Verb-argument integration in primary progressive aphasia: Real-time argument access and selection

Jennifer E Mack; M-Marsel Mesulam; Emily J Rogalski; Cynthia K Thompson

doi:10.1016/j.neuropsychologia.2019.107192

. Author manuscript; available in PMC: 2020 Nov 1.

Published in final edited form as: Neuropsychologia. 2019 Sep 13;134:107192. doi: 10.1016/j.neuropsychologia.2019.107192

Verb-argument integration in primary progressive aphasia: Real-time argument access and selection

Jennifer E Mack ¹, M-Marsel Mesulam ^2,³, Emily J Rogalski ^2,⁴, Cynthia K Thompson ^1,^2,³

PMCID: PMC6858499 NIHMSID: NIHMS1051971 PMID: 31521633

Abstract

Background.

Impaired sentence comprehension is observed in the three major subtypes of PPA, with distinct performance patterns relating to impairments in comprehending complex sentences in the agrammatic (PPA-G) and logopenic (PPA-L) variants and word comprehension in the semantic subtype (PPA-S). However, little is known about basic combinatory processes during sentence comprehension in PPA, such the integration of verbs with their subject and object(s) (verb-argument integration).

Methods.

The present study used visual-world eye-tracking to examine real-time verb-argument integration in individuals with PPA (12 with PPA-G, 10 with PPA-L, and 6 with PPA-S) and neurotypical older adults (15). Two baseline experiments probed eye movement control, using a non-linguistic task, and noun comprehension, respectively. Two verb-argument integration experiments examined the effects of verb meaning on (a) lexical access of the verb’s direct object (argument access) and (b) selection of a semantically-appropriate direct object (argument selection), respectively. Eye movement analyses were conducted only for trials with correct behavioral responses, allowing us to distinguish accuracy and online processing.

Results.

The eye movement control experiment revealed no significant impairments in PPA, whereas the noun comprehension experiment revealed reduced accuracy and eye-movement latencies in PPA-S, and to a lesser extent PPA-G. In the argument access experiment, verb meaning facilitated argument access normally in PPA-G and PPA-L; in PPA-S, verb-meaning effects emerged on an atypical time course. In the argument selection experiment, significant impairments in accuracy were observed only in PPA-G, accompanied by markedly atypical eye movement patterns.

Conclusion.

This study revealed two distinct patterns of impaired verb-argument integration in PPA. In PPA-S, impaired verb-argument integration was observed in the argument access experiment, indicating impairments in basic semantic combinatory processes which likely relate to damage in ventral language pathways. In contrast, listeners with PPA-G showed marked impairments of argument selection, likely relating to damage to left inferior frontal regions.

Keywords: primary progressive aphasia, eye-tracking, syntax, semantics, verb-argument structure

Introduction

Comprehension of simple sentences such as Amy will close the jar involves accessing individual words and combining them into syntactic and semantic units. Syntactic combinatory processes serve to combine words into phrases (e.g., combining a verb (e.g., close) and a direct object noun phrase (e.g., the jar) to create a verb phrase) and semantic combinatory mechanisms integrate the meanings of individual words into larger units, such as events (e.g., close (Amy, jar)). Verb-argument integration refers to the processes that combine verbs with their arguments (grammatical subjects and objects), and is an important component of the interface between syntax and semantics. Verb-argument integration is grounded in constraints that verbs place on their arguments. For example, the verb tour requires a direct object (She toured the castle vs. *She toured in the castle), which must be inanimate (e.g., toured the castle vs. *toured the duke) and have specific semantic properties (e.g., toured the castle vs. ?toured the garage). Eye-tracking studies have shown that neurotypical adult participants use the syntactic and semantic requirements of verbs to speed argument access, e.g., access to the meaning of the verb’s object(s) (Altmann & Kamide, 1999; Boland, 2005; Borovsky, Elman, & Fernald, 2012; Kamide, Altmann, & Haywood, 2003; Kukona, Fang, Aicher, Chen, & Magnuson, 2011; Milburn, Warren, & Dickey, 2016; Staub, Abbott, & Bogartz, 2012). These verb-based requirements are also used to select arguments, e.g., to choose syntactically and semantically appropriate objects for a verb, in sentence production and completion tasks (Hayes, Dickey, & Warren, 2016; Mack, Ji, & Thompson, 2013b). In the present study, we used eye-tracking to examine verb-argument integration (access and selection) in primary progressive aphasia (PPA).

PPA is a neurodegenerative disease characterized by a gradual, progressive onset of aphasia, with other cognitive abilities relatively preserved in early disease stages (Mesulam, 1982). There are three major clinical subtypes of PPA. The agrammatic subtype (PPA-G) is characterized by grammatical impairments in language production, often accompanied by motor speech deficits, and neurodegeneration is typically observed in left-hemisphere frontal regions, including the inferior frontal gyrus (IFG), premotor, and supplementary motor cortical areas, as well as dorsal white matter tracts such as the superior longitudinal fasciculus (SLF)/arcuate fasciculus (AF) and the frontal aslant tract (Catani et al., 2013; Galantucci et al., 2011; Gorno-Tempini et al., 2004; Mandelli et al., 2014; Mesulam et al., 2009b). The logopenic subtype (PPA-L) is associated with deficits in word retrieval and phonological working memory, and neurodegeneration typically centers on the left temporo-parietal junction (TPJ) – including the posterior middle and superior temporal gyri, the angular gyrus, and the supramarginal gyrus – as well as portions of dorsal white matter tracts such as the SLF/AF (Galantucci et al., 2011; Gorno-Tempini et al., 2008; Mesulam et al., 2009b; Teichmann et al., 2013). Word comprehension is relatively intact in both PPA-G and PPA-L. In contrast, the semantic subtype (PPA-S) is characterized by impaired word comprehension and production, reflecting damage to lexical-semantic representations, and neurodegeneration typically occurs in the anterior temporal lobes (ATLs), often to a greater extent in the left hemisphere, and ventral white matter tracts such as the inferior longitudinal fasciculus and uncinate fasciculus (Catani et al., 2013; Galantucci et al., 2011; Gorno-Tempini et al., 2004; Mandelli et al., 2014; Mesulam et al., 2009b).

