sen- for one additional subject, whose mean
tences each contained an object relative word-by-word reading times were twice
clause, with eight sentences in each of the that of the other subjects, were discarded
four conditions: sentences either had prag- without further analysis. All subjects
matic cues for both verbs, the relative whose reading span scores were not previ-
clause verb, the main clause verb, or nei- ously known were given the Reading Span
ther verb. All of the experimental sentences Test before they did the primary experi-
were 13-15 words in length, with the last mental task. Subjects were classified as be-
four to six words of the sentence simply ing High or Low Span as in Experiment 1
supplying a prepositional or adverbial and the results reported below derive from
phrase modifier. An example target sen- 24 High Span and 24 Low Span subjects.
tence from the both condition would be The Data analysis. As in Experiment 1, all
robber that the fireman rescued stole the reading times that exceeded 1800 ms (less
jewelry from the heavy steel safe. than 2% of the total) were trimmed to that
There were equal numbers of probes test- value. For each of the experimental trials,
ing comprehension of the relative and main the mean reading time per word was calcu-
clauses of the experimental sentences, and lated for each of the same four areas of the
equal numbers of “True” and “False” object relative clause as in Experiment 1:
items were used in each condition. The
The [robber that the fireman] [rescued]
probe items in this experiment were con-
1 2
structed in the same way as the probes in
Experiment 1. Sixty-eight other sentences [stole] [the jewelry] from the heavy
of similar length (but not containing object 3 4
relative clauses) were used as fillers. Com-
steel safe,
prehension items for fillers often tested for
information that was not explicitly given in Results and Discussion
order to discourage subjects from adopting
artificial reading strategies. Comprehension Accuracy
Procedure. The 100 sentences were pre- As the amount of pragmatic information
sented word-by-word on the screen of an decreases, readers have to rely more on the
IBM-PC/XT, using the moving window strictly syntactic cues present, and compre-
program described in Experiment 1. Other- hension accuracy should decrease. Figure 5
wise, the procedure was simpler because shows that comprehension for all subjects
memory load effects were not being as- was inferior following sentences in which
sessed. The initial display for each trial neither verb was pragmatically biased.
consisted of three lines of dashes (where Overall comprehension accuracy for trials
letters would appear) and blanks; the first with both verbs biased was 90% and was
two lines contained the sentence (the break almost as high when either the main or rel-
being located in the sentence-final phrase) ative clause verbs were biased (88% and
594 KING AND JUST

L---l
SEMANTIC CUES FOF4 SOTH. THE RELATIVE CLAUSE, THE MAIN CLAUSE. OR NEITHER VERB
FIG. 5. Mean percentage of comprehension items answered correctly as a function the pragmatic
bias, which fell on both verbs (B), the relative clause verb (R), the main clause verb (M), or neither
verb (N). On the left, comprehension is plotted for trials where the relative clause verb was queried,
and on the right for trials where the main clause verb was queried. The parameter on the curve, “H”
or “L, ” indicates whether comprehension is being plotted for High or Low span subjects, respec-
tively.

87%, respectively). When neither verb was subjects were not affected by the presence
pragmatically biased, however, compre- or absence of pragmatic information when
hension accuracy dropped to 78%. This the relative clause was tested, although
pattern of performance led to a significant their comprehension accuracy was lower
effect of pragmatic bias overall, with overall when the relative clause was tested.
F(3,138) = 8.90 (p < .OOl). This uneven contribution of verb bias infor-
Comprehension of the main and relative mation led to a significant two-way interac-
clauses differed reliably. When the infor- tion of pragmatic information condition and
mation being tested was from the main the clause probed (F(3,138) = 2.94, p <
clause, subjects responded accurately over .05).
