Caplan, David and Waters, Gloria (1998) Verbal Working Memory and Sentence Comprehension

#This is the unedited penultimate draft of a BBS target article that has been accepted for publication (Copyright 1998: Cambridge University Press) and is being circulated for Open Peer Commentary. This preprint is for inspection only, to help prospective commentators decide whether or not they wish to prepare a formal commentary. Please do not prepare a commentary unless you have received the hard copy, invitation, instructions and deadline information.

For information on becoming a commentator on this or other BBS target articles, write to:

For information about subscribing or purchasing offprints of the published version, with commentaries and author's response, write to: (North America) or (All other countries).#

Verbal Working Memory and Sentence Comprehension

David Caplan

Neuropsychology Laboratory,
Vincent Burnham 827,
Massachusetts General Hospital
Fruit Street, Boston, MA

Gloria S. Waters
Department of Communication Disorders
Boston University


working memory, syntactic processing, sentence comprehension


This target article discusses the verbal working memory system used in sentence comprehension. We review the idea of working memory as a short duration system in which small amounts of information are simultaneously stored and manipulated in the service of a task and that syntactic processing in sentence comprehension requires such a storage and computational system. We inquire whether the working memory system used in syntactic processing is the same as that used in verbally mediated tasks involving conscious, controlled processing. Various forms of evidence are considered: the relationship between individual differences in working memory and individual differences in the efficiency of syntactic processing; the effect of concurrent verbal memory load on syntactic processing; and syntactic processing in patients with poor short term memory, poor working memory, or aphasia. The experimental results suggest that the verbal working memory system specialized for assigning the syntactic structure of a sentence and for using that structure in determining sentence meaning is distinct from the working memory system that underlies the use of sentence meaning to accomplish further functions. We present a theory of the components of the verbal working memory system and suggestions as to its neural basis.

Verbal Working Memory and Sentence Comprehension

Research in psychology has provided considerable evidence for a division between long-term memory, in which memories of large numbers of facts and autobiographical events are maintained for up to years, and short-term memory, which is capable of retaining small amounts of information for very short periods of time (Squire and Zola-Morgan, 1991). Baddeley and his colleagues introduced the powerful concept that the appropriate way to characterize short-term memory is as a "working memory" system (Baddeley, 1976, 1986, 1995; Baddeley & Hitch, 1974). Working memory is conceived of as a short-duration limited-capacity memory system capable of simultaneously storing and manipulating information in the service of accomplishing a task (Baddeley, 1995). Appeal to the notion of a limited-capacity working memory system (or to equivalent concepts such as "processing resources") to account for features of human cognitive performance is widespread in cognitive psychology, with respect to both normal functions (Just and Carpenter, 1992) and the abilities of subjects with developmental and acquired cognitive disorders (Gathercole and Baddeley, 1993).

Baddeley and his colleagues proposed the first model of the functional architecture of human working memory. In his model working memory is made up of three main components -- the central executive, the articulatory loop, and the visual-spatial scratch pad (Baddeley, 1986; 1990; Baddeley and Hitch, 1994; Baddeley, Logie, pessi, Della Sala, and Spinnler, 1986; Gathercole and Baddeley, 1993). The articulatory loop and the visual-spatial scratch pad are "slave systems" in which verbal and visual information respectively are stored when the central executive is overloaded. One can conceive of these components as responsible for maintaining short-term information availability. The central executive is the workhorse and mastermind of human cognition. It allocates attention to a task and performs information storage and computational functions within a given task.

To determine the working memory requirements of a task, Baddeley's theory of the functional architecture of working memory needs to be supplemented by specific models of the computational demands that individual cognitive processes make on the central executive. Current models of cognitive processes provide such measures. Implemented models provide specific quantitative measurements of the computational and storage demands of a task. For instance, in procedural models such as SOAR (Lewis, 1996), a measure of the working memory requirements of a computation might be the number of procedures required to reach a subgoal and the number of elements maintained in each procedure. In contrast, measures of working memory requirements derived from neural net models would be quite different, and might consist of the number of steps that are needed to gravitate towards an attractor in a neural network (Tabor et al, 1997). The working memory demands exerted at a given point of processing in a task are usually taken to be the sum of the working memory requirements of the functions that are active at that point in the task (Just and Carpenter, 1992). Many valuable models of cognitive phenomena are not implemented, but still provide guides to the relative complexity of one operation, or set of operations, over another (Gibson and Thomas, 1996). Research into the role of working memory in cognitive processes regularly appeals to all these types of models to provide the basis for the determination of the processing demands of a task and of those demands at a particular point of processing within a task.

Many basic questions about working memory in human cognition remain unresolved. The issue we take up in this paper is that of possible specializations within working memory. There is considerable evidence for a division of the central executive of the working memory system into visual and verbal components (Shah and Miyake, 1996). We focus on a finer division -- the possibility that the verbal working memory system is composed of subsystems devoted to different types of verbal tasks.

Our particular interest is in the distinction between the extraction of meaning from a linguistic signal, which we call "interpretive processing," and in the use of that meaning to accomplish other tasks such as storing information in long term semantic memory, reasoning, planning actions, and other functions, which we call "post-interpretive processing." By "interpretive processing," we refer to the processes of recognizing words and appreciating their meanings and syntactic features, constructing syntactic and prosodic representations, and assigning thematic roles, focus and other aspects of propositional and discourse-level semantics (see below for more discussion). Many linguists and psycholinguists have argued that the processes involved in interpretive processing are distinct from those involved in other verbally mediated functions (e.g., Fodor, 1983; Forster, 1979; Frazier, 1990). Arguments regarding the "modularity" of interpretive processing have largely centered on the issue of what types of information are used in the initial determination of linguistic form and meaning (MacDonald, 1997). We will address the question of a cognitive specialization for interpretive processing from the point of view of the structure of working memory.

Our empirical focus is on the process of sentence comprehension. We will review the results of a variety of experiments that suggest that the working memory system that is called on in interpretive processing at the sentence level -- assigning the syntactic structure of a sentence and using that structure to determine the meaning of the sentence -- constitutes a separate subsystem within verbal working memory. After presenting this evidence, we will set the topic in a poader context, and advance the hypothesis that this subsystem of verbal working memory is involved in the set of related operations that is responsible for identifying linguistic elements and structures and determining the preferred literal meaning of an utterance.

This paper is divided into four main sections. First, we will review the evidence that processing resources are involved in sentence comprehension. We will then review evidence from normal subjects that bears on the nature of the working memory system involved in syntactic processing in sentence comprehension, followed by a presentation of similar evidence from several groups of patients with C.N.S. disease. We will end the paper with a statement of our theory, including its neurological aspects.

 1. Sentence comprehension and working memory There are several pieces of evidence that indicate that comprehending sentences requires verbal working memory resources. To begin with anecdotal evidence, consider the following sentence:

 1. The man that the woman that the child hugged kissed laughed.

In sentence (1), most readers having trouble assigning thematic roles (who did what to whom). They can assign thematic roles much more easily in the two sentences that combine to form it, shown in (2):

 2 a. The man that the woman kissed laughed.

 b. The woman that the child hugged kissed the man.

The trouble normal English users have understanding sentence (1) is thought to arise because they do not have sufficient working memory to retain the intermediate products of computation that are produced in building the complex syntactic structure of this sentence (Chomsky 1957; Johnson, 1996; Gibson, 1997).

A variety of implemented models of syntactic processing (parsing) have been developed. These models produce measurements of the computational demands of assigning syntactic structure and using it to determine aspects of sentence meaning. These models have taken many forms, ranging from deterministic parsers associated with Chomsky's (1981, 1986, 1993) model of syntactic representations (Berwick & Weinberg, 1984) through Abney's (1989) principles-and-parameters parser, Johnson's (1996) logical parser, connectionist models (Tabor et al, 1997), and various hypid models such as Just and Carpenter's CC-READER (Just and Carpenter, 1992). Non-implemented models of how sentences are parsed have also been developed (e.g., Frazier and Clifton, 1996; Gibson, 1997). Though there are considerable differences among the types of operations and complexity metrics found in these models, there is a remarkable degree of similarity in the measurements they produce of the relative complexity of the parsing process in many different types of sentences.

In part exploiting these models, experimental results have provided evidence beyond intuitions that syntactic processing in sentence comprehension requires the allocation of processing resources. Considerable research has found evidence that sentences that have more complex syntactic structures are more difficult and time consuming to understand (see papers in MacDonald, 1997). The evidence that syntactic structural complexity is associated with increased difficulty in sentence processing extends to measurements made internal to the sentence processing process. Eye fixation durations, self-paced word-by-word and phrase-by-phrase reading and lexical decision times, and self-paced listening times increase at points in a sentence where models of sentence processing predict an increased processing load (Frazier and Rayner, 1982; Ford, 1983; King and Just, 1991; Caplan, Hildepandt and Waters, 1994; Ferreira, Henderson, Anes, Weeks and MacFarlane, 1996; Ferreira, Anes, and Horine, 1996). Lexical decision reaction times to visually presented words and detection times for extraneous noises and phonemes increase at these points during the auditory presentation of a sentence (Zurif et al, 1995; Frauenfelder, Segui and Mehler, 1980; Cohen and Mehler, 1996), consistent with the view that local processing load is higher at these points. These and many more studies suggest that syntactic analyses and the use of these analyses to assign sentence meaning require processing resources.

 As noted above, we contrast the process whereby a listener or reader extracts the meaning of the sentence from the signal from the process of using the meaning that has been extracted to perform other tasks. To appreciate the difference intuitively, consider sentence (3):

 3. Please pick up four tomatoes, a pound of apricots, prune juice, shallots, six apples and a bag of carrots on the way home.

 Compared to sentence (1), it is relatively easy to understand what sentence (3) means. However, certain operations that one could perform on the meaning of the sentence would be difficult, such as carrying out the request from memory. We draw a distinction between these types of operations, which include remembering the content of a sentence, using the meaning of a sentence to plan action, reasoning on the basis of sentence meaning, and other aspects of what we call post-interpretive processing, and sentence interpretation itself.

Utilizing the propositional content of a sentence to accomplish tasks usually involves controlled, conscious processing, as opposed to the largely unconscious processes involved in sentence interpretation. Controlled, conscious processes constitute the domain of verbally mediated tasks that is thought to require processing resources (Schneider and Shiffrin, 1977; Shiffrin and Schneider, 1977). There are many different purposes to which the propostional content of a sentence can be put, and it is widely thought that, to the extent to which they involve controlled, conscious processing, they all make demands on verbal working memory (Baddeley, 1986).

In many experiments on working memory, researchers have contrasted performance on pairs of sentences that are selected to vary in their complexity. In most of these experimental manipulations, there is some effect of the change in sentence type on both interpretive and post-interpretive processing. Some of these contrasts, however, increase the working memory load primarily at the interpretive stage of sentence processing; others at the post-interpretive stage. Two contrasts in particular will recur in the presentation of data to follow: that between sentences with subject- vs. object-relatives (The boy that hugged the girl kissed the baby; The boy that the girl hugged kissed the baby) and that between sentences with one proposition and two propositions corresponding to each of two verbs (The boy hugged the girl and the baby; The boy hugged the girl and kissed the baby). Though both of these contrasts involve changes in both the syntactic structure of a sentence and aspects of its meaning, we will present evidence that indicates that the first of these contrasts primarily increases processing load at the syntactic level and the second primarily increases processing load associated with different aspects of post-interpretive processing.

 2. Sentence Comprehension and Working Memory Capacity in Normal Subjects As noted above, our focus is on the nature of the working memory system involved in the interpretive aspect of sentence processing. One possibility, advanced by Just, Carpenter and their colleagues, is that humans have a set of verbal processing resources that can be devoted to all verbal tasks (Just and Carpenter 1992; King and Just 1991; MacDonald, Just, and Carpenter 1992; Miyake, Just and Carpenter, 1994). An alternative view is that part of the verbal working memory system is specialized for interpretive aspects of sentence comprehension; specifically, assigning syntactic structure and using it to determine the meaning of a sentence (Caplan and Waters, 1995, 1996; Waters and Caplan, 1996a, b, c; Waters, Caplan and Rochon, 1995).

There have been two basic approaches to investigating the possible specialization of working memory for sentence interpretation. One has been to determine the relationship of individual differences in verbal working memory capacity to the efficiency of sentence interpretation. A second has been to investigate the pattern of mutual interference (or non-interference) of sentence interpretation and concurrently holding a verbal load in short-term memory.

