Reading Research Quarterly 16, 1980-81, 261-315. Restored and revised July 2005

Design criteria for process models of reading

ROBERT DE BEAUGRANDE

With summaries in English, French, and Spanish

Since reading is a complex activity, many of whose participating processes are not observable in the laboratory; a crucial factor in research progress is theoretical model design. Unfortunately, design principles are often left implicit in discussions of the results obtained by individual approaches. This paper undertakes to clarify the issues involved by offering and applying a set of design criteria to characterize ten alternative models of reading or understanding: Chomsky, the Clarks, Gibson, Kintsch, Meyer, Frederiksen, Schank, Rumelhart, Woods, and the author’s own. The final section deals with some future research directions that seem to merit exploration.

Critères d’étude pour des modèles de procédé de lecture.

Étant donné que la lecture est une activité complexe et que beaucoup de ses procédés de participation ne sont pas observables en laboratoire, une étude de modèle théorique s’avère être un facteur crucial dans le progrès de recherche. MaTheureusement, les principes d’études sont souvent implicites dans les discussions de résultats obtenus par des approches individuelles. Ce compte rendu se propose de clarifier les débouchés impliquées dans la proposition et l’application d’un ensemble de critères d’étude pour caractériser dix modèles alternatifs de lecture ou de compréhension: Chomsky, les Clark, Gibson, Kintsch, Meyer, Frederiksen, Schank, Rumelhart, Woods, et celui de l’auteur. La dernière section traite des directions de recherche future qui semblent mériter d’être examinées.

Criterios de diseño para modelos de proceso de lectura

Debido a que el proceso de lectura es una actividad compleja, muchos de los procesos integrantes no siendo observables en el laboratorio — un factor crítico para el avance de investigación es el diseño del modelo teórico. Hay que lamentar el hecho de que las normas de diseño con frecuencia se dejan implícitas en la discusión de los resultados obtenidos por intentos individuales. Este ensayo procede a clarificar los problemas planteados, presentando y aplicando un conjunto de criterios de diseño para delinear diez alternativas de modelos de lectura o comprensión: Chomsky, los Clark, Gibson, Kintsch, Meyer, Frederiksen, Schank, Rumelhart, Woods y la propia del autor. La sección final discute algunas posibilidades de investigación en el futuro que parecen merecer exploración

*     *     *     *     *     *     *     *     *     *     *     *     *     *     *

Although they had long been considered self-evident, the notions of “theory” and “observation” in scientific enquiry have recently been opened to dispute. In the traditional view, a theory is a set of statements among which logical inference relations hold (Stegmüller, 1976, p. 2). The theory allows us to make exact predictions about the domain of application and to test them via direct observation of independently available facts. It has gradually been admitted that science does not and indeed cannot work in this fashion, as shown by considerations of three kinds: historical, philosophical, and psychological. Thomas Kuhn’s (1970) historical survey demonstrates that scientists have not usually rejected a theory because of falsifying evidence. Wolfgang Stegmüller’s (1976) philosophical survey marshals the arguments for supposing that all observation is “laden with theory,” so that “facts” cannot be independently available to impartial discovery. And new explorations in cognitive psychology compel us to concede that all forms of knowledge acquisition, even sensory apperception of the immediate environment, are heavily dependent upon our internal systems of belief — our “models of reality” (cf. survey in Beaugrande, 1980a).

Some alarming implications of this state of affairs have been hotly contested: (a) that scientific theories are immune to falsification; (b) that scientific progress is not cumulative, because each new theory makes us see an entirely new world; and thus (c) scientists have no rational grounds for preferring one theory over another, since theories are incommensurable. These implications cannot be argued away, but their most troubling aspects might be brought under control. The whole conception of “theory” and “observation” should be replaced by a conception I call the interaction of functional diversification with functional consensus (Beaugrande, Note 1). The scientist constructs a theory by designing a complicated conceptual apparatus to be applied to an envisioned domain of inquiry. The theory is progressively validated by applying it to steadily more numerous and diversified domains; the theory is “functional” for any domain when it satisfies our need for orientation and account to a reasonable degree. “Functional consensus” can be attained when the theory projects a unifying viewpoint across all of its domains of application. The problem .of circularity between theory and observation is not eliminated. But every time a new domain is modelled, the points where a scientist can inject theoretical pre-dispositions and the means for doing so are structurally and functionally altered. Hence, purely artefactual findings dictated by the theory are not usually transferable from one domain to another, so that theories can be weeded out after failing to combine diversification with consensus. In the words of William Estes (1979, p. 47), “the measure of success in moving toward scientific explanation is the degree to which a theory brings out relationships between otherwise distinct and independent clusters of phenomena.”

The acceptance of this approach would impel us to accord a more suitable degree of consideration to design criteria in scientific theories and models. The conviction carried by any theory or model no longer rests solely on the thorny question of direct, objective falsification; instead, scientists can be convinced by the potential for representing elaborate processes not accessible to non-theoretical discovery. Scientific progress can thus be cumulative even though radical shifts of empirical content occur, provided that the continuity of design manifests a favourable evolution. By the same token, scientists can transfer allegiance from one theory or model to another in recognition of its design improvements — even though available evidence is not conclusive, and the new theory may have fewer successful applications to its credit. In this fashion, the alarming implications of studies like Kuhn’s may be resolved.

I shall accordingly propose a set of design criteria that can serve to construct comparative profiles of process models for reading research. These criteria were selected not because they are logically unified in any axiomatic sense, but because they are operationally compelling for the model designer who strives toward a functional, diversified theory of reading. Some of the criteria have been expressly addressed in the literature, as my references suggest. Others were postulated on the same grounds as any theory (in the sense proposed above): they are useful in promoting functional consensus by allowing diverse domains to be characterized in a common descriptive idiom.

1. Criteria for Designing Process Models

Processor Contributions

One decisive aspect is the manner in which the processor-in this case, the understander reading the text-applies stored knowledge and prior expectations. In bottom-up models, these contributions are limited to the analysis of letters, words, phrases, or sentences, while in top-down models, the processor applies integrative hypotheses about the text, drawing freely from knowledge and experience of the world to constrain understanding and fill in materials

Memory Storage: Abstraction, Construction, or Reconstruction

If understanding is linked to memory, the role of previously stored materials and the integration of new ones must be clarified. Three main approaches have crystallized (cf. Royer, 1977). The abstractive approach maintains that a processor merely extracts “features” or “traces” from the presentation and stores them away; recall is done by reviving these traces of the original experience (cf. Gomulicki, 1956; Neisser, 1967). The constructive approach assumes that the processor begins integrating stored knowledge with the presentation right away, so that memory receives an expanded, modified version of the experience and presents this when recall is required (cf. Bransford, Barclay, & Franks, 1972; Ortony & Anderson, 1975). In the reconstructive approach, further contributions are still entering after the experience is stored in memory; recall is thus based on the current state of storage being assembled by means of a general organizational pattern (cf. Bartlett, 1932; Spiro, 1977).

Utilization

Also worthy of investigation is the extent to which a processor utilizes presented materials. At one extreme, models call for thorough, complete utilization: noticing, classifying, and dealing with every element on every language level. At the other extreme, models view the processor as using the presentation only occasionally to find cues and confirm predictions.

Automatization

Due to the speed and complexity of understanding, it is reasonable to suppose that at least some automatic operations are done  not conflicting with others and not requiring attention (cf. LaBerge & Samuels, 1974). Attentional operations may remain within the reach of conscious control and thus conflict with others. A further question is under what conditions control can be imposed upon normally automatic processes if occasion arises.

Decomposition

Some models show the processor decomposing text elements (words, meanings, etc.) into configurations of primitives: minimal building blocks not capable of further reduction. Between the two extremes-everything is always decomposed versus nothing ever is there are many positions holding that some decomposition is done for particular motives.

Processing Depth

Processing depth depends not on readers nor on texts, but, rather on the tasks assigned. It has be en proven that the amount stored in memory and recalled varies markedly along this dimension. Words are remembered better after disambiguating them than after judging their correctness of spelling (Bobrow & Bower, 1969); finding anomalous meanings aids recall better than watching for particular sounds or letters (Treisman & Tuxworth, 1974); fitting words into contexts elicits “deeper processing” than finding rhymes for them (Craik & Tulving, 1975); and making follow-up statements for a sample is more effective than judging whether the sample is meaningful as it stands (Mistler-Lachman, 1974).

Scale

Some models address elements only on a local scale, e.g., recognition of sounds, letters, words, and sentences. Others apply on a global scale, e.g., the formation of a gist or summary for whole texts. Most current models are mixed, looking at elements and operations on various scales. The question then arises of how to correlate the local and global factors.

Power

The term “power” is used by Minsky and Papert (1974, p. 59) for the capacity to apply general operations and typologies to a wide range of occurrences. The prime illustration of a high-powered model is the “general problem solver” developed by the Carnegie-Mellon group of cognitive scientists (cf. Ernst & Newell, 1969; Newell & Simon, 1972). The prime illustration of a low-powered approach is the stimulus-response model in which every activity is elicited by a specific occurrence from the environment.

Modularity versus Interaction

In a modular model there is little communication among the levels of language, (e.g., phonemes, graphemes, syntax/grammar; semantics, and pragmatics) or among operations that understanding traverses in real time. In an interactive model, there is steady interchange among levels and operations, each one helping to guide the processes in another. Modular models are easy to design and to test with straightforward experiments, but, in exchange, their operations are manifestly cumbersome (Winograd, 1975, p. 192). Operations in an isolated phase or level readily become explosive if they lack ways for eliminating alternatives. For example, many syntactically possible ambiguities are at once disallowed by the semantic level in actual communication. Interactive models are difficult to design, and render experiments complicated and imprecise, but in exchange, they profit from internal interchange to eliminate many needless dead-end pathways.

Serial versus Parallel Processing

A related (though not identical) distinction must be made between doing operations one after the other and doing them concurrently. Serial models have been popular because of their simplicity for making measurements; later, the serial computer added a suggestive note. More recently, the need has been increasingly recognized for parallel models of processing complex tasks performed in limited time. The fact that serial and parallel models are mathematically equivalent (cf. Townsend, 1974), though allowing us to postpone a final commitment, does not settle our questions about real human activities (Kintsch, Note 2, p. 37).

Freedom

Once a basic process model has been set down, how much freedom does it foresee among different readers under varying conditions? How can we deal with idiosyncratic or solipsistic readings of particular individuals? Can operations be re-ordered or modified on occasion, and can they break down? Such questions are crucial for a consistent treatment of empirical evidence.

Openness versus Closedness

A closed model is not amenable to representing any other kinds of operations than those it already encompasses (e.g., “autonomous syntax” in linguistic theory). An open model can be expanded as new insights accrue, without requiring a fundamentally new design (e.g., models based on general cybernetics). The implication often is that open models are to be construed as merely representative of understanding operations, rather than as exhaustive statements. Here, the design of the models themselves is often more revealing than the claims of the researchers involved.

Logical versus Procedural Adequacy

The “adequacy” of a model is its ability to operate consistently and free of obstacles or break-downs on particular tasks. Logical adequacy designates the ability to meet the demands of conventional logic~ (especially predicate calculus): formal proofs, derivations, quantification, identity, and the like. (See especially Anderson, 1976,’ for discussion.) Procedural adequacy requires that a model be a workable representation of the operatlons people might plausibly be performing when they use language (cf. Schank & Wilensky, 1977, p. l37). These two kinds of adequacy can dictate quite divergent priorities to a designer.

Learning

A learning model adapts its operations as the understander progresses through a task. Reading a text appears to be accompanied by various kinds of learning: refining one’s predictions and expectations, applying knowledge gained early toward the processing of later sections, becoming familiar with content and style; and so forth. But this factor naturally complicates model design still further.

Typology of Materials

While most researchers agree that understanding takes account of different types of materials (e.g., narrative versus expository prose), few have expounded conclusive measures for setting up such a typology. Accordingly, many models are not in fact built to adapt to such differences, or only to a few obvious ones.

Status of Programming

This criterion simply designates the status of the model as a functioning computer program, as opposed to being based solely on observation of humans. At one extreme, researchers want all models stated as programs (cf. Newell & Simon, 1972; Anderson’& Bower, 1973). At the other extreme, the computer is flatly rejected for emotional, ethical, or sceptical motives, or because the distinctions between computers and minds is deemed irreconcilable. (See discussion in McCorduck, 1979.)

2. Some illustrative models

In this section the criteria outlined in Section 1 are applied to a selection of process models of understanding and reading, each model loosely centred around its principal investigator(s). The selection is intended to be illustrative rather than complete, and to draw the attention of reading researchers toward some whose design criteria raise interesting issues. For example, the discussion of Chomsky and the Clarks suggests the conclusions that might be drawn generally about models where understanding is factored in terms of linguistic analysis (e.g., Holmes & Singer, 1961). The treatment of Gibson’s work might be emblematic for approaches focusing on the bottom-up recognition of surface features (e.g., Gough, 1972; Venezky & Massaro, 1979). Rumelhart’s recent inquiries are essentially comparable to many "story grammars" (e.g., Mandler & Johnson, 1977; Thorndyke, 1977, but compare Beaugrande 1983). I have also chosen models from a broad range, including for instance memory storage, world knowledge; inferencing, etc., rather than restrictive accounts centred, say, on word recognition.

The outcome of these explorations is assembled in Table 1. In cases where I felt uncertain about some criterion, I was usually able to consult the investigators themselves.¹ Sometimes the criteria were in fact not applicable (“N/A” in Table 1), because the model in question had never taken or implied a stand on the matter. Still, this mode of description may prove a useful complement to surveys that take each model on its own terms and claims (e.g., Samuels, 1979).

Click here for a rotated vertical view of Table 1

The Chomsky School

For a number of years, the "generative transformational grammar" expounded by Noam Chomsky was construed as a model of human language understanding, and occasionally even of reading (e.g., Smith, 1971). A "transformational grammar" is any language model that runs by converting structures into other structures through formal rules of arrangement. This basically neutral format was linked by the Chomskyans to a set of far-reaching assumptions about the human capacities.²

The processor contributions were subsumed in the standard model (Chomsky, 1965) under the notion of "competence" as the understander’s “tacit knowledge of the language”. The design of the model foresaw a central syntactic processing unit whose input or output was “interpreted” by subsidiary units for sound and meaning. The basic mechanism of understanding was the conversion of presented materials (“surface structure”) into basic patterns (“deep structure”). This mechanism called for total utilization, since every element had to be accounted for, and for total decomposition into units incapable of further reduction. The use of the notion of “abstract automata” by the Chomsky school suggested that all operations were deemed automatic, but shunned the task of spelling them out.