90% of the time, but when the information Differences in working memory capacity
was from the relative clause, subjects were had a large effect on the comprehension re-
correct only 80% of the time, F(1,46) = sponses to more difficult relative clause
26.03 (I, < .OOl). This latter result corre- probes, but not on responses to easier main
sponds well with the 85% accuracy rate clause probes. Thus, as Fig. 5 shows, High
found in a sentence paraphrase task, which Span readers showed better comprehension
requires that both clauses be paraphrased than Low Span readers primarily when the
(Larkin & Burns, 1977). As predicted, com- relative clause is probed, and the interac-
prehension performance was poorest (74%) tion of working memory capacity and the
on trials that queried the relative clause clause probed is reliable (F( 1,46) = 5.04, p
when neither verb was pragmatically bi- < .05). This result contrasts with the
ased. This compares with an average com- smaller overall comprehension advantage
prehension rate of 82% in the three other of High Span readers, which is not quite
biasing conditions. In the case of main reliable (F( 1,46) = 3.42, p < .08). Figure 5
clause comprehension, the difference be- also indicates that pragmatic bias is quite
tween the neither verb biased condition helpful to Low Span subjects in their com-
(82%) and the other three biased verb con- prehension of relative clauses, but not at all
ditions (93%) was even larger. High span helpful to High Span subjects. This pattern
 DIFFERENCES IN SYNTACTIC PROCESSING 595

contrasts with the comprehension data for

the main clause probes, where both High
600
and Low Span readers react similarly to
pragmatic information. Because the differ-
I
ence in the helpfulness of pragmatic bias to
High and Low Span groups is only apparent
for relative clause probes, the overall inter-
action between working memory and prag-
matic bias is unreliable (F(3,138) = 1.84, p
> .l). None of the other interactions in this
analysis approach reliability.
Experiment 2, therefore, has enabled us
to further restrict the locus of the compre-
hension difficulty experienced by Low
Span subjects. They have difficulty com-
prehending the information in the relative
clause, particularly if no pragmatic infor-
mation is available. High Span subjects
comprehend well if either or both verbs
provide pragmatic information. Even when FIG. 6. Mean reading time per word for successive
there is no pragmatic information, they still areas of the sentence,comprehending subjects only.
perform better than Low Span subjects (As in Fig. 3, parenthesizedwordsare not included in
when answering questions about the con- the plotted points.) As before, “H” and “L” are used
tent of the relative clause, but no better to plot reading times for High and Low span subjects,
respectively.
when answering questions about the con-
tent of the main clauses. This pattern of the main clause 116 ms faster than Low
results suggest High Span subjects can use Span subjects at the verb of the main
pragmatic information in their construction clause. An ANOVA indicates that the only
of either the clause where the information is significant difference between High and
provided or the clause where it is absent, Low Span subjects occurs at the verb of the
whereas Low Span subjects can only use main clause, where F( 1,46) = 4.68, p < .05.
pragmatic information locally. To sum up, While differences in working memory ca-
the comprehension results provide strong pacity had large and reliable effects on
evidence that working memory capacity reading times, differences in pragmatic bi-
plays an important role in the use of prag- asing had only weak effects on reading
matic information in comprehending diffi- times. In particular, if both verbs or the rel-
cult syntactic structures. ative clause verb were pragmatically bi-
ased, readers spent an average of 597 ms on
Reading Times the main clause verb, while if only the main
The results of Experiment 1 indicated clause verb or neither verb was pragmati-
that the difficulty in processing object rela- cally biased, readers spent an average of
tive sentences was restricted to the critical 624 ms, an effect of only marginal reliability
area of the relative clause and main clause (F(1,46) = 2.93, .05 <p < .lO). There was
verb, and that Low Span subjects had rel- no evidence of any interaction between
atively more difficulty in the critical area working memory capacity and pragmatic
than High Span subjects. Figure 6 shows bias in the reading time data; indeed, the
that the reading times for each group col- pragmatic bias effect on main verb reading
lapsed over the four biasing conditions. times was virtually identical for both High
High Span subjects processed the verb of and Low Span readers.
5% KING AND JUST

In summary, the reading time results rep- ever, seems less compelling if there is only
licate the basic effect of Experiment 1, a single simple noun phrase to be main-
which is that Low Span readers spend reli- tained while parsing the inner clause, al-
ably more time processing syntactically dif- though Wanner and Maratsos (1978) have
ficult sentences at the point where the pro- found measurable loading effects. Storage-
cessing is most difftcult, but still compre- based accounts usually assume that the
hend more poorly than those subjects who length of the interruption is important in or-
have greater working memory capacity, der to explain why subject relative sen-
particularly when the relative clause is que- tences (which have a short interruption) are
ried and when there is no pragmatic infor- easier to process than object relative sen-
mation available. tences (which have a longer interruption).