With respect to the individual-differences approach, the single-resource (SR) theory predicts that having a low working memory capacity will reduce the resources available for sentence processing and make it less efficient; a separate-sentence-interpretation-resource (SSIR) theory predicts that performance on general verbal working memory tasks will not predict language processing efficiency. The single-resource model is therefore supported by significant correlations between measures of working memory capacity and measures of sentence processing efficiency. Many studies correlate working memory capacity and performance on verbally mediated tasks (see Daneman and Merikle, 1996, for review). However, almost no research into the role of working memory in sentence processing has used correlational analyses. Instead, research into this question that uses the individual-differences approach has been based on experiments in which performance on a working memory task serves to divide subjects into those with high and low (and sometimes medium) working memory capacity and sentence processing performance is measured. The single-capacity model predicts that, in the absence of ceiling and floor effects, there will be a main effect of working memory capacity in experiments of this type, with high-capacity subjects performing better on the sentence processing task than low-capacity subjects. The SSRI theory might be thought to predict that there will be no main effect of group in such analyses. However, there is another issue that complicates this prediction: low capacity subjects would not be expected to be as adept at accomplishing many psycholinguistic tasks as high-capacity subjects. For instance, King and Just (1991) had subjects do a self-paced reading task while simultaneously remembering the last word of each of a set of sentences (sets ranged from 1 - 3 sentences). Poorer performance by low-span subjects could be a reflection of difficulties they had with dividing their attention in this task. Even tasks that are relatively simply in experimental psycholinguistics, like self-paced reading coupled with answering questions at the end of each sentence (MacDonald et al, 1992) or rapid serial visual presentation (RSVP) with end-of-sentence acceptability judgments (Waters and Caplan, 1996b), might be harder for subjects with low working memory capacities. In general, poorer performance by low-capacity subjects is not necessarily due to their inability to accomplish the processing that is the focus of an experiment, but may represent difficulties these subjects have with other aspects of task demands. Hence, group effects in sentence processing may be compatible with the SSIR theory, under some circumstances.

 In most experiments sentence materials have been constructed to represent two points along a complexity scale rather than a single sentence type. This is because most researchers have assumed that both low- and high-capacity subjects have sufficient resources to assign structure and meaning in simple sentences and that the limitations imposed by low working memory capacity are felt only or mostly in complex sentences. (e.g., King and Just, 1991; Miyake et al, 1994) Though there has been little discussion of this assumption in the literature, it seems at least plausible given the number of sentences that most language users appear to process without difficulty. Given this assumption, the single-resource theory predicts that there will be an interaction between syntactic complexity and working memory capacity: sentence complexity will affect low capacity subjects more than high capacity subjects, and the differences between performance of low- and high-capacity subjects on complex sentences will be greater than the differences between the groups on simple sentences.[1] Even if the SSIR theory is compatible with a main effect of group in such experiments for the reason outlined above, it predicts that no such interactions will be found. An interaction between sentence complexity and working memory group therefore favors the SR theory.[2]

A second approach is to investigate the pattern of mutual interference (or non-interference) of verbally mediated tasks. According to the SR model, verbal memory loads that are imposed external to the comprehension task (such as a concurrent digit span task) and sentence comprehension draw on the same pool of resources. According to SSIR theory, "interpretation-external" and "interpretation-internal" processes draw on different resource pools. As with the relationship between working memory capacity and sentence processing, research using this approach has assumed that a concurrent load will affect the processing of simple sentences less than that of complex sentences (Baddeley and Hitch, 1974; Miyake et al, 1994). Making this assumption, SR theory predicts an interaction between external load and sentence complexity and SSIR theory does not. In research using this approach, the concurrent task must utilize central executive resources. Digit span has been the most widely used task for this purpose.[3]

These two approaches can be combined. In SR theory, comprehending more complex sentences and maintaining a larger load in verbal memory both require more processing resources from the same pool, and this pool is smaller in some subjects. Thus, it would be supported by the finding that the expected impairment of low-capacity subjects on more complex sentences is exacerbated by a concurrent memory load. In contrast, the SSIR theory holds that maintaining a digit load in memory calls on a resource pool that is standardly measured by working memory tests, and that comprehending sentences calls on another pool of resources. This model therefore predicts that low-capacity subjects should perform less well than high-capacity subjects under conditions of increased verbal memory load, but it does not predict that this effect should be greater for syntactically complex sentences.[4]

We review the evidence relevant to each of these lines of argument here. For published papers that contain a large number of analyses, we present summaries of the results. For work in our laboratory that has not yet been published, we present results and illustrations of representative and critical findings.

2.1 Individual Differences in Speed and Accuracy of Syntactic Processing

 Relating individual differences in verbal working memory capacity to syntactic processing abilities requires that we be able to measure each individual's working memory capacity. Daneman and Carpenter (1980) developed a "reading span" task that has become the standard method of assessing verbal working memory. In this task, subjects are required to read aloud increasingly longer sequences of sentences and to recall the final word of all the sentences in each sequence. A subject's working memory capacity is defined as the longest list length at which they are able to recall the sentence-final words on the majority of trials. There are various versions of this test (Turner and Engle, 1987; Waters, Caplan and Hildepandt, 1987; Tirre and Pena, 1992), as well as other tests with quite different sets of operational demands such as random digit generation (Petrides et al, 1993) or backwards digit span (Kemper, 1988), that have been used to measure verbal working memory. Most research into the performance of groups that differ in working memory capacity in processing sentences with different structures has used the Daneman and Carpenter task as the measure of working memory, and the results that we will report follow this practice.

 The first set of results deals with interpreting sentences with relative clauses. There is a considerable amount of evidence that assigning the structure of an object-relativized sentence such as (4) is more demanding than a subject-relativized sentence such as (5), and this increased load occurs at the verb of the embedded clause (push).

 4. The boy that the girl pushed kissed the baby

 5. The girl pushed the boy that kissed the baby

 King and Just (1991, exp 1) reported self-paced word-by-word reading times for high and low span subjects for these two sentence types. In this task, a sequence of words that forms a sentence is presented visually, one word at a time, with the subject pressing a response key to see each successive word. Reading time increases at points of syntactic complexity in this task. King and Just (1991) present a graph, reprinted in Just and Carpenter (1992), that is said to show that the biggest reading time differences between high and low span subjects were in the syntactically critical area of the object relative sentences. However, no statistical analyses were reported in King and Just (1991) to support the contention that the difference in reading times between high and low-span subjects is specifically localized to the region of object relative sentences where there is the greatest processing load. Moreover, the data presented by King and Just (1991) and Just and Carpenter (1992) did not isolate performance on sentences in which the subjects did not have to retain the sentence-final words (no analyses were reported by King and Just on the sentences in the zero load condition alone).

 Our laboratory has failed to find differences between high- and low-span subjects in on-line processing of these types of sentences. In one study, we used the "auditory moving windows" task, in which spoken sentences are recorded, digitized and segmented into constituents, and subjects press a key to hear each successive constituent (Ferreira et al, 1996). As in the study by Ferreira et al (1996), we found significant increases in listening times for the embedded verb of object-relativized sentences compared to subject-relativized sentences. However, in a group of 100 subjects, this effect was no larger in low capacity than in high capacity subjects. The results are displayed in Figure 1, which shows the differences in listening time for the same phrases in the complex object-relativized sentences and the simple subject-relativized sentences. As Figure 1 shows, listening times increase at the embedded verb and, to a lesser degree, at the main verb -- both points of syntactic processing load -- and these increases do not differ in the three groups of subjects. We have also used a continuous visual lexical decision task (Ford, 1983), in which subjects were required to judge whether each successive word of the sentence was a real word. We replicated Ford's finding of a statistically significant increase in reading time on the relative clause verb in object- as compared to subject-relative clauses. In a group of 98 undergraduates divided into high, medium, and low span subjects using the Daneman and Carpenter task, this effect was no larger in low span than in high span subjects. We have obtained similar results in a replication of this study with 63 subjects in which the sentences were presented auditorily rather than visually.

To summarize these data, we know of four studies that compare the performance of high and low capacity subjects on measures of word processing as a function of sentence type (object- vs. subject-relativization) and region within sentence -- the self-paced reading study of King and Just (1991) in which subjects had to recall sentence-final words as well as process the sentences, our self-paced auditory and visual lexical decision studies, and our self paced listening (auditory moving window) study. None of these studies showed an interaction between span group, sentence type and region. In our studies, there were no interactions of span group and sentence type in the region at which there is the greatest processing load.

A second type of sentence that has heavy processing demands are so-called "garden path" sentences. These sentences are ones that are locally ambiguous, and that eventually are resolved in favor of the less preferred interpretation. For instance, the sequence The experienced soldiers warned about the dangers is ambiguous. The phrase warned about the dangers could be a main verb with a prepositional phrase adjunct, as in (6), or a relative clause in which the relative pronoun has been omitted, as in the garden path sentence (7):

 6. The experienced soldiers warned about the dangers before the midnight raid.

 7. The experienced soldiers warned about the dangers conducted the midnight raid.

 MacDonald, Just and Carpenter (1992) claim to have found differences in high- and low-capacity subjects' self-paced reading times and accuracy in answering questions about sentences such as (7). MacDonald et al's results are very complex, and the reader is referred to Waters and Caplan (1996a) for a review and detailed analysis. In general, for garden path sentences like (7), group differences in reading times and accuracies were not statistically significant and differences in reading times were not found while subjects were reading the ambiguous portions of the sentence.

We have examined the performance of high- and low-capacity subjects on sentences like (7), as well as two other types of garden path sentences (Waters and Caplan, 1996b). We carried out three experiments in which we compared the ability of high and low span subjects to comprehend these three different types of garden path sentences both with visual presentation of whole sentences and with rapid serial visual presentation (RSVP) at increasingly faster rates. Although garden path sentences were significantly more difficult to interpret than non-ambiguous control sentences, these effects were not larger in subjects with low working memory capacity. We also replicated the MacDonald, Just and Carpenter (1992) self-paced experiment using their materials in 91 subjects and failed to find memory span group differences in reading times ( Figure 2) -- a replication of MacDonald et al's result, which, as noted above, did not find significant group differences in reading times in the ambiguous region of these sentences.

MacDonald et al (1992) also examined self-paced reading times for sentences like (6), which are locally ambiguous but are resolved in favor of the preferred interpretation and the simpler syntactic structure. They found that high-capacity subjects had longer reading times than low-capacity subjects and they accounted for this superficially paradoxical result by suggesting that high-capacity subjects maintain both interpretations of a sentence in mind until it is disambiguated whereas low-capacity subjects lack the resources to do so and drop the less preferred analysis. However, their data provide little evidence to support the view that subjects with different working memory capacities perform differently in processing these sentences. MacDonald et al. did find that reading times were longer in high-capacity than in low-capacity subjects but this was true only for the last word of sentences such as (6) and not throughout the ambiguous region. Moreover, low-capacity subjects made more errors in answering questions. This pattern may simply represent different speed-accuracy trade-offs in the decision-making process in the different capacity groups. Pearlmutter and MacDonald (1995) reported a similar result, in which high- and low-capacity subjects used different cues as to the likelihood that the sentence would end with a main clause or a reduced relative clause interpretation. All of MacDonald and her colleagues' results must be interpreted cautiously, because she had subjects answer questions about the thematic role played by the first noun after each sentence. This may have alerted the more verbally-adept high-span subjects to the presence of the ambiguity, and any differences between groups may be due to strategic factors operating in these experiments. We have repeated MacDonald et al's experiment in 91 subjects using the exact methods and materials from MacDonald et al (1992). We did not find poorer performance in high- than low-capacity subjects on sentences like (6) ( Figure 3). In a follow-up experiment, we eliminated the questions that were posed after the sentences. This experiment, with 75 subjects, also failed to find differences between high- and low-capacity subjects.

 A third line of research has investigated the interaction of semantic and syntactic information in sentence interpretation in subjects with different working memory capacities. Just and Carpenter (1992) studied garden path sentences similar to those in (7) but in which the animacy of the first noun constrains the possible interpretation of the sentence, as well as sentences that were unambiguous because of the presence of a relative pronoun. The sentence types are shown in (8) - (11):

8. Syntactic garden path; semantically unconstrained:

 The defendant examined by the lawyer shocked the jury.

9. Syntactically unambiguous; semantically unconstrained:

 The defendant that was examined by the lawyer shocked the jury.

 10. Syntactic garden path; semantically constrained:

 The evidence examined by the lawyer shocked the jury.

 11. Syntactically unambiguous; semantically constrained:

 The evidence that was examined by the lawyer shocked the jury.

 Their data show that both high- and low-capacity subjects have longer first-pass fixation durations on the by-phrase in the sentences with reduced relative clauses (sentences 8 and 10) than in the corresponding sentences with the overt relative clause marker (sentences 9 and 11), regardless of the animacy of the first noun. This provides evidence for similar degrees of modularity of syntactic processing in both high- and low-span subjects (Ferreira, 1992; Waters and Caplan, 1996a).

Just and Carpenter also found that the high-capacity subjects showed longer fixation times on the by-phrase in relative clause sentences with animate first nouns (sentences 8 and 9) than in the corresponding relative clause sentences with inanimate first nouns (sentences 10 and 11), while this was not true for the low-capacity subjects. This result is difficult to interpret. It indicates that high-capacity subjects were unable to use the clear, unambiguous cue to syntactic structure afforded by the relative pronoun and the auxiliary verb (that was) of the relative clause in sentence (9), and took the defendant as the subject of the active form of the verb examined in the phrase The defendant that was examined, because of the animacy of the word defendant. It is hard to understand why the more verbally adept high-capacity subjects would not use this clear cue, and other data show the opposite pattern. King and Just (1991, exp 2) studied the ability of high- and low-capacity subjects to comprehend object relative sentences that contained verbs that either did or did not provide strong pragmatic cues as to which of the two potential actors in the sentence was the agent of a given verb (sentences 12 - 15):

12. Main verb constrained; embedded verb constrained:

 The robber that the fireman rescued stole the jewelry.

