Chapter V
Coherence
1. If meaning is used to designate the potential of a language expression (or other sign) for representing and conveying knowledge (i.e., virtual meaning), then we can use sense to designate the knowledge that actually is conveyed by expressions occurring in a text. Many expressions have several virtual meanings, but under normal conditions, only one sense in a text. If the intended sense is not at once clear, non-determinacy is present. A lasting non-determinacy could be called ambiguity if it is presumably not intended, or polyvalence if the text producer did in fact intend to convey multiple senses at the same time. Though not yet well explained, the human ability to discover intended senses and preclude or resolve ambiguities is one of the most amazing and complex processes of communication (cf. for example Hayes 1977).
2. A text “makes sense” because there is a continuity of senses among the knowledge activated by the expressions of the text (cf. Hörmann 1976). A “senseless” or “nonsensical” text is one in which text receivers can discover no such continuity, usually because there is a serious mismatch between the configuration of concepts and relations expressed and the receivers prior knowledge of the world. We would define this continuity of senses as the foundation of coherence, being the mutual access and relevance within a configuration of concepts and relations (cf. I.6). The configuration underlying a text is the textual world, which may or may not agree with the established version of the “real world” (cf. VII.18.1), i.e., that version of the human situation considered valid by a society or social group. Note, however, that the textual world contains more than the sense of the expressions in the surface text: cognitive processes contribute a certain amount of commonsense knowledge derived from the participants’ expectations and experience regarding the organization of events and situations. Hence, even though the senses of expressions are the most obvious and accessible contribution to the meaningful-ness of texts, they cannot be the whole picture.
3. Knowledge is not identical with language expressions that represent or convey it, though confusion on this point is rife in linguistics and psychology.1 This confusion arises from the enormous difficulty in envisioning and describing knowledge and meaning without constantly relying on language expressions. Many researchers agree that a language-independent representation would be highly desirable (cf. for example Schank, Goldman, Rieger & Riesbeck 1975). But so far, we cannot seem to agree on any representational mode already proposed. This stalemate is no mere accident: instead, it reflects the nature of the entities we are trying to systemize (in the sense of III.3).
4. As argued in I.6, a concept can be defined as a configuration of knowledge that can be recovered or activated with more or less consistency and unity. This definition is operational, based on the indisputable fact that language users, when employing or being confronted with a particular expression, tend to activate roughly the same chunk of knowledge (i.e. place the chunk in active storage, cf. III.26; IV.2).2 Variations among different language users do not seem to be substantial enough to occasion disturbances very often. It should follow from here that the meaning of a concept is the sum of its possible uses (cf. Schmidt 1968). Unfortunately, many concepts are so adaptable to differing environments that they remain quite fuzzy in regard to their components and boundaries.3 Therefore, defining concepts involves working with comparative probabilities: weaker or stronger likelihood that the concept will subsume certain knowledge when actualised in a textual world, where each concept appears in one or More relations to others, e.g. “state of”, “attribute of”, and so on (cf. V.26). These relations constitute the linkage which delimits the use of each concept.
5. If concepts can indeed subsume different knowledge elements according to the conditions of activation, then concepts cannot be primitive, monolithic units. Instead, concepts must have their own components held together by a particular strength of linkage. Components essential to the identity of the concept constitute determinate knowledge (e.g. all humans are mortal). Components true for most but not all instances of the concept constitute typical knowledge (e.g. humans usually live in communities). Components which happen to be true of random instances only constitute accidental knowledge (e.g. some humans happen to appear blond).4 As Loftus and Loftus (1976: 134) point out, this gradation is also fuzzy. Very few components, for example, turn out to be absolutely determinate: birds can be birds even if they can’t fly or if their feathers are stripped off; tables might have all kinds of shapes and all numbers of legs; and so on. Labov (1973) tested the borderlines where people were willing to call a presented shape a ‘cup’ as opposed to other vessels (‘jar’, etc.), and found only partial agreement. Still, some gradation in strength of linkage is probably necessary if concepts are to be operational. The concept is, after all, designed to handle normal instances rather than bizarre counter-examples dreamed up in peculiar situations (e.g. philosophers’ debates).
6. It is one thing to agree that concepts can be decomposed into more basic units; it is quite another to agree upon what those units might be (cf. le Ny 1979).5 Even straightforward cases can become entangled in irresolvable debates. For example, it ought to be fully reasonable to view the concept ‘kill’ as composed of ‘cause’, ‘become’, ‘not’, and ‘alive’; yet even here, controversy rages. And texts can be found where this simple analysis fails utterly:
[83] And though I kill him not, I am the cause
His death was so effected.
(All’s Well That Ends Well III ii 118-19)
Evidently, the components of concepts are themselves not entirely stable, whether they be called “features”, “markers”, “primitives”, “semes”, “sememes”, or whatever.6
7. Even if we could agree on the units which constitute concepts, we would not have shown that the decomposition of concepts is a routine activity of text processing. Evidence for such routines is at present slight (Kintsch 1974: 242; J. Anderson 1976: 74; Hayes-Roth & Hayes-Roth 1977). And the unsettled questions are alarming. How many units would be needed for all possible concepts? Would the same set of units work for concepts and expressions? Given that people communicate via expressions, how are units acquired? How can we define the units without having recourse to the same kinds of expressions or concepts that we are trying to decompose? Are there units which, in the worst case, are needed for only one concept or expression in the whole language?