Yet other interrupting structures, such as a
GENERAL DISCUSSION relatively long appositive phrase in the mid-
This research shows that individual dif- dle of a sentence like this one, do not seem
ferences in effective working memory ca- nearly as difficult to process and compre-
pacity can cause systematic differences in hend. Similarly, an increase in the interrup-
the processing of complex syntactic struc- tion length might explain the increase in
tures. In both experiments, low capacity reading time at and following the relative
readers showed poorer comprehension de- clause verb as a difficulty in reactivating
spite spending more time processing the the memory trace of the head noun, but in
critical portions of object relative sen- fact the accessibility of the first actor in a
tences. In addition, the results of Experi- sentence is actually quite high, even when
ment 1 showed that high capacity readers that actor is not the agent of the action de-
showed reduced comprehension of object picted and accessibility is tested near the
relative sentences when an increased work- end of a sentence (Gernsbacher, Har-
ing memory load reduced the resources greaves, & Beeman, 1989). It is unlikely,
available for comprehension. Experiment 2 however, that this representation is kept
showed that an increase in the availability active without some pressure on working
of pragmatic information helped all readers, memory, as the model presented below will
but particularly Low Span readers, in their indicate.
comprehension of object relative sen- Role assignment. Models of syntactic
tences. In this discussion, we will describe processing that emphasize the assignment
how individual accounts that emphasize of syntactic and thematic roles of constitu-
storage, role assignment, and perspective ents naturally localize these operations to
maintenance could partially explain these just the places where elevated reading
results and how aspects of these accounts times are observed-at the relative clause
have been combined in a larger computa- and main clause verbs. Indeed, a strictly
tional model (Just & Carpenter, 1991). syntactic account together with an assump-
Storage. The classical explanation for tion about processing difficulty can explain
the observed difficulty in parsing object rel- many of our results. As explained in the
ative constructions primarily rests on the introduction, extra computational re-
difficulty of maintaining a representation of sources must be used to determine that the
the head noun in working memory while gap in the relative clause has been found,
processing the interrupting clause (e.g., and then to make multiple role assignments
Miller & Chomsky, 1963). Indeed, the dif- in case of the object relative clause and a
ficulty of a double center embedding (The single role assignment in the case of the
reporter that the senator that the schoolboy subject relative clause. The role assignment
. . .) , would seem to arise from this double account correctly predicts that readers who
interruption. A strict storage account, how- have fewer computational resources avail-
 DIFFERENCES IN SYNTACTIC PROCESSING 597

able would have more difficulty at this relative sentences, since there is only one
stage, and the model presented below in- role assignment to make in both cases. Sim-
cludes a similar mechanism. Pragmatic in- ilarly, it is unclear why differences in read-
formation, if available, can be used to help ing times explained by working memory ca-
assign roles to the actors represented by pacity should appear here, where process-
noun phrases at some point in the process. ing should be relatively easier than at the
Reading time results from Experiment 2 relative clause verb. One explanation is
suggest that relatively high-level pragmatic that readers may “miss” the gap in the rel-
biasing does not have an early effect on the- ative clause and realize their mistake only
matic processing, but it does have an even- when reading the following main clause
tual effect; pragmatic information weaker verb, but then it would be difficult to see
than animacy constraints are probably not why reading times are also elevated at the
used on-line, but can help lessen the com- relative clause verb, unless the extra time
prehension gap between high and low ca- was being used to find the gap and begin
pacity readers. making role assignments to the unattached
Although the role assignment account noun phrases. Thus a complete computa-
makes accurate predictions about some of tional model will have to include other pro-
the data, there are two problems with this cesses in addition to simple role assign-
account, considered alone. The first prob- ment.
lem is that it is unclear why the assignments Resolving conjkfs in thematic assign-
made in the object relative clause should be ment and perspective maintenance. It has
particularly difficult per se. In English, an been long assumed that it is difficult to
embedded clause of the form NPI- make two different and conflicting thematic
(that)-NP2-V always has the directly pre- assignments to a single noun phrase/actor
verbal noun phrase (NP2) as its syntactic (Bever, 1970, provides one early account).