 13. Main verb unconstrained; embedded verb constrained:

 The robber that the fireman rescued watched the program.

14. Main verb constrained; embedded verb unconstrained:

 The robber that the fireman detested stole the jewelry.

 15. Main verb unconstrained; embedded verb unconstrained:

 The robber that the fireman detested watched the program.

King and Just found that the comprehension of high-capacity subjects did not improve in the pragmatic bias condition, but that of low-capacity subjects did. Although these findings reflect accuracy on end-of-sentence judgments rather than eye fixation durations during on-line sentence processing, they are the opposite of the Just and Carpenter (1992) results.

A fourth line of research that has been pursued involves increasing the processing demands associated with sentence comprehension by changing the perceptual demands of the task. Miyake, Carpenter and Just (1994) reported a series of experiments using RSVP, in which low-, medium- and high-capacity subjects were required to indicate the actor or answer questions about sentences that required syntactic analyses to be understood. Miyake et al. reported significant interactions between span group and sentence type. In order to determine whether these interactions reflected differences in the ability of subjects with different working memory capacities to utilize syntactic structures, we reanalyzed the data from the one experiment for which sufficient data were available for reanalysis, and found that subjects with different working memory capacities did not perform differently as a function of syntactic complexity (Caplan and Waters, 1995). The group by sentence type interaction was due to poorer performance by low-span subjects on sentences with two propositions compared to sentences with one proposition. The larger effect of the number of propositions in low-capacity subjects suggests that these subjects have more difficulty than high-capacity subjects retaining information about the actor in each proposition in memory. Thus, this study provides evidence that increasing the perceptual demands of a task does not differentially affect the abilities of subjects with different working memory capacities to process syntactic structure, but may make certain post-interpretive processes more difficult for low-capacity subjects.

We conclude this section with a pief discussion of the effects of age upon sentence processing. Most studies have found that working memory declines with age (Salthouse, 1991), so the study of age-related changes in sentence comprehension provides indirect evidence about the effect of changes in working memory capacity on this function.

 The effects of aging on sentence comprehension abilities have been examined using a variety of tasks such as object manipulation (Feier and Gerstman, 1980), question answering (Emery, 1985; Davis and Ball, 1989), acceptability judgment (Kemper, 1988; Obler, Fein, Nicholas, and Albert, 1991), and sentence recall (Norman, Kemper, Kynette, Cheung and Anagnopoulous, 1991). Some studies have found that elderly subjects perform more poorly overall, but that they are not differentially impaired on syntactically more complex sentences (e.g., Feier and Gerstman, 1980). Several studies have measured sentence processing as a function of age using on-line tasks. Baum (1991) and Waldstein and Baum (1992) used a word monitoring paradigm, in which subjects were required to monitor for a word that followed an ungrammaticality in a sentence, to investigate the sensitivity of younger and older adults to local and long distance ungrammaticalities. Although processing times and error rates of the elderly were higher overall than those of the younger subjects, there was no evidence that elderly subjects were more reliant on sentential context or that they were less sensitive to ungrammaticalities. Kemtes and Kemper (1996) also found that the effect of syntactic ambiguity on word by word reading times did not differ in older and younger subjects, although the older subjects were more affected on an off-line measure of accuracy in responding to questions. We analyzed the data from the garden path experiment described above (Waters and Caplan, 1996b) for effects of age, and found no such effects.

Other studies have found decreased comprehension of more complex syntactic structures in the elderly (e.g., Obler et al., 1991; Davis and Ball, 1989). However, in terms of the distinction we made above between interpretive and post-interpretive processing, the tasks on which aging subjects have shown such effects are ones that require elaborate post-interpretive processing, such as retaining and re-ordering large amounts of material in memory (Kemper, 1986; Light, 1990) or interpreting implausible sentences (Davis and Ball, 1989; Obler et al., 1991). Thus, in many of these studies, the effect of age may be attributed to difficulties subjects may have with processes other than those involved in sentence interpretation. In the domain of on-line processing, using a cross-modal naming paradigm, Zurif, Swinney, Prather, et al (1995) found that older subjects were delayed at establishing the connection between a gap in a relative clause and the head of the clause (e.g., determining that it is the man who is kissed in a sentence such as The man who the woman kissed was embarrassed). Zurif et al's data are difficult to interpret, however, since a comparison group of younger subjects was not tested on the same materials. We have failed to find differences as a function of age in listening times for phrases in high-load positions in more complex sentences in the auditory moving window task described above. Overall, the evidence indicates that there is little, if any, effect of aging on syntactic processing in comprehension in aging, and therefore indirectly supports the view that differences in working memory capacity are not associated with differences in the efficiency of syntactic processing in sentence comprehension.

In summary, there are several studies that examine the performances of subjects with different verbal working memory capacities in syntactic processing in sentence comprehension tasks. Most of these studies have not shown differences as a function of span group, even in processing sentences with very complex syntactic structures. A few studies have reported group differences, but these differences may be due to strategic effects (e.g., different speed-accuracy trade-offs in different capacity groups) and are inconsistent across a wider set of studies. In all, this literature provides evidence that the speed and accuracy of syntactic processing does not differ in a systematic way in normal subjects with different working memory capacities as measured by performance on sentence span tasks. In contrast, there is a suggestion in the data that low-capacity subjects may be more affected than high-capacity subjects by some sentential features, in particular, the number of propositions in a sentence. Given that subjects with different working memory capacities do not differ in their ability to use syntactic structure to determine sentence meaning, the larger effect of the number of propositions in low- than in high-capacity subjects suggests that low-capacity subjects have more difficulty with retaining information about the propositional content of a sentence in memory. We shall present data pertaining to other post-interpretive functions below.

2.2 Effects of an External Memory Load on Speed and Accuracy of Syntactic Processing

 The second approach to the issue of the possible specialization of working memory for syntactic processing is to investigate the pattern of mutual interference of verbally mediated tasks. According to the SR model, verbal memory loads that are imposed external to the comprehension task (such as a concurrent digit span task) and sentence interpretation draw on the same pool of resources, and therefore predicts that there will be interference between the two. In SSIR theory, "interpretation-external" and "interpretation-internal" factors draw on different resource pools, hence the theory predicts main effects of each but no interference between the two. We have undertaken several studies; all found the second pattern.

 In the first study (Waters, Caplan and Rochon, 1995), we had subjects match sentences to pictures in a no interference condition and in two concurrent verbal load conditions: while retaining a random sequence of digits equal to their span and equal to one less than their span. Sentences were all semantically reversible and varied in their syntactic complexity. The effect of syntactic complexity was examined by comparing performance on pairs of sentences that were matched for number of words, propositions, verbs, and thematic roles but that differed with respect to whether a sentence contained a noun phrase that had been moved, according to Chomsky's (1986, 1993) theory, creating a non-standard order of thematic roles. For instance, object-relativized sentences (such as 4: The boy that the girl pushed kissed the baby) were syntactically complex compared to subject-relativized sentences (such as 5: The girl pushed the boy that kissed the baby) according to this measure. There was an effect of load on the sentence accuracy scores, but no effect of syntactic complexity and no interaction of syntactic complexity with load in these scores, or in the digit recall data. We re-examined this question by changing the task to enactment, which may be more demanding (Waters and Caplan, under review). Subjects had to enact the thematic roles in sentences with the same structures used in the picture matching task, in a no interference and a digit load condition. Again, there was an effect of condition, but no interaction of load and syntactic complexity.

These studies used off-line measures of accuracy to assess sentence comprehension, which may not be sensitive to the effects of a concurrent load. We have carried out one dual-task study in which subjects' performances were timed as well as scored for accuracy (Waters, Caplan and Hildepandt, 1987). In this study, subjects made judgments about the plausibility of sentences that differed in their syntactic complexity (object- vs. subject-relativization) in no interference and digit load conditions (recalling a random sequence of six digits). There was an effect of syntactic type and of load; the interaction of load and syntactic complexity was significant in the sentence data (RTs and accuracy) in the analyses by subjects but not by items.

Although the above results suggest that digit span does not interfere with syntactic processing in sentence comprehension, other studies indicate that digit span does interfere with several verbally mediated tasks (Baddeley and Hitch, 1974), and several researchers have interpreted their results as showing that digit span affects subjects' abilities to process syntactically complex sentences. However, a review of this literature indicates that the effect of a digit load on syntactic processing only occurs under some task conditions. We summarize this literature here; more complete discussions can be found in Waters et al (1995) and Caplan and Waters (1996).

 Baddeley and Hitch (1974) found that a concurrent random six-digit load interfered more with subjects' performance on passive than on active sentences in a "reasoning task" in which subjects had to indicate whether a statement about a sequence of letters corresponded to the actual presentation of letters (e.g., A is not followed by B -- BA). However, many factors (order of mention of the letters and the letter pair; presentation of letters in their alphabetical order in either the proposition or the pair; proactive interference; etc.) that may have affected the results were not considered in the analyses, and the interaction of load and voice may have been qualified by higher-order interactions that were not considered in this early work on this topic. Several investigators (Wright, 1981; Gick, Craik, & Morris, 1988; Morris, Gick, & Craik, 1988; Morris, Craik & Gick, 1990) have studied younger and older subjects on tasks that combine digit load with a sentence plausibility (or truth) judgment task. Many of these studies have found significant 2-way interactions between the size of a concurrent memory load and sentence complexity. The sentence complexity metric in these experiments, however, was the presence of a negative element in the sentence (e.g., A cat does not hunt mice), which significantly affects the complexity of the judgment task but has little effect on the complexity of the syntactic structure of the sentence. These studies have not shown interactions of syntactic structure with load.

 Other experiments have investigated the effect of a concurrent digit load that has been presented as the requirement to recall the final word of each of a series of sentences. King and Just (1991) had subjects read object- and subject-relative sentences word by word in a self-paced task while retaining one, two, or three sentence-final words in memory, and answer questions about the sentences. There was an interaction between syntactic complexity and memory load in the recall data. There was no such interaction in the accuracy data, and the reading time data were not analyzed for the effect of size of memory load. A second experiment using the Daneman and Carpenter paradigm was reported by Carpenter and Just (1988). In this experiment, sentences were considered more difficult if they contained less frequent and less concrete words, and thus the effect of a memory load on syntactic processing was not investigated. Results using a variant of the Daneman and Carpenter paradigm were reported by our group (Waters et al., 1987; Waters and Caplan, 1996b). We found an effect of syntactic complexity (object vs. subject relativization) on the number of sentence-final words recalled in a sentence acceptability judgment task. Finally, Wanner and Maratsos (1978) interrupted sentence presentation to allow a set of words to be displayed, and reported poorer recall of words in object-relativized than in the less complex subject-relativized sentences.

 This pattern of results suggests that whether or not a concurrent digit load interferes with syntactic processing depends upon the relationship of the recall and sentence processing tasks. The published literature is consistent with the generalization that when the stimuli in both the sentence and recall tasks are presented in an uninterrupted fashion -- as when the digit load is presented prior to each sentence and must be recalled once the sentence has been understood -- no interference is found, but when the presentation of the stimuli in one task is interrupted by the presentation of a stimulus relevant to the second -- as when the sentence is interrupted by a series of words while it is being presented, or the presentation of the to-be-recalled items is staggered across the sentence task (as in sentence-final word recall) -- a digit load of sufficient size interferes with processing syntactically complex sentences more than with processing syntactically simple sentences. This pattern of results suggests that the reason for the effect of a digit load on syntactically complex sentences is not that retaining a sequence of digits and structuring a sentence syntactically compete for the same resource pool, but rather that the attentional shifts associated with interrupting each task interfere with subjects' abilities to structure sentences syntactically or to use that structure to assign sentence meaning. We have suggested that this may be a secondary effect of disruptions of lexical access (Caplan and Waters, 1996). An anonymous BBS referee pointed out that interruption of a task makes it harder to retrieve information associated with that task, which also may affect the ability to construct more complex structures.

 An important point to note is that the tasks in which an interaction between digit load and syntactic complexity has not been found are not simply so easy that this interaction cannot be detected. Testing for digit recall at each subject's predetermined span guarantees that each subject will be below ceiling on the recall task. In fact, in the research cited above, the number of trials in which all digits were produced in correct serial position in the recall task was usually around 50%. The subjects who have been tested in the literature cited above also performed below ceiling on at least some sentences types in which the effects of syntactic complexity could be examined. Nonetheless, a concurrent digit load did not lead to an effect of syntactic complexity except when the presentation of the sentence or the digit sequence is interrupted by the other stimulus.

