Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Sentence Comprehension in Aphasia
David Glenn Clark*
University of Alabama at Birmingham
Abstract
Aphasia is a disorder of language that sometimes occurs in the setting of brain damage. Aphasic
disturbances of syntactic comprehension have been described only in recent decades, and have
been characterized as the loss of the ability to understand non-canonical sentences. Several interesting hypotheses have been advanced to explain this phenomenon. Some of these hypotheses rest
on the assumption that heuristics play a role in producing the aphasic comprehension pattern. Evidence from the past decade suggests that heuristic processes do play a role in normal sentence
comprehension, and act in concert with slower processes that more closely resemble the syntax of
linguistic theory. The slower syntactic mechanisms appear to be more readily disturbed by brain
damage, leading to greater difficulty with sentences that cannot be interpreted heuristically. However, the dichotomy between canonical and non-canonical sentences accounts for little of the variance in the performance of aphasic individuals. Instead, the comprehension of each sentence
structure is disrupted according to the complexity of the structure and the severity of the individual’s aphasia. This observation argues against hypotheses that posit a deficit of a specific linguistic
operation and favors reduction of a hypothetical cognitive resource that is necessary for normal
comprehension. In this case, the chief contribution from linguistic theory is the syntactic tree
itself, which affords a concise characterization of the difficulty of parsing and interpreting a given
sentence. The field now faces several challenges. The neural basis of heuristic, parsing, and interpretive mechanisms must be elucidated. Achievement of this objective entails a neural characterization of the cognitive resource that is disrupted in sentence comprehension disorders and of the
semantic representations onto which morphemes and sentences are mapped.
Introduction
APHASIA
Aphasia is an acquired disorder of language that is caused by brain damage. Common
causes of aphasia include stroke, trauma, brain tumor, and neurodegenerative diseases.
Clinicians who care for patients with aphasia routinely classify patients parametrically
according to the presence or absence of impairment of four gross linguistic skills: naming,
fluency, repetition, and comprehension. Each of these parameters is a complex language
act that requires numerous processing steps, all of which are supported by a distributed
network of neurons. Naming impairment of some sort is assumed to be universal among
patients with aphasia, while relative sparing or impairment of the other three abilities
yields eight clinical categories of aphasia that have more or less well-defined cerebral
localizations (see Table 1) (Damasio 1992; Mendez and Clark 2008).
Linguists impose divisions on language which differ greatly from the parameters that
generate the clinical taxonomy of aphasic disorders. Linguistic theory breaks language
behavior down according to levels of representation, including phonetics, phonology,
morphology, syntax, semantics, and pragmatics. Any given aphasic individual may exhibit
ª 2011 The Author
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
Sentence Comprehension in Aphasia
719
Table 1. Eight types of aphasia, defined by four parameters of spoken ⁄ auditory language
performance. All aphasia syndromes listed here are associated with impaired naming, but
the other three parameters may be relatively spared or impaired. ‘X’ indicates relative
impairment; ‘ok’ indicates relative sparing.
Naming
Fluency
Repetition
Comprehension
Lesion
Anomic
X
ok
ok
ok
Broca’s
X
X
X
ok
Wernicke’s
X
ok
X
X
Conduction
X
ok
X
ok
Transcortical
sensory
X
ok
ok
X
Transcortical
motor
X
X
ok
ok
Transcortical
mixed
X
X
ok
X
Global
X
X
X
X
Almost anywhere in the cerebral
cortex, left hemisphere more
likely
Usually anterior, especially
inferior frontal
May involve posterior superior
and middle temporal gyri,
inferior parietal regions
Inferior parietal, superior
temporal sulcus, subcortical
white matter
Posterior watershed region
between territories of middle
and posterior cerebral arteries;
involves parieto-occipital and
temporo-occipital regions
Either large dorsolateral frontal
or mesial frontal, involving
supplementary motor area and
anterior cingulate gyrus
Large, possibly outlining the
perisylvian region through
involvement of watershed
zones anterior and posterior to
area supplied by MCA
Large, involving majority of left
perisylvian region
MCA, middle cerebral artery.
deficits which impact several levels of linguistic representation. This review focuses on
syntax, which is generally considered to be ‘central’ to language function, constituting a
sort of interface between sounds and meanings. Due to the functional interposition of
syntax between sound and meaning, the discussion of sentence comprehension deficits in
the literature often encroaches on the syntax-like field of morphology (the grouping of
meaningful sound segments into words), and on the more abstract field of semantics (the
representation of linguistic meanings in the mind).
The study of brain-language relationships would be greatly simplified if there were
anatomically defined neural centers that transparently supported clinical aphasia parameters
(e.g., fluency or naming) or levels of linguistic representation (e.g., phonology or syntax),
but this is not the case. Instead, the brain comprises a vast network of microscopic neurons, all of which are processing information simultaneously and transmitting results to
other neurons. During the last 30 years a view of the global organization of the brain has
emerged that blends localizationist views (in which parts of the brain are devoted to single, complex functions) with equipotential views (in which the entire brain could be
involved in any function). The dominant view among neurologists is that cognition
results from the activity of large-scale, partially distributed networks (Mesulam 1990).
