Disintegration at the Syntax-Semantics Interface in Prodromal Alzheimer’s Disease: New Evidence from Complex Sentence Anaphora in Amnestic Mild Cognitive Impairment (aMCI)

Abstract

Although diverse language deficits have been widely observed in prodromal Alzheimer’s disease (AD), the underlying nature of such deficits and their explanation remains opaque. Consequently, both clinical applications and brain-language models are not well-defined. In this paper we report results from two experiments which test language production in a group of individuals with amnestic Mild Cognitive Impairment (aMCI) in contrast to healthy aging and healthy young. The experiments apply factorial designs informed by linguistic analysis to test two forms of complex sentences involving anaphora (relations between pronouns and their antecedents). Results show that aMCI individuals differentiate forms of anaphora depending on sentence structure, with selective impairment of sentences which involve construal with reference to context (anaphoric coreference). We argue that aMCI individuals maintain core structural knowledge while evidencing deficiency in syntax-semantics integration, thus locating the source of the deficit in the language-thought interface of the Language Faculty.

1. Introduction

Early detection of Alzheimer’s disease (AD) during its long prodromal period is of paramount importance clinically. Understanding how language changes during this period may provide a deeper understanding of early progression of AD and strengthen early detection when treatments may most effectively halt disease progression (Golde et al., 2018; Selkoe, 2021). At the same time, such investigation can inform our understanding of language as a cognitive system and crucially advance our knowledge of fundamental brain-language relations.

Study of the neuroscientific foundations of the progression of AD, including its prodromal preclinical periods, has advanced widely (Braak & Braak, 1991, 1996; Braak et al., 2011; Dickerson et al., 2008, 2011, 2013; Dickerson & Walk, 2012; Putcha et al., 2019; Karaman et al., 2022; Sperling et al., 2011). Behaviorally and clinically, Mild Cognitive Impairment (MCI) is defined as a prodromal stage in which there is evidence of cognitive decline that often continues to advance to dementia. Primary difficulty with memory during this period, referred to as MCI amnestic subtype (aMCI), is most likely to be attributable to Alzheimer’s disease (AD) pathology (Petersen, 2003, 2004, Wolk et al., 2011; Vos et al., 2013, Rye et al., 2022). However, while memory deficits are primary, there are also decades of research revealing that deficits in other cognitive domains, including language, also occur during prodromal and clinical periods of AD (see Sherman et al., 2021 for recent review). Within the research on language, there are differing views as to the degree and manner in which linguistic factors, syntactic, semantic, or both (Chapin et al., 2022; Kemper et al., 1993; Ahmed et al., 2013; Joyal et al., 2020) contribute to this decline.

In spite of advancing evidence for linguistic decline in early and prodromal AD, the variation in research data and methods, including a relative lack of hypothesis driven experimental research, significantly challenges neuroscientific investigations which would explain language deterioration and lay a foundation for bridging brain and behavior in a precise way. (See for example Roark et al., 2007 for evaluation of methods seeking to characterize deterioration in syntactic complexity).

1.1. This paper

In this paper we report results of experimental research designed to expand the assessment of language in aMCI, specifically targeting anaphora in complex sentences.¹ By concentrating on anaphora, an integral component of language computation which critically integrates syntax and semantics, we seek a more precise explication of the nature of linguistic deterioration in prodromal stages of AD. We report results of two experimental tests of language production. The first experiment tests two types of anaphora in sentences with adverbial subordinate clauses (ASC) (Table 1), and the second does so in coordinate clauses (CS) (Table 2). We assess the relationship between language production and memory performance.We supplement our report with supplementary materials where relevant, denoted as [SM].

Table 1.

Experimental Sentences by Factorial Design: Adverbial Subordinate Clauses (ASC).^1,2,3

Semantically Plausible (+SP)
RB2	F	A33	The dressmaker mended the costume when she encountered the actress.
		B2	The electrician fixed the light switch when he visited the tenant.
	B	A1	He triggered the alarm when the ambassador saw the intruder.
		B3	She answered the telephone when the receptionist heard the caller.
LB	F	A2	When the professor understood the scholar, he explained the theory.
		B5	When the announcer introduced the visitor, she read the greeting.
	B	A4	When she remembered the orphan, the composer wrote the sonata.
		B6	When he summoned the tourist, the gentleman indicated the road.
Semantically Implausible (−SP)
RB	F	A2	The programmer concealed the pen when she contacted the candidate.
		B3	The navigator imagined the plant when he offended the queen.
	B	A3	He labeled the painting when the astronomer pushed the resident.
		B5	She measured the artifact when the storyteller paid the boyfriend.
LB	F	A6	When the employee discouraged the artist, she exchanged the bucket.
		B4	When the disciple abandoned the teenager, he dropped the coffee.
	B	A5	When she respected the hunter, the drummer buried the replica.
		B6	When he misled the neighbor, the deputy inherited the ball.

¹Bold sentences block anaphoric construal by Binding Theory Principle C.

²RB: Right-Branching; LB: Left-Branching; F: Forward; B: Backward Pronoun.

³A/B: item number in the two sentence batteries.

Table 2.

Experimental Sentences by Factorial Design: Coordinate Sentences (CS) with NP, Pronoun or Elided Coordinations.¹

Semantically Plausible (+SP)
NP	A3	The student opened the backpack, and the student erased the blackboard.
	B1	The postman sorted the mail, and the postman delivered the package.
Pronoun	A2	The server provided the menu, and she brought the appetizer.
	B3	The firefighter organized the equipment, and he started the pump.
Elided	A1	The babysitter emptied the bottle and prepared the formula.
	B2	The mechanic replaced the transmission and repaired the suspension.
Semantically Implausible (−SP)
NP	A5	The driver offered the onion, and the driver misplaced the checkbook.
	B4	The plumber damaged the book, and the plumber requested the apple.
Pronoun	A6	The banker accepted the guitar, and he watched the escalator.
	B5	The hitchhiker selected the document, and she mentioned the farm.
Elided	A4	The perpetrator cherished the filter and noted the stereo.
	B6	The astronaut purchased the alcohol and recalled the hurricane.

¹Bold sentences block anaphoric construal by coreference and require binding.

As illustrated in Figure 1, design of the Language Faculty integrates syntactic computation with external components, both sensori-motor and cognitive, at its interfaces. Critically, the Language Faculty must relate language and thought (conventionally termed the Conceptual-Intentional (C-I) interface). At this interface, syntactic computation must integrate with general cognition, e.g., attention and memory, which are necessary for reference to entities in the world, and subsequent inference based on reference. Current developments in linguistic theory pursue the nature of these interfaces (e.g., Chomsky, 2011; Berwick et al., 2013). “Correctly capturing the interface is the crucial adequacy criterion of any syntactic theory.” (Reinhart, 2006, p. 2).

Figure 1.

Language Faculty Interfaces.

Based on our experimental results we argue that in accord with the architecture of the Language Faculty, early language deterioration in prodromal AD, i.e., in aMCI, targets the C-I interface, even while leaving essential aspects of syntactic computation relatively intact. Our analyses of two forms of anaphora computation reveal particular deterioration in anaphoric construal which integrates coreference in sentence syntax.

1.2. Anaphora

We use the term “anaphora” here in the general sense: “any case where two nominal expressions are assigned the same referential value or range” (Reinhart, 1999, p. 20; cf., Wasow, 1986; Reuland, 2011). For example, in (1), the pronoun his reflects an open variable whose value must be determined in order for the sentence to be interpreted and its truth value determined (e.g., Reinhart, 1983a,b, 1986, 2006; Johnson, 2013; Collins & Postal, 2012).

• 1The electrician repaired his equipment.

  Through a form of anaphoric construal, his may be interpreted as the electrician. Alternatively, the open variable of the pronoun may be closed by reference to external context, e.g., the electrician may have repaired an inspector’s equipment, e.g., in accord with discourse topic or a speaker-hearer’s presupposition.

1.2.1. Two forms of anaphoric construal

It has long been recognized that language reveals a fundamental dissociation between two forms of anaphoric construal: binding and coreference (Reinhart 1983a,b, 1986 and earlier, e.g., Keenan, 1971; Evans, 1980; see Reuland, 2011, p. 27–30; Reuland & Evereart, 2001 for review, Everaert et al. 2010 for studies relevant to their psychological reality). While binding is directly computed through syntax, computation of coreference reaches outside of core grammar; it must integrate with contextual, i.e., semantic and pragmatic, reasoning. Although all sentence production and comprehension involves a wide complex of syntactic, semantic and pragmatic knowledge, two forms of anaphoric construal dissociate these factors.

The two forms of anaphoric construal are shown in classic examples such as (2), where the pronoun his (reflecting an open variable) allows the two interpretations represented in simplified form in (2a) and (2b).

