Theoretical analysis of word production deficits in adult aphasia

Downloaded from on December 29, 2014
Theoretical analysis of word production
deficits in adult aphasia
Myrna F. Schwartz
Moss Rehabilitation Research Institute, Elkins Park, PA, USA
Cite this article: Schwartz MF. 2014
Theoretical analysis of word production deficits
in adult aphasia. Phil. Trans. R. Soc. B 369:
One contribution of 17 to a Discussion Meeting
Issue ‘Language in developmental and
acquired disorders: converging evidence
for models of language representation
in the brain’.
Subject Areas:
behaviour, cognition, neuroscience
aphasia, naming, repetition, errors, semantic,
Author for correspondence:
Myrna F. Schwartz
e-mail: [email protected]
The cognitive analysis of adult language disorders continues to draw heavily
on linguistic theory, but increasingly it reflects the influence of connectionist,
spreading activation models of cognition. In the area of spoken word production, ‘localist’ connectionist models represent a natural evolution from
the psycholingistic theories of earlier decades. By contrast, the parallel distributed processing framework forces more radical rethinking of aphasic
impairments. This paper exemplifies these multiple influences in contemporary cognitive aphasiology. Topics include (i) what aphasia reveals about
semantic-phonological interaction in lexical access; (ii) controversies surrounding the interpretation of semantic errors and (iii) a computational
account of the relationship between naming and word repetition in aphasia.
Several of these topics have been addressed using case series methods,
including computational simulation of the individual, quantitative error patterns of diverse groups of patients and analysis of brain lesions that
correlate with error rates and patterns. Efforts to map the lesion correlates
of nonword errors in naming and repetition highlight the involvement of
sensorimotor areas in the brain and suggest the need to better integrate
models of word production with models of speech and action.
1. Introduction
Aphasia in adults is caused by stroke, trauma or degenerative pathology that
compromises brain networks for language. Arguably, the most pervasive
symptom of aphasia, within and across etiologies, is the inability to produce
known words in a timely and accurate manner. The word production deficit
is readily detected by tests of picture naming, where it manifests in hesitations,
word-finding gaps and/or commission errors. The focus of this paper is the
model-driven analysis of naming errors in patients and what this reveals
about the functional architecture of word production.
Recent decades have witnessed a change in both the models and methods of
cognitive aphasiology. In place of the traditional box-and-arrow diagrams, contemporary models feature connectionist networks, with units corresponding to
localist or distributed representations that are primed and retrieved through the
mechanism of spreading activation. Computer-implemented models are necessarily highly specific in their representational and processing commitments and
so have taken on an important role in theorizing about aphasia, as they have in
normal language research.
The research designs used to collect aphasia data have also changed. The
single-subject approach has evolved into case series methods, in which multiple
individuals are studied on the same set of tasks with the goal of understanding
why the patients differ from one another. In small case series studies, each individual can be studied intensively with multiple different tasks, in the manner of
single case studies. Large case series studies trade depth of assessment for the
opportunity to test a larger, more diverse, and sometimes more representative,
sample of patients. The objective is then to explain the behavior of interest by
analysing patterns of covariation within and across tasks (see [1]; and ensuing
commentary in Cognitive Neuropsychology, 28/7, 2011). Data generated with case
series methods have been simulated computationally. They have also been used
in conjunction with advanced lesion-mapping methods to localize the lesions
& 2013 The Author(s) Published by the Royal Society. All rights reserved.
Downloaded from on December 29, 2014
that explain the pattern of symptom variation. We will see
examples in the sections that follow.
Table 1. Taxonomy of error types (from [13]).
2. The two stages of word production
apple ! ‘worm’).
Mixed—Real word response that qualifies as a semantic error
and that meets the criterion for phonological similaritya
(e.g. cow ! ‘cat’; skirt ! ‘shirt’).
Formal—Real word response that meets the criterion for phonological
similaritya (e.g. dog ! ‘dock’; pirate ! ‘tire’).
Unrelated—Real word response that is neither semantically nor
phonologically similar to the target (e.g. banana ! ‘camp’;
well ! ‘car’).
Nonword—String of phonemes that does not constitute a word in the
language (e.g. camel ! /kæ[email protected] r/ (‘kah-mer’); piano ! /pInæno/
(‘pih-nah-no’). Most such errors pass the phonological similarity
criteriona, and, depending on the study, this can be a requirement
for inclusion in the category.
Response and target must share at least one phoneme in corresponding
syllable or word position or two phonemes in any position, not counting
unstressed vowels.
whose error pattern was indicative of a problem at or
before lemma selection (i.e. the patient’s errors were predominantly of the semantic type). That a stage-1 production
problem was ameliorated by a variable that operates at
stage 2 is strong evidence for interaction across the stages.
As noted earlier, computational modelling has come to
play an increasingly important role in aphasia theorizing. In
studies that simulate individual patients’ error patterns
with implemented models, the best fits have been obtained
with models that incorporate interaction. Importantly,
though, the most successful of these models contain features
designed to set limits on the interactive flow of information
[11]. The next section describes in detail one such model
and the evidence that supports it.
3. The interactive two-step model of
lexical access
This model, developed by Dell and co-workers [13,22], has
been used to fit data from patients in numerous studies of
word production deficits [13,23 –30]. Researchers interested
in applying the model to new data can access the automated
data-fitting program at
The model contains a three-level lexical network consisting of semantic features, words, and phonemes (figure 1)
and weighted connections that transmit activation both topdown and bottom-up. All activation is positive; there is no
inhibition. All words are CVC syllables, with phonemes
marked for syllable position. This is a model of a stable
system after learning has been completed: connection weights
are comparable and fixed for all words and all phonemes and
the model does not represent variations in lexical frequency
Phil. Trans. R. Soc. B 369: 20120390
During the 1970s, linguistics-inspired analysis of speech error
corpora produced a theory of sentence production that
greatly influenced research in aphasia [2]. An enduring
legacy of that theory is the idea that words are retrieved in
two, discrete, serial stages, one concerned with selecting a
word or ‘lemma’ [3] from the mental lexicon, the second
with attaching form to the selected word. The first stage is
responsive to meaning and grammar; the second to phonological structure and content [4–8].