To date, most research on sentence comprehension in PPA has focused on the effects of syntactic complexity (e.g., non-canonical word order). Deficits in complex sentence comprehension are a characteristic feature of PPA-G and are sometimes observed in PPA-L (Amici et al., 2007; Thompson et al., 2013; Wilson et al., 2010a; Wilson, Galantucci, Tartaglia, & Gorno-Tempini, 2012). These deficits may be linked to impairments in grammatical processing in PPA-G and phonological working memory in PPA-L (Thompson & Mack, 2014; Wilson et al., 2012). In PPA-S, comprehension of complex structures is relatively preserved (Amici et al., 2007; Rochon, Kave, Cupit, Jokel, & Winocur, 2004; Wilson et al., 2014b). Due to the focus on syntactic processing, in most of these studies lexical demands were minimized.

In contrast, relatively little is known about how patients with PPA perform on tasks that place greater demands on combinatory lexical processing, such as verb-argument integration. One previous study (Peelle, Cooke, Moore, Vesely, & Grossman, 2007) found impaired sensitivity to verb-argument structure violations (e.g., *The blankets view movies vs. The passengers view movies) across PPA subtypes, although a methodologically similar study found intact processing of these violations in individuals with “progressive non-fluent PPA” (PPA-G or PPA-L) and individuals with PPA-S were not included (Price & Grossman, 2005). In language production, speakers with PPAG have difficulty producing verbs with complex (vs. simple) argument structures, and make frequent verb-argument structure errors, whereas verb-argument production is relatively preserved in PPA-L and PPA-S (Thompson, Ballard, Tait, Weintraub, & Mesulam, 1997; Thompson et al., 2012a; Thompson, Lukic, King, Mesulam, & Weintraub, 2012c). Thus, these studies suggest that verb-argument integration may be impaired in some individuals with PPA, but leave open the question of how these deficits are manifested across PPA subtypes and linguistic tasks.

Given the relative paucity of research on verb-argument integration in PPA, our predictions in the present study were informed by previous research on stroke-induced aphasia, as well as neuroimaging studies with neurotypical adult participants. Much relevant research on stroke-induced aphasia has focused on agrammatic (Broca’s) aphasia. Agrammatic aphasia shares many linguistic features with PPA-G including non-fluent language production, grammatical production deficits and difficulty comprehending complex sentences (Thompson et al., 2013). In sentence comprehension, individuals with stroke-induced agrammatic aphasia show online sensitivity to verb-argument structure distinctions (Piñango & Zurif, 2001; Shapiro, Gordon, Hack, & Killackey, 1993; Shapiro & Levine, 1990) and are relatively good at detecting verb-argument structure violations (Grodzinsky and Finkel (1998); Kim and Thompson (2000, 2004); though see Kielar, Meltzer-Asscher, and Thompson (2012)), suggesting intact access to verb-argument structure representations. However, verb-argument structure production is impaired (see Thompson and Meltzer-Asscher (2014) for a review), indicating difficulty with selecting verb-arguments and sequencing them accurately.

Eye-tracking studies have provided additional evidence for argument selection deficits in agrammatic aphasia. In a previous study (Mack et al. (2013b), based on Altmann and Kamide (1999)), people with stroke-induced agrammatic aphasia and neurotypical older adults completed two eye-tracking experiments. In the argument access experiment (Experiment 1 in Mack et al. (2013b)), participants viewed an array of four object pictures (e.g., jar, plate, pencil, stick), heard a simple sentence (e.g., Tomorrow Amy will close/break the jar), and indicated whether there was a picture in the array corresponding to the verb’s direct object (e.g., a picture of a jar). Semantically restrictive verbs (e.g., close) were compatible only with the direct object (e.g., the jar) whereas unrestrictive verbs (e.g., break) were compatible with all four objects. The experiment tested whether semantically-restrictive verbs, as compared to unrestrictive verbs, speeded access to the meaning of the verb’s direct object, as indicated by eye movement patterns. Indeed, it did, in both neurotypical adults and adults with aphasia. In a second argument selection task (Experiment 2 in Mack et al. (2013b)), participants viewed object arrays, heard sentence fragments (e.g., Tomorrow Amy will open/break the …), and were required to click on the picture that best fit the sentence, thus probing the use of semantically-restrictive verbs to select an appropriate argument. In contrast with Experiment 1, in Experiment 2 participants with agrammatic aphasia evinced reduced accuracy and processing delays. Most of the participants with aphasia in that study had lesions that included the left inferior frontal gyrus (IFG), which is characteristic of Broca’s/agrammatic aphasia (Fridriksson, Fillmore, Guo, & Rorden, 2015; Mohr et al., 1978; Thompson & Meltzer-Asscher, 2014; Yourganov, Smith, Fridriksson, & Rorden, 2015) and associated with lexical selection in neurotypical adults (Hickok & Poeppel, 2007; Thompson-Schill, D’Esposito, Aguirre, & Farah, 1997), suggesting that the left IFG plays a critical role in argument selection.

Patients with stroke-related agrammatic aphasia also often evince damage within the left temporoparietal region, including the superior and middle temporal and angular gyri (STG, MTG, AG, respectively) and the TPJ (Barbieri, Mack, Chiappetta, Europa, & Thompson, in press; Rogalsky et al., 2018). Notably, these regions also have been implicated in verb-argument processing and processing of thematic relations in healthy people. Here, the term “thematic” is used in a broad sense to refer to event-based relationships (e.g., dog-leash, in which the two words are from different semantic categories but often co-occur in the same events), including verb-argument relationships (e.g., eat-cake). Processing of these thematic relationships, which engages left temporoparietal regions, stands in contrast with feature-based or “taxonomic” relations (e.g., cat-dog, in which the semantic features overlap), associated with the ATL (review: Mirman, Landrigan, and Britt (2017)).