• 2
  The electrician repaired his equipment and the inspector did too.

  1. Electrician (λx (x repairs a’s equipment)) & inspector (λy (y repaired a’s equipment)) [“his” = electrician, a referent given in the sentence; alternatively, some male individual in the context, e.g., discourse]
  2. Electrician (λx (x repaired x’s equipment)) & inspector (λy (y repaired y’s equipment)) [“his” = electrician or inspector; each repaired their own equipment]

In one interpretation, (2a), the open variable of the pronoun his is free; it is closed by a referent which is available from sentential and/or discourse context. A relation of coreference links pronoun and antecedent². In contrast, in (2b), the pronominal variable is “bound” by each sentence subject. Here, a structural relation of c-command of the open variable by the sentence subject, either electrician or inspector, provides the core syntactic domain for anaphoric binding. Because it allows the pronoun his to be bound, (2b) allows the interpretation where both the electrician and the inspector repaired their own equipment. No unique referent through external context is assigned to the open variable.³ We will assume the fundamental definitions of binding and coreference in (3a) and (3b).

• 3
  1. Anaphoric Construal by Binding (derived from Chomsky 1981,1980) α binds β iff α and β are coindexed and α c-commands β.
  2. Anaphoric Construal by Coreference

If pronouns are not bound, they are free. Their reference, and coreference between antecedent and pronoun, may be interpreted in terms of context.

In sum, pronouns are essentially ambiguous: “at the C-I interface pronouns can be translated either as expressions receiving a value directly from a discourse storage or as variables to be bound by an antecedent” (Reuland, 2001, p. 447).

1.2.2. At the interface

The two forms of anaphoric construal straddle the C-I interface in complementary ways. Construal by binding is directly determined by a structural relation between antecedent and variable through c-command, an essential property of syntax involving hierarchy and dominance in sentence representation. Construal through coreference is also constrained by syntax, but in this case, syntax determines where anaphoric construal can not occur; as we summarize in (4) (cf., Reinhart, 1983b, p. 75).

• 4
  1. Bound anaphora is possible if a given NP c-commands a pronoun.
  2. Anaphora (bound anaphora and coreference) is impossible if a given NP is c-commanded by a pronoun.

For example, in (5a), anaphoric construal between the pronoun he and the electrician is disallowed; the subject pronoun structurally c-commands the electrician. [SM] In 5b or c, the pronoun is not so constrained and is free to reach out to semantic and pragmatic context, e.g., discourse or sentence context.

• 5
  1. #He visited the tenant when the electrician fixed the light switch.⁴
  2. When he visited the tenant, the electrician fixed the light switch.
  3. The electrician fixed the light switch when he visited the tenant.

In Binding Theory(BT), Principle C (6) states this constraint (e.g., Chomsky, 1981; Büring, 2005a).

• 6Binding Theory Principle C

  A referential expression must be free (i.e., not bound).

Decades of linguistic and psycholinguistic study have pursued explanation for anaphoric construal, and the “interface strategies” it necessitates.⁵ For example, it has been proposed that resolving anaphoric dependency by variable binding is “more readily accessible”, “less costly” or “more economical” than establishing coreference by using discourse storage (see Reuland, 2001, p. 448, 2010, p. 271, 2011, p. 127 for review, Reinhart, 1986, 2006).⁶

2. Leading hypothesis

If language deterioration characterizes prodromal AD, i.e., Mild Cognitive Impairment of the amnestic type (aMCI), then we predict that this deterioration will distinguish the performance of aMCI individuals from comparable Heathy Aging (HA), and/or Healthy Young (HY) on complex sentences with anaphora. More specifically, if the source of this prodromal language deterioration involves the challenge of syntax-semantics integration at the Conceptual-Intentional (C-I) interface of the Language Faculty with broad cognition (Lust et al., 2015a, 2017, Sherman et al 2021), then this deterioration in aMCI should differentiate two forms of anaphoric construal. Deterioration should be revealed most directly in sentences involving anaphoric construal through coreference where syntax and external reference must be integrated in the computation of anaphoric construal. At the same time, aMCI individuals may show competence in the essential syntax of binding, i.e., in the structural foundations of binding, reflected in BT Principle C, and in comparable forms of bound anaphora, where anaphoric construal does not require external reference. Finally, as in previous research (Sherman et al., 2021), we predict that this deterioration is to some degree independent of deficits in memory which characterize the aMCI population.

3. Method

3.1. Participants [SM]

Through cross-institutional interdisciplinary collaboration (Cornell University, Massachusetts Institute of Technology (MIT), Massachusetts General Hospital (MGH) and Institutional Review Board approval for human subjects research at three sites including informed consent, we tested three populations: aMCI, HA and HY, as summarized in Table 3. All completed a socio-demographic questionnaire.

Table 3.

Summary of Participants and Basic Metadata across Groups.

Group	n	Mean Age	Age Range	Females	Mean Years Education	Years Education Range
HY	10	23	21–27	5	15.6	14–18
HA	24	74	62–87	13	17.1	12–20
aMCI	61	78	58–98	26	16.4	10–20
Total	95	71	21–98	44	16.5	10–20
Participant Subgroups
HA
Subgroup	n	Mean Age	Age Range	Females	Mean Years Education	Years Education Range
HA (Cornell)	14	71	65–80	7	16.7	12–20
HA (MGH PAC)	10	77	62–87	6	17.6	16–20
HA Totals	24	74	62–87	13	17.1	12–20
aMCI
Subgroup	n	Mean Age	Age Range	Females	Mean Years Education	Years Education Range
aMCI (MGH PAC)	22	75	68–88	10	16.8	12–20
aMCI (MGH ADRC)	39	79	58–98	16	16.3	10–20
aMCI Totals	61	78	58–98	26	16.4	10–20

aMCI participants in the psycholinguistic tests include both an MCI sample of 22 recruited from (MGH) Psychology Assessment Center (PAC) and a sample of 39 additional aMCI subjects from MGH Alzheimer’s Disease Research Center (ADRC). ADRC participants received only the CS experiment, not the ASC or the cognitive tests.⁷ aMCI from MGH PAC, participated in both the ASC experiment and the CS experiment as well as in the two memory assessments (ACE-R, BP) ( See section 4.6). The aMCI participants were identified based on their performance on a battery of neuropsychological tests, specifically the Uniform Data Set from the National Alzheimer’s Coordinating Center which documented primary deficits in the domain of memory [SM].

The 24 HA participants included 14 assessed at Cornell University and 10 at MGH PAC. Ten HY, presumably at the height of memory skills (Park et al., 2002), were recruited from MIT (six students, four employees, i.e., research assistant, lab technicians, administrative assistant).

3.2. Elicited imitation (EI) task

We elicited language production through an experimentally designed and controlled EI task. In the EI task, participants are simply asked to ‘say the sentence just as I say it”. Standardized criteria for administration and analyses are applied. Extensive prior investigation has revealed that performance on this task requires analysis and reconstruction of both syntactic and semantic factors involved in a sentence model, thus significantly revealing generative linguistic computation (e.g., Blume & Lust, 2017, Lust et al., 1996, 1986; Dye & Foley, 2021). In integrating syntactic and semantic factors, successful generation of a sentence model in the EI task directly reveals production but also requires comprehension. Patterns of successful performance and errors (i.e., failures to reproduce the model sentence according to standardized criteria) evidence this computation. Experimental factorial designs allow assessment of hypothesized factors involved in this linguistic computation. The task is accessible to participants from a wide cognitive range, including those in early stages of dementia (e.g., Whitaker, 1976; Miller et al., 2021).

Participant responses were recorded and later transcribed and scored for accuracy and for the nature of errors, using standardized procedures for administration, transcription, reliability and scoring (Blume & Lust, 2017; Lust et al., 1996). In general, any major change in syntactic or semantic factors was scored as incorrect. Reformations of the model were analyzed.⁸

4. Design

4.1. Adverbial Subordinate Clause (ASC) Experiment

Our first experiment (ASC) tests anaphora in sentences with adverbial subordinate “when” clauses with lexical pronoun subjects (Table 1), which productively allow coreference construals (e.g., 7a-c). In 7a-c the pronoun is free to optionally refer to a referent from sentence or discourse context. We contrast these with one condition where coreference is grammatically blocked (by BT Principle C) bolded on Table 1, e.g., 7d, where the named subject is c-commanded by the pronoun and may not be anaphorically construed with it. [SM]

• 7
  1. The electrician fixed the light switch when he visited the tenant. (RBF)
  2. When the professor understood the scholar, he explained the theory.(LBF)
  3. When she remembered the orphan, the composer wrote the sonata. (LBB)
  4. #She answered the telephone when the receptionist heard the caller. (RBB)

As Table 1 displays, in the ASC experiment, our sentence design varies two syntactic factors: Branching Direction (BD), i.e., direction of adjunction of subordinate clause relative to the main clause, either right (R) or left (L), i.e. following or preceding the main clause, and Pronoun Direction (PD) (whether potential nominal antecedent within the sentence precedes, forward (F), or follows, backward (B), the pronoun). In this way, we evaluate the critical linguistic factors, structure (BD) and linearity (PD), in anaphora computation. BT Principle C, for example, critically involves integration of pronoun linearity with its structural context, e.g., RBB sentence (7d).

