Assessing abstract thought and its relation to language with a new nonverbal paradigm: Evidence from aphasia

Abstract

In recent years, language has been shown to play a number of important cognitive roles over and above the communication of thoughts. One hypothesis gaining support is that language facilitates thought about abstract categories, such as democracy or prediction. To test this proposal, a novel set of semantic memory task trials, designed for assessing abstract thought non-linguistically, were normed for levels of abstractness. The trials were rated as more or less abstract to the degree that answering them required the participant to abstract away from both perceptual features and common setting associations corresponding to the target image. The normed materials were then used with a population of people with aphasia to assess the relationship of abstract thought to language. While the language-impaired group with aphasia showed lower overall accuracy and longer response times than controls in general, of special note is that their response times were significantly longer as a function of a trial's degree of abstractness. Further, the aphasia group's response times in reporting their degree of confidence (a separate, metacognitive measure) were negatively correlated with their language production abilities, with lower language scores predicting longer metacognitive response times. These results provide some support for the hypothesis that language is an important aid to abstract thought and to metacognition about abstract thought.

Introduction

Most animals think, in some sense of the word ‘think.’ What, if anything, is distinctive of human thought? A common answer is that humans are uniquely capable of abstract thought, reflected in the abstract lexicon of our spoken languages. For example, in English, words like ‘integrity,’ ‘philosophy,’ ‘love,’ ‘justice,’ and ‘mind’ mark intuitively abstract notions. This invites the question: what is the relationship between abstract thought and the words we use to express it?

An early and still influential answer to this question assigns to language a merely communicative role, with words enabling the expression of thoughts, while being inessential to abstract thought itself (Fodor, 1975; Pinker, 1994). On this view, someone lacking language could think all the same thoughts as a fluent speaker without being able to put their thoughts into words—not even “in their head,” as inner speech. Recently, however, a competing hypothesis sees language as not only a vehicle for communicating thoughts but as a resource that supports or enables certain forms of thought (Binder, Westbury, McKiernan, Possing, & Medler, 2005; Borghi et al., 2017; Condry & Spelke, 2008; Dove, 2014; Langland-Hassan, Gauker, Richardson, Dietz, & Faries, 2017; Lupyan & Bergen, 2016; Wang, Conder, Blitzer, & Shinkareva, 2010; Yee, 2019). In particular, language is seen by many as a crucial support or tool for abstract thought (Borghi et al., 2017; Boroditsky, 2001; Davis & Yee, 2019; Thibodeau, Hendricks, & Boroditsky, 2017). On some of these views, linguistic labels and word associations provide a kind of cognitive shortcut to abstract conceptual information (Barsalou, 2008; Wilson-Mendenhall, Simmons, Martin, & Barsalou, 2013). According to others, language facilitates thoughts about abstract relationships—especially when a relationship's holding between two entities can be difficult to grasp in purely sensory terms (Borghi et al., 2019; Boroditsky, 2001; Dove, 2018; Gentner & Boroditsky, 2001; Lupyan & Lewis, 2019; Reilly, Westbury, Kean, & Peelle, 2012). On other views, language serves a more fundamental role as a representational medium, providing an “internalized amodal symbol system” needed for abstract thought (Dove, 2014, Dove, 2019). These by no means exhaust the current menu of options (for review, see Borghi et al. (2017), Yee (2019), and Bolognesi and Steen (2018)).

The present study aims to build on this existing work by investigating the relation of language to a specific form of abstract thought—namely, the ability to abstract away from the perceptible features and salient thematic associations of two items to grasp some other commonality. To that end, we compare the performance of a group of people with aphasia (“PWA”) to that of controls on a novel set of non-verbal stimuli, designed and normed for the purpose of assessing this kind of abstract thought. This paper reports both the norming process itself and the subsequent experiment with PWA. In the balance of this introduction, we will situate the study's conception of abstract thought with respect to other common conceptions in the literature, as a means to motivating the measure of abstract thought developed in the norming study. A general theme is that existing measures of abstract thought are typically tied to words in a way that makes them unideal for use in a setting where the relationship of abstract thought to language itself is being assessed. We then explain (Section 1.7) why we predict that a capacity for the kind abstract thought assessed here—and for metacognitive awareness of such abstract thought—would be disrupted by a loss of language abilities.