In contrast to the failure of a concurrent digit load to affect processing of syntactically more complex sentences (except when either stimulus is interrupted), the same external verbal memory load has significant effects on sentences with two propositions compared to sentences with one. Waters et al (1987) found that a concurrent load of six random digits interfered more with acceptability judgments about sentences with two propositions than sentences with one. Waters et al (1995) reported that this was also the case in a sentence-picture matching task when the concurrent digit load was equal to each subject's span, and Waters and Caplan (under review) have found similar results using an enactment task. This suggests that, unlike syntactic processing in sentence comprehension, operations on the propositional content of a sentence such as matching it to knowledge in semantic memory or depictions of events or using it to plan and execute actions share resources with span tasks.

2.3 Effect of the Combination of External Memory Load and Individual Differences in Working Memory Capacity on Speed and Accuracy of Syntactic Processing

 A third approach to the question of the specialization of verbal processing resources is the combination of the first two approaches -- looking for the combined effects of individual differences in working memory capacity and of external load on language processing efficiency. The single-resource theory asserts that comprehending more complex sentences and maintaining a larger digit load in memory both require more processing resources from the same pool, and that this pool is smaller in some subjects. Thus it predicts that the expected impairment of low-capacity subjects on more complex sentences will be exacerbated by a concurrent memory load. In contrast, the separate-sentence-interpretation-resource theory holds that maintaining a digit load in memory calls on a resource pool that is standardly measured by tests such as the Daneman and Carpenter reading span task, and that comprehending sentences calls on another pool of resources. Therefore this model predicts that low-capacity subjects should perform less well than high-capacity subjects under conditions of increased verbal memory load, but it does not predict that this effect should be greater for syntactically complex sentences.

Relevant data come from several studies. King and Just (1991) had high-, medium-, and low-capacity subjects read object- and subject-relative sentences word-by-word in a self-paced task, while retaining one, two, or three sentence-final words in memory. In the recall task, King and Just reported significant interactions between group and size of memory load and between sentence type and size of memory load, but no interaction between group, syntactic complexity, and memory load. In the probe question comprehension results, there was a statistically significant main effect of group and the effect of sentence type was marginally significant, but none of the interactions involving memory load approached statistical significance. The analyses reported in King and Just (1991) thus do not show the interactions between group, sentence type, and load that the SR theory predicts.

Carpenter and Just (1988) investigated the number of words that high-, medium-, and low-capacity subjects recalled in a Daneman and Carpenter-type reading span task when the sentences were either "easy" or "difficult." All subjects performed the task at set sizes 2, 3, and 4 under conditions in which the stimulus materials contained 0, 1, or all difficult sentences. High capacity subjects were only influenced by difficult sentences at set size 4, while medium capacity subjects were affected by difficult sentences at set sizes 3 and 4. However, the authors did not report on the presence or absence of a three-way interaction between span group, set size, and sentence type, so it is unclear whether high- and medium-capacity subjects were differentially affected at either set size 3 or 4 by sentence complexity. Contrary to the predictions of the capacity theory, the performance of low capacity subjects was unaffected by sentence difficulty at any set size. In addition, in this experiment, sentences were considered more difficult if they contained less frequent and less concrete words, so the experiment does not bear directly on syntactic processing.

We investigated the possibility that syntactic processing would be more affected by a concurrent verbal memory load in low-capacity than in high-capacity subjects in enactment (Waters and Caplan, under review), sentence picture matching, and speeded sentence acceptability judgment tasks. Figure 4 shows the data from the enactment task. The effect of syntactic complexity was greater in the low- than in the medium-capacity group, but did not differ in the low- and high-capacity groups. This effect did not increase more in the low- than in the medium- or high-capacity subjects in the dual-task conditions. Similar results showing no increase in the magnitude of a syntactic complexity effect in low capacity subjects compared to high-capacity subjects as a function of the presence of a digit load were found in the other tasks.

 2.4 Some considerations regarding studies of normal subjects

The results of the studies reviewed above are consistent with SSIR model. However, they are subject to several caveats, which we will discuss here.

 The first is that many of the arguments made in the preceding review are based upon null results -- the failure to find differences between different capacity groups in syntactic processing, to find effects of a concurrent verbal memory load on syntactic processing, or to find differential effects of load on syntactic processing in low- vs. high-capacity subjects. Issues of power and related concerns arise in interpreting non significant results. Concerns about the interpretation of null results can only be completely laid to rest by running a much larger number of subjects than is feasible (often hundreds or more: Krueger, 1994). However, they can be partially countered by noting there were interactions between sentence type and condition and sentence type and group in many of the experiments we have reported, but that the sentence type variable that entered into these interactions was not one that reflected syntactic complexity but rather the number of propositions in a sentence. This suggests that the designs had sufficient power to generate interactions involving the sentence type factor but that varying the syntactic form of a sentence does not differentially affect low-capacity subjects or performance under dual task conditions.

Other considerations pertain to the dual task experiments. One issue is that most work has used the digit span task as a concurrent verbal memory load designed to compete for resources with sentence comprehension. The span task has the advantages that load can be equated across subjects by testing each subject at his or her span, and that span represents a patient's ceiling and thus calls on all resources that can be devoted to immediate serial recall. Nonetheless, span may be largely based on a specialized auditory-to-articulatory information transfer mechanism (Freidrich, 1990), and additional research using concurrent tasks such as random number generation, which Baddeley (1993) has argued is less highly automatized than span and thus requires more of the limited resources of working memory, is needed to investigate the effect of a concurrent memory load on comprehension of syntactically more complex sentences. Preliminary results in our laboratory indicate that this task also does not interact with syntactic complexity. A second issue that arises regarding our interpretation of the dual task experiments is that we have argued that the effects of a concurrent memory span task on syntactic processing are limited to situations in which the presentation of the stimuli relevant to one task is interrupted by the presentation of the stimuli relevant to the second. This suggests that the interference is secondary to switching attention, retrieving information when a task is interrupted, or other control processes, rather than shared resources. These possible explanations of the data need to be tested directly.

 Finally, our review of the literature on individual differences in working memory capacity and concurrent verbal memory loads on syntactic processing in sentence comprehension reveals that many findings with normal subjects have been hard to replicate. One possible reason for this difficulty may be related to the measurement of subjects' working memory capacity. The most widely used measures of verbal working memory are sentence span tasks, based on the Daneman and Carpenter (1980) reading span. These tasks have several characteristics that affect how they may be interpreted. One is that, while these tasks are designed to require both processing and storage of verbal material, only the latter is usually measured. When measurements are made of performance on the sentence processing component of a sentence span task, the measurement is usually only of accuracy, not reaction time, and this measurement does not enter into the determination of which span group a subject belongs to. Because of this, subjects' spans may in part reflect any number of trade-offs between allocating attention and processing resources to the sentence and recall components of the sentence span task. A second consideration is that the verbal memory load in a sentence span task is unrelated to the sentence processing that is required in these tasks. This introduces a dual-task element into these tasks and allows them to be heavily affected by the capacity for dividing attention. Other working memory tasks (self-ordered number generation, in which subjects have to produce a random series of the digits 0 - 9 without repeating a digit, and externally ordered number generation, in which subjects have to recognize which digit is omitted in a random series of the digits 0 - 9: Petrides et al, 1993) require the memory and processing components of the task to be related to each other, as they are in naturally occurring tasks that require working memory. For all these reasons, performance on a sentence span task may not be a reliable assessment of working memory capacity.

We have investigated the relationship between different working memory measures and the reliability of several of these measures (Waters and Caplan, 1996c). We tested 94 subjects on the Daneman and Carpenter (1980) reading span test, a variant of that test we developed in which RTs and accuracies on a sentence acceptability judgment task as well as sentence-final word recall were measured (Waters et al, 1987), and self-ordered and externally-ordered number and design generation tasks (Petrides et al, 1993). There were significant correlations between all of these measures of working memory capacity (r between .52 and .58), other than the externally-ordered design generation task. This level of correlation is consistent with other reports in the literature (Engle, 1995; Daneman and Merikle, 1996). However, test-retest reliability was not high. Forty-four subjects were retested on a subset of these measures at an average of a 1 month interval. Sentence-final word recall scores at the two testing sessions were weakly correlated (r =.41 for the Daneman and Carpenter task). Of the 44 subjects who participated in the follow-up study, 18 subjects (41%) changed in terms of their classification as high-, medium- or low-capacity subjects at time 2, with equal numbers of subjects improving and declining and 22% of the subjects originally classified as high capacity being reclassified as low capacity on retesting. These figures call into question the stability of the most commonly used measures of working memory capacity. Difficulties in replicating effects of working memory on syntactic processing may thus reflect the insecurity of subject classification.

 One way to deal with this problem is to use measures of working memory capacity that are more stable over time. We found that test-retest reliability improved when we used a composite measure of working memory capacity that took into account performance on the sentence processing as well as the recall portion of a sentence span test (r = .88; Waters and Caplan, 1996c). However, little research has been done in which this type of measure is used as the basis for subject grouping. Another approach is to study patients with C.N.S. disease as members of a low capacity group, since their performances are characteristically stable. There are a few studies that use this approach, to which we turn.

 3. Studies of Patients with Reduced Working Memory Capacity

 3.1 Patients with Short-Term Memory Disorders

 One group of patients whose sentence comprehension has been studied fairly extensively are patients with intact long term memory but specific auditory verbal short-term memory impairments. Most of these patients have had problems with rehearsal and/or passive storage functions (see Vallar and Shallice, 1990, for representative cases). If working memory capacity plays a central role in language comprehension, then one might expect these patients' extreme limitations in the mechanisms that support central executive functions to have important effects on their language comprehension abilities. The literature clearly establishes that this is not the case. In a review of all short-term memory patients published up to 1990, we found no evidence linking short-term memory impairment to difficulties with syntactic processing in sentence comprehension (Caplan and Waters, 1990). The literature since that time bears out this pattern: although both storage and rehearsal impairments can affect certain aspects of comprehension (Martin, 1990), these components of short-term memory have not been related to syntactically-based sentence comprehension difficulties. Results that are typical of the dissociation between impaired STM and intact comprehension are found in a patient, B.O., whom we tested quite extensively (Waters, Caplan and Hildepandt, 1991). This patient had a memory span of only 2 to 3 items when tested both on recall and probe recognition tasks. When tested on the Daneman and Carpenter task she had a working memory span of only one, despite the fact that she had no difficulty understanding the sentences used in this task when they were presented in isolation. B.O.'s performance on visual and auditory enactment tasks and a speeded visual whole sentence acceptability judgment task that tested comprehension of many complex syntactic forms was similar to that of normals and much better than aphasic patients who have had a left hemisphere stroke. We have also tested B.O.'s comprehension of garden path sentences under timed auditory and written whole sentence presentation conditions. She performed as well as normals on these syntactically complex materials ( Figure 5). Parenthetically, B.O. showed a reliably larger effect of the number of propositions on comprehension than controls, in keeping with the idea that using the propositional content a sentence to accomplish a task such as making a judgment about the plausibility of the sentence makes use of a short-term memory system that is not required for determining the sentence's syntactic form and propositional content.

 However, as noted, the majority of patients with STM impairments have impairments in rehearsal and/or storage functions. The working memory system used in language comprehension has been claimed to correspond more closely to the central executive of a Baddeley-type model. The comprehension abilities of patients with limitations in the central executive component of this system thus constitute more direct and critical tests of the relationship between general-purpose working memory and comprehension.

3.2 Patients with Limitations in the Central Executive

Patients with Dementia of the Alzheimer Type (DAT) have been found to have intact functioning of the articulatory rehearsal and phonological storage components of working memory but impairments on tasks that require central executive functions (Baddeley et al., 1991; Morris, 1984; 1986, 1987; Morris and Baddeley, 1988; Waters, Rochon, and Caplan, in press). We have found that, when tested on the Daneman and Carpenter task with simple sentence structures that they have no difficulty understanding in isolation, all the DAT patients we have tested have had Daneman and Carpenter spans of one or less. Thus, the working memories of these patients are more reduced than those of normal subjects in whom the relationship between working memory and syntactic processing has been tested. Moreover, these patients' performances on working memory tests have been stable or have deteriorated over time (Baddeley et al, 1991). If the single-resource model is correct, then DAT patients should have particular difficulty with syntactically more complex sentences, and they should perform differentially poorly on the more difficult syntactic structures when a concurrent memory load is imposed. Our work has not found these results in this population.

In our first study, we tested 22 DAT patients and age- and education-matched controls for their ability to comprehend nine different syntactic structures using a sentence-picture matching test (Rochon, Waters and Caplan, 1994). The stimulus materials were similar to those used in the sentence-picture matching and object manipulation tasks in the dual-task experiments described above, and were designed to contrast sentences that were matched for length and other relevant variables but that differed in syntactic complexity, and sentences that were matched for syntactic complexity but differed in terms of number of propositions. The results showed an effect of group, with DAT patients performing more poorly than controls, and a group by sentence type interaction. Analysis of this interaction showed that DAT patients did not perform more poorly on the syntactically more complex sentences, but rather that their performance was poorer than controls on sentences with two propositions. This pattern of an absence of an effect of syntactic complexity for DAT patients on sentence picture matching has been replicated in several studies (Waters, et al., 1995; Waters, Rochon and Caplan, in press; Rochon and Saffran, 1995). We have also found the effect in speeded acceptability judgment tasks. Figure 6 shows the results of one experiment from our lab comparing performance of DAT patients and matched controls on two proposition simple and complex sentences in an acceptability judgment task. In this task, both patients and controls showed an effect of syntactic complexity on reaction times. However, the effect was not greater in the patients than in the control subjects.

