You are on page 1of 15

AJSLP

Research Article

Part of Speech Production in Patients


With Primary Progressive Aphasia:
An Analysis Based on Natural
Language Processing
Charalambos Themistocleous,a Kimberly Webster,b
Alexandros Afthinos,a and Kyrana Tsapkinia,c

Background: Primary progressive aphasia (PPA) is a tagging. We then implemented a twofold analysis: (a) linear
neurodegenerative disorder characterized by a progressive regression to determine how patients with nonfluent PPA,
decline of language functions. Its symptoms are grouped semantic PPA, and logopenic PPA variants differ in their
into three PPA variants: nonfluent PPA, logopenic PPA, POS productions and (b) a supervised classification analysis
and semantic PPA. Grammatical deficiencies differ depending based on POS using machine learning models (i.e., random
on the PPA variant. forests, decision trees, and support vector machines) to
Aims: This study aims to determine the differences between subtype PPA variants and generate feature importance (FI).
PPA variants with respect to part of speech (POS) production Outcome and Results: Using an automated analysis of a
and to identify morphological markers that classify PPA short picture description task, this study showed that content
variants using machine learning. By fulfilling these aims, the versus function words can distinguish patients with nonfluent
overarching goal is to provide objective measures that can PPA, semantic PPA, and logopenic PPA variants. Verbs
facilitate clinical diagnosis, evaluation, and prognosis. were less important as distinguishing features of patients
Method and Procedure: Connected speech productions with different PPA variants than earlier thought. Finally, the
from PPA patients produced in a picture description task study showed that among the most important distinguishing
were transcribed, and the POS class of each word was features of PPA variants were elaborative speech elements,
estimated using natural language processing, namely, POS such as adjectives and adverbs.

M
orphosyntactic deficits have been identified as & Mack, 2014). PPA is characterized by substantial vari-
one of the key symptoms of primary progres- ability of symptoms as an effect of the degree of neuro-
sive aphasia (PPA; Mesulam et al., 2014, 2012; degenerative decline, underlying pathology, and areas of
Mesulam & Weintraub, 2014), a neurodegenerative condi- brain damage (Mesulam, 2013; Thompson, Lukic, et al.,
tion that results in a considerable deterioration of speech 2012; Thompson & Mack, 2014). To understand the symp-
and language skills (Thompson, Lukic, et al., 2012; Thompson toms, recently established consensus criteria classify pa-
tients into three main PPA variants: the nonfluent PPA
variant (nfvPPA), the logopenic PPA variant (lvPPA), and
a
Department of Neurology, Johns Hopkins School of Medicine, the semantic PPA variant (svPPA; Gorno-Tempini et al.,
Baltimore, MD 2011; Gorno-Tempini & Pressman, 2016).
b
Department of Otolaryngology, Johns Hopkins Medicine, Morphosyntactic production is key for language
Baltimore MD communication, and as it becomes impaired in PPA, it can
c
Department of Cognitive Science, Johns Hopkins University, provide objective markers for the classification of patients
Baltimore MD with PPA into variants, for clinical evaluation, prognosis
Correspondence to Charalambos Themistocleous: cthemis1@jhu.edu of the condition, and intervention. Patients with nfvPPA
Editor-in-Chief: Melissa Duff are characterized by sparsity of function words and ab-
Editor: Dirk B. Den Ouden normal syntax due to peak atrophy at the posterior inferior
Received September 16, 2019 frontal gyrus (Broca’s area; Gorno-Tempini et al., 2011;
Revision received February 14, 2020 Thompson et al., 1997, 2013; Thompson, Cho, et al., 2012).
Accepted May 13, 2020
https://doi.org/10.1044/2020_AJSLP-19-00114
Publisher Note: This article is part of the Special Issue: Select Papers Disclosure: The authors have declared that no competing interests existed at the time
From the 49th Clinical Aphasiology Conference. of publication.

466 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021 • Copyright © 2020 American Speech-Language-Hearing Association

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Patients with lvPPA are characterized by deficits in single- speech productions is labor intensive, as it requires the
word retrieval and naming. Repetition of long words and identification of the part of speech (POS) for every word
phrases is also impaired, and their speech is characterized that appears in the transcriptions of connected speech.
by phonological errors, such as paraphrases. Atrophy in Our study offers a less labor-intensive automatic analy-
individuals with lvPPA is related to brain volume loss in sis of POS from connected speech productions and em-
the left posterior temporal and inferior parietal regions ploys the analysis to classify patients into lvPPA, svPPA,
(Gorno-Tempini et al., 2008, 2011; Gorno-Tempini & and nfvPPA. Automatic methods were employed success-
Pressman, 2016). Patients with svPPA are characterized fully in earlier research for Alzheimer’s disease diagnosis,
by impaired object knowledge, impaired confrontation classification, and so forth (Bucks et al., 2000; Fraser
naming, and limited single-word comprehension. They et al., 2018; Guinn & Habash, 2012; Orimaye et al., 2014;
are also characterized by impaired semantic memory of Rentoumi et al., 2014; Themistocleous et al., 2018; Thomas
familiar objects, “empty speech” in verbal production, et al., 2005) and in the automatic identification of patients with
and surface dyslexia, as an effect of damage at the anterior svPPA, nfvPPA, and healthy controls (Fraser et al., 2014).
temporal lobe (Gorno-Tempini et al., 2011; Gorno-Tempini The aims of this study were (a) to quantify the pro-
& Pressman, 2016; Wilson et al., 2010). duction of POS in patients with different PPA variants,
Morphosyntactic production in PPA has been ana- (b) to determine whether and which POS markers can clas-
lyzed qualitatively (Kertesz et al., 1998; Mesulam, 2001) sify patients into PPA variants automatically, and (c) to
and quantitatively (Thompson, 2012; Thompson et al., provide a methodological approach that can be employed
1997, 2014; Wilson, Brandt, et al., 2014; Wilson, DeMarco, in clinical settings to enable clinicians to quantify mor-
et al., 2014; Wilson et al., 2010). Several early studies phosyntactic productions in patients with nfvPPA, lvPPA,
have shown that patients with nfvPPA produce propor- and svPPA in a timely manner. Our overarching goal is to
tionally fewer function (also known as grammatical) words identify objective morphosyntactic measures that can be
than content words from patients with lvPPA and svPPA employed in clinical diagnosis, evaluation, and prognosis.
(Thompson, Cho, et al., 2012; Thompson & Mack, 2014; We employed samples of connected speech from picture
Thompson et al., 2013; Wilson et al., 2010). Function description tasks from all three PPA variants and auto-
words are words with grammatical meaning and syntactic matically extracted measures of POS production. Subse-
purpose, such as conjunctions and articles. In contrast, quently, we employed computational morphosyntactic
content words have semantic meaning, such as nouns and analysis, namely, morphosyntactic tagging, to identify the
verbs. However, studies, such as Fraser et al. (2014), did morphosyntactic categories automatically.
not find a pronounced deficit in function words in individ-
uals with nfvPPA in comparison to controls and patients
with svPPA. Method
In addition to a specific deficit in content versus func-
tion words, patients with nfvPPA have been characterized Participants
by selective impairment in naming verbs relative to naming Recordings of the Cookie Theft picture description
nouns (Thompson, Lukic, et al., 2012). The opposite of task from 52 patients diagnosed with PPA were analyzed
this pattern was found to exist in patients with svPPA, who (Goodglass et al., 2001). Diagnosis of PPA was provided
are characterized by selective impairment in naming nouns by neurologist(s) according to consensus criteria (Gorno-
relative to verbs (Cappa et al., 1998; Daniele et al., 1994; Tempini et al., 2011). Variant subtyping was based on
Hillis et al., 2004; Kertesz et al., 1998; Thompson, Lukic, magnetic resonance imaging results, clinical and neuropsy-
et al., 2012). Overall, patients with PPA display substantial chological examination, and speech and language evalua-
heterogeneity with respect to content word production, espe- tions following consensus criteria. Nineteen participants were
cially nouns and verbs. For example, Thompson, Cho, et al. subtyped as nfvPPA, 18 were subtyped as lvPPA, and 15
(2012) did not find significant differences between patients were subtyped as svPPA. Unclassified patients with PPA and
with nfvPPA, patients with lvPPA, and healthy controls in apraxia of speech–only cases were excluded (see Table 1).
noun-to-verb ratio. Only patients with svPPA displayed dif- Data were collected as part of a clinical trial conducted
ferences in noun-to-verb ratio from both patients with at Johns Hopkins University (NCT02606422; Tsapkini
nfvPPA and controls (Thompson, Cho, et al., 2012). et al., 2018). Inclusion criteria comprised progressive speech
Previous studies examined connected speech produc- and language deficits, native English speakers, minimum of
tions in PPA using picture description tasks (see Faroqi- high school education, and normal or corrected vision.
Shah et al., 2020; Wilson et al., 2010). Connected speech Exclusion criteria comprised absence of any developmental
reveals a wide array of naturally occurring morphosyn- or nondegenerative neurological disorder (e.g., stroke)
tactic content and function word productions, such as the and absence of primary deficits in other cognitive domains,
production of adjectives, adverbs, pronouns, and conjunc- such as memory. The present data are from baseline neuro-
tions, and can determine the distribution of content words psychological and neurolinguistic evaluations.
and function words across PPA variants (Ash et al., 2006; One-way analysis of variance tests showed that there
Graham et al., 2004; Kavé et al., 2007; Meteyard & Patterson, were no significant differences between patients with nfvPPA,
2009). Nevertheless, the manual analysis of connected lvPPA, and svPPA for gender (F = 5.82, df = 2, p =.09), age

Themistocleous et al.: Part of Speech Production in PPA 467


Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Table 1. Demographic and neuropsychological data for each participant group.

