discussion

night nurse

chapter7kellogg.docx

Home >Psychology homework help >discussion

Chapter 7 Imagery and Knowledge Representation

Learning Objectives

· Explain the difference between a propositional representation of knowledge and an image representation.

· Describe the uses and limitations of visual imagery in cognition.

· Contrast rule-governed concepts with object concepts based on family resemblance.

· Describe the neocortical areas that mediate the conceptual representation of artifacts (e.g., tools), natural kinds (e.g., animals), and abstractions (e.g., numbers).

· Discuss how semantic memory is organized and used in answering factual questions.

In this chapter, the fundamental questions about the representation of knowledge in semantic memory are examined. As discussed in Chapter 1, some mental representations are unavailable to consciousness. We will see in this chapter that our knowledge of concepts about the world—such as knowing that a chicken has wings and can fly and is a member of a larger category called birds—is stored as a dense semantic network of facts in long-term memory. Such knowledge is propositional in that a given assertion can be judged as true or false. It is true that a chicken has wings, but it is false that a chicken can bark like a dog. Although we can use the propositions stored in the semantic network to answer such factual questions, the network itself is a vast collection of unconscious mental representations. As an analogy, we can use a search engine on the internet to find some fact of interest at the moment, but the vast web of connections that comprise the internet is unavailable to our inspection.

By contrast, some mental representations can be conscious and take the form of images that seem similar to perception. For example, close your eyes and imagine you are sitting outside on a park bench under a large, shady tree on a warm, sunny day. Pretend you can feel the light breeze enveloping you in comfort and contentment. The fragrance of nearby roses wafts past. The green grass has never seemed greener, nor the sky bluer, than you see them right now. With this act of imagination, the brain activates perceptual systems to recreate visual images of the bench, the tree, the grass, and the sky. Tactile imagery allows you to feel, in a sense, the light breeze, just as olfactory imagery imitates the perception of a rose’s fragrance. Exactly what is the nature of these conscious images? This question has been hotly debated in the field of cognitive psychology. One side of the debate has contended that images retrieved from long-term memory are, in fact, like perceptions. An opposing point of view has held that images are instead stored in memory as abstract propositions. According to this view, the fundamental representation of all knowledge is propositional and is stored together in a semantic network that is unavailable to conscious inspection. From this propositional knowledge, images might be constructed, but they are not really the same as perceptions. What seems like a perceptual image is only an illusion of how the propositional knowledge is used.

The present chapter first addresses the question of whether concrete images are represented differently than abstract conceptual knowledge is weighed. Next, a detailed discussion of the nature and uses of visual imagery is provided. There is a large literature on imagery that shows the strengths and limitations of these conscious forms of mental representation. Third, the nature of concepts is considered in detail. It is necessary to contrast rule-governed concepts with object concepts of natural kinds and artifacts. The neural correlates of these conceptual representations in the temporal lobe of the brain are also discussed. Finally, the chapter closes with work on how knowledge is retrieved from semantic memory. Cognitive psychologists have modeled how people use semantic memory to answer simple factual questions, much as we use a web browser to retrieve facts from the internet.

7.1 Imagery and Abstract Representation

In Chapter 1, the concept of a mental representation was introduced, but the specific nature of the code used was left unaddressed. Similarly, in Chapter 5, declarative memory was discussed without attempting to specify the codes used for representing a personal experience stored in episodic memory or a concept stored in semantic memory. Exactly how is declarative knowledge coded in the mind? What is its format? One hypothesis in cognitive science is that the concepts of semantic memory are represented in an amodal fashion. The representations are an abstract and symbolic means for storing relevant properties and relationships apart from the modalities of perception and motor action (E. E. Smith, 1978). Abstract propositions provide such a means for storing knowledge, as will be seen shortly. They provide essentially a verbal-like description of semantic knowledge. However, they are not actually words associated with the oral modality of speech or the visual modality of reading. Rather, they are amodal and abstract in character. A good analogy is to think of the code used to store knowledge in a digital computer. The digital computer stores knowledge in an abstract digital code of strings of ones and zeros (bits of information). The computer is a symbol processing system and has the capacity to represent all kinds of modality-specific symbols, such as a written word, or a spoken word, or a visual picture. But, in each of these cases, a computer codes these symbols in terms of ones and zeros. A CD or DVD can portray visual pictures, spoken words, or written words using the same abstract code that has nothing to do with the modalities of the symbols. By analogy, it has been hypothesized that the brain also uses an amodal code for all kinds of mental representations, including those that give rise to conscious visual imagery.

An alternative possibility is that perceptual and motor systems provide modality-specific codes for representing concepts in semantic long-term memory. The perceptual properties of an object (e.g., how it looks in size, color, and shape) and the motor properties relevant to interacting with the object (e.g., how to touch, grip, turn, sit upon, or otherwise make use of it) are modal representations. These codes are not symbolic substitutes; rather, they provide a direct mental analog of the object. Everyday applications that use similar representations are a bit out of date today, but photographic negatives or the grooves inscribed on a vinyl record are both analog representations. In terms of mental representations, these types will be defined below as amodal propositional codes versus modal imaginal codes. As an extension of this hypothesis, the dual coding theory of memory, introduced in Chapter 4, assumes that episodic memories can either be coded as images or as words. Verbal and imaginal codes are thus viewed as two kinds of modality-specific codes. The theory specifically rejects the claim that there is such a thing as an amodal code (Sadoski & Paivio, 2001).

An imaginal code is a concrete means of mental representation that directly conveys perceptual qualities. For example, if an object is perceived through the visual sensory modality, it is then possible to form a mental image of the object that seems like the original perception. A propositional code is an abstract means of mental representation that is not linked to any sensory modality. A propositional code instead represents the features of an object and its relationships to other objects without forming an image. To illustrate, consider the concept of a bird. Shown in Panel (a) of Figure 7.1 is an imaginal representation of a robin, the prototype of the category. The prototype can also be represented as a propositional code, as illustrated in Panel (b). Note that propositional codes are similar to the verbal codes hypothesized in dual-coding theory: Both assume that the name of an object is stored in long-term memory (see Chapter 4). They differ, however, in that verbal codes are associated with the perceptual qualities of speech and writing, whereas propositional codes are purely abstract (Paivio, 1986).