For example, in the RBF (Right Branching Forward Pronoun) condition, e.g, 7a, either the electrician or someone else, possibly under discussion in the discourse, e.g., an inspector, may have visited the tenant. Here the pronoun may optionally be construed with, and its reference optionally determined by, either the named main clause subject, or a potential external referent. However, in the RBB (Right Branching Backward) condition 7d, in accord with BT Principle C, it cannot be the “receptionist” who answered the telephone. This RBB condition contrasts with all other ASC sentences, where anaphora through free reference is structurally allowed, including other sentences with backward pronouns,e.g., 7c.

4.2. Coordinate Sentence (CS) Experiment

In the second experiment (CS) (Table 2), we also test pronominal anaphora with potential coreference. Here we contrast coordinate sentences such as 8a which allow coreference with sentences with binding, e.g, 8b (bolded on Table 2). The elided coordinate sentences like 8b involve bound anaphora. (They involve coordinate verb phrases (VPs) with null subjects).⁹

• 8
  1. The server provided the menu and she brought the appetizer.
  2. The babysitter emptied the bottle and Ø prepared the formula.

In sentences such as (8b), only the babysitter is possibly construed as who prepared the formula; no external referent is possible.^10,11 This contrasts with 8a where the pronoun is free to refer to the named antecedent (the server) or with someone else.

To dissociate the role of anaphora from complex sentence formation in general, we also test comparable coordinate sentences which do not involve anaphora, but repeat the subject noun phrases (NPs), e.g., 9.¹²

• 9The student opened the backpack and the student erased the blackboard

4.3. Semantic Plausibility

As tables 1 and 2 show, both ASC and CS experiments also vary in Semantic Plausibility (SP) (+/−), i.e., whether the sentence can more easily be interpreted to “make sense” intuitively, or not. Given that the EI task requires integration of syntax and semantics, this factor allows us to evaluate whether syntactic and semantic effects are independent.

4.4. Summary of ASC and CS experimental designs

As Table 1 indicates, ASC sentences with pronoun subjects in finite subordinate clauses varied by 2×2×2=8 conditions: Semantic Plausibility (SP) (+/−) × Branching Direction (BD) (right or left position of subordinate clause relative to main clause) × Pronoun Direction (PD) (forward or backward relative to potential sentence antecedent). Group was a Between Groups factor. As Table 2 shows, CS Sentences varied by 2×3=6 conditions: SP(+/−), Coordination Type, i.e, the nature of the coordinate subject NPs: a lexical Pronoun, a null proform in Elided VP coordination, or a repeated NP. Group was a Between Groups factor.

4.5. Sentence controls

With one replication item for each condition, each participant imitated 24 ASC sentences and 12 CS sentences, with the order of ASC and CS experiments and the order of + or −SP sentences counterbalanced across the subjects.¹³ In both experiments, sentences were administered in two randomized counterbalanced batteries (A and B) containing replication items, counterbalanced within group. All sentences within and across the two linguistic experiments were controlled in length (17 syllables), as well as in lexical frequency, lexical semantic associativity, and all other aspects of structure other than targeted variables. Lexical frequency was determined through logfreq-HAL (Elexicon http://elexicon.wustl.edu/). Semantic Plausibility was based on University of South Florida Free Association Norms for the lexicon (http://w3.usf.edu/FreeAssociation). Pronoun gender was systematically varied across replication items. Animacy of NPs was controlled and counterbalanced across conditions.

4.6. Cognitive tests of memory [SM]

We analyze our populations with regard to the relation between their responses on the ASC and CS experiments and their performance on two memory tests.¹⁴ The episodic memory measure is included within a standardized multi-domain cognitive screening measure sensitive to assessment of neurocognitive disorders (including MCI and dementia), the Addenbrooke’s Cognitive Examination-revised (ACE-R), which includes brief assessments of several different forms of episodic memory, as well as other domains of cognition. The other measure specifically tests verbal working memory: the (Brown-Peterson (BP) Consonant Trigram test, also referred to as the BP procedure in Brown (1958)) “requires on-line maintenance of information no longer available from sensory input in the face of simultaneous activation of cognitive processes that compete for attentional resources” (Strauss et al., 2006, p. 705). The BP procedure has been found to be sensitive to MCI in AD progression (Belleville et al., 2007). Both ACE-R and BP assess general cognitive processes involved in memory, which may be assumed to underlie performance on our linguistic tasks.

4.7. Design-based predictions

Our experimental design allows us to test our leading hypothesis. It allows us to evaluate potential linguistic deficit in aMCI, in contrast to HA and HY. Specifically, our ASC and CS designs allow us to evaluate two forms of anaphoric construal, coreference and binding. They allow us to test (i) potential deficit in coreference computation, doing so across two structures (subordinate and coordinate); (ii) to contrast this with cases where coreference is not possible, i.e., with cases ruled out by BT constraint (in the ASC experiment); (iii) and with cases where binding is obligatory (in the CS experiment); and (iv) to assess effects of syntactic factors of hierarchical structure (Branching Direction) and linearity (Pronoun Direction) which underlie computation of BT in syntax. Finally, our designs allow us to assess (v) effects of Semantic Plausibility (+/−), for its potential interaction with syntactic effects.

To the degree that foundational aspects of syntactic knowledge remain intact, we predict evidence in aMCI for obeyance of BT (Principle C) and sensitivity to the structural principles which underlie the BT; accordingly, Linearity effects (Pronoun Direction) should be modulated by structure (Branching Direction). If anaphoric construal by coreference indicates a locus of deterioration during prodromal AD (i.e., aMCI), then we predict deficit in production of the ASC sentences, with exception of the Principle C condition where coreference is blocked. In CS we predict that sentences with Pronoun subjects would be significantly deficited in contrast to Elided sentences which reflect obligatory binding. If anaphora is a focal domain for participant deficit, over and above general sentence syntax, and if complex sentence formation remains accessible, then sentences with repeated NPs, which do not involve anaphora, may not reveal this exceptional deficit. To the degree that the linguistic deficit in aMCI is independent of cognitive deficits including memory decline, we expect the observed pattern of results not to be predicted by memory decline as measured by standardized tests of memory, in keeping with earlier results (Sherman et al., 2021).

4.8. Analyses

Quantitative analyses of language data resulting from our factorial designs were conducted with logistic-linear mixed models with binomial error assumption and a logit link function. Qualitative analyses supplemented quantitative analyses through linguistic examinations of imitation reformations, which were counted as errors.

We first analyzed a series of models to examine the results of the (ASC and CS) tests administered. The models to examine performance on these tasks (the dependent variable was the proportion of correct responses) included Group (3 levels—HY, HA, aMCI) and repeated measures factors resulting from the sentence structure of the task (in ASC and in CS) as fixed classification factors; the interactions of these factors; and individuals as levels of a random classification factor. There is a replication on each sentence type, but replication was not included in the model as a classification factor. Incorrect or correct (0 or 1) is the score for each response. Analysis was by logistic mixed models with binomial error (Wolfinger & O’Connell, 1993; Schall, 1991). The logistic specification is chosen because of the dichotomous outcome variable. Mean differences in proportion correct response were examined in a mixed model including fixed classification factors for Group and the Sentence Types; the interactions among these factors; and individuals as levels of a random factor.

A second set of models examined regressions of the psycholinguistic test (ASC and CS) outcomes on various ACE-R and BP scores, with those regressions specified separately by Group, and the homogeneity of the regressions tested (Henderson, 1982). In these models, the dependent variables are summary versions of linguistic task outcomes—the total score correct across all sentence types and the totals for +SP and −SP (the latter two are not independent of the first and are not interpreted as such)—and the independent variables are Group (3 levels) and a particular ACE-R or BP variable as a covariate (with regressions conducted separately by Group). A key test is the significance of the regressions in the HA and aMCI groups. These models were carried out for ACE-R total score and subcomponent variables, total fluency, both semantic and phonemic fluency subcomponents, memory, and visuospatial; and for BP, total and delays of 0, 9, 18, and 36 seconds. In this paper, we report the ACE-R memory subcomponent and the BP total score. Because the ADRC subset of MCI participants did not receive the ASC experiment or the ACE-R and BP tests, they are not included in the analyses that include those tests.