Variable nfvPPA lvPPA svPPA

Demographics
No. of participants 19 18 15
Gender 9F 10M 8F 10M 7F 8M
Age 69 (6) 68 (8) 67 (6)
Education 16 (1) 17 (2) 16 (2)
Severity
Language severity 2.77 (0.5) 1.39 (0.8) 2.27 (0.6)
Total severity 10 (6.5) 11 (6.3) 12 (5.5)
Language production
Boston Naming Test (30) 17 (5.6) 14 (6.5) 13 (11.1)
Category fluency (fruits, animals, vegetables) 22.9 (10.1) 20.8 (14.3) 7.8 (6.3)
Learning and memory
Rey Auditory Verbal Learning Test 33 (14) 19 (12) 18 (8)

Note. Recorded materials and speakers’ demographic information. Shown are the number of participants,
gender, age, education, language severity, and total severity. The numbers in parentheses indicate the standard
deviation. PPA = primary progressive aphasia; nfvPPA = nonfluent PPA variant; lvPPA = logopenic PPA variant;
svPPA = semantic PPA variant; F = female; M = male.

(F = 0.354, df = 2, p = .705), education (F = 0.162, df = 2, The automatic transcriptions were compared word by word
p = .853), language severity (F = 0.154, df = 2, p =.86), with the transcripts of the two independent evaluators and
and total severity (F = 1.162, df = 2, p = .33; see Table 1). marked as true (“T”) or false (“F”) for each word based
Participants’ age, education, and language severity were on whether the word was correctly transcribed by the soft-
considered in data analysis (Riello et al., 2018). Language ware (T) or whether the software provided a different tran-
severity in Table 1 is a five-level rating subscale in the Clin- scription or no transcription (F). Using Cohen’s kappa,
ical Dementia Rating–Modified Frontotemporal Lobar the agreement of the automatic transcription with the first
Degeneration Scale (Knopman et al., 2008). This scale was transcriber was .95 ( p < .0001), and the agreement with
found effective in representing the progression of language the second transcriber was .96 ( p < .0001); the agreement
deficits due to neurodegeneration, distinguishing patients between the two transcribers was .97 ( p < .0001). After
based on a corresponding language severity outcome measure establishing that transcriptions were close to human tran-
(Knopman et al., 2008). The Clinical Dementia Rating– scriptions, no further transcription modifications were
Modified Frontotemporal Lobar Degeneration Scale showed employed as we were interested to see how well the tran-
that participants in this study were matched for overall lan- scriptions perform without manual modifications. Most
guage severity. fillers (e.g., “um” and “uh”) were transcribed (and ana-
lyzed) but were not considered word utterances, that is,
they were not included in the total word count. Repetitions
Materials and false starts were transcribed in Roman alphabet; repe-
The Cookie Theft picture description task was part titions of words were included in the total word count.
of the oral evaluation (Goodglass et al., 2001). This word- Neologisms were transcribed using standard orthography
less picture shows three people performing different activi- in Roman alphabet. Table 2 shows the mean number of
ties: two children trying to steal cookies from a cupboard words per variant. Individuals with nfvPPA, lvPPA, and
as their mother is washing the dishes. A clinician presented svPPA produced different numbers of words; individuals
the picture to the participant and prompted the partici-
pant following the standard Boston Diagnostic Aphasia
Examination–Third Edition (Goodglass et al., 2001) instruc- Table 2. Number of words and mean number of words per participant
in each variant in Cookie Theft picture description task.
tions (“Tell me everything you see going on in this picture”).
The patient is instructed to speak in complete sentences,
Materials produced nfvPPA lvPPA svPPA
making reference to the objects, people, and activities
shown. No prompts or cues are provided. Speech produc- Total no. of words 1,816 4,121 3,126
tions were audio-recorded with a digital audio recorder. Mean words 121 (85)*,a 217 (165) 165 (97)
The audio recordings were all converted into 16000 Hz
Note. PPA = primary progressive aphasia; nfvPPA = nonfluent
mono format and transcribed using speech-to-text technol- PPA variant; lvPPA = logopenic PPA variant; svPPA = semantic
ogy implemented in an in-house software platform, Themis PPA variant.
(Themistocleous et al., 2018; Themistocleous & Kokkinakis, a
Significantly different from the logopenic primary progressive
2018). Twenty percent of the transcriptions were selected aphasia variant. *p < .05.
randomly and evaluated by two independent evaluators.

468 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
with nfvPPA produced fewer words than patients with Statistical Analysis
svPPA and lvPPA.
To determine whether patients with nfvPPA, svPPA,
and lvPPA differ in the production of POS, generalized
linear models were conducted with the PPA variants as pre-
Natural Language Processing
dictors and the POS measures as response variables (see also
and Morphosyntactic Measures Lien et al., 2017). Estimated marginal means (also known
The transcripts were analyzed morphosyntactically as least squares means) provided contrasts of the three
using the NLTK Python library (Bird et al., 2009). Morpho- PPA variants. The statistical analysis was performed in
syntactic analysis involved the following steps: (a) identi- R (R Core Team, 2016), using lmer4 (Bates et al., 2015),
fication of words and (b) sentence structure analysis and lmerTest (Kuznetsova et al., 2017), and emmeans (Russell,
POS identification in the text (for an overview of these 2018) packages in R.
methods, see Jurafsky & Martin, 2009). From the tagged
corpus, both content words and function words were
counted, and their corresponding proportions of the main Machine Learning Classification
POS categories (POS category/number of word count)
To determine whether variants differed with respect
were estimated. In addition to proportions, the ratios of
to morphosyntactic measures and identify the morpho-
all POS classes were calculated. Specifically, we calculated
syntactic measures that enable variant classification, a
the following POS measures:
machine learning classification task was conducted. Specif-
1. Content words, namely, the proportion of nouns, ically, a binary one-against-all classification was employed,
verbs, adjectives, and adverbs in the text; content where the model distinguishes one variant from the other
words have a specific lexical meaning as they refer two variants: nfvPPA versus (lvPPA and svPPA), svPPA
to objects, actions, and so forth. versus (nfvPPA and lvPPA), and lvPPA versus (nfvPPA
2. Function words, namely, the proportion of conjunc- and svPPA). Three machine learning models were employed:
tions (e.g., and, or, but); prepositions (e.g., in, of ); decision trees (DTs), random forests (RFs; Breiman,
determiners and predeterminers (e.g., the, a/an, both); 2001; Breiman et al., 1984), and support vector machines
pronouns such as he, she, and it and wh– pronouns (SVMs; Cortes & Vapnik, 1995). DTs split the data in
such as what, who, and whom; modal verbs (e.g., a binary manner and provide a classification output rep-
can, should, will); possessive ending (’s); adverbial resented as a dendrogram (tree diagram). The main dis-
particles (e.g., about, off, up); and conjunction (e.g., advantage of DTs is that they often tend to create very
infinitival to, as in to do). large and complicated trees that cannot generalize to un-
known data. RFs aim to address DTs’ disadvantage by
3. The corresponding ratio of nouns and verbs with
producing several tree models that are subsequently av-
other POS: noun/verb ratio, noun/adjective ratio, noun/
eraged. SVMs employ hyperspaces, that is, demarcating
adverb ratio, noun/pronoun ratio, noun/preposition
planes, to distinguish different groups, using both linear
ratio, noun/conjunction ratio, verb/adjective ratio,
and nonlinear kernels and often perform better than DTs
verb/adverb ratio, verb/pronoun ratio, verb/preposi-
and RFs.
tion ratio, verb/conjunction ratio, adjective/adverb
For the classification task, predictors were first indi-
ratio, and content word/function word ratio.
vidually scaled, such that they ranged between zero and
The analysis of ratios identifies the morphosyntactic one. To find the best performing models during evaluation,
selectivity in the speech, as ratios can determine specific several models were tested using multiple cross-validations
preferences in the use of one POS category over another from 5-, 10- (…) to 35-folds (Hastie et al., 2009). We opti-
POS category. For example, the reduction of attributes mized the number of trees in RF, the size of the kernels
assigned to nouns (e.g., “the red car” vs. “the car”) can in SVM models, and the initial random state for DTs. The
be expressed by the noun/adjective ratio. The verb/adverb optimization process showed that the output of the 35 cross-
ratio provides information about the use of adverbs with validation provided better results overall.
respect to verbs in PPA productions “(she) reads” versus Machine learning models were trained and evaluated
“(she) reads fast,” and the adverb/adjective ratio provides 35 times using a different set of 34 folds as a training set
information about the selective use of adverbs with respect and a different single fold as an evaluation set. During the
to adjectives. As the overall data size was small, POS ra- evaluation phase, the machine learning models provided
tios provide further information that can improve the a prediction of the PPA variant of each patient. After a
automatic classification. The automatic POS tagging was trained model predicts the variant of a new and unknown
compared to the manual tagging provided by three raters (to the model) patient, the prediction of the model is com-
with background in linguistics. Cohen’s kappa shows that pared to the actual PPA variant from the neurological
there was substantial agreement between the automatic evaluation to estimate whether the model made a correct
morphological analysis and the manual one provided from prediction. From the comparison of the predicted variant
Rater 1 (κ = .983, p < .0001), Rater 2 (κ = .985, p < .0001), and the actual variant, the following cases are estimated:
and Rater 3 (κ = .983, p < .0001). (a) the true positive (TP) cases (i.e., the cases when patients

