Rational adaptation in word production: Strong conceptual ability reduces the effect of lexical impairments on verb retrieval in aphasia

Abstract

Language users rely on both linguistic and conceptual processing abilities to efficiently comprehend or produce language. According to the principle of rational adaptation, the degree to which a cognitive system relies on one process vs. another can change under different conditions or disease states with the goal of optimizing behavior. In this study, we investigated rational adaptation in reliance on linguistic versus conceptual processing in aphasia, an acquired disorder of language. In individuals living with aphasia, verb-retrieval impairments are a pervasive deficit that negatively impacts communicative function. As such, we examined evidence of adaptation in verb production, using parallel measures to index impairment in two of verb naming’s critical subcomponents: conceptual and linguistic processing. These component processes were evaluated using a standardized assessment battery designed to contrast non-linguistic (picture input) and linguistic (word input) tasks of conceptual action knowledge. The results indicate that non-linguistic conceptual action processing can be impaired in people with aphasia and contributes to verb-retrieval impairments. Furthermore, relatively unimpaired conceptual action processing can ameliorate the influence of linguistic processing deficits on verb-retrieval impairments. These findings are consistent with rational adaptation accounts, indicating that conceptual processing plays a key role in language function and can be leveraged in rehabilitation to improve verb retrieval in adults with chronic aphasia.

1. Introduction

When people process language, they make use of various types of information, including linguistic representations and conceptual knowledge (Caramazza & Zurif, 1976; Dresang et al., 2018; Dresang, Warren, et al., 2021; Gibson et al., 2013; McRae & Matsuki, 2009; Warren & Dickey, 2021). According to models of speech production and lexical retrieval, linguistic and conceptual processing are both engaged even in the retrieval of a single word (Foygel & Dell, 2000; Roelofs, 2014; Ueno et al., 2011; Walker & Hickok, 2016; see model discussion below). A shared assumption of these models is that activation spreads simultaneously between lexical and conceptual layers, and critically does so in both directions. However, the differential weighting of linguistic versus conceptual processing can vary based on context, task demands, and disease states (Amichetti et al., 2016; Gibson et al., 2013; Rezaii et al., 2022). Consistent with this view, there is growing evidence that adults with aphasia, an acquired language disorder, may rely less on linguistic information and more heavily on conceptual, world, or event knowledge when accomplishing language tasks (Dresang, Warren, et al., 2021; Gibson et al., 2015; Hayes et al., 2015; Warren et al., 2017; see rational adaptation discussion below). In the current work, we look for supporting evidence in verb retrieval, using parallel measures to index impairment in two of verb retrieval’s critical subcomponents: conceptual and linguistic action processing.

Of note, investigating verb-retrieval impairments is important clinically. Verb-retrieval impairments are common results of both stroke and neurodegenerative disorders (Bak, 2013) and can significantly impede everyday communicative function (Rofes et al., 2015). There is substantial evidence that both linguistic and conceptual processing contribute to verb retrieval in both healthy adults and adults with aphasia. However, most of the evidence relevant to aphasia examines linguistic- and conceptual-processing contributions to action/verb comprehension. It is critical to extend this question to verb production, given that approximately 70% of individuals with aphasia experience verb-retrieval deficits on action picture naming tests (Mätzig et al., 2009). Therefore, one objective of this study was to characterize how linguistic and conceptual deficits each contribute to action-naming impairments in post-stroke aphasia. Doing so can shed light on the cognitive systems that support verb retrieval and underlie action-naming impairments, supporting both more specific diagnosis and more effective treatment of these impairments.

Figure 1 shows a version of Foygel and Dell’s (2000) model that illustrates the key stages of lexical access and the flow of information between these stages. Although Foygel and Dell’s model was developed for object naming or noun production, it should in principle apply to action naming as well. The processes or levels in this model (conceptual-semantic, lexical, and phonological) are commonly assumed to be shared between lexical production and comprehension, particularly the conceptual-lexical interface (Walker & Hickok, 2016). Although models differ as to whether lexical-to-phonological processes are shared between comprehension and production (Foygel & Dell, 2000) or are distinct (Ellis et al., 1988; Walker & Hickok, 2016), there is broad agreement that there is only a single conceptual-to-lexical mapping that supports both production and comprehension. Importantly, lexical-production impairments have been modeled computationally as deficits in the spread of activation between levels (Foygel & Dell, 2000; Schwartz et al., 2006), capturing the effects of conceptual, lexical, and phonological impairments on lexical production.

Figure 1.

Lexical access stages assessed in each task.

Notes: This model was adapted from Foygel and Dell’s model of impaired lexical access in speech production (2000). Black regions depict stages engaged during each task.

Figure 1A depicts the stages required for verb production in an action picture-naming task: picture input engages semantic processing, a lexical word representation (lemma) is activated, and then the participant must select phonemes to produce the verb aloud. Figure 1B depicts the stages required for lexical tasks, including word comparison and word attribute tasks (see Methods section for full description). These lexical tasks provide participants with lexical inputs, which activate corresponding lexical word representations, and then engage semantic representations associated with each word so that participants can make a semantic decision (e.g., “Which action would make the loudest sound?”). Finally, Figure 1C depicts the stages required for conceptual tasks, including picture comparison and picture attribute tasks (see Methods). These conceptual tasks provide participants with picture inputs, which engage semantic processing associated with the image so that participants can make a semantic decision.

In the current work, we take advantage of a set of carefully matched assessment measures that employ parallel stimuli in picture-based action/event-processing and word-based verb-processing tasks to index impairment in these distinct kinds of processing (Fiez & Tranel, 1997). The specific tasks used in this assessment battery are noted in Figure 1A–C above and are further described in the Methods below. The conceptual and lexical action-processing that are indexed by these measures should both contribute to action naming/verb retrieval given widely accepted models of lexical access (e.g., Foygel & Dell, 2000: see further discussion below). The standardized assessment battery (Fiez & Tranel, 1997; Kemmerer et al., 2001, 2012) we use was designed to measure both conceptual and linguistic processes relevant to action knowledge and verb production and comprehension. The battery consists of both lexical (language-based) and conceptual (picture-based) tasks. Kemmerer and colleagues (2012) tested a large sample of individuals with a variety of acquired brain injuries on this battery and found that across participants, there were widely varying patterns of associations and dissociations across the lexical and conceptual measures. This battery is congruent with a model of lexical access that assumes: 1) conceptual representations can be accessed from picture-based stimuli; 2) lexical representations can be accessed from word-based stimuli; and 3) bidirectional connections exist between lexical and conceptual representations, thus enabling picture naming (mapping from conceptual to lexical representations) or semantic judgments about words (mapping from lexical to conceptual representations).

Whereas the activation of word or phoneme/grapheme processing stages is necessary for the action naming and lexical tasks, it is not obligatory for conceptual tasks. However, these stages could still be activated, given evidence that people automatically retrieve the names and phonology of pictured objects even when they do not need to name them (e.g., Meyer et al., 2007; Oppermann et al., 2010). Indeed, people sometimes activate representations that are not obligatory when completing a task. For example, participants can accomplish lexical decision tasks using purely lexical processing, but they often show semantic effects especially for slower and more difficult decisions, suggesting that semantic representations are activated (Armstrong & Plaut, 2016). Whether word and/or phoneme/grapheme stages are activated during conceptual tasks is likely related to how quickly and automatically the representations in these stages are activated and how useful they are likely to be.

