Dual Coding Theory and Education

words from school subjects will be more difficult to define and categorize, will be more confusable, and will show other theoretically and pedagogically interesting differences from concrete words. With respect to the meaning of larger verbal units, such as sentences and paragraphs, experimental research has also demonstrated a major role for imagery processes. Paivio and Begg (1971b) showed that comprehension and imagery latencies are similar in magnitude for concrete sentences, and highly correlated. Relative to abstract sentences, concrete sentences generally are understood faster (e.g., Jorgenson and Kintsch, 1973; Klee and Eysenck, 1973), although not always significantly so (e.g., Paivio and Begg, 1971b). O'Neill and Paivio (1978) also found that the random exchange of words in sentences was more disruptive to the meaning of concrete than abstract sentences. Specifically, abstract sentences with words randomly substituted across unrelated sentences were rated as more comprehensible and sensible than comparable concrete sentences. One interpretation of this finding is that the meanings of abstract sentences are more vague (i.e., less definite) than the meanings of concrete sentences, and therefore less disrupted by random substitutions. In addition to research on literal language, concreteness and imagery have also been shown to be important in processing figurative sentences (Katz et al., 1988; Paivio, 1971, 1986; Paivio and Begg, 1981; Paivio and Clark, 1986). Katz et al. (1988), for example, found that the rated imagery value of 484 metaphors correlated with a number of semantic attributes rated by university students (e.g., .79 with ease of comprehension and .80 with ease of interpretation). It has been hypothesized that figurative language may serve to concretize concepts by relating abstract topics to concrete vehicles (e.g., governments are elephants). Educational research has confirmed the importance of imagery and concreteness for the comprehension of sentences and larger textual units. Image generation and supplementary pictures generally benefit text comprehension (see Denis, 1984), although it is not always easy to distinguish comprehension effects from closely related memory effects discussed later. Reading of educational materials seems to elicit substantial amounts of spontaneous (i.e., uninstructed) imagery. Long et al. (1989), for example, questioned grade 5 students about their thoughts at various points during the reading of passages from textbooks and obtained indicators of imagery on over 60% of the trials. Imagery may be especially common for emotionally arousing passages, a topic to which we return later (e.g., Sadoski, 1983, 1985; Sadoski and Goetz, 1985). Other findings indirectly support the hypothesized role of imagery in reading. Glenberg et al. (1987), for example, showed that university students read text faster if objects that would be spatially contiguous in an imaginal representation of the passage were mentioned together in the text, rather than separated. Dual Coding Theory and Education 161 The educational importance of concreteness and imagery has been further demonstrated by research on text readability. Flesch (1950), for example, measured text concreteness by the percentage of definite words (e.g., names of people, nouns that indicate a specific time) and found a correlation of -.55 between this measure and the average grade level of children who could correctly answer half of the test questions. In an experimental study, Wharton (1952; see Wharton, 1980, 1985) substituted "picture forming" words into history passages, and observed the effects on comprehension and interest. To illustrate, the sentence "With England in control of the seas and France invincible on land the war became an economic contest" was revised to "With England sweeping the seas and France overrunning the land the war lapsed into an economic tussle." University students found the high imagery texts more interesting and scored higher on a comprehension test than with the originals, even though both texts were equated on traditional aspects of readability (e.g., sentence length) and the substituted words were less familiar than the originals. Educational research has also demonstrated a relation between text readability and counts of abstract or concrete nouns (see Klare, 1974/1975; Morris and Halverson, 1938, cited in Gilliland, 1972). Other readability formulae may reflect concreteness indirectly. Some methods, for example, count the frequency of basic words in the text, but basic word lists contain few abstract words. Moreover, Cloze measures of readability, which require subjects to guess deleted words from passages (e.g., McKenna and Robinson, 1980), implicate various associative correlates of concreteness discussed earlier, such as associative strength (e.g., Cramer, 1968; deGroot, 1989; Paivio, 1968). Although the positive effects of imagery and concreteness on comprehension are consistent with DCT, such effects are not universal (e.g., Long et al., 1989). Rather than discrediting DCT, however, these qualifications often involve interactions with individual differences and other specific effects that demonstrate the strengths of DCT even more clearly than the simple effects of imagery and concreteness. Consider, for example, the fact that university students who image narrative passages read the passages more slowly than students who do not use imagery (Denis, 1982). This finding, which only appears to contradict the hypothesized benefits of imagery for comprehension, is explained by the DCT assumption that concreteness and related effects result from activation of quasi-perceptual imagery representations. Being perceptual, images can compete with reading for shared perceptual processes resulting in a slowed reading rate (Denis, 1982). Teachers who instruct their students to image while reading might therefore expect a decline in reading rate. Conversely, an emphasis on reading speed during instruction might decrease the likelihood of imaginal processing of text, with its associated benefits on comprehension and, as we shall see, memory. 