Themistocleous et al.: Part of Speech Production in PPA 469


Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
belong to variant X [e.g., nfvPPA] and the model predicts Distribution of Content Words and Function Words
correctly that they belong to that variant X), (b) the false
Content Versus Function Word Production
negative (FN) cases (the cases that the model predicts that
In nfvPPA, the proportion of content words is higher
patients do not belong to variant X, when in actuality their
than the proportion of function words (see Figures 1A and
variant is X), (c) the false positive (FP) cases (the cases
1B and Table 3). Patients with nfvPPA also produce a sig-
that the model predicts that patients belong to variant X,
nificantly higher content word/function word ratio com-
when in actuality they do not belong to variant X), and
pared to patients with svPPA and lvPPA, β = 1.720, t(47) =
(d) the true negative (TN) cases (i.e., when the model pre-
13.97, p < .0001 (see Figure 2C and Table 3). Estimated
dicts that a patient does not belong to variant X correctly).
marginal means (also known as least squares means) for
These cases are employed to calculate the following evalua-
pairwise differences for the content word/function word
tion metrics:
ratio show significant differences for nfvPPA and lvPPA
1. Precision (Precision = TP/[TP + FP]): The precision ( p < .05), and nfvPPA and svPPA ( p < .05), but there is
is a measure of the true positives divided by the sum no significant contrast between lvPPA and svPPA. These
of true positives and FPs; if there are many FPs, the findings demonstrate that content word/function word ra-
precision is lower. tio distinguishes patients with nfvPPA from svPPA and
2. Recall (Recall = TP/[TP + FN]): The recall is a mea- lvPPA.
sure of the true positives divided by the sum of true
positives and FNs, that is, a low recall indicates that Noun Production
there are many FNs. Patients with nfvPPA produce proportionally more
nouns than patients with lvPPA and svPPA (see Figure 2).
3. Accuracy and balanced accuracy (Accuracy = Consequently, patients with PPA differ significantly in
[TP + TN]/[TP + TN + FP + FN]): The accuracy is the proportion of nouns, F(2, 47) = 15.7, p < .01. Both
a measure of the correct predictions made by the
lvPPA (t = −4.06, p < .01) and svPPA (t = −5.30, p < .01)
model divided by the total number of all cases. The
are statistically different from nfvPPA. However, pairwise
balanced accuracy is the average of recall obtained
comparisons reveal that patients with lvPPA and svPPA
on each class. The accuracy is a number from 0 to
do not differ in the proportion of nouns (nonsignificant).
100; an accuracy of 50 means that the model is cor-
rect half of the times, and a 100% accuracy means
Verb Production
that the model is always correct.
Patients with PPA do not significantly differ in verb
4. The F1 score (F1 score = 2 × [(Precision × Recall)/ production (nfvPPA, M = 0.24, SD = 0.06; lvPPA, M = 0.26,
(Precision + Recall)]): The F1 score measures the SD = 0.6; svPPA, M = 0.28, SD = 0.05). The comparison
weighted average of Precision and Recall. It is a score of verb production in patients with nfvPPA and svPPA
from 0 to 100 that is similar to accuracy, which is shows that patients with nfvPPA produce fewer verbs (t =
adjusted if the design is unbalanced. 2.2, p = .03), but the overall model was not significant (see
5. ROC/AUC curve: The receiver operating character- Figure 3).
istic (ROC) and the area under the curve (AUC)
are two evaluation measures that display the per- Adjective and Adverb Production
formance of a model. The ROC is a curve that is Patients with nfvPPA produced fewer adjectives and
created by plotting the TP rate against the FP rate. adverbs than patients with svPPA and lvPPA (see Figures 4A
The ROC/AUC measures the entire two-dimensional and 4B). This resulted in an overall statistically significant
area under the entire ROC curve from (0, 0) to (1, 1). effect of the variant on both adjectives and adverbs (see
An optimal model has an ROC/AUC closer to 1, Figures 4A and 4B and Table 3).
whereas a bad model whose predictions are 100%
wrong has an ROC/AUC of 0 (see also Baldi et al., Modal Verb, Conjunction, Determiner,
2000; Chicco, 2017, for a discussion on the selection and Preposition Production
of evaluation metrics in machine learning). There were fewer productions with modal verbs in
patients with nfvPPA (M = 0.002, SD = 0.003); patients
with lvPPA (M = 0.012, SD = 0.014) and patients with
svPPA (M = 0.012, SD = 0.013) produced approximately
Results the same number of modal verbs. Patients with lvPPA
First, we present the overall distribution of morpho- (β = 0.011, t = 2.94, p < .01) and patients with svPPA (β =
syntactic words into content words and function words. 0.01056, t = 2.76, p < .01) produced significantly more
Second, we present the distribution of POS features in PPA modal verbs than patients with nfvPPA. Although patients
variants and the results of the regression analysis. Third, with nfvPPA produce fewer conjunctions, determiners,
we present the machine learning models mentioned above particles, and prepositions from patients with svPPA and
(RF, SVM, and DT) that classify patients into PPA vari- lvPPA, the statistical models for these morphonsyntactic
ants using these features. categories were not significant.

470 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Figure 1. Violin plots with the proportion of content (A), function words (B), and the content/function word ratio (C). The shape of the plots
indicates the probability density of the data. Dots indicate the means and error bars indicate 95% confidence intervals. PPA = primary progressive
aphasia; nfvPPA = nonfluent PPA variant; lvPPA = logopenic PPA variant; svPPA = semantic PPA variant.

The Appendix shows the noun/verb ratio, noun/ of patients with nfvPPA, lvPPA, and svPPA, we employed
adjective ratio, noun/adverb ratio, noun/pronoun ratio, noun/ three machine learning models, RFs, SVMs, and DTs, in a
preposition ratio, and noun/conjunction ratio. Figure 5 supervised one-against-all classification task. The accuracy,
shows the results for the noun/adverb ratio (Figure 5A) and balanced accuracy, F1 score, precision, recall, and ROC/
the verb/adverb ratio (Figure 5B). These ratios distinguish AUC produced in supervised one-against-all classification
patients with nfvPPA, svPPA, and lvPPA significantly from RF, SVM, and DT are shown in Table 4; the ROC/
manifesting different structural effects in the patterns of AUC plots are shown in Figure 6. Patients with nfvPPA
POS in PPA individuals. and svPPA are identified better with 79% and 77% classifi-
cation accuracy, respectively, than patients with lvPPA who
are identified with 64% classification accuracy. Although,
Machine Learning Classification SVM achieved 84% classification accuracy, it was accom-
To evaluate the contribution of POS markers for panied with a lower ROC/AUC. As such, we are presenting
identifying patients with different PPA variants and deter- the feature hierarchy results from the RF model, which
mine POS features that enable an automatic classification indicated both the highest ROC/AUC and accuracy. It is

Themistocleous et al.: Part of Speech Production in PPA 471


Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Table 3. Regression models showing the statistical effects of primary progressive aphasia (PPA) variants on the distribution of part of speech
(POS).

