1 Semantic Macro-Structures and Knowledge Frames in Discourse Comprehension Teun A. van Dijk University of Amsterdam INTRODUCTION This paper outlines a theory of macro-structures within a larger framework of a model of cognitive information processing. We will focus on semantic structures and processes of discourse comprehension; in particular, we will formulate the macro-rules underlying the global interpretation of discourse. Although the focus is on discourse comprehension, the basic principies of macro-processing also are valid for other cognitive domains, such as vision, action, thinking, and problem solving. The notion of macro-structure is introduced as a partial explication of such notions as `schema' or 'plan' as they are currently used in cognitive psychology and artificial intelligence. Semantic macro-structures will be distinguished from other kinds of `schematic' structures of discourse, such as narrative super-struc- tures. Finally, we will elucidate the relationship between macro-structures and fr ames. In linguistics, macro-structures have been postulated in order to account for the 'global meaning' of discourse such as it is intuitively assigned in terms of the 'topic' or 'theme' of a discourse or conversation. The assumption is that these notions cannot be accounted for in terms of current logical, linguistic, and cognitive semantics for isolated sentences or sequences of sentences. In disci- plines such as rhetorics and narrative theory, macro-structures may constitute the semantic basis for specific categories and rules. For instance, the setting of a narrative should be defined at a macro-level of analysis and cannot be defined in terms of individual sentences (or their underlying propositions). Similarly, well- known categories such as `premise' and 'conclusion' in an argument also operate 4 VAN DIJK on global structures of discourse. Besides defining the global coherence of discourse, macro-structures also contribute to 'local' coherence at the micro-level of connections between propositions in composite sentences and successive sentences. In this paper, we will show that some fundamental problems of cognitive psychology can be accounted for in terms of macro-processing of complex semantic information. If a discourse is at all long, subjects are unable to process it at a micro-level alone: not only are they unable to store and retrieve such discourse verbatim, but they are also unable to retrieve the constituent proposi- tions. At the same time, if the sequences of sentences can be assigned a macro-structure, they will be recalled much better than scrambled sequences. Thus, comprehension (as well as production) probably takes place at several levels, such that lower-level information is organized, reduced, and represented at higher levels. These processes involve the use of macro-rules; the input to the macro-rules is the micro-structure, and the output is the macro-structure. Macro- structures help to explain the ability to summarize discourse, and in general to use information from discourse for other cognitive tasks, even if the individual propositions of the discourse are no longer accessible. Similarly, macro-struc- tures may provide further insight into the structure and use of frame-like representations of conventional knowledge in discourse comprehension and other cognitive tasks. SEMANTIC STRUCTURES OF DISCOURSE We will deal primarily with semantic macro-structures, although other kinds of global structures (e.g., narrative and pragmatic ones) may also play a role in the global organization and comprehension of discourse. One possible way to make semantic structures explicit is to use a formal language, such that expressions of this language can be given a suitable interpretation in the formal semantics of that logical language.' This formal semantics may be of the usual model-theo- retical sort, (e.g., involving notions such as possible worlds). One of the diffi- culties with other kinds of representations (e.g., graphs) is that they do not yet have an associated formal semantics (cf. Woods, 1975). Micro-structures and Coherence Micro-structures, the sequence of propositions underlying the sequence of sen- tences of the discourse, constitute the input to the macro-rules. The micro-struc- tures themselves require some preliminary semantic description. We will ' For an introduction to formal semantics (especially of modal logics), see Hughes and Cresswell (1968); for application in natural language, see Cresswell (1973). A brief survey is given in van Dijk (1977). See also Keenan (1975). 1. SEMANTIC MACRO-STRUCTURES 5 introduce the abstract concept of a text to refer to the abstract structure of a discourse. We can characterize the structures of texts at different levels of description.2 At one level of description, a text is simply an ordered sequence of propositions, which under various pragmatic, stylistic, and other constraints is mapped onto a sequence of sentences. Propositions are construed in the usual way, namely, as n-place predicates followed by n arguments which may be bound by quantifiers. Propositions may be modalized by various kinds of operators (tense, knowledge, belief, obligation, etc.), and connectives may be used to make compound propositions. The semantics of the formal language representing propositions provides recursive truth (or satisfaction) conditions in a constructive way. That is, interpretations of larger units depend on the interpretation of smaller units. Predicates are interpreted as properties or relations and arguments as individuals. Modal operators are interpreted in terms of possible worlds; for example, It is necessary that p is true in a world w iff p is true in all possible worlds which are accessible i from w . Sentences have as their referents the facts in possible worlds. Thus, the i sentence Peter is ill refers to the fact that Peter is ill now, i.e., in the actual