8. Perhaps it would be more productive to try working in the other direction: rather than asking how expressions or concepts can be cut into the tiniest possible pieces, we might inquire how expressions are assigned conceptual senses, and how senses are put together into large configurations of a textual world. Certainly, the building of textual worlds is a documented routine activity in human communication. This reversal of outlook would shift attention away from questions which adjudication a priori cannot solve (such as those in V.7) toward questions which can be pursued empirically (e.g. via reading and recall of texts, cf. IX.24ff). The fuzziness and instability of concepts and their possible components should become steadily less prominent when they appear in more and more determinate contexts of communication. In that perspective, the sense of an expression or the content of a concept are definable as an ordered set of hypotheses about accessing and activating cognitive elements within a current pattern. To describe such a sense or content, one would have to stand at that point in the configuration of concepts and relations and look out along all pathways (cf. Quilban 1966).9 The study of language meaning via this approach is the concern of a recent trend known as procedural semantics (cf. Miller & Johnson-Laird 1976; Winograd 1976; Johnson-Laird 1977; Levesque 1977; Schneider 1978; Levesque & Mylopoulos 1979). It is recognized that in addition to declarative knowledge (statements of facts or beliefs about the organization of events and situations in the “real world”), communication requires procedural knowledge (facts or beliefs stated in formats intended for specific types of uses and operations) (cf. Winograd 1975; Winston 1977: 390ff; Goldstein & Papert 1977; Bobrow & Winograd 1977). The meaningfulness of language in texts is just a special case of the acquisition, storage, and utilization of knowledge in all manner of human activities. Since language use is highly differentiated and reasonably well regulated by social agreement, the special case is perhaps the most promising approach to the general one (cf. X.7).
10. When expressions are used in communication, the corresponding concepts and relations are activated in a mental workspace we can hence term active storage (cf. III. 29; IV. 2; V.4). George Armitage Miller (1956) reported that this workspace seems limited to only about seven items at a time. It follows, he observed, that efficiency would be promoted if the items were large, well-integrated chunks of knowledge rather than single, unrelated elements. Consequently, the knowledge that underlies textual activities would normally figure as global patterns which are matched and specified to accommodate current output (in production) and input (in reception) (cf. V. 16). The difficulty in processing non-expected or discrepant occurrences (cf. VII.13) presumably arises because they cannot be handled as parts of well-integrated stored patterns and must be held separately in active storage until they can be fitted in and accommodated.
11. These patterns of knowledge might well look different according to the demands of current processing tasks. Text receivers would use patterns for building and testing hypotheses about what the major topic is (cf. V.23) and how the textual world is being organized. It follows that the topic pattern would be utilized more richly than subsidiary patterns of marginal usefulness for the text at hand (cf. V.16). Another scale of difference would be the importance and relevance of the text for the receiver’s situation: as these factors rise, utilization of knowledge would become more detailed and thorough (cf. III.3 1).
12. When some item of knowledge is activated, it appears that other items closely associated with it in mental storage also become active (though perhaps not so active as the original item). This principle is often called spreading activation (see Collins & Loftus 1975) and mediates between the explicitly activated concepts or relations and the detailed richness which a textual world can assume. In production, spreading activation might work outward from concepts or relations toward natural language expressions that could be preferentially used (cf. III.23). In reception, spreading activation makes it possible to form elaborate associations, to create predictions and hypotheses, to deploy mental images, and so forth, far beyond what is actually made explicit in the surface text. Determinate and typical knowledge should be especially prone to spreading activation (cf. V.5), though accidental knowledge might also be involved if imprinted forcefully enough in one’s own experience.
13. There is some evidence of two different principles of storing and utilizing knowledge. Endel Tuiving (1972) introduced the notions of episodic memory vs. semantic memory to account for the distinction. Episodic memory contains the records of one’s own experience (‘what happened to me’), while semantic memory at least in the most appealing sense of the term7 reflects the inherent patterns of the organization of knowledge, e.g. the structures of events and situations (‘what is true about the world at large and how it all fits together’). Of course, one’s experiences continually feed into one’s general views about the world, while the latter impose organization upon experience. Still, episodic knowledge would be heavily tied to the original contexts of encounter and would thus manifest many accidental traits. Semantic knowledge, in contrast, would be more dominantly organized in terms of the characteristics which all or most individual instances have in common.
14. Since the times of Plato and Aristotle up through the Middle Ages even into the present, the comparative importance of experience vs. human reasoning powers in the acquisition of knowledge has been hotly debated. Whether concepts can exist independently of all particular instances of them (as Plato believed), or whether they must all be extracted from personal experience (as empiricists asserted), are questions which may be irresolvable in the framework of the usual discussions. Any view which denies either innate human reasoning powers or the effects of real experience would prove untenable if subjected to unbiased comprehensive investigations of human conduct—a recourse which generations of philosophers hardly seem to have considered. The utilization of texts almost certainly involves steady interactions and compromises between the actual text materials being presented, and the participants’ prior disposition, according to conditions which, though flexible and variable, are by no means unsystematic (cf. discussion in IX.37ff.).1
15. In a procedural approach, arguments in favour of one model of knowledge over another should be couched in terms of tasks and operations. Consider for example the question of economy. On the one hand, each item of knowledge might be stored in a system only once, no matter how many configurations would contain the item. There would be either a very dense interlocking of configurations, or else a given configuration would have to be assembled every time need arose. This kind of system offers great economy of storage, but heavy expenditure on search. On the other hand, items could be redundantly stored in each of the configurations which include them. This system would work very rapidly on search, but would be horrendously wasteful on storage. As Walter Kintsch (1977a: 290f.) notes, this trade-off between economy of storage and economy of search is probably resolved by compromise. Frequently used configurations would be stored as wholes, in spite of the redundancy involved; unusual, seldom required configurations would be assembled via searching out component items only when occasion arises.