POS PPA Estimate SE t Pr(>|t|) F test

Content word/function word ratio nfvPPA 1.72 0.123 13.97 < .0001*** F(2, 47) = 4.9, p < .01, R2 = 17%
lvPPA −0.466 0.174 −2.67 .0103*
svPPA −0.502 0.186 −2.69 .0098**
Content word nfvPPA 0.6401 0.0234 27.36 .0001*** F(2, 49) = 1.47, p < .24, R2 = 2%
lvPPA −0.0571 0.0335 −1.7 .095
svPPA −0.0342 0.0352 −0.97 .336
Function word nfvPPA 0.3800 0.0261 14.54 .0001*** F(2, 49) = 3.74, p < .05, R2 = 13%
lvPPA 0.0996 0.0375 2.66 .011*
svPPA 0.0706 0.0393 1.79 .079
Noun nfvPPA 0.3606 0.0284 12.71 .0001*** F(2, 49) = 11.4, p < .001, R2 = 32%
lvPPA −0.1468 0.0407 −3.61 .0007***
svPPA −0.1901 0.0427 −4.45 .0001***
Adverb nfvPPA 0.0337 0.0113 2.98 .0044** F(2, 49) = 6.99, p < .01, R2 = 22%
lvPPA 0.0315 0.0162 1.94 .0577
svPPA 0.0636 0.0170 3.73 .0005***
Adjective nfvPPA 0.0215 0.0142 1.51 .138 F(2, 49) = 3.16, p < .05, R2 = 11%
lvPPA 0.0232 0.0204 1.13 .263
svPPA 0.0539 0.0214 2.51 .015*
Pronoun nfvPPA 0.2243 0.0172 13.04 .0001*** F(2, 49) = 1.45, p = .25, R2 = 6%
lvPPA 0.0351 0.0247 1.42 .16
svPPA 0.0385 0.0259 1.49 .14
Verb nfvPPA 0.2243 0.0172 13.04 .0001*** F(2, 49) = 1.45, p = .24, R2 = 6%
lvPPA 0.0351 0.0247 1.42 .16
svPPA 0.0385 0.0259 1.49 .14
Determiner nfvPPA 0.1172 0.0130 8.99 .0001*** F(2, 49) = 2.66, p < .01, R2 = 10%
lvPPA 0.0344 0.0187 1.84 .072
svPPA −0.0076 0.0196 −0.38 .702
Modal verb nfvPPA 0.0017 0.0025 0.68 .501 F(2, 49) = 5.6, p < .001, R2 = 19%
lvPPA 0.0106 0.0035 3.02 .004**
svPPA 0.0098 0.0037 2.67 .01*
Preposition nfvPPA 0.1049 0.0112 9.41 .0001*** F(2, 49) = 1.53, p = .2, R2 = 6%
lvPPA 0.0279 0.0160 1.74 .088
svPPA 0.0119 0.0168 0.71 .481
Conjunction nfvPPA 0.0521 0.0078 6.69 .0001*** F(2, 49) = 0.78, p < .46, R2 = 3%
lvPPA 0.0063 0.0112 0.56 .58
svPPA 0.0147 0.0117 1.25 .22
Particle nfvPPA 0.0144 0.0037 3.88 .0003*** F(2, 49) = 0.54, p = .58, R2 = 2%
lvPPA −0.0039 0.0053 −0.74 .4657
svPPA 0.0017 0.0056 0.31 .76
Possessives nfvPPA 0.0040 0.0015 2.67 .01* F(2, 49) = 0.94, p = .39, R2 = 3%
lvPPA −0.0006 0.0021 −0.27 .79
svPPA −0.0029 0.0022 −1.32 .19

Note. The intercept of the model corresponds to the nonfluent PPA variant (nfvPPA). It also provides the standard error (SE), the t value, and
the p value, Pr(>|t|). The last column provides the overall analysis of variance score of the model and its corresponding R2. lvPPA = logopenic
PPA variant; svPPA = semantic PPA variant.
*p < .05, **p < .01, ***p < .001.

important to note that feature hierarchies depend on the indicates that highly ranked features have a greater contri-
respective machine learning model. Therefore, different bution to the final classification output; this denotes the
models might result into different featural rankings. Regard- relative importance of a POS measure. The FIs shown in
less of rankings, they demonstrate the properties that Table 5 are standardized and add up to one. The highest six
influence the model, which can be indicative of which mor- rankings for each variant are shown.
phological features influenced the classification outcome This output demonstrates the morphosyntactic selec-
of the machine learning model. tivity in patients with nfvPPA, lvPPA, and svPPA variants.
The POS features with the greatest impact on the The noun/adverb ratio; the proportion of nouns, adjectives,
machine learning classification model for the identification and adverbs; and the noun/adjective ratio distinguish pa-
of each PPA variant are shown in Table 5. The relative tients with nfvPPA from patients with the other two PPA
importance of POS as a measure of PPA variance is calcu- variants. The high predominance of content words, especially
lated from the relative FI ranking, which is employed as nouns in the productions of patients with nfvPPA, demon-
a decision node in trees generated by the RF algorithm. FI strates the overall impact of content words for the classification

472 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Figure 2. Violin plots with the distribution of nouns normalized to the total number of words produced by
patients with nfvPPA, lvPPA, and svPPA. The shape of the plots indicates the probability density of the
data. Dots indicate the means and error bars indicate 95% confidence intervals. PPA = primary progressive
aphasia; nfvPPA = nonfluent PPA variant; lvPPA = logopenic PPA variant; svPPA = semantic PPA variant.

of this variant. The machine learning model employed dif- for the quantitative assessment of connected speech and the
ferent POS measures for the identification of lvPPA. Specif- comparative evaluation of morphosyntax in all three PPA
ically, the proportion of determiners, the adverb/preposition variants using natural language processing. Our findings
ratio, the verb/pronoun ratio, the content word/function show substantial heterogeneity of POS production in patients
word ratio, and the proportion of adjectives contribute to with PPA. Grammatical selectivity is manifested in the dis-
the classification of patients with lvPPA. Finally, the svPPA tribution of content words (particularly nouns) and func-
classification model employed the proportions of nouns and tion words in distinguishing patients with different variants.
adverbs, the noun/adverb ratio, the adverb/preposition ra- In short, automatic morphosyntactic tagging of connected
tio, and the noun/verb ratio as predictors. This FI suggests speech samples from an isolated picture task corroborated
that the proportion of nouns, the noun/adverb ratio, and earlier findings that were elicited using a time-consuming
the noun/verb ratio have a high impact for the classification battery of tasks. It also provided novel results that iden-
of patients with svPPA. tify the effects of PPA on the production of adjectives and
adverbs and highlights differences in verb production in
patients with PPA compared to previous studies.
Discussion
Grammatical production in PPA (Mesulam et al., 2014, POS Production in nfvPPA
2012; Mesulam & Weintraub, 2014) varies considerably Patients with nfvPPA produced more content words
between patients with different variants. Despite the long- than function words. Specifically, nouns were more fre-
standing interest in PPA morphosyntactic production, re- quent in patients with nfvPPA than in patients with lvPPA
search has not led to conclusive morphosyntactic patterns and svPPA. This finding corroborates earlier studies that
for the systematic classification of PPA variants. This arti- suggest naming is not as impaired in nfvPPA as in svPPA
cle aimed to address the need to determine morphosyntactic (i.e., making reference to objects, people, etc., in a picture
production in patients with PPA using POS measures and description task; e.g., Bastiaanse & Jonkers, 2012; Gorno-
highlighted their role for subtyping PPA using machine Tempini et al., 2011). By contrast, the production of function
learning. This was accomplished by charting a procedure words was significantly diminished in patients with nfvPPA

Themistocleous et al.: Part of Speech Production in PPA 473


Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Figure 3. Violin plots with the distribution of verbs normalized to the total number of words produced by
patients with nfvPPA, lvPPA, and svPPA. The shape of the plots indicates the probability density of the
data. Dots indicate the means and error bars indicate 95% confidence intervals. PPA = primary progressive
aphasia; nfvPPA = nonfluent PPA variant; lvPPA = logopenic PPA variant; svPPA = semantic PPA variant.