subject and the other noun phrase (NPl) as This hypothesis accounts for the relative
its object; there is never any ambiguity. In- difficulty of object relative sentences, in
deed, adult speakers of English are very which the agent of the main clause is the
consistent in their use of word order infor- patient of the relative clause, when com-
mation, which is both highly predictive and pared to subject relative sentences, in
highly available even when it conflicts with which the agent of the main clause is also
cues like animacy or verb agreement the agent of the relative clause. The princi-
(MacWhinney, Bates, & Kliegl, 1984). Fur- ple of perspective maintenance gives a
ther, if the clause is active and the verb is slightly different (and more general) expla-
strictly transitive (as in the present experi- nation for this difficulty in the main clause
ments), then the thematic roles of agent and of object relative sentences. This principle
patient directly map onto the syntactic sub- suggests that the difficulty in parsing cer-
ject and object, respectively. Despite these tain relative clause constructions arises
facts, however, making these assignments from the need to switch perspective from
in object relative sentences took a substan- one actor to another in the construction of
tial amount of time in the current experi- the referential representation for a sentence
ments and led to a large number of regres- (MacWhinney & Pleh, 1988). In a subject
sive eye movements in the study done by relative sentence, the first actor is the sub-
Holmes and O’Regan (1979). ject (and often the agent) of both clauses,
A more important difficulty with this ac- and a reader can build a consistent referen-
count of our data is that it does not directly tial representation of the sentence by sim-
explain why reading times should also be ply adding new facts to what is already
elevated on the main clause verb of the ob- known about that actor. In an object rela-
ject relative sentence compared to subject tive clause, however, the actor described in
598 KING AND JUST

the initial noun phrase is not the subject or Most productions do not do their full work
agent of the relative clause, and since read- in a single cycle, but reiteratively increment
ers tend to follow the point of view of the the activation levels of their target working
agent as a sentence progresses, a reader memory elements over several cycles until
would begin building a referential represen- some specified threshold value is reached.
tation centered on the second actor. But These unconventional features make the
when the main verb is reached, the reader processing of both READER and CC
must shift perspective back to the first ac- READER closely resemble the processing
tor. Attentional shifts draw resources away of connectionist networks.
from other processes, degrading their per- The greatest difference between CC
formance; this is particularly a problem for READER and older versions of the
Low Span readers, who have fewer re- READER model is that CC READER di-
sources to spare. The principle of thematic rectly imposes constraints on processing by
conflict and the principle of perspective limiting the total amount of activation avail-
maintenance both reflect the importance of able on each cycle. When the limit on a
higher level processing to a sentence pro- given cycle would be exceeded, the actual
cessing model that can include an account activation that each production propagates
of individual differences. is scaled back in proportion to the amount
A capacity constrained model of parsing. requested. As one would expect, this limi-
The three accounts discussed above are ac- tation on activation applies primarily when
tually not mutually exclusive, but can be there are either a large number of partial
integrated into a single capacity con- products that must be maintained in work-
strained parsing model that implicitly in- ing memory or when many different pro-
cludes aspects of each of them (Just & Car- ductions are firing at the same time. (In this
penter, 1991). The new model is a modified, system, both storage and active processing
capacity constrained version of the are the result of production firings.) CC
READER simulation model (Thibadeau, READER can model individual differences
Just, & Carpenter, 1982; Just & Thibadeau, in working memory capacity either by vary-
1984) called Capacity Constrained ing the total amount of activation available,
READER (CC READER for short). CC or by varying the efficiency with which par-
READER, like READER before it, is ticular productions can activate their target
based on a production system model that working memory elements, or both. In
performs computations by matching pro- practice, the first of these is generally used
duction rules to the contents of a working and is discussed below.
