Boundaries to grounding abstract concepts

Abstract

Grounded theories of cognition claim that concept representation relies on the systems for perception and action. The sensory-motor grounding of abstract concepts presents a challenge for these theories. Some accounts propose that abstract concepts are indirectly grounded via image schemas or situations. Recent research, however, indicates that the role of sensory-motor processing for concrete concepts may be limited, providing evidence against the idea that abstract concepts are grounded via concrete concepts. Hybrid models that combine language and sensory-motor experience may provide a more viable account of abstract and concrete representations. We propose that sensory-motor grounding is important during acquisition and provides structure to concepts. Later activation of concepts relies on this structure but does not necessarily involve sensory-motor processing. Language is needed to create coherent concepts from diverse sensory-motor experiences.

This article is part of the theme issue ‘Varieties of abstract concepts: development, use and representation in the brain’.

1. Introduction

Grounded theories of cognition face a major challenge in explaining how abstract concepts can be grounded in sensory-motor experiences [1,2]. Most, if not all, concepts are abstract in some way. Some concepts have abstract meanings in the sense that they represent entities without physical features, such as justice or reason. Other concepts have concrete meanings but are abstractions to some degree because they generalize across specific exemplars. Apples and cars come in all kinds of physical formats, and can be further abstracted into fruits and vehicles or natural objects and artefacts [3–6]. Different mechanisms have been proposed to explain grounded representation of abstract concepts, sometimes dealing with a particular subset of abstract concepts [4]. Recent approaches have considered conceptual representation as a combination of sensory-motor and language processing [2,7,8]. Such hybrid models of concepts are attractive because they may explain evidence for grounded concepts but also evidence that is problematic for the grounded view. They may even offer a mechanism for the flexibility of concepts that seem to be grounded in some circumstances but not in others. Before discussing these hybrid models, let's consider two of the main other explanations of abstract concepts, Conceptual metaphor theory and Situated cognition, and to what extent they might explain sensory-motor grounding.

2. Conceptual metaphor theory

Cognitive linguists have proposed that linguistic metaphors reveal the structural relations between concrete and abstract concepts [9–14]. For example, people may understand the process of solving a problem in terms of a journey. In this metaphor, problem solving involves travelling from a starting point (the problem situation) to a destination (the solution) along a path (the method that is used to solve the problem). The use of this metaphor is illustrated by expressions such as to get sidetracked or to have something get in one's way. According to this view, the concrete concept (journey) is grounded in sensory-motor processes, and by lending its structure to the abstract concept (problem solving) also provides grounding to the abstract concept. Some theorists have argued that the mental representation of the concrete concept is necessary to fully understand the abstract concept ([13–15]; but see [16]). An important claim is that the metaphorical relation is conceptual rather than linguistic, and thus does not depend on metaphorical language. Most of the concrete concepts used for such metaphorical mappings are basic image schemas, which are schematic mental representations of simple spatial relations that can be applied in many situations, such as source–path–goal, containment, up–down, or balance [12,15]. These image schemas could be grounded in embodied experiences.

Experimental studies have shown that image schemas might indeed be activated by abstract concepts even in the absence of metaphorical language. The up–down image schema, for example, has been linked to abstract concepts such as power [17,18] and valence [19,20]. To avoid explicit activation of the metaphorically related concrete concept, experiments have been set up such that no linguistic references are made to the metaphor [18,21–24]. Results indicated that the abstract concept (e.g. valence) directed attention towards vertical locations that were consistent with the image schema.

In these studies the image schema was not task-relevant, suggesting that the image schema was activated automatically, consistent with the view that the image schema is an essential part of the concept. Similar studies (e.g. [25]) with concrete concepts, however, suggest that, even for concepts that have a strong association to up or down spatial positions (e.g. moon, roof, basement, fall), spatial information is not activated automatically. Thornton et al. [26] showed that no spatial congruency occurred if participants did not have to make a binary spatial response selection, suggesting that the effect of spatial congruency is due to the task requirement to select a response.