Qualitative analyses of imitations scored incorrect complemented quantitative analyses of successful imitation performance. Along with variation in “correct” reproductions which cohere with experimental sentence design, response reformations provide evidence of each individual’s representations underlying their performance, and support inference regarding the linguistic computation underlying production.

5. Results

5.1. ASC Experiment

Table 4 reports the mean success values for production of the ASC sentences across design conditions, across the three groups, HY, HA and aMCI. Figure 2 displays mean correct production of ASC sentence types for each group, HY, HA and aMCI, conflating over the Semantic Plausibility factor, and reports group comparisons across these.

Table 4.

Mean Correct Imitation of Sentences with Adverbial Subordinate Clauses (Table 1): across Healthy Young, Healthy Aging and aMCI. (SE in parentheses.)¹

Healthy Young: 0.802 (0.095)	+SP: 0.879 (0.100)	Right Branching: 0.879 (0.109)	Forward Pronoun Backward Pronoun	0.800 (0.124) 0.958 (0.124)
		Left Branching: 0.879 (0.109)	Forward Pronoun Backward Pronoun	0.852 (0.124) 0.905 (0.124)
	−SP: 0.725 (0.099)	Right Branching: 0.700 (0.108)	Forward Pronoun Backward Pronoun	0.650 (0.123) 0.750 (0.123)
		Left Branching: 0.750 (0.108)	Forward Pronoun Backward Pronoun	0.700 (0.123) 0.800 (0.123)
	SP Conflated	Right Branching: 0.789 (0.099)	Forward Pronoun Backward Pronoun	0.725 (0.108) 0.854 (0.108)
		Left Branching: 0.814 (0.099)	Forward Pronoun Backward Pronoun	0.776 (0.108) 0.853 (0.108)
Healthy Aging: 0.632 (0.061)	+SP: 0.708 (0.064)	Right Branching: 0.792 (0.069)	Forward Pronoun Backward Pronoun	0.729 (0.079) 0.854 (0.079)
		Left Branching: 0.625 (0.069)	Forward Pronoun Backward Pronoun	0.604 (0.079) 0.646 (0.079)
	−SP: 0.555 (0.064)	Right Branching: 0.577 (0.070)	Forward Pronoun Backward Pronoun	0.631 (0.080) 0.523 (0.080)
		Left Branching: 0.534 (0.070)	Forward Pronoun Backward Pronoun	0.566 (0.080) 0.501 (0.080)
	SP Conflated	Right Branching: 0.684 (0.064)	Forward Pronoun Backward Pronoun	0.680 (0.070) 0.688 (0.070)
		Left Branching: 0.579 (0.064)	Forward Pronoun Backward Pronoun	0.585 (0.070) 0.573 (0.070)
aMCI: 0.489 (0.064)	+SP: 0.592 (0.067)	Right Branching: 0.699 (0.073)	Forward Pronoun Backward Pronoun	0.580 (0.084) 0.818 (0.084)
		Left Branching: 0.485 (0.073)	Forward Pronoun Backward Pronoun	0.509 (0.084) 0.461 (0.084)
	−SP: 0.386 (0.067)	Right Branching: 0.398 (0.073)	Forward Pronoun Backward Pronoun	0.386 (0.083) 0.409 (0.083)
		Left Branching: 0.375 (0.073)	Forward Pronoun Backward Pronoun	0.455 (0.083) 0.296 (0.083)
	SP Conflated	Right Branching: 0.548 (0.067)	Forward Pronoun Backward Pronoun	0.483 (0.073) 0.614 (0.073)
		Left Branching: 0.430 (0.067)	Forward Pronoun Backward Pronoun	0.482 (0.073) 0.378 (0.073)

¹+SP=Semantically Plausible; −SP=Semantically Implausible

Figure 2.

Percent correct imitation: Adverbial Subordinate Clauses by Factorial Conditions across three groups: Healthy Young, Healthy Aging, and amnestic Mild Cognitive Impairment (aMCI).

5.1.1. Group comparisons

In the ASC experiment, Group is significant, F(2, 799) = 3.90, p = 0.021. As can be seen in Figure 2, HY and HA do not differ significantly from each other, t(799) = 1.51, p = 0.132. However, aMCI individuals perform significantly lower than HY: HY vs. aMCI, t(799) = 2.74, p = 0.006, and lower than HA, t(799) = 1.61, p = 0.107 although not significantly.

These results confirm significant linguistic deterioration in aMCI on the ASC sentences, supporting our first/main hypothesis. The lack of significant difference between HY and HA suggests that this linguistic deterioration is not simply due to aging per se, but rather to specific deficits in MCI, though the fact that there is only a trend towards significance between HA and aMCI suggests that there is some effect of aging on sentence production. Since these sentences involve anaphora computation in complex sentences where coreference is productively allowed, (except for one condition), this deficit documents that this area provides a general area of language decline in aMCI.

Semantic Plausibility is significant, F(1,799) = 38.62, p < 0.0001. Production is lower in implausible conditions (−SP), without interaction with Group, F(2,799) = 0.5, p = 0.61.[SM] The fact that Semantic Plausibility does not significantly interact with Group suggests that the source of aMCI deterioration lies at least partially in the syntactic factors they involve over and above their general semantic plausibility.

5.1.2. Structural factors: Branching Direction and Pronoun Direction

Overall, the structural factor of Branching Direction, which was dissociated from linear Pronoun Direction in design, has a significant effect, F(1, 799) = 5.87, p = 0.016, without significant interaction with Group; RB sentences are more successfully produced (overall mean RB = 0.674, LB = 0.608). The Pronoun Direction factor is not significant, F(1, 799) = 1.95, p = 0.163, and also does not interact significantly with Group, F(2, 799) = 1.12, p = 0.328. Both of these factors underlie an effect of the Principle C sentences.

5.1.3. Binding Theory constraint: Principle C and Linearity

As can be seen in Figure 2 (which reports success across sentence types in each group conflating over Semantic Plausibility), there was superior production of the right branching structures with backward pronouns (RBB) where BT Principle C blocks coreference (RBB), compared to RBF sentences which allow coreference. Although relative success on RBB sentences characterizes all groups, as Figure 2 shows, this superiority is significant only within the aMCI group. The difference between RBB and RBF sentence success in aMCI is, t(799) = −2.28, p = 0.023; compared to HA, t(799) = −0.15, p = 0.881; or HY, t(799) = −1.52, p = 0.129.

As Figure 2 shows, the superior performance by aMCI on the RBB sentence type contrasts with their performance on sentences which allow coreference. The aMCI group is depressed in the production of sentences LBF, RBF and LBB, which allow coreference (LBF = 0.482, RBF = 0.483, LBB = 0.378), compared to HA (mean = 0.585, 0.680, 0.573) and HY (mean = 0.776, 0.725, 0.853). In contrast, in spite of their deficited performance on these sentences across the design, the aMCI group performs as well as HY and HA on the RBB sentences that grammatically block coreference by Principle C, (aMCI = 0.614, HA = 0.688, HY = 0.854). aMCI vs. HA: t(799) = 0.74, p = 0.459; vs. HY: t(799) = 1.84, p = 0.066. In aMCI, not only are RBB sentences significantly superior to RBF sentences, unlike either HY or HA, but aMCI individuals achieve significantly greatest success on RBB (0.614), where BT prevents coreference, and lowest on LBB (0.378), which also involves a backward pronoun, and where coreference is an option (RBB vs. LBB in aMCI: t(799) = 4.12, p < 0.0001). Pronoun Direction does not significantly affect performance on LB structures where both forward and backward pronouns are possible. aMCI individuals perform similarly on LBB and LBF (mean = 0.378 vs. 0.482, t(799) = 1.81, p = 0.071). LB sentences with backward pronouns do not differ significantly from RB forward pronouns (mean = 0.378 vs. 0.483, t(799) = 1.84, p = 0.066); both of which allow anaphoric construal by coreference and are depressed. Therefore, the greater success of the grammatically constrained RBB sentences, accentuated in aMCI, does not simply reflect the backward pronoun linearity they involve, but the integration of this linearity with structure; specifically when this integration blocks anaphoric construal by coreference.

5.1.4. ASC Experiment and Semantic Plausibility [SM]

Figure 3 displays the effect of the Semantic Plausibility factor across ASC sentence conditions. Although it did not interact significantly with Group, Semantic Plausibility significantly interacted with both of the factors, Branching Direction and Pronoun Direction, F(1, 799) = 4.95, p = 0.026, and F(1, 799) = 4.28, p = 0.039, respectively. Not only did Semantic Implausibility (−SP) generally depress performance across the board in all groups, but RB superiority was linked to its Semantic Plausibility. Whereas with semantically plausible sentences, aMCI individuals have highest performance on the RBB BT Principle C constrained sentences, this effect is eliminated in Semantically Implausible (−SP) conditions, as can be seen in Figure 3. Whereas with +Semantic Plausibility (+SP), on RBB sentences the aMCI group does not differ significantly from either HY or HA, without Semantic Plausibility (−SP), aMCI individuals perform significantly lower than either HY or HA.