The possibility that lexical representations might be activated during conceptual processing tasks and might contribute to their accomplishment could help explain why Kemmerer et al. (2012) found such variability in patterns of impairment on matched lexical vs. conceptual tasks. If the lexical and action naming tasks engage representations of semantics, words, and graphemes/phonemes, whereas conceptual tasks engage only semantics, then individuals with language but not semantic impairment, e.g. many people with aphasia, should perform well on the conceptual tasks and poorly on the lexical and action naming tasks. But Kemmerer et al. found little to no evidence of clean patterns of impairment like this. Their findings of variability in impairment are more consistent with accounts under which multiple kinds of representations can be drawn upon to accomplish a task, and different people might use strategies that rely on different representations to different degrees (see Hirshorn et al., 2020 for evidence of this in reading).

One mechanism that could enable the Foygel and Dell (2000) model and its successors to capture the observed variability in task performance is bidirectional spreading activation. Bidirectional spreading of activation through the lexical network is a broadly accepted phenomenon (Goldrick & Rapp, 2002; Rapp & Goldrick, 2000; Roelofs, 2014), and it is key in the model of lexical access laid out above. Spreading activation would allow lexical and phonological/orthographic representations to be activated from conceptual representations, even when they are not obligatory for the task at hand. For example, the finding that lexical decisions are faster to more semantically contentful words has been attributed to their greater semantic activation driving more feedback to orthographic nodes (e.g., Pexman et al., 2002). Indeed, bidirectional spreading activation can create resonance or relatively stable constellations of activation within a network that correspond to frequently encountered patterns of activation across neural systems or representations (Kumaran et al., 2016; McClelland et al., 1995). Such constellations provide a potential mechanism of adaptative or compensatory support for an impaired system, because stronger or more precise activation patterns at one representational level may help guide and potentially even stabilize weaker or noisier activation patterns at another level. In this way, reliance on intact and accessible knowledge could increase and help support activation of less intact or accessible knowledge. More generally, this mechanism could underlie rational adaptation, an information theory principle that states that a cognitive system can optimize its behavior under different conditions or disease states by modifying the degree to which it relies on different types of information to accomplish a task (Anderson, 1991; Gibson et al., 2015; Howes et al., 2009; Warren et al., 2017).

If we extend this idea to people with aphasia and assume that spreading activation between lexical and conceptual representations may allow relatively preserved conceptual representations to engage/co-activate more-impaired lexical representations, this might suggest that heightened spreading activation would be adaptive in at least some people with aphasia. There is longstanding evidence that some people with aphasia can show a high degree of priming during lexical processing, even greater than age-matched controls (lexical priming paradigms: Milberg & Blumstein, 1981; Janse, 2006; visual-world paradigms: Mirman et al., 2011). Furthermore, there is evidence from verb processing specifically (lexical priming paradigms: Dresang et al., 2021) that individuals with apahsia show greater spreading activation between (i.e., priming from) associatively related nouns and verbs. These findings are consistent with the hypothesis that a heightened level of spreading activation – present in at least some people with aphasia – is an adaptive response of the impaired language-processing system.

Most previous approaches to investigating potential rational adaptation in aphasia have involved investigating whether people with aphasia, who by definition have language impairment, rely more on conceptual or world knowledge during language processing than neurotypical comparison participants do. Most of these studies have also focused on sentence-level performance in aphasia. Gibson et al. (2015) provided evidence consistent with this hypothesis from an act-out task indexing the interpretation of plausible and implausible prepositional-object and double-object sentences: participants with aphasia were more likely to settle on an interpretation in which event plausibility overrode linguistic cues than neurotypical controls were. Yet, the critical interaction between aphasia status and plausibility was not fully reliable in their study. Warren et al. (2017) replicated and extended Gibson et al.’s (2015) findings with a larger sample and a sentence-picture matching task, and in their data, the critical interaction was reliable. However, the binary nature of the sentence-picture matching task Warren et al. used raises concerns: if a participant chose the picture of a sensible event when the sentence they had heard described a bizarre event, it could indicate that they weighted event plausibility more heavily than linguistic cues and assumed the sentence they heard had been corrupted or it could indicate that the participant simply failed to process the sentence. Using a variant of the visual-world paradigm task, Hayes et al. (2015) found that under conditions in which young neurotypical adults showed evidence of relying on both linguistic cues and event knowledge cues to anticipate upcoming prepositional phrases, people with aphasia only showed evidence of relying on event knowledge cues. However, they tested many more neurotypical adults than people with aphasia, so power could have potentially played a role in this finding.

Dresang, Warren, et al. (2021) investigated potential tradeoffs between linguistic and conceptual representations in lexical processing in aphasia. They had participants with and without aphasia read aloud written verbs that were preceded (primed) by either a conceptually event-related noun (pencil – WRITE), a lexical collocate noun (name – WRITE), or an unrelated noun (water – WRITE). In accuracy measures, consistent with rational adaptation, adults with aphasia showed amplified priming effects from conceptual cues compared to the age-matched neurotypical comparison group, whereas the neurotypical group demonstrated greater priming from lexical cues. However, these findings are qualified by the fact that the accuracy of the neurotypical comparison group was very close to ceiling. The findings of each of these papers hint towards rational adaptation in aphasia, but each has a weakness.

Another approach to investigating rational adaptation in aphasia has been to measure an individual’s ability to do linguistic processing and conceptual processing, and to use each of these indices to predict their behavior on a language processing task. This approach has the benefit of being more targeted; indexing impairment in both conceptual and linguistic processing allows more precise predictions regarding their tradeoffs. Usually, linguistic impairment is measured via a standardized aphasia battery like the Comprehensive Aphasia Test (Swinburn et al., 2004) and conceptual impairment is measured via picture-based tasks like Pyramids and Palm Trees (Howard & Patterson, 1992), Kissing and Dancing Task (Bak & Hodges, 2003), or the Event Task (Dresang et al., 2019). Research using this approach in aphasia has also been mixed. Colvin et al. (2019) found that linguistic impairment marginally predicted sensitivity to a linguistic anomaly and conceptual impairment predicted sensitivity to semantic anomalies, but didn’t look for tradeoffs across them. Warren et al. (2017) found that increased linguistic impairment was associated with greater reliance on meaning than linguistic form in sentence interpretation, but found no relationship between conceptual impairment and reliance on meaning or linguistic form in sentence interpretation.

The current study was designed to focus on verb retrieval specifically, using the model of lexical access laid out in Figure 1A–C to test the relative contributions of lexical versus conceptual processing to verb retrieval in aphasia. The study’s design was also intended to address the limitations of prior research examining potential rational adaptation in aphasia, which has tended to use all-purpose measures of language and conceptual impairment that are only loosely related to the processes engaged in the studied tasks. To address this limitation, we look for evidence of adaptation in verb retrieval, using parallel measures to index impairment in two of verb retrieval’s critical subcomponents: conceptual and linguistic action processing. Using these more direct measures of impairment in processes critical to the experimental task will provide a better test than general-purpose measures of whether impairment in linguistic processing increases reliance on conceptual processing and vice versa. In the current study, we will relate impairment on these assessments to performance on a picture-based action-naming task, using these comparisons to examine interactions of lexical and conceptual processing in verb retrieval, and to look for evidence of rational adaptation in lexical processing in aphasia. Understanding these systems and the ways they might rationally adapt could also inform treatment strategy in patients with action-naming/verb-retrieval impairments. For example, if an individual with relatively severe linguistic deficits adapts to rely more on conceptual processing to support verb retrieval, then their naming ability might derive particular benefit from treatments that target conceptual-semantic processing, such as pantomimed gesture use (Dresang et al., 2023) or verb network strengthening treatments (Edmonds, 2016).