A difficulty in studying abstract thought lies in specifying what qualifies a form of thought as “abstract” in the relevant sense (Gilead, Trope, & Liberman, 2020). If we adopt the common view that concepts are the building blocks of thought—with thoughts being composed of concepts—we can ask what it is that makes a concept abstract. This question belies a misunderstanding, however, as all concepts are abstract in a sense, insofar as grouping any set of individuals under a single concept involves abstracting away from properties they do not share to focus on ones they do (for review, see Borghi et al., 2019; Gilead, Trope, & Liberman, 2020; Yee, Jones, & McRae, 2018). To recognize both a black Poodle and a Golden Retriever as dogs, for instance, we need to abstract away from—i.e., ignore—their differences in color and shape to focus on their shared doghood.

Nevertheless, a relative sense of abstractness, whereby some concepts are more abstract than others, can be defined. A common means for doing so is by appeal to the relative concreteness or imageability of words associated with the concept (Brysbaert, Warriner, & Kuperman, 2014; Coltheart, 1981; Cortese & Fugett, 2004). Concreteness is typically understood as the extent to which a particular item or event can be experienced by the senses; a concreteness rating for a word is generated by averaging the values participants give when asked to assess how easy it is to perceive the item named by the word (Brysbaert et al., 2014; Medler & Binder, 2005). This way of defining abstract thought gives rise to the well-known “concreteness effect,” whereby words with higher concreteness ratings are processed faster in lexical decision tasks and are associated with better performance in naming and recall (Begg & Paivio, 1969; Kounios & Holcomb, 1994; Schwanenflugel & Shoben, 1983). Comparable results have been found with respect to the related measure of imageability, where imageability is understood as the subjective ease with which a word gives rise to a related sensory-motor mental image (Cortese & Fugett, 2004; Paivio, 1971). (A third related, but less commonly used, measure is Juhasz and Yap's (2013) Sensory Experience Ratings (SER).¹)

A straightforward way to assign a degree of abstractness to a concept is to associate its abstractness with the concreteness, imageability, or SER rating given to the word that expresses the concept. A limitation of this method is that, in being tied to specific words, such ratings are of limited use in assessing the relation of abstract thought to words themselves. If words—with their associated ratings—are used in experimental stimuli, there is a risk of conflating a capacity to process linguistic items with a capacity for abstract thought. On the other hand, if the experimental stimuli exclude words, it can be difficult to assess which concepts are triggered by a stimulus—and, thus, which if any concreteness, imageability, or SER ratings should be used to rate the abstractness of the stimulus. The novel measure of abstractness and concreteness developed below (which we call “Trial Concreteness”) aims to overcome these problems, while respecting motivations of these word-related rating systems.

A second commonly discussed notion of abstract thought derives from hierarchical relationships among categories, with concepts of superordinate (i.e. more inclusive) categories being rated as more abstract than concepts of subordinate (i.e. less inclusive) categories (Rosch, 1978; Yee, 2019). For instance, in the hierarchical sense of abstractness, the concept object is more abstract than the concept missile, because the category of objects is superordinate to that of missiles: all missiles are objects, but not all objects are missiles. Concepts themselves can be seen as stores of semantic knowledge arrived at through processes of abstraction from regularities in sensory input and related motor outputs (Yee, 2019). A concept of a superordinate category—such as object—will tend to be more abstracted from such sensorimotor correlates than a concept corresponding to one of its subordinate categories (such as chair), rendering the concept itself “more abstract.” In short, the more abstracted a store of knowledge is from its sensorimotor correlates, the more abstract the concept is that constitutes that knowledge.