The two-stage assumption has been implemented in localist connectionist models, where nodes at one level of the
network stand in one-to-one correspondence with words.
Within this class of models, some accept the discreteness of
the stages [9], whereas others postulate an interactive flow
of information between them [7,10]. They also differ in how
many levels of representation are proposed and how they
are characterized (for review [11,12]).
Along with experimental and speech error evidence from
normal speakers, data from aphasia have contributed to the
development and evaluation of competing models. In aphasia, error production tends to be high even in single-word
production tasks like picture naming, and this has enabled
researchers to use error data from these simple, controlled
paradigms to address theoretical issues. A case in point is
the debate about interactivity.
Table 1 shows some of the error types that researchers
study. Error classification is invariably theory-driven, so the
categories and definitions of error types vary somewhat
from study to study. The definitions in the table are those
that my colleagues and have used in the case series studies
referenced in this article.
Most individuals with aphasia produce a variety of these
error types in their naming performance. However, much interest has centered on the occasional selective presentation, e.g.
where one patient produces mostly phonological errors
(formal errors and nonwords) and another, of comparable
severity, produces mostly semantic errors. Double dissociations
like these might be viewed as supporting the two-stage theory
and, more specifically, the proposal that the semantic and
phonological stages are discrete, i.e. non-interacting.
On the other hand, in studies of naming errors aggregated
from multiple patients the interactive account has received
support. The evidence here is that mixed errors and formal
errors, both of which involve substitution of a word phonologically related to the target, occur at significantly higher
frequencies than the discreteness account predicts [14–16].
This indicates that phonological information does not operate
exclusively at the phonological stage of production; rather, it
is accessed prior to lemma selection and plays a role in which
lemma gets selected.
A recent study adds to the evidence for interaction by
showing that a phonological variable, phonological neighbourhood density, can bias lemma selection in favour of the
correct target [17]. This study corroborated previous evidence
that naming accuracy is higher for words with many than
few phonological neighbours [18–21]. Its novel contribution
was to show this neighbourhood density effect in a patient
target (e.g. pig ! ‘sheep’; church ! ‘building’; shoe ! ‘Nike’;
Semantic—Real word response that is a synonym, category
coordinate, superordinate, subordinate or strong associate of the
Downloaded from on December 29, 2014
semantic features
Vo Co
Figure 1. Structure of the interactive two-step model. (Adapted from [23] with permission from Elsevier.)
(but see [30]). A naming trial begins with external activation
supplied to the target’s semantic features. Activation flows
freely for a designated number of time steps, after which
the most activated word is selected. This completes step-1.
Step-2 begins with external activation supplied to the selected word, designed to simulate grammatically triggered
phonological encoding. This activation propagates down to
the phonemes of the selected word and back up from each
phoneme to all words in the lexicon that contain that phoneme in that syllable position. Activation reverberates
throughout the network until, after a fixed number of time
steps, the most activated segments from the onset, vowel
and coda segment clusters are selected, terminating step-2
and the trial, and defining the model’s response.
As noted, activation flows bidirectionally during each of
the two steps of lexical access. However, the design of the
model was influenced by experimental evidence showing
that early in a word retrieval episode, semantic influences
dominate, whereas later in the episode, phonological influences dominate [31]. In keeping with this, the model’s
default activation parameters were set so that step-1 selection
is primarily influenced by top-down activation from semantics. Phonological feedback during step-1 has its primary
impact on the cohort that competes for word selection,
where it privileges candidates that share the target’s phonemes. This effect of feedback explains the above-chance
incidence of mixed and formal errors, noted above.
Now consider step-2: in a model where the reciprocal feedback between semantics and phonology fully integrates these
two information sources, the phonemes of words in the semantic neighbourhood of the target would be expected to enter into
the competition for selection at step-2, along with those of the
target. The interactive two-step model exhibits little competition of this kind, for the reason that the external activation
supplied to the selected word at the start of step-2 boosts its
segments over those of other words in the semantic cohort
[11]. The weakness of competition for phoneme selection is
key to why having many phonological neighbours helps
rather than hinders phonological access [19,20].
By design, then, the interactive two-step model operates
in a way that is globally modular (semantic and phonological
information sources are largely restricted to step-1 and step-2,
respectively) but locally interactive (feedback from later to earlier levels influences the competition for selection) [22]. The
model has been tested for its ability to simulate individual
differences in the naming response patterns of diverse groups
with aphasia [13,23,32]. The naming response pattern is the
proportional breakdown of total naming responses into correct
responses and the five error types shown in table 1. According
to the model, semantic, mixed, formal and unrelated errors are
the product of incorrect word selection at step-1. Nonword
errors result from faulty segmental selection at step-2.
Here is how the simulations worked: the model was first
implemented to account for the normal naming pattern
(mostly correct with a small number of semantic errors),
and then assumptions were made about how aphasia alters
the normal system [13,23,32]. The semantic-phonological
account [32] proposes that aphasia alters the weight of connections in the lexical network, weakening the flow of
activation between semantics and words (s-weight lesion)
or between words and segments (p-weight lesion), or both.
In a computational case series study with 94 patients,
Schwartz et al. [23] found that the semantic-phonological
account explained 94.5% of the variance in the individual
naming response patterns.
As discussed by Schwartz et al. [23], the key reason why
the semantic-phonological account is able to explain diverse,
individual naming response patterns in aphasia is that it
allows errors generated at step-1 (lexical errors) and those
generated at step-2 (sublexical errors) to dissociate. My colleagues and I extended the evidence for the distinctiveness
of these error generators in a series of studies that mapped
the neural correlates of particular error types and modeldriven weight parameters [33– 37]. The key finding, for
present purposes, is that the neural correlates of semantic errors (the prototypical step-1 error type) and nonword
errors (step-2) were different and non-overlapping. As these
lesion-mapping studies will be mentioned again later, some
details are in order.