162 Clark and Paivio Other specific phenomena consistent with DCT concern the interaction between individual differences and the use of imagery in text comprehension. As noted earlier, DCT assumes that people vary in the ease and skill with which they use nonverbal, imaginal processes. Individual differences should contribute to comprehension processes and interact with other variables, such as instructions to image. A study by Denis (1982) illustrates this line of inquiry. University students were classified as high or low imagers on the basis of a questionnaire and read descriptive passages under instructions that emphasized speed, required imagery, or were self-paced. All subjects read faster under speed than under image instructions, and there were no differences between high and low imagers in these conditions. Under self-paced instructions, however, only high imagers read at a slower rate than under speed instructions. This finding is consistent with the hypothesis that high imagers spontaneously imaged the stories, which slowed their reading rate. The reviewed work confirms a central role for nonverbal processes in the representation and processing of text, including educational materials. Nonverbal processes presumably play an even more important role in processing the large amount of educational knowledge that is nonverbal in nature. Visual information in schools would include maps, graphs, science apparatus, geometric shapes and principles, and theoretical models for many constructs in biology and other sciences (e.g., cells, atoms, chemical molecules). Moreover, imagery is implicated in the associative structure of knowledge, to which we now turn. Associative Structure of Knowledge According to DCT, words derive meaning from their semantic relations with other words, as well as from images. Associative relations contribute to the meaning of all words, and are the primary source of meaning for abstract words that lack object referents (e.g., knowledge, syntax, schema, force, strategy). The essential idea is that associative relations connect words to one another, and activation of this associative structure contributes to the meaning of the words. Although similar to other models of semantic memory (e.g., schema theory, hierarchical structure), DCT emphasizes diverse associative relations between verbal representations rather than highly restricted relations (e.g., superordinate, property) between amodal abstract representations. Moreover, images can be incorporated into the meaning of concepts, especially concrete concepts, although here we emphasize verbal associative aspects of meaning. In general, associative knowledge is viewed as a complex collection of diverse associations between Dual Coding Theory and Education 163 verbal representations, with additional referential connections to nonverbal components of meaning. Much of the cognitive and educational research on the structure of semantic memory has emphasized superordinate networks, scripts, and other hierarchical structures in which instances or component parts converge on superordinate nodes (e.g., Bower et al., 1979; Collins and Loftus, 1975; J. M. Mandler, 1984). From a DCT perspective, hierarchical structures reflect associations in which multiple words converge on superordinate labels (Clark and Paivio, 1984). Verbal representations for "dog," "cat," "lion," and other animal words, for example, become associated with the verbal representation for "animal." Early research on verbal concept learning demonstrated that activation of superordinate and related category information depends on such variables as the strength of the individual links between instances and concepts, the spacing between presentation of instances from the same category, and instructions (e.g., Underwood and Richardson, 1956). Instructions to find what is shared by a set of instances, for example, increases the likelihood that the shared category or other common feature is identified. Many facets of educational knowledge can be conceptualized in terms of verbal associative networks. Indeed, the prototypical examples of a hierarchical network are biological and other science taxonomies in which words are linked into an associative hierarchy with multiple levels that are themselves labeled (e.g., genus: lion--, family: Felidae--, order: Carnivora --, class: Mammalia --, phylum: Vertebrata --, kingdom: Animalia). Hierarchical structures have also been proposed for the representation of text (e.g., Kintsch and Van Dijk, 1978). Some texts activate preexisting semantic structures, but even when prior schemata are not available, comprehension processes construct an organized hierarchy of statements using linguistic cues to