While it may be tempting to view concepts that are relatively abstract, in this hierarchical sense, as corresponding to words with low concreteness or imageability ratings, the relation between the two is not straightforward. Whether the members of a given category share many, or only a few, salient perceptible similarities—and thus whether that category is highly abstracted from related sensorimotor information—is not what subjects are asked to assess when providing imageability or concreteness ratings for a word. Following Paivio (1968), Cortese and Fugett solicit imageability ratings by informing participants that “any word that…arouses a mental image…very quickly and easily should be given a high imagery rating,” and “any word that arouses a mental image with difficulty or not at all should be given a low imagery rating” (2004, p. 387). Similarly, in soliciting concreteness ratings, Brysbaert et al. (2014) explain to subjects that they should assign a high concreteness rating to a word to the extent that it “refers to something that exists in reality; you can have immediate experience of it through your senses,” while low concreteness should be assigned to words that “refer to something you cannot experience directly through your senses or actions” (2014, p. 906). Prima facie, there is no reason it should be more difficult to form images of members of superordinate as opposed to subordinate categories, and no reason members of a subordinate category should be judged to exist in reality, and to be perceptible, to a greater degree than members of superordinate categories. After all, to form an image of (or to perceive) a chair is simultaneously to form an image of (or to perceive) a member of the subordinate category ‘chair’ and the superordinate category ‘object.’

Nevertheless, participants often appear to answer imageability and concreteness prompts as though they are being asked to assess hierarchical relations. The noun ‘object,’ for example, has imageability and concreteness ratings of 408 and 487, respectively, compared to its subordinate ‘cat’, which has corresponding ratings of 617 and 615 (Coltheart, 1981), despite the fact that cats are themselves objects and therefore cannot be more real, or more readily perceptible than objects. Likewise, ‘object’ receives only a 2.2 SER rating, while ‘missile’ has a 4.45 SER rating, even though missiles are themselves objects and thus cannot be easier to form a sensory image of than an object. (Similarly, on the Brysbaert et al. (2014) concreteness ratings, ‘object’ receives a score of 3.66, whereas ‘missile’ has a score of 4.83.) Thus, whether or not it is warranted by the nature of the prompts used to solicit such ratings, the kind of abstractness expressed by concreteness, imageability, and SER ratings aligns fairly well with the kind that is associated with hierarchical relations among categories, where concepts of superordinate categories are more abstract than concepts of subordinate categories, because they are stores of knowledge whose sensorimotor correlates abstract-away from past experience to a greater degree.²

As with imageability, concreteness, and SER scores, there are limitations to using the superordinate-to-subordinate relation as a measure of abstractness when assessing the relation of language to abstract thought. One limitation is that such relations are far clearer within particular hierarchies than across them, limiting the ability to compare a participant's performance on a wide range of concepts with different levels of abstractness. ‘Labrador’ is subordinate to ‘dog,’ for instance, and therefore less abstract in the hierarchical sense. But which of ‘Labrador’ or ‘screwdriver’ is more abstract? We are not aware of any objective ratings that would enable such comparisons. A second limitation is that it is not always possible to place concepts within meaningful hierarchies, and thus not possible to assign hierarchy-based abstractness ratings. This is most obvious for verb and adjective concepts, but extends also to numerous putatively abstract noun concepts, such as concept, hope, philosophy, and justice. A third limitation is that, in the context of a non-verbal stimulus, it can be difficult to discern which category concept is elicited by a stimulus, and thus which hierarchical level of category is relevant to rating the abstractness of the stimulus. The measure of abstract thought developed below overcomes these limitations, while again preserving the connection between the notion of abstract thought and the process of abstracting away from past perceptual experience.

A third approach to defining abstract thought—especially influential to the present study—can be found in Lupyan and Mirman (2013) and Perry and Lupyan (2017). In place of an abstract/concrete concept distinction, Lupyan and Mirman (2013) explore differences in what they term “low-dimensional” and “high-dimensional” categories, comparing performance between a population with aphasia and controls. Their distinction between high and low dimensional categories itself echoes the notion of dense versus sparse categories (Sloutsky, 2010) and the distinction between resemblance (or association)-based categories and rule-based categories (Couchman, Coutinho, & Smith, 2010; Minda, Desroches, & Church, 2008). On Sloutsky's account, statistically dense categories “have multiple intercorrelated (or covarying features) relevant for category membership,” while sparse categories have members that share “very few relevant features” (2010, p. 1250). Similarly, a high-dimensional category, on Lupyan and Mirman's understanding, is one where the things united under the category share many salient features, while a low-dimensional category groups items that share only one or a few salient characteristics.

