Primary progressive aphasia (PPA) is a degenerative disease affecting language while leaving other cognitive facilities relatively unscathed (Mesulam et al., 2012). Three distinct subtypes of the disorder have been identified with different disease etiology and progression, as well as different profiles of spared and impaired language function. The agrammatic subtype of PPA (PPA-G) is characterized by the production of ungrammatical sentences and impaired comprehension of syntactically complex sentences, with spared single-word comprehension. In contrast, the logopenic and semantic subtypes (PPA-L, PPA-S, respectively) do not show grammatical impairments in production or comprehension. Rather PPA-L is characterized by impaired naming and word-finding with spared single-word comprehension, and PPA-S is characterized by impaired naming and single-word comprehension but spared speech production (Mesulam et al., 2009; Gorno-Tempini et al., 2011).

In this paper, we focus on the on-line mechanisms underlying the comprehension of complex noncanonical sentences in people with agrammatic PPA. Noncanonical sentences subvert the dominant agent-verb-theme order in English, such that the theme precedes the verb in these structures. The agent (the do-er of the verb’s action) and theme (the participant affected by the verb) refers to thematic roles that specify the semantic relation between the verb and its arguments. For example, in an object-relative clause structure [e.g., The sandwich that the girl ate (gap) was tasty], the theme (the sandwich) has been displaced from the object position following the verb [at the (gap) site] to a position preceding the verb (ate) and the agent (the girl). This displacement creates a filler-gap dependency, in which the displaced element (the sandwich) is the filler for the gap. In a subject-relative clause [The girl that (gap) ate the sandwich ran away], the agent (the girl) has been extracted from the subject position of the relative clause [i.e., the (gap) position], resulting in a preserved canonical agent-verb-theme order.

In sentences with noncanonical word order, linking the filler to the gap is essential for correct assignment of thematic roles, which in turn, is essential for comprehending who did what to whom within a sentence. While this process depends on syntax, many sentences can be understood even in the absence of syntactic processing. Thus in sentences with “non-reversible” relative clauses such as The ball that the boy is kicking is round, the thematic relations between the verb (kicking) and its arguments can be puzzled out based on real-world knowledge. In contrast, thematic relations in “reversible” relative clauses, where either participant could plausibly be linked to either thematic role, are difficult to understand without syntax, as real-world knowledge does not favor one interpretation over another. For example, the eagle in the sentence The eagle that was chased by the hawk was faster could plausibly be the agent or the theme of the relative clause verb (chased), likewise for the hawk. However, the syntax of this particular sentence indicates that the eagle (the subject) must be the theme and the hawk must be the agent. Without syntax, misunderstandings of such sentences would be common.

Sentences with reversible clauses, therefore, serve as a useful indicator of syntactic comprehension deficits. Indeed, deficient comprehension of reversible noncanonical structures is often found in agrammatic aphasia after a stroke (Caramazza and Zurif, 1976; Grodzinsky, 1986; Love et al., 2008; Thompson et al., 2013). In agrammatic PPA as well, the comprehension of reversible noncanonical constructions is consistently reported to be impaired, across a variety of experimental tasks, including sentence-picture matching (Wilson et al., 2010; Charles et al., 2013; Thompson et al., 2013; but see Zimmerer et al., 2014), sentence-picture verification (Kinno et al., 2017), and visual sentence decision (deciding if the agent is male or female; Cooke et al., 2003).

However, the results from these studies have all been based on off-line methods, in which the measurement occurs after the end of the sentence. Such off-line methods are temporally insensitive and allow for strategic, metalinguistic processing, as well as opportunities for self-correction and reflection, and so do not necessarily reflect the same (or even similar) processes that underlie automatic syntactic processing as a sentence unfolds over time (Swinney and Osterhout, 1990). While prior results may therefore speak to impaired sentence comprehension in agrammatic PPA, they do not reveal when or how it may have gone wrong.

On-line methods such as eye-tracking or event-related potential techniques are capable of measuring automatic sentence comprehension in real-time but have not been previously brought to bear on the comprehension of noncanonical structures in agrammatic PPA. On-line methods have been used to investigate language processing in agrammatic stroke aphasia, and given the similar profiles of off-line performance in the two disorders, theories of agrammatism developed in the stroke aphasia literature may well provide insight into the nature of on-line sentence comprehension deficits in agrammatic PPA.