Identifying precisely how visual, auditory, and other kinds of images are mentally represented and neurally coded by the brain, however, has been an ongoing and difficult challenge (Kolers, 1983). Some have argued that images are actually based in propositional representations to which one does not have conscious access (Pylyshyn, 1973, 1981). On the other hand, some evidence clearly suggests that visual images behave like perceptions, not propositions. The key distinction is that visual images possess a spatial or analog quality that cannot, in principle, be explained in terms of an abstract, propositional, or verbal representation.

Figure 7.1 An imaginal and a propositional code for the concept of a robin.

Visual Imagery

Shepard and his colleagues pioneered the study of visual images with an extensive series of experiments on mental rotation (Metzler & Shepard, 1974; Shepard & Cooper, 1983; Shepard & Metzler, 1971). The researchers presented people with pictures of three-dimensional objects such as those shown in Figure 7.2. They presented a pair of objects and asked the viewers to decide whether the objects were identical except for their orientation. For example, compare the two objects in Panel (a). Are they the same or different? By mentally rotating the object on the left, you can verify that it, in fact, matches the object on the right. In Panel (b), the same response is again called for after rotating the figure. However, in Panel (c), the left object can be rotated 360 degrees and it will still not match the right object. Thus, this trial calls for a different response.

Shepard and his colleagues systematically varied the angle of rotation required to determine that the pair of objects was the same. They reasoned that if mental images are like real objects in the mind’s eye, then the time required to make a same/different decision ought to increase as a linear function of the angle of rotation. If, instead, the decision is based on a propositional representation of the objects, then making a determination after the object is rotated 180 degrees should take no longer than when the object is rotated 90 or 45 degrees. In each case, the degrees of rotation would simply be a variable in one of the propositional arguments (e.g., left object, clockwise rotation, picture plane, 180 degrees). The results are illustrated in Figure 7.3 (Shepard & Metzler, 1971). Decision times increased as a linear function of the angle of rotation. Thus, there is a striking resemblance between rotating a real object and rotating a mental representation of that object. The images behave in a manner analogous to the real objects.

The rotation of a mental object is a kind of spatial imagery that activates regions in the parietal lobe (Zachs, 2008). It appears that transformations such as mental rotation depend on a different brain network than does the use of spatial imagery to specify the location of an object. Thompson, Slotnick, Burrage, and Kosslyn (2009) used the same stimuli in one of two kinds of imagery tasks. In the location-based task, participants first studied a spatial array of letters for 30 seconds. Next, in the test phase, they were cued to remember the area of the array where a particular target letter had been located. By contrast, in the transformation-based task, participants in the test phase were required to rotate one of the studied letters a certain number of degrees. The rotation task activated the parietal lobe, replicating Zach’s (2008) finding. By contrast, specifying the location of a letter instead activated (a) the sulcus at the border of the parietal and occipital lobe, (b), a closely related cortical region known as the precuneus, and (c) the posterior cingulate cortex. The important point is that two subjectively similar processes involving visual imagery—transforming an object by rotating it and identifying the location of the object—engage different networks in the brain.

Figure 7.2 Try to mentally rotate the left object of each pair so that it matches the right object.

Figure 7.3 Mental rotation time increases with the degrees of rotation required to judge the objects as the same.

The ability to transform objects using spatial imagery could be particularly useful in specific occupations. One study assessed the mental rotation ability of airline pilots (Dror, Kosslyn, & Wang, 1993). Compared with others of the same age, sex, and degree of education, the airline pilots showed a superior ability to make spatial judgments and mentally rotate objects. Perhaps the occupational training and job performance aided the development of mental rotation. Alternatively, it could be that those with superior aptitude for mental rotation are more likely to pass the selection and training procedures required to become a pilot in the first place.

Analog Properties.

Kosslyn (1980, 1981) and his colleagues explored the analog aspect of mental imagery in detail. For example, Kosslyn (1975) showed that the time to scan a mental image depended on “how far” one must scan. Participants memorized a picture of an object, such as a boat, as shown in Figure 7.4. They were then instructed to visualize the object and to scan the mental image to find a feature from a particular starting point. They pressed a button as soon as they “found it” in the image. The results showed that when scanning an image of the boat from the left, participants took longer to find the anchor than they did to find the porthole. Just as the time to scan the picture would take longer with the greater real distance, the same result applied to the mental scan time as well.

Figure 7.4 A study of scanning mental images.

SOURCE: From Kosslyn, S. M. (1973). Scanning visual images: Some structural implications. Perception and Psychophysics, 14(1), 90–94. Reprinted with permission of the Psychonomic Society.

Similarly, small mental images are harder to scan to find features than are large images. Kosslyn (1975) instructed people to imagine a target animal, such as a rabbit, next to a small animal (a fly) or a large animal (an elephant). Try this yourself. Most people report visualizing a very small image of a rabbit next to an elephant and a much larger image of a rabbit next to a fly. When then asked to search for a property of the target (the rabbit’s ears), people take about 200 milliseconds longer when “looking” at the small image than when “looking” at the large image. Because the image is an analog of the real rabbit, a small image is harder to visualize clearly.

Finally, Kosslyn, Ball, and Reiser (1978) asked participants to memorize several locations on a map. Participants were asked to form a mental image of the map and focus attention on a specific starting location. The experimenters then named a second location and asked the participants to mentally find their way, in a straight-line path, to the new location. The results showed that the time needed to perform this mental scan increased proportionally to the real distances on the map. If the new location was close to the starting location, then the scan time was brief. If it was far away, then more time was needed.

The format of knowledge representations may be a concrete, perceptual-like imaginal code or an abstract, verbal-like propositional code.

The functional equivalence hypothesis claims that visual imagery uses the same mental representations, processes, and neural structures as does visual perception.

Imagery Versus Perception.