On its face, the proposal that low dimensional categories are more abstract than high dimensional aligns well with the idea that superordinate categories are more abstract than subordinate ones. Superordinate categories compare to low dimensional categories in that their members tend to share fewer salient features than subordinate (and high dimensional) categories. However, there is an important difference in these two ways of rating abstractness. The superordinate/subordinate distinction applies to categories. Whereas, properly understood, high and low dimensionality apply to individual trials, and not to specific categories.³ Each of Lupyan and Mirman's trials presented participants with twenty images of familiar objects such as foods, vehicles, tools, and animals. Participants were then asked to select images that met a certain criterion. In an example of a low-dimensional trial, participants were asked to identify, from among the twenty images, all and only the things that are blue. Here the idea was that objects so grouped would have little or nothing in common other than their color, making the trial low-dimensional. To successfully group all of the blue items in Lupyan and Mirman's “things that are blue” trial, the blue objects' many differences must be ignored—they must be abstracted-away from—while only their color remains a point of focus. By contrast, in an example of a high-dimensional trial, participants were asked to identify all the pictured fruit. Unlike ‘things that are blue,’ members of the category ‘fruit’ share multiple salient properties, such as being edible, sweet, found in the produce department, and alive.

But consider now a hypothetical trial where participants are asked to group ‘things that are yellow’ (another of Lupyan and Mirman's low-dimensional trials), yet where all of the yellow items among the choices are bananas, and all of the distractor items are vehicles. The trial could in that case be considered high dimensional, insofar as many different properties (shape, flavor, use, common setting) could serve to anchor selection of the correct (yellow) items. It is only in a context where, relative to the distractor items, the correct choices share no salient similarities other than their color that ‘things that are yellow’ becomes a low-dimensional trial. Going in the other direction, suppose that, on the ‘fruit’ trial, all of the distractor items are vegetables that share rough visual similarities with fruits and that fruits with uncommon flavor characteristics (e.g., tomatoes and eggplants) are among the correct choices. What was formerly a high-dimensional trial is now low dimensional due to the similarity of the distractor items to the correct choices. The close perceptual and thematic similarities of the distractor items to the correct choices forces participants to abstract away from almost all of the salient properties of the correct choices to focus on just one that unites them—viz., being a fruit—in order to arrive at the correct grouping.

Note, however, that while the degree of abstractness pertaining to high and low dimensionality may be trial-relative (i.e., relative to distractor items), there remains a close kinship between this notion of abstractness and that pertaining to hierarchical category relations. In both cases, the degree of abstractness increases as there are fewer salient features uniting a class of items. We have simply observed that, in some cases, features of a context—and not simply the uniting category itself—can play a role in determining how many salient features must be abstracted away from in order to appreciate a commonality between two or more things. In the next section, we introduce the term ‘Trial Concreteness’ as a measure of this sort of trial-relative abstract thought and relate it more definitively to traditional measures of (what we will call) ‘Concept Concreteness,’ which is the concreteness of the (trial-independent) concept that links the target and match.

The trial-relativity of low and high dimensionality is especially important to the present study, as it is likely to occur within many non-verbal assessments of abstract thought. Such assessments are, in turn, important to investigating the relation of language to abstract thought. In particular, trial-relativity occurs within standard pictorial semantic memory tasks, which serve as a framework for the test of abstract thought developed here. On a pictorial semantic memory task—such as the Camel and Cactus Test (CCT) (Bozeat, Lambon Ralph, Patterson, Garrard, & Hodges, 2000)—a target image is shown with four choice images below it, and the participant is asked to indicate which choice image best goes with the target image. While existing semantic memory tasks of this sort are not rated for the level of abstract thought they require, it is natural to think that some such trials will require more abstraction from past perceptual regularities—and, correspondingly, use of concepts that are themselves more abstract—than others. The aim of our norming study was to generate abstractness ratings for multiple trials of that sort, spanning a wide range of abstractness levels. To do so properly, the relativity of abstractness level—as a function of the distractor items—must be taken into account.