memory and performing the action(s) of When parsing sentences with relative
each matching production to alter the con- clauses, CC READER reflects the partial
tents of working memory. An unconven- truth of both storage-based and role assign-
tional aspect of this production system ment-based models. Maintaining the head
model is that it allows all of its productions of an object relative clause in working
whose conditions are satisfied on a given memory is a source of difficulty for CC
cycle to fire in parallel, with each produc- READER because the representation of the
tion accomplishing its function by incre- head of the object relative receives less
menting or decrementing the activation lev- maintenance support from other items in
els of various elements in working memory. working memory until it is well integrated
Individual productions change the activa- into higher level structures. Similarly, it is
tion level of a given working memory ele- more difficult to perform simultaneous mul-
ment by computing a weighted function of tiple role assignments because the produc-
the activation levels of the working mem- tions that perform the assignments end up
ory elements matched by the production. competing against each other for activa-
 DIFFERENCES IN SYNTACTIC PROCESSING 599

tion. And it is more difficult to assign a sec- productions that either draw away re-
ond, conflicting thematic role to the head of sources from other levels of processing or
the object relative clause because the two provide data that arrives too late to be fully
different thematic structures also end up integrated into higher levels of processing.
competing for maintenance activation. CC Indeed, there are slight differences in the
READER solves this last difficulty while it way the lexical access processes of High
is still processing the main verb of an object and Low Span readers respond to a work-
relative sentence by gradually building a ing memory load (Carpenter & Just, 1989),
more complex thematic representation that although higher level differences are more
adequately supports both role assignments. pronounced.
This more complex structure is in some The differences between High and Low
ways equivalent to the consolidation of dis- Span readers in their comprehension of ob-
course perspective predicted by the per- ject relative clauses imply that the supply
spective maintenance model, although the and efficient use of cognitive resources may
referential level of CC READER is still differ from person to person. The diversity
quite limited. and sometimes great specificity of acquired
CC READER also successfully models language processing deficits also supports
individual differences between High and this hypothesis, and some recent theories
Low Span readers in their processing of ob- of cognitive functioning before and after
ject relative sentences. If the activation ca- brain damage make explicit mention of the
pacity limit is set high enough, the model allocation and use of cognitive resources
will always make the correct thematic role (Shallice, 1988). It is possible that such dif-
assignments within a reasonable number of ferences in resource use could produce sub-
cycles, but if the activation capacity limit is stantial variation in “normal” people as
lowered, the model begins to take substan- well. Individual differences in syntactic
tially longer and can even make errors. Just processing exist, just as individual differ-
and Carpenter (1991) include more informa- ences in mechanical reasoning or writing
tion on the qualitative fit of the CC ability exist, because complex cognitive
READER model not only to the data from phenomena arise from the interaction of
this study, but also from other studies in- many different processes and subpro-
vestigating the role of working memory ca- cesses, each of which makes demands on
pacity in language comprehensions. limited cognitive resources.
Although this research has focused on
higher level processes, individual differ-
APPENDIX 1
ences are certainly not confined to this
MEANREADINGTIMESFORTWO-WORD
level of processing, and the CC READER
LOADTRIALS IN EXPERIMENT 1
model is in principle extendable to other
(STANDARD ERRORS IN PARENTHESES)
levels of processing. For example, one
could postulate in the model that word- Group
encoding and lexical access processes are span
Area High Low span
indirectly responsible for the comprehen-
sion differences observed in our experi- Subject relative sentences
Initial 529 (27) 574 (29)
ments and see if the modified CC READER Rel. clause 619 (53) 713 (55)
provided a better or worse tit of the data. Main clause 631 (45) 809 (75)
The speed of lexical access has a small but Final 530 (28) 641(34)
reliable correlation with measures of read- Object relative sentences
ing proficiency (Hunt, Lunneborg, & Initial 502 (26) 581 (33)
Lewis, 1975; Jackson & McClelland, 1979), Rel. clause 665 WI 758 (70)
which suggests that some Low Span read- Main cause 708 (47) 767 (39)
ers could have less efficient lexical access Final 588 (32) 676 (59)
600 KING AND JUST

APPENDIX 2: MATERIALS USED IN Main Clause Verb Phrase Biased

EXPERIMENT 2
The broker that the banker admired pur-
Both Verb Phrases Biased chased the stock in the new hi-tech com-
The banker that the teacher instructed paw.
The fireman that the lawyer insulted ex-
approved the loan after asking a few ques-
tinguished the blaze in the old abandoned
tions . house.