compared to patients with lvPPA and svPPA. Our findings and lvPPA. This suggests that verb production might not
showed that patients with nfvPPA produced fewer adjec- be a good predictor for PPA variant classification (Fraser
tives and adverbs (recognized as content words) than patients et al., 2014; Kim & Thompson, 2000). However, when
with svPPA and lvPPA. Patients with nfvPPA are known verbs are taken together with nouns, the noun/verb ratio
to have impaired production of function words, so this find- is higher in patients with nfvPPA than in patients with
ing suggests that they have impairment with certain content lvPPA and svPPA. This ratio can distinguish patients with
words as well, namely, adverbs and adjectives. However, all three PPA variants (for a different finding, see Thompson,
the reduced use of adverbs by patients with nfvPPA com- Cho, et al., 2012).
pared to patients with svPPA does not imply that the lan- Modal verbs (e.g., can, should) were analyzed sepa-
guage of patients with nfvPPA is grammatically less dense rately from matrix verbs, as modal verbs are grammatical
or less meaningful than that of patients with svPPA. In words that constitute a closed class and are employed in
fact, speech production in individuals with svPPA is often speech to convey deontic modalities (e.g., the expression
indefinite and recognized as “empty.” The linear model of wishes, commands; Bybee & Fleischman, 1995; Bybee
showed that the ratios of nouns with other POS (i.e., noun/ et al., 1994). We found a different distribution of modal
adjective, noun/preposition, noun/adverb, noun/verb) can verbs from matrix verbs in PPA. Patients with nfvPPA
reliably and significantly distinguish patients with nfvPPA produced significantly fewer modal verbs than patients with
from other variants. The contribution of these ratios to lvPPA and svPPA. This finding on modal verbs can be inter-
the identification of patients with PPA was further sup- preted as a manifestation of the agrammatism that charac-
ported by the subsequent classification task. This finding terizes patients with nfvPPA (Knibb et al., 2009; Thompson
highlights the significant predominance of nouns and the & Bastiaanse, 2012; Thompson, Cho, et al., 2012; Wilson
interplay of nouns with adjectives, prepositions, adverbs, et al., 2010).
and verbs as a characteristic to the speech of patients with
nfvPPA.
Given the central role of verbs in sentence structure, POS Production in svPPA
it was surprising that patients with nfvPPA did not differ in SvPPA is characterized by a combination of word
the production of verbs compared to patients with svPPA comprehension deficits, fluent aphasia, and a particularly

474 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Figure 4. Violin plots with the distribution of adjectives (A) and the distribution of adverbs (B) normalized to the total number of words produced
by patients with nfvPPA, lvPPA, and svPPA. The shape of the plots indicates the probability density of the data. Dots indicate the means
and error bars indicate 95% confidence intervals. PPA = primary progressive aphasia; nfvPPA = nonfluent PPA variant; lvPPA = logopenic
PPA variant; svPPA = semantic PPA variant.

Figure 5. Violin plots showing the noun/adverb ratio (A) and the verb/adverb ratio (B). The shape of the plots indicates the probability density
of the data. Superimposed are error bars showing 95% confidence intervals from the mean (black circle in the middle). PPA = primary progressive
aphasia; nfvPPA = nonfluent PPA variant; lvPPA = logopenic PPA variant; svPPA = semantic PPA variant.

Themistocleous et al.: Part of Speech Production in PPA 475


Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Table 4. Mean and standard deviation (in square brackets) of model accuracy, balanced accuracy, F1 score, precision, recall, and area under the
curve (AUC) produced in supervised one-against-all classification from random forests (RF), support vector machines (SVM), and decision trees (DT).

PPA RF SVM DT

nfvPPA *Accuracy 79% [36%] 84% [31%] 79% [36%]


Balanced Accuracy 71% [40%] 71% [40%] 68% [41%]
F1 score 71% [41%] 70% [41%] 68% [41%]
Precision 71% [40%] 71% [42%] 71% [42%]
Recall 71% [40%] 70% [40%] 68% [41%]
AUC 78% [25%] 71% [36%] 78% [25%]
lvPPA *Accuracy 64% [42%] 57% [43%] 57% [42%]
Balanced Accuracy 64% [42%] 57% [43%] 57% [42%]
F1 score 63% [43%] 55% [44%] 58% [42%]
Precision 64% [45%] 56% [46%] 63% [46%]
Recall 64% [42%] 57% [43%] 57% [41%]
AUC 64% [23%] 57% [17%] 43% [32%]
svPPA *Accuracy 77% [40%] 69% [43%] 73% [38%]
Balanced Accuracy 77% [40%] 69% [43%] 73% [38%]
F1 score 76% [41%] 70% [43%] 72% [39%]
Precision 76% [42%] 72% [44%] 74% [41%]
Recall 77% [40%] 69% [43%] 73% [38%]
AUC 87% [22%] 81% [35%] 75% [25%]

Note. Shown with boldface are the models that provided the highest classification accuracy. The star symbol (*) indicates nonweighted
measures; weighted measures consider the unbalance in a design. PPA = primary progressive aphasia; nfvPPA = nonfluent PPA variant;
lvPPA = logopenic PPA variant; svPPA = semantic PPA variant.

Figure 6. Mean receiver operating characteristic (ROC) curves from severe anomia, whereas the acoustic or phonological char-
the classification models of nvfPPA, lvPPA, and svPPA. We run acteristics are viewed as unimpaired (Wilson et al., 2010).
three separate supervised one-against-all classification models We found that patients with svPPA produced fewer nouns
using rain forests (RFs), support vector machines (SVMs), and decision than patients with nfvPPA and lvPPA, which aligns with
trees (DTs). Shown in the figure is the ROC with the mean area under
the curve (AUC) for the DT, SVM, RF models and the corresponding our expectations: Patients with svPPA are characterized by
SD. The best classification is the one whose ROC approaches the impaired naming abilities (Gorno-Tempini et al., 2011). Also,
right upper corner of the plot and that has the highest AUC; here, they frequently substituted nouns with pronouns, which
the best mean AUC resulted from RFs for nfvPPA, RFs for lvPPA, was manifested in the noun/pronoun ratio (e.g., I see him).
and SVM models for svPPA.PPA = primary progressive aphasia;
nfvPPA = nonfluent PPA variant; lvPPA = logopenic PPA variant;
This can be attributed to accommodation or compensation
svPPA = semantic PPA variant. strategies for noun (e.g., I see a child) production difficul-
ties (Wilson et al., 2010). As such, the noun/pronoun ratio
can be employed to distinguish patients with svPPA from
patients with nfvPPA.
Verb production did not constitute a marker that
could differentiate patients with svPPA from patients with
lvPPA and svPPA. This finding replicates earlier studies on
verb production in patients with svPPA (Fraser et al., 2014;
Thompson, Cho, et al., 2012). Patients with svPPA pro-
duced more adverbs than patients with nfvPPA and lvPPA.
Adverbs modify the sentence and/or the verb phrase by
providing information about manner, place, time, frequency,
degree, and so forth. Adverbs were found to be important
in distinguishing individuals with svPPA from healthy con-
trols, but not from nfvPPA (Fraser et al., 2014). Unlike
patients with nfvPPA, patients with svPPA produced a
higher number of function words than patients with other
PPA variants (Meteyard & Patterson, 2009).

