Lawrence W. Barsalou and Katja Wiemer-Hastings
Roughly speaking, an abstract concept refers to entities that are neither purely physical nor spatially constrained. Such concepts pose a classic problem for theories that ground knowledge in modality-specific systems (e.g., Barsalou, 1999, 2003a,b). How could these systems represent a concept like TRUTH?1 Abstract concepts also pose a significant problem for traditional theories that represent knowledge with amodal symbols. Surprisingly, few researchers have attempted to specify the content of abstract concepts using feature lists, semantic networks, or frames. It is not enough to say that an amodal node or a pattern of amodal units represents an abstract concept. It is first necessary to specify the concept's content, and then to show that a particular type of representation can express it. Regardless of how one might go about representing TRUTH, its content must be identified. Then the task of identifying how this content is represented can begin.
The primary purpose of this chapter is to explore the content of three abstract concepts: TRUTH, FREEDOM, and INVENTION. In an exploratory study, their content will be compared to the content of three concrete concepts - BIRD, CAR, and SOFA - and also to three intermediate concepts that seem somewhat concrete but more abstract than typical concrete concepts -COOKING, FARMING, and CARPETING. We will first ask participants to produce properties typically true of these concepts. We will then analyze these properties using two coding schemes. Of particular interest will be the content of abstract concepts, and how it compares to the content of concrete and intermediate concepts.
We will not attempt to provide evidence that modality-specific systems represent abstract concepts. Once we have assessed their content, however,
1 Italics will be used to indicate concepts, and quotes will be used to indicate linguistic forms (words, sentences). Within concepts, uppercase words will represent categories, whereas lowercase words will represent properties of categories.
we will speculate on how modality-specific systems could represent it. Notably, though, recent studies have obtained evidence for modality-specific representations in abstract concepts. Glenberg and Kaschak (2002) found that abstract concepts contain motor information, as did Richardson, Spivey, Barsalou, and McRae (2003). For evidence that other types of concepts are grounded in modality specific systems, see recent reviews by Barsalou (2003b), Barsalou, Niedenthal, Barbey, and Ruppert (2003), and Martin (2001).
related issues in representing abstract concepts
The issue of modal versus amodal representation is not the only important issue related to abstract concepts. Previous researchers have raised other important issues. In particular, previous researchers have suggested that situations, word associations, and metaphors are potentially important aspects of how abstract concepts are represented. We address each in turn.
In a series of studies, Schwanenflugel, Shoben, and their colleagues demonstrated that it is often difficult to think of a situation in which an abstract concept occurs (for a review, see Schwanenflugel, 1991). For example, what is a situation in which TRUTH occurs? Although a court trial might eventually come to mind, or a child confessing to a parent, it often takes a while to retrieve such situations. In contrast, situations seem to come to mind much more readily for concrete concepts. For CHAIR, situations like living rooms and classrooms come to mind quickly.
Schwanenflugel, Shoben, and their colleagues explored the role of situation availability across a variety of cognitive tasks. To do this, they first demonstrated the general advantage of concrete over abstract concepts (also see Paivio, 1986). Specifically, they showed that (1) lexical access is faster for concrete words than for abstract words (e.g., Schwanenflugel, Harnishfeger, & Stowe, 1988), (2) word comprehension is faster for concrete words than for abstract words (e.g., Schwanenflugel & Shoben, 1983; Schwanenflugel & Stowe, 1989), and (3) memory is better for concrete words than for abstract words (e.g., Wattenmaker & Shoben, 1987). Then, in the same studies, these researchers demonstrated that situation availability played a major role in these differences by manipulating the presence
2 In the original work on this topic, Schwanenflugel and her colleagues referred to what we're calling "situation availability" as "context availability." We use "situation" instead, first, because the construct of a situation has played a central role in much of our recent work and in the work of other researchers (see Yeh and Barsalou, 2004, for a review), and second, because situations can be viewed as one possible form of context.
versus absence of a relevant situation. For example, participants might first read about a court trial before studying "truth" for a memory test, or first read about a living room before studying "chair." When relevant situations were present, abstract words were processed as well as concrete words. Participants accessed and understood both types of words equally quickly, and remembered them just as well.