Figure 3.

Effect of Semantic Plausibility on Adverbial Subordinate Clause (ASC) Sentence Imitation in amnestic Mild Cognitive Impairment (aMCI).

Across sentence types on aMCI, the highest decrease due to Semantic Implausibility (−SP) characterized the Principle C constrained RBB sentences (+SP: 0.818, −SP: 0.409), t(799) = 5.05, p < 0.0001. In aMCI, no sentence condition other than RBB showed a significant effect of Semantic Plausibility. This selective RBB sensitivity to Semantic Plausibility also characterized the HA. As illustrated in Figure 3, in semantically implausible sentences (−SP), aMCI individuals no longer differentiate among sentence conditions except for falling to their lowest performance on LBB sentences (−SP: 0.296).

5.1.5. ASC Experiment: summary

Results from the ASC experiment show that language decline in aMCI relative to HY and HA is specifically reflected in their differential production of sentence types which allow anaphoric construal by coreference. At the same time, aMCI individuals, like HA, are sensitive to structure which underlies binding, viz., Branching Direction. i.e, they are sensitive to a specific structural constraint of the BT Principle C. The high success of aMCI individuals on the BT constrained sentence type RBB (82% with Semantic Plausibility, 61% conflated over both semantic types) reveals general psycholinguistic competence for complex sentences when they do not involve anaphoric construal through coreference.

aMCI like HA access semantic as well as syntactic factors in their productions. Without Semantic Plausibility, sentence computation, in particular anaphora computation, is disabled. −SP selectively depresses the condition involving binding, i.e., the BT Principle C constrained RBB, and eliminates the superior performance by aMCI on the BT constrained RBB sentences.

5.2. CS Experiment

Table 5 reports mean success values for production of CS sentences (Table 2), across design conditions for the three groups, HY, HA and aMCI. Figure 4 displays performance on three Coordination Types for the three groups.

Table 5.

Mean Correct Imitation of Coordinate Sentence Types (Table 2) across Healthy Young, Healthy Aging and aMCI. (SE in parentheses.)

Healthy Young: 0.865 (0.068)	+SP: 0.860 (0.078)	Pronoun NP Elided	0.682 (0.161) 0.946 (0.051) 0.860 (0.100)
	−SP: 0.870 (0.073)	Pronoun NP Elided	0.806 (0.123) 0.946 (0.051) 0.806 (0.123)
	SP Conflated	Pronoun NP Elided	0.749 (0.122) 0.946 (0.041) 0.835 (0.093)
Healthy Aging: 0.765 (0.065)	+SP: 0.846 (0.054)	Pronoun NP Elided	0.789 (0.084) 0.959 (0.027) 0.657 (0.108)
	−SP: 0.658 (0.086)	Pronoun NP Elided	0.685 (0.115) 0.712 (0.099) 0.571 (0.115)
	SP Conflated	Pronoun NP Elided	0.740 (0.080) 0.884 (0.049) 0.615 (0.096)
aMCI: 0.499 (0.055)	+SP: 0.577 (0.058)	Pronoun NP Elided	0.373 (0.069) 0.739 (0.059) 0.601 (0.071)
	−SP: 0.421 (0.058)	Pronoun NP Elided	0.316 (0.065) 0.585 (0.071) 0.370 (0.069)
	SP Conflated	Pronoun NP Elided	0.344 (0.057) 0.666 (0.057) 0.485 (0.063)

¹+SP =Semantically Plausible; −SP=Semantically Implausible

Figure 4.

Percent Correct Imitation: Coordinate Sentences by Factorial Conditions (Elided, NP, or Pronoun) across three groups: Healthy Young, Healthy Aging and amnestic Mild Cognitive Impairment (aMCI).

5.2.1. Group comparisons

In Coordinate Sentences (Table 5, Figure 4), Group is again significant, F(2, 1024) = 7.11, p = 0.001, as is Semantic Plausibility, F(1, 1024) = 5.50, p = 0.019, with no significant interactions between them. [SM] Overall, again there is no significant difference between HY and HA, t(1024) = 0.99, p = 0.324, but aMCI individuals are significantly depressed compared to both HY and HA, t(1024) = 2.99, p = 0.003, and, t(1024) = 2.80, p = 0.005, respectively. Overall mean percent correct performance on CS decreases from 87%, and 77% in the HY and HA respectively to 50% in aMCI.

5.2.2. Coordination Type

As Figure 4 displays, Coordination Type (CT) is significant, F(2, 1024) = 11.58, p < 0.0001 and does not interact significantly with Group, F(4,1024) = 2.11, p = 0.077. Sentences with repeated NP subjects, had the highest success within all groups including aMCI, with or without Semantic Plausibility.

The CT variable differentiating sentences with coreference (Pronoun subjects) and those with binding (Elided) significantly differentiated the aMCI from HA and HY. aMCI individuals perform significantly below HA on coordinate sentences with a Pronoun subject, which allow coreference, t(1024) = 3.48, p = 0.001, and below the HY, t(1024) = 2.50, p = 0.013 (means aMCI, HA and HY, respectively, 0.344, 0.740, 0.749). At the same time, aMCI group does not differ significantly from HA in producing coordinate sentences which involve binding (Elided), t(1024) = 1.11, p = 0.268 (means HY: 0.806, HA: 0.615, aMCI: 0.485 (conflated over Semantic Plausibility)). As can be seen in Figure 4, in contrast to HA, aMCI individuals perform lower on the CS sentences with Pronoun subjects than they do on Elided coordinations, t(1024) = 2.67, p = 0.008 in aMCI, t(1024) = −1.61, p = 0.109 in HA, t(1024) = 0.87, p = 0.384 in HY. Neither HA or HY revealed a significant contrast between Elided and Pronominal Coordinations, either with or without Semantic Plausibility. For HA,+SP, t(1024) = −1.28, p = 0.199, −SP, t(1024) = −0.99, p = 0.324; for HY, +SP, t(1024) = 1.23, p = 0.219, −SP: t(1024) = 0.00, p = 1.000).

5.2.3. CS Experiment and Semantic Plausibility [SM]

Without Semantic Plausibility (−SP), the aMCI group performs significantly below HA on sentences with coordinate Pronoun subjects, t(1024) = 2.73, p = 0.006 for HA vs. aMCI, as they do with Semantic Plausibility. However as in ASC results, without Semantic Plausibility, aMCI individuals no longer differentiate between binding and coreference (Elided vs. Pronoun coordination). −SP sentences with binding (Elided coordinations) are no longer significantly superior to those with coreference (Pronoun subjects) (means 0.370 vs. 0.317 respectively), t(1024) = 0.78, p = 0.437, compared to those +SP, t(1024) = 3.01, p = 0.003. In aMCI, Semantic Plausibility has a significant effect on CS sentences which involve binding (Elided Coordinations) (means +SP: 0.601, −SP: 0.370); but not on sentences which involve coreference (Pronoun subjects) (−SP: 0.316 vs. +SP: 0.373), t(1024) = 3.04, p = 0.002 for Elided; t(1024) = 0.80, p = 0.422 for Pronoun subjects. This difference for either HY or HA was nonsignificant.

5.2.4. CS Experiment Summary

CS results converge with our essential hypothesis that anaphoric construal by coreference in particular, may be significantly deteriorated in aMCI, even while binding may be relatively preserved. The results converge with those from the ASC experiment in showing a deficit in aMCI related to coreference, and show that this effect generalizes across distinct structures (coordinate vs. adjoined subordinate clauses in CS and ASC experiments). In contrast to HA and HY, aMCI show a significantly greater reduction in successful production of sentences which involve coreference, compared to those with binding. Relatively greater success in aMCI on CS repeated NP sentences confirms basic overall competence for complex sentence production, independent of anaphora. As with the ASC experiment, −SP eliminated the crucial distinction between anaphoric construal types, both within the aMCI group and in its contrast to other groups. As with ASC, the SP factor significantly affected sentences involving binding (Elided coordinations).

5.3. Reformations: Changes to model sentences

Participant reformations target the anaphoric relation in both ASC and CS experiments. Table 6 provides examples of responses by two aMCI individuals across sentences in the ASC experiment allowing either coreference (an LBB sentence), or a Binding Theory Principle C constrained sentence blocking coreference (RBB); and in the CS experiment, across a coordinate sentence with Pronoun subject which allows coreference, and an Elided coordination which involves only binding..

Table 6.

Examples of Differential Sentence Production by two aMCI Individuals across Coreference and Binding in Adverbial Subordinate Clause (ASC) and Coordinate Sentence (CS) Experiments.