Patients with Parkinson's Disease (PD) also have impairments in executive functions (pown and Marsden, 1988; 1991; Lees and Smith 1983; Taylor et al., 1987). We tested 17 non-demented PD patients and age- and education-matched elderly controls on a battery of tasks that tapped components of the short-term memory system (articulatory rehearsal and phonological storage), verbal working memory capacity (reading span, self-ordered and externally-ordered number generation), and various aspects of executive functions (Stroop color-word interference, Wisconsin card sorting, verbal fluency). Subjects were then tested for the comprehension of sentences differing in syntactic complexity and number of propositions, using a sentence-picture matching task. Verbal working memory spans were significantly reduced in PD patients compared to controls, but rehearsal and storage functions were normal in span tasks in these patients. The performance of PD patients differed from that of controls on the sentence comprehension task, but only on sentences that contained more propositions ( Figure 7). Comparisons of sentences that differed in syntactic complexity but that held other factors constant, such as subject- vs. object-relativized sentences, were not significant. These data suggest that, like DAT patients, PD patients do not have impairments in structuring sentences syntactically.

 We have also found that the ability of DAT and PD patients to structure sentences syntactically is not affected any more than that of normals by a concurrent memory load. In one study, DAT patients performed the sentence picture matching task outlined above in a no interference condition and while retaining a concurrent memory load that was equivalent to their span or one less (Waters et al, 1995). Overall performance was poorer with the digit load, but comparisons of length-matched sentences showed that the concurrent memory load exacerbated the effect of number of propositions, but not the effect of syntactic complexity, in the patient group. We have repeated this study with the PD group described above. As with the DAT patients, the concurrent memory load did not exacerbate the effect of syntactic complexity in the PD group ( Figure 8).

 These studies used untimed accuracy measures of sentence comprehension. Though they are not subject to the criticism that effects of syntactic structure were missed because of ceiling effects (because some comparisons of sentences that differed in syntactic complexity were made on pairs of sentences on which performance was below ceiling), measures of on-line processing may be more sensitive than an accuracy result. Using a cross-modal naming task to study on-line processing in DAT patients, MacDonald and her colleagues found that DAT patients were as sensitive as age matched controls to grammatical violations and as able as age matched controls to use frequency information and semantic and syntactic contexts to resolve syntactic ambiguities (Lalami, Marblestone, Schuster, Andersen, Kempler, Tyler, & MacDonald, 1996; Stevens, Kempler, Andersen, & MacDonald, 1996). The entire pattern of performances suggests that DAT patients have relatively preserved abilities to structure sentences syntactically in sentence interpretation. Preliminary results in our laboratory indicate that same appears to be true of PD patients: a group of 12 PD patients showed the same pattern of performance as normal controls on the auditory moving windows task.

 We have combined the use of timed measures with dual task conditions in a study of the ability of patients with DAT to make speeded acceptability judgments to auditorily presented sentences that varied in their syntactic complexity under no interference and two concurrent tracking conditions (Waters and Caplan, 1997). In a dot tracking condition, subjects were required to press the switch on a mouse that corresponded to the location of a dot that appeared on the computer screen (left, middle, or right). In a digit tracking condition, a continuous random sequence of the digits "1", "2", and "3" was presented in the center of the computer screen and subjects indicated which digit had been presented by pushing one of three switches on the mouse. DAT and control subjects' error rates were equated on each of the secondary tasks by varying the duration of the dot and digit stimuli and the interval between these stimuli. Patients showed an effect of syntactic complexity in the RT and sentence accuracy measures, which did not increase under the tracking conditions. Patients also showed an effect of condition, which was not greater for the more syntactically complex sentences. Both patients and controls showed an effect of syntactic structure in the secondary task performance measures, which was not larger for the patients. Hence this experiment found that a concurrent task failed to affect processing of syntactically complex sentences more than syntactically simple sentences under very demanding concurrent task conditions in low-capacity subjects, even when sensitive timed measures of syntactic processing efficiency were employed.

 Our experiments with DAT and PD thus indicate that patients with severely reduced working memory capacity retain extremely good syntactic processing capacities in sentence comprehension. We note, however, that some authors have asserted that sentence comprehension is impaired in patients with DAT (Tomoeda, Bayles, Boone, Kaszniak, and Slauson, 1990; Kontiola, Laaksonen, Sulkava, Erkinjuntti, 1990; Emery, 1985) and PD (Grossman et al., 1991, 1992; Lieberman et al., 1990; Natsopoulos et al., 1991). Examination of these studies suggests that DAT patients may have performed poorly because of impairments they have with aspects of post-interpretive processing, such as deficiencies in their ability to access semantic knowledge, to enact responses, and to accomplish other post-interpretive requirements of many of these tasks (see Rochon, Waters, and Caplan, 1994, for discussion). For instance, real world knowledge is necessary to understand and to respond to questions that are posed in some comprehension tasks (Emery, 1985, 1986) and DAT patients may have done poorly in these tasks because of impairments affecting semantic memory even if they understood the stimulus sentences. Enactment is a response requirement of many of the tasks used (e.g., the Token Test: DeRenzi & Vignolo, 1962), and it requires visuospatial, perceptual and practice abilities that are often abnormal in DAT (Christensen,1974). Memory abilities are also required in many of the tasks used in existing studies (Hart, 1988). The studies that show little or no sentence comprehension impairments in DAT have tended to use tasks with simpler demands, such as sentence-picture matching (Rochon et al, 1994; Schwartz, Marin and Saffran 1979; Sherman et al., 1988; Smith, 1989; although see Grober and Bang, 1995). The same observations hold regarding studies in which sentence comprehension impairments have been found in PD patients (Grossman et al., 1991, 1992; Lieberman et al., 1990; Natsopoulos et al., 1991). For instance, Natsopoulos et al. (1991) had PD patients match a sentence to one of six pictures. Poor performance on these tasks could therefore reflect problems in handling these task demands.

We have suggested at several points in our presentation that the effect of the number of propositions in sentence-picture matching, enactment, and acceptability judgment tasks arises at the post-interpretive stage of sentence processing. Our finding of an increased effect of the number of propositions in DAT is consistent with the view that we advanced above that these patients' poor performance on some sentence comprehension tasks is due to their having impairments with such processes. Because of the importance of the claim that the effect of number of propositions reflects post-interpretive processing, we have examined the effect of manipulating the non-linguistic visual and memory demands of the sentence picture matching task on the effect of syntactic complexity and number of propositions in DAT patients and controls (Waters et al., in press).

In one experiment, subjects matched a spoken sentence to one of two pictures that appeared either before or immediately following the presentation of the sentence. The second experiment used a video-verification task in which subjects were required to determine whether a spoken sentence matched a videotaped depiction of the action in the sentence or a syntactic foil. In this task, the spoken sentence sometimes ended before the action was completed, thereby requiring the propositional content of the sentence to be maintained in memory while the action in the video unfolded. In the third experiment, in different conditions, subjects were required to determine whether a spoken sentence matched a single picture or to chose the picture that matched the sentence from an array of two or three pictures. In all tasks, DAT patients were affected by the number of propositions in the presented sentence, but not by the syntactic complexity of the sentence. Comparison across the one-, two, and three-picture versions of the task showed that the magnitude of the effect of number of propositions increased for both the DAT and control subjects as the number of pictures in the array increased ( Figure 9). In addition, analysis of the data from each of the tasks separately showed that the effect of number of propositions did not occur when the foil depicted an incorrect lexical item but only when subjects were attempting to match the target to a foil that required a syntactic analysis by depicting reversed thematic roles. These results support the view that the effect of number of propositions arises at a post-interpretive stage of processing at which the thematic roles in the sentence are held in memory and matched against an analysis of a picture.

 Finally, in one additional study we have found evidence suggesting that the effect of number of propositions seen on the sentence-picture matching task is related to the functioning of a general-purpose verbal working memory system. We examined the relationship between the magnitude of the proposition effect and DAT patients' performance on tasks that assessed primary memory (digits forward, Corsi block, auditory and visual word span), articulatory rehearsal/phonological storage processes in STM (auditory and visual phonological similarity and word length), and measures of working memory function (digits backward, working memory span, dual-task performance). The magnitude of the effect of number of propositions was unrelated to all of the measures of primary memory and phonological storage/articulatory rehearsal (except visual word span) but was significantly correlated (r=-.62-.70) with all three working memory measures.

To summarize the results of our studies of the number-of-propositions effect, we have found that: (1) "intrinsic" limitations in WM capacity in normal subjects (in whom the measurement of WM capacity may be inaccurate) and in patients with DAT and PD are associated with an increase in the number-of-propositions effect in several tasks; (2) an "extrinsic" verbal memory load imposed by a concurrent task often increases the number-of-propositions effect; and (3) the magnitude of the number-of-propositions effect correlates negatively with measures of working memory in DAT patients. This set of findings provides support for the view that the number-of-propositions effect arises at a stage of processing that shares resources with working memory tests such as the reading span task. In addition, increases in non-linguistic task demands such as the presence of a larger number of pictures in sentence-picture matching increases the number-of-propositions effect. This suggests that these resources are deployed at the post-interpretive stage of sentence processing.

 We conclude this section with a pief note on sentence production in patients with DAT. Although not directly concerned with sentence comprehension, the results of several studies of this function are suggestively similar to those we have reviewed in the domain of comprehension.

Fluent syntactically well-formed speech typically characterizes the conversational output of patients with DAT and has been found in DAT patients in picture description (Hier, Hagenlocker and Shindler, 1985; Kemper, Anagnopoulos, Lyons and Heberlein, 1994) and sentence construction tasks (Schwartz, et al, 1979), suggesting to many researchers that syntactic abilities are preserved in these patients, at least in the early stages of the illness (Irigaray, 1973; Kempler, Curtiss, Jackson, 1987; Hier, Hagenlocker and Shindler, 1985; Bayles, 1982; Schwartz, Marin and Saffran, 1979; Kirchner, 1982; Illes, 1989; Blanken, Dittmann, Haas, and Wallesch, 1987). DAT patients have also been found to correct errors of syntax and phonology, but not semantic errors, in anomalous sentences (Bayles, 1982), and to make better use of syntactic than semantic cues in disambiguating spoken homophones while writing them to dictation (Schwartz et al, 1979; Kempler et al., 1987).

While DAT patients have been shown to produce both oral and written language that is syntactically well structured, the language production of these patients does differ from that of age-matched controls in other important respects. In one study, over 350 DAT patients were asked to write a single sentence and these productions were scored for their length in words and clauses, as well as for the number of propositions produced (Kemper et al., 1993). The sentences were also scored for six categories of grammatic constituents including pronouns, main verbs, secondary verbs, negatives, conjunctions, and questions. Results showed that 89% of the variance in the clinical rating of dementia severity could be accounted for by sentence length in clauses and propositional content. This suggests that the ability to produce propositions in written form is reduced as the severity of dementia increases. Furthermore, estimates of these linguistic abilities early in life appear to be powerful predictors of cognitive function and Alzheimer's disease in later life. In a longitudinal study, Snowdon et al. (1996) analyzed autobiographies written by 93 nuns when they entered a convent and evaluated the cognitive performances of these women when they were between 75 and 95 years of age (a span of some 58 years). Alzheimer's disease was assessed neuropathologically in a sample of 25 participants who died. A stronger and more consistent association was found between cognitive function later in life and the density of propositions in these early narratives than between cognitive functions and the grammatical complexity of the narratives. Strikingly, low idea density was present in the autobiographies of 90% of women with neuropathologically proven Alzheimer's disease but only in 13% of those without Alzheimer's disease. Further studies have shown that measures of idea density and grammatical complexity are highly stable over the life span (r=.62-.74) (Kemper, Snowdon and Greiner, 1996). The authors hypothesize that low linguistic ability early in life may be an early expression of Alzheimer's disease neuropathology. From our point of view, these results suggest a finer distinction. The greater relationship between measures of propositional density and DAT than between measures of grammatical well-formedness and DAT is consistent with the view that these patients are impaired in formulating concepts but are able to use the forms of language to convey the concepts that they do activate. This is a division of function related to language production that corresponds to that we have suggested between post-interpretive and interpretive aspects of the comprehension process.


3.3 Studies of patients with reduced resources for syntactic processing

 We may also approach the question of the relationship between working memory capacity and syntactic processing in sentence comprehension from another angle, by examining the effect of a concurrent memory load on syntactic comprehension in patients in whom there is evidence for a reduction in the resources available for syntactic processing in sentence comprehension.