The lawyer that the doctor examined The reporter that the salesman ignored
drafted the contract before leaving the of- wrote the editorial about the increase in
fice. fraud.
The salesman that the accountant audited The barber that the waiter disliked sharp-
sold the merchandise at a rather large dis-
ened the razor with a fine oiled stone.
count. The comedian that the detective despised
The waiter that the comedian amused delivered the punchline at the crowded
opened the wine carefully to avoid spilling night club.
any. The accountant that the judge liked bal-
The detective that the barber shaved in- anced the books only with great difficulty.
vestigated the accident at the suburban The doctor that the robber phoned pre-
shopping center. scribed the medication despite the risks.
The judge that the reporter interviewed The teacher that the governor praised
dismissed the charge at the end of the hear- graded the essay in the large busy office.
ing.
The robber that the fireman rescued stole Neither Verb Phrase Biased
the jewelry from the heavy steel safe.
The governor that the broker advised The banker that the barber praised
signed the legislation just before the end of climbed the mountain just outside of town.
the year. The lawyer that the reporter phoned
cooked the pork chops in their own juices.
Relative Clause Verb Phrase Biased The salesman that the fireman liked dom-
inated the conversation about the pennant
The comedian that the banker funded an- race.
swered the telephone after the third ring. The waiter that the broker despised
The accountant that the lawyer sued read drove the sportscar home from work that
the newspaper on the train to New York. evening.
The doctor that the salesman assisted The detective that the teacher disliked
watched the movie later with several clipped the coupons out with the dull scis-
friends. sors.
The teacher that the waiter served The judge that the doctor ignored
clipped the coupons out of the newspaper. watched the movie about Colombian drug
The broker that the detective questioned dealers.
drove the sportscar through the tiny village. The robber that the accountant insulted
The fireman that the judge sentenced read the newspaper article about the fire.
dominated the conversation in the bar that The governor that the comedian admired
evening. answered the telephone in the fancy restau-
The reporter that the robber mugged rant.
cooked the pork chops in a large metal fry-
ing pan. REFERENCES
The barber that the governor endorsed
climbed the mountain on the following BADDELEY, A. D. (1986). Working memory. Oxford:
weekend. Clarendon Press.
 DIFFERENCES IN SYNTACTIC PROCESSING 601

BEVER, T. G. (1970). The cognitive basis for linguistic What does it mean to be high verbal? Cognitive
structures. In J. R. Hayes (Ed.), Cognition and Psychology, 2, 194-227.
the development of language. New York: Wiley. JACKSON, M. D., & MCCLELLAND, J. L. (1979). Pro-
CARPENTER, P. A., L JUST, M. A. (1989). The role of cessing determinants of reading speed. Journal of
working memory in language comprehension. In Experimental Psychology: General, 108, 151-181.
D. Klahr & K. Kotovsky (Eds.), Complex infor- JUST, M. A., & CARPENTER, P. A. (in press). A ca-
mation processing: The impact of Herbert A. Si- pacity theory of comprehension: Individual differ-
mon. Hillsdale, NJ: Erlbaum. ences in working memory. Psychological Review.
CLIFTON, C., JR. & FERREIRA, F. (1987). Modularity JUST, M. A., & CARPENTER, P. A. (1980). A theory of
in sentence comprehension. In J. L. Garfield reading: From eye fixations to comprehension.
(Ed.), Modularity in knowledge representation Psychological Review, 87, 329-354.
and natural language understanding. Cambridge, JUST, M. A., & CARPENTER, P. A. (1987). The psy-
MA: MIT Press. chology of reading and language comprehension.
Newton, MA: Allyn & Bacon.
DANEMAN, M., t CARPENTER, P. A. (1980). Individ-
ual differences in working memory and reading. JUST, M. A., CARPENTER, P. A., & WOOLLEY, J. D.
Journal of Verbal Learning and Verbal Behavior, (1982). Paradigms and processes in reading com-
19, 450-466. prehension. Journal of Experimental Psychology:
General, 111, 228-238.
DANEMAN, M., C CARPENTER, P. A. (1983). Individ-
ual differences in integrating information between JUST, M. A., & THIBADEAU, R. H. (1984). Developing
and within sentences. Journal of Experimental a computer model of reading times. In D. E.