All of these studies suggest that visual imagery is functionally the same as visual perception. The functional equivalence hypothesis states that visual imagery, while not identical to perception, is mentally represented and functions the same as perception (Finke, 1989). For example, close your eyes and imagine in your mind’s eye the room in which you are now seated. Once you have an image of the room, imagine that a close friend enters the room and walks toward you. Your experience is likely to be similar, although not identical, to actually perceiving the room and the movement of your friend walking toward you. The objects and spatial relations among them seem to be represented much the same way as they are in viewing reality. According to this hypothesis, experiencing a change in a mental image ought to be analogous to experiencing a change in perception. The spatial relations of a visual image ought to be analogous to the spatial relations of a real environment.

Recent neuroimaging studies further support the view that imagery uses the same neural machinery as does perception. Positron emission tomography (PET), functional magnetic resonance imaging (fMRI), and event-related potential (ERP) studies indicate that brain activity during imagery is localized in the areas known to be used in vision. For example, when a person is reading a list of concrete nouns (e.g., cat) and forming an image of each object, greater ERP recordings are obtained from the primary visual cortex in the occipital lobe than when the person is just reading the nouns and not forming an image of each object (Farah, Peronnet, Gonon, & Girard, 1988). Lesions in the visual areas not only disrupt sight but also disrupt imagery (Farah, 1988). For example, a patient with bilateral lesions in the ventral occipitotemporal areas known to be involved in face perception had trouble imagining faces well known to him. When asked to imagine the face of Abraham Lincoln, it appeared to him with a distorted length and shape (specifically, short and round). Finally, Kreiman, Koch, and Fried (2000) directly recorded the activity of single neurons in the medial temporal lobes of epilepsy patients who had failed to respond to drug treatments. They identified neurons that showed selective changes in firing rates while viewing a particular object (e.g., a face, a baseball) and again when imagining the same object. The activity recorded from the regions of the brain known to be involved in declarative memory might have reflected either the retrieval of the object from long-term memory or the maintenance of the image in working memory.

There is an important difference between perception and imagery, however. Images must be maintained in working memory, whereas perceptions are continuously available without relying on memory. Finding a hidden part of an object is much easier when one can perceive the object than when it is mentally imagined (Reed, 1974). When participants performed the task illustrated in Learning Activity 7.1, they correctly identified the hidden parallelogram only 14% of the time when they had to rely on a mental image held in working memory. It was much easier to detect the rectangle. Reed concluded that the object was verbally labeled as the Star of David and stored propositionally in working memory (Star, David) rather than as an image that might be readily manipulated and searched. Although it was possible to identify a simple shape, such as a rectangle, the complexity of the parallelogram remained hidden in the mental image. Verbally labeling the image would encourage a propositional representation rather than a purely visual-spatial representation.

Learning Activity 7.1 The Limitations of Mental Images

Study the object shown below. Now, form a mental image of the object, and without looking back at the book, try to find a rectangle in the image. Could you find one? What about a parallelogram?

The Star of David demonstration of the limitations of visual imagery.

Mental Maps.

One common use of imagery is to construct spatial mental models. As Tversky (1991) explained,

There are many simple, everyday tasks, such as following road directions, using instructions to assemble a bicycle, reading a novel, or helping to solve your child’s geometry homework, that seem to entail constructing a spatial mental model from a description. In order to comprehend Go straight till the first light, then turn left, go down about three blocks to Oak, and make a right, it is useful to have a spatial representation. (p. 109)

Mental maps reveal one’s beliefs about the environment. Which city is farther west—San Diego, California, or Reno, Nevada? The mental map that most people generate places San Diego farther west (Stevens & Coupe, 1978). After all, California is west of Nevada, so surely San Diego must be farther west (see Figure 7.5). It turns out not to be the case. Similarly, most people regard Philadelphia, Pennsylvania, as farther north than Rome, Italy, but this is incorrect as well (Tversky, 1981). Rome is associated with a warm Mediterranean climate and so is assumed to be farther south than it really is.

Figure 7.5 Mental maps of the United States versus real maps.

The distortions in our mental maps reflect our beliefs about how the world is organized. Tversky (1981) argued that these beliefs and associated images are simplified by using two heuristics, or rules of thumb. One is the alignment heuristic. The odd shapes of the United States and Europe are difficult to imagine exactly, so people tend to align their shapes at the same latitude. Doing so would definitely place Rome at a southern location relative to Philadelphia. But in reality, Europe lies at a more northern latitude relative to the United States.

People also use a rotation heuristic, according to Tversky (1981). This suggests that states or countries that are tilted in reality are visualized as more vertical than they really are. They are rotated to the perpendicular, in other words. This explains the San Diego/Reno confusion. California is imagined in the mind’s eye as more or less vertical, forming the western boundary of much of the United States. In fact, California is rotated so that San Diego lies much farther east than most people realize. Reno, sitting on the border of California and Nevada, actually lies slightly west of San Diego. Because of the rotation heuristic, the mental map constructed is at odds with reality.

Just as people can make errors in geographic judgments by consulting their mental maps of the world, it has commonly been found that reliance on images in visual and spatial working memory distorts judgments about the size of objects (Smith, Franz, Joy, & Whitehead, 2005). For example, most people overestimate the size of the loaf of bread, as illustrated in Learning Activity 7.2. The functional equivalence hypothesis leads us to expect that visual perception and visual imagery ought to be closely linked. To illustrate, an object that is placed at a distance from a viewer is perceptually smaller; similarly, when one imagines the same object at a distance, it is also smaller in the mind’s eye. Smith et al. hypothesized that images of common objects that are often manipulated close to the body would be stored in long-term memory and retrieved in the task. The images stored in long-term memory might thus distort the size judgment in this task much in the same manner that a mental map can distort geographic judgments. To test their hypothesis, the investigators reasoned that frequent visual experience ought to be essential to storing such images in long-term memory. They compared, then, individuals who had been blind for a portion of their lives with sighted people in performing this task. The blind would need to rely on manual rather than visual representations, leading to the prediction that they should be more accurate at manually gauging the size of a loaf of bread or other common objects, compared with the sighted. The findings supported their hypothesis. Thus, mental maps and object images stored in long-term memory can produce some interesting kinds of errors in judgment, despite the fact that visual imagery is closely linked with perception.