The design of the model also required working on a local scale, with serial operations being carried out by absolutely modular components. The model was closed against such domains as world-knowledge and human goals. Logical adequacy was exacted by the fact that the whole model was inspired by formal logic to begin with (Chomsky, 1964; on some implications of this bias, cf. Beaugrande, 1981). Procedural adequacy was quite poor (Woods, 1970). Power was low because every operation step had to be expressly done with corresponding “rules” incapable of generalization across wide ranges of tasks.

Since Chomsky (1965, pp. 3ff) explicitly excluded memory as “irrelevant,” no stand on storage was taken. Processing depth was not an issue either: The concept of “deep structure” was not related to this notion as expounded in Section 1. No provisions were made for variations in tasks, and all language users were grouped together as a “homogenous community.” That view also dispenses with freedom, and to a great degree, with learning. The processing of any sentence was uniform (handled as a random member of the “infinite set of sentences” that the model was claimed to describe). By the same token, no typology of materials was envisioned.

The Chomskyan model has fallen into discredit for various reasons though it has retain a group of die-hards who just keep tinkering with the design to keep empirical issues ay bay (Beaugrande 1998). The standard version could not be effectively programmed, though some improved versions were (e.g., Marcus, 1980). It proliferates alternative formattings to an alarming degree with no routine processing advantages from converting structures to other structures of the same type (Beaugrande: 1980a, pp. 33ff.). It discovers many ambiguities no reasonable human would be likely to consider (Riesbeck & Schank, 1978, p. 251). And, as already noted, it is closed to many factors that obviously play important roles in human communication.

Herbert and Eve Clark

Like the Chomsky school, the Clarks and their associates at Stanford also work on the supposition that understanding is in some ways comparable to the analysis done by professional linguists on single sentences. Their work is broadly eclectic, conforming to linguistic trends and to the exigencies of tightly controlled laboratory experiments.

The basic activities of understanding include parsing a sentence into “constituents” (noun phrase, verb phrase, etc.) and building underlying “propositions.” These propositions are very close to the surface sentences, though their status is rather confused. At one point, the reader is told that “each proposition consists of a verbal unit plus one or more nouns” (perhaps not a bad working definition of a sentence); later, propositions are written in bracketed notation with “functions” that “are expressed as verbs” or other parts of speech; still later the reader is shown “the proposition ‘Mary bought the book from John’, which is represented as ‘Buy (Mary, book, John)’“ (Clark & Clark, 1977, pp. 11, 46, and 114, respectively). Moreover, “propositions” inherit the “arrangement” of constituents in the sentence, and are “combined” by means of surface syntax, e.g., conjunctions and relative clauses (Clark & Clark, 1977, pp. 13ff.). It is therefore not at all clear whether “proposition” is occasionally just a vague name for “sentence” and thus a syntactic rather than semantic unit.

All the same, the processor contributions in the Clarks’ model are not limited to this managing of constituents and “propositions” in sentences. World-knowledge is admitted in many studies of how sentence organization is affected when describing visual scenes (Clark, Carpenter, & Just, 1973; Clark & Chase, 1974). “Inferencing” is admitted on a modest scale to keep track of the identity of referents (i.e., different ‘designations for the same thing) and of relations like instance-class, part-whole, and cause-effect (Clark, 1977). To keep inferencing in bounds, only those inferences are considered that the text producer has presumably “authorized” (Clark, 1978) — a standard explained by examples, not by a clear definition.

Herbert Clark (in a personal communication) describes his outlook on memory storage as agreeing with construction and reconstruction. But he has combated the “constructivist view” as “too inclusive” (Clark, 1978, p. 295). And l cannot see how provisions could be made for reconstruction on any scale. Instead, he seems to disallow the view when he states that “comprehension is conceived to be the process by which people arrive at the interpretation the speaker intended them to grasp for that utterance in that context” (Clark, 1978, p. 295, appealing to the philosophers H.Paul Grice, Stephen Schiffer, and Jonathan Bennett, none of whom relies on empirical evidence from real “speakers”). At most, a reader could “reconstruct” the author’s “intentions”-not the spirit in which the notion of “reconstruction” was introduced. Due to the strong concern for linguistic analysis, utilization is heavy in this model, being done serially on a local scale. Power is low, since, as we saw, sentences and propositions are so closely intertwined; the same factor of course encourages interaction of syntax and semantics. A strong stand was taken on decomposition (E. Clark, 1973), now softened considerably (Clark & Clark, 1977).

The categories of automatization, freedom, and logical and procedural adequacy have apparently not been addressed at all, and processing depth only in passing (Clark & Clark, 1977, pp. 151ff.). Learning might be possible when understanders adapt to distributions of “given” and “new information” (Clark & Haviland, 1977). No typology of materials is offered beyond the sentence. And none of the model has been programmed.

The Clarks’ model is comparatively versatile within the limitations of conventional linguistics and experimental psychology. They characteristically prove their assumptions by trying to show that some understanding operation (building internal representations of a sentence, resolving ambiguities, making inferences, etc.; cf. Clark & Clark, 1977) takes time. Accordingly, the model has been steadily accruing in an exemplary bottom-up fashion that leaves little room for doubting that such operations exist. The question is rather whether those operations are sufficient to capture understanding.

Eleanor J. Gibson

Eleanor Gibson’s approach to reading has been decisively influenced by the work of James J. Gibson, who flatly denies that human processors contribute anything to what they understand (J. Gibson, 1966). The only operations allowed are those of detecting, extracting, and utilizing the “features” within the presented materials. A model of “word perception” was accordingly set forth as a series of independent phases of feature extraction in this order: phonemic/ graphemic, then syntactic, and finally, semantic (E. Gibson, 1971). Later, an elaborate model of reading was proposed along comparable lines (E. Gibson. & Levin, 1975).

The abrogation of processor contributions has important consequences for the model’s design. Utilization is heavy (especially for skilled readers) and, as we saw, modular. Decomposition Is advanced to a status of marked prominence,³ though it is conceded that words are not broken down into individual letters (Gibson & Levin, 1977, p. l64) — an extreme view espoused on occasion by Philip Gough (1972). Scale is naturally local, and power quite low. Significantly, most of the Gibson-Levin volume is devoted to sounds, letters, spelling, grammar, and syntax. Meaning figures as little more than “adaptive significance” (Gibson & Levin, 1975, p. 20) that is “learned as part of a situational context in which it is invariant with an event of interest” (E. Gibson, 1977, p. l68).

Provisional conclusions might be drawn about criteria l did not find being treated. Automatization figures only in adult readers (Gibson & Levin, 1975, p. 9), though the feature-extracting operations have an automatic quality in general. Processing depth might be accounted for as a truncation of processing before the semantic stage causing a detriment to recall. Memory storage would be abstractive. (The Gibson-Levin volume has little to say about memory, despite frequent recourse to “learning” and “storing” features.) The illustrations suggest a strong leaning toward serial processing, both in the order of phases and the stressing of serial media (letters, word components, phrases, etc.). Learning is accounted for as skill acquisition through experience and transfer (Gibson & Levin, 1975, pp. 63ff.). A typology of materials could be implied as a grouping of feature repertories.

Other criteria seem still more doubtful. Logical and procedural adequacy are both equally remote. None of the model is programmed, and the computer is viewed with mistrust (Gibson, 1977).

Freedom is accounted for only in the different developmental stages of children. The model is open toward further operations of feature extraction, but emphatically closed toward any information-processing operations (Gibson, 1977).

The Gibsonian denial of processor contributions is undoubtedly a minority view at present, regarded with discomfort even by cordially disposed professional colleagues (e.g., Neisser, 1976). In emphasizing the role of contrastive features and differentiation, the Gibsonian model is left with few methods for depicting the connection, unification, and integration that readers must perform; a simple assembly of features into steadily larger units would most likely be explosive owing to a lack of hypotheses about what kind of units should be built in a given case.

Walter Kintsch

We turn now from predominantly bottom-up models to consider some models more favourably inclined toward top-down operations. Among the best-known is that developed by Walter Kintsch and co-workers at the ‘University of Colorado. His approach reflects his extensive expertise on theories of learning, coding, and memory. (See Kintsch, 1977a.) By working with texts rather than sentences, he has gradually moved away from the classical procedures of experimental psychology and has applied elaborate techniques of measurement (described in Miller & Kintsch, 1980).

For a long time, Kintsch’s model did not begin with the presented text at all. A hand-coded proposition list was prepared as the basis for study. (Techniques are described in Turner & Greene, 1977.) Recently, however, he has depicted the treatment of the text in terms of segmenting into “chunks” according to the number of underlying “propositions,” Le., predicate plus one or more arguments (Miller & Kintsch, 1980). Further processing is done by a. “coherence graph generator” that connects propositions into a network, a “fact organizer” that matches up input to stored “world knowledge,” an “inferencer” that “fills in missing propositions,” and a “control structure” guiding the “macro-operators” that produce the “gist” of the text (Kintsch, Note 2, p. 13). In particular, “schemas” are applied to give an abstract outline of a story or event sequence (cf. Kintsch, 1977a, pp. 373ff.; 1977b; Kintsch & van Dijk, 1978)..

We can see that utilization is heavy, and that memory storage will be likely to be both constructive and reconstructive. Propositions are studied both on local and global scales, e.g., when one proposition takes others as its arguments, leading to a hierarchical network (cf. Kintsch & Vipond, 1979). Power is high as a result of the general operation types, such as “deleting” (removing), “generalizing” (subsuming), and “constructing” (restating) of materials during the formation of a steadily more global representation (cf. Kintsch, 1977b; Kintsch & van Dijk, 1978). The model runs in parallel, with emphasis on procedural adequacy. Learning is manifested in the cumulative operations that lead to the gist.

Rather surprisingly, Kintsch (Note 2, p. 46) describes his own model as modular. This designation is based on the need to test only one part of it in a given experiment. In the domain of theory, the model is decisively integrative: How, for instance, could human readers separate their knowledge of “facts” from their “inferencing”? Modularity is also needed because only some parts of the model have been programmed: the “coherence graph generator” (running) and the ‘‘fact organizer” (in progress) — no parser or inferencer is available.

Kintsch agrees that factors such as automatization, processing depth, freedom, and a typology of materials are vital considerations, but he has not yet tried to account for them in his own work. Regarding decomposition, he argues that it can be done if required, but is not carried out routinely (Kintsch, 1974). He consistently notes that his model is open. It would not be hard to add on a parser, for example.

The appeal of Kintsch’s model lies especially in its unusual breadth that correlates functional diversification with functional consensus (Anderson, 1976, p. 55). His handling of semantics — an important weak point in the models examined previousl — is elegant, powerful, and humanly reasonable. Unlike the Clarks, Kintsch is willing to disregard effects of surface syntax to some degree, and classifies sentences only on the basis of their number of underlying propositions. l do not agree that his experiments “are not tests of strict deductions of the theory,” but only “studies in search of a theory” (Kintsch, 1974, p. 243). As argued at the outset of this paper, the notion of “theory” should be modified. It might be argued that different validation procedures are required in this kind of research; clinging to the old paradigm of statistical predictions followed by direct observation would make us discard some of our best data and insights.

Bonnie June Francis Meyer

The work of Bonnie Meyer and associates at the Arizona State University takes its distinctive cast from the positions adopted in regard to processor contributions, scale, and power. She postulates that each text possesses an inherent “structure of content” yielding an ordered hierarchy of importance. Reading operates most efficiently if this hierarchy is discovered and utilized. She began by constructing intuitive hierarchies for sample science texts and asking independent judges to rate the importance of the constituent propositions; there was fairly high agreement among the judges about relative importance, and the higher-ranking propositions by and large were best and most often recalled by diverse groups of readers (Meyer & McConkie, 1973).

The act of reading is envisioned along the lines of Joseph Grimes’s (1975) discourse analysis, using a “case grammar” of verbs and verb adjuncts to handle the surface text. The words of the text (at least the verbs and their adjuncts) are taken as “lexical predicates” whose “arguments” are “ideas from the content of the text”; these predicate-argument combinations are “lexical propositions” that in turn become arguments of “rhetorical propositions” (cf. Meyer, 1975, pp. 26ff.; Meyer, 1977, p. 317). By making those “rhetorical propositions” into arguments of more and more general propositions, a hierarchy of importance is eventually obtained.

In consequence of having changed from an abstractive outlook (e.g., in Meyer & McConkie, 1973) to a constructive or reconstructive one (Meyer, personal communication), she is evasive about whether the “top-level structures” (which she equates with “schemas” in the sense of Richard Anderson, 1977)5 are situated in the text or in the reader. Some processor contributions of readers are doubtless hidden away in the “importance judgements” elicited regarding passage structure. Meyer admits that a text may follow the “top-level structure” in a “normal” or “distorted” way, as has in fact been shown by Perry Thorndyke 1977), so that readers must be able to make such distinctions. For example, having the main ideas at the beginning of a text is considered “normal” (Meyer & Freedle, 1979, p. 3). Meyer’s solution is to appeal to the schema used by the author: Readers who fail to use the same schema (for instance, because they reject the author’s arguments, or are simply unskilled) will not perform as well (Meyer, 1980). Meyer’s model is apparently most valid for readers who follow the author’s guidance. Her group surmounts the technical difficulties (knowing what the author intended) by authoring their samples themselves around various “top-level structures.”

In agreement with this overall framework, utilization is viewed as medium heavy, and scale and power are mixed as needed to build a hierarchy. Interaction is possible at least between the “lexical” and “rhetorical” components, and parallel processing is implied (though not clearly demonstrated). Decomposition is not postulated.