As illustrated in Table 6, both participants reveal breakdown completing the anaphora relation between the NP and the pronoun on an ASC sentence which allows coreference (LBB); while at the same time, they successfully reproduce the RBB sentence where coreference is blocked. In the CS experiment, they successfully reproduce the Elided coordinate sentence requiring binding but reform the sentence including a Pronoun subject allowing coreference, revising it to a bound anaphora form, an Elided type. Given the controlled design of all sentences in the experiments, equalizing length, lexical complexity and other structural factors, this error pattern instantiates the increased deficit on sentences involving coreference as opposed to those disallowing coreference.

5.3.1. Group analyses of reformations: ASC Experiment [SM]

Errors involving reformations of the antecedent NP (head of an anaphora relation) and/or the pronoun were analyzed for each participant for each sentence across three populations, according to standardized Scoring Criteria. As Table 7 summarizes, while aMCI individuals operate on both components of the potential anaphora relation in the ASC sentences, there are more reformations of the NP, the antecedent of the potential anaphora relation, than of the pronoun.¹⁵ A pattern of predominant change to the NP antecedent in aMCI, replicates across all three groups, and across both + and −SP. Table 8 exemplifies aMCI changes of NPs to pronouns or other NPs; pronouns to nulls or NPs.

Table 7.

Amount of Changes of NP or Pronoun in Adverbial Subordinate Clause (ASC) Experiment. Percent of errors in each group (percent of items in parentheses)¹

	Semantically Plausible (+SP)		Semantically Implausible (−SP)
	Pronoun	NP	Pronoun	NP
HY (n = 10)	0 (0)	14.29 (2.5)	19.05 (5)	28.57 (7.5)
HA (n = 24)	7.89 (1.56)	21.05 (4.17)	7.69 (2.6)	38.46 (13.02)
aMCI (n = 22)	3.28 (1.14)	22.95 (7.95)	8.33 (4.55)	31.25 (17.05)

¹See Table 8 for examples of change types. Only errors which maintained two clauses are included here.

Table 8.

Examples of Anaphora Reformations by aMCI Individuals in Adverbial Subordinate Clause (ASC) Experiment. Changes of NP Antecedent or Pronoun.^1,2,3

I. NP

NP > Pronoun

(1) R: When she remembered the orphan, the composer wrote the sonata. (LBB, +SP)

P: When she remembered … the something … she remembered the sonata. I missed parts of that. (TS005)

P: When she remembered the orphan, she …encountered…the orphan’s sonata (TS007)

(2) R: When she respected the hunter, the drummer buried the replica. (LBB, −SP)

P: When she respected the hunter, she buried the hunter. (TS006)

P: When she respected the hunter, they buried the piano? (TS202–2)

NP > NP

(3) R: She measured the artifact when the storyteller paid the boyfriend. (RBB, −SP)

P: She measured the artifact when the somebody uh … something to the boyfriend. (TS005)

P: She measured the artifact when the uh>..when the uh boyfriend inspected the…(TS 029)

P: Ah, she,ig..ah..examined the artifact while the somebody paid the boyfriend (TS 011)

(4) R: He triggered the alarm when the ambassador saw the intruder. (+SP)

P: He triggered the alarm when the… in> the intruder startled the intruder. (TS006)

(5) R: When she respected the hunter, the drummer buried the replica (LBB, −SP)

P: When she respected the hunter, …ah...something buried the…(TS001)

P: When she respected the hunter, the…somebody buried the rec-replica (TS003)

(6) R: When she remembered the orphan, the composer wrote the sonata (LBB, +SP)

P: When she remembered the orphan, the dressmaker…oh boy…remembered the sonata (TS026)

P: When she remembered the>orphan…the writer wrote the sonata. (TS002)

II. Pronoun

Pronoun > Null

(7) R: The electrician fixed the light switch when he visited the tenant. (RBF, +SP)

P: The electrician fixed the light switch when Ø visiting the lieutenant. (TS001)

(8) R: The navigator imagined the plant when he offended the queen. (RBF, −SP)

P: The navigator inspected the plant after Ø discussing with the queen. (TS029)

Pronoun > NP

(9) R: The navigator imagined the plant when he offended the queen. (RBF, −SP)

P: The navigator offended the plant when something talked to the t-queen, or something like that. (TS 202–2)

(10) R: When he misled the neighbor, the deputy inherited the ball. (LBB, −SP)

P: When the neighbor inherited, the deputy inherited the ball. (TS019)

P: Uh when the deputy’s uh something to the neighbor, ah, somebody dropped the ball. (TS011)

¹Bold signifies RBB models where Principle C blocks anaphora.

²R: Researcher; P: Participant

³+SP=Semantically Plausible; −SP=Semantically Implausible

In the Principle C-blocked RBB sentences, which showed a high successful production in the aMCI group (as well as in HA and HY), aMCI individuals never changed the pronoun. They did not convert the NP to a pronoun as they do in Table 8 examples 1–4. As the Principle C examples 3–4 ( bolded) suggest, here participants often changed the NP to another NP with indefinite reference, e.g., somebody. These NP substitutions cohere with an absence of construal of an anaphora relation.¹⁶

5.3.2. Reformations in CS Experiment

As Table 6 exemplifies, in the CS experiment, both participants convert the CS sentence with a pronoun subject which allows coreference to one with a null subject, which disallows coreference and involves only binding. In contrast to ASC sentences, in coordinate sentences with a Pronoun subject which allow coreference, group analyses confirm that aMCI productively and predominantly (41% of items, 95% of errors) convert the CS Pronoun subject to a null subject, ie. an Elided coordination, as in the examples on Table 6. This coheres with their superior performance on the Elided sentences. It contrasts with their responses to the ASC sentences with pronouns where changes to the NP antecedent were predominant. Thus, where individuals have the structural option to convert to unambiguous binding, as in these coordinate sentences, they do so.

5.4. Memory and anaphora in HA and aMCI [SM]

As Table 9 shows, regression tests of total linguistic scores conducted against total scores on the two tests of memory (ACE-R and BP) showed that neither memory test significantly predicted overall performance on the ASC sentences, whether Semantically Plausible or not, for either the aMCI or the HA. For CS as well, neither memory test significantly predicted the linguistic performance by the aMCI. In HA, in the CS condition with Semantic Implausibility (-SP) there was a limited effect of the ACE-R memory test and an overall effect of the BP test. These results confirm that the language deficits we found in aMCI vs. Healthy Adults are not simply explained by memory deficits in these individuals.¹⁷

Table 9.

Estimated Regressions of Successful Production on Adverbial Subordinate Clauses (ASC) and Coordinate Sentences (CS) on Tests of Memory: Addenbrooke’s Cognitive Examination Revised (ACE-R) and Brown-Peterson Verbal Working Memory Test (BP) in the Healthy Aging (HA) and Amnestic Mild Cognitive Impairment (aMCI) groups.^1,2

Test		ASC TOT		ASC_+SP TOT		ASC_−SP TOT
	Group	Estimate (SE)	p	Estimate (SE)	p	Estimate (SE)	p
ACE-R	HA	0.023 (0.020)	0.249	0.025 (0.019)	0.197	0.031 (0.023)	0.182
Memory	aMCI	0.003 (0.014)	0.844	−0.001 (0.014)	0.945	−0.002 (0.015)	0.877
BP Total	HA	0.004 (0.007)	0.572	0.003 (0.007)	0.629	0.008 (0.008)	0.324
	aMCI	0.012 (0.008)	0.165	0.003 (0.010)	0.772	0.012 (0.009)	0.197
Test		CS TOT		CS_+SP TOT		CS_−SP TOT
	Group	Estimate (SE)	p	Estimate (SE)	p	Estimate (SE)	p
ACE-R	HA	0.043 (0.018)	0.022	0.027 (0.017)	0.138	0.060 (0.021)	0.005
Memory	aMCI	0.002 (0.013)	0.849	−0.004 (0.013)	0.762	−0.001 (0.015)	0.921
BP Total	HA	0.012 (0.007)	0.077	0.007 (0.006)	0.256	0.017 (0.008)	0.030
	aMCI	0.007 (0.008)	0.348	−0.009 (0.010)	0.357	0.009 (0.009)	0.294

¹Bold signifies p < .05.

²ASC Total (TOT) scores include the 8 sentences with nonfinite adjunct clauses (see footnotes 12, 16, 19). Subsequent tests on just ASC finite sentences do not significantly change the value of results.

6. Summary of ASC and CS experimental results

The experimental studies reported in this paper provide three major results in accord with our hypotheses.

First, our results converge across experiments to confirm significant language impairment in production of complex sentences with anaphora in aMCI in contrast to HA although there are effects of aging in our data. Throughout our results, HA and HY do not differ significantly from each other overall, at the same time that they differ from aMCI. This suggests that aging alone is not responsible for linguistic decline in aMCI.