Neuroimaging studies have also found temporoparietal regions to play an important role in verb-argument integration because they encode the meanings (or control access to the meanings) of verbs/events (e.g., Pallier, Devauchelle, and Dehaene (2011)). In one fMRI study, activity in the left AG correlated with verb identity, but not noun identity in two-word phrases (e.g., eat meat/eat quickly vs. eat meat/tasty meat) (Boylan, Trueswell, & Thompson-Schill, 2015). In several other fMRI studies, activation in the left AG increased with the number of arguments encoded by the verb (i.e., ditransitive > transitive > intransitive), indicating sensitivity to event complexity and its syntactic correlates (den Ouden, Fix, Parrish, & Thompson, 2009; Meltzer-Asscher, Mack, Barbieri, & Thompson, 2015; Thompson et al., 2007; Thompson, Bonakdarpour, & Fix, 2010). The left ATL also appears to be important for verb-argument integration, but for a different reason: it has been shown to support basic combinatory processing (Bemis & Pylkkanen, 2011; Bornkessel-Schlesewsky & Schlesewsky, 2013; Lau, Phillips, & Poeppel, 2008; Lau, Weber, Gramfort, Hamalainen, & Kuperberg, 2016; Matchin, Hammerly, & Lau, 2017; Westerlund, Kastner, Al Kaabi, & Pylkkanen, 2015; Westerlund & Pylkkanen, 2014). Although some neurobiological models associate the ATL with syntactic and semantic combinatory processing, evidence from PPA-S – in which syntactic processing is demonstrably intact despite left ATL atrophy – suggests that the ATL may specifically support semantic combinatory processes (Wilson et al., 2014b). It follows, then, that damage to temporoparietal regions is associated with verb-argument integration deficits (Meltzer-Asscher, Mack, Barbieri, & Thompson, 2015; Thompson, Bonakdarpour, & Fix, 2010) and impaired processing of thematic relations (Kalenine & Buxbaum, 2016; Mirman & Graziano, 2012; Mirman et al., 2017; Schwartz et al., 2011), whereas damage to the ATL and ventral language pathways, both in stroke (Schwartz et al., 2011) and in PPA (Hurley, Paller, Rogalski, & Mesulam, 2012; Mesulam et al., 2009a; Mesulam et al., 2013) disrupts taxonomic processes as well as semantic combinatory processes.

The primary experiments in the present study used the argument access and argument selection paradigms from Mack et al. (2013b) to examine verb-argument integration in PPA. For PPA-S and PPA-L, two alternative hypotheses were considered. The first hypothesis was that verb-argument integration (both access and selection) would be intact in these subtypes of PPA. This hypothesis was based on previous research indicating intact verb-argument structure production (Thompson et al., 2012a; Thompson et al., 2013). The second hypothesis, based on previous research on the neural basis of verb-argument integration, was that deficits would be observed in one or both subtypes. Damage to the left ATL in PPA-S may result in deficits in semantic combinatorial processes, whereas damage to temporoparietal regions (i.e., the TPJ) in PPA-L may be expected to impair access to verb representations. Given the linguistic similarities between agrammatic aphasia caused by stroke and PPA-G, we expected that individuals with PPA-G would show the same pattern that we previously observed in stroke-related agrammatism: intact use of verb-argument structure information to access arguments, but impaired argument selection (Mack et al., 2013b).

In addition to these experiments, we also examined eye movements in a non-linguistic task (eye movement control) as well as in a single-word comprehension task (noun comprehension) in order to ascertain the presence of oculomotor abnormalities, which have been observed in a wide range of neurodegenerative diseases (Leigh & Zee, 2015), and to better understand lexical contributions to verb-argument integration in PPA, respectively. If abnormalities were found on these tasks, we considered them in relation to performance on our primary experiments of verb-argument processing.

Several previous studies have examined eye movement control in PPA, indicating that low-level eye movements, such as saccades made reflexively in response to visual stimuli, are generally preserved (Boxer et al. (2006); Garbutt et al. (2008), although cf. Burrell, Hornberger, Carpenter, Kiernan, and Hodges (2012), who reported a trend towards increased saccadic latencies). However, eye movement abnormalities have been observed in more complex tasks, such as the anti-saccade task (which requires inhibition of visually-guided saccades; Boxer et al. (2006); Garbutt et al. (2008)) and predictive smooth pursuit tasks (which require the eyes to “follow” a target that momentarily disappears from a visual display; Coppe, Orban de Xivry, Yuksel, Ivanoiu, and Lefevre (2012)). Further, the cortical regions that support cognitive control of eye movements, including the frontal eye fields (located at the junction of the precentral, middle frontal, and superior frontal gyri), supplementary eye fields (located anterior to the supplementary motor area), dorsolateral prefrontal cortex, and parietal eye fields (located in the intraparietal sulcus) (Leigh & Kennard, 2004), neighbor many of the areas of cortical atrophy often observed in PPA, particularly PPA-G and PPA-L. Therefore, when using eye movement measures to probe language processing in PPA, it is important to also examine performance in a non-linguistic task. Because we used a simple reflexive saccade paradigm (albeit in the context of complex visual stimuli, which were matched to those used in our primary experiments), we expected normal eye movement patterns in the PPA participants in the eye-movement control experiment.