Meyer’s undecided stance regarding the relationship of texts and readers makes some of the model criteria difficult to apply: Automatization, freedom, and adequacy have either not been addressed or are onIy now coming under investigation. In exchange, she places great emphasis on a typology of materials designed to reflect “top-level structures,” e.g., “adversative” (comparing a favoured view to an opposing one), “covariance” (relating preconditions to their outcomes), “response” (stating a problem and offering a solution), “attribution” (depicting the traits of an object or event), and so forth. She suggests on occasion that these different types may elicit variations in processing depth (Meyer & Freedle, 1979, p. 19). Learning is also acknowledged to be a factor of text organization: she indicates that “recall is facilitated when passages [of a text read progressively] have the same structure but different content, while it is inhibited if passages have different structures but the same content” (Meyer, 1977, p. 3l0). She suggests that traditional learning experiments on “serial position,” “primacy,” and “recency,” or “proactive and retroactive inhibition”-in effect, whether recall depends on the order of presentation-were inconclusive because of the failure to consider hierarchical structure of content (Meyer, 1977, pp. 308ff.).

Meyer’s model is the last one considered here which has not been programmed at all. In her case, as in that of the Clarks and Gibson, the possibility exists that creating a program would force a clarification of issues now left indistinct (cf. pp. 46ff.). Still, Meyer’s model is open to admitting more reader activities, and empirical results are impressive and intriguing. Her exploration of global organization was clearly a pioneering effort at a time when few other researchers had realized the importance of this factor. 

Carl Frederiksen

Like Meyer, Carl Frederiksen (now at McGill University) and his associates are concerned with the evolution of propositions from a presented text. But Frederiksen’s model differs from Meyer’s in significant ways, notably in the development of more detailed typologies of materials. He defines “propositions” as “semantic relations” that either “identify objects or actions” with respect to space, time, attributes, parts, and the like, or “specify the elements of an event”: the “case roles” such as agent, object, patient, or the like (Frederiksen, Frederiksen, Humphrey, & Ottesen, Note 3, p. 4). He sees propositions as moving upward in a scheme of “ranks,” including “event frames” (whatever participates in an event), “relative systems” (propositions connected by a function), and “dependency systems” (propositions connected by logical, causal, or conditional relationships) (Frederiksen, 1977, p. 65).

The act of reading is represented as follows: The reader approaches a text segment (usually, but not obligatorily, a sentence), finds the propositions, and assigns them to classes. There follows a “first stage” of “inferences” that resolves questions of reference (e.g., pronouns, substitutions) and ambiguity. Then more elaborate inferences are made in three basic categories: “Connective” inferences “relate a current proposition to propositions derived from the prior text”; “extensive” inferences “extend current propositions by generating new propositions based on prior knowledge and discourse context”; and “structural” inferences “segment and organize a text, building a coherent model of the text as a whole” (Frederiksen et al., Note 3, pp. 5ff.).

Utilization of the presented text materials is variable according to the presence of “discourse features that require or elicit inferences,” such as conjunctions, deictics, pronouns, and so forth.

It is obvious that processor contributions are heavy, since most of the organization that Meyer sees as inherent to a text is done here by the reader’s inferencing operations. This approach lends Frederiksen’s model an unusual flexibility in regard to bottom-up or top-down processing. Indeed, the find of “schema” or “top-level structure” that applies here constitutes itself directly from the interaction of text features and reader operations. Such a view is opposed to reading as a matching of static patterns with fixed slot configurations. Memory ~storage is naturally constructive and reconstructive...

Scale is mixed according to the scheme of “ranks” (a notion of Michael Halliday). Considerable power resides in the notion of a “system” as a functioning repertory of options, such that any structure is viewed as a series of decisions (Fredericksen, 1977, p. 59). Frederiksen also relies on high-powered typologies capable of subsuming the details of semantic units and relations on all levels. And he now recognizes a still higher-powered reader awareness of one’s own procedures, allowing for modification and adaptation during operation.

The model’s heavily interactive design renders it open to treating far more issues than it now encompasses. Procedural adequacy dominates, but does not exclude logical.⁶ Learning is called for constantly when readers consult previous propositions in understanding and organizing subsequent ones (Frederiksen, 1977, p. 70). Freedom is treated by comparing recalls of children at different ages in order to measure prior knowledge and reading practice (Frederiksen et al., Note 3).

To avoid committing his model to a fixed set-up of compulsory procedures, Frederiksen (in a personal communication) takes no stand on automatization, processing depth, or serial vs. parallel processing. He would admit decomposition as a possibility, but his model does not use it.

The programmed version of the model is far from completed. It segments the text, but queries a human user about presence and classification of propositions for each segment. A “text file” is built from these propositions, and an “inferential review” notices “common elements” and creates a coherent network. Some samples are reprinted in Frederiksen (1977).

Like Kintsch and Meyer, Frederiksen is actively engaged in reading education. He uses “natural experiments” in which children’s reading is studied in “a wide variety of naturally occurring discourse situations” (Frederiksen et al., Note 3, p. l0). His samples are actual school texts children normally read. Such research is an important counterbalance to the restrictive, artificial laboratory tasks with isolated sentences.

Roger Schank

Unlike the models discussed hitherto, the approach of Roger Schank and co-workers (now at Yale) was devoted, from its earliest inception, to building functioning computer understanders. Openly repudiating the low-powered, bottom-up models of conventional linguistics and psychology, they designed understanders that reason extensively about the organization of actions and events in order to read stories and answer questions about or summarize what was read. They also rejected the view that memory is “semantic” — that is, a neat dictionary of classifications — and argued in favour of “episodic memory” for systemizing one’s own personal experiences from day to day (Schank, 1975a).

The first large-scale program was called MARGIE (Memory Analysis and Response Generation in English) (Schank, Goldman, Rieger, & Riesbeck, 1975). The understander would start reading words, recovering underlying concepts, drawing inferences, and making predictions (“requests”) about what should be expected — all this before the end of a sentence was reached (Riesbeck, 1975). The mainstay was the stored knowledge about what constitutes actions and events: how they update the state of the world, and what participants (“conceptual cases” such as “objective, directive, recipient, or instrument”) should be anticipated and used to fill slots (cf. Schank, 1975b, pp. 30ff.). Rich inferencing was done right away about causes, effects, intentions, and motivations (Rieger, 1975).

In more recent programs such as SAM (Script Applier Mechanism), the processor contributions were upgraded in scale and power to the applying of “scripts” as global sequences of routine actions commonly done in human affairs (e.g., a visit to a restaurant) (Schank & Abelson, 1977; Cullingford, 1978) In PAM (Plan Applier Mechanism), another technique was to reason about the goals of story characters in general (Wilensky, 1980). These approaches lent more direction and purpose to the activities of understanding, and reduced the volume of local inferencing to be done.

Schank’s approach is so heavily constructive that the utilization of the actual text is quite light. Riesbeck’s (1975) MARGIE parser did no syntactic analysis except what was useable right away for conceptual understanding. In a new version called IPP (Integrated Partial Parser), the parser reads only some words, skips others and saves them, or else skips them altogether (Schank, Lebowitz, & Bimbaum, 1978). Words are judged “interesting” if they generate expectations troublesome issue for models like Gibson (1971) and Gough (1972).

The recognized words are fed into a “parser” containing knowledge about “case frames,” i.e., patterns of units associated with a word (usually a verb) (Rumelhart & Levin, 1975, pp. 171ff.). The parser builds a format called an “augmented transition network” (cf. Woods, 1970; Kaplan, 1975) (shown later in Figures 1 through 7). This network has states (nodes) representing words, and arcs (links) representing syntactic relationships. The parser invokes the “interpreter” that gradually assembles a semantic network composed of basic primitives like “do”, “cause”, “change’’, and so on (cf. Rumelhart & Levin, 1975). Each event or action has a “schema” of these primitive components to organize incoming knowledge (Norman & Rumelhart, 1975, pp. 406ff; Rumelhart, 1977b, p. 266).

Like Schank’s group, Rumelhart and his associates felt a need for some higher-powered, more global way of guiding understanding. The notion of “schema” was stipulated to represent the general line of a story with “settings” (statement of time, place, and characters of the story) and “episodes” (event plus the protagonist’s reaction) (Rumelhart, 1975). Soon, this definition was expanded to incorporate problem-solving procedures (Rumelhart, 1977b; Rumelhart & Ortony, 1977). Now, the story line is driven by the protagonist’s “trying” to attain a “goal” set up in response to some “event.” The story outline is thus a hierarchy of episodes (event + response of setting up goal + try for the goal + outcome) that cause or contain each other. It was shown that readers making summaries used the materials corresponding to the upper levels of this hierarchy (Rumelhart, 1977b ) — a finding Meyer obtained from her own standpoint. (See above.)

The processor contributions in Rumelhart’s model are therefore stored patterns often termed “schemas.”⁸ This approach leads to a constructive and reconstructive view of memory storage (Norman & Rumelhart, 1975a, p. 23). Such factors as scale and serial vs. parallel processing hinge upon the size of schemas and their order of application (on parallel use of multi-scale independent “knowledge sources” via a “message center,” cf. Rumelhart, 1977c). Schemas may work for both global and local input, and be applied concurrently; with or without knowledge of each other. The strong stand on decomposition taken in the 1975 volume has been attenuated: Decomposition would be done only if the components of a schema being matched were smaller than the input units. Much of the schema-matching is believed to be done automatically.9

As was remarked in connection with Schank’s work (p. 280), processing depth is humanly plausible, but hard to capture in a processing model. Rumelhart and Levin (1975, pp. 203ff.) suggest a “variable depth strategy” which “allows information to be stored at different levels in the decomposition process.” They envision “partial comprehension” that leaves ambiguities unresolved in hopes of finding out more later on. Another approach would be to allow for differences in utilization according to the reader’s motivation to account for the input with more or less completeness and thoroughness: Here, the matching of schemas would simply be less demanding. Freedom could easily enter on a local scale when different readers are all working for only a superficial or general account; on the global scale, however, he found that readers’ hypotheses tend to converge as soon as a story becomes reasonably unambiguous (Rumelhart, Note 4).

The appeal to abstract pattern-matching operations in all areas of understanding lends Rumelhart’s model substantial power: At some degree, all processes can be viewed as essentially the same. In such a framework, interaction is easy to introduce (Rumelhart & Norman, 1975b, p. l59; cf. Rumelhart, 1977c); and the model is open to many extensions. Once attained in one domain, procedural adequacy can readily be generalized to others.

Gibson and Levin (1975, p. 40l) misunderstand the notion of “schema” as a denial that readers can “extract any new information from what we read.” Rumelhart has pointed out that learning is always possible: The abstract outlines of schemas can be filled or combined in totally unique new ways. Moreover, schemas .can evolve as they are “tuned” to new experiences (e.g., becoming more general or specific, or refining their range of expected variables);new schemas can be induced from experience or patterned after old ones, and even old schelpas do not preclude memory traces of particular occurrences. (Discussion is given in Rumelhart, 1980.)

As of 1980 present, Rumelhart’s model was far from running as a complete computer program. For one thing, the model lacks access to sufficient machine memory. The surface text is hand-parsed into a case-like representation. And only a few programs have been written to apply story schemas to simple examples.

Rumelhart’s approach, like that of the groups around Kintsch and Schank, is commendably broad. But unlike those groups, his concerns extend to the most basic mechanisms of recognition and storage.¹⁰ Reading is therefore seen to be situated within the domain of human information processing at large. It is greatly to be hoped he will eventually provide an integrated, start-to-finish version of his model.

William Woods

Though designed to run entirely on the computer, the understanding models developed at Bolt, Beranek, and Newman (not by accident in Cambridge, MA) under the supervision of William Woods follow quite a different line of reasoning from those in Schank’s group. Woods (1970) was among the first to recognize that the augmented transition network (see Figure 1-7 below) was superior to the dominant transformational model for parsing English syntax. Lately, he has been working to generalize this format for all the subsystems of language understanding in parallel: “one for acoustic phonetic recognition, one for lexical retrieval (word recogni­tion), one for syntax, one for semantics, and one for subsequent discourse tracking” (Woods, 1978b, p. 35). Each subsystem should be able to consult the findings of the others about ongoing structures, and thus to drastically reduce the number of alternatives to be considered. In fact, operations that would yield structurally comparable configurations in different subsystems (or along different paths of the same system) are merged on the spot and thus done only once-a scheme Woods (1978b) calls “cascading networks.”

According to the most recent version I could obtain (Brachman, 1979a), the reading process would run as follows. A user types English sentences into the computer to be syntactically analysed with a parser designed by Rusty Bobrow (1978). This parser works through the premodifiers in a noun phrase or verb phrase until the head is found. The results are sent to the semantic stage in the reverse order (head first). Then any post-modifiers are picked up until the phrase is completed. Finally, specifications are sent along: determiner and number for a noun; or tense, aspect, and (if present) negation for a verb. All the word classes are identified with the help of a syntactic taxonomy (list of types). By consulting the semantic stage early during syntactic processing, the understander avoids setting up syntactically acceptable, but semantically anomalous readings.

The semantic stage applies concepts with an intricate internal organization designed by Ronald Brachman (1978). Its salient traits include specifying the “roles” of a concept (e.g., parts of an object or “participants in an event) — a kindred approach to Rumelhart’s “cases” and Schank’s “conceptual dependencies” — and an elaborate capacity for “inheritance” of knowledge between a class and an instance (e.g., human beings/Napoleon) or between a superclass and a subclass (e.g., animals/human beings) (cf. also Fahlman, 1979). These traits lend the semantics a remarkable power combined with a compact, easily searchable memory store.¹¹

The outcome of the semantic stage is a “literal” interpretation of the meaning of sentences in a structured network. The “discourse expert” then takes care of pronouns (anaphora), quantification, and speech-act recognition. This last operation (worked out for computers especially by James Allen, 1979) is done by hypothesizing a plan that the user should be following — or even by “inferring what the user wants the system to think is his or her plan” (Brachman, 1979a, p. 20).

Processor contributions in this model are not just prior knowledge: More important are skilled techniques for organizing operations on the spot. Consequently, the model attains unrivalled power, interaction, and learning capabilities. The grammar of procedures continually refines itself the longer it is used, and the specific tasks at any moment refine themselves still further (Woods, 1978, p.22). Thus, although the model is open, it may not require any more major theoretical additions. Woods (1978, pp. 63-70) shows how it can be generalized beyond -the domain of language understanding to encompass such areas as visual scene analysis (e.g., light intensity or. hue) and acoustic frequency recognition. Noting is needed besides ways to “structure” and “index” a “perceptual domain” with “operations” that “probe transitions” to identify “states” and’ “constituents” (especially “initial” and “final” states in a unit), and to keep records of what is hypothesized and found.