By evaluating how well lexical and conceptual task performance predict action-naming in a sample of people with and without aphasia, the current study expands on previous work targeting verb comprehension impairments in aphasia (Saygin et al., 2004). Furthermore, it evaluates three previously untested predictions. The first prediction is that deficits in lexical task performance (see Figure 1B) will be predictive of action-naming performance in individuals with aphasia. This prediction is consistent with previous findings that lexical factors influence verb production in aphasia (Cho-Reyes & Thompson, 2012; Gordon & Dell, 2003; Thorne & Faroqi-Shah, 2016), and is consistent with spreading activation accounts such as the model in Figure 1A–B (Foygel & Dell, 2000). If individuals are unable to activate target lexical representations or to map between corresponding lexical and semantic representations, then their action naming should also be impaired.

The second prediction is that deficits specific to conceptual task performance (i.e., the semantic processing stage, Figure 1C) will be predictive of action-naming performance in individuals with aphasia. This prediction is consistent with previous findings that verb-production impairments in aphasia are associated with performance on picture-based semantic-processing tasks (Dresang et al., 2019) and damage to motor-related white-matter tracts (Dresang, Hula, et al., 2021). It is also consistent with the lexical-production model in Figure 1A. If individuals are unable to activate target semantic representations, the first step in successful action naming, they will be impaired in their verb production.

The final prediction is the rational adaptation prediction that lexical and conceptual impairments will interact to predict verb production impairments in aphasia, with conceptual-processing ability attenuating the effects of lexical impairments on action naming and vice-versa. Rational adaptation models claim that the relative noisiness of a process or representation (in this case, lexical vs. conceptual processes) will affect whether it is used to accomplish a given task, with less noisy (more intact) processes having greater influence on task performance than noisier (less intact) processes. The specific mechanism behind the rational adaptation that we predict here is spreading activation from stronger to weaker conceptual, lexical, and phonological representations. The rational adaptation prediction is also consistent with the rationale behind aphasia treatments that leverage action concepts to improve verb production, including gesture-based treatments (e.g., Rose & Sussmilch, 2008) and Semantic Feature Analysis for actions (Wambaugh et al., 2014). Such treatments assume that increasing activation of relatively intact action concept representations can help remediate, compensate for, or circumvent language-specific impairments.

2. Methods

2.1. Participants

Participants were 17 individuals with chronic aphasia due to unilateral left hemisphere stroke and 15 age-matched neurotypical controls. All participants were 1) native English speakers, 2) able to provide informed consent, 3) 42–78 years old, 4) (pre-morbidly) right-handed, 5) had no significant hearing loss or vision impairment that prevented them from completing the experimental tasks, 6) had no pre-existing or subsequent brain injury/stroke (e.g., to right-hemisphere regions for individuals with aphasia), and 7) had no history of progressive neurological or psychiatric disease, drug, or alcohol dependence, or significant mood or behavioral disorder. In addition, all neurotypical participants passed a line-bisection visual screening, a binaural pure-tone hearing screening (0.5, 1, 2, and 4 KHz at 40 dB), a Mini-Mental State Examination cognitive screen (required 27/30; Folstein et al., 1975), and Raven’s Coloured Progressive Matrices non-linguistic cognitive screen (required 30/36; Raven, 1965).

All participants with aphasia were more than six months post-onset (range: 19–265 months; M=95.8, SD=62 months), had a Comprehensive Aphasia Test (CAT; (Swinburn et al., 2004) Naming Modality T-score ≥ 40, Cognitive Screening for recognition memory T-score ≥ 30, and an overall mean T-score < 70. Cognitive screening and general language assessment measures, including the CAT, were already available for the participants with aphasia, who had previously participated in Hula and colleagues’ study (Hula et al., 2020). Institutional Review Board approval was obtained, and all participants provided informed written consent and were compensated for their time. Demographic participant characteristics are reported in Table 1 for participants with aphasia and Table 2 for age-matched controls.

Table 1.

Demographic characteristics of participants with aphasia

Participant ID	Age	Sex	Education Level	Years of Education	Months Post-Onset	Years Post-Onset
7201	59	F	Graduate degree	20	132	11
7202	63	M	Bachelor’s degree	14	265	22.08
7203	61	F	Master’s degree	17	60	5
7204	55	M	High school	12	53	4.42
7205	52	M	High school	12	136	11.33
7206	78	F	Some graduate	13	114	9.5
7207	70	F	Some college	14	45	3.75
7208	76	M	Some college	14	138	11.5
7209	77	M	Law degree	19	53	4.42
7210	54	M	Bachelor’s degree	16	83	6.92
7211	71	M	Some college	14	26	2.17
7212	55	M	Bachelor’s degree	16	19	1.58
7213	68	M	High school	12	184	15.33
7214	53	F	Bachelor’s degree	17	81	6.75
7215	71	M	Bachelor’s degree	16	87	7.25
7216	72	M	Some college	14	60	5
7217	72	M	Some college	15	93	7.75
Summary	M=65.12	5 F; 12 M		M=15	M=95.82	M=7.99
	SD=9.11			SD=2.35	SD=62	SD=5.17

Table 2.

Demographic characteristics of age-matched control participants

Participant ID	Age	Sex	Education Level	Years of Education
7001	42	M	Tech college	14.5
7002	59	M	High school	12
7003	74	M	Bachelor’s degree	16
7004	52	M	Bachelor’s degree	16
7005	54	M	Bachelor’s degree	16
7006	57	M	High school	12
7007	72	F	Master’s degree	18
7008	64	F	Master’s degree	18
7010	74	M	Master’s degree	20
7011	68	F	Master’s degree	22
7012	72	M	Bachelor’s degree	16
7013	65	M	Law degree	19
7014	71	F	Master’s degree	17
7015	52	M	Master’s degree	22
7016	69	F	Master’s degree	18
Summary	M=63	5 F; 10 M		M=17.1
	SD=9.82			SD=3

2.2. Materials

All participants completed a set of behavioral language and conceptual knowledge assessments to test verb and action processing. A standardized action knowledge battery (Fiez & Tranel, 1997) was used to probe participants’ abilities to name actions as well as to retrieve lexical and conceptual information pertaining to actions. The action knowledge battery consisted of (1) a verb retrieval task (action picture naming), (2) two conceptual, picture-based tasks (picture attribute and picture comparison judgments), and (3) two linguistic, verb-based tasks (word attribute and word comparison judgments). See section 2.3 and Fiez and Tranel (1997) for task details. These tasks provide: a measure of verb retrieval via action naming, and a comprehensive and detailed classification of linguistic versus conceptual verb and action impairments via the matched conceptual and linguistic tasks. Following earlier work with these assessments (Kemmerer et al., 2012), performance on these tasks will henceforth be referred to as lexical versus conceptual action-processing abilities and impairments.