Impaired noun comprehension is one of the diagnostic features of PPA-S, whereas in PPA-G and PPA-L, noun comprehension accuracy is generally preserved (Hillis et al., 2006; Thompson et al., 2012c). However, relatively little work has examined the time course of noun comprehension in PPA – a critical question for the present study, given that our primary dependent variable was eye-movements to an object (noun) picture during sentence comprehension. One recent eye-tracking study examining noun comprehension (Faria, Race, Kim, & Hillis, 2018) found substantially fewer fixations on the target picture in PPA-S, as compared to neurotypical controls and listeners with PPA-G and PPA-L, even though accuracy was only mildly impaired. Another eye-tracking study of noun comprehension (Seckin et al., 2016) revealed impaired speed and accuracy in PPA-S, and eye-movement patterns indicated “taxonomic blurring”, i.e., the loss of lexical-semantic features that distinguish between objects of the same superordinate category (Hurley et al., 2012; Mesulam et al., 2009a; Mesulam et al., 2013). In contrast, participants with PPA-G and PPA-L showed slowed eye movements to the target picture, but normal-like accuracy and no evidence of taxonomic blurring. Therefore, listeners with PPA-G and PPA-L may show subtle lexical processing impairments in online tasks, though offline comprehension patterns are preserved (cf. studies on noun production which have found subtle impairments of semantic and phonological processing in PPA-G and PPA-L (Mack et al., 2013a; Thompson et al., 2012b)). In the present study, linguistic stimuli were presented auditorily rather than visually (as in the Seckin et al., 2016 study). This is relevant given that auditory processing deficits have been reported across PPA subtypes, which may affect the time course of word comprehension (Grube et al., 2016). Therefore, for the noun comprehension task we expected to find impaired accuracy in PPA-S (but not PPA-G or PPA-L), but slowed eye-movement latencies in all participant groups.

Method

Participants

Twenty-eight participants with PPA (12 PPA-G, 10 PPA-L, 6 PPA-S) and 15 neurotypical older adult control participants (“controls”) took part in the study. Four participants were excluded from E1 only: one participant with PPA-G and two controls had poor eye-tracking data quality that was limited to this task, and one participant with PPA-G did not make any mouse-click responses. All participants were right-handed native speakers of English. The control participants passed a neuropsychological test battery that screened for perceptual and cognitive impairments, and reported no prior history of neurological, language or learning disorders. The study was approved by the Institutional Review Board at Northwestern University and all participants gave informed consent.

Demographic, neuropsychological, and language measures (Table 1) were compared across groups using pairwise Mann-Whitney U Tests (significance thresholds of p <. 05). The four participant groups did not differ significantly with respect to age; however, the PPA-L group had a greater number of years of education as compared to the other three groups. The participants with PPA were diagnosed and classified into PPA subtypes following Mesulam, Wieneke, Thompson, Rogalski, and Weintraub (2012); see also Gorno-Tempini et al. (2011). The three PPA groups did not differ with respect to symptom duration. Scores on the Mini Mental-State Examination (MMSE; Folstein, Folstein, and McHugh (1975)) and the Clinical Dementia Rating (CDR; Morris (1993)) also did not differ across PPA groups and indicated mild (if any) non-verbal cognitive impairment. However, MMSE scores were lower in all PPA groups compared to controls, likely reflecting language impairments (Osher, Wicklund, Rademaker, Johnson, & Weintraub, 2008). Working memory deficits, as measured by Digit Span Forward and Backward tests on the Wechsler Memory Scale-III (Wechsler, 1997) were evident in the PPA-L and PPA-G, but not PPA-S groups.

Table 1.

Demographic, neuropsychological, and language measures for study participants.

		Neurotypical older adult controls	PPA-G	PPA-L	PPA-S
		N=15	N = 12	N = 10	N = 6
Demographic Measures
Age (years)	M	65.5	61.8	65.8	61.8
	SD	8.6	5.8	4.4	5.5
Education (years)	M	15.6^L	15.8^L	17.6	14.3^L
	SD	2.4	2.2	1.3	2.7
Gender (M:F)		9:6	8:4	5:5	5:1
Neuropsychological Measures ^#
Symptom Duration (months)	M	N/A	40.8	43.5	47.8
	SD		24.0	29.6	23.0
MMSE (30)	M	29.3	23.6^*	26.7^*	24.5^*
	SD	0.8	5.0	2.4	3.8
CDR Sum of Boxes (18)	M	N/A	1.3	0.7	0.7
	SD		1.1	0.5	0.6
Digit Span Forward (WMS-III)	M	7.9	4.8 ^*^,S	4.8 ^*^,S	6.8
	SD	0.7	1.4	1.5	1.0
Digit Span Backward (WMS-III)	M	5.8	3.2^*^,S	3.3^*^,S	4.8
	SD	1.3	0.8	0.8	1.2
Language Measures ^#
WAB-R Aphasia Quotient (100)	M	N/A	81.6	89.2	81.6
	SD		9.0	8.3	12.0
PPVT (36)	M	35.5	32.8^*	34.1	18.3 ^*^,G,L
	SD	0.5	2.9	2.7	7.0
PPT Pictures (52)	M	51.3	48.5^*	50.1^*	42.3 ^*^,G,L
	SD	0.8	3.1	1.3	5.0
NAVS Verb Comprehension Test (%)	M	N/A	95.1	100.0	98.5
	SD		17.1	0.0	3.7
NNB Noun Naming (%)	M	N/A	93.2	94.4	46.9^G,L
	SD		9.4	14.0	35.1
NNB Verb Naming (%)	M	N/A	85.4^L	93.1	68.8
	SD		12.0	11.9	32.6
WAB Repetition Subset (66)	M	65.3	46.3^*	50.1^*	57.5^*
	SD	1.1	12.9	9.8	3.5
NAVS SCT Canonical Sentences (%)	M	N/A	93.9^S	98.7	100.0
	SD		8.1	4.2	0.0
NAVS SCT Non-canonical Sentences (%)	M	N/A	92.1	95.3	95.6
	SD		9.8	7.1	5.4
NAVS SPPT Canonical Sentences (%)	M	N/A	87.9	96.7	98.9
	SD		26.1	6.5	2.7
NAVS SPPT Non-canonical Sentences (%)	M	N/A	65.5	80.0	91.1
	SD		31.0	15.7	11.7
NAT Canonical Sentences (%)	M	100.0	90.6^*	96.0^*	95.6
	SD	0.0	9.2	4.7	8.1
NAT Non-canonical Sentences (%)	M	99.3	49.4 ^*^,L,S	80.7^*	78.9^*
	SD	2.1	18.5	13.1	22.1
Cinderella: Words Per Minute	M	N/A	86.6	96.6	126.2
	SD		25.6	31.8	47.5
Cinderella: Grammatical Sentences (%)	M	N/A	68.3	83.6	76.8
	SD		24.5	8.7	28.6