To see this vividly, suppose that the target image on a pictorial semantic memory trial is of a pear and the image it is to be matched with is of an apple (both being fruit). In a situation where the three distractor images are all of tools—a hammer, wrench, and screwdriver, for example—there are many salient features the pear shares with the apple that can facilitate linking one with the other. For that reason, we could say that it is a high-dimensional and (to extend ordinary usage somewhat) relatively concrete trial. But now consider a situation where the three distractor images are of vegetables—an artichoke, carrot, and onion—that share many features with fruits. In that context there are fewer salient characteristics shared only by the pear and apple to facilitate linking one with the other. It has become a lower-dimensional, more abstract trial. Nevertheless, in both cases, the concept linking target and match is the same—namely, fruit. Thus, the degree of abstraction required to arrive at the correct choice in a trial can vary while the concept, label, or category linking the target and match remains the same.

It is helpful, then, to distinguish two senses of concreteness, each with its corresponding notion of abstractness. First, there is Concept Concreteness, which is concreteness linked to individual words or concepts. Concreteness ratings, imageability ratings, and SER ratings are all ways of measuring concept concreteness, so understood. In addition, while there are no standard numerical scores corresponding to the hierarchical notion of concreteness earlier discussed, a category's place in a hierarchy can be considered an additional measure of the concreteness of the concept corresponding to that category.

But we can also speak of Trial Concreteness, where a semantic memory trial (of the sort just described) has high concreteness if there are many salient features shared by the target and match that, in the context, can be used to distinguish them as falling under a single category, and is abstract to the extent that there are very few features that can be used to so distinguish them. The two imagined versions of a fruit-related semantic memory task just described have different levels of Trial Concreteness, due to the difference in the distractor items. Yet, in each case, the concept (fruit) linking the target and match is the same. In that sense, we can say the Concept Concreteness of each task is the same.

The ability of Trial Concreteness ratings to differ as a function of context meshes with the idea that concepts themselves are context-dependent, insofar as there is no static representational structure exploited across contexts that triggers (what otherwise might seem to be) the same concept (Barsalou, 1987; Casasanto & Lupyan, 2015; Yee & Thompson-Schill, 2016). Instead, it is proposed, concepts are “constantly changing” and “inextricably linked to their context” (Yee & Thompson-Schill, 2016). The difference in Trial Concreteness in the two fruit trials described above tracks some of this context-relativity, insofar as it captures the way in which exercising (or triggering) a concept may require more or less abstraction from present sensory input in different contexts.

In the norming study described below, we generated Trial Concreteness ratings for 86 pictorial semantic memory trials, with lower Trial Concreteness ratings corresponding to more abstract trials. We identified two dimensions along which a trial might differ in Trial Concreteness. First, a trial could be more concrete to the extent that the target and match share many visually perceptible similarities, in comparison to the target and distractor items. This aligns well with the sense in which subordinate categories are more concrete than superordinate ones (on the assumption that members of subordinate categories share more perceptible features than members of superordinate categories). It also aligns with the sense in which categories with low imageability, concreteness, and SER scores are relatively abstract, at least insofar as categories are rated lower in concreteness or imageability when their members share fewer salient perceptible features.

A second dimension of abstractness relevant to semantic memory tasks consists in whether the two matching items are often found in a common setting and, for that reason, are strongly associated. A fork and a plate, for instance, do not share many visual similarities. However, they are thematically associated due to their typically appearing together at a commonly experienced type of event. Thematic connections of this sort are highly salient and strongly shape perceiver expectations. They are learned earlier than superordinate categorical relationships (Markman, 1981, Markman, 1990), and both children and adults default toward sorting items by thematic relationships (as opposed to categorical or taxonomic ones) when not given a word by which to sort (Lin & Murphy, 2001; Markman, 1990). Further, adults tend to sort items more quickly by thematic relation than by functional category (such as “footwear”) (Kalénine et al., 2009; Kalénine, Mirman, Middleton, & Buxbaum, 2012). And, more generally, the presentation of a word or image primes recognition of thematically related items (Estes & Jones, 2009; Mirman & Graziano, 2012).