There are many different perspectives on these comprehension deficits in aphasia (for review, see Kolk and Weijts, 1996; Dickey and Thompson, 2004; Patil et al., 2016). On one broad view, comprehension deficits in aphasia reflect a breakdown in grammatical knowledge. One prominent view from this perspective, the trace deletion hypothesis, posits that impaired comprehension of filler-gap structures reflects an inability to create gaps. Without the gap, the filler cannot be assigned a thematic role, and so comprehenders rely on heuristic mechanisms to interpret the role of the filler (Grodzinsky, 1986, 1995, 2000). However, tests of this hypothesis have been made almost exclusively with off-line measures such as sentence-picture matching, that do not measure how the sentence was processed before the response (Grodzinsky, 2000).

Alternative hypotheses focus on aspects of real-time processing during the comprehension of a sentence. One set of hypotheses suggest that the timing of key processes is disrupted in agrammatic aphasia. On these views, grammatical knowledge is posited to remain intact, but processing is slowed to the extent that the comprehension of filler-gap constructions is impaired. According to the slow-syntax hypothesis, the brain damage that results in agrammatic aphasia leads to a reduction in processing resources, such that syntactic structure building is slowed during real-time sentence comprehension (Burkhardt et al., 2003; Avrutin, 2006). To explain why the delayed but otherwise intact structure building does not ultimately succeed given enough time, on this view the slowed syntactic processing leads to interference from other processes that lead comprehension astray. On-line evidence from a cross-modal interference task suggests delayed processing of a filler at a gap during comprehension of noncanonical filler-gap constructions in agrammatic aphasia (Burkhardt et al., 2008).

According to the lexical slow-rise hypothesis, syntactic processing is neither impaired nor slowed in agrammatism. Rather, evidence from lexical priming studies suggests that lexical activation peaks later than normal in agrammatic aphasia (Swinney et al., 1996). This delayed lexical activation, therefore, feeds rapid automatic syntactic processing too slowly during normal-rate speech (roughly four to six syllables per second), leading to breakdowns in comprehension. Studies of sentence comprehension using cross-modal lexical priming report delayed priming effects both for the initial activation of a word and for the reactivation of the filler at a gap (Love et al., 2008; Ferrill et al., 2012). Also consistent with this view, presenting auditory sentences with a slowed input rate (less than four syllables per second) results in on-time re-activation of the filler at the gap and improved off-line comprehension in agrammatic listeners (Love et al., 2008).

On another view, difficulty comprehending complex syntactic structures in agrammatic aphasia reflects deficient thematic integration of verb arguments into the syntax (Thompson and Choy, 2009). This results, first, from failure to predictively assign the agent argument role to the first noun encountered (i.e., the filler) in noncanonical sentences, as do unimpaired listeners (Mack and Thompson, 2017) and, second, to re-assign the theme role to that argument when the gap is encountered. The results of several studies using eye-tracking paradigms with stroke aphasic agrammatic participants show eye movement patterns consistent with both thematic prediction and integration impairments (Meyer et al., 2012; Mack et al., 2013, 2016; Hanne et al., 2015). In contrast, the processes underlying the creation of the gaps in the syntax and reactivation of the filler at the gap position have been found to proceed normally (Dickey et al., 2007; Dickey and Thompson, 2009; Thompson and Choy, 2009).

In sum, while these hypotheses consistently predict off-line comprehension deficits in agrammatic aphasia for non-canonical structures, the time-course and nature of the predicted on-line deficits vary. Examining these predictions in agrammatic PPA therefore requires an on-line method with a fine-grained temporal resolution. Here, we use a visual world eye-tracking paradigm with reversible object-relative and subject-relative sentences presented auditorily at a normal speech rate. The paradigm and materials were adapted from those previously used in a study of agrammatic stroke aphasia, which reported on-time gap filling in sentences with object-relative clauses (Dickey et al., 2007; Dickey and Thompson, 2009).

According to the trace-deletion hypothesis, no evidence of reactivation of the filler at the gap site should be seen. According to the slow-syntax and lexical slow-rise hypotheses, gap-filling in sentences presented at a typical rate of speech should be qualitatively normal but delayed. According to the thematic deficit hypothesis, reactivation at the gap location in a noncanonical structure should be on-time in agrammatic PPA, but thematic role processes should be abnormal, both for thematic prediction (before the gap) and for thematic integration (at or downstream from the gap).

Finally, we include PPA-L as a comparison group. Offline comprehension deficits have been observed in PPA-L for noncanonical sentences, though such deficits may reflect verbal working memory deficits rather than grammatical deficits per se, and are not typically as severe as in PPA-G (Thompson and Mack, 2014). We are not aware of any on-line studies of noncanonical sentence comprehension in PPA-L.