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
720 David G. Clark
Each neuron is relatively specialized for processing a certain kind of information, such as
sounds within a narrow frequency range, lines of a certain orientation, or more complex
stimuli such as faces or entire words. Situating neurons with similar response properties
close together reduces the total volume of wiring connections and might render the system more efficient (Buzsaki, et al. 1991; Chklovskii and Koulakov 2004). Tight groups
of neurons that process the same kind of information form anatomically distinct nodes in
a given cognitive network. Selective linkage of distant nodes permits the binding together
of distinct streams of information, and results in the overall structure of a small-world
network (Bassett and Bullmore 2006; Watts and Strogatz 1998). Language functions are
supported by a vast cerebral network, most of which is situated in the left hemisphere,
around the sylvian fissure (Caplan et al. 1996). Brain lesions that impact linguistic or
syntactic performance are usually located in this region.
AGRAMMATISM
Aphasiologists commonly refer to morphological or syntactic linguistic disturbances as
agrammatism. The term has been in use since the late 19th century and is used to describe
the language output of certain (but not all) non-fluent patients (Goodglass 1993). Agrammatic individuals suffer from a deficit of language production at the level of phrases or sentences. Their speech is effortful and consists of brief phrases that contain mainly nouns,
often with omission of grammatical features such as auxiliary verbs, prepositions, or
inflectional morphemes (Damasio 1992; Goodglass 1993).
Caramazza and Zurif are the first to apply the term agrammatism to faulty language
comprehension in their report of a dissociation that pertains to language understanding at
the level of phrases or sentences (Caramazza and Zurif 1976). Specifically, they find that
patients with certain aphasic syndromes, who initially appear to have intact auditory comprehension, actually have difficulty understanding the specifics of semantically reversible,
non-canonical sentences. The experiment involves the presentation of sentences with
center-embedded relative clauses, similar to sentences (1–4), below.
(1)
(2)
(3)
(4)
The
The
The
The
boy that eats the apple is wearing a green shirt.
boy that kicks the girl is wearing a green shirt.
apple that the boy eats is green.
boy that the girl kicks is wearing a green shirt.
The subject and the contents of the relative clauses from sentences (2) and (4) are semantically reversible, because boys may kick girls or vice versa. Sentences (3) and (4) are noncanonical, because they do not follow the typical subject-verb-object (SVO) word order
of English. However, only (4) is both semantically reversible and non-canonical. Caramazza and Zurif report that when patients with Broca’s aphasia or conduction aphasia are
asked to select a picture (from several distracters) that accurately depicts the meaning of
this type of sentence, they often confuse who is doing the kicking and who is being
kicked (a process referred to as ‘thematic role assignment’), and generally perform at a
level that does not differ statistically from chance. They surmise that language comprehension processes include both simple rules that may be applied to a string of words
(heuristics) and processes that act on a more complex mental representation of groups of
words (which they term algorithms). They suggest that these aphasic patients suffer from
disruption of syntax-like algorithmic processes, and that the unchecked heuristic processes
lead the patients to false conclusions regarding the meanings of the sentences, particularly
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
Sentence Comprehension in Aphasia
721
the assignment of thematic roles. The general pattern of agrammatic comprehension has
been observed with other non-canonical structures, such as passive voice sentences (see,
for example, Grodzinsky 2000b).
Despite the attraction of syntactic explanations for agrammatic comprehension, however, patients with this pattern of comprehension performance appear to have preserved
sensitivity to syntactic structure. Specifically, they retain the ability to make accurate
judgments of sentence grammaticality, at least with certain sentence structures (Linebarger
et al. 1983; Wilson and Saygin 2004). In light of this discovery, Linebarger et al. (1983)
propose two ideas to explain the agrammatic comprehension pattern. The first is that
agrammatic patients are able to construct syntactic representations, but are unable to use
these representations to derive sentence meaning. The second idea is that the patients suffer from a resource limitation that results in a combination of reduced parsing efficiency
and reduced capacity for compositional semantic interpretation, such that different tasks
produce qualitatively different patterns. That is, when subjects are not required to make
semantic judgments on sentences, they are able to devote more resources to accurate
syntactic analysis. The dissociation between syntax comprehension and grammaticality
judgments remains a topic of debate.
THE MODERN APPROACH TO THE STUDY OF APHASIA
Two important developments in the previous century set the stage for the modern psycholinguistic approach to studying syntax in the brain. First, Noam Chomsky revolutionized the study of linguistics by rejecting behaviorism and introducing elements of formal
language theory to characterize mental representations of syntactic structures (Chomsky
1957, 1959). The goals of linguistics are generally couched in computational terms (e.g.,
to devise a formal system that generates all and only the sentences of human languages),
and most linguists consider language itself to be their object of study, rather than the
minds that use language or the neural tissue that gives rise to it. Nevertheless, linguistics
provides a foundation for language description that is essential for evaluating aphasia and
for designing experiments that seek to elucidate relationships between language and the
brain. Basic grammatical constructs, such as nouns and verbs, are now known to be differentially impacted by brain damage and to be supported by anatomically dissociable cortical networks (Bak et al. 2001; Hillis et al. 2002, 2004). It remains to be seen, however,
which abstract theoretical constructs that have been devised by linguists are suitable for
generating or evaluating neuroscientific hypotheses.