In view of the high saliency of thematic associations, we propose that, just as a process of abstraction is required in order to sort perceptually dissimilar items together into a superordinate category (where one abstracts away from the salient perceptual features of the target to match it with another item), so, too, is abstraction involved in linking items that do not tend to occur together in a commonly experienced type of event. In the latter case, the participant must abstract away from the perceptual features not of things of the same categorical kind, but of things that, in one's experience, are commonly found together with the item. While this is not the very same notion of abstractness that is tracked by imageability and concreteness ratings, it is relatively well-aligned with the hierarchical notion of abstractness. Plausibly, the further a category moves in the direction of being superordinate in relation to others, the less likely it is that its members will share salient perceptual or theme-related features.

Thus, while we recognize this as a novel approach, we propose that when considering the overall abstractness level of a pictorial semantic memory task, one should take account of both dimensions—visual similarity and common setting—simultaneously. One reason that “abstracting away from common settings” is sometimes overlooked in assessments of abstract cognition is that existing measures of Concept Concreteness are linked to single words (as opposed to phrases), and, as Markman (1990) observes, we typically lack single words to refer to thematic relationships, or event types. This leaves any rating system linked to individual words unable to measure the sort of abstraction involved in abstracting away from common thematic relationships.

(Notably, Barsalou's (1983) category of ad hoc concepts includes many thematic categories; however, there are no standard concreteness or imageability ratings available for ad hoc concepts. Note, also, that we should not expect all thematic categories to be equally concrete or abstract. For those who frequently camp, we can expect the ad hoc category ‘Things to take on a camping trip’ to be relatively concrete, in the sense that its members are very closely associated due to their typically appearing together in a frequently experienced kind of event. By contrast, the ad hoc category ‘things to take from one's home during a fire,’ will, for most, not unite items that are frequently experienced together in a common setting, simply because such events are rarely experienced. However, for someone who never goes camping but has the misfortune of experiencing many house fires, the category ‘things to take from one's home during a fire’ will have more strongly associated members—and linking them will require less abstraction-away from regularities in past experience—than ‘things to take on a camping trip.’)

Accordingly, in the norming study described below, we assigned Trial Concreteness ratings to individual semantic memory trials by summing a visual similarity score (relating to how visually similar the target and stimulus are, relative to the target and distractor items) with a common setting score (relating to how frequently the target and match are found together in a common setting, relative to the target and distractor items). In our view, this allows for a more complete measure of the degree of abstraction-from-past-experience required for properly answering a trial than either taken alone. When a trial has low Trial Concreteness ratings—and is therefore highly abstract—there will be few salient features shared by the target and match that can serve to alert the participant to the fact that they go together. By contrast, were one only to take account of visual similarity in generating Trial Concreteness ratings, relatively simple trials where, for instance, a fork is matched with a plate, could be rated as “highly abstract” due to the lack of visual similarity between the two. Intuitively, this is the wrong result. The frequent cooccurrence of forks and plates within a common theme or setting suggests, to the contrary, that trials linking them should be viewed as relatively concrete, in the sense that answering them requires relatively little abstraction from past experience (modulo the distractor items).

Finally, it bears noting that, in the context of a pictorial semantic memory task, trials will tend to become more difficult as Trial Concreteness decreases. This is because, as trials decrease in Trial Concreteness, there are fewer salient perceptual or theme-related features shared only by the target and match to serve as cues for making the correct selection. (The same is likely true as the category linking target and match becomes more superordinate with respect to others.) While it may be possible to dissociate difficulty from Trial Concreteness (see Section 4.1), we did not attempt to keep difficulty constant while manipulating Trial Concreteness in our stimuli. Instead, we used a mediation analysis (Section 3.4.3) to explore whether Trial Concreteness affects performance over and above the contribution of difficulty.

The trials developed and normed for Trial Concreteness in the norming study do not incorporate words as stimuli or require them as responses. This makes them suitable for use with a language-impaired population in assessing whether their language deficits lead to corresponding deficits in abstract thought. The main experiment presented here compares a group of people with aphasia (“PWA”) to age-, education-, and gender-matched controls on their performance in selecting the correct matching image on the normed trials. Should PWA, despite their language impairment, show facility with categorizations on trials that are very low in Trial Concreteness, this would provide some evidence for the language-independence of abstract thought. Likewise, if PWA show greater difficulties than controls with stimuli low in Trial Concreteness compared to those high in Trial Concreteness, this would give reason to think that language is an important resource for abstract thought.