2.3. Testing procedure

The full battery of tasks was presented on a computer monitor, in the same order presented by Kemmerer and colleagues (2012): Naming, Picture Comparison, Picture Attribute, Word Comparison, and Word Attribute. Participants were given detailed instructions and practice items to ensure they understood each task. Brief descriptions of the tasks are provided below. Further details are reported in Fiez and Tranel (1997) and Kemmerer, Tranel, and Barrash (2001). For a summary of previous studies that have used tests from this battery, see Kemmerer and colleagues (2012). Participants viewed stimuli on computer screen. They were allowed an unlimited amount of time to respond to each item. An external microphone recorded naming responses using Audacity^® software, and trial-level accuracy was scored by hand for each task.

Naming Task (N = 100 items):

For each item, the participant was shown a single color-photograph of a person or animal performing an action. The participant was instructed to provide a single word/verb that describes what the person or animal was doing. The participant’s first response was recorded. Following Fiez and Tranel (1997), experimenters prompted participants for a second response if the participant did not provide a verb (e.g. “Can you tell me what the person is doing?”) or if the participant provided a description (e.g. “Can you give me a single word that best describes what the person is doing?”). Target responses provided after a prompt were also scored as correct. Alternate forms of a target verb were accepted as correct (e.g., run, running, ran). Normative data (from Fiez & Tranel, 1997): M = 85.0% correct; SD = 5.0.

Picture Comparison Task (N = 24 items):

For each item, the participant was shown three colored photographs depicting actions. The participant was instructed to determine which picture showed an action that was most different in meaning from the other two (e.g., cutting a pie, cutting a piece of paper, folding a piece of paper). This task is analogous to the Word Comparison Task. Normative data (from Fiez & Tranel, 1997): M = 83.6% correct; SD = 8.3.

Picture Attribute Task (N = 72 items):

For each item, the participant was shown two colored photographs depicting actions and an attribute judgment question. The experimenter read the question aloud. The participant was instructed to indicate which picture best answered the question (e.g., “Which action would make the loudest sound?”). This task is analogous to the Word Attribute Task. Normative data (from Fiez & Tranel, 1997): M = 91.7% correct; SD = 4.8.

Word Comparison Task (N = 44 items):

For each item, the participant was shown three printed verbs. The experimenter read the verbs aloud. The participant was instructed to determine which word was most different in meaning from the other two. This task is analogous to the Picture Comparison Task. Normative data (from Fiez & Tranel, 1997): M = 88.7% correct; SD = 8.1.

Word Attribute Task (N = 62 items):

For each item, the participant was shown two printed verbs and an attribute judgment question. The experimenter read the question and the verbs aloud. The participant was instructed to indicate which verb best answered the question (e.g., “Which action would make the loudest sound?”). This task is analogous to the Picture Attribute Task. Normative data (from Fiez & Tranel, 1997): M = 94.8% correct; SD = 2.6.

Participant performance on each task is reported in Table 3 for participants with aphasia and Table 4 for age-matched controls.

Table 3.

Conceptual and lexical action-processing performance (percent accuracy) for participants with aphasia

Participant ID	Action Naming	Conceptual Action-Processing		Lexical Action-Processing
		Picture Comparison	Picture Attribute	Word Comparison	Word Attribute	CAT Modality Mean T Score
7201	0.63	0.88	0.93	0.80	0.82	56.33
7202	0.31	0.58	0.58	0.34	0.53	47.83
7203	0.56	0.50	0.83	0.70	0.92	59.67
7204	0.77	0.79	0.86	0.75	0.84	56
7205	0.44	0.29	0.85	0.73	0.84	48.33
7206	0.89	0.63	0.89	0.86	0.94	66.67
7207	0.52	0.33	0.76	0.77	0.97	59
7208	0.13	0.25	0.69	0.39	0.60	47.83
7209	0.51	0.33	0.85	0.75	0.81	52.83
7210	0.59	0.75	0.81	0.34	0.73	51.33
7211	0.67	0.46	0.79	0.59	0.69	50.33
7212	0.53	0.92	0.88	0.77	0.81	51.17
7213	0.43	0.63	0.78	0.64	0.76	49.83
7214	0.67	0.54	0.86	0.57	0.90	62
7215	0.02	0.29	0.64	0.48	0.56	53.17
7216	0.35	0.54	0.86	0.70	0.82	42.5
7217	0.44	0.46	0.74	0.68	0.94	58.67
M	0.50	0.54	0.80	0.64	0.79	53.73
SD	0.22	0.21	0.09	0.16	0.13	6.13

Notes: CAT = The Comprehensive Aphasia Test (Swinburn et al., 2004) assesses aphasia severity. All other measures come from the verb and action processing battery (Fiez & Tranel, 1997).

Table 4.

Conceptual and lexical action-processing performance (percent accuracy) for age-matched controls

Participant ID	Action Naming	Conceptual Action-Processing		Lexical Action-Processing
		Picture Comparison	Picture Attribute	Word Comparison	Word Attribute
7001	0.9	0.96	0.94	0.93	0.97
7002	0.99	0.83	0.90	0.95	0.98
7003	0.9	0.92	0.92	0.86	0.94
7004	0.86	0.75	0.96	0.98	0.92
7005	0.94	0.88	0.93	0.80	0.98
7006	0.8	0.63	0.82	0.84	0.92
7007	0.92	0.92	0.94	0.95	0.94
7008	0.95	0.92	0.99	1	0.95
7010	0.92	0.79	0.94	0.93	0.98
7011	0.98	0.96	0.94	0.93	0.90
7012	0.95	0.96	0.92	0.91	0.98
7013	0.87	0.79	0.93	0.98	0.97
7014	0.93	1	0.94	0.93	0.95
7015	0.94	0.83	0.93	1	0.95
7016	0.89	0.79	0.94	0.95	0.98
M	0.92	0.86	0.93	0.93	0.95
SD	0.05	0.10	0.04	0.06	0.03

Notes: All measures come the verb and action processing battery (Fiez & Tranel, 1997).

2.4. Analyses

Data were analyzed using Bayesian multilevel logistic regression models, which were created in the Stan computational framework (Carpenter et al., 2017; http://mc-stan.org/) and accessed with the brms package (Bürkner, 2017). The dependent variable in these models was item-level action naming accuracy expressed as a dichotomous variable. Standardized composite scores for conceptual action-processing performance (Picture Comparison, Picture Attribute) and lexical action-processing performance (Word Comparison, Word Attribute) were entered as fixed-effect predictors. These composite scores were developed as follows: Individual proportion correct scores for the Picture Comparison, Picture Attribute, Word Comparison, and Word Attribute tasks underwent empirical logit transformations, and these empirical logit scores were z-score transformed and assessed in terms of the assumptions of normality, homoscedasticity, linearity, and the presence of outliers. The z-scores were then averaged for Picture Comparison and Picture Attribute to create the conceptual action-processing composite score and for Word Comparison and Word Attribute to create the lexical action-processing composite score. This model structure allows examination of the effects of conceptual versus lexical action-processing skills on verb retrieval (action naming) ability. Grouping performance on conceptual versus lexical tasks is supported both by the design of the standardized battery (Fiez & Tranel, 1997) and by a factor analysis conducted in a large sample of neurologically-impaired adults, which indicated that the conceptual (picture-based), lexical (verb-based), and action naming tasks each loaded on separate factors (Kemmerer et al., 2001). Refer to the results and Appendix A for correlations in the current sample.