Second, as described above, Caramazza and Zurif (1976) discovered that patients with
Broca’s or conduction aphasia exhibit difficulty assigning thematic roles in sentences with
object-relative clauses, despite continuing to perform well with subject-relative clauses.
This finding has been replicated and extended, and has led researchers to pose a number
of interesting questions pertaining to the neural and computational bases for sentence
production and comprehension.
Agrammatic Comprehension
THE TRACE DELETION HYPOTHESIS AND THE DOUBLE DEPENDENCY HYPOTHESIS
Most investigators use abstractions from linguistic theory to characterize or explain the
pattern of deficits of comprehension among agrammatic individuals. The Trace Deletion
Hypothesis is one such approach, and probably generates more discussion and controversy
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
722 David G. Clark
than any other single hypothesis (Beretta and Munn 1998; Beretta, Pinango, Patterson, &
Harford, 1999; Berndt et al. 1997; Caplan et al. 2006; Caramazza et al. 2005; Drai and
Grodzinsky 2006; Grodzinsky 2000a,b). The term ‘trace’ refers to a phonologically null
element in the syntactic representation that occupies the former position of a syntactic
constituent that has been moved to another location in the sentence. For example, some
linguistic analyses of sentence (4), repeated here, include a trace following the verb
‘kicks’, which is decorated with the same subscript (co-indexed) with the moved element
‘the boy’.
(4)
The boyi that the girl kicks ti is wearing a green shirt.
From a psychological standpoint, the implication is that when normal subjects listen to a
sentence, they unconsciously construct a syntactic representation that describes the dependencies among the constituents of the sentence. Since traces are phonologically null, their
presence must be inferred from other cues in the signal (e.g., the presence of the complementizer ‘that,’ and the absence of an object following the verb ‘kicks’). Traces must then
be linked or associated with appropriate moved constituents in the sentence in order for
the listener to derive the correct meaning from the sentence. Syntacticians represent this
mental association with co-indexation.
Grodzinsky (1995a,b, 2000a,b) proposes that the deficit underlying agrammatic comprehension is the deletion of traces from the syntactic representation. Thus, when individuals with agrammatic aphasia hear sentence (4), they are unable to establish a link
between the verb ‘kicks’ and the moved constituent ‘the boy’. These patients maintain
knowledge of the meaning of ‘kick’, including the fact that it takes two arguments (an
agent who kicks and a theme or patient that gets kicked). However, in the absence of normal syntactic mechanisms, Grodzinksy argues that they follow a ‘default’ (heuristic) strategy in which they always assign the agent role to the first noun phrase (NP) in the
sentence. (For an explication of the default strategy, see Grodzinsky 1990: 97). If a subsequent NP is assigned the role of agent through intact syntactic mechanisms, they are led
to the conclusion that there are two agents, and then are left with no option but to guess.
Sentences with fewer moved elements or with only vacuous movement (such as active
voice sentences and sentences with subject-clefts or subject-relative clauses) are therefore
comprehended fairly well, while those in which movement disrupts the canonical word
order of the sentence are comprehended poorly. Thus, performance falls to chance level
on passive voice sentences and sentences with object-clefts or object-relative clauses.
The TDH captures much of the observed data from sentences with theoretically moved
constituents.
One source of controversy around the TDH regards the specifics of the default strategy. Part of the argument against the use of the default strategy is that such heuristics are
not a part of syntactic theory. This criticism loses some of its force in light of more recent
psycholinguistic evidence that normal sentence comprehension is facilitated by heuristic
strategies (Ferreira 2003; Ferreira and Patson 2007; Ferreira et al. 2002; Townsend and
Bever 2001). A more specific objection is that the default strategy, as defined by proponents of the TDH, should lead agrammatic patients to interpret verbs in certain sentence
structures as having two agents. Thus, one might expect an aphasic person using this
strategy (to interpret a non-canonical sentence) to prefer pictures in which both NP referents are performing an action over pictures that correctly depict the event. This prediction is not supported by experimental work in which individuals with agrammatic
comprehension are tested with materials that allow for this choice. The agrammatic
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
Sentence Comprehension in Aphasia
723
individuals rarely select the double-agent pictures (Beretta and Munn 1998). In addition,
the default strategy causes the TDH to make incorrect predictions regarding the interpretation of certain passive constructions in Venezuelan Spanish (Beretta et al. 1999).