Second, across experiments, the aMCI are characterized not only by general depression relevant to HY and HA, but by an amplified dissociation between forms of anaphoric construal. In both ASC and CS experiments, language in aMCI was found to selectively reveal impairment in sentences with anaphora which allow coreference. Their sentence production was depressed on sentences which involved linguistic computation requiring free establishment of a context-based external referent, i.e, either sentences in the ASC experiment (except for the Principle C constrained RBB) or sentences with pronouns in CS. At the same time, core linguistic properties of binding, a syntactic structure-dependent form of anaphora computation, were relatively preserved. aMCI individuals performed similarly to the healthy adult groups on sentences which involved either syntactic constraint against coreference in accord with BT Principle C in the ASC experiment, or obligatory binding as in Elided coordinations in the CS experiment. The results of the ASC experiment confirmed that the aMCI population demonstrated structure-dependence required for binding, distinguishing structural variation (i.e., Branching Direction), and allowing it to override pronoun linearity (Pronoun Direction). In both experiments, the aMCI group revealed a general competence for the syntax of complex sentence formation without anaphora, i.e, in coordinate sentences with repeated NP subjects, as well as in high performance on the Principle C BT constrained RBB sentences in the ASC experiment. They revealed general access of semantic representation (the Semantic Plausibility factor).

Participant reformations of model sentences in the (EI) task confirmed that anaphora construal underlies these results. In ASC sentences, individuals in all groups target the antecedent NP, head of the anaphora relation, more than the pronoun in their sentence reformations. With CS where either a coreferential or an obligatory binding option was available, aMCI individuals productively converted a coordinate sentence allowing coreference (those with pronoun subjects) to a sentence which allowed only binding (VP coordination with null subjects) in keeping with their superior performance with anaphoric construal engaging binding.

In both ASC and CS experiments, results critically depend on Semantic Plausibility, confirming the nature of the EI task which requires syntax-semantics integration in sentence generation. Reformations (Table 8) show that even in the case of Semantic Implausibility, participants attempt to resolve the syntactic anaphora relation. Semantic Implausibility in model sentences not only significantly reduced correct performance in all groups, increasing sentence reformations, but significantly interacted with sentence syntax. With Semantically Implausible ASC and CS sentences, aMCI no longer dissociated anaphoric construal types; they no longer exhibited advantage in sentences involving either BT constraint by Principle C (RBB) in the ASC experiment or obligatory binding with elided coordination in the CS experiment. These converging results provide evidence that anaphoric construal reflects syntax-semantics integration and that aMCI language deterioration lies specifically in a deficit in this integration.

Third, regression analyses between two standardized memory tests and linguistic performance in both experiments confirmed that our linguistic results were not simply determined by memory deficits characterizing the aMCI population.

7. Discussion

Advances in identifying brain-behavior relations in prodromal AD and explanatory theories of these relations must approach a fundamental “mapping” problem. Precise and fundamental linguistic concepts must be linked to precise and relevant neuroscientifically sound representations of brain systems, in spite of “incommensurability” across neural systems and behavioral domains (Poeppel, 2012, p. 49; Embick & Poeppel, 2014). As has been suggested, “it might turn out that it is linguistics that will have to meet the challenge of ‘decomposing’ its primitive representations and computations in different ways, if we’re going to be able to unify linguistics with what we discover about the nature of how the human nervous system implements storage and computation, and so on … “ (Poeppel & Idsardi, 2022). In our research program, we have pursued the hypothesis that refining the linguistic representation of language deterioration in prodromal AD can contribute to this challenge.

By our experimental analyses of anaphora in complex sentences and by engaging linguistic theory in our experimental designs, our results provide new and more precise evidence that aMCI language deterioration lies in the integration of syntactic and semantic knowledge at the C-I interface of the Language Faculty (cf., Sherman et al., 2021; Lust et al., 2017). However, before considering the implications of these results for future research, we must first consider several potential counterexplanations.

7.1. Alternative Explanations

7.1.1. Ambiguity

Our results raise the possibility that deterioration in sentences which allow coreference arise not from the allowance of this anaphoric construal type, but rather from “ambiguity” in these sentences.¹⁸ Once a coreference construal is possible, multiple potential referents are possible. In addition, in our adverbial subordinate clause (ASC) experiment, in sentences like RBF (Table 1), the syntactic structure of the sentence not only allows anaphoric construal by coreference but also by binding. The same c-command structure which would rule out both forms of anaphoric construal by Binding Theory Principle C in RBB, allows either binding or coreference in sentences such as RBF. [SM]. Therefore aMCI individuals’ depressed performance on sentences which allow coreference, involves deterioration in production of sentences which allow coreference but do not require it.¹⁹

However, there is no indication that two different interpretations are in play in processing these sentences, as an effect of ambiguity would predict. There is no context provided for the experimental sentences (e.g., no discourse or pragmatic lead) which would foster a potential interpretation, and no indication that a particular interpretation, or more than one, is being entertained by participants. In general, aMCI individuals do not eschew simple indeterminacy in pronouns, as shown by their high performance on the RBB Principle C sentences, which have an initial pronoun with undetermined content, and as suggested by the phenomenon of productive “empty pronouns” in natural speech of prodromal and AD individuals (e.g., Almor et al., 1999). The source of deterioration in sentences which allow coreference then appears to engage the syntactic structure of the sentences; a syntactic structure which allows more than one linguistic computation. If “ambiguity” of some form is involved in aMCI language deterioration, it is an “ambiguity” between two potential anaphoric construals defined syntactically, not between two or more interpretations.

7.1.2. Reference and context

The fact that aMCI deficit specifically involves coreference computation is not surprising, given the complex cognitive computation involving both semantics and pragmatics, which is necessary to free reference and coreference resolution (e.g., Gundel et al., 1993; Gundel & Heidberg, 2008), the extensive cognitive load, involved in computation over context (Stalnaker, 2014), and the general deficits of aMCI in a wide range of cognitive functions (e.g., memory, executive attention and reasoning). However, our results do not document a cognitive difficulty in reference itself. Specific referential interpretations provided by context are not engaged in our current results.

7.1.3. A pronoun problem

Our results also do not support an alternative hypothesis that “pronouns” per se, as a lexical or morphological unit, are the main source of language deficit in aMCI, although pronoun comprehension impairments exist in prodromal and early AD speech (e.g., Sherman et al., 2015). Various hypothesized explanations of the source of this “pronoun problem” range from lexical access problems to memory deterioration (Almor et al., 1999). Our results suggest, however, that it is not the pronoun as a morphological entity which is the source of such deterioration. Rather, it is the structural context in which pronouns appear and the anaphora construals they potentially reflect. aMCI individuals maintain pronouns where they do not involve free reference anaphora (e.g., the Principle C constrained RBB sentence), even while they productively eliminate them where omission leads to obligatory bound anaphora (e.g., in the CS experiment). In the ASC experiment, reformations mainly concern the NP antecedent in an anaphora relation, not the pronoun.²⁰,²¹

7.1.4. Binding principles

Our results add to a developing literature concerning the psychological reality of the linguistically defined dissociation between types of anaphora resolution, but force us to review the nature of this linguistic knowledge (Everaert et al., 2010; Branco et al., 2005). They confirm not only that anaphora construal by coreference and by binding can be dissociated, but that binding is “basic” (Reinhart, 1983a,b, 1986, 2006; Reuland, 2003).²² However, our results do not document a binding construal as the first “most economical” option chosen in anaphoric construal as executed in sentence processing. If they did, in the ASC experiment individuals could have simply bypassed the coreference option in sentences like RBF in favor of a construal by binding, eliminating the deficit in these sentences. Our results, therefore, reveal binding as a primitive of knowledge rather than of processing strategies.

Syntactically determined knowledge, i.e, knowledge that anaphoric construal by coreference is an option to construal by binding, appears to underlie the increased difficulty in aMCI. This implicates what has been termed “reference set computation”, which “requires holding two or more derivations open” (Reinhart, 2006, p. 11), and “comparing a set of derivations…” (Reinhart, 2006, p. 19), so that the speaker-hearer can meet the thought interface and attain propositional truth value.²³ “Computing this type of reasoning requires construction of a reference set which consists of <d, i> pairs of a derivation and interpretation” (Reinhart, 2006, p. 244). However, while our results implicate some form of ‘reference set computation’, our results do not evidence comparison of distinct interpretations as a means of resolving this computation. In addition, the failure of memory tests, including episodic and working memory, to predict linguistic performance in our aMCI population is inconsistent with the memory load required for reasoning required by such construal comparisons (see Reinhart 2006 for discussion). See also converging evidence from experimental study of effects of memory on sentence comprehension in individuals with AD (Rochon & Waters, 1994; Caplan et al., 2011, p. 439; Caplan & Waters, 2013).