These findings demonstrate two points about the processing of words. First, the meanings of words are not established in isolation. A word's meaning is typically not a stand-alone package of features that describes its associated category. Instead, words are typically understood and represented against background situations (cf. Murphy & Medin, 1985). When a situation is not available, a concept is difficult to process. Much early work on language comprehension reached this same conclusion (for reviews, see Bransford & Johnson, 1973; Bransford & McCarrell, 1974). Much recent work echoes this theme (e.g., Barsalou, 2003b; A. Clark, 1997; H. Clark, 1992; Yeh & Barsalou, 2004). In general, situations provide much useful information for understanding concepts. Understanding what CHAIR means relies not only on the physical properties of the object, but also on the settings in which it is found (e.g., classrooms) and the activities performed in them (e.g., attending lectures). Knowledge of chairs is inadequate if one does not know how they are used in relevant situations. For this reason, situations often appear central to the representation of concepts.
Second, retrieving situations for abstract concepts appears more difficult than retrieving situations for concrete concepts. At least the following two factors may be responsible. First, abstract concepts may be associated with a wider variety of situations than concrete concepts (Galbraith & Underwood, 1973). As a result of greater interference between competing situations, retrieving a single one may be more difficult than for a concrete concept. Second, when people process abstract concepts in the real world, there may typically be a relevant situation already in place. People may not typically entertain a concept like TRUTH unless a relevant situation is in place to which the concept applies. As a result, the conceptual system becomes oriented toward retrieving information about abstract concepts with relevant situations already in place. Conversely, because it is relatively unusual to process abstract concepts in situational vacuums, people draw blanks initially when receiving them out of context.
People may not exactly draw a blank when processing abstract concepts in isolation. Instead, the word for an abstract concept may trigger highly associated words. Because no situation comes to mind immediately, other associated information becomes active. Because the memory cue is a word, other associated words come to mind, following the principle of encoding specificity (Tulving & Thomson, 1973). Note, however, that these words don't necessarily constitute conceptual meaning. Instead, as word associates become active, only their surface-level phonological forms are retrieved, accompanied by minimal meaning. Certainly, more semantic content could be retrieved on occasion. Nevertheless, meaning may often be processed minimally.
Several sources of evidence support the conjecture that abstract words encountered out of context initially activate superficial word associations. First, Glaser (1992) reviews much evidence that words in general often generate word associates, prior to the activation of conceptual information. Consider an example. Solomon and Barsalou (2004) found that participants used the strength of word associations in a property verification task when associative strength was a diagnostic cue for responding - they did not typically access conceptual knowledge. Participants adopted this strategy because the object and property words on true trials tended to be associated (e.g., "horse" - "mane"), whereas the object and property words on false trials tended to be unassociated (e.g., "chair" - "feathers"). Because word association strength predicted the correct response, participants did not need to access conceptual knowledge. Kan, Barsalou, Solomon, Minor, and Thompson-Schill (2003) reported fMRI evidence consistent with Solomon and Barsalou's behavioral findings. Glaser (1992) offers many such examples of surface-level word forms playing central roles in "conceptual" tasks. Together, these findings indicate that word associations can become activated quickly with minimal retrieval of conceptual information.
A second source of evidence further suggests that processing abstract concepts in isolation may initially produce word associations. As Pulvermuller (1999) reviews, lesion and neuroimaging studies have localized the processing of abstract concepts in left frontal areas. In these studies, participants usually receive isolated words for abstract concepts not linked to particular situations. Thus, retrieving situations should be difficult, and word associations could fill the conceptual void. Consistent with this account, left-frontal areas tend to be implicated with word generation processes (e.g., Thompson-Schill, D'Esposito, Aguirre, & Farah; 1997; Peterson, Fox, Posner, Mintus, & Raichle, 1989). The proximity of these areas to Broca's area further implicates word generation. To our knowledge, no neuroscience research has assessed the processing of abstract concepts in situations. It would be interesting to see if situational processing shifted brain activation outside word generation areas.