The method we used is called voxel-based lesion-symptom
mapping (VLSM) [38]. Each of our studies started with case
series behavioral data, in which the N’s ranged between 64
and 106. After tracing each patient’s lesion on the high-quality
structural brain scan obtained under a research protocol,
Phil. Trans. R. Soc. B 369: 20120390
Downloaded from on December 29, 2014
(e) ABCD
4. Alternative accounts of semantic errors
The interactive two-step model of word retrieval has engendered considerable debate. One aspect of the debate centres
on the model’s assumptions about interactivity. Rapp &
Goldrick [11] proposed an alternative model, the restricted
interaction account (RIA). While RIA, too, implements the
two-step assumption in a localist connectionist architecture, it
differs from Dell’s model in a few key respects, among which
are that feedback from phonology reverberates between
the phonological and lexical levels but does not extend to the
semantic level. Rapp and Goldrick compared the computational fit of these two models, and some others, with the
individual data from three contrasting aphasia cases and
found that RIA produced the best fits. Additional arguments
and evidence about the interaction assumptions in these two
models can be found in the earlier studies [27,28,39].
The Dell and Schwartz model-fitting studies have been criticized for their single task focus. The concern is that by failing
to consider patients’ performance on tasks other than oral
naming, we missed the opportunity to independently verify
the model’s account of where the deficit resides and, more
importantly, to obtain evidence that might falsify model
assumptions [39]. This is a valid criticism. For example, the
model assumes that all semantic errors arise from weak
s-weights and so does not recognize the possibility that semantic-level processing itself could be impaired. Yet, further study
of patients who produce semantic errors in naming has shown
that some do have semantic-level deficits (e.g. confusion
between semantically similar items in comprehension);
others, however, do not [11,40–44]. This indicates that the
model is flawed in its assumption that all semantic errors are
errors of lexical selection. RIA [11] potentially escapes this
problem by allowing for damage (in the form of noisy activation) to the model’s semantic level, in addition to its lexical
and phonological levels.
In a more radical departure from Dell’s model, it has been
proposed that all semantic errors in production are the consequence of damage at the semantic level. The motivation for
this is the parallel distributed processing (PDP) account of
language production, which rejects the two-step assumption
altogether. In the PDP framework, there is nothing in the
structure of the system that corresponds to individual
words. What appear to be lexical effects are properties that
emerge from the functional interplay between semantics
and phonology, which are fully interactive [45 –48].
Semantics and phonology, in this account, are ‘primary
systems’, each comprised representations that were learned
in childhood under pressure to concurrently produce, comprehend and imitate language [46,49]. Having been shaped
in this way, these representations are amodal in the sense
that they support all manner of language tasks and modalities in adulthood. It follows, then, that patients’ accuracy
and error patterns in naming might be predictable from
their performance on other tasks.1 This prediction was
tested by Lambon Ralph et al. [47] in a paper entitled,
‘Anomia is simply a reflection of semantic and phonological
impairments’. It reported case series data from 21 patients
with wide-ranging anomia, including evidence that the
variance in accuracy in naming was predicted by a linear
combination of composite semantic and phonological
Phil. Trans. R. Soc. B 369: 20120390
Figure 2. Results of VLSM of naming errors in chronic, post-stroke aphasia. Voxels in which lesion status correlated significantly with semantic errors are shown in
purple. Those correlated with phonological (nonword) errors are shown in blue. In each analysis, the critical threshold was based on the correction for multiple
comparisons (false discovery rate, q ¼ 0.05). There was no overlap in the two lesion maps. (Adapted from [36] with permission from Oxford University Press.)
we used specialized software to transform the traced lesions
onto a standard brain template. That made it possible to superimpose the lesions and statistically analyse areas of overlap. In
each voxel that was lesioned in some minimum number of
patients, the software measured the effect size for the association between the presence/absence of a lesion and the
magnitude of the dependent variable (e.g. number of semantic
errors). A threshold for significance was set that controls for the
many thousands of voxels tested, and the results were
displayed in colour-coded brain maps like that shown in
figure 2, where the area rendered in purple is the identified
locus for semantic errors, and the area in blue is the locus for
nonword errors. As the two lesion maps do not overlap, we
conclude that separate brain regions are implicated in the
genesis of these two error types.
Downloaded from on December 29, 2014
Competing models of word production are challenged
to explain other types of production tasks in addition to
naming, most notably word repetition. Naming and repetition deficits frequently co-occur in aphasia, but they can also
dissociate. Some have used such evidence as a means to
distinguish phonological deficits that originate within the
lexicon from those that originate in a post-lexical phonological/phonetic stage of processing. One proposal is that an
impairment in lexical-phonological processing will impair
naming, but not word repetition, because repetition can be
performed without lexical mediation, as is the case when one
Phil. Trans. R. Soc. B 369: 20120390
5. The interactive two-step account of word
repeats a nonword. On this view, comparable performance
in naming and word repetition is evidence of a post-lexical
phonological deficit [12]. Space precludes further consideration of the lexical/post-lexical distinction (see [61] for an
interesting perspective). However, the relationship between
production mechanisms in repetition and naming will be
considered in some detail.
Most researchers would agree that while a phonological deficit of some sort lies at the heart of patients’ difficulties with
repetition, lexical and semantic factors also play a role, even at
the level of single words [62,63]. This motivates the effort to
explain repetition within a word production model that
includes semantic, lexical and phonological representations,
and interactive activation between them. The rest of this section
describes efforts to apply the interactive two-step model to the
task of single-word repetition. For an account of repetition
that adopts the primary systems view, see Ueno et al [55].