 In such a bigh-powered approach, handling scale and procedural adequacy posed no separate difficulties. Utilization is somewhat heavy, but never wasteful. Decomposition is much less important than assembly. Parallel operation is built in with unmatched elegance. In return, the design implies no decisions regarding automatization, processing depth, freedom, or typology of materials. Plainly, such an elaborate program must, for the present anyway, rely on uniform treatment of readers and their memories. Only a restricted amount of the English language can be dealt with so far.

Unfortunately, Woods’ understander is known chiefly to specialists, and has influenced just a few reading models (e.g., Rumelhart’s and mine). Its merits emphatically recommend it to a wider audience, particularly its extraordinary power and economy. Such a model is manifestly able to encompass a far wider domain than it has been managing to date, and suggests how a host of issues might be handled that have scarcely yet been recognized by many researchers.

Robert de Beaugrande

My own search for a process model began in 1974, when l undertook an account of the operations involved in translating poetry. Although an outline model was eventually propounded (Beaugrande, 1978a) in terms of reader aesthetics, many of the issues required a more detailed and general theory of text production and comprehension, stated within the framework of cognitive science rather than old-style linguistics.

l have now assembled far more complex outline models for texts and text processing (Beaugrande 1980a, 1984, 1997). Unusual emphasis has been consistently devoted to issues of integrative design: Reading is seen as an interaction of phases of processing dominance, i.e., as a correlation of processing types sharing the processor’s cognitive resources in varying distributions. The phases are: parsing (identifying the grammatical dependencies of the surface text), concept recovery (associating language expressions with cognitive content), idea recovery (building the central conceptual configuration that organises content) and plan recovery (identifying the plans and goals that the text is intended to pursue) (cf. also Beaugrande & Dressler,.1980, ch. 3). Any dominant phase freely consults the results of non-dominant ones, so that grammar is continually correlated with meaning, meaning with action planning, and so on. l do not see these phases receiving dominance in a neatly fixed sequence; instead, dominance is probably passed back and forth frequently, as particular aspects enter the focus of processing, e.g., because results seem unsatisfying.

To support functional diversification and consensus, l have argued that the main activity is connectivity search: Given any textual occurrence (be it a letter, word, clause, concept, discourse action, or whatever), how does it relate to other occurrences in that subsystem or in other subsystems? The duration, intensity, and thoroughness of this search are always dependent on the reader’s demands and motivation, and the relevance of the text to current tasks and goals. The model should accordingly address some idealized degree of processing while being adaptable to a range of approximations among individual readers. Here, the ideal representation is clearly the network, not the tree, the lists, the set, or the formula: A network can retain connectivity even if some elements are left unlabeled or omitted (e.g., skipped in reading or forgotten afterward) — precisely what readers seem to do quite often (cf. Masson, 1979).

The initial processing unite is not the sentence (after all, many texts are not composed exclusively of sentences, especially in speech), but rather the stretch of text that can be comfortably held in working memory under current limitations of attention, familiarity, and interest. It would be important to learn whether that stretch is or is not usually a sentence, but we must not assume so in advance. Other candidates deserve consideration: a clause, a group of short sentences, a long noun phrase or verb phrase, a chunk of propositions, a discourse action, and so on. The goal of processing cannot be syntactic analysis, but rather building a model of a textual world. By this l mean that a reader reconstitutes a more or less extensive “world” containing the situations and events depicted by a text along with various amounts of prior knowledge and assumptions that might reasonably apply to such a world. Like any “model,” the text-world model of a particular reader may have varying degrees of correspondence to whatever is postulated as complete or ideal understanding of the text. Yet all readers’ models will plausibly share some “family resemblance” with each other and with the text producer’s own model. In this perspective, criteria for model design enter both into the theory of reading and into the ‘actual object of investigation.

Imagine someone reading the following passage from a children’s story:¹²

 (l) A great black and yellow rocket stood in a desert. Empty, it weighed five tons. For fuel it carried eight tons of alcohol and liquid oxygen.

(2) Everything was ready. Scientists and generals withdrew to some distance and crouched behind earth mounds. Two red flares rose as a signal to fire the rocket.

 The ideal first step is to recognize the words of the surface text and parse them into a structure of grammatical dependencies (a structure with at least two grammatical elements, one of which cannot stand alone). That task can be represented with an “augmented transition network” that moves through a sequence trying to predict and identify the elements being encountered. Presumably, the phrase is the usual unit, since it provides a fairly well-defined pattern in English. The opening phrase “a great black and yellow rocket” might be parsed as shown in Figure 1.

Figure 1. An augmented transition network for a noun-phrase macro-state

On encountering the determiner “a,” the processor enters a macro-state of “noun phrase” and keeps predicting the “head”; when the head is not found. tbe next highest search priority is for a "modifier."¹³ This sequence is represented by dotted lines for failed predictions and solid lines for the confirmed ones. The lines at the top of the figure are for the grammatical dependencies such as "modifier-to-head," giving us the labeled links for a dependency network.

A comparable network could be set up for the verb phrase "stood in a desert." Eventually, the grammatical dependency network shown in Figure 2 could be obtained by discarding the left-to-right chaining and keeping only relational links.

Fig. 2. A grammatical dependency network

The arrows are pointers to the nodes designated by the abbreviated labels. This much seems straightforward enough, but, in fact, a number of disturbing questions are hidden away here. Does a reader build a network for a whole sentence without consulting the meaning? Or is meaning consulted after each phrase (as is apparently called for in the Norman & Rumelhart version)? Or does each dependency correspond to a list of candidate relations in the conceptual domain? Could a reader bypass this kind of analysis and go straight to the meaning? Or are all of these methods possible?

My own tentative conclusion is that all of these methods are possible, but, according to the circumstances, not equally plausible nor workable. Readers probably have a set of preferences for correlating grammatical relationships with conceptual ones (cf. Beaugrande, 1980a, p. 88), and the organization of phrases or clauses could easily be consulted in modelling the organization of situations or events. But interactions like these are asymmetrical, i.e., not usually one-to-one and hence never fully reliable (cf. p. 295); still, we customarily have constraints imposed by one domain upon another.

To build a text-world model, the reader would set up configurations of concepts connected by relations, e.g., "agent-of," "location-of," "attribute-of," etc. (see more extensive listing in Beaugrande, 1980a, pp. 82ff.) Here, the reader can use networks whose nodes are not surface words, but rather concepts for which the words we see are only names, that is, instructions to activate knowledge in memory. For the material underlying the opening sentence, the reader would have the text-world fragment shown in Figure 3.¹⁴ The ensuing processing would steadily expand this configuration by adding onto already created nodes where feasible, in this case, mostly to the "rocket" node. Thus, the material underlying the whole first paragraph would yield a model space shown in Figure 4.

The dense linkage around the "rocket" designates it as the topic concept: an operational control center for storing, using, and recalling the text's content.

Thus far, the text-world model might appear to contain no concepts not explicitly activated by the text. However, such a restriction could not be upheld very long. The second paragraph opens with material whose relationship to the foregoing is not obvious: what is subsumed by “everything,” and how do “scientists” and “generals” enter the scene? Two solutions are possible: to set up fragments and wait for some later account of relationships, or to use inferences for connecting the material as it comes along. Let us assume that the latter solution is more plausible. Two reasonable inferences can be entered (using square brackets in the nodes involved) that “everything” includes whatever enables the rocket’s “take-off,” and that the “scientists” and “generals” were there to “observe” the rocket.¹⁵ If linkage is added to the effect that the “ready” state was the reason for the “withdrawing,” “crouching,” and “signaling,” the integrated configuration shown in Figure 5 is obtained.

Figure 5. Two merged model spaces with inference nodes

Proceeding in this fashion, a network for an entire text can be modelled (see illustrations in Beaugrande 1980a, 1980b), but this design should suffice to illustrate the advantages and disadvantages of the approach. A main advantage is that an ideal pattern of concepts and relations is obtained against which the patterns of readers’ reports of the content can be matched. For example, one reader began a protocol like this:

In a desert, a rocket waited to be launched: it weighed five tons empty. The generals and technicians stepped back behind dirt mounds and launched two red flares signalling the launch of the rocket.

An account of this report in terms of surface texts or single propositions would be quite intricate, and not very general, But a matchable pattern with a few simple conventions can be designed: Put recovered nodes in the positions corresponding to the original model, mark alternative concept names with “*”, set inferred nodes in square brackets, and the outcome is represented in Figure 6.

Fig. 6 Section of a reader’s recall protocol

Continuing in this fashion for a set of reader reports, l can get some graphic impression of the typical ways that readers make sense out of whatever they understand and remember when connecting it into a coherent configuration. For instance, it appeared that readers tended to lose whole spaces, rather than isolated nodes and links (in this sample, the material regarding the fuels and “everything” being “ready”); that readers tend to fill in links among the concepts they do retain (in this sample, the launching” of ‘‘flares’’ was assigned to “generals and technicians”); and that, by and large, readers are more concerned with maintaining coherence than with reproducing faithfully what has been read.

Another technique l have used is to design for the text-world model a matching “world-knowledge correlate” containing facts readers would be likely to know before encountering the text: that rockets takeoff and bum fuel, that burning causes heat and thus makes people seek shelter, that generals use rockets to attack, while scientists use them to explore outer space, and so forth. A correlate for the sample might look like Figure 7.

Fig. 7. A section of a world-knowledge correlate

Again, nodes not present in the text are identified with square brackets. And a distinction is made between determinate relations (marked with a Greek delta) for necessary components, and merely typical relations (marked with a Greek tau) for usual, but not indispensable components of commonsense facts.¹⁶

Knowledge of these particular facts is no doubt influential, but it does not guide processing with any- major scale or power. Included is a schema for “flight” as a sequence of events: “take off,” “ascend,” “reach a peak,” “descend,” and “land” are the indispensable components shown in Figure 8.

Figure 8. A schema for flight

Suppose that upon encountering a “rocket,” readers activate the “flight” -schema and predict that the schema events will be mentioned. Notice, however, that the opening text fragment does not in fact mention any of these events, but rather describes the scene and some incidental preparations. When subjects reported what they read, they very often moved some mention of the “take-of’ event right up to the beginning of their protocols: Such was done in the first sentence of 13 protocols within one group of 42 readers. This kind of processing is anticipated by inserting that inference node for “take-off’ in Figure 5. It is hypothesized that the events of the schema are regularly inserted into a text-world model as control centers for integrating and organizing what the text actually says. (This graphic is given in Beaugrande, 1980a, p.

The disadvantages of the representation are not minor. Numerous tests showed that readers tend to visualize whole scenes: The setting was specified as “sandy plains” “under a bright sun,” where a rocket was on its “launching pad.” If slides of rocket launchings were shown while the text was read aloud, these scenes were influenced; yet the presented network fails to capture this “visuality” of discourse, which was not integrated in terms of the mainframe of text study until my New Introduction (Beaugrande 2004).

Another drawback is that the words of the text are retained as concept names, so that the underlying meanings (e.g., of “black” or “fuel”) are not specified. If primitives were used (e.g., those proposed by Schank and Lehnert), the configuration would look quite different, especially the links. (For instance, the connection between “fuel” add “fire” or “fire” and “flares” would become obvious.) Surely some portion of coherence is attributable to relationships among the basic components of knowledge (cf. Gentner, 1978).

A third disadvantage is that the model remains static, even though the text world changes every time an event takes place. For instance, when the take-off happens (in a later part of this text), most of the facts stated in these opening paragraphs are no longer true. Some means should be developed to capture the updating of a textual world as reading progresses (e.g., that would convert the location of “rocket” from “desert” to “sky”). It might emerge that the changes readers make when building their own models of text content reflect the effects of updating in regular ways.

One final drawback deserves notice: that the representation does not account for “reference” as a correspondence of concepts to some real or possible objects and events in the world of everyday experience. If the text were a report of an actual flight, and readers had been present there, they might be judging the various assertions as true vs. false, or accurate vs. inaccurate. One conjecture is that reference is done for whole model spaces, not for individual words and concepts, because context is required for efficient matching. But since humans often do not have such access to the reality, they are skilled in building models of that domain also. Hence, reference would be  a matching of models, and truth (or accuracy) a threshold of correspondence that people in communication are disposed to demand.

One group of readers was tested some six weeks after they recalled the “rocket” text. They were asked to decide the “truth” of statements like “A rocket always burns fuel” and “A rocket is often used to attack military targets.” The outcome was disrurbing. Only two statements were received with unanimous agreement (“Without fuel, a rocket can’t fly” and “Launching sites usually are located on the earth”). In all other cases, people couldn’t agree whether the facts were always, often, or sometimes true. Average US readers are certainly uninformed about rocketry — they even have the “rockets’ red flare” in their National Anthem. Yet people do much of their reading about areas in which they are not experts, and reading models must respect that factor. Traditional logicians’ models of reference are unrealistic. Our concern should rather be with “general design principles, which appear to be important in the effective utilization and integration of large amounts of intrinsically uncertain knowledge” (Allport, 1979, p. 61).

I have deliberately attacked my own model to show the pressing need for further evolution. In that regard, one major contribution of any  theory or model (beyond its potential for representation and discovery) is its ability to reveal the mode of its own transcendence. Endel Tulving (1979, pp. 29ff.) observes: 

There is every reason to believe that all current ideas, interpreta­tions, expIanations, and theories are wrong, in the sense that sooner or later they will be modified or rejected . . . . Supporting, retaining, and affirming existing theories longer than necessary more often than not stands in the way of progress. Experiments aimed at existing theories should be designed to find out how, where, and in what sense these theories need revision or why they should be rejected.

3. Toward an Ideal Model?

To adequately explore reading, a necessary first step is a firm definition of the notion of “text”. It cannot be not just a series of sentences, as one is often required to assume. A text is often both more and less than the sum of its parts. And people communicate quite successfully with texts even though these parts are fuzzy and non-determinate in isolation, e.g., the meanings of individual words and concepts being subject to dispute. The definition posited here is: the text is an actual system, that is, a working system in which decisions and selections have been made such that the various occurrences have some function(s) in contributing to the operations of the whole. This system is quite different from a virtual system like syntax, which merely states the available options. In an actual system, occurrences mutually constrain each other and thus constrain what text users can do. To disregard these constraints — for instance, by assigning one’s own personal meanings to words, like Humpty Dumpty — is to abuse the system and endanger its operation. Some latitude is of course allowed, according to a principle that might be compared to cybernetic regulation: A change in a system elicits some compensatory actions elsewhere. (For instance, speakers can invent their own meanings if they stop and explain them.).