Patients with aphasia provide such cases. Research into the nature of the sentence comprehension impairments seen in patients who are aphasic subsequent to a left hemisphere stroke has shown that many such patients have disturbances affecting their ability to use syntactic form to determine the meaning of a sentence (Caplan, Baker and Dehaut, 1985; Caplan, Hildepandt and Makris, 1996; see Berndt et al, 1996, for review). Several aspects of the performance of these patients suggests that one reason for this impairment is a reduction in the processing resources that a patient can apply to this task. One piece of evidence that favors this view is that groups of patients have been shown to have difficulty understanding sentences with more complex syntactic structures (Caplan, Baker and Dehaut, 1985). Second, factor analyses have shown that a single factor on which all sentence types load accounts for about two-thirds of the variance in many syntactic comprehension tasks in aphasic groups (Caplan et al, 1985, 1996, 1997). Third, cluster analysis shows that patients tend to be grouped according to their overall level of performance in these tasks, with performance in more impaired clusters showing greater effects of syntactic complexity (Caplan et al, 1985, 1997). Finally, some patients have been able to interpret sentences when either of two syntactic features was present, but not when both were found in a sentence (Hildepandt, Caplan and Evans, 1987). These patterns of performance are consistent with the hypothesis that the problem in syntactic processing in sentence comprehension that is seen in many aphasic patients results in part from reductions in their ability to allocate processing resources to the syntactic comprehension task.

 According to SR theory, any reduction in resources available for syntactic processing comes from a pool that is shared with other verbal tasks and therefore that syntactic complexity effects will be increased in aphasic patients under a concurrent verbal memory load condition. In SSIR theory these pools are separate, and there is the strong prediction that aphasic patients will not show an increase in syntactic complexity effects under a concurrent load condition.

We examined the effect of a concurrent digit load on the sentence comprehension performance of aphasic patients (Caplan and Waters, 1996). Over 200 aphasic patients were screened to ensure that we only tested patients who showed effects of syntactic complexity on a sentence-picture matching task, whose performance was below ceiling and above chance on that task, and whose abilities to repeat single words permitted them to be tested on a digit span task. We selected ten patients who met these criteria, and tested them on a sentence picture matching task in no-interference and concurrent load conditions (span and span-1). Although aphasic patients showed large effects of syntactic complexity when tested on the sentence picture matching test without a concurrent load, these effects were not exacerbated by the addition of a memory load. Their performance on the concurrent memory task was poorer with larger digit loads, but the effect of syntactic complexity was not exacerbated. In the digit recall data, there was an effect of number of propositions but not of syntactic complexity. These results provide striking evidence for the separation of the resources used in syntactic processing in sentence comprehension and those required for span tasks.

 3.4 Summary of Studies with Patient Populations

 The studies reviewed in this section provide important evidence against the single processing resource model. Patients with several etiologies of C.N.S. disease who have reduced verbal working memory have been shown to do well on tests of syntactic comprehension. Where effects of syntactic structure arise in these groups, the effects are comparable to those found in normal subjects. Concurrent verbal loads do not disproportionally affect comprehension of syntactically more complex sentences in these patient groups. These results are based on patients' performances on demanding concurrent tasks and include measurements of RTs as well as accuracy. They have documented a domain of retained functional capacity in the midst of the many cognitive and executive limitations found in these patients. In the aphasic population, where several arguments can be made that there is a reduction in processing resources used for syntactic comprehension, a concurrent verbal memory load does not exacerbate the effect of syntactic complexity. All these results are consistent with the view that the resources that are used in syntactic processing in sentence comprehension are not reduced in patients with reduced verbal working memory capacity, and are not shared by the digit span task. In contrast, patients with reduced working memory capacities have shown larger effects of the number of propositions in a sentence than control groups, and these effects were sometimes increased under concurrent load conditions. This suggests that the ability to match a proposition to a picture or to check its plausibility against information in semantic memory is affected in these patients, and affected by a concurrent digit span task. Though other accounts must be considered, this pattern of results is consistent with the view that the reduction in working memory seen in these patients affects these post-interpretive processes and that the resources used in these post-interpretive tasks are shared by immediate serial recall tasks.

 4. Discussion: Fractionating Verbal Working Memory

 The studies reviewed above provide strong evidence that subjects whose verbal working memory capacity is reduced on standard tests of this function can retain the ability to use syntactic structure to determine sentence meaning. This is not only true in normal subjects (i.e., those without neurological disease), in whom the measurement of working memory may be unstable over time and lead to misclassification, but also in patients with Alzheimer's Disease and Parkinson's Disease, whose verbal working memory and executive control functions are significantly impaired compared to normal subjects. In many experimental paradigms, the processing of syntactically more complex sentences is not disproportionately affected by a concurrent verbal memory load, either in normal subjects or in subjects with extremely reduced working memory capacity or in aphasic patients. These results provide evidence that the working memory system involved in sentence interpretation is separate from that measured by standard tests of working memory.

If there is a specialization within verbal working memory for the assignment of syntactic structure and its use in sentence interpretation, this specialization may not be restricted to this one aspect of the sentence interpretation process. Syntactic processing is one of a set of related operations that transform the acoustic signal into a discourse-coherent semantic representation in normal, everyday, conversations. A partial list of other types of operations in this process includes acoustic-phonetic conversion, lexical access, recognition of intonational contours, and determination of discourse-level semantic values such as topic, focus, co-reference, causality, and temporal order of events. These operations normally act in concert: a listener computes lexical, propositional, and discourse meanings in every normal communicative act (Marslen-Wilson, 1987). Most researchers consider that the integrated set of operations that activate representations in sentence and discourse interpretation has processing characteristics that are part of its usual successful functioning. Processors are thought to be obligatorily activated when their inputs are attended to. They generally operate unconsciously, and they usually operate quickly and remarkably accurately. The operations of the language interpretation process are thus integrated in two senses: they always compute items within the same restricted set of representational types, and they do so in a particular manner. Moreover, they are among the most highly practiced of human cognitive functions. Because of their integration and degree of overpractice, we have suggested that one resource system is utilized by all these different types of processes that combine in the interpretation process (Caplan and Waters, 1997; Waters and Caplan, 1996a). Because of its greater inclusiveness, we will refer to this hypothesis as the "separate-language-interpretation-resource" (SLIR) hypothesis, an extension of the "separate-sentence-interpretation-resource" (SSIR) hypothesis.

 According to SLIR hypothesis, the entire set of operations that compute a coherent discourse meaning depends upon a single working memory resource. It would therefore be supported by the finding that the efficiency of these operations is positively correlated across normal subjects and that the effects of increasing the working memory demands of different interpretive operations are interactive. Operationally, the entire system need not be tested for this hypothesis to begin to be investigated; these predictions can be applied to pairs of operations. For example, the theory could begin to be tested by investigating the effect of variables that affect lexical access, such as frequency, and of variables that affect sentence interpretation, such as syntactic complexity. In the model, the efficiency of post-interpretive operations is not related to that of interpretive operations. It therefore predicts that the efficiency of interpretive and post-interpretive operations is not correlated and that the effects of increasing the working memory demands of interpretive and non-interpretive operations are not interactive. The SLIR hypothesis predicts that working memory capacity, as measured on a task that emphasizes controlled conscious manipulation of verbal information, will not correlate with processing efficiency for any component of the interpretation process, that a concurrent verbal memory load will not interfere with more demanding processing within any interpretive component, and that each language processing component involved in interpretive processing will be preserved in some neurological subjects with reduced working memory.

The view that we have presented for an integrated set of language processing operations in the interpretation process has been challenged (Just, Carpenter, and Keller, 1996). One argument that has been made is that the comprehension process is not always as automatic and seamless as we have represented it to be. The sentence with which we began this paper --The man that the woman that the child hugged kissed laughed -- can be interpreted, but its successful interpretation is not usually obligatory, unconscious, and fast. Reanalysis of ambiguous lexical items (We hated the cheap hotel room because of all the bugs we saw in it. We realized our conversations would not be private) and of syntactic structures (The aggressive trial lawyer questioned in minute detail by the judge hesitated) also often involves conscious, slow, controlled processing. So does drawing certain types of complex inferences (Harvey frequently invested in junk bonds. He now sells pencils at the corner of poad and Main), or revising inferences (John returned home from the auction $500 poorer. He swore he would never attend a function in a rough part of town again). Listeners appear to accomplish sentence interpretation in two ways: in the usual obligatory, unconscious and fast mode, or in a mode in which conscious, controlled processes are applied to the task. We do not conceive of the latter type of processing as belonging to the set of operations that we suggest utilize a specialized resource pool. There have been several models of the change from one processing mode to another, especially in the area of syntactic reanalysis (e.g., Sturt and Crocker, 1996). A research program could be directed towards the question of whether the divisions made by these models predict the nature of the working memory system involved in processing different types of sentences.

 A second issue that has been raised is that the interpretation process -- specifically, syntactic processing -- is influenced by non linguistic factors such as pragmatic expectations (Trueswell et al, 1994) and the frequency with which particular lexical items co-occur in a language (MacDonald, Perlmutter and Seidenberg, 1994; see papers in MacDonald, 1997). These findings do not undermine the analysis presented here. The fact that the interpretation process can accept such information as input does not entail that it cannot be distinguished from other verbally mediated functions. One way to distinguish it is on the basis of the combination of its input, its intermediate representations, and its output compared to those of post-interpretive tasks. We may think of interpretive processes as a function that maps the acoustic signal onto a representation of the preferred literal meaning of sentences in a coherent discourse, that computes an intermediate set of linguistic representations (phonemes, words, syntactic structures, etc.). No other function accomplishes this mapping. Most post-interpretive processes are functions between the output of this process and logical entailments, confirmations of the presence of a proposition in a memory system, plans for action, or some other endpoint. Very few functions other than language interpretation compute any of the intermediate representations that are computed by the interpretation process, and none compute all the representations that are routinely computed by this process. In short, encapsulation in Fodor's (1983) terms was only one hypothesized feature of the interpretation process. The interpretation process does not have to be encapsulated to be cognitively separable or to rely on a specialized resource pool (Caplan, 1985).

 If we tentatively accept the view that there is a specialization in the verbal working memory for language interpretation, we may try to place this specialization within a neurobiological context. One can only speculate about the origin of such a specialization. It may be the product of innately determined neural development (Rakic, 1988), practice (Grafton et al, 1992, 1995), or both. There are no data presently available that determine this issue.

 Somewhat more is known about the neural systems that may support such a specialization, although knowledge in this area is also very incomplete. There is strong evidence that the dominant perisylvian association cortex is critical for all language interpretive functions (Caplan, 1987). This area is the best candidate for the location of the neural system that supports the processing resources used in language interpretation. Focusing specifically on syntactic processing in sentence comprehension, lesions throughout this region affect syntactic processing in sentence comprehension in ways that suggest a reduction in available processing resources (Caplan et al, 1985, 1996, 1997). It has been suggested that one region in the left perisylvian cortex -- the pars opercularis and traingularis of the third frontal convolution, podmann's areas 44 and 45 (poca's area) -- may be particularly important in this function. Evidence supporting this localization comes from the finding that on-line processing of sentences with relative clauses is abnormal in patients with poca's aphasia, who tend to have lesions that involve this region, but not in patients with fluent aphasias, whose lesions tend to spare this area (Zurif, Swinney, Prather, et al., 1993; Swinney and Zurif, 1995; Swinney et al, 1996). Studies of event-related potentials have also identified an early negative wave in the left frontal region associated with processing object-relative clauses (Neville, Nicol, Barss, et al., 1991; Kluender and Kutas, 1993) and two studies have shown a localized increase in regional cerepal blood flow (rCBF) in the pars opercularis when subjects made acceptability judgments about sentences with object- as opposed to subject-relative clauses (Stromswold, Caplan, Alpert and Rauch, 1996; Caplan, Alpert and Waters, in press). These results are consistent with a more narrow degree of localization of processing resources used in at least one aspect of syntactic comprehension in poca's area. Not all available data support this localization, however. Just, Carpenter, Keller, Eddy, and Thulborn (1996) have reported increased rCBF in both the left frontal and the left temporal lobe (and, to a lesser degree, in the homologous contralateral cortical regions) in a question-answering task using sentence types that were very similar to those tested using acceptability judgment by Stromswold et al (1996). These differences across studies are not yet explained. Overall, the picture that emerges is that association cortex in the dominant perisylvian region is supported in all likelihood by sustaining projections from thalamus and basal ganglia and probably also the contralateral non-dominant perisylvian association cortex, is responsible for language interpretation and the best candidate for the locus of the resources that are used in this aspect of language processing.

In contrast, many studies show that working memory tasks that involve conscious, controlled processing of language representations activate dorsolateral frontal regions of the left hemisphere rostral to the perisylvian association cortex (for representative results, see Jonides, Smith, Awh, Minoshima, & Monton, 1993, and Petrides, Alivisatos, Meyer, & Evans, 1993). The division of neural structures into a perisylvian cortical region involved in language interpretation and a more anterior region that supports conscious, controlled verbal working memory functions is consistent with the division within the verbal working memory system that we have advocated here.