A third line of work also implicates word association in the processing of isolated abstract concepts. Krauth-Gruber, Ric, Niedenthal, and Barsalou (2004) had participants generate information about isolated abstract and concrete concepts under one of three instructional sets. One group produced word associations for the abstract and concrete concepts. A second group constructed an image of what each concept referred to and then described the image. A third group produced properties that are typically true of each concept. Of interest was whether property generation participants more resembled word association participants or imagery participants in the information produced. For concrete concepts, property generation participants produced essentially the same information as imagery participants, consistent with the conclusion that property generation participants used images to represent these concepts (also see Wu & Barsalou, 2004). For abstract concepts, however, property generation participants essentially produced the same information as word association participants, consistent with the conclusion that property generation participants initially accessed word associations for the isolated abstract words. When a background situation was not present, the initial information retrieved for the abstract concepts appeared to be word associations.
Finally, it is worth noting that the retrieval of word associations during initial lexical access is consistent with theories like LSA (Landauer & Dumais, 1997) and HAL (e.g., Burgess & Lund, 1997). According to these theories, a word becomes associated to other words it cooccurs with in texts, with associative strength reflecting frequency and proximity of cooccurrence. Subsequently, when a word is encoded, its network of word associates becomes active. According to LSA and HAL, these associates constitute the word's meaning. Alternatively, these associates may simply be word forms that point to underlying concepts.
Some theorists have argued that the meanings of abstract concepts are grounded in concrete domains (e.g., Gibbs, this volume; Lakoff & Johnson, 1980; Lakoff & Turner, 1989). For example, the abstract concept, ANGER, is grounded in concrete phenomena, such as boiling water exploding out of a closed pot. We agree that metaphors often augment the meanings of abstract concepts, and make certain aspects of their conceptual content salient (e.g., Boroditsky & Ramscar, 2002).
Nevertheless, direct experience of abstract concepts appears central to their content (Prinz, Chapter 5, this volume). One reason is that people have considerable amounts of direct experience with abstract concepts. Consider ANGER. People have much experience with the external situations that trigger anger, what anger feels like subjectively, and how one acts and looks when angry. Indeed, norms for emotion concepts like ANGER contain detailed features of this experience (e.g., Fehr & Russell, 1984). Notably, these norms don't contain much mention of metaphors.
Direct experience of abstract concepts is important for another reason. A concrete metaphor can not be mapped into an abstract concept, if the abstract concept doesn't have it's own structure (e.g., Murphy, 1997). If an abstract concept has no structure based on direct experience, the concrete metaphor would have nothing to map into. Certainly, metaphors may interpret direct experience and add new material to it. The point is, however, that metaphors complement direct experience of abstract concepts, which often appears extensive.
Thus, our focus will be on the direct knowledge that people have of abstract concepts. Later, when we report an exploratory study, we will focus exclusively on direct knowledge. Indeed, we found little mention of metaphors when people described the content of abstract concepts.
hypotheses about conceptual content
A common assumption is that abstract and concrete concepts have little, if anything, in common. With respect to their content, they constitute two completely different kinds of concepts. In contrast, we propose that concrete and abstract concepts share important similarities. In particular, we propose that they share common situational content (Hypothesis 1). Where concrete and abstract concepts differ is in their focus within background situations, with concrete concepts focusing on objects, and abstract concepts on events and introspections (Hypothesis 2). As a result of these different foci, the representation of abstract concepts is more complex, being less localized in situational content than the content of concrete concepts (Hypothesis 3). Finally, because the content of abstract concepts is grounded in situations, this content can be simulated in modality-specific representations (Hypothesis 4). We address each hypothesis in turn.
Hypothesis 1. Concrete and Abstract Concepts Share Situational Content
As reviewed earlier, situations appear important for accessing and representing both abstract and concrete concepts. Consider the concrete concept, HAMMER. If people only know the physical parts of HAMMERS (e.g., head, handle), they lack an adequate concept of the category. Instead, people must also know the settings where the objects are used, such as the presence of two boards and nails, along with an agent for using them. People also need to know the actions that the agent performs (e.g., swinging the hammer), and also the events that result from these actions (e.g., the hammer head driving the nail into the boards). Finally, people need to know about the mental states of the agent, including goals (e.g., bind two boards together) and affective reactions (e.g., satisfaction when the nail is pounded in correctly). Only when all of the relevant situational content about HAMMERS is known does someone approach a full understanding of the concept. For a detailed account of how situational knowledge underlies the semantics of artifacts and natural kinds, see Barsalou, Sloman, and Chaigneau (in press).