We ran two models, one including both the aphasia group and a neurotypical comparison group and one with only the aphasia group. In each model, we tested our predictions related to rational adaptation interactions and main effects of lexical and conceptual action-processing on verb retrieval. In Model 1, naming performance was compared between participant groups.¹ Fixed effects were a comprehensive three-way interaction effect (group × conceptual z-scores × lexical z-scores), such that all main effects and interactions among the following variables were examined: group assignment (participants with aphasia coded as 0 vs. neurotypical controls coded as 1), conceptual action-processing composite z-scores, and lexical action-processing composite z-scores. Random intercepts were included for subjects and items. Model 1 parallels previous approaches (Dresang, Warren, et al., 2021; Gibson et al., 2015; Warren et al., 2017) that investigated whether people with aphasia (i.e., individuals with language impairment) rely more on conceptual processing than neurotypical comparison participants do.

Model 2 examined aphasic performance in greater detail, while controlling for overall aphasia severity.² Fixed effects were a comprehensive three-way interaction effect (severity × conceptual z-score × lexical z-score), such that all main effects and interactions among the following variables were examined: aphasia severity (CAT Modality Mean T Score), conceptual action-processing composite z-scores, and lexical action-processing composite z-scores. Random intercepts were included for subjects and items. The structure of Model 2 parallels Model 1, with aphasia severity taking the place of group because only participants with aphasia were included in Model 2. Model 2 thus controls for overall aphasia severity to examine the specific effects, and potential rational adaptation, of conceptual and lexical abilities on action naming.

For both models, each parameter was given dispersed starting values and a vague prior using the default options in brms. There was a student t(3,0,2.5) prior distribution on the model intercepts and a flat uniform prior on the betas (see Appendix B for Stan model code and prior specification). This allowed the Bayesian estimation process to explore the full parameter space and provide conservative estimates of posterior distributions (McElreath, 2018). For each model, four Hamiltonian Bernoulli family Markov chain Monte Carlo (MCMC) chains were run for 20,000 samples, with half of the iterations discarded as warm-up and 10,000 iterations monitored for convergence and parameter estimation. MCMC convergence was assessed graphically by inspection of the autocorrelation and trace plots (Appendix C) and statistically using the Gelman-Rubin potential scale reduction statistic (Ȓ) and the number of effective samples (ESS; Tables 5 and 6). The Ȓ statistic is a ratio of the variance within each chain to the variance pooled across chains. Ȓ values close to 1 indicate satisfactory convergence of the chains to a stable distribution (Gelman et al., 2013). ESS factors out the autocorrelation in the observed MCMC chains and estimates the number of independent samples that would achieve the same degree of precision for the parameter estimates (Carpenter et al., 2017). Large ESS values indicate satisfactory convergence. The posterior distributions are summarized by the estimated parameters and 95% highest density credible intervals (HDI). The HDI is conceptually similar to the frequentist confidence interval but is defined such that it contains only values with higher probability density than all values outside the interval (Meredith & Kruschke, 2020).

Table 5.

Model 1 population-level effects for participants with aphasia and age-matched control participants

	Estimate	Est. Error	Lower 95% HDI	Upper 95% HDI	Ȓ	Bulk ESS	Tail ESS
(Intercept)	0.05	0.28	−0.49	0.59	1	4048	5900
Group	3.26	0.38	2.54	4.07	1	6655	6545
Conceptual action-processing	0.82	0.39	0.09	1.62	1	4479	5873
Lexical action-processing	0.31	0.38	−0.45	1.03	1	4021	4686
Group : Conceptual	−0.54	0.58	−1.77	0.55	1	4944	5132
Group : Lexical	−0.3	0.53	−1.32	0.76	1	4159	4539
Conceptual : Lexical	−0.26	0.44	−1.13	0.60	1	3284	4700
Group : Conceptual : Lexical	−0.11	0.65	−1.47	1.12	1	3159	4559

Notes: Composite scores for conceptual action-processing and lexical action-processing were calculated by an empirical-logit and z-score transformation per participant. HDI=Highest density interval. Ȓ=The potential scale reduction factor on split chains (at convergence, Ȓ= 1). ESS=Effective sample size.

Table 6.

Model 2 population-level effects for participants with aphasia

	Estimate	Est. Error	Lower 95% HDI	Upper 95% HDI	Ȓ	Bulk ESS	Tail ESS
(Intercept)	−0.02	0.32	−0.66	0.62	1	1431	1895
Aphasia severity	0.14	0.35	−0.52	0.85	1	1128	1225
Conceptual action-processing	0.93	0.52	−0.01	2.00	1	1455	1601
Lexical action-processing	−0.02	0.49	−1.00	0.99	1	1511	1146
Severity : Conceptual	1.05	0.82	−0.52	2.76	1	1039	942
Severity : Lexical	0.42	0.52	−0.60	1.49	1	1292	1064
Conceptual : Lexical	−1.15	0.68	−2.53	0.15	1	884	1156
Severity : Conceptual : Lexical	−0.03	0.62	−1.25	1.27	1	1671	1784

Notes: Aphasia severity was z-score transformed CAT modality mean score. Composite scores for conceptual action-processing and lexical action-processing were calculated by an empirical-logit and z-score transformation per participant. HDI=Highest density interval. Ȓ=The potential scale reduction factor on split chains (at convergence, Ȓ = 1). ESS=Effective sample size.

3. Results

The current sample of participants with aphasia demonstrated correlations between the conceptual action-processing tasks (Picture Comparison and Picture Attribute: r = 0.78, p < 0.001) and between the lexical action-processing tasks (Word Comparison and Word Attribute: r = 0.78, p < 0.001). These results support the chosen approach to calculate conceptual and lexical action-processing z-scores. In addition, correlations were examined between study variables that were entered as regression predictors. Although action naming was correlated with both conceptual (r = 0.67, p = 0.003) and lexical z-scores (r = 0.63, p = 0.007), the correlation between the conceptual and lexical z-scores did not reach statistical significance in this sample of participants with aphasia (r = 0.42, p = 0.10). At the same time, Bayesian analysis of the conceptual and lexical z-score correlation suggests a 91% probability that the correlation is positive. See Appendix A for a full task correlation matrix.

3.1. Model 1: Effect of group on action naming performance

To illustrate the magnitude of group differences in performance, Figure 2 shows untransformed group performance (percent accuracy) on each set of tasks: action naming, lexical tasks, and conceptual tasks. These data are also reported at the participant level in Tables 3–4.

Figure 2.

Group-level performance across tasks.

Notes: All group differences were statistically significant (action naming: t = 7.31, p < 0.001; lexical tasks: t = 6.12, p < 0.001, and conceptual tasks: t = 6.01, p < 0.001).

The trace plots for all parameters demonstrated rapid convergence and were stationary relative to the parameter means. There were no divergent transitions. The autocorrelation plots corroborated this assessment and showed minimal autocorrelation. These plots are provided in Appendix C1. The Ȓ statistic and number of effective samples for each parameter indicated satisfactory convergence and MCMC mixing. These statistics are reported in Table 5. Table 5 also provides the point estimates, estimated error (EE), and 95% highest density credible intervals for each parameter. A graphical posterior predictive check of the extent to which the observed data could plausibly have been generated by the estimated model is shown in Figure 3. Histograms of the posterior distributions for the estimates discussed below are provided in Figure 4(A–E).

Figure 3.

Posterior predictive check for Model 1

Figure 4.