The Double Dependency Hypothesis (DDH, Mauner et al. 1993) has been advanced
with the goal of explaining the agrammatic pattern of performance without resorting to
heuristic strategies that have not been a part of the traditional linguistic theory. According
to this hypothesis, patients with agrammatism are unconstrained when co-indexing
dependencies in the syntactic representation. This lack of constraint does not affect the
interpretation of sentences with only one referential dependency, such as simple active
voice sentences and subject-clefts. However, it leads agrammatic individuals to ambiguous
conclusions when attempting to understand sentences with two or more referential
dependencies, such as object-relative sentences. For example, consider sentence (4) once
more, now decorated further with an additional trace motivated by the Verb PhraseInternal Subject Hypothesis (Koopman and Sportiche 1991). (For a version of the TDH
that takes into account the VP-Internal Subject Hypothesis, see Hickok et al. 1993).
(4)
The boyi that the girlj tj kicks ti is wearing a green shirt.
This sentence requires the listener to establish two chains of referential dependencies
(shown here enclosed in angle brackets):
(5)
<[the boy]i, ti> and <[the girl]j, tj>
According to the DDH, however, patients with agrammatic comprehension are not constrained by normal co-indexation patterns when establishing chains and therefore produce
the following aberrant chains in addition to the correct ones:
(6)
<[the boy]i, tj> and <[the girl]j, ti>
The aphasic person, faced with these two contradictory possibilities, is therefore required
to guess which one is correct, and therefore performs at ‘chance’ level on non-canonical
material. Subject-relative sentences and other canonical constructions have only one relevant dependency that requires co-indexation. Because no ambiguity is generated for these
sentences, patients perform ‘above chance’.
While the DDH has the advantage of a stronger foundation in theoretical syntax (i.e.,
no requirement for atheoretical heuristics), and partitions the data from scrambled passives
in Venezuelan Spanish appropriately (Beretta et al. 1999), evidence from two Mandarin
Chinese-speaking aphasic individuals offers some support to the TDH (Su et al. 2007), or
at least supports the use of heuristics in deriving sentence meaning. The study of Mandarin Chinese is important, because it happens that the TDH and the DDH make opposite
predictions regarding the comprehension of relative clauses in Mandarin. As in English,
when analyzed with the VP-Internal Subject Hypothesis, the Mandarin object-relative
sentence is expected to have two referential dependencies. The DDH therefore predicts
that individuals with agrammatic aphasia will perform at chance on this structure, but
above chance on subject-relative sentences. However, although Mandarin is an SVO language, relative clauses in Mandarin precede the head noun, while relative clauses in western languages typically follow the head noun. Sentence (7) is an example of a Mandarin
subject-relative sentence, and sentence (8) is an example of a Mandarin object-relative
sentence (see sentences 7 and 8).
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
724 David G. Clark
(7)
ti
zhui
gou de
maoi hen
xiao
chase dog COMP cat
very small
The cat that chased the dog was very small.
(8)
gouj tj zhui
ti de
maoi hen
xiao
dog
chase
COMP cat
very small
The cat that the dog chased was very small.
Because of this difference in relative clause placement, Mandarin subject-relative sentences
are non-canonical, while object-relative sentences are canonical. For Mandarin, therefore,
the TDH predicts worse performance with subject-relative sentences than with objectrelative sentences, a reversal of the pattern predicted by the DDH. The authors observe
the pattern predicted by the TDH in two out of six Mandarin-speaking individuals with
aphasia, but do not observe the pattern predicted by the DDH.
Further evidence that heuristic processes play a role in sentence comprehension arises
from Su, Lee, & Chung’s (2007) observations of patient performance when interpreting
the predicate adjectives of Mandarin sentences like (7) and (8) or their English translations.
Note that both Mandarin sentences (7) and (8) contain the substring ‘cat very small’, which
has a meaning consistent with the meaning of the matrix clause. In contrast, the English
translation of sentence (7) contains the substring ‘the dog was very small’ (which has a
meaning quite different from the meaning of the matrix clause). The presence of these
local sequences may provide a simple explanation for the observation that Mandarin
patients perform well on interpretation of the predicate adjectives for both sentence structures, while English-speaking patients exhibit difficulty interpreting predicate adjectives in
the matrix clause of center-embedded subject-relative sentences (Hickok et al. 1993).
RESOURCE REDUCTION
A separate body of work, largely due to David Caplan et al. suggests that sentence comprehension failure that occurs in the setting of aphasia is the result of a single defect: specifically, a shortage of some resource that supports the computational processes necessary
for constructing syntactic representations and deriving interpretations of them. One possibility for this resource is a specialized subset of verbal working memory (Caplan and
Waters 1999), but other possibilities have been suggested (Caplan 2006; Caplan et al.
2007). This approach contradicts the more linguistically-motivated hypotheses, each
of which implicates damage to a specific feature of the grammar (e.g., trace deletion).
A consequence of this specificity is that these hypotheses predict that agrammatism will
impact sentences with certain syntactic structures, while other sentences will not be
affected. In contrast, the resource reduction hypothesis predicts that all sentences will be
affected by aphasia to varying degrees, depending on their overall complexity.