7.2. Leading question now

Although our results suggest that two forms of anaphora are dissociated and available to the speaker-hearer at the Language Faculty interface, the mechanisms by which this linguistic knowledge links to language production remain illusive. These analyses go beyond what we can pursue in this paper. However, independent psycholinguistic and neurolinguistic studies of sentence processing have revealed several facts which cohere with our results and should be pursued in future research.

(i) A neuropsychological model has provided evidence that first stages of auditory sentence processing are completed within 500–1000 ms (Friederici, 2002, p. 79). Syntactic parsing is abstract early in the timeline of sentence processing, first providing only a skeleton of full sentence representation. Syntactic integration with sentential semantic relations, i.e., thematic relations derived from the lexicon and subject-predicate relations, occurs only late in the processing timeline, there integrating general memory and other cognitive resources. (ii) Contextual interpretation, necessary to coreferential anaphora, has been discovered to be delayed. For example, in “resolution of ambiguous pronouns…” “there appears to be no initial and automatic commitment to a certain interpretation of a pronoun upon encountering it” (see Karimi & Ferreira, 2016, p. 1024 for review). In reference assignment in sentences with either pronominal or elliptical anaphora, the delayed time allows pragmatic context to inform inference necessary for interpretation. (cf., Nagano et al., 2020; Mak & Sanders, 2010; Stevenson & Vitkovitch, 1986). Thus, in an initial syntactic parse, pronouns function as uninterpreted variables, supporting an abstract linguistic representation of pronouns in initial sentence parsing as proposed for example by Johnson (2013, p. 165).²⁴

(iii) Independent research has localized an AD deficit in the integration of reference in discourse, i.e., cross sentence anaphoric computation. This would also locate the deficit outside of essential sentence syntax psycholinguistically and neurally (e.g., Bornkessel-Schlesewsky, 2007). Seminal work by Kempler et al. (1999) reported AD individuals to be “impaired in their ability to maintain discourse (pronominal) dependencies over comparable distances” (p. 238) rather than in linguistic or memory deficits (cf., Sandoz et al., 2020; Berisha et al., 2015). More recently, a picture description task with AD and MCI found: “definite NP anomalies most commonly occurred in the use of pronouns, and often involved problems where a reference could not be linked, or was incorrectly linked, to a previously introduced referent, confirming a problem with anaphoricity” (Chapin et al. 2022, p.8, Lofgren & Hinzen, 2022).

Thus, in a comprehensive model of language knowledge and use, the linguistic structure-dependent knowledge allowing and dissociating two forms of anaphoric construal must be abstract and determined early in sentence processing. Straddling the interface, that linguistic knowledge must subsequently be integrated with contextual reasoning allowing reference to be determined, truth value to be computed and discourse to be conducted.

7.3. Clinical implications

Our results can potentially strengthen clinical evaluation in prodromal AD, complementing advances in biotyping (e.g., Sperling et al., 2011; Jang et al., 2022). More refined diagnosis of prodromal AD can be achieved by extending current standardized language testing to research-informed tests of complex sentences and the linguistic computation they involve. The controlled EI task is amenable to systematized application and automated scoring. More generally, clinical diagnosis can be strengthened by assessment that is based on consideration of the general cognitive models involved in particular cognitive functions, rather than on individual components of knowledge. (See for example, Sherman et al., 2021, 2020).

7.4. Limitations and future directions

Our cross sectional study investigated only one period of prodromal AD, i.e., MCI, and only one form of MCI, i.e., aMCI. Assessing prognostic value of our results will require longitudinal studies, and extended samples given the heterogeneity of both MCI and HA populations. To pursue mapping between language and neural foundations, comparisons can now be made to other forms of dementia, e.g., Primary Progressive Aphasia (Dickerson et al., 2008, 2011; Dickerson & Wolk, 2012; Mesulam et al., 2014; Katsumi et al., 2023) and other language pathologies (e.g., Broca’s aphasia, e.g., Vasić, 2006; Vasić et al., 2006), each revealing specific neurological foundations. Results can be compared across diverse methodologies (e.g, comprehension tasks such as in Sambin et al, 2012).

Pronominal anaphora which involves only coreference and not binding (e.g., Koornneef, 2010), should now be tested, as should other forms of binding. Further analyses are necessary comparing pronominal anaphora to various forms of null anaphora, as in nonfinite adverbial subordinate clauses (cf., footnote 20), thus further dissociating the role of the lexical pronoun from anaphora type.²⁵ The significant effects of Semantic Plausibility in this study, which was designed in terms of word association norms, motivate further study now in the domain of semantic representation, e.g. by defining semantic ‘plausibility’ in terms of predicates or propositions, or assessing the nature and role of potential contextual reference.

Finally, our experimental results motivate replication and extension of neuroscientific study. For example, brain scan analyses were conducted with a small number (6) of our aMCI population in contrast to a matched group of healthy adults (from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database) (Lust et al., 2016, 2017) following Friederici’s (2002, 2017) serial model of auditory sentence processing. That preliminary study began to test the hypothesis that aMCI language deterioration in syntax-semantics integration would predict the neural structure of prodromal AD atrophy. This pilot research can now be integrated with independent studies which, using event related MRI and/or ERP, currently investigate the neural foundations of reference and other semantic and pragmatic factors in sentence comprehension. (e.g., Nieuwland et al., 2007; Nieuwland & van Berkum, 2008; McMillan et al., 2012; Piñango & Burkhardt, 2004; Joyal et al., 2020).

8. Conclusions

In this paper, we have provided evidence for a specific linguistic deficit in language of amnestic Mild Cognitive Impairment (aMCI). This deficit is generally independent of effects of healthy aging; and of memory deficits in aMCI, although both aging and memory deficits exist in our population. Factorial experimental designs assessing anaphora in individuals’ production of complex sentences have clarified the nature of this deficit. This deficit resides not in complex sentence formation in general, not in anaphoric dependency relations in general, not in reference in general, not in pronouns per se, but specifically in the integration of reference into sentence syntax, i.e., in specific aspects of syntax-semantics integration. This deficit predominantly targets one form of anaphoric construal, that involving coreference, while preserving crucial aspects of anaphoric construal by binding. Our results confirm a dissociation between linguistic performance and memory, converging with results from earlier study of relative clauses in aMCI (e.g., Sherman et al., 2021) and suggesting that the observed deficits in aMCI across the three experimental studies (adverbial subordinate clause adjuncts, coordinate sentences and relative clauses) reveal a specifically linguistic effect.

We locate the source of deterioration in a general theory of the Language Faculty, i.e., the human competence for language, implicating deterioration in integration of the language-thought interface of the Language Faculty with general cognition, while aspects of syntax (the essence of the Language Faculty) remain relatively intact. In doing so, they help to locate aMCI deterioration in the general cognitive architecture of the mind and brain. These results have implications not only for informing clinical investigations of prodromal Alzheimer’s disease (AD) but for refining both current linguistic theory and neurolinguistic models of brain-language and brain-behavior mapping (e.g, Roger et al 2022, Friederici 2017, Hinzen 2013).

Figure 5.

Structure of a Pronoun (Johnson 2003, p. 165).

Highlights.

• Amnestic Mild Cognitive Impairment individuals have selective deficits in language

• aMCI show preservation of syntactic aspects of language

• aMCI show deficits in relating pronouns and their contextual referents

• Selective deterioration in aMCI shows a breakdown of language-thought relations

6. ABBREVIATIONS

ACE-R

Addenbrookes’s-revised (Addenbrook Cognitive Examination-Revised)

AD

Alzheimer’s Disease

ADL-Q

Activities of Daily Living–Questionnaire

ADRC

Alzheimer’s Disease Research Center

ADNI

Alzheimer’s Disease Neuroimaging Initiative

aMCI

amnestic subtype/MCI

ASC

adverbial subordinate clauses

B

Backward

BD

branching direction

BP

Brown-Peterson

BT

Binding Theory

CDR

clinical dementia rating

C-I

Conceptual-Intentional

CITRA

Cornell Institute for Translational Research on Aging

CS

coordinate sentences

CT

Coordination Type

EI

elicited imitation

F

Forward

HA

healthy aging

HY

Healthy Young

IPL

inferior parietal lobule

L

Left

MCI

Mild Cognitive Impairment

MGH

Massachusetts General Hospital

NART

North American Reading Test

NP

Noun Phrase

PAC

Psychology Assessment Center

POB

primitives of binding

PPA

Primary Progressive Aphasia

R

Right

RBB

right-branching backward

RB

right-branching

SP

semantic plausibility

STG

Superior Temporal Gyrus

UDS

Uniform Data Set

Data Availability

Experimental Designs, methods and materials have been archived at: Virtual Linguistics Lab Experiment Bank archive (https://testvcla.cac.cornell.edu/the-vll/experiment-bank/).