Open in a new tab

Note: M: Mean; SD: Standard deviation; MMSE: Mini Mental State Examination; CDR: Clinical Dementia Rating; WMS-III: Wechsler Memory Scales, Third Edition; WAB-R: Western Aphasia Battery-Revised; PPVT: Peabody Picture Vocabulary Test; PPT: Pyramids and Palm Trees Test; NAVS: Northwestern Assessment of Verbs and Sentences: SCT: Sentence Comprehension Test; SPPT: Sentence Production Priming Test; NAT: Northwestern Anagram Test;

= significantly lower than controls (p<.05, uncorrected);

^G/L/S

= significantly lower than PPA-G, PPA-L, or PPA-S, respectively (p<.05, uncorrected); N/A = not applicable or not administered;

= Neuropsychological and language measures were available for only 10 controls. The other five controls completed a different neurocognitive test battery.

With respect to language profiles, overall aphasia severity did not differ across the three PPA groups, based on Aphasia Quotients from the Western Aphasia Battery-Revised (WAB-R; Kertesz (2006)). Compared to the other PPA groups, the PPA-S group showed greater impairments in word comprehension (items 157–192 of the Peabody Picture Vocabulary Test (PPVT); Dunn and Dunn (2006)), semantic knowledge/association (Pyramids and Palm Trees Test (PPT), pictures version; Howard and Patterson (1992)), and confrontation naming of nouns/objects (Northwestern Naming Battery (NNB); Thompson and Weintraub (2014)), indicating lexical-semantic deficits. However, comprehension of high-frequency verbs was relatively intact in this group (Verb Comprehension Test from the Northwestern Assessment of Verbs and Sentences (NAVS); Thompson (2011)). In PPA-G and PPA-L, lexical-semantic processing was relatively preserved, but mild impairments as compared to controls were observed in PPA-G for word comprehension (PPVT) and in both groups for semantic knowledge/association (PPT). Repetition of phrases and sentences (measured using a subset of items from the Repetition subtest of the WAB-R) was impaired in all PPA groups.

Participants with PPA-G showed significant impairments in verb/action naming (NNB) and grammatical sentence construction, particularly complex sentences (non-canonical items on the Northwestern Anagram Test (NAT); Thompson, Weintraub, and Mesulam (2012d)), reflecting their agrammatic language profiles. Numerically (although not significantly) poorer performance was also observed in PPA-G as compared to other PPA groups in other grammatical tasks, such as the Sentence Comprehension Test and Sentence Production Priming Test of the NAVS. In narrative language production (Cinderella story re-tell, analyzed using the Northwestern Narrative Language Analysis system; Hsu and Thompson (2018); Thompson et al. (2012a)), the PPA-G group also had the lowest speech rate (words per minute) and proportion of grammatical sentences, although these did not differ significantly from other PPA groups. Grammatical processing was relatively preserved in PPA-L and PPA-S, although both groups showed impairments relative to controls on the NAT.

All participants with PPA included in this study passed a neurological exam that screened for abnormalities in vision and eye movement control (which are often associated with other neurodegenerative diseases, such as progressive supranuclear palsy; Leigh and Zee (2015)), indicating that eye movement control was grossly normal. Notably, one individual with PPA was screened for the study but excluded due to gross abnormalities in oculomotor control.

Structural MRI data

For participants with PPA, structural MRI data were acquired using a 3T Siemens TIM Trio System with a 12 channel birdcage head coil, at the Center for Translational Imaging at Northwestern University. Structural MRI data were unavailable for one participant with PPA-S. Structural T1-weighted images (3D MP-RAGE sequences) were acquired with the following parameters: repetition time (TR) = 2300ms, echo time (TE) = 2.91ms, inversion time (TI) = 900ms, flip angle = 9°, field of view (FOV) = 256mm; 176 slices with a slice thickness of 1.0 mm.

Structural MR images were processed using FreeSurfer (version 5.1.0; http://surfer.nmr.mgh.harvard.edu/) using procedures described previously to derive cortical thickness maps for each participant group (Mesulam et al., 2009b; Mesulam et al., 2012; Rogalski et al., 2011) contrasted with those from a cohort of 35 previously described right-handed, age- and education-matched controls (18 females) (Rogalski et al., 2014). The analyses employed a general linear model (GLM) for each vertex along the cortical surface. False Discovery Rate (FDR; threshold of 0.001) was used to correct for multiple comparisons (Genovese, Lazar, & Nichols, 2002).

All PPA groups showed left-lateralized patterns of cortical atrophy (Figure 1). In PPA-G (n=12), the peak areas of atrophy in the left hemisphere included posterior portions of the inferior, middle, and superior frontal gyri, the temporo-parietal junction, and much of the left lateral temporal lobe, extending to the ventral surface of the temporal lobe. In PPA-L (n=10), the left-hemisphere regions with the greatest atrophy included the superior temporal gyrus and the temporo-parietal junction, with a lesser degree of atrophy in the posterior inferior and middle frontal gyri. In PPA-S (n=5), the focus of atrophy in the left hemisphere was the anterior temporal lobe, extending into the insula and ventral frontal lobe.

Eye-tracking Experiments 1 (E1; Eye movement control) and 2 (E2; Noun comprehension)

E1 and E2 examined eye movement control and noun comprehension, respectively. We expected to find intact performance in all PPA subtypes in E1, but impaired accuracy in PPA-S and delayed eye-movement latencies in all PPA groups in E2.