Several lines of evidence support the resource reduction hypothesis (Caplan, Hildebrandt, & Makris, 1996; Caplan 2006; Caplan et al. 2007). First, split-half reliability estimates reveal higher correlations across sentence types than within sentence types.
Linguistic-specific hypotheses predict the opposite pattern, with scores on affected sentence types showing the strongest correlations. Second, factor analysis of the performance
of groups of individuals with aphasia reveals that most of the variance is explained by a
single factor on which all sentence types load. Linguistic-specific hypotheses, in contrast,
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
Sentence Comprehension in Aphasia
725
predict that only affected sentences should load on the first factor. Third, ANOVA
reveals an interaction between the overall level of patient performance (reflecting resource
availability) and sentence type (reflecting resource demands). In addition, an individual’s
aphasic syndrome (e.g., Broca’s aphasia) does not seem to predict any particular pattern of
sentence comprehension deficit (Caplan et al. 2007). Specific linguistic hypotheses do not
predict strong task effects, yet performance on different sentence comprehension tasks
may vary substantially within individuals (Cupples and Inglis 1993; Caplan et al. 2007;
Salis and Edwards 2009). Both the DDH and the TDH imply that performance on certain structures falls to chance levels because subjects are incapable of computing certain
structures and therefore guess. However, evidence from eye-tracking (Dickey and
Thompson 2009) and from self-paced reading (Caplan et al. 2007) reveals that aphasic
individuals exhibit evidence of normal processing on the occasions when they answer
correctly. This suggests that aphasia disrupts language processing without completely
blocking specific linguistic computations, and that a component of randomness or
stochastic noise may be necessary to explain variability in performance within individuals.
A further line of evidence arises from studies in which non-brain damaged individuals are
made to perform language comprehension or grammaticality judgment tasks with spectrally and temporally degraded auditory input. The performance of these subjects on a
range of sentence structures is quantitatively similar to the performance of aphasic subjects
with the same (undegraded) sentence material (Wilson et al. 2003; Dick et al. 2003).
THE TREE PRUNING HYPOTHESIS AND SYNTACTIC HIERARCHIES
Neither the TDH nor the DDH can account for other observations of agrammatic comprehension performance that can be parsimoniously addressed by considering the hierarchical
constituent structure of syntactic trees. Patients with agrammatic comprehension generally
perform worse on sentences with non-canonical word order than on sentences with canonical word order, but this dichotomy accounts for only a fraction of the variability seen among
sentence structures. For example, Friedmann (2006) points out that (across several studies),
object-relative sentences are significantly more difficult to comprehend than passive voice
sentences. This distinction can be easily explained by taking into account the fact that
object-relative sentences are more deeply embedded than passives [although number of
propositions might have explanatory value as well, see Caplan et al. (1998)and Caplan et al.
(1997)]. Thus, if the syntactic tree is ‘pruned’ at the level of the tensed or agreement phrase
(TP or AgrP), patients will fail to build a representation that contains the relative clause,
which requires projection to the complementizer phrase (CP, located above the inflectional
phrases in the tree – see Figure 1). Such observations regarding the complexity of syntactic
trees have also been fruitful for research on agrammatic language production, indicating a
potential common computational substrate for the two processes.
A computational model along these lines (Clark 2009 – see Figure 2) is proposed as an
explanation of the performance of a sample of aphasic individuals on twelve sentence
structures, as reported in the literature (Caplan et al. 1996). This approach is consistent
with a resource reduction approach, and makes use of a component of random noise,
such that the model sometimes produces accurate output even when ‘lesioned.’ The
model is a modification of freely available software that accompanies an introductory
computational semantics textbook (Blackburn and Bos 2006). The modifications include
(i) extension of the grammar to include sentences relevant to the study of aphasia, and (ii)
incorporation of a variable representing aphasia severity. This severity variable is used to
compute the probability of successfully parsing any given node in the tree by the formula
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
726 David G. Clark
Fig 1. The chief maximal projections of the syntactic tree, including CP (complementizer phrase), AgrP (agreement
phrase), TP (tense phrase), and VP (verb phrase). An arc is shown to represent the level of ‘pruning’ identified by
proponents of the Tree Pruning Hypothesis (TPH).
p ¼ 100 hs;
where h represents the height of the node in the syntactic tree, s represents the aphasia severity coefficient, and p represents the percent chance of successfully parsing the node (i.e.,
including it in the phrase structure representation). At each step in the parse, a random number between zero and 100 is generated and compared to the value p. The parse moves forward only if the step is grammatically legal and the random number is less than p. Thus,
constituents that are lower in the syntactic tree are more likely to be parsed and semantically
interpreted than constituents higher in the tree. The model generates output in the form of
sentences from first-order logic, assembled via expressions from lambda calculus.