Situations appear even more central to abstract concepts. Consider Barsalou's (1999) semantic analysis of an everyday sense of TRUE, namely, the sense that an agent's claim about the world is accurate. To represent this sense requires a situation that contains the following event sequence. First, a speaker makes a verbal claim about some state of the world to a listener (e.g., it's raining outside). Second, the listener constructs a mental representation of the speaker's claim (e.g., what raining outside might look like). Third, the listener perceives the part of the world that the claim is about (e.g., the weather outside). Fourth, the listener determines whether the represented claim matches or doesn't match the current state of the world. Fifth, if the claim matches the world, the listener concludes that the speaker's claim has the property of being TRUE; otherwise, it's FALSE. As this example illustrates, a complex situation is necessary to represent the meaning of TRUE, including multiple agents (e.g., speaker, listener), physical events (e.g., communication, states of the world), and mental events (e.g., representing, comparing). Without a situation, it would be impossible to represent the meaning of this abstract concept. Barsalou (1999) provides additional examples of how situations underlie other abstract concepts, such as OR, and also ad hoc categories, such as THINGS TO STAND ON TO CHANGE A LIGHT BULB.
As these examples illustrate, situations appear central to both concrete and abstract concepts. Thus, there should be strong similarities between their content. When we ask participants to describe the content of concepts, we should observe extensive mention of situations for not only for abstract concepts, but also for concrete ones.
Hypothesis 2. Concrete and Abstract Concepts Differ in Situational Focus
Where we propose that concrete and abstract concepts differ is in their focus on situational content. For concrete concepts, attention should focus on the respective objects against their background situations. In representing HAMMERS, the focus should be on hammer objects in their situations of use. Even though information about events and introspections exists in the representation, attention focuses on the critical objects and their specific properties.
In contrast, the focus for abstract concepts should be distributed across other types of situational content. Specifically, abstract concepts should focus on event and introspective properties. One basis for this prediction is Barsalou's (1999) preliminary analyses of abstract concepts, where these two types of properties played central roles. Complex configurations of event and introspective properties generally appeared necessary to capturing the meaning of an abstract concept. Further bases for this prediction are empirical findings from Wiemer-Hastings and her colleagues. In one line of work, the similarity between the linguistic contexts in which abstract concepts occur predicted the similarity of the concepts' meanings (Wiemer-Hastings & Graesser, 2000; see also Wiemer-Hastings & Graesser, 1998). Because these contexts were defined as verbs and prepositions, this correlation supports the proposal that events and settings are central for abstract concepts. In another study, introspective properties were especially important for predicting the differential abstractness of abstract concepts (Wiemer-Hastings, Krug, & Xu, 2001).
Hypothesis 3. Abstract Concepts are More Complex Than Concrete Concepts
According to Hypothesis 2, concrete concepts focus on objects in situations, whereas abstract concepts focus on events and introspections. An implication of this proposal is that the representations of abstract concepts should be more complex than those of concrete concepts. For concrete concepts, the focus is on a relatively local, spatially circumscribed region of a situation. In a HAMMER situation, for example, the focus is on the region that the hammer occupies.
For abstract concepts, however, the focus is on multiple components that are not localized but distributed widely. In a TRUE situation, for example, the focus includes the speaker's claim, the listener's representation of the claim, and the listener's assessment of the claim. All these components, and the relations between them, must be represented and integrated to evaluate TRUE's focal content. We have increasingly come to believe that abstract concepts seem "abstract" because their content is distributed across situations. Another contributing factor may be the centrality of introspective information, which may be more subtle to detect than entity information. Thus, in the study to follow, we expected to see more complex representations for abstract concepts than for concrete ones.
Hypothesis 4. The Content of Abstract Concepts Could, in Principle, Be Simulated
All of the conceptual content that we have discussed so far could, in principle, be simulated in the brain's modality specific systems. It is well known from both imagery and neuroscience research that people construct images of objects, settings, and events (see Barsalou, 2003b; Martin, 2001). Thus, all of this situational content could in principle be simulated as people represent concrete and abstract concepts.