We used a visual world eye-tracking paradigm to examine the real-time processing and comprehension of complex sentences in individuals with PPA and healthy controls. Participants listened to short three-sentence stories about two participants (e.g., bride, groom), and then heard a final sentence asking them to point at one of the participants by clicking on the target picture in a visual array of four pictures. The final sentence was structured with either a subject-relative clause [Point to the one (that_— tickled the groom) in the mall] or an object-relative clause [Point to the one (that the bride tickled_—) in the mall]. We measured response accuracy as well as eye-movements during the target sentences to examine how individuals with PPA process the structural gaps (indicated by “_—” in the examples) in these relative clauses in real-time.

Concerning response accuracy, both patient groups showed poorer performance compared to healthy controls for both sentence types, as expected, with a more severe impairment for the PPA-G group, consistent with their performance on standard offline measures (i.e., comprehension and production accuracy at canonical sentences from the NAVS SPPT and the NAT, though this only reached significance for sentence production on the NAT; Table 1). This finding is also similar to that of a prior study showing (numerically) reduced accuracy of subject-relative clause comprehension in both PPA-G and PPA-L (Thompson et al., 2013). For the object-relative sentences, the PPA-G group also showed reduced response accuracy compared to the control participants, consistent with prior findings of noncanonical sentence comprehension deficits in PPA-G (Cooke et al., 2003; Wilson et al., 2010; Charles et al., 2013; Thompson et al., 2013; Kinno et al., 2017; but see Zimmerer et al., 2014). The PPA-L group also had impaired response accuracy relative to the control participants and were not statistically different from the PPA-G group.

Turning to the eye movement patterns, for the sentences with subject-relative clauses, the healthy control group steadily increased their rate of gazing towards the correct target immediately after the gap in the sentence, before it leveled off in the sentence-end region. The eye-movement patterns for the PPA-G and PPA-L groups were not statistically different from those of the control group, either in the post-gap region or the sentence-end region, particularly given the correction for multiple comparisons.

For the object-relative clause sentences, the healthy controls had a high (but non-increasing) rate of looks to the filler of the gap (e.g., the groom) in the post-gap region (e.g., in the mall). The PPA-L group was not different from the controls in this region, and also showed a high but steady rate of looks to the filler. In contrast, the PPA-G group showed a significantly decreasing rate of looks to the target, relative to both of the other groups. Additional spline regression analysis revealed that the slope changed and began to decrease at the gap position.

We argue that the thematic integration deficit account of agrammatic aphasia accounts for these results (Thompson and Choy, 2009). The finding that changes in the eye-gaze slopes began at the gap site is consistent with the prediction of intact filler-gap structure building. Also, the decreasing looks to the target in the post-gap region for the agrammatic comprehenders are consistent with the deficient interpretation of the filler’s thematic role at this point. That is, we take the change in processing that begins at the gap position as evidence that the gap itself was structured. If the gap had not been created, then it is not clear what kind of process could have led to the sudden change that we observed at that point. These results are consistent with prior findings from agrammatic stroke aphasia with this paradigm, which report normal gap-filling in the face of a reduced advantage for looks to the theme (relative to the agent) after the gap in stroke agrammatism for object-extracted wh-questions and cleft structures (Dickey et al., 2007; Dickey and Thompson, 2009). The correspondence in the patterns across populations is particularly clear when comparing the gazes over time to the theme. Just as for the current results for agrammatic PPA, the agrammatic stroke results from those studies also indicate a sharp downward turn for theme looks after the gap in object-relative sentences (Dickey and Thompson, 2009, Figure 2B).

A second aspect of the thematic integration deficit account that is supported by the current data is that predictions regarding the thematic role of the filler are impaired in agrammatic comprehension. A broader view of the eye-movement patterns before the gap in the object-relative sentences speak to this issue. During the relative clause subject (“the bride”), healthy control participants increased their gazes both to the filler (groom) and to the relative clause subject (bride, Figure 3A). This pattern is consistent with an agent-first strategy, given the potential for both animate participants (groom, bride) to be agents. Note that correctly predicting a theme role for the filler at this point would be unexpected—the sentence does not yet contain any information about which participant is linked to which thematic role. However, this pattern changed during the relative clause verb (e.g., “was tickling”), where looks to the agent (bride) decreased and looks to the theme (groom) continued to rise. This indicates that during the verb, healthy listeners correctly assign the agent role to the relative clause subject, leaving the correct prediction that, when encountered, the theme role would be assigned to the filler. This reflects a reanalysis process, whereby the initial assumption of an agent role is changed to an expected theme role for the filler. Confirmation of the filler’s theme role at the gap is also consistent with the finding of a steady rate of target-looks in the post-gap region.