Stimuli

The visual stimuli for both E1 and E2 consisted of 120 gray-scale drawings of animate (n=10) and inanimate (n=110) objects. The nouns corresponding to the objects had high lexical frequency (mean (SD) frequency per million in the Corpus of Contemporary American English (Davies, 2008): 73.9 (81.2)) and were short (mean (SD) length in syllables: 1.3 (0.5)). All pictures were normed for name agreement as part of a previous study (Mack et al., 2013b). E2 also included auditory word stimuli, which were recorded by a female native English speaker at a normal speech rate and presented at a comfortable listening volume (60–70 dB). Mean word duration was 744 ms (SD = 147; min = 400 ms, max = 1031 ms).

For each of these two experiments, the 120 object pictures were arranged into 30 different visual arrays, which contained an object picture in each of the four corners and a cross at the center (see Figure 2). Each object picture subtended 9.0 degrees of visual angle vertically and 13.1 degrees horizontally. There was no semantic or phonological relationship between the four objects within an array in either experiment. Target and distractor object nouns did not differ with respect to lexical frequency or length (two-tailed t-tests, p’s>.1). The arrays were normed for visual prominence with 10 young adult participants, who were presented with the visual arrays and asked to indicate whether any of the pictures “stood out” from the others. Arrays that were flagged by three or more participants were revised and re-normed. The location of the target picture in the four corners of the visual array was counterbalanced across trials. There were seven trials in which the same target object was used in both E1 and E2, whereas in the remaining trials (23 in each experiment) the target object was not repeated across experiments. E1 and E2 were presented at the beginning of the eye-tracking session, with their order counterbalanced across participants.

Figure 2. — Example stimuli and task procedures from the eye movement control (E1; left) and noun comprehension (E2; right) experiments.

Procedure

Participants were instructed to look at the pictures on the computer screen and to click as quickly as possible when a circle appeared around one of the pictures for E1. For E2, they were instructed to click on the picture corresponding with the word they heard. Both experiments began with a three-item practice session, followed by 30 experimental trials. The experimental stimuli were presented in a randomized order. Each trial (see Figure 2) began with the presentation of a cross and participants clicked on the cross, which remained on the screen for an additional 1500 ms. Then, the four-object visual array was presented. After an additional 500 ms, either a short (30 ms) pure-tone beep was presented auditorily, and simultaneously, a red circle appeared around one of the object pictures for E1 or the target auditory word was presented for E2. Participants then clicked on the circled picture (E1) or the picture that matched the auditory word (E2), with a 5000 ms response limit.

Eye movement data were acquired using an Applied Sciences Laboratories (ASL) EYE-TRAC 6000 desk-mounted eye-tracker, recording data from one eye (typically the left eye, unless tracking of the left eye was unsuccessful) at 60 Hz. The eye-tracker was calibrated before each experiment, following a nine-point static calibration procedure. Testing took place in a dark room, and participants’ heads were stabilized using a chin rest to prevent motion artefacts. Participants heard auditory stimuli over speakers and viewed pictures on a 16” by 10” computer monitor (approximately 20” from the participant).

Data Analysis

Accuracy and eye movement data were analyzed using mixed-effects regression in R (Bates, Maechler, Bolker, & Walker, 2015; R Core Team, 2017). Linear models were estimated using the lmer function and logistic models with the glmer function. Continuous predictors were centered and categorical predictors were simple-coded, i.e., each level of the predictor was compared to a reference level, with the intercept being the grand mean (UCLA: Statistical Consulting Group (2011)). The advantage of this coding method is that pairwise comparisons between all levels can be obtained by shifting the reference level, without affecting other predictors. In this paper, we report pairwise comparisons between all levels of categorical variables (extracted using the function glht in R package multcomp; Hothorn, Bretz, and Westfall (2008)). Unless otherwise stated, the maximal converging random effects structure was included. Z and p-values are reported.

Eye movement data were preprocessed into a series of fixations using Eyenal (ASL Laboratories). Fixations were defined as gazes within 1 degree of visual angle that lasted at least six tracking samples (~ 100 ms). Each fixation was assigned to an area of interest within the visual array (i.e., target picture, distractor pictures, “other” for fixations that fell outside the four pictures). Subsequent eye movement data analysis included only trials in which the participant had responded correctly.

In E1, accuracy was at ceiling so it was not analyzed statistically. To analyze the eye movement data, we computed the latency of the first fixation to the target picture for each trial. This measure was selected in order to facilitate comparison with previous studies of visually-guided saccade tasks in PPA, which have reported saccadic onset latencies (Boxer et al., 2006; Burrell et al., 2012; Garbutt et al., 2008). Only fixations starting more than 100 ms after the appearance of the critical stimulus (the circle) were included, as it typically takes approximately 200 ms to execute a saccade (Munoz, Broughton, Goldring, & Armstrong, 1998). The inverse-transformed first-fixation latencies were entered as dependent variables in mixed-effects linear regression analyses. The fixed predictors included in these analyses were participant group, target location (i.e., the location of the target picture in the four corners of the visual array), and stimulus onset fixation location (i.e., the location of the fixation (if any) taking place between 0 and 100 ms after stimulus onset: target picture, distractor picture, cross, other on-screen area, or none). The latter two variables were included in order to account for possible effects of visuospatial factors on first fixation latencies. Random intercepts for participant and item were included. For visualization purposes, the cumulative likelihood of a target fixation was plotted in 50 ms bins time-locked to stimulus onset.

For E2, accuracy was analyzed using a mixed-effects logistic regression model containing a main effect of participant group and random by-participant and by-item intercepts.

The eye movement data were analyzed using a time-course approach, similar to many previous studies of single-world comprehension using visual-world eye-tracking (e.g., Farris-Trimble and McMurray (2013)). Based on the results of E1 (which indicated that the earliest stimulus-driven fixations occurred approximately 300 ms after stimulus onset), word onset was shifted forwards 300 ms to adjust for the time needed to generate a fixation. Thus, we used a study- and population-specific value for temporally shifting the data. The data were aggregated into 50 ms bins time-locked to the (adjusted) onset of the auditory word. For each participant and time bin, we computed (a) the number of fixations to the target picture vs. the distractor pictures (target advantage) and (b) the number of fixations to the pictures vs. other on-screen fixations (workspace advantage). To visualize the data, for target advantage we plotted the proportion of target fixations, out of all fixations to pictures, whereas for workspace advantage we plotted the proportion of fixations to pictures, out of all on-screen fixations.