Each subsystem of a text (e.g., lexicon, grammar/syntax, concepts, plan steps, discourse actions, and so forth) runs partly on internal principles and partly on requirements of feedback from the others. This interaction of subsystems is regular, but asymmetrical, lacking fixed one-to-one correlations of operations, functions, or elements (cf. p. 289). When many elements in one subsystem would each suffice alone to signal a higher-power single function in another subsystem, redundancy results.

The stability of a text as a system is constituted by a functional continuity of occurrences within or among its interacting subsystems: The function of each occurrence (its contribution to the system’s operation) must somehow be relevant to that of others, for instance, by providing access, imposing constraints, signalling contingent decisions, and so on. The significance of occurrences or of configurations of these, such as word meanings, sentence formats, speech-act purposes, and so on, can be viewed as an ordered set of hypotheses about appropriate processing actions. When these actions have been carried out to the satisfaction of the processor, a threshold of termination is reached, and operations are directed eIsewhere.

It is difficult to believe that the staid appeal to the “author’s intentions” is really decisive. (Compare the use of the notion in Frederiksen, 1977; Clark, 1978; Freedle & Meyer, 1979.) The author must not only “intend” to elicit processing actions; he or she must also know how to do so successfully. The intentionality must somehow be actualised in the text as a configuration that imposes a characteristic demand for reading tasks to be done with certain results. The author can influence the probability of those results, but cannot force them to come about.

The same point can be made at the readers’ end. Many researchers fear that by downgrading the author’s intention, we open the door to utter anarchy, as in Herbert Clark’s (1978, p. 295) whimsical case of someone being “adventitiously reminded of her mad Uncle Harry.” But, as l already noted, abusing a system renders it dysfunctional, and communication could break down. And how are readers to know the author’s intention except by enacting the text?

l have tried to specify the nature of the text-system according to seven principles of textuality. Cohesion is the principle that connectivity should obtain among the surface elements of the text (sounds, letters, syntax). Coherence is the principle that connectivity should obtain among the underlying concepts and relations. Intentionality is the text producer’s attitude that a cohesive, coherent text is being created for some goal; acceptability is the corresponding attitude on the text receiver’s part. Informativity results from the extent to which text occurrences are not probable nor predictable in their context. Situationality is the relevance of a text to a current or recoverable situation. Intertextuality is the principle whereby the utilization of a given text depends on knowledge and experience gained from using other texts (specifically or in general).

For an extensive discussion of these principles, l can only refer to my previous work (Beaugrande, 1980a, 1984, 1997, 2004; Beaugrande & Dressler, 1980). It must suffice to note here that each of these principles is chiefly devoted to continuity, which, in most cases, is not openly enacted in the surface text. Accordingly l have appealed to the notion of problem-solving as a general search and discovery method for connecting systemic states (NewelI & Simon, 1972). A problem is defined here simply as a state from which the attainment of a desired successor state is not certain or obligatory. Problems may be trivial if their solution is easy to find, or serious if the chances for failure (non-attainment of the envisioned state) outweigh those of success.

It must be stressed that a text with no problems at all, hence, a system with total stability, is neither possible nor desirable. No text can make every connection explicit. And every text is in some ways unique. But even a text with only trivial problems and occurrences of high predictability would be of little value, and readers would lose interest very soon. At the other extreme, a text with numerous serious problems and almost no predictability would overload the processing capacities of normal readers. Thus, the topmost control factor for creating a text must be to keep a suitable degree of unpredictability and problematicity for the expected audience. It is in this sense that the author sets tasks for the readers and speculates on the outcome. Communication is presumably driven by the constant removal and restoration of satisfactory (but not total) stability. There may even be some neurological impulse to escape the stagnation of cognitive systems; boredom is naturally unpleasant.

It should now be clear that no model of reading can dispense with assuming heavy processor contributions, even to deal with the surface text. Phrases, clauses, sentences, coherence, purpose — all these must be enacted, not just found. Moreover, there must be operations for contributing additional material. Spreading activation ensues without special command when some element is placed in working memory, that is, its active status “spreads” to other items closely associated in stored knowledge (Collins & Loftus, 1975). Inferencing is done whenever there is a specific discontinuity to be bridged. (See Warren, Nicholas, & Trabasso, 1979.) The distinction that might be drawn (and not all researchers do so) is one of directedness and conscious control. Though usually cited in dealing with conceptual coherence, these two operation types might be envisioned as running in any textual subsystem, and with any degree of scale or power. It would follow that part of the problem-solving posited in my model is performed without effort through spreading activation (cf. Ortony, 1978).

We must now confront the issue of automatization. Following LaBerge and Samuels (1974, p. 295), we can define “automatic” processes as those done “while attention is directed elsewhere.” If “attention” is in turn defined as an expenditure of processing resources that inhibits the potential for other tasks at the same time (Keele, 1973), we can conclude that automatic processes do not compete with whatever other operations the mind must perform simultaneously. Clearly, at least some areas of understanding must run automatically, since the brain has a limited capacity for complex tasks (cf. Posner& Snyder, 1975, pp. 64ff.). But it is probably too strong to claim that it is all automatic (as Anderson & Bower, 1973, suggest for memory). As a working hypothesis, it might be argued that conscious control is retained over high-powered processing, while low-powered tasks are done by automatized subroutines. Between the two domains there would be a flexible threshold that moves up or down depending on how much time and attention can be made available. A possible corollary is that most laboratory experiments tend to manipulate this threshold and hence to yield unnatural results (cf. p. 302).

Now, if certain reading operations are in fact done automatically while others are going forward, the former might appear to take no time. The actual time spent on reading would represent only an accumulation of those tasks that require conscious attention. We have no way of measuring the operations running concurrently in terms of time variables. This raises questions about the experiments like those done by the Clarks in which understanding operations must be proven in terms of the time used up (cf. p. 272).

Since George Miller’s (1956) pioneering tests, it has been accepted that short-duration working memory has a fixed capacity of about seven chunks. More recently, Kintsch and Polson (1979) suggest that this capacity is more flexible, depending on how much additional drain is being made on cognitive resources at the time. Thus, the capacity would be larger at the beginning of a task and shrink considerably as processing load increases later on. It seems plausible that reading operations are well-chunked in any case to avoid overload under most normal conditions (barring high unpredictablity as noted above). Yet the same automatized subroutines suggested here as consuming no time might also consume no space in working memory! It has been conclusively shown that long-term memory can be contacted by automatically perceived input (even if one tries to ignore it). Steven Keele (1973), who reviews the evidence on this point, suggests that long-term memory contact is unlimited: the bottleneck is rather in one’s ability to react. That factor might play some role in readers’ ability to understand more than they can report, though other factors must be involved as well.

Research on learning and memory also indicates that processing is uniformly more efficient if the occurrence is cued in advance or otherwise predictable (cf. Laberge, Van Gelder, and Yelllot, 1970; Keele & Boies, Note 5). Following the proposal of Kintsch and Polsen just cited, it is conceivable that a reader might begin processing with a fairly large working memory capacity employed to gather a sampling of input and to designate tasks to the proper subroutines, many fully automatic. Thereafter, working memory shrinks but experiences no absolute overload because the automatic activities consume no space. This effect would explain why the placement of a main idea at the beginning of a paragraph or text is so useful for organizing and recalling content (findings in Kintsch, Kozminsky, Streby, McKoon, & Keenan, 1974; cf. also above, pp. 275-276).

It can be objected that since texts are nearly always unique in some ways, automatization must have major limits. But the question arises whether this uniqueness is not largely finalized by low-powered operations on local structures: At a higher degree of power and scale, there may be much less differentiation. Suppose then that the higher-powered processes make the major decisions and selections that call up packages of automatic subroutines. These subroutines could run in such a way as to intersect at strategic points and influence each other's behaviour without further instructions from high-powered controls. For example, the interaction of conceptual meanings that engenders context from configurations of words might well operate via intersection of knowledge pathways in memory (cf. Quillian, 1966; Collins & Loftus, 1975; Rieger, 1975; Woods, 1978; Fahlman, 1979). Bizarre occurrences might block the subroutines and force the processor to apply high-powered resources in order to retain control. But inconsequential anomalies such as misspelled words or minor ambiguities might cause at most tiny slowdowns without leaving the domain of automaticity (and hence accounting for findings such as MacKay, 1966, on syntax; Marslen-Wilson, 1973, on corrections made during shadowing; and Gough & Cosky, 1975, on letter recognition).

For operations like these to be feasible, powerful methods of pattern-matching are needed in order to recognize and manage vast numbers of partially similar and partially diverse configurations. Rumelhart (in a personal communication) views pattern-matching as largely automatic. Readers might set a given standard of "goodness of fit" that the matching process should attain in order to count as successful, depending on the motives for reading and using that text; processing depth might then arise from the goodness of fit being demanded, as reflected in better or poorer memory for the text itself. Alternately, some patterns might have their own built-in demands; for example, a mismatch of determinate or typical knowledge (cf. p. 292) would be more troublesome than one of accidental or neutral knowledge. Strategic intersection points of automatized subroutines could be centered upon determinateness and typicality.

The possibility that automatic processes elude usual experimental techniques, as I have speculated, places special emphasis on the subtlety of model design. Laboratory settings and brief, uninteresting language samples probably impel readers to devote attention to many operations normally done automatically. In natural settings, reading would run with greater smoothness and rapidity, while performance on recognition or recall would be greatly relaxed if it did chance to be asked for. This relaxation is exactly what would be predicted if the lower-powered processes involved here (mostly, recognition of words and phrases) are indeed carried out without attention.

Available data strongly indicates that memory storage must involve abstraction, construction, and reconstruction to some degree (cf. data in Beaugrande, 1980b; Beaugrande & Miller, 1980). As suggested in the foregoing paragraph, abstraction is not usually necessary for recalling texts read in natural settings —people can report what they read without being able to get back exact words or sentences. But there is apparently some abstraction under way nonetheless: Bits and pieces of surface structure persistently find their way into the recall protocols l have obtained under various conditions. Some of these pieces could be reconstructions, especially if there is only one normal way to express a particular concept (e.g., "rocket") (cf. Beaugrande, 1980a, pp. 175,220, 228); yet not all of them can be accounted for in this way (Walter Kintsch, personal communication). Here also, a controlled laboratory setting, by moving the threshold of attention toward lower-powered processes, might encourage an artificially high degree of abstraction. Everyday readings of long texts would be more prone to result in constructive and reconstructive storage.

Utilization is in all likelihood variable, depending on the readers' interests, motivations, and available time or resources. Anything as thorough as linguists' analysis is hardly likely to be performed under normal conditions. The usual case would presumably be casual utilization. Readers could pass over a certain amount of surface material in search of a global gist (DeJong, 1977; Schank, Lebowitz, & Birnbaum, 1978; Masson, 1979). Omitted materials could be regarded as non-essential or supplied from stored patterns as defaults and inferences. When building a gist, readers might not even experience the omitted materials as gaps: By raising scale and power, continuity can be satisfied using only the control centers or macro-states, e.g., the central concept within a whole conceptual space, as suggested by my networks.

Redundancy among interacting language subsystems would allow for casual utilization in any case. The “function words” in syntax (e.g. determiners, prepositions, conjunctions) are often so slurred and indistinct in speech that people are obviously able to supplant their functions by inferring conceptual relations. (Compare data in Woods  & Makhoul, 1973.) It is therefore plausible when some of the models reviewed above (Meyer, Frederiksen, Schank) do not foresee heavy utilization of function words.

My own suggestion, aired above, that reading consists of simultaneous processing phases could be relevant here. There might be an “attention-sharing” mechanism (comparable to “time-sharing” on the computer) along the lines suggested by LaBerge and Samuels (1974): Attention could be shifted so rapidly from one domain or phase to another that the illusion of attending to them all at once could arise. Casual utilization could be done wherever attention happened to be absent at the moment; casualness could reach a high degree if some extreme drain on attention were imposed, such as a drastic disturbance in the readers' vicinity. Still, Woods’ cascading network model indicates how domains that are being denied attention might be reasonably well controlled by merging their processing with that of domains in focus.

The long and unresolved disputes over decomposition are symptomatic of the divergence between classificatory or logical models and operational models. The frequent claim of linguists that word meanings are decomposed into “semantic features” (e.g., E. Clark, 1973) has seldom indicated how readers could process the meaning of extended discourse in any orderly and rapid fashion. It has never been convincingly proven that one set of semantic features could encompass all aspects of the meaning for all the words in a language; on the contrary, most examples in the literature seem eminently unsuited for capturing subtle, individual aspects of meaning (mental imagery, social or personal values, etc.), and many proposed features (like “+ gaseous pressure” or “+ without-a-mate-at-breeding-time”) seem intuitively forced and arbitrary.

A more operational approach to decomposition would be to focus not on a listing of individual features, but rather on procedures that can identify and utilize basic shared meaning components during realistic understanding tasks. Features would figure as content types whose detailed specification depends on the current context of cognition or communication. Semantic features would thus be operational tendencies during appropriate processing actions, such as: analogy (using one configuration as a model for another; cf. Rumelhart, 1979), inheritance (passing features from some class or superclass to a specific instance or subclass; cf. Brachman, 1978; Fahlman, 1979), and creativity (modifying some system in response to a particular motivation; cf. Beaugrande, 1979).

There would accordingly be no principled answer to the question: Do readers decompose word meanings? Presumably, a need for decomposition arises if the components of patterns (e.g., schemas) to be matched happen to be more specific or basic than the components provided by the presented text (Rumelhart, personal communication). In one instance, for example, it might be relevant to know that a “rocket” has certain meaning components as compared to “aircraft,” “missile,” or “flying object” (e.g., being unpiloted, lacking landing gear and large wings, etc.); in another instance, readers might have no use for such knowledge. Reading experiments do show cases where recall is based on components both more specific and more general than individual word meanings (Beaugrande, 1980á, p. 228), but a more careful study of contextual motivation is required: We cannot resolve issues of breaking down meanings without simultaneously considering those of assembling meanings.