integrate sentences that vary in importance or centrality (e.g., Halliday and Hasan 1976; Kintsch and van Dijk, 1978). This research has produced detailed models of the cognitive processes that underlie mental outlines of text, such as identifying both main ideas, which are assigned central roles, and extraneous material, which can be excluded from the associative network or placed at lower levels (e.g., Williams et al., 1981; van Dijk and Kintsch, 1983). Hierarchical and related associative structures relevant to different school subjects have been measured and studied using a variety of techniques (e.g., Preece, 1976; Shavelson, 1974). In an early study, Johnson (1964) had high school students with different science backgrounds freeassociate to physics terms (e.g., mass, momentum, velocity). He found that the associative structure of the terms varied with the amount and recency of instruction in science. For example, students with relevant classroom in164 Clark and Paivio struction were more likely to respond with other terms on the list that were related to the stimulus word by scientific principles, whereas students lacking instruction were more likely to respond with nonscience terms not on the stimulus list. A variety of methods have now been developed to obtain raw data on cognitive structures (e.g., sorting words, rating similarity). Preece (1976) has demonstrated that the different methods lead to similar conclusions about the underlying cognitive structures. Increasingly sophisticated measures are also being developed to extract various properties of associative structures from the raw data. Measures of depth (e.g., number of levels), complexity (e.g., number of branches), and other properties of cognitive structure have been related to such educational factors as instruction, expertise, and age (e.g., Fisher, 1988; Nagy, 1984). A study by Naveh-Benjamin et aL (1986) illustrates the direction of this research. University students and their instructor sorted 16 terms (e.g., senile dementia, retrieval, integrity, intimacy) from a Psychology of Aging course into any order. Student sortings were done at the beginning, middle, or end of the course, and on each occasion sortings were done four times. The sortings were analyzed to produce measures for degree of organization (lack of randomness), similarity to instructor, and other aspects of cognitive structure. These derived measures were in turn correlated with class grades and showed expected changes from preto postinstruction. For example, associative structures became more complex (less random) and more similar to the associative structures of the instructor, and these effects were stronger for students who did well in the course than for students who did poorly. In their study, Naveh-Benjamin et al. (1986) represented the cognitive structures of students and instructors as tree diagrams in which higherlevel words are at the uppermost level, specific terms at the lowest level, and intervening category names between these extremes. Students and teachers often make similar outlines in the form of diagrams (see later discussion). This frequent use of spatial methods to represent relations among the verbal elements in cognitive structures (e.g., hierarchical trees, multidimensional scaling results) illustrates the intimate relation between verbal associative knowledge and imagery. Specifically, cognitive maps use spatial relations to represent the associative links among verbal representations in a nonverbal way (i.e., as a diagram). In an outline of an associative hierarchy, for example, lines between words represent class-inclusion relations and indicate that one element is a subordinate or superordinate of the other. These spatial representations for associative hierarchies are considered again when we examine memory and related educational phenomena. Dual Coding Theory and Education 165 The use of imagery to represent verbal associative knowledge demonstrates the strengths of DCT's emphasis on the collective and interactive effects of the verbal and nonverbal systems. The important interplay between words and images is also supported by the robust correlation between concreteness and categorizability that was discussed earlier (e.g., Kintsch, 1974; Toglia and Battig, 1978). Concrete, perceptual referents may somehow facilitate acquisition of superordinate categories and hierarchical structures. This section has reviewed educational and cognitive phenomena that are consistent with the DCT characterization of the structure of knowledge. In short, meaning and cognitive structure result from the separate and collective actions of the imagery and verbal associative systems. These same processes explain learning and memory effects of interest to educational researchers and teachers. LEARNING, MEMORY, AND STUDY SKILLS The successful transmission of new skills and knowledge depends on student learning and memory processes that have received much attention from educational and cognitive researchers. From a DCT perspective, learning and remembering involve the same imagery and associative processes discussed in the preceding section on the structure of knowledge. Indeed, comprehension and memory are investigated with similar tasks, and the boundary between the two processes is often fuzzy. For example, it is unclear how much time must pass between reading a passage and answering questions before the questions tap memory rather