It is worth noting that there are areas of the pain that may be involved in both types of tasks. For example, the anterior cingulate gyrus is activated in a wide variety of tasks -- both those that involve language interpretation and those that involve conscious, controlled verbal working memory -- and has been hypothesized to be a structure that is involved in setting subjects' level of arousal thereby allowing them to devote resources to a task (Posner et al, 1988). SLIR theory does not deny the existence of such regions, but its support comes from the existence of regions that contribute to working memory in only one of these types of language functions.

 In summary, we have presented evidence that the use of processing resources in assigning syntactic structure and using that structure to determine the meaning of a sentence is separable from a subject's verbal working memory capacity as measured by standard working memory tasks. We have accordingly argued that the working memory system contains specializations for different verbal processes and that one such specialization is used for the integrated processes involved in the determination of the meaning of discourse. Our hypothesis requires additional specification that can only come as models of both language processing and working memory become better developed and justified, but it can be tested in the context of present-day knowledge about both language processing and working memory.

Author Notes

 This research was supported by grants from the National Institute of Aging (AG09661) and from the Medical Research Council of Canada (MA9671) to David Caplan and Gloria Waters.

Reference NotesReferences

 Abney, S. P. (1989). A computational model of human parsing. Journal of Psycholinguistic Research, 18, 129-144.

 Baddeley, A.D. (1976). The Psychology of memory. New York: Basic Books.

 Baddeley, A. D. (1986). Working Memory . New York: Oxford University Press.

 Baddeley, A. (1995). Working Memory. In M. S. Gazzaniga (Ed.), The Cognitive Neurosciences (pp. 755-764). Campidge, Mass.: The MIT Press.

 Baddeley, A. D. (1993). Working memory and conscious awareness. In A. F. Collins, S. E. Gathercole, M. A. Conway, & P. E. Morris (Ed.), Theories of memory Hove: Lawrence Erlbaum Associates Ltd.

 Baddeley, A. D., & Hitch, G. J. (1974). The Psychology of Learning and Motivation . New York: Academic Press.

 Baddeley, A. D., & Hitch, G. J. (1994). Developments in the concept of working memory. Neuropsychology, 8, 485-493.

 Baddeley, A., Logie, R., pessi, S., Della Sala, S. D., & Spinnler, H. (1986). Dementia and working memory. The Quarterly Journal of Experimental Psychology, 38A, 608-618.

 Baddeley, A. D., pessi, S., Della Sala, S., Logie, R., & Spinnler, H. (1991). The decline of working memory in Alzheimer's disease. pain, 114, 2521-2542.

 Baum, S. (1991). Sensitivity to syntactic violations across the age span: Evidence from a word monitoring task. Journal of Clinical Linguistics and Phonetics, 5, 317-328.

 Bayles, K. A. (1982). Language function in senile dementia. pain and Language, 16, 265-280.

 Berndt, R., Mitchum, C., & Haendiges, A. (1996). Comprehension of reversible sentences in "agrammatism": a meta-analysis. Cognition, 58, 289-308.

 Berwick, R. C., & Weinberg, A. (1984). The Grammatical Basis of Linguistic Performance: Language Use and Acquisition . Campidge, Mass.: MIT Press.

 Blanken, G., Dittmar, J., Haas, J. C., & Wallesch, C. W. (1987). Spontaneous speech in senile dementia and aphasia: Implications for a neurolinguistic model of language production. Cognition, 27, 247-274.

 pown, R. G., & Marsden, C. D. (1988). Internal versus external cues and the control of attention in Parkinson's disease. pain, 111, 323-345.

 pown, R. G., & Marsden, C. D. (1991). Dual task performance and processing resources in normal subjects and patients with Parkinson's disease. pain, 114, 215-231.

 Caplan, D. ( 1985): Comments on Fodor's The Modularity of Mind: A NeoCartesian Alternative. Behaviour and pain Sciences; 8: 6-7.

 Caplan, D. (1987): Neurolinguistics and Linguistic Aphasiology: An Introduction, Campidge University Press: Campidge, U.K.

 Caplan, D. & Waters, G. S. (1990). Short-term memory and language comprehension: A critical review of the neuropsychological literature. In G. Vallar, & T. Shallice (Ed.), Neuropsychological Impairments of Short-Term Memory (pp. 337-389). Campidge: Campidge University Press.

 Caplan, D., & Waters, G. (1995). Aphasic disturbances of syntactic comprehension and working memory capacity. Cognitive Neuropsychology, 12, 637-649.

 Caplan, D., & Waters, G. (1996). Syntactic processing in sentence comprehension under dual-task conditions in aphasic patients. Language and Cognitive Processes, 11, 525-551

 Caplan, D., Baker, C., & Dehaut, F. (1985). Syntactic determinants of sentence comprehension in aphasia. Cognition, 21, 117-175.

 Caplan, D., Hildepandt, 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., Hildepandt, N., & Makris, N. (1996). Location of lesions in stroke patients with deficits in syntactic processing in sentence comprehension. pain ,119, 933-949

 Caplan, D., Hildepandt, N., & Waters, G. S. (1997). Determinants of Sentence Comprehension in Aphasic Patients in Sentence-Picture Matching Tasks, 40, 542-555

 Caplan, D., Alpert, N. and Waters, G.S. (in press). Effects of Syntactic Structure and Propositional Number on Patterns of Regional Cerepal Blood Flow, Journal of Cognitive Neuroscience

 Carpenter, P. A. & Just, M. A. (1988). The role of working memory in language comprehension. In D. Klahr & K. Kotovksy (Eds.), Complex information processing: The impact of Herbert A. Simon. Hillsdale, N.J.: Erlbaum.

Chomsky, N. (1957). Syntactic Structures . The Hague: Mouton.

 Chomsky, N. (1981). Lectures on Government and Binding. Dordrecht: Foris.

 Chomsky, N. (1986). Knowledge of Language . New York: Praeger.

 Chomsky, N. (1993). The Minimalist Program, Campidge, MA: MIT Press.

 Christiansen, M.H. (1996). Recursive Sentence Structure in Connectionist Networks, presented at the Ninth Annual CUNY Conference on Sentence Processing, CUNY, NY.

Christensen, A.L. (1974). Luria's neuropsychological investigation: Text. Copenhagen: Munksgaard.

 Cohen, L., & Mehler, J. (1996). Click monitoring revisited: An on-line study of sentence comprehension. Memory & Cognition, 24, 94-102.

 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. Psychological Bulletin and Review, 3, 422-433.

 Davis, C. A., & Ball, H. E. (1989). Effects of age on comprehension of complex sentences in adulthood. Journal of Speech and Hearing Research, 32, 143-150.

 DeRenzi, E. & Vignolo, L. (1962). The token test: A sensitive test to detect receptive disturbances in aphasics. pain, 85, 665-678.

 Emery, O. B. (1985). Language and aging. Experimental Aging Research (Monograph), 11, 3-60.

 Emery, O. B. (1986). Linguistic decrement in normal aging. Language & Communication, 6(1/2), 47-64.

 Engle, R. (1995), Overview of Working Memory, Conference on Cognitive Aging, Atlanta

 Feier, C. D., & Gerstman, L. J. (1980). Sentence comprehension abilities throughout the adult life span. Journal of Gerontology, 35, 722-728.

 Ferreira, F. (1992). Linguistic Processing is Modular, talk presented at the University of Wisconsin-Milwaukee Conference, Milwaukee, WS

 Ferreira, F., Anes, M. D., & Horine, M. D. (1996). Exploring the use of prosody during language comprehension using the auditory moving window technique. Journal of Psycholinguistic Research, 25, 273-290

 Ferreira, F., Henderson, J. M., Anes, M. D., Weeks Jr., P. A., & 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 & Cognition, 22, 324-335.

 Fodor, J. A. (1983). The Modularity of Mind . Campidge, Mass.: MIT Press.

 Ford, M. (1983). A method for obtaining measures of local parsing complexity throughout sentences. Journal of Verbal Learning and Verbal Behavior, 22, 203-218.

 Forster, K. I. (1979). Levels of Processing and the Structure of the Language Processor. In W. E. Cooper and E. C. T. Walker (Ed.), Sentence Processing (pp. 27-85). Hillsdale, N.J.: Erlbaum.

 Frauenfelder, U., Segui, J., & Mehler, J. (1980). Monitoring around the relative clause. Journal of Verbal Learning & Verbal Behavior, 19, 328-337.

 Frazier, L. (1990). Exploring the Architecture of the Language-Processing System. In G. T. M. Altmann (Ed.), Cognitive Models of Speech Processing: Psycholinguistic and Computational Perspectives (pp. 409-433). Campidge, MA: MIT Press (padford Books).

 Frazier L. and Rayner. K (1982): Making and correcting errors during sentence comprehension: eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 143, 178-210

 Frazier, L., & Clifton, C. (1996). Construal . Campidge, MA: MIT Press.

 Freidrich, F. J (1990): Multiple phonological representations and verbal short-term memory, in T. Shallice and G. Vallar (eds.): The Neuropsychology of Short-Term Memory, p.p. 74 - 93, Campidge University Press, Campidge, UK.

 Gathercole, S. E., & Baddeley, A. D. (1993). Working Memory and Language . Hillsdale, NJ: Erlbaum.

 Gibson. (1997). Syntactic complexity: Locality of syntactic dependencies. manuscript submitted for publication.

 Gibson, E., & Thomas, J. (1996). The processing complexity of English center-embedded and self-embedded structures. In C. Schutze (Ed.), Proceedings of the NELS 26 workshop on language processing: MIT Working Papers in Linguistics, Campidge, MA:

Gick, M. L., Craik, F. I. M., & Morris, R. G. (1988). Task complexity and age differences in working memory. Memory & Cognition, 16, 353-361.

 Grafton, S.T., Mazziota, J.C., Presty, S., Friston, K., Frackowiak, R.S.J., and Phelps, M.E. (1992): Functional anatomy of human procedural learning determined with regional cerepal blood flow and PET, Journal of Neuroscience, 12, 2542-2548

 Grafton, S.T., Hazeltine, E., and Ivry, R. (1995): Functional mapping of sequence learning in normal humans, Journal of Cognitive Neuroscience, 7, 497-510

 Grober, E., & Bang, S. (1995). Sentence comprehension in Alzheimer's disease. Developmental Neuropsychology, 11, 95-107.

 Grossman, M., Carvell, S., Gollomp, S., Stern, M. B., Vernon, G., & Hurtig, H. I. (1991). Sentence comprehension and praxis deficits in Parkinson's disease. Neurology, 41, 1620-1628.

 Grossman, M., Carvell, S., Stern, M. B., Gollomp, S., & Hurtig, H. (1992). Sentence comprehension in Parkinson's disease: The role of attention and memory. pain & Language, 42, 347-384.

 Hart, S. (1988). Language and dementia: A review. Psychological Medicine, 18, 99-112.

 Hier, D. B., Hagenlocker, K., & Schindler, A. G. (1985). Language disintegration in dementia: Effects of etiology and severity. pain and Language, 25, 117-133.

 Hildepandt, N., Caplan, D., & Evans, K. (1987). The mani lefti without a trace: a case study of aphasic processing of empty categories. Cognitive Neuropsychology, 4, 257-302.

 Illes, J. (1989). Neurolinguistic features of spontaneous language production dissociate three forms of neurodegenerative disease: Alzheimer's, Huntington's, and Parkinson's. pain and Language, 37, 628-642.

 Irigaray, L. (1973). Le langage des dements . The Hague: Mouton.

 Johnson, M. (1996). Resource sensitivity in grammar and processing. Presentation at The Ninth Annual CUNY Conference on Human Sentence Processing,

Jonides, J., Smith, E. E., Awh, E., Minoshima, S., & Monton, M. A. (1993). Spatial working memory in humans as revealed by PET. Nature, 363, 623-625.

 Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99(1), 122-149.

 Just, M.A., Carpenter, P.A., and Keller, T. (1996) Working Memory: New Frontiers of Evidence, Psychological Review, 103, 773-775

 Just, M. A., Carpenter, P. A., Keller, T. A., Eddy, W. F., & Thulborn, K. R. (1996). pain activation modulated by sentence comprehension. Science, 274, 114-116.

 Kemper, S. (1986). Imitation of complex syntactic constructions by elderly adults. Applied Psycholinguistics, 7, 277-287.

 Kemper, S. (1988). Geriatric psycholinguistics: Syntactic limitations of oral and written language. In L. L. Light, & D. M. Burke (Ed.), Language, Memory, and Aging (pp. 58-76). New York: Campidge University Press.

 Kemper, S., & Rash, S. (1988). Practical Aspects of Memory: Current Research and Issues . London: Campidge University Press.

 Kemper, S., Snowdon, D. A., & Greiner, L. H. (1996). Linguistic ability across the life span: Findings from The Nun Study. Paper presented at the Cognitive Aging Conference, Atlanta

 Kemper, S., Anagnopoulos, C., Lyons, K., & Heberlein, W. (1994). Speech accommodations to dementia. Journal of Gerontology, 49, 223-229.