Similar conclusions could be drawn regarding scale. In traditional linguistics, researchers would set up a taxonomy of “ranks” (or “levels”) with units of various size: morphemes, lexemes, clauses, sentences, etc. Like semantic features, these ranks are classificatory rather than operational, and may not be obligatory units of processing at all — disturbing possibility in view of all the research on “letter identification,” “word identification,” etc. (e.g. Smith, 197l). At the very least, the long stress on local scale is far too restrictive. Most studies have been devoted to “sentence comprehension” (see the monumental survey in Levelt, 1978), though virtually no real discourse is just one sentence long. Clark and Clark rationalize this practice (1977, p. l54): “The studies of unrelated sentences tell us what to look for in ‘normal’ situations.” This assertion is at best unproven, and at worst downright dysfunctional. Sentence formats are discarded early during reading, e.g., before the reader constructs a running gist of the text. The selection of processing units depends on memory capacity, attentional resources, expertise, and the like, not on the grammatical categories of linguists. Indeed, language users are often not confident about where sentences begin and end in speech (Broen, Note 6), and the acceptability of sentences is a function of their context (Greenbaum, 1977).

The Clarks place their reliance upon laboratory tests where readers see a single sentence and then “press a button when they feel they have understood the sentence” (Clark, 1978, p. 314). By varying the sentence with regard to complexity, ambiguity, number of propositions, inferences, or whatever, the Clarks try to prove some operation by showing that it takes extra time. But as was argued already, many reading operations would not necessarily consume extra time when they run automatically in parallel with others. Confronted by an isolated sentence, readers suddenly find their high-powered procedures for distributing attention, allotting memory space, co-ordinating sub-routines, etc., inappropriate. Resources are abruptly freed for careful execution of tasks that would normally be done quite casually, e.g., noticing surface structures, word classes, possible ambiguities or traps, and so on. The outcome is a misleading picture of unnaturally thorough syntactic analysis of the kind linguists are inclined to postulate. l cannot see how this picture can be transferred to “normal situations.” Even the “pressing of a button” upon having “understood a sentence” seems disturbingly unnatural and awakens ominous associations with a stimulus-response situation people do not encounter in real life.

The recent concern for global scales is therefore greatly to be welcomed. Care should be exercised to differentiate the status of global patterns (Beaugrande, 1980a, pp. 163ff.). The frame is an array signalling what elements are in principle related and mutually accessible. The schema is a progression showing in what order elements are likely to occur in a context. The plan is a progression of steps that lead to a goal. The script is a statement of the routine actions and events in a stereotyped situation having a goal structure, with a specification of participant roles. Thus, schemas are frames placed in sequential progression, plans are goal-directed schemas put into application, and scripts are socially stabilized plans. For example, people might have a “house-frame” stating what houses are like, what they are made of, etc.; a “house-touring schema” or a “house-building-schema” stating what order would be traversed in walking through or constructing a house; a “house-building plan” for actually making a house step by step; and a “house-building script” for professional contractors who routinely build houses.

A major difficulty with global entities is the degree to which they are flexible versus fixed. A fixed pattern would be of little use except in cases of unusually great coincidences; a flexible pattern presents the problem of how to decide where it applies at all. As Wilensky (1978) has shown, the script is too inflexible to be used when reading about the wide variety of typical human situations and actions, so that a more versatile plan theory is called for. No doubt there are gradations among the global entities of each type. There may be frames and schemas whose major function is to offer formatting rather than content, especially when readers are uncertain about the text’s topic (e.g., while reading the opening passage of a text or chapter). These frames and schemas would offer substantial power for managing variety and novelty of content.

However, it is essential not to confuse scale with power. Processors ought to have both low-powered and high-powered operations on both local and global scales. In the local operations of sentence parsing categories like “head:' “modifier:' or “verb” would be more powerful than “count noun,” “descriptive adjective,” or “transitive active verb”; the processor could get by most of the time with the powerful categories because syntax interacts with other processing phases. (Compare the models of the Clarks, Kintsch, Schank, and Woods in particular.) In concept recovery, the processor could use high-powered categories like “object,” “event,” “action,” “location.” and so forth, specifying them as needed to construct the content activated by particular textual expressions. On a global scale. a “social gathering-schema” would be more powerful than a “party-schema” or a “birthday-party-schema:' a “room-frame” more powerful than a “kitchen-frame”; a “steal-plan” more powerful than a “bank-robbing-plan”; and so on. The normal processing tendency should be to rely at first on the most powerful entity and work down to the less powerful only as needed for the context. I suggested above that attention and memory space are chiefly reserved for higher-powered operations, while lower-powered ones are done by automatic subroutines with strategic intersections: This principle should hold for both local and global scales.

On the other hand, considerations of processing depth and logical vs. procedural adequacy must be kept in mind. Processing depth appears to vary according to the threshold imposed between attentional and automatic operations. In normal reading, the priority of getting at the general meaning and relevance of a text clearly draws attention away from noticing and storing away the exact words and sentences. Consequently, recall tends to be approximative. Yet special tasks like proofreading spellings, watching for certain sounds, or finding rhymes (cf. the literature cited under “processing depth” in Section 1) can drain attention away such that lower-powered processing is apparently quite possible without a firm base in higher-powered (the reverse of the normal case). Recall and recognition are poor because working memory has been wastefully cluttered with details not firmly integrated into the powerful organizational patterns used by long-term memory (Kintsch. 1974. l 977a). In contrast, tasks like disambiguation, detecting anomalous meanings, fitting sentences into sequences, or devising follow-up sentences (again examples from the literature) impel readers to work with higher power, keeping down memory load and promoting organization suitable for long-term memory. Moreover. these activities are more commonplace in communication and can thereby draw upon more efficient routines. Trade-offs of this kind are likely to emerge in all reading models with a carefully integrative design.

In logical adequacy, the trade-off is between power and flexibility — an unusual source of antagonism between factors that are normally in agreement. The traditional predicate logic descending variously from Aristotle, Frege, Russell, and Carnap bought its power by strictly limiting the contextual relationships that the system could represent. Thus, mechanical decisions could be made about the truth or falsity of any statement (predication) or — using conjunction rules — any sequence of statements. Reference was viewed as a property of noun phrases quantified either as individual objects or as universal sets. In exchange, procedural adequacy for realistic models of understanding is quite unsatisfactory here. We need to study not formal proofs and quantifiers, but rather cognitive thresholds at which readers believe statements and recognize objects or sets. Only after having clarified these issues could some form of (fuzzy) logic be re-introduced in a meta-perspective of reconstruction (e.g., the use of deductive theorem-proving in abstract pattern-matching, cf. Beaugrande, 1981).

Future research can be expected to uncover some kind of “logical consultant” residing in cognitive and communicative processors. It is responsible for matters like identity, conditionality, consistency, set membership, induction, and deduction. Yet its advice is not always asked, and when asked, not always taken, by the top-powered controls. Thus, readers entertain no discomfort in reading about impossible “facts” in imaginary worlds. Indeed, contradictions, paradoxes, disallowed transmutations of identity, and such things, exercise a strong fascination upon readers. Still, the readers could recognize and explain why logic is being violated if they were asked — presumably by recourse to the “logical consultant” whose advice they were ignoring before. Perhaps the research data used to support claims about the logical essence of thought and meanings are merely a partial and artefactual picture obtained by making humans fall back unnaturally on the logical consultant.

The remaining criteria from Section 1 reveal the dependency of research not so much on observable evidence as on our capacities for representation and design. All the researchers (except Chomsky and Gibson) said that they view reading as interactive and parallel, capable of freedom and learning, and sensitive to types of materials. Yet they often conceded that their own models do not capture these factors. It appears that such factors can only be attained after the more essential aspects of design (e.g., scale and power) have been worked out. For example, a model of automatic reading must begin with modular development of components that are integrated later on (Schank et al., 1978); human experiments must similarly focus on individual components of reading (Kintsch, Note 2). But the ultimate success of such models requires a design that makes the modules readily amenable to interactive operation on real-time tasks (Rumelhart, 1977c). Conventional linguistic models have been constructed such that either (a) all language subsystems look totally different, or (b) low-powered, local notions (minimal features, binary oppositions, grammaticality of sentences) are unsuccessfully carried across subsystems. Consequently, narrowly linguistic reading models have been rather unproductive.

The priorities of both computer hardware and laboratory statistics favour serial processing over parallel. Parallel models can raise horrendous problems of scheduling and factoring for the investigator, but they also seem much more plausible accounts of reading (Rumelhart, 1977c, p. 574). Memory space is quite limited, and many single elements cannot be assigned appropriate functions without consulting the results of other processes that would have to run at about the same moment. Here again, scientific progress has been hampered by inadequate model design. A more circumspect consideration of processor contributions, automatization, processing depth, and power would enable us to attack the design problems posed by parallel processing in terms of the distribution of resources and contingencies. On some level, serial and parallel processing are equivalent, both mathematically (Townsend, 1974) and cognitively (Fahlman, 1979). However, that equivalence does not free us from the obligation to design representations of interactive, parallel reading operations humans might reasonably perform.

The same case must be made for freedom and learning as factors openly sacrificed to commonplace aspirations of researchers. Transformational grammarians and some experimental psychologists expressly proclaim their interest in only those operations common to all humans (or, even to all animals) under all conditions. Freedom was accordingly either denied and ignored (by linguists) or laboriously factored out (by psychologists) as uninteresting and disturbing. Yet the comparatively few studies that take freedom seriously (e.g., ethnomethodology, children’s language disabilities) reveal it to be among the most essential and intriguing aspects of cognition and communication, with crucial importance for the evolution and expression of the personality.

By the same token, learning models are unpopular because their conditions steadily evolve. Yet reading any stretch of text is vitally linked to what has been read so far: refinement of predictions and expectations, discovery and application of patterns (e.g., schemas), distribution of attention and interest, and much more. The tactics necessary for excluding these factors, such as working only with isolated, personally irrelevant sentences, can be no more than a provisional stop-gap crying out for replacement with more natural discovery procedures. For example, story understanding models should explore the ways in which one stretch of the story provides a specific model for a later one (cf. Beaugrande & Miller, 1980).

The restriction to sentences was no doubt intended to evade the question of text types as well. Many text types are vague, intuitive heuristics used by readers to tailor their processing to individual examples. It follows that reading models will have to find control points in the reading process where text-type priorities can be inserted and respected. High-powered categories such as “narrative,” “descriptive,” “argumentative,” and the like could be postulated and specified for lower-powered ones like “children’s story,” “science report,” “political attack,” and so forth, with implications for both formatting and content.

The conclusion to be drawn from the discussion in this whole section is essentially the one suggested at the outset of the paper. The aspirations of scientists ought to be shifted away from sole reliance on traditional methods of observation and falsification on minute, tightly controlled tasks and toward a continual improvement and evolution of theory and model design. Theories can be validated in a lasting way only through functional diversification and consensus. Hence, any theory of reading ought to be just one domain within a general theory of cognition and communication. In the meantime, however, emphasis should fall upon models that deliberately provide for integration and expansion. Among the models reviewed above that in fact evince this striving (Kintsch, Frederiksen, Schank, Rumelhart, Woods, Beaugrande), that of Woods is most dominated by it in fine detail. He expressly works toward a model so powerful that it is “capable of characterizing arbitrary classes of structured entities” (Woods, 1978, p. 24). It could, for instance, readily incorporate the power of general problem solving (Newell & Simon, 1972), which it resembles already (e.g., his “search from both ends” compared to their “means-end analysis”).

On the other hand, we must not rush to the other extreme and sacrifice the experimental literature and its results. Many issues in reading are relevant to topics in established psychology, such as concept learning, generalization, coding, and memory storage. The proper course is rather to re-examine the literature in light of more integrative, powerful models. (See for example Beaugrande, 1980a, pp. 233ff.) In so complex a domain as reading, we cannot afford to disregard hard evidence, even if we still debate about what processes have engendered that evidence. The approach advocated here “would not mean an end to laboratory experiments, but a commitment to the study of variables that are ecologically important rather than those that are easily manageable” (Neisser, 1976, p. 7).

The intensity of reading research in recent years signals that it may well be the test case for a new outlook on a science of cognition and communication. We are caught in the transition of methods and ambitions arising from the fact that our data is data only after substantial interpretation on the part of the investigator. Without such interpretation, the observable substrate (e.g., verbal reports of what was read or remembered) is at best neutral and at worst meaningless. When I examine the reports of children who have read a story, I find their messy, noisy data fascinating in all its freedom. Our research will achieve enduring success only if reasonable sense can be made of the data: what was remembered, changed, combined, or omitted, and why. A model with the appropriate design can reveal the regular, powerful strategies underlying it all. And only then can we make a really substantive contribution to the entire learning process of young readers from all groups and backgrounds.

Recourse to computer simulation seems inescapable in view of the complexities in new model design. Artificial intelligence made a great leap forward when it was discovered that, by using “weak methods and experiments instead of proof,” the computer can be “adapted to the exploration of poorly structured task domains” (Simon, 1979, p. l086) — reading among them. The same difficulties facing the reading researcher have also been crucial to the programmer: storage, search, memory, scheduling, on-the-spot evolution and debugging, and formulating complicated operations as steps and routines. The computer scientists therefore adopted the outlook I have suggested here for reading research: to view scientific puzzles at least partly as issues of design.

The age-old quandary of scientific progress and loyalty to theories, raised at the beginning of this paper, may then become more tractable. In closing, let me propose some values that might be pursued when preferring one theory or model over another, or when determining where progress has been made:

(a) Diversification: A design is preferable if it encompasses more domains than another.

(b) Consensus: A design is preferable if it brings out the correlations between seemingly diverse domains.

(c) Economy: A design is preferable if it uses a few powerful concepts rather than many non-powerful ones for its foundations.

(d) Dynamics: A design is preferable if it enables a progression from sets or systems toward processes, procedures, and simulations.

(e) Efficiency: A design is preferable if it allows for operations to be carried out with less expenditure than another.

(f) Computability: A design is preferable if its computer simulation program is more rapid, compact, and general than another.

(g) Commensurability: A design should allow us to compare and contrast it with previous, alternative designs in a common conceptual framework.

(h) Ecological validity: A design is preferable if it can be extended from the laboratory or computer to socially real situations of everyday life.

(i) Plausibility: A design is preferable if it respects our intuitions more prominently than another.