There is an additional stipulation that was inspired by Friederici and Gorrell’s Structural
Prominence Hypothesis (SPH, Friederici and Gorrell 1998). These authors point out that
aphasic individuals have an increased tendency to assign the thematic role of agent to the
highest NP in the syntactic tree. This is implemented in the model as a heuristic method
for deriving the model’s ‘behavior’ (i.e., a decision of how it would perform in an enactment task) based on the formal logic strings it produces. Specifically, when a transitive
verb interpretation occurs with only one argument, the argument is assigned the thematic
role of patient. This is, essentially, the contrapositive of the SPH, since it nearly always
results in the assignment of the patient role to the lowest NP in a tree structure when
the model fails to generate a complete parse.
When the model is run 20 times at each of 21 levels of aphasia severity (setting s to
the values 0–20), its cumulative performance is strongly correlated with that of aphasic
patients on the same structures (r = 0.85, 72% of the variance). The model does not
directly address the question of whether the deficit of agrammatism is primarily syntactic
or involves the mapping from syntactic representations to semantic ones, and the effect of
number of propositions is dealt with only implicitly.
Conclusion
Current evidence points to three avenues by which our understanding of syntax in aphasia – and thus our understanding of language in the brain – may be advanced. The first is
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
Sentence Comprehension in Aphasia
727
Fig 2. A parse of the active sentence ‘A man chases a dog’ as described in Clark (2009). Each node in the syntactic
tree contains (1) an interpretation expressed in first-order logic and lambda calculus, as described in Blackburn and
Bos (2006), (2) a syntactic constituent label, (3) a height, which equals the number of steps from the node to the
most distant of its constituent words, and (4) a probability (p), determined by the equation 100)hs, where h is the
height of the node and s (representing aphasia severity) is set to 3. The cumulative probability of correctly producing the entire tree is equal to the product of all the probabilities for all the nodes (approximately 80%).
Det = determiner, N = noun, V = verb, NP = noun phrase, VP = verb phrase, S = sentence, $ = ‘there exists’ (from
first-order logic), @ signifies the application of a lambda expression to its argument. Lambda calculus is a deceptively simple computational system in which a variable bound by a lambda operator (k) can be replaced by any
argument. For example, application of the expression (kn. n · n) to any number yields the square of the number:
(kn. n · n)@2 = 2 · 2 = 4.
to focus on certain findings from psycholinguistic studies of normal subjects. The example
most relevant to this review is the finding that heuristic processes, which may be semantically driven, appear to play an important role in normal language processing (Ferreira and
Patson 2007; van Herten et al. 2006). The nature and neural basis for such heuristic processes should receive more attention.
The second important consideration is to create models of agrammatic comprehension
that generate explicit semantic representations from syntactic structure. It is possible that
much of the variation in patterns of agrammatism can be explained by taking into
account compositional semantics (and its reliance on hierarchical structure), along with
heuristic mechanisms. Faulty interpretive mechanisms, rather than parsing mechanisms,
may underlie the phenomena observed in agrammatic comprehension (Linebarger et al.
1983). Therefore, falsifiable models that make these processes more explicit will aid in
dissociating the contributions of each process to the empirical observations. However,
identification of a plausible representation for semantic structures is problematic. One
approach to characterizing the semantics of words is to classify them according to their
co-occurrences with other words (see, for example, Landauer and Dumais 1997). However, this approach does not offer an unambiguous method for the composition of
semantic structures associated with sentences. In addition, semantic representations
derived from word co-occurrences are sometimes insufficient for explaining the behavior
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
728 David G. Clark
of human subjects in settings that require consideration of the perceptual knowledge of
objects (Glenberg and Robertson 2000). This observation supports the contention that
symbols are grounded in perceptual experience (Barsalou 1999). Future work in aphasiology should consider the degree to which semantic structures rest on embodied cognition
(see Mahon and Caramazza (2008) for a more nuanced view).
The third consideration for advancing our understanding of sentence comprehension
pertains to the ‘resource’ that appears to be reduced in the setting of brain damage. While
current evidence supports this conjecture, there is little consensus regarding what the
resource actually is, although a reasonable hypothesis is that it is some form of working
memory (Caplan and Waters 1999; Haarmann et al. 1997). A great deal of effort has
already been applied to using neuroscientific methods as a means of anchoring psychological observations to more physically measurable phenomena. These methods include
structural and functional brain imaging, transcranial magnetic stimulation, electro- and
magnetoencephalography, and multi-cellular recordings in behaving animals. From the
standpoint of disturbed language comprehension, two fruitful avenues for investigation
would be better neural characterization of (i) the semantic representations onto which
words and sentences are mapped, and (ii) the events underlying sentence parsing and
interpretation. The first of these challenges might provide necessary scaffolding for the
second, i.e., the mapping from sound to meaning might be more transparent if we have a
realistic model of the brain’s representation of meaning. Better understanding of the sentence processing resource that is reduced by brain damage should be present, at least
implicitly, in any theory that answers these challenges.