16. Some types of global patterns would be stored as complete chunks because of their usefulness in many tasks. Frames are global patterns that contain commonsense knowledge about some central concept, e.g. ‘piggy banks’, ‘birthday parties’, etc. (cf. Charniak 1975b; Minsky 1975; Winograd 1975; Petbfi 1976; Scragg 1976; Metzing (ed.) 1979). Frames state what things belong together in principle, but not in what order things will be done or mentioned. Schemas are global patterns of events and states in ordered sequences linked by time proximity and causality (cf. Bartlett 1932; Rumelhart 1975, 1977b; Kintsch 1977b; Mandier & Johnson 1977; Rumelhart & Ortony 1977; Spiro 1977; Thorndyke 1977; Kintsch & van Dijk 1978; Beaugrande & Colby 1979). Unlike frames, schemas are always arrayed in a progression, so that hypotheses can be set up about what will be done or mentioned next in a textual world. Plans are global patterns of events and states leading up to an intended goal (cf. Sussrnan 1973; Abelson 1975; Sacerdoti 1977; Schank & Abelson 1977; Cohen 1978; McCalla 1978; Wilensky 1978a; Allen 1979; Beaugrande 1979a, b). Plans differ from schemas in that a planner (e.g. a text producer) evaluates all elements in terms of how they advance toward the planner’s goal. Scripts are stabilized plans called up very frequently to specify the roles of participants and their expected actions (cf. Schank & Abelson 1977; Cullingford 1978; McCalla 1978). Scripts thus differ from plans by having a pre-established routine. The importance of these kinds of global patterns has become recognized in the procedural attachment of producing and receiving texts: how a topic might be developed (frames), how an event will progress in a sequence (schemas), how text users or characters in textual worlds will pursue their goals (plans), and how situations are set up so that certain texts can be presented at the opportune moment (scripts). Different pattern types might share the same basic knowledge in a variable perspective (e.g. a frame for ‘structure of a house’ versus a plan for ‘building a house’). Using global patterns greatly reduces complexity over using local ones, and allows retaining much more material in active storage at one given time. We provide some demonstrations later on.8
17. A further issue in procedural models of knowledge is that of inheritance: the transfer of knowledge among items of the same or similar types or sub-types (cf. Falhman 1977; Hayes 1977; Brachman 1978; Levesque & Mylopoulos 1979). At least three kinds of inheritance should be noted. First, an instance inherits all the characteristics of its class9 unless expressly cancelled (Fahlman 1977). We assume that Napoleon had toes, to use a familiar example from Walter Kintsch (1974), even though nobody (but Walter) has ever told us so, because Napoleon is an instance of the class ‘human beings’. If he had had no toes, there would doubtless be some historical anecdote to cancel our assumption. Second, subclasses inherit from superclasses only those characteristics that the narrower specification of the subclasses allows. For example, the subclass of ‘ostriches’ differs from the superclass of ‘birds’ in being unable to fly, but able to run extremely fast. Third, entities can inherit from those with which they stand in analogy, i.e. they are of different classes but comparable in some useful respects. For instance, researchers in cognitive science and artificial intelligence are making assumptions about the human mind by analogy to the computer (cf. X.26ff.). Without claiming that minds and computing machines are the same thing, we can still discover comparable characteristics that are helpful in building complex models of cognition.
18. Inheritance relates to the consideration of economy raised in V.15. If knowledge about classes/instances, subclasses/ superclasses, or analogies were stored in a neat hierarchy, predictions should be possible about time needed to access certain facts. For example, [84a] should take longer to judge “true” or “false” than [84b] because the superclass ‘animal’ is higher up in the hierarchy than the subclass ‘bird’, and thus to connect them demands at least one more step:
[84a] A chicken is an animal.
[84b] A chicken is a bird.
However, testing failed to confirm such predictions (cf. Collins & Quillian 1972). For one thing, [84c] was regularly confirmed faster than [84b], though ‘chicken’ and ‘robin’ should be on the same plane in the hierarchy:
[84c] A robin is a bird.
Smith, Shoben, and Rips (1974) explain this effect in terms of “features” as basic components of concepts like ‘bird’: the more determinate and typical features (cf. V.5) an instance or subclass has, the quicker it will be judged a member of a class or superclass. ‘Robins’, who fly and sing well, are thus easier to judge as ‘birds’ than are ‘chickens’, who do not. In a like fashion, people are more prone to misjudge [84d] as true than [84e]:
[84d] A bat is a bird.
[84e] A stone is a bird.
because of the shared feature ‘can fly’ which sets ‘bats’ and ‘birds’ into an analogy. Rosch and Mervis (1975) argue that “family resemblances” are responsible for such effects rather than defining features, because it is extremely difficult in many cases to decide what features every member of a class must have (cf. examples in V. 5).
19. We can readily see that the procedural considerations we have outlined—activation (V.4, 10), strength of linkage (V.5), decomposition (V.6-7), spreading activation (V.12), episodic vs. semantic memory (V.13), economy (V.15), global patterns (V.16), and inheritance (V.17-18)—all depend upon each other. They all must be treated in terms of whatever are taken as the basic units and operations upon knowledge. A very simple, limited model might accommodate the results of experiments on judging sentences like [84a] through [84e] and still tell us little about that larger question (Kintsch 1979b). Symptomatic of this disparity is the attempt to separate off a neatly organized “lexicon” or “dictionary” of words or concepts from the vast, messy maze of an “encyclopedia” of world knowledge (cf. Smith 1978). As Kintsch (1979b) points out, such a separation is a research fiction which hinders the development of really powerful, general models and breaks down eventually in view of a wider range of realistic data.