Learning Activity 7.2 Estimating Object Size

Try to form a vivid mental image, with your eyes closed, of a loaf of bread. Once you have this image clearly in mind, while still keeping your eyes closed, use your hands to scale the length of the loaf by separating your left from your right hand by the correct amount. Finally, without moving your hands, open your eyes and check on the distance between your hands. Is it accurate or in error? If it is in error, did you overestimate the size of the loaf of bread or underestimate it?

Propositions

Knowledge can be coded in terms of abstract representations called propositions instead of as images. In logic, a proposition refers to the smallest unit of knowledge that one can sensibly judge to be true or false. “Fred is tall” is a proposition; “tall” or “is tall” is not. A proposition is an assertion that may be understood and evaluated. It is an abstract representation of the meaning conveyed by language—by all words, phrases, sentences, paragraphs, and whole speeches and documents.

One way to show the elements of a proposition is in a list format (Kintsch, 1974). The list begins with a relation followed by a set of arguments. Verbs, adjectives, and other words that convey a relationship are listed first. The arguments follow in a specified order; they define the agent of an action (e.g., who does X?), the object of the action (e.g., does X to what?), the time of the action, and other elements of a meaningful assertion.

The propositional code, then, breaks down knowledge into individual factual assertions and represents how they all are related to one another. Consider a few of the sentences that Bransford and Franks (1971) presented to participants in their experiment concerned with integration in memory encoding. As shown in Figure 7.6, each sentence can be represented by a set of abstract propositions in which each is expressed as a relation and a set of arguments.

Propositions provide an abstract and elemental representation of the meaning of verbal information. They provide the mental code for language. The outcome of Bransford and Franks’s (1971) study illustrates this point well. Recall from Chapter 6 that they tested recognition memory for sentences of the type illustrated in Figure 7.6. People recognized sentences as old if they contained a large number of propositions expressed in the study sentences, even though the specific sentences presented on the test were, in fact, new. For example, this new sentence–“The ants in the kitchen ate the sweet jelly which was on the table”–was falsely recognized by nearly all of the participants. It represented the prototype sentence that captured the propositions that occurred frequently in the study sentences.

Figure 7.6 Propositional representations of the sentences used in an experiment on integration in memory encoding.

Latent semantic analysis (LSA) is a mathematical procedure for automatically extracting and representing the meanings of propositions expressed in a text. It allows one to compare the similarity of the propositions in two texts without going through the tedious task of manually determining the propositions of each sentence (Kintsch, 1998). It can, for example, represent the propositions of an entire encyclopedia in a computer database. The procedure works by representing the co-occurrence of words and their contexts (Landauer & Dumais, 1997). For example, the word model occurs in numerous contexts in this book. LSA represents the associations between this word and its contexts in order to provide a usage-based meaning of model. It does this for every word it encounters in the text. Words that occur infrequently in the same contexts (e.g., model) are well defined by their usage, whereas words that occur frequently in many contexts are less so (e.g., the). The net result allows one to compute the similarity in meaning between two words, between a word and a paragraph, or between two paragraphs.

Once LSA has extracted and represented the meaning of a text, it can use this representation to answer questions about it. For example, the mathematical procedure used in LSA was applied to the 4.6 million words that compose Grolier’s Academic American Encyclopedia. These words came from 30,473 articles written at a level appropriate for young students. The model’s knowledge of word meanings was then tested using the Test of English as a Foreign Language (TOEFL). Each item presents a test word along with four alternatives that are more or less close in meaning. The task is to select the word closest in meaning to the test word. Because LSA can compute the similarity in meaning of two words, it is well suited for taking this test once it has been trained by “reading” an encyclopedia. In fact, LSA was about as well prepared as typical applicants to U.S. colleges from non-English-speaking countries, who score an average of 64.5% correct on the TOEFL. When LSA was tested, it picked the correct choices 64.4% of the time. By learning the similarity of word meanings from where and how often they occur in an encyclopedia, LSA behaves as if it knows English vocabulary.

A proposition is coded as a relation and a set of arguments specifying an assertion that may be true or false.

LSA is a mathematical procedure for extracting and representing the propositions expressed in a text. It allows one to compare two texts in terms of the similarity of their propositional content and to answer questions about a text.

Conclusion

The debate over the nature of knowledge representation spawned experimentation that settled one point conclusively. Prior to the experimentation discussed above on the analog nature of visual imagery, a convincing case could be made that our subjective experience of visual imagery was not a reliable guide to the underlying mental representation (Pylyshyn, 1973). A symbolic, amodal code could, in theory, store the properties and relations needed to generate conscious images. As shown above, numerous experiments rejected this line of argument against modality-specific analog representations, such as those seen in visual imagery. For more than 35 years now, both propositional and imaginal codes have been regarded as essential to our understanding of how knowledge is represented in the brain (J. R. Anderson, 1978).

In the past decade or so, however, some skepticism has been raised about the psychological validity of amodal, propositional codes (Barsalou, 1999; A. Martin, 2007; Rubin, 2006). For example, in the Bransford and Franks (1971) experiment discussed in the previous chapter, it appeared that people stored abstract ideas or a list of propositions as a result of reading a list of 30 sentences. These ideas—stored as propositional codes—were then subject to the process of integration, leading to errors in recognition memory. An alternative explanation of their results is that the words, per se, in each sentence were stored in the form of a modality-specific verbal code (i.e., as heard or seen words). Given that propositions are verbal-like in nature, it is difficult to distinguish between the words themselves and the abstract propositions underlying the words (Sadoski & Paivio, 2001). It should be pointed out that, although LSA can be conceptualized as a means for extracting propositions from written text, the actual units of analysis are in fact verbal. That is to say, LSA analyzes the co-occurence of the words in a text.

The growing evidence from cognitive neuroscience emphasizes that the conceptual knowledge of semantic memory is often grounded in the perceptual and motor activity of the brain. For example, processing the category of animals activates the modalities involved in perceiving and interacting with animals; animals activates the brain areas for visual form and animate motion. Tools, on the other hand, activates regions involving action and inanimate motion (A. Martin, 2007). It is too soon to rule against the validity of propositional codes, but further work on how the brain recruits perceptual and motor systems for the storage of knowledge in semantic memory ought to be revealing and of interest.