Dell and co-workers [24,30,64] carried out a series of computational studies of repetition in which they modified
the basic naming model to instantiate different accounts
of the relationship between naming and repetition, and then
fit the alternative models to word repetition data from individual subjects or case series data. As a general feature, the
implemented models included separate units and connections
for phonological input and output processing ([65]; see [66]
for a different approach), and the input processes were assumed
to be intact, both in the model and in the patients whose data
were simulated.2
Two repetition models were compared. Both commit to the
proposition that word repetition involves retrieval of lexicalphonological representations via the second step of the lexical
access model. The lexical-only account of repetition proposes
that this is the sole mechanism by which the adult speaker
repeats words. By contrast, the summation dual-route account,
which was inspired by Hillis & Caramazza [67], proposes
that phonological access in repetition is accomplished jointly
by the lexical route and a second, non-lexical route. Figure 3
presents schematic diagrams of these two repetition models
along with the naming model.
The lexical-only model makes the strong prediction that
repetition can be accurately predicted from naming performance at the individual level. Simulating a patients’ repetition
performance under the lexical-only model started by setting
the naming model’s s- and p-weights in accordance with that
individual’s naming response pattern. Each simulated repetition trial began with external activation supplied to the
target lemma, implementing successful word recognition. As
in naming, the activation was allowed to reverberate among
the phonological, lexical and semantic units until eventually
the most active phoneme units were chosen to represent the
repetition response. Errors and accuracy were tabulated across
trials to generate the predicted outcome for the patient.
To implement the summation dual-route account, a nonlexical route was grafted on to the interactive two-step model
[24]. This route consisted of a single input node outside the
model, which connected to the model’s output phonemes via
connections whose weights represent the strength of the nonlexical route (figure 3). This parameter (‘nl’, for non-lexical)
was estimated separately for each patient according to how
accurately the patient performed a nonword repetition test.
To run the summation dual-route model for a particular patient,
the s-weight, p-weight and nl parameters were set based on
the patient’s naming and nonword repetition, respectively.
measures (60– 70% of the variance explained). Moreover,
examination of the bivariate correlations between composite
scores and error types showed that consistent with expectations, greater semantic deficit was associated with more
semantic errors and omissions, whereas greater phonological
deficit was associated with more nonwords.
In recent years, the debate between the two-step and PDP
camps has moved into the cognitive neuroscience arena.
Details are beyond the scope of this paper, but we can see
some of what is at issue by looking back at figure 2. Recall
that the area mapped in purple is where lesion status correlated with semantic errors in our VLSM of naming errors.
It includes an area in the left temporal lobe that stretches
from the mid-part of the middle temporal gyrus forward to
the temporal pole. This area sits within the anterior temporal
lobe (ATL), a region made famous by the ‘distributedplus-hub’ theory of semantic memory [50]. The ATL is
the hypothesized hub, proposed as the storage site for the
amodal semantic representations that support language in
multiple modalities [49]. While there are semantic hubs
in both hemispheres, the one on the left is thought to play
an especially strong role in production by virtue of its
proximity to left-lateralized phonological areas [51].
At first blush, our finding that semantic errors in naming
localize to the ATL would appear to accord well with this primary systems account. However, there is an important wrinkle
to the story. The VLSM of semantic errors statistically controlled for errors that might have arisen at the semantic level
by regressing out scores on comprehension measures. As my
colleagues and I have argued [33,35], the evidence therefore
favours the two-step account over the primary systems account
by identifying a brain locus for production only semantic errors,
exactly as would be expected if semantic errors arose in the process of accessing lemmas from semantics. Within the wider
cognitive neuroscience literature, support can be found for
both these theoretical positions [52–57].
Referring again to figure 2, we see that the lesion map for
semantic errors ( purple area) also includes a region in the
left prefrontal cortex. It is possible that lesions to this area
generate semantic errors by compromising cognitive control
mechanisms involved in language production. There is substantial evidence that damage to left ventrolateral prefrontal
cortex impacts aphasics’ ability to resolve competition for
semantic and lexical selection under experimental conditions
that exaggerate competition [58 –60]. Our prefrontal finding
may indicate that this competition-resolving mechanism
plays a critical role even in a simple naming task.
Downloaded from on December 29, 2014
output phonology
output phonology
NL node
lexical-route repetition
output phonology
summation dual-route
Figure 3. Relationship between the picture naming model (top) and the two repetition models described in the text: the lexical-only model (left) and the
summation dual-route model (right). (Adapted from [68] with permission from Elsevier.)
obtained repetition
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
predicted repetition (SP model)
Figure 4. Predicted and obtained proportion correct repetition for 59 research
participants with aphasia whose repetition was fit with the interactive twostep model of repetition (lexical-only model). The solid line represents perfect
prediction and the dotted lines represent boundaries where the deviation
between predicted and obtained is 0.20. (Adapted from [64] with permission
from Elsevier.)
On each trial, external activation was applied to the external
input node, triggering non-lexical-route processing, and also
to the target lexical node, triggering lexical-route processing.
After a designated time period, the activation in the output
units from both routes was summed to produce the response.
Once again, errors and accuracy were tabulated across trials
to arrive at model predictions, and these were compared with
the actual word repetition scores of the patient.
Figure 4, from [64], shows that the lexical-only model did
quite well in predicting word repetition accuracy. For most of
the 59 patients studied, the deviation between predicted and
obtained proportions was less than 0.20 (dotted lines in the
figure). However, there were seven patients, represented as
points above the upper dotted line, whose repetition was
better than the lexical-only model predicted. These patients
were more accurately fit by the summation dual-route
model. However, for the majority of patients, the summation dual-route model predicted higher accuracy than was
actually observed (see also [24]).
Nozari et al. [30] took a different approach to comparing
the lexical-only route and summation dual-route repetition
models. Their study was based on the well-founded assumption that the frequency-sensitivity of word repetition can be
used to gauge the involvement of the lexical route. The
study is complicated, and so I will limit my discussion to
the major findings, which relate to the incidence and
frequency-sensitivity of nonword errors in word repetition,
compared with picture naming.