20. From here, some basic conclusions can be drawn. First of all, instead of trying to cut language off from everything else, we should strive to build models in which the use of language in real texts is explainable in terms comparable to the processes of apperception and cognition at large (cf. Minsky 1975; Miller & Johnson-Laird 1976; Kintsch 1977a; Rumelhart 1977a; Beaugrande 1980a). The restrictions upon research which reduce all issues to a matter of variations in time for performance on unrealistic tasks (including sentence judgements along the lines depicted in V.18) run counter to the main incentive for such an undertaking. We must work toward a diversity of experiment types among which the everyday utilization of texts plays a leading role.
21. A second conclusion is that efforts to encompass the study of texts and knowledge into the framework of logic since Aristotle may prove a mixed blessing. We should rather reverse our priorities by first building humanly plausible models and then inquiring after types of logic that can serve as formalisms (Petöfi 1978: 44f.). Humans are evidently capable of intricate reasoning processes that traditional logics simply cannot explain: jumping to conclusions, pursuing subjective analogies, and even reasoning in absence of knowledge (Collins 1978). For example, when confronted with a possible fact, people might say to themselves: ‘If this were true, I ought to know about it; since I don’t know, it is probably false’— the lack of knowledge inference described by Collins. The important standard here is not that such a procedure is logically unsound, but rather that the procedure works well enough in everyday affairs.
22. A third conclusion is that, as we have already stressed (V.8), knowledge and meaning are extremely sensitive to the contexts where they are utilized. We would like to pursue some implications of that view for a candidate model of text coherence. Basically, the combination of concepts and relations activated by a text can be envisioned as problem-solving in the sense of III.17. Given some fuzzy, unstable units of sense and content, text users must build up a configuration of pathways among them to create a TEXTUAL WORLD (V.2). Only certain characteristics or "features" of the concepts involved are really necessary and relevant for these operations. Such procedures as decomposition, spreading activation. inferencing, and inheritance will be carried out in accordance with the current conditions of processing. The central question is how to classify and systemize those conditions (and not how to prove that all text users do the same things all the time). In this line of inquiry, we could ask: how do people extract and organize content from texts for use in storing and recalling? What factors of the interaction between the presented text and people's prior knowledge and disposition affect these activities? What regularities can be uncovered by varying factors such as the style of the surface text or the user groups to whom the text is presented? What is the role of expectations?
23. An initial step toward exploring these and similar questions is to find a basic representation for the coherence of texts. We shall suggest at least one possible means analogous to our proposals for a procedural model of syntax in IV.5-10. Coherence will be envisioned as the outcome of combining concepts and relations into a network composed of knowledge spaces centred around main topics. Our demonstration text will be the 'rocket'-sample [4] from I.i, used already in our brief discussion of cohesion (IV.7ff., 24, 29) as well as in some earlier research.10
24. Before tackling the text itself, we should call to mind the requirements for representing the processing of texts. We focus now on reception rather than production, though, as we stressed in III.29, there are undoubtedly important similarities between the two activities. The imposition of coherence on any stretch of text should be performed along the lines suggested in III.29ff. The surface text is parsed onto a configuration of grammatical dependencies, as depicted in IV.5-10. The surface expressions are taken as cues to activate concepts (V. 4, 10). This phase cannot involve a straightforward lookup in a mental “dictionary” (cf. V.19). Instead, the concepts are treated as steps in the construction of a continuity of sense (V.2), and the extent of processing expended will vary according to whatever is required and useful for that task. Attention would be directed particularly toward the discovery of control centres, i.e. points from which accessing and processing can be strategically done.
25. The most likely candidates for control centres can be termed primary concepts:
(a) objects: conceptual entities with a stable identity and constitution;
(b) situations: configurations of mutually present objects in their current states;
(c) events: occurrences which change a situation or a state within a situation;
(d) actions: events intentionally brought about by an’ agent.11
26. The other concepts would be assigned to a typology of secondary concepts. The following set is taken from Beaugrande (1980a), where a more elaborate justification is offered:
(a) state: the temporary, rather than characteristic, condition of an entity;
(b) agent: the force-possessing entity that performs an action and thus changes a situation (cf. V.25(d));
(c) affected entity: the entity whose situation is changed by an event or action in which it figures as neither agent nor instrument;
(d) relation: a residual category for incidental, detailed relationships like ‘father-child’, ‘boss-employee’, etc.,
(e) attribute: the characteristic condition of an entity (cf. “state”);
(f) location: spatial position of an entity;
(g) time: temporal position of a situation (state) or event (cf. I.10);
(h) motion: change of location;
(i) instrument: a non-intentional object providing the means for an event;
(j) form: shape, contour, and the like;
(k) part: a component or segment of an entity;
(l) substance: materials from which an entity is composed;
(m) containment: the location of one entity inside another but not as a part or substance;
(n) cause: see 1.7;
(o) enablement: see 1.7;
(p) reason: See 1.8;
(q) purpose: See 1.9;
(r) apperception: operations of sensorially endowed entities during which knowledge is integrated via sensory organs;12
(s) cognition: storing, organizing, and using knowledge by sensorially endowed entity;
(t) emotion: an experientially or evaluatively non-neutral state of a sensorially endowed entity;
(u) volition: activity of will or desire by a sensorially endowed entity;
(v) recognition: successful match between apperception and prior cognition;