7.2 Representing Concepts

factual and conceptual knowledge constitute key ingredients in speaking and listening, in reading and writing, and in problem solving and thinking. Without an ability to acquire, represent, and use knowledge about the world and its meaningful symbols, high-level forms of human cognition would not be possible. It is through semantic memory that we categorize the world, allowing us to ignore the details and to see a specific object as a general kind. Lakoff (1987) aptly underscored the importance of categorization as follows:

There is nothing more basic than categorization to our thought, perception, action, and speech. Every time we see something as a kind of thing, for example, a tree, we are categorizing. Whenever we reason about kinds of things—chairs, nations, illnesses, emotions, any kind of thing at all—we are employing categories. Whenever we intentionally perform any kind of action, say something as mundane as writing with a pencil, hammering with a hammer, or ironing clothes, we are using categories. The particular action we perform on that occasion is a kind of motor activity. . . . They are never done in exactly the same way . . . , yet . . . we know how to make movements of that kind. Any time we either produce or understand any utterance of any reasonable length, we are employing dozens if not hundreds of categories: categories of speech sounds, of words, or phrases and clauses, as well as conceptual categories. Without the ability to categorize, we could not function at all, whether in the physical world or in our social and intellectual lives. (pp. 5–6)

Concepts are the general ideas that enable the categorization of unique stimuli as related to one another. The concept of a tree, for example, allows for considerable variations in the individual examples of trees that fit the general idea. A tree concept includes a set of variable dimensions, such as the types of roots, trunks, branches, and leaves, that may be instantiated in a large number of ways. Many unique trees all fit the general concept. The stimuli categorized may be concrete objects or abstract ideas. Mathematical (e.g., imaginary numbers), philosophical (e.g., free will), and psychological (e.g., depression) concepts are among those that group together highly abstract entities.

Rule-Governed Concepts

The classical approach to categorization assumed that concepts are defined by a set of singly necessary and jointly sufficient features. In other words, the defining features of a concept were governed by a conjunctive rule stating that each and every feature must be present for an object to fit the concept. For example, the concept of a cow could be understood in terms of the features animate, four-legged, hoofed, female, and adult and other distinctive features that set the object apart from, say, a buffalo. Rule-governed concepts specify the features and relations that define membership in the class on an all-or-none basis (Bourne, 1970; Bruner, Goodnow, & Austin, 1956). The classical view of categorization regarded all concepts as rule governed.

Some abstract concepts can be viewed as rule governed. The definitions of real numbers, gravity, grand larceny, and a touchdown in American football can be specified by mathematicians, physicists, lawyers, or referees, for example. When a football is carried by the runner so that it crosses the vertical plane above the goal line of the opponent without being dropped, a touchdown is scored. This is not to say that reasonable people cannot argue whether a particular play in question resulted in a touchdown, but rules of football spell out the concept unambiguously. Fans may disagree about the features of the play, but not about the concept itself.

Not all abstract concepts fit the classical view very well, however (Lakoff & Johnson, 1980). For example, what are the defining features of truth or justice? Moreover, the objects encountered every day in the environment do not belong to rule-governed concepts (Rosch & Mervis, 1975; Smith & Medin, 1981).

Object Concepts

Object concepts refer to natural kinds (biological objects) and artifacts (human-made objects). They are often organized hierarchically in subordinate, basic, and superordinate categories. An example of a hierarchy of natural kinds would be robin (subordinate), bird (basic), and animal (superordinate; Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976). Similarly, claw hammer, hammer, and tool illustrate a hierarchy of artifacts.

In contrast to rule-governed concepts, there did not appear to be any set of defining features for object concepts. The philosopher Ludwig Wittgenstein was the first to articulate that some linguistic categories seemed to defy the classical description by lacking defining features. For instance, take the concept of a “game.” As Lakoff (1987) observed,

Some games involve mere amusement, like ring-around-the-rosy. Here there is no competition—no winning or losing—though in other games there is. Some games involve luck, like board games where a throw of the dice determines each move. Others, like chess, involve skill. Still others, like gin rummy, involve both. (p. 16)

Games do have a “family resemblance” to one another. Each member shares some features in common with other members, but no set of defining features is common to all. Rather, what makes games a category is that the examples show similarities to one another in a wide variety of ways. Object concepts are ill-defined in that it is difficult to decide whether or not a particular poor example falls within the legitimate range of membership.

The fuzzy boundary of class membership can readily be seen from the results of Labov (1973), who presented people with the cup-like objects seen in Figure 7.7. Exactly at what point does an object cease to be a cup and become a bowl, a glass, or a mug? The first four items varied in terms of the ratio of the width of the cup to its depth. Labov’s results showed that when participants were thinking about food, the probability of their calling the object a bowl increased gradually—not abruptly—as this ratio increased. Similarly, the probability of their calling the object a cup decreased gradually as the ratio of the width of the object to its depth increased. But for ratios falling in the middle of the overall range, both responses often occurred, reflecting the fuzzy boundary of concept membership. In a neutral context, the cup response became less likely, but people were more reluctant to call Items 3 or 4 a bowl. This difference in judgment with context shows the flexible boundary of object concepts; they shift depending on which other concepts are active in memory.

Prototypes.

An object differs from other members of a concept in terms of its features and the frequency of occurrence of these features in the category (Attneave, 1957). For example, birds differ from one another in terms of size, color, and shape of wings. The range and frequency of occurrence of these features are also represented; one easily understands that a turkey has rare features, whereas a sparrow has more common features. The prototype is the best or most typical member of a category and serves as an important mental representation of the concept. Moving away from the prototype, there is a gradient of membership. At least for residents of North America, the robin assumes the role of prototype for the bird concept (see Figure 7.8). The turkey lies far out on the gradient from the prototype, whereas the blue jay and sparrow lie closer. The crow or the hawk fall in between these extreme cases.

Figure 7.7 The boundary of the object concept of cup is fuzzy.

SOURCE: From Labov, W., The boundaries of words and their meanings, in C. J. N. Bailey & R. W. Shuy (Eds.), New Ways of Analyzing Variations in English, copyright © 1973. Reprinted with permission of Georgetown University Press.