Other types of sensory-motor processes also seem optional rather than central to concepts. Consistent with the idea that motor action is central to cognition [27], Tucker & Ellis [28] showed that participants made faster hand responses to pictures of objects (e.g. a cup with handle) if the graspable part was on the same side as the response hand than if it was on the other side. The position of the handle was irrelevant for the task, which suggests that participants automatically activate a grasping response towards the object that facilitates the actual response. However, the alignment effect is obtained also when responses are made with the feet instead of the hands [29], or with two fingers of the same hand [30]. Roest et al. [31] found that the alignment effect disappeared in a go/no-go task in which participants made a response with only one hand, eliminating the need to select among response alternatives. Such results are hard to reconcile with the explanation that the object picture activates actions that are afforded by the object. The alignment effect is present whenever the response set contains a left–right dimension even when the response is very dissimilar from a grasping action, as in the case of foot presses, and is not present when the response is an actual grasping action without left–right dimension in the responses set. This suggests that the effect depends on similarities between the object picture and response at a more abstract level, such as left–right dimension, rather than a concrete motor action [32]. Likewise, image schemas are also abstract representations of space [11,33]. An image schema such as up–down may refer to experiences in different modalities such as seeing the top and bottom of a tree or the feeling of reaching towards a high shelf. Thus, the image schema might be a multimodal or supramodal representation of space and it must be devoid of specific sensory-motor experiences in order to map it onto an abstract concept such as power [34].

Finally, the idea that image schemas are central to abstract concepts presents some logical problems. To map an image schema onto an abstract concept, the structures of both concepts need to be aligned, which requires that there is already a representation for each concept [16]. Moreover, an explanation is needed for how people distinguish between the image schema and the abstract concept, and related, how people distinguish between different concepts when the same image schema is mapped on these concepts. Such an explanation will likely involve unique features for each abstract concept and the question is whether these features also need to be grounded in sensory-motor processing.

3. Situated cognition

A related proposal is that abstract concepts are grounded by situating them in concrete events [35–37]. According to Barsalou & Wiemer-Hastings [38], situations are important for both concrete and abstract concepts. Concept understanding requires knowing in what kind of situations the concept occurs. Although concrete concepts, such as bike, can be known by their physical properties, it also seems essential to know when, where, and by whom bikes are used. Abstract concepts, such as truth, also need to be situated in events in order to be understood. Participants provided substantial proportions of situational properties for concrete, intermediate and abstract concepts [38,39]. Event information, such as thematic roles and temporal sequence, is activated for concepts [40–42] and the ease with which words evoke a context is related to naming latencies [43]. Providing a context reduces the difference in reading times for abstract and concrete sentences [44]. Moreover, Wilson-Mendenhall et al. [36] found that the patterns of brain activity when participants processed abstract concepts reflected the type of context (physical danger or social evaluation) in which the concepts had been situated. Even superficial situational details that may seem unrelated to the central concept can be important. In problem solving and other effortful cognitive tasks that require processing abstract principles, people often rely on irrelevant surface features ([45–51]; but see [52]).

A role of situations is consistent with exemplar models of categorization and memory [35,53,54]. As exemplar models show, many findings on concept learning and representation can be explained by assuming that individual experiences are stored in memory and that abstraction across experiences occurs only during retrieval. On this account, an abstract concept is represented in many memory traces for specific events in which that concept occurred. When memory is probed by the abstract concept, many traces are activated. A concept is the weighted summary of the activated traces, which forms a computed, temporary representation based on situated experiences [40,42,55]. Thus, summary representations need not be stored in memory directly but emerge from the activation of several traces during retrieval. Thus, sensory-motor features from previous experiences contribute to the summary representation. More recent memory models [56,57] include a mechanism where traces store features from many different experiences, thus forming semantic rather than episodic traces. Rather than being abstract, however, such traces contain many features from different contexts. Thus, although these models do not focus on sensory-motor grounding, they show the importance of contextual features and thereby support the notion that concepts are represented by situations. Because many elements of situational experiences have sensory-motor features, this account could provide grounding for abstract concepts.

Wiemer-Hastings & Xu [39] noted, however, that the properties that participants named for abstract concepts were themselves also abstract. Whereas for concrete concepts participants listed parts, actions, locations, and so on, for abstract concepts they listed person, mental state, relation to a state of affairs, and so on. Moreover, the idea that links to concrete concepts will ground abstract concepts in sensory-motor processing hinges on the assumption that concrete concepts are represented by sensory-motor simulations. Several empirical findings, however, suggest that this may not be as clear-cut as it seems. Much of the evidence for a role of sensory-motor processing might also be explained by spreading activation from cognitive processing to sensory-motor processing [58,59]. Stronger evidence for a causal role of sensory-motor processing can be obtained by investigating the effect of reduced availability of sensory-motor processes on cognition [60,61]. Indeed, some studies have shown that concurrent motor interference reduces performance in object naming [62] and picture memory [63,64] more for manipulable objects (hammer) than for nonmanipulable objects (elephant). Other studies, however, systematically failed to find evidence that performance for manipulable objects was disrupted more by concurrent motor tasks than performance for nonmanipulable objects ([61,65–67], see also [68]). Similarly, reduced availability of the cortical motor system seems to have little effect on object representation [6,69].