(w) communication: activity of expressing and transmitting cognitions by a sensorially endowed entity;
(x) possession: relationship in which a sensorially endowed entity is believed (or believes itself) to own and control an entity;
(y) instance: a member of a class inheriting all non-cancelled traits of the class (cf. V. 17);
(z) specification: relationship between a superclass and a subclass, with a statement of the narrower traits of the latter (cf. V. 17);
(aa) quantity: a concept of number, extent, scale, or measurement;13
(bb) modality: concept of necessity, probability, possibility, permissibility, obligation, or of their opposites;
(cc) significancie: a symbolic meaning assigned to an entity;
(dd) value: assignment of the worth of an entity in terms of other entities;
(ee) equivalence: equality, sameness, correspondence, and the like;
(ff) opposition: the converse of equivalence;
(gg) co-reference: relationship where different expressions activate the same text-world entity (or configuration of entities) (cf. IV.21);
(hh) recurrence: the relation where the same expression reactivates a concept, but not necessarily with the same reference to an entity, or with the same sense (cf. IV. 12-15).14
27. Most of these concept types are familiar from “case grammars”15 that undertook to classify language relationships according to the organization of events and situations (cf. Fillmore 1968, 1977; Chafe 1970; Grimes 1975; Longacre 1976; Frederiksen 1977). At some point, these schemes tend to become a classification of knowledge and its organization, reflected in other domains besides language (cf. Kintsch 1974; Charniak 1975a; Schank et al. 1975; Woods 1975; Wilks 1977b). We incorporate some further concepts to encompass mental operations (apperception, cognition, emotion, volition, communication, possession), class inclusion (instance, specification), and notions inherent in systems of meaning per se (quantity, modality, significance, value, equivalence, opposition, co-reference, recurrence). We do not claim that this typology is exhaustive, or superior to others proposed before. It is merely useful for labeling the links among concepts, e.g. that one concept is the “state of” another, or the “agent of” another, etc.; and through various combinations, we can capture the notions of other typologies we have examined so far. One might easily work with typologies having greater or lesser detail than ours.16
28. In addition to a typology of concepts for labelling links, we might need a set of operators which further specify the status of linkage. For example, we could introduce operators for strength of linkage in the sense of V. 5: (a) a determinateness operator [d] for components necessary to the identity of a concept; and (b) a typicalness operator [t] for frequent, but not necessary components. These operators would apply to configurations of world knowledge, as shown in V.39. Furthermore, we could introduce operators for linkage which involves boundaries: (a) an initiation operator [i] for an entity ‘just being created or enacted; (b) a termination operator [†] for the converse; (c) an entry operator [ε] for an entity coming about on its own; and (d) a exit operator [χ] for the converse of entry. Finally, two operators are useful for dealing with approximative or contrary-to-fact linkage: (a) the proximity operator [π] for relationships with some distance or mediation (cf. temporal proximity in 1. 10, causal proximity in V.36, etc.); and (b) the projection operator [ρ] for relations which are possible or contingent, but not true in the textual world (cf. IV.48). To distinguish operators from the link labels formed out of the beginning letters of concept names (e.g. “ca” for cause, “ti” for time, etc.), operators are Greek letters from the first or second position of the respective words, joined to the other labels with the sign “¸”.17
29. The motives and applications for our typologies presented above should become clearer through a demonstration. We start with the opening paragraph of our ‘rocket’ text:
[4] [1.1] A great black and yellow V-2 rocket 46 feet long stood in a New Mexico desert. [1.2] Empty, it weighed five tons. [1.3] For fuel it carried eight tons of alcohol and liquid oxygen.
The control centre for this passage is clearly the object concept ‘rocket’, to which are assigned attributes (‘great’, ‘black’, ‘yellow’, ‘long’), a specification (‘V-2’), and a state (‘stood’) with its locations (‘New Mexico’, ‘desert’); the attribute ‘long’ has the quantities ‘46’ and ‘feet’. We could put all of these conceptual relations into a network such as that shown in Figure 6.
Link labels announce the type of concept that is attained by traversing links in the directions shown by the arrows. The operations would be comparable to those depicted for the transition networks in IV.5-10. The processor works from a current state to a following state by trying to identify the type of the node to be attained. The strategies of problem-solving, (III.17) would apply, assisted by spreading activation (V.12), inferencing (cf. V. 32ff ), and global patterns (V.16).
30. It is important to compare and contrast the conceptual network in Fig. 6 with the grammatical network in Fig. 4 in IV.10. Although we still use English words in the notation of Fig. 6, we are now representing concepts rather than surface expressions. It might be desirable to have some other representation, but at present, researchers are manifestly unable to agree on any one. Notice that the general pattern of the two networks is similar: the access routes from node to node are much the same. Hence, it seems reasonable that text processing should make use of structural similarities on different levels as far as is expedient (cf. R. Bobrow 1978; Walker (ed.) 1978; Woods & Brachman 1978). For example, a hypothesis that grammatical heads are usually primary concepts would be confirmed often enough to merit its general application. Similarly, one could postulate that grammatical modifiers are attributes, states, locations, etc., in a certain preference order (on preferences, cf. III.18) as indicated by the nature of the primary concept at the control centre. Such hypotheses and preferences could serve to augment the transitions among nodes in the sense of IV.5. Where possible, the uncovering of grammatical and conceptual dependencies would interact heavily or even run in parallel rather than as two separate phases, though there would nearly always be some asymmetry involved, since the grammatical repertory is smaller than the conceptual (cf. III.18, 25). In other terms: the problems on one level could be solved with the aid of more readily solvable or already solved pathways on another level.