POS Production in lvPPA


Patients with lvPPA were found to differ significantly
from nfvPPA with regard to function words, modal verbs,
and nouns. Patients with lvPPA produced more modals

476 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Table 5. Feature importance (FI) of the classification models for patients with nfvPPA, lvPPA, and svPPA from random forests.

nfvPPA lvPPA svPPA


Ranking Features FI Features FI Features FI

1 Noun/adverb ratio 0.11 Determiners 0.10 Nouns 0.12


2 Nouns 0.09 Adverb/preposition ratio 0.08 Noun/adverb ratio 0.09
3 Adjectives 0.07 Verb/pronoun ratio 0.07 Adverbs 0.07
4 Adverbs 0.07 Content/function word ratio 0.07 Adverb/preposition ratio 0.05
5 Noun/adjective ratio 0.06 Adjectives 0.06 Noun/verb ratio 0.05
6 Content/function word ratio 0.04 Noun/verb ratio 0.06 Noun/adjective ratio 0.05

Note. The FI ranking is a standardized measure with positive values that sum to 1; the table shows the first six highest ranked features.
PPA = primary progressive aphasia; nfvPPA = nonfluent PPA variant; lvPPA = logopenic PPA variant; svPPA = semantic PPA variant.

and function words compared to patients with nfvPPA. everything they see is that patients are inclined to elicit
Interestingly, patients with lvPPA produced morphosyntac- labeling rather than narration and verbs produced are
tic constituents that overlapped with both patients with mostly in the present tense (past tense and future tense are
nfvPPA patients and patients with svPPA (i.e., their pro- relatively uncommon). Storytelling and discourse settings
portions were usually in-between those with svPPA and (e.g., free conversation) can facilitate the production of
nfvPPA). Therefore, it is not surprising that the classifi- POS and also differ in their POS distributions as shown
cation accuracy of patients with lvPPA was lower—just by several studies of speech genres (Bhatia, 1993; Swales,
64% classification accuracy—than in patients with nfvPPA 1990). Future research is important to employ computa-
(71% classification accuracy) and svPPA (77% classifica- tional analysis of morphosyntactic production in different
tion accuracy). types of discourse settings.
With regard to the aim of classifying PPA variants
Diagnostic Utility of the Automatic Analysis of POS based on morphological (POS) markers, this study is lim-
ited by the small corpus used. Cookie Theft description
The automatic analysis of POS employed in this study of 52 participants with PPA may be a respectable sample
and in previous studies (e.g., Fraser et al., 2014) highlights for PPA given the rarity of the syndrome, but it is a small
the contribution of computational methods for analyzing sample for totally unbiased unsupervised machine learn-
language production in patients with PPA and bridging the ing methods. A larger corpus would enable us to character-
gap between computational linguistic analysis and manual ize POS productions with greater precision and reduce
evaluation of language productions in PPA. In clinical set- variability. A related methodological limitation is that the
tings, the automatic analysis of POS in the form of a com- machine learning models do not classify all PPA variants
puter application enables physicians, neuropsychologists, at the same time but classify one variant against the other
and speech-language pathologists to elicit POS measures of two. The decision for one against all classification was ne-
connected speech, discourse, and so forth. At the same time, cessitated by the limited number of data.
it provides objective markers that estimate patients’ gram- Future work should employ unsupervised machine
matical competence. Clinicians can employ these objective learning models that offer simultaneous classification of all
markers for diagnosisng PPA, subtyping variants, and esti- PPA variants against each other. Outputs should be used
mating linguistic changes due to speech-language therapy in different types of discourse settings, and multifactorial
and/or the natural progression of degenerative condition. designs should be implemented with predictors that include
It also enables clinicians to measure additional POS catego- speech acoustics, grammar, and so forth, to capture both
ries that can inform individualized therapeutic programs speech and overall language deficits.
(e.g., targeting increased use of specific POS productions).
Finally, the methodological approach of this study as a
diagnostic tool can be employed in other conditions affect- Acknowledgments
ing language, such as aphasia due to stroke (Fyndanis & We are grateful to our participants and referring physicians
Themistocleous, 2019). for their interest in our study. We thank Olivia Hermann and
Bronte Ficek for their assistance in our study. This work was sup-
Limitations and Future Directions ported through grants from the National Institutes of Health (Grant
NIH/NIDCD R01 DC014475) and the Science of Learning Insti-
A major limitation in determining the difference tute at Johns Hopkins University to K. T.
between PPA variants in POS production is the type of
task. Grammatical productions, specifically the distribution
of POS, verb tense, and so forth, depend on the nature References
of the task. A common criticism of picture description Ash, S., Moore, P., Antani, S., McCawley, G., Work, M., &
tasks (e.g., Cookie Theft) that instruct patients to discuss Grossman, M. (2006). Trying to tell a tale: Discourse impairments

Themistocleous et al.: Part of Speech Production in PPA 477


Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
in progressive aphasia and frontotemporal dementia. Neurol- transcripts. Cortex, 55, 43–60. https://doi.org/10.1016/j.cortex.
ogy, 66(9), 1405–1413. https://doi.org/10.1212/01.wnl.0000210435. 2012.12.006
72614.38 Fyndanis, V., & Themistocleous, C. (2019). Are there prototypical
Baldi, P., Brunak, S., Chauvin, Y., Andersen, C. A., & Nielsen, H. associations between time frames and aspectual values? Evi-
(2000). Assessing the accuracy of prediction algorithms for dence from Greek aphasia and healthy ageing. Clinical Lin-
classification: An overview. Bioinformatics, 16(5), 412–424. guistics & Phonetics, 33(1–2), 191–217. https://doi.org/10.1080/
https://doi.org/10.1093/bioinformatics/16.5.412 02699206.2018.1480657
Bastiaanse, R., & Jonkers, R. (2012). Linguistic accounts of agram- Goodglass, H., Kaplan, E., & Barresi, B. (2001). BDAE-3: Boston
matic aphasia. In R. Bastiaanse & C. K. Thompson (Eds.), Diagnostic Aphasia Examination–Third Edition. Lippincott
Perspectives on agrammatism (pp. 17–33). https://doi.org/10.4324/ Williams & Wilkins.
9780203120378 Gorno-Tempini, M. L., Brambati, S. M., Ginex, V., Ogar, J.,
Bates, D., Mächler, M., Bolker, B. M., & Walker, S. C. (2015). Dronkers, N. F., Marcone, A., Perani, D., Garibotto, V., &
Fitting linear mixed-effects models using lme4. Journal of Sta- Miller, B. L. (2008). The logopenic/phonological variant of
tistical Software, 67(1), 1–48. https://doi.org/10.18637/jss. primary progressive aphasia. Neurology, 71(16), 1227–1234.
v067.i01 https://doi.org/10.1212/01.wnl.0000320506.79811.da
Bhatia, V. K. (1993). Analysing genre: Language use in professional Gorno-Tempini, M. L., Hillis, A. E., Weintraub, S., Kertesz, A.,
settings. Longman. Mendez, M., Cappa, S. F., Ogar, J. M., Rohrer, J. D., Black,
Bird, S., Klein, E., & Loper, E. (2009). Natural language process- S., Boeve, B. F., Manes, F., Dronkers, N. F., Vandenberghe, R.,
ing with Python: Analyzing text with the natural language toolkit. Rascovsky, K., Patterson, K., Miller, B. L., Knopman, D. S.,
O’Reilly Media. Hodges, J. R., Mesulam, M. M., & Grossman, M. (2011).
Breiman, L. (2001). Random forests. Machine Learning, 45(1), Classification of primary progressive aphasia and its variants.
5–32. https://doi.org/10.1023/A:1010933404324 Neurology, 76(11), 1006–1014. https://doi.org/10.1212/WNL.
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). 0b013e31821103e6
Classification and regression trees. Wadsworth. Gorno-Tempini, M. L., & Pressman, P. (2016). Introduction to
Bucks, R. S., Singh, S., Cuerden, J. M., & Wilcock, G. K. (2000). primary progressive aphasia. In S. L. Small & G. Hickok (Eds.),
Analysis of spontaneous, conversational speech in dementia of Neurobiology of language (pp. 935–952). Elsevier. https://
Alzheimer type: Evaluation of an objective technique for ana- doi.org/10.1016/B978-0-12-407794-2.00075-4
lysing lexical performance. Aphasiology, 14(1), 71–91. https:// Graham, N. L., Patterson, K., & Hodges, J. R. (2004). When more
doi.org/10.1080/026870300401603 yields less: Speaking and writing deficits in nonfluent pro-
Bybee, J., & Fleischman, S. (1995). Modality in grammar and dis- gressive aphasia. Neurocase, 10(2), 141–155. https://doi.org/
course. Benjamins. 10.1080/13554790490497256
Bybee, J., Perkins, R., & Pagliuca, W. (1994). The evolution of Guinn, C., & Habash, A. (2012). Language analysis of speakers with
grammar. Tense, aspect, and modality in the languages of the dementia of the Alzheimer’s type. AAAI Fall Symposium—
world. The University of Chicago Press. Technical Report, FS-12-01 (Yaruss 1998), 8–13.
Cappa, S. F., Binetti, G., Pezzini, A., Padovani, A., Rozzini, L., & Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements
Trabucchi, M. (1998). Object and action naming in Alzheimer’s of statistical learning: Data mining, inference, and prediction
disease and frontotemporal dementia. Neurology, 50(2), 351–355. (2nd ed.). Springer. https://doi.org/10.1007/978-0-387-84858-7
https://doi.org/10.1212/wnl.50.2.351 Hillis, A. E., Oh, S., & Ken, L. (2004). Deterioration of naming
Chicco, D. (2017). Ten quick tips for machine learning in compu- nouns versus verbs in primary progressive aphasia. Annals of
tational biology. BioData Mining, 10(1), 35. https://doi.org/ Neurology, 55(2), 268–275. https://doi.org/10.1002/ana.10812
10.1186/s13040-017-0155-3 Jurafsky, D., & Martin, J. H. (2009). Speech and language pro-
Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine cessing: An introduction to natural language processing, compu-
Learning, 20, 273–297. http://doi.org/10.1007/BF00994018 tational linguistics, and speech recognition (2nd ed.). Pearson
Daniele, A., Giustolisi, L., Silveri, M. C., Colosimo, C., & Gainotti, G. Prentice Hall.
(1994). Evidence for a possible neuroanatomical basis for Kavé, G., Leonard, C., Cupit, J., & Rochon, E. (2007). Structur-
lexical processing of nouns and verbs. Neuropsychologia, ally well-formed narrative production in the face of severe con-
32(11), 1325–1341. https://doi.org/10.1016/0028-3932(94) ceptual deterioration: A longitudinal case study of a woman
00066-2 with semantic dementia. Journal of Neurolinguistics, 20(2),
Faroqi-Shah, Y., Treanor, A., Ratner, N. B., Ficek, B., Webster, K., 161–177. https://doi.org/10.1016/j.jneuroling.2006.06.003
& Tsapkini, K. (2020). Using narratives in differential diagnosis Kertesz, A., Davidson, W., & McCabe, P. (1998). Primary progres-
of neurodegenerative syndromes. Journal of Communication sive semantic aphasia: A case study. Journal of the Interna-
Disorders, 85, 105994. https://doi.org/10.1016/j.jcomdis.2020. tional Neuropsychological Society, 4(4), 388–398. https://www.
105994 ncbi.nlm.nih.gov/pubmed/9656612
Fraser, K. C., Lundholm Fors, K., Eckerström, M., Themistocleous, Kim, M., & Thompson, C. K. (2000). Patterns of comprehension
C., & Kokkinakis, D. (2018). Improving the sensitivity and and production of nouns and verbs in agrammatism: Implica-
specificity of MCI screening with linguistic information. In Pro- tions for lexical organization. Brain and Language, 74(1), 1–25.
ceedings of the LREC 2018 Workshop “Resources and process- https://doi.org/10.1006/brln.2000.2315
ing of linguistic, para-linguistic and extra-linguistic data from Knibb, J. A., Woollams, A. M., Hodges, J. R., & Patterson, K.
people with various forms of cognitive/psychiatric impairments (2009). Making sense of progressive non-fluent aphasia: An
(RaPID-2)” (2015) (pp. 19–26). http://lrec-conf.org/workshops/ analysis of conversational speech. Brain, 132(10), 2734–2746.
lrec2018/W31/pdf/3_W31.pdf https://doi.org/10.1093/brain/awp207
Fraser, K. C., Meltzer, J. A., Graham, N. L., Leonard, C., Hirst, G., Knopman, D. S., Kramer, J. H., Boeve, B. F., Caselli, R. J., Graff-
Black, S. E., & Rochon, E. (2014). Automated classification Radford, N. R., Mendez, M. F., Miller, B. L., & Mercaldo, N.
of primary progressive aphasia subtypes from narrative speech (2008). Development of methodology for conducting clinical