The investigators first analysed the naming and repetition
responses of 59 diverse aphasic speakers using hierarchical, multinomial logistic regression analyses. This revealed
that patients produced significantly fewer nonword errors
in repetition than naming, but the frequency-sensitivity
of nonword errors was as great in repetition as it was in
naming. The latter is a striking finding, and strong indication
that the lexical route is routinely used in repeating words. In
the computational portion of the study, Nozari et al. adjusted
lexical connection weights to implement frequency differences, and then ran the two repetition models, along with
the naming model, to compare frequency-sensitivity across
the models. The counterintuitive finding here was that the
two repetition models yielded frequency effects that were
equally strong and comparable to the naming model. The frequency effect in the dual-route repetition model was not
diminished by the presence of the non-lexical route, because
Phil. Trans. R. Soc. B 369: 20120390
Downloaded from on December 29, 2014
3.49 3.50
3.96 3.97
Figure 5. VLSM analyses. (a) Maps voxels in which lesion status correlated significantly with nonword errors in the object naming task (this corresponds to the blue
map in figure 2). (b) Maps voxels that correlated with nonword errors in the word repetition task, after covarying out errors generated in nonword repetition (see
text for explanation). In both maps, voxels that exceeded the threshold (false discovery rate correction, q ¼ 0.01) are rendered in a red-to-yellow scale. Those below
threshold are rendered in green (closer to threshold)-to-blue. The same participants (n ¼ 103) participated in both studies.
effects of the two routes were additive (summed). This meant
that the two routes combined without any loss; when the
non-lexical route was added, the lexical route contributed
as much as when the non-lexical route was absent.
Whereas the involvement of the non-lexical route did not
diminish the size of the model’s frequency effect, it did
reduce the model’s production of nonword errors and, consequently, made its repetition more accurate. With the summation
dual-route model, as with the patient data, nonword errors
were less numerous in repetition, compared with naming, but
these errors were equally sensitive to frequency across tasks.
In this study, then, the summation dual-route model provided
the best account of the overall findings. At the same time, in
summarizing findings for individual patients in this and
previous comparisons of the repetition models, Nozari et al.
emphasize that there are important individual differences in
whether and to what degree the non-lexical route is added
when repeating words. What accounts for these individual
differences? Clues can be found in a recent meta-analysis that
showed that patients who repeat by the lexical-only route
tend to have better comprehension, whereas dual-routers
have better phonological working memory [68].
6. Brain localization for nonword errors in
naming and repetition
A central claim of the interactive two-step model of lexical
access is that impairment in the step-2 mapping between lexical and phonological units is a cause of nonword errors in
both naming and word repetition. VLSM offered interesting
possibilities for testing this claim.
Earlier, I mentioned the VLSM analysis we ran to identify
the neural correlate of nonword errors in picture naming
[36]. The results of that analysis, which figure 2 displayed
in blue, are reproduced in figure 5a using a more informative
colour scale. The accompanying figure 5b shows the results
of a VLSM of nonword errors in word repetition, based on
same group of 103 patients.3 Importantly, for the latter analysis, scores on nonword repetition were regressed out to
control for possible impairments in non-lexical-route processing (input or output) that might have influenced the word
repetition data. This amounted to statistically isolating the
contribution of lexical-route damage to production of nonword errors in word repetition.
Although the errors mapped in the two VLSM analyses
derive from different tasks, according to the model they share
a common functional locus and so should converge on the
same region of brain. Confirming this prediction, the maps in
figure 5a,b are strikingly similar. In both cases, the significant
or nearly significant voxels localized to the anterior parietal
lobe (supramarginal and post-central gyri), extending forward
into pre-central and pre-motor cortices.
These fronto-parietal regions have been implicated in
many prior neuroantomical studies of phonological processing [69,70], including a large lesion-mapping study of
phonological errors in acute aphasia [71]. On the other
hand, in neuroimaging research aimed at identifying the
neural substrate of lexical-phonological retrieval, the area
most commonly identified is located more posteriorly, in
temporal and temporo-parietal cortices inclusive of Wernicke’s area [53,72,73]. It was unexpected that the VLSM of
nonword errors would fail to pick out this temporal lobe
region while identifying strong effects in fronto-parietal
regions most often associated with the planning and regulation of speech and action.
What lessons should we draw from this? Perhaps, we
should be looking to theoretical accounts of language production that are more grounded in sensorimotor processes,
for example, the theory of articulatory phonology that postulates that units that enter into phonemic speech errors are
actually units of articulation [74]; see also [75]. Certainly,
we need a better understanding of the mechanisms that transform static lexical representations into a temporally extended,
ordered sequence of articulatory gestures. The PDP framework seems to offer an interesting way forward [46,55].
Also, there may be useful convergence with research looking
at how the brain solves the serial-ordering problem in other,
related domains, such as sign-language production and the
gestures that accompany spoken speech [76].
7. Conclusion
This paper has presented a selected review of research on
word production deficits in adult aphasia, emphasizing the
Phil. Trans. R. Soc. B 369: 20120390
Downloaded from on December 29, 2014
Task differences are also important to the primary systems account.
Production/naming is considered the most vulnerable to semantic
damage, because it rests solely on the mapping from semantics to
phonology. By contrast, other tasks provide a second, non-semantic
avenue for activating phonology (e.g., orthographics-to-phonology
mapping in reading; acoustic-to-phonology mapping in repetition).
Input processing was tested in the patients who contributed to
these studies, and those who did not achieve high accuracy on phoneme discrimination and other auditory input tasks were treated
The data were collected for a published study [37] that mapped
voxels that correlated with model parameters derived from patients’
repetition performance (i.e., p-weight, nl weight parameters). Here, in
a re-analysis of the repetition data, we analysed the lesion correlates
of the errors themselves.
Schwartz MF, Dell GS. 2010 Case series
investigations in cognitive neuropsychology. Cogn.
Neuropsychol. 27, 477– 494. (doi:10.1080/
Saffran EM. 1982 Neuropsychological approaches to
the study of language. Br. J. Psychol. 73, 317–337.
Kempen G, Huijbers P. 1983 The lexicalisation
process in sentence production and naming: indirect
election of words. Cognition 14, 185–209. (doi:10.