To summarize, although there is little doubt that abstract concepts derive their meaning from being situated, the question is whether this provides sensory-motor grounding. Given that sensory-motor processes are involved only under specific circumstances even for concrete concepts, it seems unlikely that in the context of understanding abstract concepts people will routinely use sensory-motor simulations to represent the concrete properties of situations.

4. Hybrid models of concepts

Thus, sensory-motor grounding is unlikely to fully explain abstract concepts. Even for concrete concepts, sensory-motor processes seem optional and context-dependent. The sensory-motor experiences associated with an abstract concept such as truth are likely much more diverse than those for a concrete concept such as bike, and may thus be unlikely to represent an abstract concept in a consistent way. Several authors have suggested that language might be crucial for abstraction. Gentner [70–72] argued that abstraction requires the use of analogy and that analogy at a structural level is promoted by using language. Abstract concepts often involve relations between entities, such as causation [73]. To recognize that a particular relation holds, one needs to compare the current situation to previous situations with the same relation. In general, however, people find it easier to notice similarities between two situations at superficial levels, such as objects they have in common, than at structural levels, such as relations between entities [70,74]. As a result, people are more likely to compare situations that are superficially similar than situations that are superficially dissimilar. This is shown in studies on problem solving where solutions can be found by comparing a new problem to previous examples at a structural level [50,75]. In these studies, participants perform better when structurally similar problems are also superficially similar than when they are superficially dissimilar. Another example is that children learn nouns earlier than verbs, probably because nouns refer to concrete objects whereas verbs often involve relations between entities [76]. Superficially similar elements may act as reminders, resulting in retrieval of superficially similar examples from memory, sometimes at the cost of structurally similar examples. When structurally similar situations are given the same linguistic label such label may also act as a cue to retrieve relevant situations from memory, allowing people to extract structural similarity between current and previous situations. This way, language promotes comparison between two superficially different situations. Thus, it is the combination of language and use of analogy that allows people to form abstract concepts.

Hybrid models propose that conceptual knowledge can be represented by a combination of sensory-motor processes and relations between words [8,77–80]. Andrews et al. [7] developed a hybrid model in which semantic representations are formed by combining experiential (i.e. sensory-motor) features listed by participants and linguistic co-occurrence data from a large corpus of text. They showed that semantic representations are richer when the two types of data are integrated in a model than when they are used independently. They argued that integration resulted in more coherent and refined representations than independent input. Moreover, the integrated model allows abstract concepts to be grounded indirectly through links with concrete concepts. This model may capture the entire developmental trajectory in concept learning. It has been argued that young children's word learning depends to a large extent on their embodied interactions with the referents (e.g. [81]) but that after some years of formal schooling learning is based less and less on direct embodied experience with the material and more on linguistic information such as textbooks [82].

A slightly different theory, LASS [8], proposes that linguistic processing and sensory-motor simulations both contribute to task performance, but the extent to which each is involved is task dependent. Linguistic processing is more important for superficial tasks, in which for example word associations are sufficient. In deeper tasks, sensory-motor simulations become more important. In LASS, meaning is mainly provided by simulations. In that respect the LASS theory differs from the proposal by Andrews et al. [7], who do assume that meaning captured by word distributions is complementary to experiential data.