(j) Increasing approximation: A design is preferable if it can be steadily tuned to capture more and more empirically discoverable traits and factors.

Whether or not these criteria and values are taken as complete or binding, l hope that by discussing and applying them, we may at least sharpen our awareness for issues that should concern both researchers and teachers of reading.  

Footnotes

1. The following researchers were interviewed in regard to the application of my criteria to their work: Herbert Clark, Carl Frederiksen, Walter Kintsch, Wendy Lehnert (for Schank’s group), Bonnie Meyer, David Rumelhart, and William Woods. I also profited from discussions of the preliminary version of the paper with Bob Carey and Jay Samuels. Compare note of acknowledgement, above.

2. The original “standard theory” foresaw understanding as based on a syntactic “deep structure” recovered from “surface structure” via transformations (Chomsky, 1965). The “extended standard theory” relocated understanding (“semantic interpretation”) as addressing both deep and surface structure (Chomsky, 1972). In the newest version, ‘‘trace theory,” understanding is now directed to surface structure annotated by “traces” arising from moving phrases around (Chomsky, 1976). My remarks apply largely to all three versions, which share many features (autonomous syntax, logical base, etc.), but my main concern was with the “standard theory” as the best known and most often applied.

3. Gibson and Levin (1975, p. 147) find the stand taken by Eve Clark (1973) on semantic features “congenial to our way of thinking about the relationships between perceptual development and language development.” Res ipsa loquitur.

4. The conventional notation for propositions is bracketed, e.g., “Socrates is Greek” comes out as (GREEK, SOCRATES), with the “arguments” following the “predicate” (cf. Kintsch, 1974, pp. 13ff.) The same notation is used on occasion by Clark & Clark (1977, p. 46) and Norman & Rumelhart (1975, p. 42). Compare different bracketings in Frederiksen (1977, pp. 62ff.) and Schank & Abelson (1977, p. 15). Network graphics are equivalent in the sense that the nodes are arguments and some of the predicates, while the links are predicates or predicate types (e.g., “attribute or’) (cf. Woods, 1975; Brachman, 1979; Beaugrande, 1980a, ch. 3).

5. However, Anderson (1977) locates the schema emphatically in the reader, rather than in the text.

6. Frederiksen (1975, p. 376) calls for a “Iogical network” of “labeled, binary logical, causal, or algebraic relations connecting propositions which are represented as substructures in the semantic network.” His proposal might be compared to my remarks on the “logical consultant” in Section 3.

7. The importance of actions and events in understanding can be argued on several grounds: They are control centers with rich linkage to other concepts (participants, time, place, causality, etc.); they correspond most often to verbs (major control centers in syntax); they figure in causal chains; they update the state of the world by changing situations; they direct the flow of narratives; they frequently correspond to steps in a plan; and so on (Beaugrande, 1980a, p. 81).

8. The term first appears as an afterthought in Norman & Rumelhart (1975, pp. 406ff.): “a framework that interrelates the different aspects of a body of knowledge.” In Rumelhart (1975), the term figures only in the paper’s title and is apparently identified with a structure of stories (“episodes” of “events,” and “settings” built from “states”); in Rumelhart (1977b), problem-solving is added. However, the broad usage survives in Rumelhart (1977a, p. 165), where a schema is “generalized knowledge about a sequence of events,” and in Rumelhart (1980), where the schema is “a data structure representing the generic concepts stored in memory,” indeed encompassing “our knowledge about all concepts.” In Section 3, I use a narrower definition that agrees with the literature since Bartlett (1932). For discussions of schema theory, see especially Beaugrande & Colby (1979); Norman (1979); and Thorndyke & Yekovich (1980).

9. These remarks on decomposition, automatization, utilization, power, and procedural adequacy. are based on my discussion with Rumelhart himself.

10. Rumelhart is concerned with “semantic memory,” rather than “episodic memory,” as can be-seen from his unequal treatment of the two (Rumelhart, 1977a).

11. There is normally a trade-off between compact storage and rapid search, because it is hard to find things stored only once in a vast network of knowledge (Kintsch, 1977a, p. 290; cf. Beaugrande, 1980a. p. 72).

12. The text originally appeared in McCall &. Crabbs (1961, Booklet C, p. 8), and the Teacher’s College Press has kindly allowed me to reprint it. I use here only a small fragment for purposes of demonstration. More extensive uses appear in Kintsch & Vipond (1979); Beaugrande (1980a, 1980b, 1981).

13. This ordering of hypothesis is intended to be only illustrative. The actual orderings used by readers must be determined empirically (cf. Stevens & Rumelhart, 1975).

14. Note that determiners like “a”, and later (in Figure 4) pronouns like “it” are suppressed in the conceptual-relational networks, since their functions are assignable only in the surface text (but cf. Beaugrande, 1980a, chs. 4-5). I place prepositions as link tags in curly brackets. Link labels are formed from the initial letters of the word naming the concept type toward which the arrows point. Thus, links are traversed by identifying and labelling each arc, just as in the grammatical network system illustrated in Figure 1. However, this traversing is probably not done in a strict time sequence; more likely, it runs in parallel at various points, some of it automatic (cf. discussion in Section 3).

15. These inferences are admitted because large numbers of my readers recalled them as parts of the text on many experimental conditions.

16. This design feature (also proposed by Hollan, 1975) is useful for treating numerous issues set forth in Beaugrande (1980a), e.g., using predictions, measuring informativity, deciding definiteness, and so on. As shown by data cited later on, readers are rather uncertain about their facts on a topic like rocketry.

17. Slides showed launchings at Cape Kennedy, Florida. The water visible in the background led some readers to recall the ending not as “the rocket plunged to earth,” but rather as “the rocket fell into the ocean.”

Reference Notes

  1. Beaugrande, R. Process model design for natural language understanders. Paper presented at the Second Cognitive Science Conference, Yale, June 1980.

2. Kintsch, W. Approaches to the study of the psychology of language. IBM Distinguished Lecture on Language, 1979.

3. Frederiksen, C., Frederiksen, J., Humphrey, F., & Ottesen, J. Discourse inference: Adapting     to the inferential needs of school texts. Paper presented at the American Research Association, Toronto, 1978.

4. Rumelhart, D. Comprehension of stories. Paper presented at the 28th Annual Meeting, National Reading Association, St. Petersburg Beach, 1978.

5. Keele, S. & Boies, S. Processing demands of sequential information. Paper presented at the Western Psychological Association, Portland, 1972.

6. Broen, P. A discussion of the linguistic environment of the young language-learning child. Paper presented at the American Speech and Hearing Convention, Chicago, 1971.  

REFERENCES

ALLEN, J. A plan-based approach to speech-act processing. Toronto: University of Toronto Department of Computer Sciences Technical Report 131/79, 1979.

ALLPORT, D.A. Conscious and unconscious cognition: a computational metaphor for the mechanism of attention and integration. In L. (Ed.), Perspectives on memory research. Hillsdale, N.J.: Erlbaum, 1979.

ANDERSON, J.R. Language, memory, and thought. Hillsdale, N.J.: Erlbaum, 1976.

ANDERSON, J.R., & BOWER, G. Human associative memory. Washington, DC: Winston, 1973.

ANDERSON, R. The notion of schemata and the educational enterprise. In R. Anderson, R. Spiro, & W. Montague (Eds.), Schooling and the acquisition of knowledge. Hillsdale, N.J.: Erlbaum, 1977.

BARTLETT, F. Remembering. Cambridge: Cambridge University Press, 1932.

BEAUGRANDE, R. Factors in a theory of poetic translating. Assen: van Gorcum, 1978. (a)

BEAUGRANDE, R. de. Textual processes in reading research. Gainesville: University of Florida English Department Technical Report NL, 1978. (b)

BEAUGRANDE, R. de. Information, Expectation, and processing. Poetics, 1978, 7, 3-44 (e)

BEAUGRANDE, R. de. Toward a general theory of creativity. Poetics, 1979, 8, 269-306.

BEAUGRANDE, R. de. Text, discourse, and process. Norwood, N.J.: Ablex, 1980. (a)

BEAUGRANDE, R. de. The status of texts in reading research. In R. Tierney, P. Anders, & J. Mitchell (Eds.), Understanding a reader’s understanding. London: Longman, 1980. (b)

BEAUGRANDE, R. de. Linguistic theory and metatheory for a science of texts. Text, 1981, 1.

BEAUGRANDE, R. de. The story of grammars and the grammar of stories. Journal of Pragmatics 6, 1982, 383-422.

BEAUGRANDE, R. de. Text Production. Norwood, N.J.: Ablex, 1984.

BEAUGRANDE, R. New Foundations for a Science of Text and Discourse. Greenwich, CT: Ablex, 1997.

BEAUGRANDE, R. de. Performative speech acts in linguistic theory: The rationality of Noam Chomsky. Journal of Pragmatics 29, 1998, 765-803.

BEAUGRANDE, R. de. A New Introduction to the Study of Text and Discourse. Published on the Internet, July 2004

BEAUGRANDE, R., & COLBY, B.N. Narrative models of action and interaction. Cognitive

Science, 1979,3, 43-66.

BEAUGRANDE, R., & DRESSLER, W. lntroduction to text linguistics. London: Longman, 1980.

BEAUGRANDE, R., & MILLER, G. Processing models of children’s story comprehension. Poetics, 1980, 9, 181-201.

BIRNBAUM, L., & SELFRIDGE, M. Problems in conceptual analyses of natural language. New Haven: Yale Department of Computer Sciences Research Report 168, 1979.

BOBROW, R. The RUS system. In B. Webber & R Bobrow, Research in natural language understanding. Cambridge: Bolt, Beranek, and Newman Technical Report 3878,1978.

BOBROW, S., & BOWER, G. Comprehension and recall of sentences. Journal of Experimental Psychology, 1969, 80, 455-461.

BRACHMAN, R. A structural paradigm for representing knowledge. Cambridge, MA: Bolt, Beranek, and Newman Technical Report 3605, 1978.

BRACHMAN, R. Research in natural language understanding. Cambridge, MA: Bolt, Beranek, and Newman Technical Report 4190, 1979. (a)

BRACHMAN, R. On the epistemological status of semantic networks. In N. Findler (Ed.)¡ Associative networks. New York: Academic, 1979. (b)

BRANSFORD, l., BARCLAY, R., & FRANKS, l. Sentence memory: A constructive vs. interpretative approach. Cognitive Psychology, 1972, 3, 193-209.

CARBONELL, l. POLITICS: Automated ideological reasoning. Cognitive Science, 1978, 2, 27-51.

CHOMSKY, N. The logical base of linguistic theory. In H. Lunt (Ed.), Proceedings of the Ninth International Congress of Linguists. The Hague: Mouton, 1964.

CHOMSKY, N. Aspects of the theory of syntax.

Cambridge, MA: MIT Press, 1965.

CHOMSKY, N. Studies on semantics in generative grammar. The Hague: Mouton, 1972.

CHOMSICY, N. Conditions on roles of grammar. Linguistic Analysis, 1976, 2, 303-351.

CLARK, E. What’s in a word? On the child’s acquisition of semantics in his first language. In T. Moore (Ed.), Cognitive development. New York: Academic, 1973.

CLARK, H. Inferences in comprehension. In D. LaBerge & S.J. SamueIs (Eds.), Basic processes in reading: Perception and comprehension. Hillsdale, N.J.: Erlbaum, 1977.

CLARK, H. Inferring what is meant. In W. Levelt & G. Bores d’Arcais (Eds.), Studies in the perception of language. New York: Wiley, 1978.

CLARK, H., CARPENTER, P., & JUST, M. On the meeting of semantics and perception. In W. Chase (Ed.), Visual information processing. New York: Academic, 1973.

CLARK, H., & CHASE, W. Perceptual coding strategies in the formation and verification of descriptions. Memory and Cognition, 1974, 2, 101-111.

CLARK, H., & CLARK, E. Language and psychology. New York: Harcourt, Brace, & Jovanovich, 1977.

CLARK, H., & HAVILAND, S. Comprehension and the given-new contract. In R. Freedle (Ed.), Discourse production and comprehension. Norwood, N.J.: Ablex, 1977.

COLLINS, A., & LOFTUS, E. A spreading-activation theory of semantic processing. Psychological Review, 1975,82, 407-428.

CRAIK, F., & TULVING, E. Depth of processing and the retention of words in episodic memory. Journal of Experimental Psychology: General, 1975, 104, .268-294.

CULLINGFORD, R. Script application: Computer understanding of newspaper stories. New Haven: Yale Department of Computer Sciences Research Report 116, 1978.

DEJONG, G. Skimming newspaper stories by computer. New Haven: Yale Department of Computer Sciences Research Report l04, 1977.

ERNST, G., & NEWELL, A. GPS: A case study in generality and problem solving. New York: Academic, 1969.

ESTES, W.K. On the descriptive and explanatory functions of theories of memory. In L. Nilsson (Ed.), Perspectives on memory research. Hillsdale, N.J.:Erlbaum, 1979.

FAHLMAN, S.E. NETL: A system for representing and using real-world knowledge. Cambridge, MA: MIT Press, 1979.

FREDERIKSEN, C. Representing logical and se­mantic structure of knowledge acquired from discourse. Cognitive Psychology, 1975, 7, 371-458.

FREDERIKSEN, C. Semantic processing units in understanding text. In R. Freedle (Ed.), Discourse production and comprehension. Norwood, N.J.: Ablex, 1977.

GENTNER, D. Testing the psychological reality of a representational model. TINLAP-2, 1978, 1-7.

GlBSON, E. Perceptual learning and the theory of word perception. Cognitive Psychology, 1971,2, 351-368.

GlBSON, E. How perception really develops. In D. LaBerge & S. J. Samuels (Eds.), Basic processes in reading: Perception and comprehension. Hillsdale, N.J.: Erlbaum, 1977.

GlBSON, E., &LEVIN, H. The psychology of reading Cambridge, MA: MIT Press, 1975.

GlBSON, J.J. The senses considered as perceptual systems. Boston: Houghton-Mifflin, 1966.

GOMULICKI, B. Recall as an abstractive process. Acta Psychologica, 1956, 12, 77-94.

GOODMAN, K. Reading: A psycholinguistic guessing game. In H. Singer & R. Ruddell (Eds.), Theoretical models and processes of reading. Newark: IRA, 1976, 497-508.