Psychology: Learning, Memory, and Cognition, Kieras & M. A. Just (Eds.), New methods in read-
ing comprehension research (pp. 34%364). Hills-
9, 561-584.
dale, NJ: Erlbaum.
FORD, M. (1983). A method of obtaining measures of
KIMBALL, J. P. (1973). Seven principles of surface
local parsing complexity throughout sentences.
Journal of Verbal Learning and Verbal Behavior, structure parsing in natural language. Cognition,
2, 15-47.
22, 203-218.
Foss, D. J., & CAIRNS, H. S. (1970). Some effects of LARKIN, W., & BURNS, D. (1977). Sentence compre-
memory limitation upon sentence comprehension hension and memory for embedded structure.
and recall. Journal of Verbal Learning and Verbal Memory and Cognition, 5, 17-22.
Behavior, 9, 541-547. MACDONALD, M. (1989). Priming effects from gaps to
FRAZIER, L., & FODOR, J. D. (1978). The sausage ma- antecedents. Language and Cognitive Processes,
chine: A new two-stage parsing model. Cognition, 4(l), 35-56.
6,291-325. MACWHINNEY, B., BATES, E., & KLIEGL, R. (1984).
GERNSBACHER, M. A., HARGREAVES, D. J., & BEE- Cue validity and sentence interpretation in En-
MAN, M. (1989). Building and accessing clausal glish, German, and Italian. Journal of Verbal
representations: The advantage of first mention Learning and Verbal Behavior, 23, 127-150.
versus the advantage of clause recency. Journal of MACWHINNEY, B., & PLEH, Cs. (1988). The process-
Memory and Language, 28, 735-755. ing of restrictive relative clauses in Hungarian.
Cognition, 29, 95-141.
GIBSON, E. (1990). Recency preferences and garden-
path effects. In The Proceedings of the Twelfh MILLER, G. A., & CHOMSKY, N. (1%3). Finitary mod-
Annual Conference of the Cognitive Science So- els of language users. In R. D. Lute, R. Bush, &
ciety, pp. 372-379. E. Galanter (Eds.), Handbook of mathematical
psychology, Vol. 2. New York: Wiley.
GLEITMAN, L. R., & GLEITMAN, H. (1970). Phrase
and paraphrase; Some innovative uses of lan- NEWELL, A. (1973). Production systems: Models of
guage. New York: Norton. control structures. In W. G. Chase (Ed.), Visual
Information Processing. New York: Academic
HOLMES, V. M. (1987). Syntactic parsing: In search of Press.
the garden path. In M. Coltheart (Ed.), Attention
and Pe$ormance XII: The Psychology of Read- PERFE’ITI, C. A., & LESGOLD, A. M. (1977). Dis-
ing. Hillsdale, NJ: Erlbaum. course comprehension and sources of individual
differences. In M. A. Just & P. A. Carpenter
HOLMES, V. M., L O’REGAN, J. K. (1981). Eye lixa- (Eds.), Cognitive processes in comprehension.
tion patterns during the reading of relative clause Hillsdale, NJ: Erlbaum.
sentences. Journal of Verbal Learning and Verbal
SHALLICE, T. (1988). From neuropsychology to men-
Behavior, 20, 417-430. tal structure. Cambridge, UK: Cambridge Univer-
HUNT, E. B., LUNNEBORG, C., & LEWIS, J. (1975). sity Press.
602 KING AND JUST

STOLZ, W. S. (1967). A study of the ability to decode WANNER, E., & MARATSOS, M. (1978). An ATN ap-
grammatically novel sentences. Journuf of Verbal preach to comprehension. In M. Halle, J. Bres-
Learning and Verbal Behavior, 6, 8674373. nan, & G. A. Miller (Eds.), Linguistic Theory and
THIBADEAU, R., JUST, M. A., & CARPENTER, P. A. Psychological Reality. Cambridge, MA: MIT
(1982). A model of the time course and content of Press.
reading. Cognitive Science, 6, 157-203.
TURNER, M. L., & ENGLE, R. W. (1989). Is working
memory capacity task dependent? Journal of (Received July 30, 1990)
Memory and Language, 28, 127-154. (Revision received December 23, 1990)