Butterworth B. 1989 Lexical access in speech
production. In Lexical representation and process
(ed. W Marslen-Wilson), pp. 108–135. Cambridge,
MA: MIT Press.
Garrett MF. 1975 The analysis of sentence
production. In The psychology of learning and
motivation (ed. GH Bower), pp. 133 –175. London,
UK: Academic Press.
Garrett MF. 1980 Levels of processing in
sentence production. In Language production
(ed. B Butterworth), pp. 177–220. London, UK:
Academic Press.
Dell GS. 1986 A spreading-activation theory of
retrieval in sentence production. Psychol. Rev. 93,
283–321. (doi:10.1037/0033-295X.93.3.283)
Levelt WJM. 1989 Speaking: from intention to
articulation. Cambridge, MA: MIT Press.
Levelt WJM, Roelofs A, Meyer AS. 1999 A theory of
lexical access in speech production. Behav. Brain.
Sci. 22, 1– 75. (doi:10.1017/S0140525X99001776)
Harley TA. 1984 A critique of top-down independent
levels models of speech production: evidence from
non-plan-internal speech errors. Cogn. Sci. 8,
191–219. (doi:10.1207/s15516709cog0803_1)
Rapp B, Goldrick M. 2000 Discreteness and interactivity
in spoken word production. Psychol. Rev. 107,
460–499. (doi:10.1037/0033-295X.107.3.460)
Goldrick M, Rapp B. 2007 Lexical and post-lexical
phonological representations in spoken production.
Cognition 102, 219 –260. (doi:10.1016/j.cognition.
Dell GS, Schwartz MF, Martin N, Saffran EM, Gagnon
DA. 1997 Lexical access in aphasic and nonaphasic
speakers. Psychol. Rev. 104, 801– 838. (doi:10.
Best WM. 1996 When racquets are baskets but
baskets are biscuits, where do the words come
from? A single-case study of formal paraphasia.
Cogn. Neuropsychol. 3, 369 –409.
Gagnon DA, Schwartz MF, Martin N, Dell GS, Saffran
EM. 1997 The origins of formal paraphasias in
aphasics’ picture naming. Brain Lang. 59, 450– 472.
Martin N, Gagnon DA, Schwartz MF, Dell GS, Saffran
EM. 1996 Phonological facilitation of semantic errors
in normal and aphasic speakers. Lang. Cogn.
Process. 11, 257–282. (doi:10.1080/0169096
Middleton EL, Schwartz MF. 2011 Density pervades:
an analysis of phonological neighbourhood density
effects in aphasic speakers with different types of
naming impairment. Cogn. Neuropsychol. 27,
401 –427. (doi:10.1080/02643294.2011.570325)
Gordon JK, Dell GS. 2001 Phonological
neighborhood effects: evidence from aphasia and
connectionist modeling. Brain Lang. 79, 21– 23.
Dell GS, Gordon JK. 2003 Neighbors in the lexicon:
friends or foes? In Phonetics and phonology in
language comprehension and production: differences
and similarities (eds NO Schiller, AS Meyer),
pp. 9–37. Berlin, Germany: Mouton de Greyter.
Chen Q, Mirman D. 2012 Competition and
cooperation among similar representations: toward
a unified account of facilitative and inhibitory effects
of lexical neighbors. Psychol. Rev. 119, 417–439.
Goldrick M, Folk JR, Rapp B. 2010 Malaprop’s
neighborhood: using word errors to reveal
neighborhood structure. J. Mem. Lang. 62,
113 –134. (doi:10.1016/j.jml.2009.11.008)
Dell GS, O’Seaghdha PG. 1991 Mediated and
convergent lexical priming in language production:
a comment on Levelt et al (1991). Psychol. Rev. 98,
604 –614. (doi:10.1037/0033-295X.98.4.604)
Schwartz MF, Dell GS, Martin N, Gahl S, Sobel P.
2006 A case-series test of the interactive two-step
model of lexical access: evidence from picture
naming. J. Mem. Lang. 54, 228–264. (doi:10.1016/
Hanley JR, Dell GS, Kay J, Baron R. 2004 Evidence
for the involvement of a nonlexical route in the
repetition of familiar words: a comparison of single
and dual route models of auditory repetition. Cogn.
Neuropsychol. 21, 147–158. (doi:10.1080/
Croot K, Patterson K, Hodges JR. 1998 Single word
production in nonfluent progressive aphasia. Brain
Lang. 61, 226–279. (doi:10.1006/brln.1997.1852)
Caramazza A, Papagno C, Ruml W. 2000 The
selective impairment of phonological processing in
speech production. Brain Lang. 75, 428 –450.
Ruml W, Caramazza A, Capasso R, Miceli G. 2005
Interactivity and continuity in normal and aphasic
language production. Cogn. Neuropsychol. 22,
131–168. (doi:10.1080/02643290442000031)
Ruml W, Caramazza A, Shelton JR, Chialant D. 2000
Testing assumptions in computational theories of
aphasia. J. Mem. Lang. 43, 217 –248. (doi:10.1006/
Schwartz MF, Brecher A. 2000 A model-driven
analysis of severity, response characteristics, and
partial recovery in aphasics’ picture naming. Brain
Lang. 73, 62 –91. (doi:10.1006/brln.2000.2310)
Nozari N, Kittredge AK, Dell GS, Schwartz MF. 2010
Naming and repetition in aphasia: steps, routes,
and frequency effects. J. Mem. Lang. 63, 541 –559.
Levelt WJM, Schriefers H, Vorberg D, Meyer AS,
Pechmann T, Havinga J. 1991 The time course of
lexical access in speech production: a study of
picture naming. Psychol. Rev. 98, 122–142. (doi:10.