In contrast, little direct evidence bears on the representation of the introspective information central to abstract concepts (also important but backgrounded for concrete concepts). Nevertheless, it seems quite plausible that introspective content could be simulated in mental images (for specific proposals, see Barsalou, 1999,2003a). It seems possible to simulate the introspective experiences of emotions (e.g., happiness), drive states (e.g., hunger), and cognitive operations (e.g., comparing two imagined objects). There is no empirical or theoretical reason for believing that introspective content could not be simulated as part of a conceptual representation. Thus, we will assume that the presence of introspective content in conceptual representations does not constitute evidence against embodied theories of knowledge.
an exploratory study
To assess these hypotheses, we asked college students to produce the properties of three abstract concepts, three concrete concepts, and three concepts intermediate in abstractness. Because we wanted to explore the content of these concepts in an open-ended manner, we did not constrain participants' protocols to any particular type of information. Thus participants were simply asked to describe the properties that they thought were characteristic of each concept. As will be seen, these probes led to a diverse collection of responses.
Two other factors besides concept abstractness were manipulated: situation availability (whether a concept was or was not preceded by the description of a situation) and concept form (whether a concept was presented as an entity, event, quality, etc.). These factors had little effect, so we do not report results for them. Because participants responded at their own pace for a full minute, they produced diverse content that obscured possible effects of these manipulations.
Analyses. To examine the content of participants' protocols, two coding schemes were applied. First, Wu and Barsalou's (2004) coding scheme established the amounts of taxonomic, entity, setting/event, and introspective content in the protocols (for additional applications of this coding scheme, see Cree & McRae, 2003; Krauth-Gruber et al., 2004; McRae & Cree, 2002). This coding scheme makes it possible to assess whether situational content occurs across all three concept types (Hypothesis 1). It also enables assessing whether different types of situational content are more important for concrete vs. abstract concepts (Hypothesis 2).
A second analysis established larger groups of properties in the protocols (e.g., extended descriptions of people, events, introspections, etc.). To capture these larger protocol units, a second coding scheme was developed. This scheme also captured the hierarchical relations that frequently integrated these larger units. As will be seen, analysis of these units and their hierarchical structure provides insight into the shared and distinctive properties of abstract, concrete, and intermediate concepts. Applying this scheme also allowed us to assess whether the representations of abstract concepts are more complex than those of concrete concepts (Hypothesis 3).
Finally, both analyses allow informed speculation about Hypothesis 4. Once these two analyses establish the content of abstract, concrete, and intermediate concepts, we can begin thinking about whether modality-specific systems could in principle simulate it.
Materials. The critical materials included three abstract concepts, three concrete concepts, and three intermediate concepts. Each abstract concept took three forms: TRUE, THE TRUTH, TRUTHFULNESS; A FREEDOM, TOFREE, FREELY; AN INVENTION, TO INVENT, INVENTIVENESS. Similarly, each intermediate concept took three forms: A COOK, TO COOK, SOMETHING THAT HAS BEEN COOKED; A FARM, TO FARM, SOMETHING THAT HAS BEEN FARMED; A CARPET, TO CARPET, SOMETHING THAT HAS BEEN CARPETED. A single form was used for the three concrete concepts (A BIRD, A SOFA, A CAR), given that these concepts did not have similar variants. As described earlier, the form of concepts had no effect. Thus, we only report results collapsed across forms (e.g., results are combined for A FREEDOM, TO FREE, and FREELY).
Six lists of the nine critical items were constructed. Each contained one variant of each abstract concept, one variant of each intermediate concept, and all three of the fixed concrete concepts. One set of three lists was counter-balanced such that each variant of a concept occurred equally often, and also such that the variants of different concepts differed maximally within a given list. Each third of the list contained one abstract concept, one concrete concept, and one intermediate concept, in random orders. The entire counterbalancing process was performed twice, to produce a second set of three lists, so that a given concept's list position varied, along with its neighbors, in the respective third of the list.
A short paragraph was constructed for each concept that described a relevant situation. For variants of TRUTH, the paragraph described a boy telling his mom that he wasn't responsible for breaking a living room vase, and his mom believing him. Similarly, for CAR, a short paragraph described a woman using her car for commuting to work, and listening to music while doing so. As described earlier, the open-ended nature of the protocols obscured any effects of this manipulation, such that it receives no further discussion here.