Posterior distributions and 95% highest density intervals (HDIs) of (A) the fixed effect of group, (B) the fixed effect of conceptual action-processing, (C) the interaction effect of group and lexical action-processing, (D) the interaction effect of group and conceptual action-processing, and (E) the interaction effect of conceptual and lexical action-processing (results from Model 1, participants with aphasia and healthy controls).

Notes: Dashed lines mark the 95% highest density intervals (HDIs) for the posterior distribution.

As expected, group (aphasia versus control) reliably predicted trial-level action naming accuracy (β = 3.26, EE = 0.38, 95% HDI = [2.54, 4.07]; Figure 4A), with participants with aphasia (M = 0.495, SD = 0.208) performing less well than controls (M = 0.916, SD = 0.047). There was also a reliable main effect of conceptual action-processing ability on action naming, such that higher conceptual processing was associated with higher action naming (β = 0.82, EE = 0.39, 95% HDI = [0.09, 1.62]; Figure 4B). In contrast, there was no reliable main effect of lexical action-processing ability on action naming (β = 0.31, EE = 0.38, 95% HDI = [−0.45, 1.03]). The credible intervals for all interaction effects contained zero (Figure 4C–D); but, the direction of the weak, unreliable effects suggests that both of the individual component processes, conceptual (β = −0.54, EE = 0.58, 95% HDI = [−1.77, 0.55]) and lexical action-processing ability (β = −0.30, EE = 0.53, 95% HDI = [−1.32, 0.76]), might have a stronger effect on action naming success for individuals with aphasia compared to neurotypicals. Similarly, there was unreliable evidence of conceptual and lexical action-processing interactions, providing weak support that higher conceptual ability may reduce the effect of lexical impairments on action naming, and vice versa (β = −0.26, EE = 0.44, 95% HDI = [−1.13, 0.60]). However again, the credible interval indicated that this interaction effect was not reliably negative (posterior probability that the effect was < 0 = 71%; Figure 4E). The full set of results is reported in Table 5.

3.2. Model 2: Conceptual and lexical action-processing associations with action-naming impairments in adults with aphasia

The trace plots for all parameters demonstrated rapid convergence and were stationary relative to the parameter means. There were no divergent transitions. The autocorrelation plots corroborated this assessment and showed minimal autocorrelation. These plots are provided in Appendix C2. The Ȓ statistic and number of effective samples for each parameter indicated satisfactory convergence and MCMC mixing. These statistics are reported in Table 6. Table 6 also provides the point estimates and 95% credible intervals for each parameter. The posterior predictive check is shown in Figure 5. Histograms of the posterior distributions for the estimates discussed below are provided in Figure 6(A–D).

Figure 5.

Posterior predictive check for Model 2

Figure 6.

Posterior distributions and 95% highest density intervals (HDIs) of (A) the fixed effect of conceptual action-processing, (B) the interaction effect of conceptual and lexical action-processing, (C) the interaction effect of aphasia severity and conceptual action-processing, and (D) the interaction effect of aphasia severity and lexical action-processing (results from Model 2, participants with aphasia).

Notes: Dashed lines mark the 95% highest density intervals (HDIs) for the posterior distribution.

Parallel to results from Model 1, conceptual action-processing ability robustly predicted action naming in participants with aphasia (β = 0.93, EE = 0.52, 95% HDI = [−0.01, 2.00]; Figure 6A), but lexical action-processing ability did not (β = −0.02, EE = 0.49, 95% HDI = [−1.00, 0.99]). Furthermore, conceptual and lexical action-processing abilities interacted to predict naming, such that higher conceptual ability reduced the effect of lexical impairments on action naming, and vice versa (β = −1.15, EE = 0.68, 95% HDI = [−2.53, 0.15]; Figure 6B). This interaction effect is further illustrated in Figure 7. Despite the small overlap of the credible interval with zero, this lexical × conceptual interaction effect was robust, with a 94.4 percent posterior probability of being less than zero. This provides substantial evidence for the presence of an interaction between lexical and conceptual processing abilities. In addition, Model 2 results indicate that the degree of aphasia severity (as measured by the CAT) did not reliably interact with either conceptual or lexical action-processing ability, but there was weak evidence that milder aphasia amplified the effect of both conceptual (β = 1.05, EE = 0.82, 95% HDI = [−0.52, 2.76]; Figure 6C) and lexical ability (β = 0.42, EE = 0.52, 95% HDI = [−0.69, 1.49]; Figure 6D) on action naming. Specifically, 92% and 83% of the posterior distributions, respectively, exceeded zero. Finally, the three-way interaction between aphasia severity, conceptual, and lexical action-processing did not reliably predict action naming in participants with aphasia (β = −0.03, EE = 0.62, 95% HDI = [−1.25, 1.27]). The full set of results is reported in Table 6.

Figure 7.

Interaction effect of conceptual and lexical action-processing on verb naming in participants with aphasia. Trend lines and standard error were computed from posterior samples of the fitted model (results from Model 2, participants with aphasia). Colored points were overlayed to show participants’ raw scores.

4. Discussion

This study examined how deficits in lexical and conceptual action-processing contribute to action-naming impairments in aphasia. We evaluated whether variability in action naming performance might relate to differences in strength of lexical and conceptual action-processing abilities. Furthermore, we investigated whether there is evidence of rational adaptation in the linguistic system of people with aphasia, such that better conceptual ability attenuates the effects of lexical impairments on action naming and better lexical ability attenuates the effects of conceptual impairment on action naming. Such findings would be consistent with the hypothesis that spreading activation from relatively strong processes (e.g., semantics) to more impaired processes (e.g. lexical levels) may facilitate verb retrieval by helping to guide weaker or noisier activation patterns. Although lexical and conceptual action-processing deficits often co-occur in neurologically impaired individuals, they are fully dissociable (Kemmerer et al., 2001, 2012), and there is substantial evidence that both lexical properties and conceptual representations help to characterize verb-retrieval deficits (Bak, 2013; Bak & Chandran, 2012). However, to our knowledge, this is the first study to directly examine how both lexical and conceptual processing abilities predict action naming in aphasia, as well as how rational adaptation might occur between these subcomponent processes of verb retrieval.

Following previous work investigating the rational adaptation prediction that compared to participants in a neurotypical comparison group, participants with aphasia will rely more heavily on conceptual processing when engaged in language tasks (e.g., Dresang, Warren, et al., 2021; Gibson et al, 2015; Warren et al, 2017), we started with an analysis comparing participants with aphasia to an age-matched neurotypical comparison group. However, it is important to note that unlike previous studies, our outcome variable did not directly index reliance on conceptual or linguistic processing during the experimental task. Instead, we investigated the degree to which conceptual or lexical impairment predicted performance on our experimental task. This means that the relatively small amount of variability in impairment in the neurotypical group could impair our ability to find effects in this analysis. As expected, participants with aphasia performed less well on the naming task, consistent with previous findings that action naming (verb retrieval) is commonly impaired in aphasia (Mätzig et al., 2009). Conceptual action-processing ability positively predicted action naming across the entire sample, but lexical action-processing did not. No interactions were reliable. However, there were trends toward two-way interactions between group and conceptual action-processing, as well as group and lexical action-processing, which provided weak evidence that both conceptual and lexical action-processing abilities are more influential for successful action naming in individuals with aphasia than in neurotypicals. There was also an unreliable trend providing weak evidence that higher conceptual ability may reduce the effect of lexical impairments on action naming, and vice versa.