These hybrid models might still suggest that sensory-motor simulations are needed to ground meaning during concept representation. A different view, however, is that sensory-motor experiences shape conceptual memory during acquisition, but that full re-enactment of such experiences is not necessary to represent well-learned concepts. A conceptual memory with central, non-sensory representations might still produce cognitive behaviour that reflects sensory-motor experience. For example, Plaut [83] showed that a connectionist model of single concepts that was trained on modality-specific features and concept names developed representations that were organized by modality but also by concept. In his model, representations (hidden units) showed similarity structures that reflected modality-specific experiences, but were not themselves sensory. The model consisted of four different sensory-motor units: vision and touch input units, and phonological and action input/output units. Each of these modality-specific units was connected to a central set of hidden units. The hidden units represented semantic information that combined information from all input/output units. The hidden units had a topographical bias that was implemented as more effective learning for short connections than for long connections. As a result, although all hidden units were involved to some degree for each modality, information in the hidden units was organized according to the modality during learning such that representations in the same modality show greater similarity than representations in different modalities. At the same time, however, the network was able to represent object identity regardless of the modality in which it was presented; that is, there were high correlations between the representations of the same object in different modalities. The resulting network thus had a ‘graded degree of modality specificity’ [83]. Lesions to the network showed that different locations of the hidden units contributed differently in different tasks. For example, lesions to the hidden units impaired action naming more as they were closer to the action units, whereas they impaired visual object naming more as they were closer to the vision units.

Plaut's model shows that a unitary set of interconnected hidden units could represent modality-specific experiential information and linguistic information at the same time. His model may therefore provide key insights into how it may be possible that cognitive behaviour sometimes seems to be the result of sensory-motor processing but the same behaviour is not impaired when sensory-motor processes are less available. One example is the mixed set of findings on the role of the motor system for object memory described earlier. In the studies that found an effect, motor actions were presented as photographs or short videos of the appropriate grasp for an object, and grasp similarity in a memory set affected performance. In the studies that did not obtain an effect, the motor system was made less available by a concurrent motor interference task. These mixed results suggest that the role of motor actions may not have been at the level of the motor system itself, which explains the absence of motor interference effect, but at some other, more abstract level such as the modality-specific organization in Plaut's hidden units.

Plaut's [83] model was intended to explain aphasia. To use it as a model of conceptual representation, adjustments are needed. First, as Plaut also noted, the modality-specific representations in his model are simplified and are not based on actual concepts, although he approximated the similarity structures between different modalities and between concepts in the same category. There was no other semantic information represented in the model such as associative or encyclopedic knowledge. Second, both situational knowledge and linguistic relations are not represented in the model. To be comparable to other hybrid models such as the one developed by Andrews et al. [7], Plaut's model [83] would need to be expanded.

Plaut's model shows, however, the importance of language for concepts as a means to bind representations from different modalities into a single concept. As Gentner [70–72] has argued, language may be critical for comparison, especially when situations do not share many superficial features. For abstract concepts, language may thus draw attention to similarities across different situations that involve the same abstract relation such as truth or disappointment. Although situations are concrete events, the similarities between situations that involve the same abstract concepts will mostly be structural and abstract rather than sensory-motor. As has been shown, however, people tend to pay attention to concrete details, so it seems likely that these may also be incorporated into the concept. Language may also activate image schemas for abstract concepts because language may refer to the image schema. For example, the expression to be in a low mood may invite people to align vertical orientation and valence.

A model of concepts that is based on input from modalities and from language may be very flexible and powerful [84]. Integration results in richer structures [7] and central structures may still show modality-specificity [83]. Semantic representations can thus be non-modal in the sense that they are not produced by sensory-motor simulations but still reflect the sensory-motor experience during acquisition. In general, representational processes that can use such rich information can therefore result in very different representations depending on the task and other contextual features. As a result, behaviour may show different levels of sensory-motor grounding. At one extreme, sensory-motor simulations may be necessary when novel situations need to be represented, for example when participants are asked to produce features for rolled-up lawn [85] or to verify whether a newspaper can be used to shield one's face from the wind [86]. In many situations, however, sensory-motor simulations may not be needed, but behaviour is still influenced by experiential features because concepts are organized by sensory-motor experiences. At the other extreme, behaviour may be based mostly on linguistic relations. For abstract concepts, this may be more important than for concrete concepts.

5. Conclusion

Abstract concepts may be grounded in image schemas and situations. This does not guarantee that they are grounded in sensory-motor simulations, however, because several studies indicate that concrete concepts may not always be grounded in sensory-motor simulations either. During development, sensory-motor experiences may shape the organization of concepts to a large extent, but increasing reliance on language after initial acquisition may focus the development of concepts more and more on structural similarities [6,76]. As a result, conceptual processing in many situations may not depend on reactivation of the perceptual states that gave rise to the development of the conceptual system. Thus, although concepts will often show traces of sensory-motor processing, once they are well established, sensory-motor simulations will not routinely be necessary for meaningful processing.

Data accessibility

This article has no additional data.

Competing interests

We declare we have no competing interests.

Funding

We received no funding for this study.