Short Biography
David Clark’s research blends cognitive neuroscience with clinical neurology. He was
drawn into neurology by an interest in aphasia and through this has pursued a more comprehensive understanding of language, linguistics, and computation. Current projects
include a study relating event-related magnetic fields to brain structure and a comparison
of semantic memory measures to blood flow measures for predicting functional decline in
patients with memory disorders. Before coming to the University of Alabama at Birmingham and the Birmingham VA Medical Center, Clark undertook a behavioral neurology
fellowship at the VA Greater Los Angeles and UCLA. He was trained in neurology at
Wake Forest University, in Winston-Salem, North Carolina, USA.
Note
* Correspondence address: David Glenn Clark, Birmingham VA Medical Center and University of Alabama at
Birmingham, 1720 7th Avenue S, SC 620C, Birmingham, AL 35293, USA. E-mail: [email protected]
Works Cited
Bak, T., D. O’Donovan, J. Xuereb, S. Boniface, and J. Hodges. 2001. Selective impairment of verb processing
associated with pathological changes in Brodmann areas 44 and 45 in the motor neurone disease-dementia-aphasia syndrome. Brain 124. 103–20.
Barsalou, L. 1999. Perceptual symbol systems. Behavioral and Brain Systems 22. 577–660.
Bassett, D., and E. Bullmore. 2006. Small-world brain networks. Neuroscientist 12(6). 512–23.
Beretta, A., and A. Munn. 1998. Double-agents and trace-deletion in agrammatism. Brain and Language 65. 404–
21.
——, M. Pinango, J. Patterson, and C. Harford. 1999. Recruiting comparative crosslinguistic evidence to address
competing accounts of agrammatic aphasia. Brain and Language 67. 149–68.
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
Sentence Comprehension in Aphasia
729
Berndt, R., C. Mitchum, and S. Wayland. 1997. Patterns of sentence comprehension in aphasia: a consideration of
three hypotheses. Brain and Language 60. 197–221.
Blackburn, P., and J. Bos. 2006. Representation and inference for natural language. A first course in computational
semantics. Stanford, CA: CSLI Publications.
Caplan, D. 2006. Aphasic deficits in syntactic processing. Cortex 42. 797–804.
——, N. Alpert, and G. Waters. 1998. Effects of syntactic structure and propositional number on patterns of
regional cerebral blood flow. Journal of Cognitive Neuroscience 10(4). 541–52.
——, G. DeDe, and H. Brownell. 2006. Effects of syntactic features on sentence-picture matching in Broca’s
aphasics: a reply to Drai and Grodzinsky (2005). Brain and Language 96(2). 157–70.
——, N. Hildebrandt, and N. Makris. 1996. Location of lesions in stroke patients with deficits in syntactic processing in sentence comprehension. Brain 119. 933–49.
——, and G. Waters. 1999. Verbal working memory and sentence comprehension. Behavioral and Brain Sciences
22. 77–126.
——, ——, G. DeDe, J. Michaud, and A. Reddy. 2007. A study of syntactic processing in aphasia I: behavioral
(psycholinguistic) aspects. Brain and Language 101. 103–50.
——, ——, and N. Hildebrandt. 1997. Determinants of sentence comprehension in aphasic patients in sentencepicture matching tasks. Journal of Speech, Language, and Hearing Research 40. 542–55.
Caramazza, A., R. Capasso, E. Capitani, and G. Miceli. 2005. Patterns of comprehension performance in agrammatic Broca’s aphasia: a test of the Trace Deletion Hypothesis. Brain and Language 94(1). 43–53.
——, and E. Zurif. 1976. Dissociation of algorithmic and heuristic processes in language comprehension: evidence
from aphasia. Brain and Language 3. 572–82.
Chklovskii, D., and A. Koulakov. 2004. Maps in the brain: what can we learn from them? Annual Review of
Neuroscience 27. 369–92.
Chomsky, N. 1957. Syntactic structures. The Hague: Mouton de Gruyter.
——. 1959. Verbal Behavior (review). Language 35(1). 26–58.
Clark, D. 2009. A computational semantics approach to aphasic sentence comprehension. Aphasiology 23(1).
33–51.
Cupples, L., and A. Inglis. 1993. When task demands induce ‘asyntactic’ comprehension: a study of sentence interpretation in aphasia. Cognitive Neuropsychology 10(3). 201–34.
Damasio, A. 1992. Aphasia. New England Journal of Medicine 326(8). 531–9.
Dick, F., E. Bates, and E. Ferstl. 2003. Spectral and temporal degradation of speech as a simulation of morphosyntactic deficits in English and German. Brain and Language 85. 535–42.
Dickey, M., and C. Thompson. 2009. Automatic processing of wh- and NP-movement in agrammatic aphasia: evidence from eyetracking. Journal of Neurolinguistics 22(6). 563–83.
Drai, D., and Y. Grodzinsky. 2006. A new empirical angle on the variability debate: quantitative neurosyntactic
analyses of a large data set from Broca’s aphasia. Brain and Language 96(2). 117–28.
Ferreira, F. 2003. The misinterpretation of noncanonical sentences. Cognitive Psychology 47. 164–203.
——, K. Bailey, and V. Ferraro. 2002. Good-enough representations in language comprehension. Current Directions in Psychological Science 11. 11–5.