Our second analysis took advantage of the fact that our sample of participants with aphasia displayed high variability in conceptual and lexical impairment, providing a critical opportunity to examine potential tradeoffs and rational adaptation in reliance on these systems. This analysis examined potential interactions between aphasia severity (rather than dichotomous absence vs. presence of aphasia), conceptual, and lexical action-processing within the group of participants with aphasia. Like in the previous analysis, conceptual action-processing ability positively predicted action naming, but lexical action-processing did not. As predicted by the principle of rational adaptation, we found that conceptual and lexical action-processing abilities interacted to predict naming, such that higher conceptual ability reduced the effect of lexical impairments on action naming, and vice versa. The evidence supporting this interaction between lexical and conceptual processing is limited, given that there was slight overlap of the credible interval with zero (94.4% posterior probability < 0). The lack of a fully reliable effect to support rational adaptation hypotheses may be attributed to the current sample size being underpowered to detect interaction effects. There was no evidence that aphasia severity interacted with conceptual or lexical action-processing abilities to predict action naming. Aphasia severity is often associated with verb-retrieval impairments, with increased aphasia severity being associated with greater verb-retrieval impairments (e.g., Saygin et al., 2004). Indeed, our sample also showed these correlations between overall aphasia severity and verb retrieval on the action naming task. However, in the regression model for this experiment, the lexical and conceptual scores indexed component processes engaged during verb retrieval; these scores are thus more proximal measures than overall language ability and consequently stronger predictors of verb-retrieval impairments. Even in the presence of a general correlation between aphasia severity and verb retrieval, our results suggest that relative strengths and weaknesses of the conceptual and lexical processes required for verb retrieval are more important in accounting for verb retrieval variability than gross severity is. Nonetheless, further investigations should consider participant samples with a greater range of aphasia severity to examine this effect more comprehensively. For example, a sample that includes individuals with both more and less severe aphasia would better characterize performance at very low lexical abilities and may be more likely to identify aphasia severity as a robust predictor.

The fact that both analyses showed a robust and reliable effect of conceptual action-processing ability on action naming, such that higher conceptual processing was associated with higher action naming, highlights the importance of conceptual-deficit contributions to aphasia’s hallmark naming impairments. This is consistent with the word-production model in Figure 1A–C: in such a model, increased activation of the relevant conceptual/semantic features during inspection of the photograph should drive greater activation of the target lexical word representations. In contrast, impaired or failed activation of the relevant conceptual/semantic representations (the first step in action naming) is more likely to result in unsuccessful action naming. However, the effects of conceptual-semantic deficits may also be offset by relative strengths in lexical processing, where strong lexical activation can spread to other layers and help compensate for deficits at the conceptual processing stage, thus facilitating action naming. As such, individuals with aphasia may rationally adapt to their specific impairments via bidirectional spreading activation between conceptual and lexical processing stages of verb retrieval, as indicated by the interaction between lexical and conceptual processing ability.

Models of speech production and lexical retrieval have traditionally accounted for aphasic performance by assuming reduced activation (e.g., lower connectionist weights) between levels of processing (Foygel & Dell, 2000; Roelofs, 2014; Ueno et al., 2011; Walker & Hickok, 2016). However, there is growing evidence that under certain conditions, some individuals with aphasia show greater lexical priming effects (i.e., greater spreading activation) than cognitively healthy peers (Dresang, Warren, et al., 2021; Janse, 2006; Mirman et al., 2011). Rational adaptation accounts posit that increased activation may not only be a feature of aphasic language impairments, but perhaps a beneficial adaptation to injury that allows for compensation between stages of cognitive processing. The current results find evidence for this specifically between conceptual and lexical action-processing abilities that support verb retrieval during picture naming. Heightened spreading activation between conceptual and lexical representations may thus enable at least some people with aphasia to use relatively preserved conceptual representations to engage relatively impaired lexical representations (or vice versa), to support action naming.

Whereas previous work has focused on characterizing conceptual and lexical action-processing deficits in aphasia and other neurological disorders, this study is the first to tease apart the independent contributions and potential interactions of these deficits in action-naming impairments in aphasia. The current findings provide promising evidence of interactivity and rational adaptation between the conceptual and lexical processing systems, with strong conceptual skills attenuating or outweighing the effects of lexical impairments on action naming, and vice versa. It is difficult to disentangle conceptual and lexical processing influences in heterogeneous patient populations such as aphasia. We expect that individuals with aphasia will show a wide variety of deficient activation (e.g., reduced lexical activation often seen in Broca’s aphasia; McNellis & Blumstein, 2001), maladaptive over-activation (e.g., impaired suppression often seen in Wernicke’s aphasia; McNellis & Blumstein, 2001), or beneficial over-activation (consistent with rational adaptation predictions). Clear variability can be observed within our aphasia sample (e.g., Figure 7). However, more definitive evidence may come from future studies that compare participants for whom lexical and conceptual processing more clearly dissociate (e.g., Kemmerer et al., 2012). The current sample was limited in its range of aphasia severity, missing individuals with very severe and very mild (latent) aphasia. There were also no participants with aphasia who had very low lexical action-processing and high conceptual action-processing abilities. It will be important for future research to test rational adaptation predictions in subjects who fit this profile. Nevertheless, the current findings underline the importance of interactions between lexical and conceptual systems to characterizing verb-retrieval performance in aphasia and other neurological conditions (Bak, 2013; Pulvermüller, 2018; Vigliocco et al., 2011).

The current results also support a critical message for rehabilitation: treatments targeting conceptual-semantic performance may be especially beneficial for ameliorating action-naming deficits in adults with stroke aphasia (at least across a heterogeneous sample). A weak trend in the data that should be followed up on in future work suggests that this may especially hold for adults with relatively mild aphasia. The current findings are consistent with the hypothesized mechanism of action underlying efficacious speech-language treatments like Semantic Feature Analysis (SFA; Boyle, 2010; Boyle & Coehlo, 1995), SFA for Actions (Wambaugh et al., 2014), Verb Network Strengthening Treatment (VNeST; Edmonds, 2016), and gesture-based approaches (Dresang et al., 2023; Rose & Sussmilch, 2008). These interventions aim to stimulate action- or event-concept representations to remediate, compensate for, or circumvent language-specific impairments and promote improved word production. The current findings suggest that better access to such conceptual representations is indeed associated with better action-naming or verb-production performance and may offset the effects of lexical-processing impairments. But these findings also hint at potential ways that taking a rational adaptation perspective might eventually lead to improved treatment-targeting. Although some aphasia treatments target linguistic processing and others target conceptual processing, it remains largely unclear who will respond best to each treatment approach. Characterizing interactions between linguistic and conceptual component deficits at the individual level may reveal patient profiles that are most likely to respond to conceptual-semantic versus linguistically based interventions for language rehabilitation, thus improving weak and inconsistent treatment outcomes.

Highlights.

• Rational adaptation (RA) affects reliance on linguistic vs. conceptual knowledge.

• We tested if RA explains verb retrieval performance in people with aphasia (PWA).

• Non-linguistic access to conceptual knowledge strongly predicted verb retrieval.

• Strong conceptual access offset linguistic-deficit effects on PWA’s verb retrieval.

• PWA rationally adapt to rely on conceptual knowledge to facilitate verb retrieval.