GOUGH, P. One second of reading. In J. Kavanagh & L Mattingly (eds.), Language by ear and by eye. Cambridge, MA: MIT Press, 1972.

GOUGH, P., & COSKY, M. One second of reading again, In J. Castellan, D. Pisoni, & G. Potts (Eds.), Cognitive theory, Vol. II. Hillsdale, NJ.: Erlbaum, 1975, 271-288.

GREENBAUM, S. (Ed.). Language and acceptability. The Hague: Mouton, 1977.

GRIMES, J. The thread of discourse. The Hague: Mouton, 1975.

HOLLAN, J. Features and semantic memory: Set theoretic or network model? Psychological Review, 1975, 82, l54-l55.

HOLMES, J., & SINGER, H. The substrata-factor theory. Berkeley: University of California, 1961.

KAPLAN, R. On process models for sentence analysis. In D. Norman & D. Rumelhart, (Eds.), Explorations in Cognition. San Francisco: Freeman, 1975.

KEELE, S. Attendance and human performance. Pacific Palisades: Goodyear, 1973.

KINTSCH, W. The representation of meaning in memory. Hillsdale, N.J.: Erlbaum.

KINTSCH, W. Memory and cognition. New York: Wiley, 1977. (a)

KINTSCH, W. On comprehending stories. In M. Jost & P. Carpenter (Eds.), Cognitive processes in comprehension. Hillsdale, N.J.: Erlbaum, 1977. (b)

KINTSCH, W., & VAN DIJK, T. Toward a model of text comprehension and production. Psychological Review, 1978,85, 363-394.

KINTSCH, W., KOZMINSKY, E., STREBY, W., MCKOON, G., & KEENAN, J. Comprehension and recall of text as a function of content variables. Boulder: University of Colorado Department of Psychology research report, 1974.

KINTSCH, W., & POLSON, P. On nominal and functional serial position curves: Implications for short term memory models? Psychological Review, 1979, 84, 407-413.

KINTSCH, W., & VIPOND, D. Reading comprehension and readability in educational practice and psychological theory. In L. Nilsson (Ed.), Perspectives in memory research. Hillsdale, N.J.: Erlbaum, 1979.

KUHN, T.S. The structure of scientific revolutions. Chicago: University of Chicago.

LABERGE, D., & SAMUELS, S.J. Toward a theory of automatic information processing in reading. Cognitive Psychology, 1974, 6, 293-323.

LABERGE, D., VAN GELDER, P., & YELLOT,J. A cueing technique in choice reaction time. Perception and Psychophysics, 1970, 7,57-62.

LEHNERT, W. Representing physical objects in memory. New Haven: Yale Department of Computer Science Research Report 131, 1978.

LEHNERT, W. The process of question answering. Hillsdale, N.J.: Erlbaum, 1979a.

LEHNERT, W. Text processing effects and recall memory. New Haven: Yale Department of Cómputer Science Research Report 157, 1979b.

LEVELT, W. A survey of studies in sentence perception: 1970-1976. In W. Levelt & G. Flores d’ Arcais (Eds.), Studies in the perception of language. New York: Wiley, 1978.

MACKAY, D.G. To end ambiguous sentences. Perception and Psychophysics, 1966,l,426-436.

MANDLER, J., & JOHNSON, N. Remembrance of things parsed: Story structure and recall. Cognitive Psychology, 1977,9, 111-151.

MARCUS, M. A theory of syntactic recognition for natural language. Cambridge, MA: MIT Press, 1980.

MARSLEN-WILSON, W.D. Linguistic structure and speech shadowing at very short latencies. Nature, 1973, 244, 522-523.

MASSON, M.E. Cognitive processes in skimming stories. Boulder: University of Colorado Institute for the Study of Intellectual Behavior Technical Report 84-ONR, 1979.

MCCALL, W., & CRABBS, L. Standard test lessons in readability. New York: Teachers College Press, 196l.

MCCORDUCK, P. Machines who think. San Francisco: Freeman, 1979.

MEYER, B.F., The organization of prose and its effects on memory. Amsterdam: North Holland, 1975.

MEYER, B.F. What is remembered from prose: A function of passage structure. In R. Freedle (Ed.), Discourse production and comprehension. Norwood, N.J.: Ablex, 1977.

MEYER, B.F. Following the author’s top-level structure. In R. Tierney, P. Anders, & J. Mitchell (Eds.), Understanding a reader’s understanding. London: Longman, 1980.

MEYER, B.F., & FREEDLE, R. Effects of discourse type on recall. Tempe: Arizona State Uni­versity Department of Educational Psychology Prose Learning Series Report 6, 1979.

MEYER, B.F., & MCCONKIE, G. What is recalled after learning a passage? Journal of Educational Psychology, 1973,65, 109-117.

MILLER, G. The magic number seven, plus or minus two. Psychological Review, 1956, 63, 81-97.

MlLLER, J. & KINTSCH, W. Readability and recall of short prose passages: A theoretical analysis. Journal of Experimental Psychology: Human Learning and Memory, 1980, 6.

MINSKY, M., & PAPERT, S. Artificial intelligence: Condon Lectures. Eugene: Oregon State System of Higher Education, 1974.

MISTLER-LACHMAN, J. Depth of comprehension and sentence memory. Journal of Verbal Learning and Verbal Behavior, 1974,13,98-106.

NEISSER, U. Cognitive psychology. New York: Appleton-Century-Crofts, 1967.

NEISSER, U. Cognition and reality. San Francisco: Freeman, 1976.

NEWELL, A., & SIMON, H. Human problem solving. Englewood Cliffs: Prentice-Hall, 1972.

NORMAN, D. Perception, memory, and menial processes. In L. Nilsson (Ed.), Perspectives on memory research. Hillsdale, N.J.: Erlbaum, 1979.

O’CONNELL, D. One of many units: The sentence. ‘ In S. Rosenberg (Ed.), Sentence production. Hillsdale, N.J.: Erlbaum, 1977.

ORTONY, A. Remembering, understanding, and representation. Cognitive Science, 1978, 2, 53-69.

ORTONY, A., & ANDERSON, R. On putting apples into bottles: a problem of polysemy. Cognitive Psychology, 1975, 7, 167-180.

POSNER, M., & SNYDER, C. Attention and cognitive control. In R. Solso (Ed.), Information processing and cognition. Hillsdale, N.J.: Erlbaum, 1975.

QUlLLIAN, R. Semantic memory. Cambridge, MA: Bolt, Beranek, and Newman Technical Report AFCRL-66-l89, 1966.

RIEGER, C. Conceptual memory and inference. In R. Schank, N. Goldman, C. Rieger, & C. Riesbeck, Conceptual information processing. Amsterdam: North Holland, 1975.

RlESBECK, C. Conceptual analysis. In R. Schank, N. GoIdman, C. Rieser, & C. Riesbeck, Conceptual information processing. Amsterdam: North Holland, 1975.

RIESBECK, C., & SCHANK, R. Comprehension by computer: Expectation-based analysis of sentences in context. In W. Levelt & G. Rores d’ Arcais (Eds.), Studies in the perception of language. New York: Wlley, 1978.

ROYER, J. Remembering: constructive or reconstructive? In R. Anderson, R. Spiro, & W. Montague (Eds.) Schooling and the acquisition of knowledge. Hillsdale, N.J.: Erlbaum, 1977.

RUMELHART, D. A multicomponent theory of the perception of briefly exposed visual displays. Journal of Mathematical Psychology, 1970, 7, 191-218.

RUMELHART, D. Notes on a schema for stories. In R. Bobrow & A. Collins (Eds.), Representation and understanding. New York: Academic, 1975.

RUMELHART, D. Introduction to human information processing. New York: Wiley, 1977. (a)

RUMELHART, D. Understanding and summarizing brief stories. In D. LaBerge & S.J. Samuels (Eds.), Basic processes in reading: Perception and comprehension. Hillsdale, N.J.: Erlbaum, 1977. (b)

RUMELHART, D. Toward an interactive model of reading. In S. Dornic (Ed.), Attention and performance VI. Hillsdale, N.J.: Erlbaum, 1977. (c)

RUMELHART, D. Analogical processes and pro­cedural representations. La Jolla: University of California Center for Human Information Processing Report 81, 1979.

RUMELHART, D. Schemata: The building blocks of cognition. In R. Spiro, B. Bruce, & W. Brewer (Eds.), Theoretical issues in reading comprehension. Hillsdale, N.J.: Erlbaum, 1980.

RUMELHART, D. & LEVIN, J. A language comprehension system. In D. Norman & D. Rumelhart, Explorations in cognition. San Francisco: Freeman, 1975.

RUMELHART, D. & NORMAN, D. The active structural network. In D. Norman.& D. Rumelhart, Explorations in cognition. San Fran­cisco: Freeman, 1975. (a)

RUMELHART, D. & NORMAN, D. The computer implementation. In D. Norman & D. Rumelhart, Explorations in cognition. San Francisco: Freeman, 1975. (b)

RUMELHART, D. & ORTONY, A. The representation of knowledge in memory. In R. Anderson, R. Spiro, & W. Montague (Eds.), Schooling and the acquisition of knowledge. Hillsdale, N.J.: Erlbaum, 1977.

SAMUELS, S.J. How the mind works when reading: Describing elephants no one has ever seen. In L.B. Resnick & P.A. Weaver (Eds.), Theory and practice of early reading. Hillsdale, N.J.: Erlbaum, 1979,343-368.

SCHANK, R. The structure of episodes in memory. In D. Bobrow & A. Collins (Eds.), Representation and understanding. New York: Academic, 1975. (a)

SCHANK, R. Conceptual dependency theory. In .R. Schank, N. Goldman, C. Rieger, & C. Riesback, Conceptual information processing. Amsterdam: North Holland, 1975. (b)

SCHANK, R. lnterestingness: Controlling inferences. New Haven: Yale Department of Computer Science Research Report 145, 1978.

SCHANK, R., & ABELSON, R. Scripts. Plans, goals, and understanding. Hillsdale, N.J.: Erlbaum, 1977.

SCHANK, R., GOWMAN, N., RIEGER, C., & RIESBECK, C. Conceptual information processing. Amsterdam: North Holland, 1975.

SCHANK, R., & KOLODNER, J. Retrieving information from an episodic memory, or: Why computers’ memories should be more like people’s. New Haven, Yale Department of Computer Science Research Report 159, 1979.

SCHANK, R., LEBOWITZ, M., & BIRNBAUM, L. lntegrated partial parsing. New Haven: Yale Department of Computer Sciences Research Report 143, 1978.

SCHANK, R., & WlLENSKY, R. Response to Dresher & Hornstein. Cognition, 1977. 5, 133-145.

SELFRIDGE, M. A process model of acquisition. New Haven: Yale Department of Computer Science Research Report 172, 1980.

SIMMONS, R., & CHESTER, D. Relating sentences and semantic networks with clausal logic. Austin: University of Texas Department of Computer Science Technical Report, 1979. In press.

SIMON, H. Artificial intelligence research strategies in the light of Al models of scientific discovery. Sixth International Joint Conference on Artificial Intelligence, 1979, 1086-1094.

SMITH, F. Understanding reading. New York: Holt, Rinehart, and Winston, 1971.

SPIRO, R. Remembering information from text: The “state of schema” approach. In R. Anderson, R. Spiro, & W. Montague (Eds.), Schooling and the acquisition of knowledge. HiIlsdaIe, N.J.: Erlbaum, 1977.

STEGMÜLLER, W. The structure and dynamics of theories. New York: Springer, 1976.

STEVENS, A., & RUMELHART, D. Errors in reading: Analysis using an augmented transition network. In D. Norman & D. Rumelhart, Explorations in cognition. San Francisco: Freeman, 1975.

THORNDYKE, P. Cognitive structures in comprehension and memory of narrative discourse. Cognitive Psychology, 1977, 9, 77-110.

THORNDYKE, P., & YEKOVICH, F. A critique of schemata as a theory of human story memory. Poetics, 1980, 9, 23-49.

TOWNSEND, J.T. Issues in models concerning the processing of a finite number of inputs. In B. Kantowitz (Ed.), Human lnformation Processing. Hillsdale, N.J.: Erlbaum, 1974.

TRIESMAN, A., & TUXWORTH, J. Immediate and delayed recall of sentences after perceptual processing at different levels. Journal of Verbal Learning and Verbal Behavior, 1974, 13, 38-44.

TULVING, E. Memory research: What kind of progress? In L. Nilsson (Ed.) Perspectives on memory research. Hillsdale, N.J.: Erlbaum, 1979.

TURNER, A., & GREENE, E. The construction and use of a prepositional text base. Boulder: University of Colorado Institute for the Study of Intellectual Behavior Technical Report 63, 1977.

VENEZKY, R.L., & MASSARO, D.W. The role of orthographic regularity in word recognition. In L.B. Resnick & P.A. Weaver (Eds.), Theory and practice of early reading. Hillsdale, N.J.: Erlbaum, 1979, 85-107.

WARREN, W., NICHOLAS, D., & TRABASSO, T. Event chains and inferences in understanding narratives. In R. Freedle (Ed.), New directions in discourse processing. Norwood, N.J.: Ablex, 1979.

WILENSKY, R. Understanding goal-based stories. New Haven: Yale Department of Computer Science Research Report l40, 1978. Also: New York: Garland Press, 1980.

WINOGRAD, T. Frame representation and the declarative-procedural controversy. In D. Bobrow & A. Collins (Eds.), Representation and understanding. New York: Academic, 1975.

WOODS, W. Transition network grammars for natural language analysis. Communications of the Association for Computing Machinery, 1970, 13, 591-606.

WOODS, W. Generalizations of ATN grammars. In W. Woods & R. Brachman, Research in natural language understanding. Cambridge, MA: Bolt, Beranek, and Newman Technical Report 3963, 1978.

WOODS, W., BATES, M., BROWN, B., BRUCE, B., COOK, C., KLOVSTAD, J., MAKHOUL, J., NASH-WEBBER, B., SCHWARTZ, R., WOLF, J., & ZUE, V. Speech understanding systems: Final report. Cambridge, MA: Bolt, Beranek, and Newman Technical Report 3438, 1976.

WOODS, W., & MAKHOUL, J. Mechanical inference problems in continuous speech understanding. Third International Joint Conference on Artificial Intelligence, 1973, 200-207.