Rosch and Mervis (1975) asked people to rapidly list all the features they could think of for a variety of common objects. The objects fit the superordinate categories of furniture (e.g., chair, piano, telephone), vehicle (e.g., car, tractor, elevator), fruit (e.g., orange, grapefruit, olive), weapon (e.g., gun, tank, screwdriver), vegetable (e.g., peas, lettuce, rice), or clothing (e.g., pants, pajamas, necklace). A total of 20 objects per category were used, and they clearly varied in typicality, as can be seen from the examples in Table 7.1.

The key outcome was that few features applied to all 20 objects. Those few that were common to all failed to distinguish the category from numerous other natural categories (e.g., “You eat it” was listed for all fruits, yet this applies to all foodstuffs). What defined the category was not a small set of defining features; rather, a large number of features applied to some–but not all–instances. Rosch and Mervis (1975) referred to this as the family resemblance structure of object concepts. The researchers computed a family resemblance score that reflected the sum of the frequency with which features applied to all instances of the category. They found that this family resemblance score correlated strongly with ratings of typicality. Thus, “orange” and “apple” achieve high typicality ratings in the “fruit” category because they exhibit features that occur frequently among all members of the category. “Tomato” and “olive” are given low typicality ratings because their features are not characteristic of most fruits.

Figure 7.8 Typicality of members in a basic-level category.

Table 7

The speed with which people categorize an example varies with how close it is to, or how far it is from, the prototype. One can correctly categorize a robin as a bird faster than one can a crow, which in turn can be categorized faster than can a turkey (Rosch, 1975). The order in which children learn the members of a category also depends on their typicality. Highly typical instances are acquired sooner than are atypical instances (Rosch, 1973). The typicality effect refers to the gradient of category membership, or differences in how well specific instances represent a concept.

Concepts as Theories.

Object concepts are also coherent in that they are integrated with other concepts through theories about how the world is structured. Put differently, a concept is embedded in and consistent with people’s background knowledge and folk theories (Keil, 1989; Medin & Ortony, 1989; Murphy & Medin, 1985). For example, a person’s knowledge and intuitions about biology dovetail with the concept of a bird. If a person believes that birds breathe, eat, and reproduce, then these abstract theoretical constraints would lead one to reject a robotic, manufactured robin as not really being a robin. It may look and act exactly like a robin, but if the person knows it is not an animate object, then it fails to fit the concept.

A concept can be coherent even when there are no obvious perceptual or functional similarities among its instances. For example, what are the similarities that put children, money, photo albums, and pets in the same category? Without considering how the concept is embedded in a person’s knowledge about the world, this makes no sense at all. However, in the context of “things to take out of the house in case of a fire,” individuals are able to list these and other examples without difficulty (Barsalou, 1983). In this case, then, the concept is coherent only by considering, in relation to the nature of fires, the importance of personal property and related world knowledge.

Similarity to perceptual and functional features, then, provides only part of the story of how human categorization works. Deeper, less accessible, theory-based features about the internal structure of natural kinds of objects (e.g., the genetic structure of a dog) constrain our categorizations, as well as do surface features (Gelman, 1988; Keil, 1989). The representation of a dog in semantic memory includes a theory about the internal structure of the organism. It is this internal structure that really makes a dog a dog. The functional uses of a dog (e.g., hunting) are less relevant to its classification, but functional uses are absolutely critical for artifacts such as a cup (e.g., drinking). Barton and Komatsu (1989) gave participants descriptions of objects that were either natural kinds or artifacts. The researchers changed either the internal molecular structure (e.g., a goat with altered chromosomes, a tire not made of rubber) or the functional features (e.g., a female goat not giving milk, a tire that cannot roll) of the examples to see how these different types of features affected categorizations. They discovered that changes in molecular structure were most important for natural kinds (e.g., a goat with altered chromosomes is not really a goat), whereas the opposite was true for artifacts (e.g., a tire that cannot roll is not really a tire).

Artifacts and natural kinds may be differentiated at a neural level of representation. When test subjects in fMRI and PET studies are shown a word that names a concept, two brain regions are consistently activated. One is the left ventrolateral prefrontal cortex (VLPFC), and the other is the ventral and lateral regions of the posterior temporal cortex. Stronger activation occurs in the left hemisphere than in the right hemisphere (A. Martin, 2007). The VLPFC mediates top-down control of semantic memory by guiding retrieval of concepts stored in the posterior temporal lobe. A category-specific loss of semantic memory can reveal the specific region of the temporal lobe involved. Patients with associative agnosia can have highly specific losses in their ability to make semantic categorizations. For example, one patient, J. B. R., lost the ability to identify the names of objects of living things, such as dogs and horses, but had little trouble with inanimate objects, such as umbrellas and chairs (Warrington & Shallice, 1984). Recall from Chapter 2 that the temporal lobe serves as the termination point for the lateral “what” pathway involved in object recognition. Patients with lesions in different regions of the temporal lobe experience deficits in naming particular categories of objects (Damasio, Grabowski, Tranel, Hichwa, & Damasio, 1996). As shown in Figure 7.9, damage to the anterior region of the left temporal cortex causes problems in naming the faces of famous people. Damage to the inferotemporal cortex, along the lower region, is associated with problems in naming animals. Last, naming tools was problematic for patients with lesions in the posterior region of the left temporal cortex. Using fMRI, researchers have identified distinct clusters of activity in the temporal and occipital cortex when subjects view tools, animals, houses, faces, or other objects (A. Martin, 2007). Recall from Chapter 2 the fusiform face area (FFA) discovered in the ventral aspect of the temporal lobe, for example. These regions can even be identified through neuroimaging while participants view complex visual scenes as opposed to pictures of single objects.

Figure 7.9 Regions in the temporal lobe are associated with representations of specific semantic categories.