Foygel D, Dell GS. 2000 Models of impaired lexical
access in speech production. J. Mem. Lang. 43,
182–216. (doi:10.1006/jmla.2000.2716)
Schwartz MF, Kimberg DY, Walker GM, Faseyitan O,
Brecher A, Dell GS, Coslett HB. 2009 Anterior
temporal involvement in semantic word retrieval:
VLSM evidence from aphasia. Brain 132,
3411– 3427. (doi:10.1093/brain/awp284)
Phil. Trans. R. Soc. B 369: 20120390
Funding statement. This review was prepared with support from grant
no. RO1DC000191-31 awarded to M.F.S. by the National Institutes
of Health/National Institute on Deafness and Other Communication
interactive two-step model of lexical access and some of
the patient studies it has inspired. I hope to have conveyed
something of the progress and promise of contemporary aphasiology, with its convergence of cognitive, computational and
neuroimaging methods. A next frontier, I believe, is to better
understand why the impact of lexical and phonological
access deficits is sometimes greatly magnified when the task
calls for production of multiword sequences [77,78]. New cognitive models of phonological and grammatical sequencing
[79–81] may point the way forward.
Downloaded from on December 29, 2014
61. Romani C, Galluzzi C. 2005 Effects of syllabic
complexity in predicting accuracy of repetition and
direction of errors in patients with articulatory and
phonological difficulties. Cogn. Neuropsychol. 22,
817–850. (doi:10.1080/02643290442000365)
62. Hanley JR, Edwards M, Kay J. 2002 Imageability
effects and phonological errors: implications for
models of auditory repetition. Cogn. Neuropsychol.
19, 193 –206. (doi:10.1080/02643290143000132)
63. Martin N, Saffran EM. 1997 Language and auditoryverbal short-term memory impairments: evidence
for common underlying processes. Cogn.
Neuropsychol. 14, 641–682. (doi:10.1080/
64. Dell GS, Martin N, Schwartz MF. 2007 A case-series
test of the interactive two-step model of lexical
access: predicting word repetition from picture
naming. J. Mem. Lang. 56, 490–520. (doi:10.1016/
65. Caramazza A. 1988 Some aspects of language
processing revealed through the analysis of acquired
aphasia: the lexical system. Annu. Rev. Neurosci. 11,
395–421. (doi:10.1146/
66. Martin N, Dell GS, Saffran EM, Schwartz MF. 1994
Origins of paraphasias in deep dysphasia: testing
the consequences of a decay impairment to an
interactive spreading activation model of lexical
retrieval. Brain Lang. 47, 609– 660. (doi:10.1006/
67. Hillis AE, Caramazza A. 1991 Mechanisms for
accessing lexical representations for output:
evidence from a category specific semantic deficit.
Brain Lang. 40, 106 –144. (doi:10.1016/0093934X(91)90119-L)
68. Nozari N, Dell GS. 2013 How damaged brains
repeat words: a computational approach. Brain
Lang.126, 327–337. (doi:10.1016/j.bandl.2013.
69. Vigneau M, Beaucousin V, Herve PY, Duffau H,
Crivello F, Houde O, Mazoyer B, Tzourio-Mazoyer N.
2006 Meta-analyzing left hemisphere language
areas: phonology, semantics, and sentence
processing. NeuroImage 30, 1414– 1432. (doi:10.
70. Rapcsak SZ, Beeson P, Henry ML, Leyden A, Kim E,
Rising K, Andersen S, Cho H. 2009 Phonological
dyslexia and dysgraphia: cognitive mechanisms and
neural substrates. Cortex 45, 575–591. (doi:10.
71. Cloutman L et al. 2009 Where (in the brain) do
semantic errors come from? Cortex 45, 641 –649.
72. Graves WW, Grabowski TJ, Mehta S, Gupta P. 2008
The left posterior superior temporal gyrus
participates specifically in accessing lexical
phonology. J. Cogn. Neurosci. 20, 1698–1710.
73. Wilson SM, Isenberg AL, Hickok G. 2009 Neural
correlates of word production stages delineated by
parametric modulation of psycholinguistic variables.
Hum. Brain Mapp. 30, 3596 –3608. (doi:10.1002/
Phil. Trans. R. Soc. B 369: 20120390
48. Wilshire CE. 2008 Cognitive neuropsychological
approaches to word production in aphasia: beyond
boxes and arrows. Aphasiology 22, 1019–1053.
49. Rogers TT, Lambon Ralph MA, Garrard P, Bozeat S,
McClelland JL, Hodges JR, Patterson K. 2004
Structure and deterioration of semantic memory: a
neuropsychological and computational investigation.
Psychol. Rev. 111, 205–235. (doi:10.1037/0033295X.111.1.205)
50. Patterson K, Nestor PJ, Rogers TT. 2007 Where do
you know what you know? The representation of
semantic knowledge in the human brain. Nat. Rev.
Neurosci. 8, 976 –987. (doi:10.1038/nrn2277)
51. Lambon Ralph MA, McClelland J, Patterson K,
Galton C, Hodges JR. 2001 No right to speak? The
relationship between object naming and semantic
impairment: neuropsychological evidence and a
computational model. J. Cogn. Neurosci. 13,
341 –356. (doi:10.1162/08989290151137395)
52. Damasio H, Grabowski TJ, Tranel D, Hichwa RD,
Damasio A. 1996 A neural basis for lexical retrieval.
Nature 380, 499–505. (doi:10.1038/380499a0)
53. Indefrey P, Levelt WJM. 2004 The spatial and
temporal signatures of word production
components. Cognition 92, 101–144. (doi:10.1016/
54. Mesulam M, Wieneke C, Hurley R, Rademaker A,
Thompson CK, Weintraub S, Rogalski EJ. 2013
Words and objects at the tip of the left temporal
lobe in primary progressive aphasia. Brain 136,
601 –618. (doi:10.1093/brain/aws336)
55. Ueno T, Saito S, Rogers TT, Lambon Ralph MA. 2011
Lichtheim 2: synthesising aphasia and the neural
basis of language in a neurocomputational model
of the dual dorsal-ventral language pathways.
Neuron 72, 385 –396. (doi:10.1016/j.neuron.2011.