31. Another distinction between the two network types just discussed is the stretch of text that they represent. It seems most unlikely that people would build grammatical networks for any whole texts other than very short ones.18 The standard procedure would more probably be to build grammatical networks only for a convenient span of text retainable in active storage while the conceptual network is being constructed; hence, only the conceptual network would be assembled for the entire text. The whole paragraph of our ‘rocket’-sample would be easy to assemble into a coherent knowledge space a conceptual macro-state in which concepts are micro-states (cf. IV.6) because the ‘rocket’ concept itself underlies something in each stretch of text. Thus, the addition of ‘Empty, it weighed five tons’ and ‘For fuel, it carried eight tons of liquid oxygen’ would merely involve attaching more content viz. states, quantities, containment, substances, and so on to the already created ‘rocket’ node. Figures 7a and 7b show the knowledge space first with separate configurations for sentence-length stretches, and then as an integrated unit.
Of course, the pro-form ‘it’ is at once suppressed, since its content is derivative from the co-referent ‘rocket’. Possibly, no such ‘it’ node is ever set up in such a straightforward case; the new material is immediately connected to the proper node of the co-referent.19 In this manner, cohesion (e.g. pro-forms) supports coherence.
32. The integration of the underlying configuration of the next paragraph is more intricate:
[4] [2. 1] Everything was ready. [2.2] Scientists and generals withdrew to some distance and crouched behind earth mounds. [2.3] Two red flares rose as a signal to fire the rocket.
Here, there are no noticeable cohesive devices among the sentences. Nor is the underlying coherence at once obvious. A state of ‘readiness’ is mentioned, followed by two kinds of motion events (‘withdraw/crouch’, ‘rise’). To bind things together, inferencing must be done (cf. 1.11). This operation involves supplying reasonable concepts and relations to fill in a gap or discontinuity in a textual world. In contrast to spreading activation (V. 12), which ensues without specific demand, inferencing is always directed toward solving a problem in the sense of III.17: bridging a space where a pathway might fail to reach.20 Reasonable inferences for our sample would be that the ‘ready’state was the “reason” for the motions; that ‘everything’ subsumes whatever was needed to “enable” the ‘take-off’ of the ‘rocket’; and that the ‘scientists’ and ‘generals’ were present in order to ‘observe’ the ‘rocket’.
If we now add this second model space to the first, with inference nodes placed in square brackets, we obtain the pattern given in Figure 8. Again, the links have labels explained in the key and marked with directional arrows.
33. Two possible objections should be noted here. First, it might be protested that the inferences we admitted are chosen arbitrarily. Yet, although our own intuition played some role in suggesting the inferences, they were heavily confirmed on empirical tests (discussed in IX.34), where readers reported them as part of what they had read. In one group of 72 readers, for instance, no less than 24 recalled the scientists ‘observing’ the rocket. Such a result suggests that the distinction between concepts directly activated by text expressions and concepts supplied for evident discontinuities may not be so clear-cut as we would like to suppose. Very few of the text-activated concepts fared as well in our tests as the inferred ‘observe’. Perhaps it would be expedient to assign probability values to inference nodes and links, but, if so, this might well have to be done for the text-activated nodes and links also (cf. Beaugrande 1980b).
34. The second objection might be that the inferences we admit are rather too few than otherwise. Text users could make many more: that the ‘fuel’ will burn, so that ‘scientists’ and ‘generals’ must seek shelter behind non-flammable ‘earth mounds’; that a count-down must be going on here somewhere; that the rocket is involved in an experiment; that the rocket’s location will be updated to its peak altitude (cf. IX. 26); and so forth. Such materials might well be supplied via spreading activation (V.12) without any particular searching. At present, it seems justified to distinguish between additions occasioned by problems (cf. Chamiak 1976) and those arising from a natural tendency to fill in situations or event sequences in general. Later on, we may hope to find out whether that distinction is psychologically plausible, i.e. whether text users have a uniform threshold for noticing and filling discontinuities and gaps. We may learn that large numbers of rather trivial inferences are being made but not reported, just as the text producer might have said a great deal more but saw no reason to do so. The question would then be: how similar are the textual worlds of the producer and that of the typical receiver? Do they, for instance, agree on what is or not worth mentioning? Are there major differences in the richness of their mental representations for text-world situations and events? Just now, these questions are far from answered; but accumulating results with the ‘rocket’-text indicate that the uniformity of different people’s textual worlds is at best approximative though still reasonably reliable.
35. The third paragraph surpasses the first in its use of cohesive devices, here: recurrence (‘flame’, ‘faster’, ‘yellow’), paraphrase (‘rose ... faster and faster’-’sped upward’), and pro-forms (‘it’):
[4] [3. 1] With a great roar and burst of flame the giant rocket rose slowly at first and then faster and faster. [3.2] Behind it trailed sixty feet of yellow flame. [3.3] Soon the flame looked like a yellow star. [3.4] In a few seconds, it was too high to be seen, [3. 5] but radar tracked it as it sped upward to 3, 000 mph.
The placement of ‘it’ is not always very clear, for example in [3.4] and [3.5], where possible co-referents might be ‘flame’
or ‘star’. Still, the tendency would be to attach doubtful pro-forms to the topic node, which is of course ‘rocket’ (cf. V.38).
36. The model space for this paragraph might look some-thing like Figure 9. The ‘rising’ motion of
the ‘rocket’ is the proximate cause of the ‘roar’ and ‘burst’ and has as quantities of motion ‘slowly’ and ‘faster and faster’ (these quantities being temporally proximate to each other). In locational proximity to the ‘rocket’, the ‘flame’ with its attribute ‘yellow’ and quantities ‘sixty feet’ has the motion ‘trail’ and proximity of apperception (‘look like’) to a ‘star’. The ‘yellow’ attribute is inherited via analogy from ‘flame’ to ‘star’ (cf. V.17). In the quantity of time ‘few seconds’, the ‘rocket’ is (state) at the location ‘high’ whose quantity ‘too’ is the cause that ‘seeing’ has the modality ‘not’ (inferred here). In opposition to that lack of apperception, the ‘radar’ still apperceives (‘tracks’) the ‘rocket’ in temporal proximity to the latter’s ‘speeding’ with a location of motion ‘upward’ and quantities of motion ‘3,000 mph’. If we now connect this whole model space to the previous configuration, merging the ‘rocket’ nodes and assuming that the ‘observers’ were those that could ‘not see’ the rocket at great heights, we obtain Figure 10.