The neocortical representations of concepts should not be thought of as limited to the temporal lobe. The concept of number, the foundation of all mathematical knowledge, has been extensively investigated. Human beings possess a rudimentary knowledge of numbers expressing amount even in infancy, and this capacity develops as children learn to count (Wynn, 1992). The concept of numerical quantity (one, two, three, and so on) or numerosity is represented in the brain by a specific location in the lateral region of the lobe called the intraparietal sulcus (IPS). Performing arithmetic problems of addition, subtraction, or multiplication with Arabic numerals activates the IPS, as does paying attention to the numerosity of a visual display of dots or an auditory display of a series of tones (Nieder & Dehaene, 2009). Changes in number but not object identity are detected by the IPS, whereas changes in object identity but not number are noted by the fusiform gyrus in the temporal lobe. In one experiment, participants were shown an array of dots (Piazza, Izard, Pinel, Le Bihan, & Dehaene, 2004). If the next display showed a different number of dots, the IPS became active. The researchers also found that the IPS is sensitive to the magnitude of the change in numerosity, showing systematically more and more activation as the numerical difference from one display to the next increases.

Object concepts have fuzzy boundaries and a gradient of category membership, with some instances being more typical than others.

7.3 Using Semantic Memory

How do people use semantic memory to answer simple questions? Is a dog a kind of animal? Is a hammer? How do you retrieve information about these object concepts to give the right answer? Models of the retrieval process have been proposed and tested by manipulating the specific kinds of questions asked and observing how the time to respond correctly was affected. In this final section of the chapter, these models are presented to show how the process of answering even simple questions is difficult to unveil.

Semantic Network Models

The subordinate, basic, and superordinate levels of concepts and their associated features can be organized into a hierarchical structure called a semantic network model. Figure 7.10 illustrates a small portion of what you know about the object concept of animals, with Level 0 representing the subordinate level, Level 1 the basic level, and Level 2 the superordinate level. Collins and Quillian (1969) proposed that retrieving information from this network involves working through the various levels as demanded by the task. For example, verifying that a canary can sing involves entering the network at Level 0 and searching the propositions attached to the canary node. But verifying that a canary can fly requires moving through the network to Level 1. Still more mental travel time is required to answer that a canary has skin, for this feature applies to all animals; it is a superordinate, or Level 2, feature.

Figure 7.10 A hierarchical network representation of concepts.

Notice that this hierarchical representation of features saves space in long-term memory. It is not necessary, theoretically, to store that a canary can fly and has skin because these facts can be deduced from the hierarchical structure of memory. The cognitive economy assumption claims that the features of a concept are represented only once at either the subordinate, basic, or superordinate level of the hierarchy. Although it may take more time to use semantic memory organized in this manner, less storage space is required than if properties were stored redundantly at all levels of the network.

Collins and Quillian (1969) examined how quickly people verified both conceptual feature questions and category questions (see Figure 7.11). Notice that, as predicted, the time needed to answer feature questions increased as a function of feature level. In addition, verifying that a canary is a canary (an identity judgment) is faster than verifying that a canary is a bird (a basic-level categorization). The most time was needed to verify membership in a superordinate category—that a canary is an animal. Note that nearly one tenth of a second is needed to move from one level to another in answering the category questions. Also note that searching the features associated with a particular level of category requires about 200 milliseconds more.

Figure 7.11 Time required to categorize a noun and to verify its properties.

Collins and Quillian’s (1969) proposal that concepts, or the words that name them, can be represented by a massive semantic network has proved to be very useful in designing a novel kind of dictionary. Instead of listing the words alphabetically, WordNet represents the semantic relations among words (Miller, 1999). A noun, verb, adjective, or adverb in the language is represented by a synset, a set of synonyms for each noun, verb, adjective, or adverb in the language. If a word has more than one meaning, then it appears in more than one synset. WordNet contains 122,000 words in 100,000 synsets. The synsets, in turn, are organized in the network by 139,000 pointers representing semantic relations. WordNet easily retrieves synonyms and antonyms for a given term. This system of organization can be very useful in information retrieval systems because it expands the number of possible items to search in a database. Furthermore, the different meanings of a word in different contexts are recognized by WordNet. Thus, the term board will not be misunderstood when one is intending to search for a surfboard versus a bulletin board.

Subsequent research on Collins and Quillian’s (1969) model muddied the waters, however. To begin, Rips, Shoben, and Smith (1973) reported that the hierarchical structure of the network was not consistent. If it were, then one should verify that “A collie is a mammal” faster than “A collie is an animal.” The class of mammals is a subset of all animals; therefore, strictly speaking, it ought to be searched faster than the superordinate class. The results did not show this, suggesting that people do not clearly represent a strict hierarchy of class relations.

Furthermore, the assumption of cognitive economy also encountered problems. The associative strength or frequency with which a property occurred proved to be more important than the level of hierarchy. Conrad (1972) reported that strong associations (e.g., an orange is edible) can be quickly answered even though the property of edible ought to require moving to a higher level in the network. Moreover, Conrad failed to find differences where Collins and Quillian (1969) expected them. For instance, the verification time remained the same for these types of statements: “A shark can move,” “A fish can move,” and “An animal can move.” In other words, if one selects a high-level property, theoretically it should take more time to move back down the network to the fish and animal levels (the reverse direction of the Collins and Quillian experiment). The results indicated no difference.

Finally, the typicality effect discussed earlier finds no place in a network model. Why should it be faster to verify that a robin is a bird than to verify that a canary is a bird? Obviously, the specific network model proposed by Collins and Quillian (1969) failed to capture the necessary phenomena. Later attempts came closer (Collins & Loftus, 1975), but at the cost of a lack of precision in their predictions (Chang, 1986).

WordNet uses a semantic network to represent the meanings of 122,000 English words. Each noun, verb, adverb, and adjective in WordNet is defined by a synset or the set of all synonymous words.

Typicality effects and violations of the cognitive economy assumption led to the rejection of Collins and Quillian’s model of semantic memory.

The Feature Comparison Model

Other researchers took a different tack from network models to account for the typicality effect observed with simple inquiries to semantic memory. Smith, Shoben, and Rips (1974) assumed that people first tried to retrieve the typical or characteristic features of a concept in order to answer a question such as “Is a robin a bird?” Such a list of features works best for the prototype, robin, and is satisfactory for other examples related to the prototype (e.g., canary). However, for cases that are very different from the prototype (e.g., penguin), the overall similarity is not satisfactory. Instead, a second stage is needed in which the defining features of the concept, which go beyond obvious perceptual similarities, must be retrieved. Thus, the sentence “A robin is a bird” will yield a high degree of similarity between the example and the characteristic features, and can be answered without further search. By contrast, “A penguin is a bird” requires a two-stage analysis to yield the correct answer. Therefore, the feature comparison model assumes that characteristic and defining features are assessed first in answer to simple categorization questions, and that defining features are then retrieved if necessary.