Target advantage is a standard eye-tracking variable, which measures the extent to which the linguistic input drives selective looks toward the target picture. Workspace advantage reflects the overall proportion of fixations to the pictures, i.e., the workspace of the task, as compared to other on-screen fixations. This measure may reflect the task’s cognitive demands. For example, Barr (2008; p. 471) describes a scenario in which “participants are slower to move off a central fixation cross in one condition than another” due to differing cognitive demands. This difference would be captured by a workspace advantage measure, but not a target advantage measure. In some previous eye-tracking studies, reduced inspection of the visual stimuli has been observed in listeners with aphasia as compared to controls (e.g., the tendency to fixate the left-most picture throughout sentence presentation in a sentence-picture matching task; Mack and Thompson (2017)), which may reflect the increased cognitive demands of the tasks for these listeners. Therefore, we included workspace advantage as a measure of cognitive demand.

For E2, two temporal windows for data analysis were selected on the basis of the linguistic stimuli (word durations) and the grand mean data (i.e., the data averaged across all participants/conditions), a data-driven approach suggested by Barr (2008). The Noun window spanned from noun onset (0 ms) to mean noun offset (750 ms). The grand mean target advantage data were inspected to identify the time-bins containing a rise in the proportion of target fixations (> 0.01 (1%) per bin). This rise began at noun onset and ended 1150 ms later. Therefore we defined a second Offset window which spanned from mean noun offset (750 ms) to 1150 ms. Within each temporal window, mixed-effects logistic regression analyses were conducted, in which the dependent variable was the target advantage (in one set of models) or the workspace advantage (in a second set of models). All models contained main effects of participant group, time bin, and their interaction.

Results

For E1, accuracy was at ceiling, with only one incorrect response recorded. For E2, accuracy data are summarized in Table 2. Control participants had greater accuracy than those with PPA-S (z = −3.487, p < 0.001) and PPA-G (z = −2.167, p = 0.030). Accuracy was also lower in PPA-S as compared to PPA-G (z = 2.306, p = 0.021) and PPA-L (z = 2.821, p = 0.005). No other significant group differences were observed (p’s > 0.05).

Table 2.

Accuracy by participant group and experiment.

	Control		PPA-G		PPA-L		PPA-S
	M	SD	M	SD	M	SD	M	SD
Eye Movement Control (E1)	100%	0%	100%	0%	100%	0%	100%	0%
Noun Comprehension (E2)	100%	1%	97%^*	4%	99%	2%	90%^*^,G,L	8%
Argument Access (E3)
Restrictive	98%	3%	97%	5%	99%	2%	93%^*^,G,L	9%
Unrestrictive	97%	5%	89%^*	9%	95%	4%	88%^*	10%
Argument Selection (E4)
Restrictive	89%	6%	77%^*	17%	85%	11%	83%	8%

Open in a new tab

Note: M = Mean; SD = standard deviation;

^{*, G, L}

= significantly impaired (p<.05) relative to Controls, PPA-G, and PPA-L, respectively.

The eye movement data for E1 are plotted in Figure 3. Mean (SD) first fixation latencies to the target picture, in seconds, were as follows: Controls: 0.474 (0.089); PPA-G: 0.528 (0.095); PPA-L: 0.467 (0.056); PPA-S: 0.440 (0.087). The statistical model of first fixation latencies is summarized in Table 3. Overall, first fixation latencies were reduced when the target picture appeared in the bottom right of the array, vs. the top left of the array (z = −2.258, p < 0.05). Latencies were also lower for trials in which the participant was fixating the cross at stimulus onset, as compared to trials in which the participant was not making a fixation at that point (z = −2.839, p < 0.01). With respect to effects of group, fixation latencies were lower in PPA-S as compared to PPA-G (z = 2.226, p < 0.05). No other significant differences were observed.

Table 3.

Statistical models of eye movement data in the eye movement control task (E1).

	z	p
Intercept	44.404	<0.001
Target Location
Bottom R vs. Bottom L	1.862	0.063
Top L vs. Bottom L	−0.335	0.737
Top R vs. Bottom L	0.713	0.476
Top L vs. Bottom R	−2.258	0.024
Top R vs. Bottom R	−1.22	0.222
Top R vs. Top L	1.089	0.276
Stimulus Onset Fixation Location
None vs. Cross	−2.839	0.005
Target vs. Cross	−1.065	0.287
Dis vs. Cross	−0.795	0.427
Other vs. Cross	−1.027	0.305
Target vs. None	0.513	0.608
Dis vs. None	1.705	0.088
Other vs. None	1.109	0.267
Dis vs. Target	0.537	0.591
Other vs. Target	0.276	0.783
Other vs. Dis	−0.304	0.761
Group
PPA-G vs. Control	−1.158	0.247
PPA-L vs. Control	0.47	0.638
PPA-S vs. Control	1.336	0.182
PPA-L vs. PPA-G	1.533	0.125
PPA-S vs. PPA-G	2.226	0.026
PPA-S vs. PPA-L	0.889	0.374

Open in a new tab

Note: L = left; R = right; Cross = cross at center of screen; Target = target picture; Dis = distractor picture; Other = other on-screen region. For pairwise group comparisons (e.g., PPA-G vs. Control), the reference level is listed second. Significant effects (p < 0.05), other than effects at the intercept, are shaded.