37. The single sentence of the final paragraph is cohesive with the preceding text via recurrences (‘fired’, ‘speed’, ‘mph’, ‘saw’) and the pro-form ‘it’ still co-referring with ‘rocket’. This ‘it’ occurs at considerable distance from ‘rocket’ (after three intervening sentences), but is again likely to be connected to the topic node as before (V.35). We can accordingly hook on this closing material by affixing the motions ‘return’ and ‘plunge’ along with their dependent quantities, locations, etc. to the original ‘rocket’ node, giving us the final text-world model in Figure 11.
Notice that we have to use arrows pointing to links this time to render quantities of time and location (‘a few minutes after’, ‘forty miles from’), because these concepts depend here on relations rather than on other concepts.21
38. This network representation for the sense of an entire text may look unduly elaborate. Yet it offers a useful topography for studying such questions as density of linkage as a manifestation of topic, and typical operations in recall or summary as a matching of patterns. Moreover, it is probably far less elaborated than human receivers’ mental representations with inferencing, spreading activation, updating in short, the total outcome of applying prior knowledge of the world.22 To be truly comprehensive, we might have to include time parameters for what is true at any moment
(e.g. after the location ‘in a New Mexico desert is no longer true until the ‘plunge into earth’) and probabilities for a wealth of incidental inferences (cf. V.34). Soon, the model would be in danger of exploding into unmanageable diffuseness and complexity. It would be better to limit the text-world model to only those concepts directly activated by text expressions and to only those inferences without which any portion of the model would not be connected to the rest at all. We shall propose to include world-knowledge being matched against the textual world in a pattern called the world-knowiedge correi.ate; and global knowledge in general patterns such as a schema about ‘flight’ that attaches text entries (cf. IX.25-8).
39. The world-knowledge correlate for our sample might look something like Figure 12, if we try to keep the same basic proportions for those elements also shown in Fig.11. Other elements are in square brackets: this material would be supplied by spreading activation or, if discontinuities were apperceived, by inferencing. We undertake as a suggestion at least to distinguish definite from typical linkage with the operators proposed in V. 28.23 For example, ‘burning’ must have ‘fuel’ and cause ‘heat’. A ‘signal’ must be ‘noticeable’ if it is to be ‘observed’. ‘Seconds’ are parts of ‘minutes’ by definition. In contrast, most links are merely typical: the ‘isolated’ ‘site’ for ‘take-off’ located in a ‘desert’ of ‘New Mexico’; the ‘explore’ activities of ‘scientists’ and the ‘attack’ activities of ‘generals’ via the instrument ‘rocket’; the locations, substances, and attributes of ‘shelter’ required for ‘danger’; the purposes and attributes of ‘flares’ and ‘signals’; and so forth, as shown in the diagram. Though fuzzy (V.5), the distinction between determinate and typical knowledge does appear useful. We also include recurrences in the sense of V.26(hh), rather than putting them in a separate graph. However, the global world-knowledge belonging to the notion of ‘flight’ as the overall framework for our text will be reserved for treatment under the notion of “schema” in IX.25-8.40. We have presented our text-world model without clarifying the notion of reference, despite the prominence of
that notion in many philosophical theories of meaning.24 In older semantics, it was hoped that meaning could be explained in terms of the “conditions” under which statements (misleadingly called “sentences”) are “true”. Thus, to know what something means is to know how to “verify” its “truth”. This viewpoint, sometimes called “verificationism”,25 has unpleasant implications: for one thing, it is obviously wrong that people cannot understand a statement unless they can tell whether it is true; for another, people have no such immediate access to the “truth” as is being implied here. In the line of enquiry we are pursuing, however, the text-world is constructed from cognitive content (“knowledge”)26 collated against one’s beliefs about the “real world” in a complex and often approximative manner. Hence, rather than saying that “words refer to objects” or the like, we prefer to say that “expressions activate knowledge”. The act of referring is then an intricate process of pattern-matching, during which text users may decide that a text-world failing to match at a given threshold is fictional. There are numerous contingent factors that can influence this act of referring: type and purpose of text; importance of the text and its implications for one’s situation; the believability of the text producer as encountered in past experience; and the topic materials in the textual world. Empirical research on these matters is still rare.
41. This chapter has been concerned with the means for exploring and representing coherence as the outcome of actualising meanings in order to make “sense”. To investigate human activities with texts, we should treat meaning and sense in terms of procedures for utilizing knowledge in a wide range of tasks. In that outlook, issues like these emerge: continuity (V.2), activation (V.4, 10), strength of linkage (V.5), spreading activation (V.12), episodic vs. semantic memory (V.13), economy (V.15), use of global patterns (V.16), inheritance (V.17f.), and compatibility between language in texts and apperception or cognition at large (V.20). Whereas the meanings of expressions or the content of concepts are highly disputable in isolation, their occurrence within a textual world where processing must be performed should be reasonably stabilizing and delimiting. We presented a mode for observing the construction of a text-world model for a sample, hoping to suggest and illustrate at least a few major factors worth pursuing (V.23-40). We pointed out some cases where prior knowledge affects text processing in this manner (for more discussion, cf. IX.31ff)
42. The study of coherence along such lines does not, of course, promise to be simple. But it is quite conceivable that the questions traditionally posed and disputed regarding meaning and sense are otherwise quite unanswerable. Certainly, dogmatic insistence upon extreme views, typical of so many discussions among philosophers and psychologists in the past, should yield to a flexible, realistic modelling of the diverse but systematic strategies people actually apply when using texts in everyday life.