Besides providing an explanation of the typicality effect, the feature comparison model also nicely accounts for another consistent result in the literature–the category size effect. D. E. Meyer (1970) reported that deciding whether a relation applies for a small category (e.g., all robins are gems) is faster than deciding whether it applies for a large category (e.g., all robins are stones). To respond “false” to these statements, one must search the categories of gems and stones in long-term memory. Meyer reasoned that the larger the category that must be searched, the more time that is needed to disconfirm the statement. The same result occurred for true categorical statements (e.g., all collies are dogs vs. all collies are animals). Of course, we saw this same outcome in the category-level effect reported by Collins and Quillian (1969).

The feature comparison model explains the category size effect by noting that Stage 2 processing would be needed only for large categories. For instance, to decide that a robin is a bird, one’s initial comparison of all features would yield a high degree of similarity. An immediate “true” response would be possible. Yet to decide that a robin is an animal would demand the extra step of comparing defining features alone. Perhaps the unusual violation of the category size effect noted earlier might stem from the similarity of features. Recall that people can judge “A dog is an animal” faster than they can “A dog is a mammal.” The overall similarity of dog and animal might be quite high, leading to an immediate response after Stage 1 processing. Because the category “mammal” is less familiar to most people than “animal,” a second stage of processing the defining features may be needed.

In sum, the feature comparison model accounts for most of the major results in the literature on semantic memory. However, it too has problems (Chang, 1986). One empirical problem is that false statements involving concepts that are highly similar ought to require Stage 2 processing and, thus, take extra time to answer. For example, “All dogs are cats” should take more time to disconfirm than “All animals are birds.” The latter statement should be disconfirmed after Stage 1 only. Yet Glass and Holyoak (1975) found that highly similar concepts, such as dogs and cats, are disconfirmed the fastest. Thus, even relatively simple uses of semantic memory are complex to model.

Nevertheless, feature overlap is clearly a central quality of how questions are answered using semantic memory. The category-based induction task illustrates this point as well as does the task presented earlier. This task can be illustrated by the following statements: “Robins have wings,” “Bluebirds have wings,” and “How likely is it that all birds have wings?” Osherson, Smith, Wilkie, Lopez, and Shafir (1990) found that this inductive argument is much stronger than one that lessens the degree of feature overlap in the premises, such as the following: “Robins have wings,” “Ostriches have wings,” and “How likely is it that all birds have wings?” The overall similarity of features is one important determinant of how people reason with the concepts of semantic memory.

The feature comparison model accounts for typicality effects and the category size effect by postulating a two-stage process of checking, first, defining and characteristic features, and second, defining features only.

Summary

1. Semantic memory contains factual and conceptual knowledge represented by means of imaginal and propositional codes. Imaginal codes are concrete and perceptual-like, whereas propositional codes are abstract and verbal-like. Each proposition represents a single, limited assertion about the world that can be judged as true or false. Propositional codes, then, break down knowledge into individual factual components and represent how they all are related to one another. An image behaves in much the same manner as the object it represents. Imagery and perception are functionally equivalent. For example, the time needed to rotate a mental image increases linearly with the angle of rotation, much like the time needed to rotate the object itself. Behavioral and neuroimaging evidence indicate that a dual-coding system is used in human semantic memory.

2. Object concepts refer to natural kinds (biological objects) and artifacts (human-made objects). An object concept categorizes objects and events of significance in the real world. As with any concept, an object concept treats objects that differ as being the same if they fall within its defining boundary. With object concepts, this membership boundary is fuzzy or flexible. Concepts are often organized hierarchically. The hierarchy of subordinate (robin), basic (bird), and superordinate (animal) concepts illustrates this point. Rule-governed concepts, by contrast, are defined by logical relations among a set of defining features. Their membership boundaries are clear-cut and inflexible. For example, the abstract mathematical concept of an integer is well defined.

3. All concepts specify the dimensions along which members differ from one another and order the members in terms of a gradient of membership or typicality. The prototype represents the best example of a given concept. Concepts may also be coherent, in the sense that they relate in a deep, theoretical manner to other concepts or other representations of world knowledge. Concepts may also be organized hierarchically, with the basic level of categorization providing the optimal amount of information about its members. Schemas organize related concepts in meaningful ways. For example, a restaurant script is a kind of schema that organizes everything we know about the routine activities of entering a restaurant, ordering food, paying for the food, and so on.

4. Two of the major models of how we retrieve information from semantic memory were presented. Network theory assumes that knowledge is represented hierarchically and that features connected at a superordinate level are not redundantly represented at lower levels. Retrieving a fact involves working through the various levels of the network and searching the relevant nodes for feature information. Feature comparison theory assumes that each concept includes a list of characteristic and defining features. Retrieving a fact first involves a comparison of overall similarity based on both feature types. If similarity is low, then a second stage of comparison is needed based on defining features only. The feature comparison model handles experimental results better, although it, too, has difficulties in accounting for some findings.

Key Terms

· imaginal code

· propositional code

· functional equivalence hypothesis

· latent semantic analysis (LSA)

· rule-governed concepts

· object concepts

· prototype

· family resemblance structure

· typicality effect

· folk theories

· semantic network model

· cognitive economy assumption

· synset

· feature comparison model

· category size effect

Questions for Thought

· In what daily activities do you use imaginal mental representations to represent familiar concepts? Can you describe activities that are more likely to draw on propositional mental representations?

· How does a judge define the legal concept of manslaughter or third-degree murder to a jury? What are the defining features of this rule-governed concept? How would the concept differ if prototype and fuzzy boundary were instead used in law?

· What is the first word that comes to mind when you think of the concept of “cold”? Now, using your response as a stimulus, what concept do you think of next? Continue this chain of associations through several steps. How do these words relate to one another when positioned in a semantic network that shows links that are labeled with the nature of the relationship (e.g., “Is an example of” or “Has the property of” or “Is opposite in meaning to” or “Is similar in meaning to”)?