56. Antonucci SM, Beeson PM, Labiner DM, Rapcsak SZ.
2008 Naming and semantic knowledge in patients
with left inferior temporal lobe lesions. Aphasiology
22, 281– 304. (doi:10.1080/02687030701294491)
57. Henry ML, Beeson PM, Alexander GE, Rapcsak SZ.
2012 Written language impairments in primary
progressive aphasia: a reflection of damage to
central semantic and phonological processes.
J. Cogn. Neurosci. 24, 261–275. (doi:10.1162/
58. Schnur TT, Schwartz MF, Brecher A, Hodgson C.
2006 Semantic interference during blocked-cyclic
naming: evidence from aphasia. J. Mem. Lang. 54,
199 –227. (doi:10.1016/j.jml.2005.10.002)
59. Schnur TT, Schwartz MF, Kimberg DY, Hirshorn E,
Coslett HB, Thompson-Schill SL. 2009 Localizing
interference during naming: convergent
neuroimaging and neuropsychological evidence
for the function of Broca’s area. Proc. Natl Acad.
Sci. USA 106, 322–327. (doi:10.1073/pnas.
60. Robinson G, Blair J, Cipolotti L. 1998 Dynamic
aphasia: an inability to select between competing
verbal responses? Brain 121, 77– 89. (doi:10.1093/
34. Schwartz MF, Kimberg DY, Walker GM, Brecher A,
Faseyitan O, Dell GS. 2011 Neuroanatomical
dissociation for taxonomic and thematic knowledge
in the human brain. Proc. Natl Acad. Sci. USA 108,
8520–8524. (doi:10.1073/pnas.1014935108)
35. Walker GM, Schwartz MF, Kimberg DY, Faseyitan O,
Brecher A, Dell GS, Coslett HB. 2011 Support for
anterior temporal involvement in semantic error
production in aphasia: new evidence from VLSM.
Brain Lang. 117, 110–122. (doi:10.1016/j.bandl.
36. Schwartz MF, Faseyitan O, Kim J, Coslett HB. 2012
The dorsal stream contribution to phonological
retrieval in object naming. Brain 135, 3799 –3814.
37. Dell GS, Schwartz MF, Nozari N, Faseyitan O, Coslett
HB. 2013 Voxel-based lesion-parameter mapping:
identifying the neural correlates of a computational
model of word production in aphasia. Cognition
128, 380–396. (doi:10.1016/j.cognition.2013.
38. Bates E, Wilson SM, Saygin AP, Dick F, Sereno MI,
Knight RT, Dronkers NF. 2003 Voxel-based lesionsymptom mapping. Nat. Neurosci. 6, 448–450.
39. Ruml W, Caramazza A. 2000 An evaluation of a
computational model of lexical access: comment on
Dell et al. (1997). Psychol. Rev. 107, 609 –634.
40. Rapp B, Benzing L, Caramazza A. 1997 The
autonomy of lexical orthography. Cogn.
Neuropsychol. 14, 71 –104. (doi:10.1080/
41. Hillis AE, Rapp BC, Romani D, Caramazza A. 1990
Selective impairment of semantics in lexical
processing. Cogn. Neuropsychol. 7, 191–243.
42. Caramazza A, Hillis AE. 1990 Where do semantic
errors come from? Cortex 26, 95– 122. (doi:10.
43. Gainotti G, Miceli G, Caltagirone C, Silveri MC,
Masullo C. 1981 The relationship between type of
naming error and semantic-lexical discrimination in
aphasic patients. Cortex 17, 401–410. (doi:10.1016/
44. Gainotti G, Silveri C, Villa G, Miceli G. 1984 Anomia
with and without lexical comprehension disorders.
Brain Lang. 29, 18– 33. (doi:10.1016/0093934X(86)90031-3)
45. Plaut DC. 1995 Double dissociation without
modularity: evidence from connectionist
neuropsychology. J. Clin. Exp. Neuropsychol. 17,
291–321. (doi:10.1080/01688639508405124)
46. Plaut DC, Kello CT. 1999 The emergence of
phonology from the interplay of speech
comprehension and production: a distributed
connectionist approach. In The emergence of
language (ed. B MacWhinney), pp. 381–415.
Mahwah, NJ: Lawrence Erlbaum Associates Inc.
47. Lambon Ralph MA, Moriarty L, Sage K. 2002
Anomia is simply a reflection of semantic and
phonological impairments: evidence from a caseseries study. Aphasiology 16, 56– 82. (doi:10.1080/
Downloaded from on December 29, 2014
77. Schwartz MF, Hodgson C. 2002 A new multiword
naming deficit: evidence and interpretation. Cogn.
Neuropsychol. 19, 263 –288. (doi:10.1080/
78. Thothathiri M, Schwartz MF, Thompson-Schill SL.
2010 Selection for position: the role of left
ventrolateral prefrontal cortex in sequencing
language. Brain Lang. 113, 28 –38. (doi:10.1016/j.
79. Gupta P, Dell GS. 1999 The emergence of language
from serial order and procedural memory. In The
emergence of language (ed. B MacWhinney),
pp. 447–481. Mahwah, NJ: Erlbaum.
80. Botvinick M, Plaut DC. 2004 Doing without schema
hierarchies: a recurrent connectionist approach to
normal and impaired routine sequential action.
Psychol. Rev. 111, 395– 429. (doi:10.1037/0033295X.111.2.395)
81. Dell GS, Chang F. 2014 The P-Chain: relating
sentence production and its disorders to
comprehension and acquisition. Phil. Trans. R. Soc. B
369, 20120394. (doi:10.1098/rstb.2012.0394)
74. Browman CP, Goldstein L. 1992 Articulatory
phonology: an overview. Haskins Lab. Status Rep.
Speech Res. 111/112, 23 – 42.
75. Hickok G. 2012 Computational neuroanatomy of
speech production. Nat. Rev. Neurosci. 13,
135–145. (doi:10.1038/nrg3118)
76. Marshall CR. 2014 Word production
errors in children with developmental
language impairments. Phil. Trans. R.
Soc. B 369, 20120389. (doi:10.1098/rstb.
Phil. Trans. R. Soc. B 369: 20120390