 Kemper, S., LaBarge, E., Ferraro, F. R., Cheung, H., Cheung, H., & Storandt, M. (1993). On the preservation of syntax in Alzheimer's disease: Evidence from written sentences. Archives of Neurology, 50, 81-86.

 Kempler, D., Curtiss, S., & Jackson, C. (1987). Syntactic preservation in Alzheimer's disease. Journal of Speech and Hearing Research, 30, 343-350.

 Kemtes, K. A., & Kemper, S. (1996). Younger and older adults' working memory and on-line processing of syntactically ambiguous sentences. Poster presented at the Cognitive Aging Conference, Atlanta,

King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580-602.

 Kirshner, H. S. (1982). Language disorders in dementia. In H. S. Kirshner, & F. R. Freeman (Ed.), The neurology of aphasia (pp. 187-196). Lisse, Holland: Swets and Zeitlinger.

 Kluender, R., & Kutas, M. (1993). Subjacency as a processing phenomenon. Language and Cognitive Processes, 8, 573-633.

 Kontiola, P., Laaksonen, R., Sulkava, R., & Erkinjuntti, T. (1990). Pattern of language impairment is different in Alzheimer's disease and multi-infarct dementia. pain and Language, 38, 364-383.

 Krueger, L. E. (1994). Mind your ps and qs when accepting or rejecting the null hypothesis. presented and the Psychonomic Society, St. Louis

 Lalami, L., Marblestone, K., Schuster, S., Andersen, E., Kempler, D., Tyler, L., & MacDonald, M. (1996). On-line vs. off-line sentence processing in Alzheimer's disease. Poster presented at the Meeting of Cognitive Aging, Atlanta,

Lees, A. J., & Smith, E. (1983). Cognitive deficits in the early stages of Parkinson's disease. pain, 106, 257-270.

 Lewis, R. (1996). A theory of grammatical but unacceptable embeddings. Journal of Psycholinguistic Research, 25, 93-116.

 Lieberman, P., Friedman, J., & Feldman, L. S. (1990). Syntax comprehension in Parkinson's disease. Journal of Nervous and Mental Diseases, 178, 360-366.

 Light, L. L. (1990). Interactions between memory and language in old age. In J. E. Birren, & K. W. Schaie (Ed.), Handbook of the Psychology of Aging (3rd edition) (pp. 275-290). San Diego: Academic.

 MacDonald, M. C. (1997). Language and Cognitive Processes: Special Issue on Lexical Representations and Sentence Processing, 12, 121-399.

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). Lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676-703.

 Marslen-Wilson, W. D. (1975). Sentence perception as an interactive parallel process. Science, 189, 226-228.

Marslen-Wilson, W. D. (1987). Functional parallelism in spoken word-recognition. Cognition, 25, 71-102.

 Martin, R. (1990). The consequences of reduced memory span for the comprehension of semantic versus syntactic information. pain and Language, 38, 1-20.

 Miyake, A., Carpenter, P., & Just, M. (1994). A capacity approach to syntactic comprehension disorders: Making normal adults perform like aphasic patients. Cognitive Neuropsychology, 11, 671-717.

 Morris, R. G. (1984). Dementia and the Functioning of the Articulatory Loop System. Cog Neuropsych, 1(2), 143-157.

 Morris, R. G. (1986). Short-Term Forgetting in Senile Dementia of the Alzheimer's Type. Cog Neuropsych, 3(1), 77-97.

 Morris, R. G. (1987). Articulatory Rehearsal in Alzheimer Type Dementia. pain and Language, 30, 351-362.

 Morris, R. G., & Baddeley, A. D. (1988). Primary and Working Memory Functioning in Alzheimer-type Dementia. Journal of Clinical and Experimental Neuropsychology, 10(2), 279-296.

 Morris, R. G., Craik, F. I. M., & Gick, M. L. (1990). Age differences in working memory tasks: The role of secondary memory and the central executive system. Quarterly Journal of Experimental Psychology, 42A, 67-86.

 Morris, R. G., Gick, M. L., & Craik, F. I. M. (1988). Processing resources and age differences in working memory. Memory and Cognition, 16, 362-366.

 Natsopoulos, D., Katsarou, Z., Bostantzopoulo, S., Grouios, G., Mentanopoulos, G., & Logothetis, J. (1991). Strategies in comprehension of relative clauses in Parkinsonian patients. Cortex, 27, 255-268.

 Neville, H., Nicol, J. L., Barss, A., Forster, K. I., & Garrett, M. F. (1991). Syntactically based sentence processing classes: Evidence from event-related pain potentials. Journal of Cognitive Neuroscience, 3, 151-165.

 Norman, S., Kemper, S., Kynette, D., Cheung, H., & Anagnopoulous, C. (1991). Syntactic complexity and adults' running memory span. Journal of Gerontology: Psychological Sciences, 46, 346-351.

 Obler, L. K., Fein, D., Nicholas, M., & Albert, M. L. (1991). Auditory comprehension and aging: Decline in syntactic processing. Applied Psycholinguistics, 12, 433-452.

 Pearlmutter, N. J., & MacDonald, M. C. (1995). Individual differences and probabilistic constraints in syntactic ambiguity resolution. Journal of Memory and Language, 34, 521-542.

 Petrides, M., Alivisatos, B., Meyer, E., & Evans, A. (1993). Functional activation of the human frontal cortex during the performance of verbal working memory tasks. Neurobiology, 90, 878-882.

 Posner, M. I., Peterson, S. E., Fox, P. T., & Raichle, M. E. (1988). Localization of cognitive operations in the human pain. Science, 240, 1627-1631.

 Rakic, P (1988): Specification of cerepal cortical areas, Science, 241, 170-6.

 Rochon, E., & Saffran, E. M. (1995). A semantic contribution to sentence comprehension impairments in Alzheimer's disease. Poster presented at the Academy of Aphasia, San Diego,

Rochon, E., Waters, G. S., & Caplan, D. (1994). Sentence comprehension in patients with Alzheimer's Disease. pain & Language, 46, 329-349.

 Salthouse, T. A. (1991). Theoretical perspectives on cognitive aging. Hillsdale, NJ: Erlbaum.

 Schneider, W., and Shiffrin, R.M. (1977) Controlled and automatic human information processing I: Detection, search, and attention, Psychological Review, 84, 1 - 66

 Schwartz, M. F., Marin, O. S. M., & Saffran, E. M. (1979). Dissociations of language function in dementia: A case study. pain and Language, 7, 277-306.

 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.

 Sherman, J. C., Schweickert, J., Growdon, J., & Corkin, S. (1988). Sentence comprehension in Alzheimer's disease. Paper presented at the 18th annual meeting of the Society of Neuroscience, Toronto,

Shiffrin, R.M., and Schneider, W. (1977) Controlled and automatic human information processing II: Perceptual learning, automatic attending, and a general theory, Psychological Review, 84, 127-190 Sidner, C. L. (1979). A computational model of co-reference comprehension in English. Ph.D. thesis,

Smith, S. (1989). Syntactic comprehension in Alzheimer's Disease. Santa Fe. NM:

Snowdon, D. A., Kemper, S. J., Mortimer, J. A., Greiner, L. H., Wekstein, D. R., & Markesbery, W. R. (1996). Linguistic ability in early life and cognitive function and Alzheimer's disease in late life: Findings from The Nun Study. Journal of the American Medical Association, 275, 528-532.

 Stevens, K., Kempler, D., Andersen, E., & MacDonald, M. C. (1996). Preserved use of semantic and syntactic context in Alzheimer's disease. Poster presented at the Meeting of Cognitive Aging, Atlanta.

Sturt, P., & Crocker, M. W. (1996). Monotonic syntactic processing: A cross-linguistic study of attachment and reanalysis. Language and Cognitive Processes, 11, 449-494.

 Squire, L. R., & Zola-Morgan, S. (1991). The medial temporal lobe memory system. Science, 253, 1380-1386.

 Stromswold, K., Caplan, D., Alpert, N., & Rauch, S. (1996). Localization of syntactic comprehension by positron emission tomography. pain and Language, 52, 452-473.

 Swinney, D., Zurif, E., Prather, P., & Love, T. (1996). Neurological Distribution of Processing Resources Underlying Language Comprehension. Journal of Cognitive Neuroscience, 8, 174-184.

 Swinney, D., & Zurif, E. (1995). Syntactic processing in aphasia. pain and Language, 50, 225-239.

 Tabor, W., Juliano, C., & Tanenhaus, M. K. (1997). Parsing in a dynamical system: A attractor-based account of the interaction of lexical and structural constraints in sentence processing. Language and Cognitive Processes, 12, 211-271.

 Taylor, A. E., Saint-Cyr, J. A., & Lang, A. E. (1987). Frontal lobe dysfunction in Parkinson's disease. pain, 109,

Tirre, W., & Pena, C. (1992). Investigation of functional working memory in the reading span test. Journal of Educational Psychology, 84, 462-472.

Tomoeda, C. K., Bayles, K. A., Boone, D. R., Kaszniak, A. W., & Slauson, T. J. (1990). Speech rate and syntactic complexity effects on the auditory comprehension of Alzheimer patients. J. Com. Disorders, 23, 151-161.

 Turner, M., and Engle, R.W. (1989). Is working memory capacity task dependent? Journal of Memory and Language, 28, 127-154.

 Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influence on syntactic processing: Use of thematic information in syntactic disambiguation. Journal of Memory and Language, 33, 265-312.

Vallar, G. & Shallice, T. (eds.) (1990). Neuropsychological impairments of short-term memory. Campidge, UK: Campidge University Press,

 Waldstein, R., & Baum, S. (1992). The influence of syntactic and semantic context on word monitoring latencies in normal aging. Journal of Speech Language Pathology and Audiology, 6, 217-222.

 Wanner, E., & Maratsos, M. (1978). An ATN approach to comprehension. In M. M. Halle, G. Miller, & J. pesnan (Ed.), Linguistic Theory and Psychological Reality Campidge, MA: MIT Press.

 Waters, G., & Caplan, D. (1996a). The capacity theory of sentence comprehension: A reply to Just and Carpenter. Psychological Review, 103, 761-772

Waters, G., & Caplan, D. (1996b). Processing resource capacity and the comprehension of garden path sentences. Memory and Cognition, 24, 342-355.

 Waters, G., & Caplan, D. (1996c). The measurement of verbal working memory capacity and its relation to reading comprehension. The Quarterly Journal of Experimental Psychology, 49A, 51-79.

 Waters, G., & 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., Caplan, D., & Hildepandt, N. (1987). Working memory and written sentence comprehension. In M. Coltheart (Ed.), Attention and Performance XII: The Psychology of Reading (pp. 531-555). London: Lawrence Erlbaum.

 Waters, G. S., Caplan, D., & Hildepandt, N. (1991). On the structure of verbal short-term memory and its functional role in sentence comprehension: Evidence from neuropsychology. Cognitive Neuropsychology, 8(2), 81-126.

 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., Rochon, E., & Caplan, D. (in press). Task demands and sentence comprehension in patients with dementia of the Alzheimer's type. pain and Language .

 Waters, G., & Caplan, D. (under review ). The effect of a digit load on syntactic processing in subjects with high and low working memory capacity. Psychonomic Bulletin and Review,

Wright, R. E. (1981). Aging, divided attention, and processing capacity. Journal of Gerontology,

 36(No. 5), 605-614.

 Zurif, E., Swinney, D., Prather, P., Solomon, J., & Bushell, C. (1993). An on-line analysis of syntactic processing in poca's and Wernicke's Aphasia. pain and Language, 45, 448-464.

 Zurif, E., Swinney, D., Prather, P., Wingfield, A., & pownell, H. (1995). The allocation of memory resources during sentence comprehension: Evidence from the elderly. Journal of Psycholinguistic Research, 24, 165-182.



[1] These predictions depend on the assumption that the slope of the relationship between resource demands and sentence complexity is monotonically increasing. This simplifying assumption has been made throughout the literature on sentence processing and working memory. We make it in interpreting the results presented below, recognizing that it will have to be empirically validated or that more complex models will have to be adopted and tested.

 [2] This conclusion depends upon the size of the resource pools in a separate-sentence-interpretation-resource model being uncorrelated. However, though the possibility of correlated sizes of resource pools provides an "out" for the separate-sentence-interpretation-resource model in the case that there are interactions between sentence complexity and working memory capacity group, it is fair to say that such interactions still favor the single-resource model, at least by making the separate-resource model adopt complex post-hoc and ad-hoc assumptions.

 [3] Span is often contrasted with concurrent articulation (repeating a single word like "the" or "double"), which is thought to interfere with only rehearsal, not compete with another task for central executive resources (Baddeley and Hitch, 1974; Just and Carpenter, 1992).

[4] These predictions also depend upon assumptions about floor and ceiling effects in experimental results. For instance, if the combination of task and sentence processing demands is less than the working memory resources available to even low-span subjects, ceiling effects would be expected to eliminate interactions between load, sentence type and span group. Much more complex patterns of results can be predicted if other assumptions about floor and ceiling effects are made (see Waters and Caplan, 1996a, footnote 3).