478 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
trials in frontotemporal lobar degeneration. Brain, 131(11), Themistocleous, C., & Kokkinakis, D. (2018). THEMIS-SV:
2957–2968. https://doi.org/10.1093/brain/awn234 Automatic classification of language disorders from speech
Kuznetsova, A., Brockhoff, P. B., & Christensen, R. H. B. (2017). signals. ESOC 2018: European Stroke Organisation Con-
lmertest package: Tests in linear mixed effects models. Journal ference, Gothenburg. https://doi.org/10.26226/morressier.
of Statistical Software, 82(13), 1–26. https://doi.org/10.18637/ 5ab8f55bd462b8029238c82d
jss.v082.i13 Thomas, C., Keselj, V., Cercone, N., Rockwood, K., & Asp, E. (2005).
Lien, D., Hu, Y., & Liu, L. (2017). A note on using ratio variables Automatic detection and rating of dementia of Alzheimer type
in regression analysis. Economics Letters, 150, 114–117. https:// through lexical analysis of spontaneous speech. In IEEE Inter-
doi.org/10.1016/j.econlet.2016.11.019 national Conference on Mechatronics and Automation (Vol. 3,,
Mesulam, M. M. (2001). Primary progressive aphasia. Annals of pp. 1569–1574). https://doi.org/10.1109/icma.2005.1626789
Neurology, 49(4), 425–432. https://www.ncbi.nlm.nih.gov/ Thompson, C. K. (2012). Northwestern Assessment of Verbs and
pubmed/11310619 Sentences (NAVS). Northwestern University.
Mesulam, M. M. (2013). Primary progressive aphasia and the Thompson, C. K., Ballard, K. J., Tait, M. E., Weintraub, S., &
language network: The 2013 H. Houston merritt lecture. Mesulam, M. M. (1997). Patterns of language decline in non-
Neurology, 81(5), 456–462. https://doi.org/10.1212/WNL. fluent primary progressive aphasia. Aphasiology, 11(4–5),
0b013e31829d87df 297–321. https://doi.org/10.1080/02687039708248473
Mesulam, M. M., Rogalski, E. J., Wieneke, C., Hurley, R. S., Geula, Thompson, C. K., & Bastiaanse, R. (2012). Introduction to agram-
C., Bigio, E. H., Thompson, C. K., & Weintraub, S. (2014). Pri- matism. In Perspectives on agrammatism. Psychology Press.
mary progressive aphasia and the evolving neurology of the https://doi.org/10.4324/9780203120378
language network. Nature Reviews Neurology, 10(10), 554–569. Thompson, C. K., Cho, S., Hsu, C. J., Wieneke, C., Rademaker,
https://doi.org/10.1038/nrneurol.2014.159 A., Weitner, B. B., Mesulam, M. M., & Weintraub, S. (2012).
Mesulam, M. M., & Weintraub, S. (2014). Is it time to revisit Dissociations between fluency and agrammatism in primary
the classification guidelines for primary progressive aphasia. progressive aphasia. Aphasiology, 26(1), 20–43. https://doi.org/
Neurology, 82(13), 1108–1109. https://doi.org/10.1212/WNL. 10.1080/02687038.2011.584691
0000000000000272 Thompson, C. K., Lukic, S., King, M. C., Mesulam, M. M., &
Mesulam, M. M., Wieneke, C., Thompson, C., Rogalski, E., & Weintraub, S. (2012). Verb and noun deficits in stroke-induced
Weintraub, S. (2012). Quantitative classification of primary and primary progressive aphasia: The Northwestern Naming
progressive aphasia at early and mild impairment stages. Brain, Battery. Aphasiology, 26(5), 632–655. https://doi.org/10.1080/
135(5), 1537–1553. https://doi.org/10.1093/brain/aws080 02687038.2012.676852
Meteyard, L., & Patterson, K. (2009). The relation between con- Thompson, C. K., & Mack, J. E. (2014). Grammatical impairments
tent and structure in language production: An analysis of speech in PPA. Aphasiology, 28(8–9), 1018–1037. https://doi.org/10.1080/
errors in semantic dementia. Brain and Language, 110(3), 121–134. 02687038.2014.912744
https://doi.org/10.1016/j.bandl.2009.03.007 Thompson, C. K., Meltzer-Asscher, A., Bachrach, A., Roy, I., &
Orimaye, S. O., Wong, J. S.-M., & Golden, K. J. (2014). Learning Stockall, L. (2014). Neurocognitive mechanisms of verb argument
predictive linguistic features for alzheimer’s disease and related structure processing. John Benjamins. https://doi.org/10.1075/
dementias using verbal utterances. In Proceedings of the Work- lfab.10.07tho
shop on Computational Linguistics and Clinical Psychology: Thompson, C. K., Meltzer-Asscher, A., Cho, S., Lee, J., Wieneke, C.,
From linguistic signal to clinical reality (pp. 78–87). Association Weintraub, S., & Mesulam, M. M. (2013). Neurocognitive mech-
for Computational Linguistics. https://doi.org/10.3115/v1/W14- anisms of verb argument structure processing. Behavioural
3210 Neurology, 26(1–2), 35–54. https://doi.org/10.1155/2013/749412
R Core Team. (2016). R: A language and environment for statistical Tsapkini, K., Webster, K. T., Ficek, B. N., Desmond, J. E., Onyike,
computing. R Foundation for Statistical Computing. https:// C. U., Rapp, B., Frangakis, C. E., & Hillis, A. E. (2018). Elec-
www.R-project.org/ trical brain stimulation in different variants of primary progres-
Rentoumi, V., Raoufian, L., Ahmed, S., de Jager, C. A., & Garrard, P. sive aphasia: A randomized clinical trial. Alzheimer’s & Dementia
(2014). Features and machine learning classification of connected (New York, N. Y.), 4(1), 461–472. https://doi.org/10.1016/
speech samples from patients with autopsy proven Alzheimer’s j.trci.2018.08.002
disease with and without additional vascular pathology. Journal Wilson, S. M., Brandt, T. H., Henry, M. L., Babiak, M., Ogar,
of Alzheimer’s Disease, 42, S3–S17. https://doi.org/10.3233/JAD- J. M., Salli, C., Wilson, L., Peralta, K., Miller, B. L., & Gorno-
140555 Tempini, M. L. (2014). Inflectional morphology in primary
Riello, M., Faria, A. V., Ficek, B., Webster, K., Onyike, C. U., progressive aphasia: An elicited production study. Brain and
Desmond, J., Frangakis, C., & Tsapkini, K. (2018). The role Language, 136, 58–68. https://doi.org/10.1016/j.bandl.2014.
of language severity and education in explaining performance 07.001
on object and action naming in primary progressive aphasia. Wilson, S. M., DeMarco, A. T., Henry, M. L., Gesierich, B., Babiak,
Frontiers in Aging Neuroscience, 10, 346. https://doi.org/10.3389/ M., Mandelli, M. L., Miller, B. L., & Gorno-Tempini, M. L.
fnagi.2018.00346 (2014). What role does the anterior temporal lobe play in sentence-
Russell, L. (2018). Emmeans: Estimated marginal means, aka least- level processing? Neural correlates of syntactic processing in
squares means. R package version 1(2). https://cran.r-project.org/ semantic variant primary progressive aphasia. Journal of Cog-
web/packages/emmeans/index.html nitive Neuroscience, 26(5), 970–985. https://doi.org/10.1162/
Swales, J. M. (1990). Genre analysis: English in academic and re- jocn_a_00550
search settings. Cambridge University Press. Wilson, S. M., Henry, M. L., Besbris, M., Ogar, J. M., Dronkers,
Themistocleous, C., Eckerstrom, M., & Kokkinakis, D. (2018). N. F., Jarrold, W., Miller, B. L., & Gorno-Tempini, M. L.
Identification of mild cognitive impairment from speech in (2010). Connected speech production in three variants of primary
Swedish using deep sequential neural networks. Frontiers in progressive aphasia. Brain, 133(7), 2069–2088. https://doi.org/
Neurology, 9, 975. https://doi.org/10.3389/fneur.2018.00975 10.1093/brain/awq129