Our main prediction was that PWA would indeed show pronounced difficulties, compared to controls, on trials with low Trial Concreteness. That is, while lower performance overall is to be expected in the PWA population, due to their having experienced a stroke, we predicted that their relative difficulties would be proportionately more pronounced for trials low in Trial Concreteness—resulting in proportionately lower accuracy, longer response times, and lower confidence as the concreteness of a trial decreases. This is because we shared with Lupyan and Mirman (2013) and Lupyan and Lewis (2019) the hypothesis that, in cases where only one or very few salient features serve to link two items, being able to produce the linguistic label for that feature will promote recognition of the link. Evidence for linguistic labels serving as a cognitive support for linking items in the absence of other salient perceptual or thematic connections is reported by a number of others, including Davis and Yee (2019), Sloutsky and Deng (2019), Louwerse (2018), and Vigliocco, Ponari, and Norbury (2018). Accordingly, we predicted that, for instance, those who cannot produce the words ‘forecast’ or ‘predict’ may have difficulty grouping two things as both involved in forecasting or predicting, provided that the target and its match lack other salient properties that could alert one to the correct grouping (such as being commonly found together, or being visually similar). Such results would cohere with Sloutsky's (2010) thesis that “language provides learners with an additional set of cues that allow them to form more abstract distinctions” (p. 1248).

If language supports abstract thought in roughly these ways, we could expect abilities with related acts of abstraction to vary as a function of one's linguistic capacity. Therefore, we also predicted that the severity of language production impairments of the PWA—as measured by sub-components of the Western Aphasia Battery-R (Kertesz, 2006)—would correlate with their accuracy, confidence, and response times on the main task.

We did not, however, expect that our various measures of Concept Concreteness would correlate with PWA or control success at categorization to the same degree as with Trial Concreteness ratings, because we see Trial Concreteness as a finer-grained and more comprehensive measure of the kind of abstract thought required by the trials (in ways discussed above). Nevertheless, we thought it worthwhile to explore the relative effect of each. Note, however, that in order to compare the effects of Trial Concreteness and Concept Concreteness on performance, it is necessary to associate a single word with each of our (non-linguistic) trials. That way, the ratings for that word can provide the relevant Concept Concreteness rating for the trial. We call this the ‘Linking Word’ for each trial, which was determined through a norming process described below. In some cases, however, associating a single word with the trial may be somewhat artificial (such as for thematic connections, as discussed above)—a point we return to in our discussion of the norming study (Section 2.5).

In addition to facilitating abstract thought, it has been proposed that language—sometimes in the form of “inner speech”—also supports increased levels of self-awareness (Carruthers, 2018; Morin, 2009), more accurate self-monitoring (Alderson-Day & Fernyhough, 2015; Jones & Fernyhough, 2007), and better and more comprehensive knowledge of one's own mental states (Bermudez, 2018; Carruthers, 2011; Clark, 1998; Langland-Hassan, 2014), including abstract concepts in particular (Borghi, 2020). An earlier study found preliminary evidence that on-line language use (in the form of inner speech) is an important resource for metacognitive self-assessments with respect to abstract categorizations (Langland-Hassan et al., 2017). To investigate this possible link, we included a second prompt on the main experimental stimuli querying participants' confidence in their responses. This metacognitive question was aimed at assessing whether language has a role in increasing metacognitive accuracy distinct from any it may play in facilitating categorization itself, and whether this role is especially acute in trials with low Trial Concreteness.

We predicted that a disproportionate effect of low Trial Concreteness on PWA would show itself in this metacognitive aspect of the trials as well, with PWA showing proportionately lower confidence, and longer response times in reporting confidence, compared to controls, as Trial Concreteness ratings decreased. This would be in keeping with the results of Langland-Hassan et al. (2017), who, using a similar paradigm, found PWA to be impaired, relative to controls, in the accuracy of their metacognitive judgments with respect to whether they had correctly categorized items by an abstract category.