possible world at the moment of the utterance. However, unlike classical formal languages, natural language is intensional. That is, we want to assign not only extensional referents, but also meanings. In particular, the extensions of an expression are assigned on the basis of its intensional meaning. Such intensions are functions, taking possible worlds as arguments and extensions as values. Thus, the phrase the book has an intensional meaning, namely the individual concept of a book, which may take various extensions, i.e., actual books referred to, in particular situations. Both intensional and extensional interpretations are necessary in an account of the semantic structures of discourse. A semantics of discourse is characterized by relative interpretations: sentences in a discourse sequence are not interpreted in an `absolute' way, but relative to the interpretation of other, mostly previous, sentences of the discourse. Se- quences satisfying the constraints of relative interpretation are called linearly coherent. One major coherence constraint is connection. Two propositions are considered to be pairwise connected if the facts they denote are related. This relation can be made explicit in terms of possible, probable, or necessary conditions, components, or consequences. Connections between propositions are typically expressed by natural connectives such as and, because, yet, so, etc. (cf. van Dijk, 1977). Sequences of propositions exhibit other coherence properties besides pairwise connections. For example, two expressions may both refer to the same facts, properties, or individuals. It should be noted, however, that in general these 'For introduction and further reference about the linguistic study of discourse, see Halliday and Hasan (1976). Petöfi and Rieser (1973), van Dijk (1972, 1977), and van Dijk and Petöfi (1977). The present introductory remarks about text structure are comple- mented by van Dijk (1977). 6 VAN DIJK conditions are neither necessary nor sufficient for coherence. Discourse co- herence is not primarily a matter of meaning, but of reference. Roughly speaking, the coherence of the discourse depends on the coherence of the possible-world fragment or course of events it represents. One of the conditions that makes a text uninterpretable is the impossibility of imagining situations where it could be satisfied. Of course, numerous formal and empirical details are omitted in this presentation. A further note of caution ís that our theoretícal abstractions and generalizations apply to an idealized discourse. Actual dis- courses that are produced, understood, and accepted do not always have a fully correct textual structure. Additional rules of pragmatics, cognitive strategies, and social conventions must account for the relevant conditions of acceptability of non-ideal discourses. There is another constraint on linear coherence that is not referential, but rather intensional or conceptual; not only must denoted facts be related, but this relation must be relative to a topic of discourse. That is, the facts must originate in the same range of semantic space. Thus, John's playing the piano may, as such, be independent from Mary's knitting, but both actions lie in the same range, namely, human leisure activities. Thus, connection conditions can be established relative to other propositions, which may or may not be explicitly stated in the text. For example, a sequence may be coherent because each fact relates to the general theme. The notion of a theme or topic of a discourse or a conversation will be reconsidered in terms of macro-structures. This means that condítíons for linear coherence may depend on condítíons for global coherence. Coherence is not only semantic, but may also be determined by pragmatic conditions. Clearly, connections between facts should be satisfied not only `objectively', but also relatíve to language users and communicative contexts. Similarly, the connections must relate not only facts but also speech acts. Thus, one speech act may constitute a condition, component, or consequence of another speech act. The details of these various pragmatic conditions on co- herence will be ignored here, as well as other principies of cooperative com- munication and interaction (cf. van Dijk, 1977). A major pragmatic constraint on discourse is that it be informative; conse- quently, information that the hearer already knows need not be expressed and asserted. This knowledge may be contextual or general. Contextual knowledge pertains to properties of the communicative situation, such as the presence of certain objects. This allows for coherence based on the context, such as the use of indexical pronouns (1, you, this, that, yesterday, now, etc.). General knowl- edge includes lexical/semantic information pertaining to the meaning postulates of the language — for example, knowing that the word "bachelor" entails un- marriedness. Altematively, the general knowledge may be conventional, in- volving shared knowledge about the `world' (e.g., knowing that Mexico is warm). Conventional knowledge contains not only actual facts, but also `possible facts' compatible with the actual world. The latter kind of conventional knowledge can be thought of as frames, which will be discussed in more detall later. 