Participants and Design. Initially, the study included 24 Emory undergraduates who participated for pay. Half received the words in isolation, and half received the words preceded by situations. Within each group, two participants received each of the six versions of the critical list.
Over the course of data analysis, three participants' data were lost due to computer problems (one situations subject and two no-situations subjects). For the remaining 21 participants, 4 of the 189 protocols (21 participants x 9 concepts) were also lost due to computer problems. No more than one protocol was ever lost per concept or participant, and these were distributed evenly across abstract, concrete, and intermediate categories. Because of this study's exploratory nature, these missing participants and protocols were not replaced. Because concept abstractness was manipulated within participants, every remaining participant's data could be assessed on this variable.
Procedure. Participants received the following open-ended instructions to produce the properties of concepts:
The general purpose of this experiment is to study people's knowledge about the world. Our specific purpose here today is to learn more about people's understanding of a few specific concepts. Before we go any further, let me stress that there are no correct responses to the questions that I am about to ask you. Thus, please don't worry about whether you've come up with the right answer or not. This is not an issue at all. Instead, what we're doing here is performing a scientific experiment, where the purpose of this experiment is to understand how normal people like yourself think about various concepts. Here's what you can do to help us learn more about this. In a moment, when I ask you about various concepts, please respond with the very first thoughts that come to mind, and then keep responding with the thoughts that continue to come to mind, until I ask you to stop.
Later in the instructions, participants practiced producing properties for TREE, BRICK, PENCIL, and CAMERA verbally. The instructions did not give any examples of possible properties that could be produced, so as to avoid biasing participants' responses. On every trial, participants received the following instruction:
Please report your thoughts as they come to mind.
What characteristics are typically true of the following concept: [concept name]
When ready, participants began producing characteristics verbally for a full 1 min. Whenever a participant paused for 5 sec, the experimenter stated, "Please continue to describe your thoughts as they come to mind." When participants were in the middle of describing a property at the 1 min point, they were allowed to complete the description. A digital video camera captured each protocol to a computer, with participants consenting to the recording process.
Prior to the nine critical trials, participants produced properties for six randomly ordered practice concepts, which they did not realize constituted practice. As for the critical materials, two concepts were abstract (INTEGRITY, VAGUE), two were concrete (A ROSE, A BRUISE), and two were intermediate (TO HAMMER, SOMETHING THAT HAS BEEN PAINTED).
Analysis 1: Establishing the Content of Protocol Elements
Each of the nine critical protocols for a given participant was transcribed and coded using the Noldus software for digital video coding. As a judge viewed a protocol, she transcribed each statement into a file and coded it with one of the 45 coding categories from the Wu and Barsalou (2004) coding scheme.3 These 45 coding categories fell into 5 general groups: taxonomic, entity, setting/event,4 introspective, and miscellaneous. A tax-onomic code was applied to statements that mentioned a taxonomic category related to the target concept. An entity code was applied to statements that described a property of a physical object. A setting/event code was applied to statements that described a property of a setting or event. An introspective code was applied to statements that described the mental state of someone in a situation.
The specific taxonomic coding categories were synonym, ontolog-ical category, superordinate, coordinate, subordinate, and individual. The specific entity coding categories were external component, internal component, external surface property, internal surface property, substance/material, spatial relation, systemic property, larger whole, entity behavior, abstract entity property, and quantity. The specific setting/event coding categories were person, nonperson living thing, setting object, social organization, social artifact, building, location, spatial relation, time, action, event, manner, function, physical state, social state, and quantity. The specific introspective coding categories were affect/emotion, evaluation, representational state, cognitive operation, contingency, negation, and quantity. The specific miscellaneous coding categories were cue repetition, hesitation, repetition, and meta-comment. For definitions of these categories, see Wu and Barsalou (2004), Cree and McRae (2003),
3 An updated version of the Wu and Barsalou (2004) scheme was used that incorporated additional categories found by McRae and Cree (2002) to be important.
4 In Wu and Barsalou (2004), McRae and Cree (2002), and Cree and McRae (2003), "setting/event properties" were referred to as "situational properties." We refer to them as "setting/event properties" here because we use "situation" more generally to include introspective information, not just external information about settings and events in the world.
Was this article helpful?