Notes
1 We use the term “knowledge” throughout to designate cognitive content of all kinds, as opposed to “meaning” and “sense” of expressions (cf. discussion in V.1f). Failure to make such a distinction leads to blurring the fact that knowledge must be selected and processed before it can be expressed and communicated.
2 The term “chunk” goes back to Miler (1956): a knowledge configuration processed as a unified block (cf. V.10).
Bock (1979) points out that psycholinguistic theories have all too often made little use of memory research.
3 Cf. Rosch (1973); Kintsch (1977a: 292ff.). One important use of “fuzzy set theory” has been the treatment of fuzzy concepts (cf. Zadeh 1972, 1975, 1979; Eikmeyer & Rieser 1979).
4 Smith, Shoben & Rips (1974) use the notions of “defining” and “characteristic features” similar to our notions here, though they insist that their model be based on set theory rather than on networks (see refutation in Holland 1975). Rosch & Mervis (1975) argue instead in favour of “family resemblances”, because “defining features” are often hard to discover (cf. V.18 and note 23 to this chapter).
5 The untenability of the initial proposals of Katz & Fodor (1963) has been repeatedly demonstrated (cf. Bolinger 1965; Hörmann 1976). We suspect that the whole undertaking of stipulating the units themselves may be misguided and hopeless; we can at best stipulate unit types, e.g. “attributes”. See X.5 for some wider implications.
6 The terms “features” and “markers” were used by Katz & Fodor (1963); primitives” by Wilks (1977a); “semes” in Greimas (1966); and “sememes” in Koch (1971). See the general survey in le Ny (1979).
7 Concerning some disadvantageous senses of the term, cf. Schank (1975); Kintsch (1979b). The term “conceptual memory” might be more useful (cf. Rieger 1975; Beaugrande 1980a).
8 Cf. VI.11-20; V.38; IX.25-.8
9 bA “cclass” is a group of entities sharing some characteristic, whereas a “set” is simply defined by the members it has. Set theory has been proclaimed as a foundation for the study of meaning (e.g. by Smith et al. 1974), but perhaps it only passes over rather than settles the real question: what classificatory procedures are used to form sets in the first place?
10 Cf. McCall & Crabbs (1961); Miller & Coleman (1967); Aquino (1969); Kintsch & Vipond (1979); Beaugrande (1979f, 1980a, b). For a new treatment in terms of computational logic, cf. Simmons & Chester (1979); Beaugrande (1981b).
11 It can easily be seen that “situations” subsume “objects”, and “events” subsume “actions;” we therefore usually speak of “situations and events” as a cover-all designation of primary concepts and their organization.
12 Though “perception” is more usual in the literature, “apperception” stresses the common case of applying prior knowledge to what we experience.
13 It might well be expedient to subdivide this category into “numericals” and “measurements” in future research.
14 In most cases, there is no special need to draw in “co-reference” and “recurrence” in text-world models, since the nodes in question are usually merged anyway. But it may prove useful to mark these relations when exploring such factors as the effects of repetition and variation within surface texts upon processing and recall. See note 19 to this chapter.
15 The notion of “case” originated in languages (e.g. Latin) which mark the role of nouns in phrase structures via surface formatting. Often, there were no unified conceptual criteria for all the uses of a single grammatical case. Fillmore’s early notion was that, in a language such as English, “cases” are “underlying” features of nouns in sentences. On Fillmore’s new outlook, see Fillmore (1977).
16 In general, linguists’ typologies have fewer categories than ours (e.g. Fillmore 1968; Chafe 1970; Longacre 1976), while those in artifidal intelligence have more (e.g. Wilks 1977a).
17 We use the symbol “†” for “termination” to prevent confusion with “τ” for “typical”.
18 See note 14 to Chapter IV.
19 We do not merge the two occurrences of ‘tons’, though it may turn out that we would be justified in doing so (compare note 14 to this chapter).
20 In terms of problem-solving procedures, spreading activation would be oriented more toward breadth-first search and inferencing more toward means-end analysis (cf. III.17).
21 We use two labels for the relation of ‘pilot’ to ‘plane’: the ‘pilot’ is in “containment of” the ‘plane’and also the “agent” acting on the ‘plane’ as “affected entity”. We could dodge the issue with a simple “relation-of’ link called “pilot-of”, but that doesn’t really settle anything.
22 We review some material added by receivers in IX.31ff.
23 To see whether a recipient group could agree on this matter, Beaugrande conducted a test in which receivers of the ‘rocket’-text were asked, six weeks after the original presentation, to select one of a set of choices in statements like: “Rockets must/should/don’t need to burn fuel when they fly”; “Rockets are always/often/seldom used to attack military targets”. As was expected, agreement was consistent, but by no means unanimous. Only these two statements elicited unanimous agreement: “Without fuel, a rocket can’t fly”; and “Launching sites usually are located on the earth”. Many replies showed that the group (all university freshmen) are poorly informed about aviation and rocketry (e.g. thinking that ‘alcohol’ is a “worthless fuel”). Such results should render us mistrustful of studies in which receivers are all presumed to be as knowledgeable as a “lexicon”.
24 Cf. note i 8 to Chapter IV.
25 For a discussion of this now discredited viewpoint, cf. Johnson-Laird (1978).
26 See note 1 to this chapter.
Return to main page
of
Introduction
to Text Linguistics