1. SEMANTIC MACRO•STRUCTURES 7 The propositions ímplied by the díscourse need not be expressed. Some propositions may remain implicit, even though they are essential in the establish- ment of linear coherence in texts. For example, they may be `presuppositions' of other propositions, or they may be necessary to allow the listener to draw certain ínferences. If we call the sequence of propositions of a discourse the text base,3 we can distinguish between an implicit and an explicit text base. The explicit text base is the theoretical construct containing all the propositions necessary to give relative interpretations of each proposition of the sequence. The explicit text base may be mapped onto the implicit text base by deletion of known or ínferrable proposítíons. It is the ímplicít text base which constitutes the input to the surface grammar. The distinction is not only important for cognitive reasons but also for grammatical reasons: the surface structure may contain 'traces' of the explicit propositions. One example of such a trace of the explicit text base is the use of pronouns without antecedents. Another is the use of defmite articles without indefinitely introduced noun phrases. This is also the reason why an explicit text base may not be too rich and contain all implications of each proposition. Only those propositions are interpolated which are im- mediate conditions for the full interpretation of each expressed proposition. Macro-structu res Beyond the linear semantíc structure of discourse, we also postulate global levels of descriptions, namely macro-structures.4 Macro-structures are assumed to be semantic structures of discourse whose meaning and reference is defined in terms of their constituents' meanings. Just as the value of a sentence is a function of its predicates, arguments, and operators, similarly the meaning of macro-structures is a function of the meaning and reference of the constítuent propositions of the explicit text base and the relations between those propositions. Since macro- structures are representations of meaning and reference at a more 'global' level of the discourse, they should satisfy the normal conditions of linear coherence as formulated aboye. That is, each macro-proposition must be interpreted wíth respect to other macro-propositions, including the implied propositions as de- fined aboye. Hence, macro-structures are explicit. A macro-meaning is the unifying property of the respective meanings of a sequence of propositions of a discourse. This means that certain combinations of 'The notion of `text base' has mainly been used by Petöfi and by Kintsch (1974) in cognitive psychology, where the experimental consequences of the distinction between implicit and explicit information are examined. 4The notion of macro-structure was first used by Bierwisch (1965) for specific (e.g., narrative) structures of díscourse (and their specific cognítíve consequences ín processing). It has been elaborated, first in theoretical poetics, then in linguistics, in our earlier papers and books (see van Dijk, 1972, for reference), and finally taken up again within a cognitive perspective in Kintsch and van Dijk (1975), van Dijk (1975a), and van Dijk and Kintsch (1977). 8 VAN DIJK propositions may be assigned, as a whole, a meaning at another level. It is evident that states of affairs, events, or actions may be represented at different levels of description. For example, the fact that somebody is building a house may be represented simply by the proposition 'Peter is building a house.' At a lower level of description, the 'same fact' may be represented by a long sequence of propositions denoting the actions involved in house building, such as plaster- ing and sawing. These actions are interrelated with respect to the global action of house building: they are executed under the scope of the same global intention or plan. Given our coherence conditions, these properties of possible worlds also determine properties of the discourse. That is, a combination of propositions denoting constituent actions of house building may be assigned the proposition 'Peter is building a house.' Since there are several possible levels of description, a global description may constitute the basis of a still more global description. For example, at a still higher level of description, house building may be a condition, component, or consequence of a more global action or event, such as `settling'. Hence, there is not one level of macro-structure, but several levels, as long as there are global concepts and facts defming the level. We will speak, however, of the macro-struc- ture of a discourse to refer to the highest level of macro-structure for the discourse as a whole. Characteristically, the relation of the macro-structure to the micro-structure involves notions such as importance or relevance: the macro-interpretation defines the most important or essential object or event denoted by a sequence of - propositions. For example, at a certain level of description, it may be more important that Peter built a house than where he actually bought the bricks. This notion of importance may be made explicit in terms of conditions, constituents, and consequences at the macro-level: building a house, but not buying bricks from some particular dealer, may be a condition for another global action, such as living in the house or selling it. If we substitute the macro-proposition for the sequence of propositions that it represents, we reduce the semantic information, in the sense of eliminating the details of the individual propositions. But the meaning or reference of a se- quence as a whole is preserved. Indeed, we may say that the macro-structure is entailed by its micro-structure. Macro-rules Since macro-propositions need not be explicitly expressed in the text, we need some mapping rules to obtain the macro-structure from the micro-structure of the discourse, in other words, rules to transform one proposition sequence into another `at another leve!' of description. This kind of semantic transformation we will call a macro-rule. The macro-rules must preserve global truth and meaning. Since macro-propositions need not be expressed in the discourse, 1. SEMANTIC MACRO-STRUCTURES 9 during comprehension they must be inferred from the explicit text base. The macro-rules, then, must yield an abstract model of there processes of inference. The `macro-interpretations' of the discourse consist of assignments of global meanings and references. The properties of macro-structures that