In contrast, the participants with PPA showed a different pattern than the controls. While they are hearing the relative clause subject (e.g., “the bride”), looks to the corresponding picture (bride) increased, but looks to the correct filler (e.g., “groom”) decreased (Figures 3B,C). This indicates that, rather than using an agent-first strategy, correctly predicting that either participant could be the agent, their looks corresponded directly with the word heard at that point in the sentence. This pattern is consistent with a thematic prediction deficit. Notably, this pattern was found in the eye movements of both PPA-G and PPA-L participants, suggesting that this aspect of processing is not specific to agrammatism. This pattern also diverges from a prior finding in agrammatic stroke aphasia, where it was observed that looks to the theme and agent were similar (i.e., there was no preference) during the subject region of object wh-questions (Dickey et al., 2007; Dickey and Thompson, 2009).

However, in the subsequent relative clause verb region (e.g., “was tickling”), both PPA groups showed increased looks to the filler (groom) at a rate that was similar to that of the controls, reflecting thematic reanalysis in anticipation of a downstream object gap. However, after the gap, looks to the filler decreased in the agrammatic group, indicating erroneous thematic integration. Notably, the PPA-L group did not show this decrease. Rather they showed sustained looks to the filler, like the healthy controls. These findings indicate that whereas participants with PPA-G show impaired thematic prediction and integration, those with PPA-L evince difficulty only with thematic prediction.

One potential issue for this interpretation concerns a processing strategy that may be possible for our materials. That is, once the subject/agent has been identified, the preceding story and the meaning of the verb enable the theme to be correctly predicted as the target, even if the sentence structure is not processed beyond that point. Thus, looks to the target should continue to increase, even if a gap is not created. Such a prediction does seem to be made during the verb region—looks to the theme (the correct target) increase in this region for all three groups (although this prediction did not start as soon in PPA as in the controls, it was nevertheless apparent during the verb). However, we think this alternative account does not explain our results. First, evidence from many other studies indicates that healthy comprehenders reactivate the filler at the gap position in these structures (Swinney and Fodor, 1989; Love and Swinney, 1996; Nicol et al., 2006; Love, 2007; and references therein). This is interpreted as a reflex of automatic language processing. Thus, we expect that the healthy control participants are correctly structuring the sentences, even though a similar pattern of target looks could be expected for our materials if no structural gap were created. The participants with logopenic PPA did not differ from controls in this region, so we also expect that their processing is unimpaired concerning the gap. However, despite apparently beginning to correctly predict the target during the verb, the participants with agrammatic PPA (who are predicted to have trouble structuring the gap on some views) began to look away from the correct target following the gap position—indicating that some element within the sentence disrupted their processing. This alternative account, which predicts continued looks to the target in the absence of a gap, therefore does not predict the pattern that we observed.

Instead, we argue that the present finding indicates that people with PPA do show evidence of gap-filling in object-relative sentences and that they do not show an agent-first strategy during online sentence comprehension. Thus our evidence does not support the predictions of the trace-deletion hypothesis (Grodzinsky, 1986, 1995). Likewise, the finding that gap-filling was on-time is contrary to the predictions of the slow-syntax and slow-rise hypotheses (Burkhardt et al., 2008; Love et al., 2008). In the current study, the speech rate of the auditory sentences was within the normal range, but participants show evidence of gap-filling at the gap site, as do healthy listeners, suggesting that this factor does not underlie sentence comprehension difficulty.

In conclusion, the eye movement patterns found in the present study suggest that the real-time processing of complex syntactic structures is impaired in agrammatic PPA, consistent with the hypothesis of deficient thematic integration. This includes deficits in both thematic prediction and post-verbal thematic integration, in the face of normal-like gap-filling processes during object-relative clause computation. These abnormal processing patterns help to explain the source of comprehension failure in PPA-G patients, and are in keeping with deficit patterns seen in stroke-induced agrammatic aphasia during the processing of syntactically complex structures. The novel finding that patients with PPA-L also show thematic prediction impairments may, at least in part, explain their difficulty in comprehending complex grammatical constructions. The clinical profiles of agrammatic and logopenic PPA, therefore, appear to overlap in some, though not all, aspects of syntactic processing. The implications of this finding for diagnosis and treatment of these different subtypes of PPA are left for future investigations.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.