Appendix A

Table A.

Correlation matrix for task performance and severity for participants with aphasia

	Action Naming	Picture Comparison	Picture Attribute	Word Comparison	Word Attribute	CAT Modality Mean
Action Naming	1
Picture Comparison	0.54	1
Picture Attribute	0.73	0.78	1
Word Comparison	0.56	0.21	0.73	1
Word Attribute	0.66	0.15	0.69	0.78	1
CAT Modality Mean	0.62	0.14	0.35	0.42	0.61	1

Notes: All correlations are significant at p<0.001 after Bonferroni corrections for multiple comparisons. Naming, Picture Comparison, Picture Attribute, Word Comparison, and Word Attribute measures are correlated based on percent of correct trials (Fiez & Tranel, 1997). CAT = The Comprehensive Aphasia Test Modality Mean T Score (Swinburn et al., 2004) measures aphasia severity.

Appendix B

Table B.1.

Prior Summary for Model 1

prior	class	coef	group	source
(flat)	b	conceptual_Z		default
(flat)	b	conceptual_Z:linguistic_Z		(vectorized)
(flat)	b	group		(vectorized)
(flat)	b	group:conceptual_Z		(vectorized)
(flat)	b	group:conceptual_Z:linguistic_Z		(vectorized)
(flat)	b	group:linguistic_Z		(vectorized)
(flat)	b	linguistic_Z		(vectorized)
(flat)	b	conceptual_Z		(vectorized)
student_t(3, 0, 2.5)	intercept			default
student_t(3, 0, 2.5)	sd			default
student_t(3, 0, 2.5)	sd		item	(vectorized)
student_t(3, 0, 2.5)	sd	group	item	(vectorized)
student_t(3, 0, 2.5)	sd	intercept	item	(vectorized)
student_t(3, 0, 2.5)	sd		participant	(vectorized)
student_t(3, 0, 2.5)	sd	conceptual_Z	participant	(vectorized)
student_t(3, 0, 2.5)	sd	intercept	participant	(vectorized)
student_t(3, 0, 2.5)	sd	linguistic_Z	participant	(vectorized)

Table B.2.

Prior Summary for Model 2

prior	class	coef	group	source
(flat)	b	CAT_Z		default
(flat)	b	CAT_Z:conceptual_Z		(vectorized)
(flat)	b	CAT_Z:conceptual_Z:linguistic_Z		(vectorized)
(flat)	b	CAT_Z:linguistic_Z		(vectorized)
(flat)	b	conceptual_Z		(vectorized)
(flat)	b	conceptual_Z:linguistic_Z		(vectorized)
(flat)	b	linguistic_Z		(vectorized)
(flat)	b	CAT_Z		(vectorized)
student_t(3, 0, 2.5)	intercept			default
student_t(3, 0, 2.5)	sd			default
student_t(3, 0, 2.5)	sd		item	(vectorized)
student_t(3, 0, 2.5)	sd	CAT_Z	item	(vectorized)
student_t(3, 0, 2.5)	sd	intercept	item	(vectorized)
student_t(3, 0, 2.5)	sd		participant	(vectorized)
student_t(3, 0, 2.5)	sd	conceptual_Z	participant	(vectorized)
student_t(3, 0, 2.5)	sd	intercept	participant	(vectorized)
student_t(3, 0, 2.5)	sd	linguistic_Z	participant	(vectorized)

Extracted Stan model code & prior specification

Stancode function output (identical for models 1 & 2):
// generated with brms 2.18.0
functions {
}
data {
 int<lower=1> N; // total number of observations
 int Y[N]; // response variable
 int<lower=1> K; // number of population-level effects
 matrix[N, K] X; // population-level design matrix
 // data for group-level effects of ID 1
 int<lower=1> N_1; // number of grouping levels
 int<lower=1> M_1; // number of coefficients per level
 int<lower=1> J_1[N]; // grouping indicator per observation
 // group-level predictor values
 vector[N] Z_1_1;
 vector[N] Z_1_2;
 // data for group-level effects of ID 2
 int<lower=1> N_2; // number of grouping levels
 int<lower=1> M_2; // number of coefficients per level
 int<lower=1> J_2[N]; // grouping indicator per observation
 // group-level predictor values
 vector[N] Z_2_1;
 vector[N] Z_2_2;
 vector[N] Z_2_3;
 int prior_only; // should the likelihood be ignored?
}
transformed data {
 int Kc = K - 1;
 matrix[N, Kc] Xc; // centered version of X without an intercept
 vector[Kc] means_X; // column means of X before centering
 for (i in 2:K) {
  means_X[i - 1] = mean(X[, i]);
  Xc[, i - 1] = X[, i] - means_X[i - 1];
 }
}
parameters {
 vector[Kc] b; // population-level effects
 real Intercept; // temporary intercept for centered predictors
 vector<lower=0>[M_1] sd_1; // group-level standard deviations
 vector[N_1] z_1[M_1]; // standardized group-level effects
 vector<lower=0>[M_2] sd_2; // group-level standard deviations
 vector[N_2] z_2[M_2]; // standardized group-level effects
}
transformed parameters {
 vector[N_1] r_1_1; // actual group-level effects
 vector[N_1] r_1_2; // actual group-level effects
 vector[N_2] r_2_1; // actual group-level effects
 vector[N_2] r_2_2; // actual group-level effects
 vector[N_2] r_2_3; // actual group-level effects
 real lprior = 0; // prior contributions to the log posterior
 r_1_1 = (sd_1[1] * (z_1[1]));
 r_1_2 = (sd_1[2] * (z_1[2]));
 r_2_1 = (sd_2[1] * (z_2[1]));
 r_2_2 = (sd_2[2] * (z_2[2]));
 r_2_3 = (sd_2[3] * (z_2[3]));
 lprior += student_t_lpdf(Intercept | 3, 0, 2.5);
 lprior += student_t_lpdf(sd_1 | 3, 0, 2.5)
  - 2 * student_t_lccdf(0 | 3, 0, 2.5);
 lprior += student_t_lpdf(sd_2 | 3, 0, 2.5)
  - 3 * student_t_lccdf(0 | 3, 0, 2.5);
}
model {
 // likelihood including constants
 if (!prior_only) {
  // initialize linear predictor term
  vector[N] mu = rep_vector(0.0, N);
  mu += Intercept;
  for (n in 1:N) {
   // add more terms to the linear predictor
   mu[n] += r_1_1[J_1[n]] * Z_1_1[n] + r_1_2[J_1[n]] * Z_1_2[n] + r_2_1[J_2[n]] * Z_2_1[n] +
r_2_2[J_2[n]] * Z_2_2[n] + r_2_3[J_2[n]] * Z_2_3[n];
  }
  target += bernoulli_logit_glm_lpmf(Y | Xc, mu, b);
 }
 // priors including constants
 target += lprior;
 target += std_normal_lpdf(z_1[1]);
 target += std_normal_lpdf(z_1[2]);
 target += std_normal_lpdf(z_2[1]);
 target += std_normal_lpdf(z_2[2]);
 target += std_normal_lpdf(z_2[3]);
}
generated quantities {
 // actual population-level intercept
 real b_Intercept = Intercept - dot_product(means_X, b);
}

Appendix C

C.1.

Trace and density plots for MCMC samples for Model 1

C.2.

Trace and density plots for MCMC samples for Model 2