we discussed aboye determine some of the properties of the macro-rules. For example, since there are several levels of description of any event, macro-rules must be recursive: they can be applied to organize global meanings hito still higher-level global meaning. From this recur- sive nature of the macro-rules it follows that the notion of macro-structure is relative to an underlying level of propositional representation, which in turn may be a macro-structure with respect to still more specific levels of representation. A macro-structure is typically more general than its corresponding micro-struc- ture. The first macro-rule, then, is GENERALIZATION. If somebody tells us he has a dog, a cat, and a parakeet, we may describe the same state of affairs with the more general statement that he has (three) pets. This rule operates not only on individual propositions but also on sequences of propositions, as in the three conjoined propositions of our example. GENERALIZATION holds for predi- cates in general, and hence applies both to nouns and verbs of the discourse. For instance, I may take the train to Paris and my friend may take a plane, but we are both going to Paris. In the example aboye, the kind of pet or the kind of transport become irrelevant at the macro-level, and can be replaced by the superordinate. Thus, the reductive nature of the macro-rule of GENERALIZA- TION is thus based on the deletion of certain essential properties which at some macro-level have become irrelevant, such that a whole set of discourse referents may be denoted by their common superset. While macro-rules must abstract and generalize, there must be some restriction on the level of abstraction. We still want macro-rules to yield specific informa- tion and not too general (and hence, uninformative) concepts. So, the rule must select the immediate or smallest superconcept. This guarantees that 'dog' and `cat' are not generalized recursively to `animate beings' or 'concrete things'. Note that this is a formal rule. In actual linguistic and other cognitive behavior we may well reduce information simply by referring to the various toys of our children lying around, for example, as `those things'. The general principie is that the least general superordinate is generated (if any exists). If there are several possible `next' superordinates, or if abstraction is possible at various levels, there are constraints that determine the correct set of possibilities. These constraints are the topic of discourse and the pragmatic, cognitive, and social contexts. Thus, in the example aboye, `pets' would be generated during a family discourse and `mammals' would be generated in a biological discourse. A second and closely related constraint is that there must be an upper bound on the application of macro-rules. Once a certain level of abstraction is attained, the tules should no longer be applied. The constraint we will use is rather simple: the application of macro-rules makes sense only if the reduced information is 10 VAN DIJK propositional. Hence, the macro-rules will apply only on sequences of two or more propositions. It follows that each proposition which cannot be reduced by some macro-rule must itself become part of the macro-structure. This occurs in those cases where macro-propositions are expressed in the discourse itself. This non-application of a macro-rule will be taken as the application of a ZERO-rule, whose input and output are identical. Such a rule corresponds to a cognitive operation that determines that a piece of information of the expressed text-base must be relevant with respect to the discourse as a whole. Let us now give a more precise definition of this first macro-rule: (1) GENERALIZATION (M-I) Given a sequence of propositions E = (yo. (al), )2(a2), • .. 9n(an)), where n 2, if there is a concept (I) such that each cD, and if there is a concept A such that each a D. A, then substitute (1) (A) for E, where D. i denotes immediate inclusion of the sets defined by the concepts. The rule allows both the predicates and the arguments to be generalized by a super-concept. In such a case the nongeneralized concept may remain identical, e.g., a = a = . . .= A. It is also understood that if the arguments are identical l 2 the predicates should be different, and conversely. This is a pragmatic constraint of informativeness on natural language propositional sequences of discourse. Finally, it is required that we indeed do have a specific super-concept for the predicates or arguments of the propositions of the text base. Since entailment is defined in terms of conceptual inclusion, we may re- formulate the rule by stating simply that a sequence E of propositions may be substituted by a proposition p if p is entailed by each of the members of E. We see that the rule is based on an entailment relation, as was required aboye. A special case of this rule is the generalization of complex predicates allowing a predicate such as `to travel by train' to be replaced by `to travel'. When the adverbial modifier is reconstructed as a separate atomic proposition, it can be deleted by the next macro-rule, to be discussed below. The same holds for other modifiers and complementa. Some examples may clarify how this rule handles sequences of sentences. We will use the symbol It to denote the macro-operation involved. (2) (John was moving the chairs, John was moving the table, John was moving the furniture John was moving the chest . . . (3) (Father was cleaning the kitchen, Mother was typing her new book, The whole family was working The children were painting the doghouse) In this last example we see that generalization is possible both of the subject terms and of the verbs. 