——, and N. Patson. 2007. The good enough approach to language comprehension. Language and Linguistics
Compass 1. 71–83.
Friederici, A., and P. Gorrell. 1998. Structural prominence and agrammatic theta-role assignment: a reconsideration
of linear strategies. Brain and Language 65. 253–75.
Friedmann, N. 2006. Generalisations on variations in comprehension and production: a further source of variation
and a possible account. Brain and Language 96. 151–3.
Glenberg, A., and D. Robertson. 2000. Symbol grounding and meaning: a comparison of high-dimensional and
embodied theories of meaning. Journal of Memory and Language 43. 379–401.
Goodglass, H. 1993. Understanding aphasia. San Diego, CA: Academic Press.
Grodzinsky, Y. 1990. Theoretical perspectives on language deficits. Cambridge, MS: MIT Press.
——. 1995a. A restrictive theory of agrammatic comprehension. Brain and Language 50. 27–51.
——. 1995b. Trace deletion, theta-roles, and cognitive strategies. Brain and Language 51. 469–97.
——. 2000a. Overarching agrammatism. Language and the brain: representation and processing – studies presented
to Edgar Zurif on his 60th birthday, ed. by Y. Grodzinsky, L. Shapiro and D. Swinney, 73–86. San Diego, CA:
Academic Press.
——. 2000b. The neurology of syntax: language use without Broca’s area. Behavioral and Brain Sciences 23. 1–71.
Haarmann, H., M. Just, and P. Carpenter. 1997. Aphasic sentence comprehension as a resource deficit: a computational approach. Brain and Language 59. 76–120.
van Herten, M., D. Chwilla, and H. Kolk. 2006. When heuristics clash with parsing routines: ERP evidence for
conflict monitoring in sentence perception. Journal of Cognitive Neuroscience 18(7). 1181–97.
Hickok, G., E. Zurif, and E. Canseco-Gonzalez. 1993. Structural description of agrammatic comprehension. Brain
and Language 45(3). 371–95.
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd
730 David G. Clark
Hillis, A., S. Oh, and L. Ken. 2004. Deterioration of naming nouns versus verbs in primary progressive aphasia.
Annals of Neurology 55. 268–75.
——, E. Tuffiash, and A. Caramazza. 2002. Modality-specific deterioration in naming verbs in nonfluent primary
progressive aphasia. Journal of Cognitive Neuroscience 14(7). 1099–108.
Landauer, T., and S. Dumais. 1997. A solution to Plato’s problem: the latent semantic analysis theory of acquisition,
induction, and representation of lexical knowledge. Psychological Review 104. 211–40.
Linebarger, M., M. Schwartz, and E. Saffran. 1983. Sensitivity to grammatical structure in so-called agrammatic
aphasics. Cognition 13. 361–92.
Mahon, B., and A. Caramazza. 2008. A critical look at the embodied cognition hypothesis and a new proposal for
grounding conceptual content. Journal of Physiology – Paris 102. 59–70.
Mauner, G., V. Fromkin, and T. Cornell. 1993. Comprehension and acceptability judgments in agrammatism:
disruptions in the syntax of referential dependency. Brain and Language 45. 340–70.
Mendez, M., and D. Clark. 2008. Neuropsychiatric aspects of aphasia and related disorders. In S. Yudofsky and
R. Hales, (eds). The American Psychiatric Publishing Textbook of Neuropsychiatry and Behavioral Neurosciences. 5th edition. Washington, DC: American Psychiatric Publishing. 519–34.
Mesulam, M. 1990. Large-scale neurocognitive networks and distributed processing for attention, language, and
memory. Annals of Neurology 28. 597–613.
Salis, C., and S. Edwards. 2009. Tests of syntactic comprehension in aphasia: an investigation of tas effects. Aphasiology 23(10). 1215–30.
Su, Y., S. Lee, and Y. Chung. 2007. Asyntactic thematic role assignment by Mandarin aphasics: a test of the tracedeletion hypothesis and the double dependency hypothesis. Brain and Language 101. 1–18.
Townsend, D., and T. Bever. 2001. Sentence comprehension: the integration of habits and rules. Cambridge, MA:
MIT Press.
Watts, D., and S. Strogatz. 1998. Collective dynamics of ‘small-world’ networks. Nature 393. 440–2.
Wilson, S., and A. Saygin. 2004. Grammaticality judgment in aphasia: deficits are not specific to syntactic structures,
aphasic syndromes, or lesion sites. Journal of Cognitive Neuroscience 16(2). 238–52.
——, ——, E. Schleicher, F. Dick, and E. Bates. 2003. Grammaticality judgment under non-optimal processing
conditions: deficits induced in normal participants resemble those observed in aphasic patients. Brain and
Language 87. 67–8.
ª 2011 The Author
Language and Linguistics Compass 5/10 (2011): 718–730, 10.1111/j.1749-818x.2011.00309.x
Language and Linguistics Compass ª 2011 Blackwell Publishing Ltd