Eye movement data for E2 are plotted in Figure 4; statistical analyses are summarized in Table 4. Starting with the target advantage results, an increase over time occurred during both time windows (Noun window: z = 15.538, p < 0.001; Offset window: z = 7.389, p < 0.001). During the Noun window, the rate of increase was greater in controls than in PPA-G (z = −3.370, p = 0.001) and PPA-S (z = −4.284, p < 0.001), and in PPA-L than in PPA-S (z = 2.493, p < 0.05). During the Offset window, the control group had an overall greater target advantage than PPA-S (z = −2.014, p < 0.05). For the workspace advantage results, during the Noun window there was an overall increase (z = 4.612, p < 0.001) and the PPA-S group also showed a greater workspace advantage than the three other groups (Control: z = 2.044, p < 0.05; PPA-G: z = −2.127, p < 0.05; PPA-L: z = −2.172, p < 0.05). No other significant differences were found.

Table 4.

Statistical models of eye movement data in the noun comprehension task (E2).

	Target Advantage				Workspace Advantage
	Noun		Offset		Noun		Offset
	z	p	z	p	z	p	z	p
Intercept	−5.605	<0.001	14.032	<0.001	16.398	<0.001	11.735	<0.001
Bin	15.538	<0.001	7.389	<0.001	4.612	<0.001	1.912	0.056
PPA-G vs. Control	−0.912	0.362	−1.092	0.275	−0.205	0.837	−0.047	0.962
PPA-L vs. Control	−1.615	0.106	−1.018	0.309	−0.328	0.743	0.046	0.964
PPA-S vs. Control	−3.621	<0.001	−2.014	0.044	2.044	0.041	0.641	0.522
PPA-G vs. PPA-L	0.713	0.476	−0.025	0.980	0.125	0.901	−0.087	0.931
PPA-G vs. PPA-S	2.781	0.005	1.090	0.276	−2.127	0.033	−0.656	0.512
PPA-L vs. PPA-S	2.100	0.036	1.080	0.280	−2.172	0.030	−0.567	0.571
PPA-G vs. Control * Bin	−3.370	0.001	0.953	0.341	0.594	0.553	−0.132	0.895
PPA-L vs. Control * Bin	−1.892	0.058	0.626	0.531	0.854	0.393	−1.010	0.313
PPA-S vs. Control * Bin	−4.284	<0.001	0.781	0.435	−0.707	0.479	0.514	0.607
PPA-G vs. PPA-L * Bin	−1.271	0.204	0.286	0.775	−0.270	0.788	0.869	0.385
PPA-G vs. PPA-S * Bin	1.462	0.144	0.041	0.968	1.124	0.261	−0.604	0.546
PPA-L vs. PPA-S * Bin	2.493	0.013	−0.211	0.833	1.313	0.189	−1.254	0.210

Open in a new tab

Note: For pairwise group comparisons (e.g., PPA-G vs. Control), the reference level is listed second. Significant effects (p < 0.05), other than effects at the intercept, are shaded.

Discussion: Experiments 1 and 2

Results of E1 indicated that first-fixation latencies did not differ between participants with PPA and controls in our eye movement control task. This finding is consistent with two previous studies that found no differences in saccadic onset latencies between PPA patients and controls (Boxer et al. (2006); Garbutt et al. (2008), though see Burrell et al. (2012)).

In contrast to the results of E1, the eye movement patterns noted in E2 indicated noun comprehension impairments in PPA-S and PPA-G. Accuracy was impaired in PPA-S as compared to all other groups, and in PPA-G as compared to neurotypical controls. With respect to eye-movement patterns in correct trials, delayed lexical access was observed in PPA-S and PPA-G, as shown by a slower increase in target advantage during the Noun window. In PPA-S, an overall reduction in target advantage persisted into the Offset window, where normal-like eye movement patterns were observed in PPA-G and PPA-L. These findings are consistent with previous studies which showed substantially delayed lexical access in PPA-S and subtle delays in PPA-G (Mack et al., 2013a; Seckin et al., 2016; Thompson et al., 2012b). In contrast with some previous studies (Mack et al., 2013a; Thompson et al., 2012b), we did not find significant differences in lexical processing between PPA-L and controls. However, we note that the rise in target advantage was numerically delayed for PPA-L in the Noun window, suggesting that subtle delays may be present albeit not detected statistically by our analyses.

Collectively, the results of E1 and E2 indicate that abnormal response patterns found in linguistic tasks using a visual world paradigm cannot be attributed to impaired eye movements. Thus, group differences found in E2 (as well as in E3 and E4; see below) can be attributed to language deficits, rather than oculomotor control.

Experiments 3 (E3; Argument Access) and 4 (E4; Argument Selection)

E3 and E4 investigated verb-argument integration, using the paradigms from Mack et al. (2013b). Participants listened to sentences containing semantically restrictive verbs which could be integrated with only one object in the visual array (e.g., close the jar, but not close the stick/pencil/plate) and unrestrictive verbs which could be integrated with any of the four objects (e.g., break the jar/stick/pencil/plate). E3 used a picture-word verification task to probe argument access. Following previous work, we expected that neurotypical adults would use verb meaning to facilitate argument access, resulting in an earlier increase in looks to the target picture in the restrictive vs. unrestrictive conditions (Altmann & Kamide, 1999; Boland, 2005; Borovsky et al., 2012; Kamide et al., 2003; Kukona et al., 2011; Milburn et al., 2016; Staub et al., 2012). E4 used a sentence completion task to probe argument selection. We expected neurotypical adults to rapidly use verb meaning, in the restrictive condition, to select an argument to complete the sentence. We hypothesized that listeners with PPA-G, like those with stroke-induced agrammatic aphasia (Mack et al., 2013b), would show intact access to argument structure information, resulting in normal effects of verb meaning in E3, but impaired argument selection in E4. For both PPA-L and PPA-S, we evaluated competing hypotheses. The first predicted intact verb-argument integration, consistent with production patterns in these subtypes (Thompson et al., 2012a). The other predicted impairments in one or both groups due to the typical neurological distribution of neurodegeneration: left TPJ in PPA-L, associated with verb-argument structure access and thematic processing, and the left ATL in PPA-S, associated with semantic combinatory processes.