Themistocleous et al.: Part of Speech Production in PPA 479


Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions
Appendix
Regression Models Showing the Statistical Effects of Primary Progressive Aphasia (PPA) Variants on the Distribution of Part
of Speech

POS ratio PPA Estimate SE t Pr(>|t|) F test

Noun/verb ratio nfvPPA 1.583 0.172 9.21 < .0001*** F(2, 47) = 7.3, p < .01, R2 = 24%
lvPPA −0.678 0.243 −2.79 .0076**
svPPA −0.939 0.26 −3.61 .0007***
Noun/adjective ratio nfvPPA 11.87 1.36 8.7 < .0001*** F(2, 43) = 7.58, p < .01, R2 = 26%
lvPPA −5.5 1.79 −3.07 .0037**
svPPA −6.89 1.86 −3.7 .0006***
Noun/adverb ratio nfvPPA 12.69 1.9 6.67 < .0001*** F(2, 46) = 8.09, p < .001, R2 = 26%
lvPPA −8.16 2.69 −3.03 .0040**
svPPA −10.51 2.78 −3.78 .0005***
Noun/pronoun ratio nfvPPA 5.414 0.952 5.69 < .0001*** F(2, 44) = 4.37, p < .05, R2 = 17%
lvPPA −2.544 1.327 −1.92 .0617
svPPA −4.043 1.394 −2.9 .0058**
Noun/preposition ratio nfvPPA 3.29 0.384 8.56 < .0001*** F(2, 46) = 5.36, p < .01, R2 = 19%
lvPPA −1.442 0.536 −2.69 .0099**
svPPA −1.678 0.572 −2.93 .0052**
Noun/conjunction ratio nfvPPA 7.302 0.941 7.76 < .0001*** F(2, 45) = 6.01, p < .01, R2 = 21%
lvPPA −2.708 1.293 −2.09 .0419*
svPPA −4.735 1.377 −3.44 .0013**
Verb/adjective ratio nfvPPA 9.223 1.585 5.82 < .0001*** F(2, 43) = 0.142, p = .86, R2 = 0.6%
lvPPA −1.021 2.079 −0.49 .6300
svPPA −0.969 2.165 −0.45 .6600
Verb/adverb ratio nfvPPA 9.25 1.25 7.4 .0000*** F(2, 46) = 5.34, p < .001, R2 = 19%
lvPPA −4.29 1.77 −2.43 .0191*
svPPA −5.64 1.83 −3.09 .0034**
Verb/pronoun ratio nfvPPA 3.779 0.699 5.41 < .0001*** F(2, 44) = 1.19, p = .3, R2 = 5%
lvPPA −0.529 0.973 −0.54 .5900
svPPA −1.558 1.023 −1.52 .1300
Verb/preposition ratio nfvPPA 2.254 0.256 8.8 < .0001*** F(2, 46) = 0.15, p = .86, R2 = 0.6%
lvPPA −0.016 0.357 −0.04 .9600
svPPA 0.174 0.381 0.46 .6500
Verb/conjunction ratio nfvPPA 5.071 0.642 7.9 .0000*** F(2, 45) = 0.31, p = = .73, R2 = 1%
lvPPA 0.234 0.882 0.27 .7900
svPPA −0.487 0.939 −0.52 .6100
Adjective/adverb ratio nfvPPA 0.7450 0.1506 4.95 0.000011*** F(2, 46) = 0.23, p = .79, R2 = 1%
lvPPA 0.0748 0.2130 0.35 0.73
svPPA −0.0753 0.2200 −0.34 0.73
Adjective/pronoun ratio nfvPPA 0.3438 0.1186 2.90 0.0058** F(2, 47) = 0.9, p = .41, R2 = 3%
lvPPA 0.2086 0.1653 1.26 0.2137
svPPA 0.0365 0.1737 0.21 0.8346
Adjective/preposition ratio nfvPPA 0.2478 0.0948 2.62 0.012* F(2, 46) = 1.74, p = .19, R2 = 7%
lvPPA 0.1579 0.1321 1.20 0.238
svPPA 0.2579 0.1410 1.83 0.074
Adjective/conjunction ratio nfvPPA 0.626 0.266 2.35 0.023* F(2, 45) = 0.87, p = .4, R2 = 3%
lvPPA 0.474 0.366 1.30 0.202
svPPA 0.165 0.390 0.42 0.673
Adverb/pronoun ratio nfvPPA 0.430 0.244 1.77 0.084 F(2, 44) = 0.93, p = .4, R2 = 4%
lvPPA 0.458 0.339 1.35 0.184
svPPA 0.185 0.357 0.52 0.607
Adverb/preposition ratio nfvPPA 0.309 0.131 2.36 0.022* F(2, 46) = 2.83, p = .07, R2 = 10%
lvPPA 0.350 0.182 1.92 0.061
svPPA 0.419 0.194 2.16 0.036*
Adverb/preposition ratio nfvPPA 0.736 0.180 4.09 0.00018*** F(2, 45) = 2.88, p = .07, R2 = 11%
lvPPA 0.571 0.248 2.31 0.02570
svPPA 0.452 0.264 1.71 0.09348

Note. The intercept of the model corresponds to the nonfluent PPA variant (nfvPPA). It also provides the standard error (SE), the t value, and
the p value, Pr(>|t|). The last column provides the overall analysis of variance score of the model and its corresponding R2. lvPPA = logopenic
PPA variant; svPPA = semantic PPA variant.
*p < .05. **p < .01. ***p < .001.

480 American Journal of Speech-Language Pathology • Vol. 30 • 466–480 • February 2021

Downloaded from: https://pubs.asha.org Iberoamericana- Instituto Universitaria on 03/24/2024, Terms of Use: https://pubs.asha.org/pubs/rights_and_permissions

You might also like