1. SEMANTIC MACRO-STRUCTURES 11 Whereas the GENERALIZATION rule abstracts from essential properties of objects or properties or relations, the next rule, DELETION, deletes full proposi- tions from a given text base. It is difficult to formulate the precise conditions for the application of this rule. Its intuitive idea is that `irrelevant' information may be deleted. The notion of `relevance' or 'importance' of a proposition, relative to the discourse as a whole, will be made explicit in tercos of interpretation conditions. We assume that each proposition expressed by a discourse should be considered as relatively unimportant if it is not a condition for the interpretation of another proposition. Thus if we describe an episode in which Mary is playing with a blue ball and she breaks a window with the ball, the fact that the ball is blue is (normally) irrelevant; the proposition representing this fact may be deleted because it is not a condition for understanding the rest of the discourse. Similarly, a proposition is indirectly irrelevant if it determines the interpretation of a proposition which is itself deleted or substituted. However, these conditions for deletion are neither sufficient nor necessary. If the last proposition of a text base is not a condition for the interpretation of other propositions, it does not mean that this proposi- tion is globally irrelevant. So we should add that a proposition should not be deleted if it denotes a consequence of an event denoted by a macro-proposition. The deletion rule is essential for the reduction of information and operates in all those cases where the other rules do not apply. We may formulate the rule, provisionally, as follows: (4) DELETION (M-II) Given a sequence of propositions E = (Pi , P2 , . , p ), n -> 2, then each n proposition p may be deleted which is neither an interpretation condition i for at least one proposition (mostly following p ), nor a consequence of a macro-proposition r, of a sub-sequence of E. i This rule suggests that there is a necessary rule ordering of macro-rules, just as in the syntactic transformation of generative grammars: the deletion rule can apply only if we already have macro-structures of previous parts of the sequence. To stay within the same transformational terminology, we may say that this rule is an example of `semantic tree pruning'. Let us pie another example of this rule: (5) (I needed some olive oil, needed some olive So I went to the store and oil, So I went to the bought some, There was a store and bought some, sale on olive oil, We had a We had a good Italian good Italian meal . . . ) meal . . . > The proposition 'There was a sale on olive oil' may be deleted because it is irrelevant to the topic of eating Italian food and needing olive oil for it: the sale does not `follow' from needing olive oil, and hence does not determine buying it, 12 VAN DIJK nor having Italian food. If the sale would condition these facts, the proposition representing it would not be deleted. A characteristic of both rule I and rule II is that the information from which the macro-structure is abstracted cannot be recovered. However, other cues (i.e., other than the macro-structure) may allow retrieval of detailed information. The next macro-rule, INTEGRATION, deals with the possibility that macro-in- formation may be directly expressed in the discourse. This rule and the next one are interesting from the point of view of discourse processing because they organize and reduce information that is coherently related. According to the INTEGRATION rule, the more specific information of the passage may be deleted by the simple fact that its global information has already been expressed in the text by the proposition that also serves as a macro-proposition. This proposition may be selected as being important to the discourse. Formally, if a passage yields a macro-proposition that is the same as a propositíon in the discourse expressing this macro-structure, the latter proposition may be deleted. An alternative way of stating it is that all detailed information may be deleted which somehow has been integrated into another proposition of the discourse. We will follow the second alternative for convenience. How does the integration take place? The answer to this question will be simplified here: we will say that a proposition p may be `integrated' into a proposition Q if p denotes either a normal condition, a normal component, or a normal consequence of the fact denoted by Q. While formal definition of these tercos will not be attempted, we must at least consider what is generally meant by "normal." The notion of normality, which has very important implications in linguistic and cognitive semantics, will be taken as denoting truth in most possible worlds compatible with our actual world. This means that given some other facts, a normally co-occurring fact is expected. A condition, component, or consequence may have different `strengths', ranging from possibility via probability to all kinds of necessity. Thus we get `possible condition', 'probable condition', `necessary condition', `possible component', 'probable component', `necessary component', `possible consequence', 'probable consequence', and `necessary consequence' (for definitions of these notions in tercos of actual theory and formal semantics, cf. van Dijk, 1977). Consider some concrete examples. A normal condition for smoking a pipe is that I light it (although I may imagine some worlds where somebody else lights my pipe, or where it is lit by putting it in the sun, under a magnifying glass, etc.). Under the circumstances the condition is necessary because smoking a pipe implies that my pipe is burning and burning tobacco necessarily pre- supposes lighting. Similarly, drawing my pipe or puffing smoke are normally necessary components of pipe smoking. The point of all this, of course, is that given the proposition 'I smoke a pipe', the other propositions may be integrated into this proposition, because the detailed information is weakly implied by it. The macro-proposition allows us to describe detailed actions and events at a