11. Teun van Dijk and Walter Kintsch¹

11.1 The volume Strategies of Discourse Comprehension (1983) (hereafter SD), co-authored by a linguist and a psychologist, marks a new ‘surge’ since ‘around 1970’ (SD ix, 1). ‘The study of discourse’ arose from the decision that ‘actual language use in social contexts’, rather than ‘abstract or ideal language systems’, ‘should be the empirical object of linguistic theories’ (SD 1f, ix) (cf. 3.1; 4.17; 5.65; 8.50, 9.6f; 13.14, 36). The study requires an ‘interdisciplinary background’ and ‘diverse’ ‘scientific approaches': ‘linguistic analysis’, ‘psychological laboratory experiments’, ‘sociological field studies’, ‘computer understanding of text’ and so on (SD 19, ix) (cf. 13.22f). We can also look to ‘historical sources': ‘classical poetics and rhetoric’, ‘Russian Formalism’, ‘Czech Structuralism’, and ‘literary scholarship’ (SD 1). More recent work comes from ‘sociolinguistics’, examining ‘forms’ and ‘variations of language use’ like ‘verbal dueling and storytelling’; and from ‘anthropology’ and ‘ethnography’, moving from ‘verbal art’ in ‘myths, folktales, riddles’, etc. to ‘a broader analysis of communicative events in various cultures’, notably in ‘conversational interaction’ (SD 2). Today, ‘we witness a major ‘integration of theoretical proposals’ in ‘the wide new field of cognitive science’ (SD 4) (cf. 11.5, 102; 13.64).

11.2 ‘Until the 1970s, modern linguistics in America rarely looked beyond the sentence boundary’, aside from ‘tagmemics’ with its ‘fieldwork on indigenous languages’ (SD 2; cf. 5.56). ‘The prevailing generative transformational paradigm focused on phonological, morphological, syntactic, and later also semantic structures of isolated, context- and text-independent sentences, ignoring’ the ‘call for discourse analysis by Harris’ (1952) (cf. 5.56; 7.73, 79). So ‘interest’ in ‘discourse’ was ‘restricted’ to ‘European linguistics’, which was ‘closer to the structuralist tradition and had less respect for the boundaries of linguistics’ and ‘of the sentence unit’, as revealed in ‘studies’ ‘at the boundaries of grammar, stylistics, and poetics’. Also in Europe, attempts to ‘account for the systematic syntactic structures of whole texts’ led to ‘text grammar’, which however ‘remained in a programmatic stage, still too close to the generative paradigm’ (e.g. van Dijk 1972).

11.3 Influenced too by the ‘generative transformational trend’, ‘psycholinguistics’ focused not on ‘discourse’ in ‘language processing’ but on ‘the syntax’ and ‘semantics of isolated sentences’ (SD 3). Since then, we have realized that ‘models of sentence recognition’ based on ‘transformational grammar should be discarded’ (SD 74; cf. 11.14ff, 34, 81; 13.19). ‘Through analysis by analysis or analysis by synthesis’, such ‘models’ ‘try to match an input string of lexical items to structures generated by grammatical rules’; yet ‘even for a moderately complex sentence, the number of possible structural descriptions (trees) is astronomic’, precluding ‘effective search’ (cf. Woods 1970). Many ‘models less close to the grammar’ (e.g. of Fodor, Bever, & Garrett 1974, Chomsky's onetime associates), also foresee a ‘sentence recognition device’ for ‘syntactic analysis’ ‘trying to discover clauses’ as ‘surface representations of underlying sentoids'² without using ‘other kinds of information, such as semantic, contextual, or epistemic’ (SD 74f) (cf. 7.73, 82). This ‘information’, so often neglected by ‘philosophers, psychologists’, and ‘linguists’ (with their tidy ‘“lexicon”’), is just what the ‘language user’ deploys to derive ‘powerful expectations about the meaning of a sentence, and therefore also about the correct surface analysis’ (SD 305, 75) (cf. 5.57; 13.55). ‘Moreover, morphophonemic surface signals for syntactic structures’ may be ‘few’ and ‘difficult to perceive in natural speech’ (SD 75; cf. 11.36, 41f, 44, 56, 81; 7.48). ‘Hence, a semantically and pragmatically based system’ is ‘more effective’, able to ‘select among alternative parses’ or ‘even to circumvent syntactic analysis altogether’ (Clark & Clark 1977: 72) (cf. 11.34; 13.53).

11.4 ‘Psychology’ also saw ‘a breach in the paradigm’ in the 1970s and a revival of ‘work on discourse in the gestalt tradition’ (with its ‘notion of schema’) (e.g. Bartlett 1932; Cofer 1941) (SD 3; cf. 11.23-28). ‘Discourse materials’ were used in experiments on ‘semantic memory’ and in ‘educational psychology’, which ‘realized’ their role in ‘learning’ (cf. 11.37, 52, 54, 71, 95f, 98ff). ‘Extensive work’ also brought together ‘text linguistics and the psychology of discourse comprehension’ (SD 79). Similar trends ‘took place in artificial intelligence’, where a ‘paradigm’ was needed for ‘the computer-simulated understanding of language’ and ‘the automatic processing of texts’ (SD 3). To be sure, many ‘discourse process models’ still have ‘serious shortcomings’, being ‘incomplete’ and ‘focused’ on ‘problems of representation rather than dynamic aspects of processing’, such as ‘how textual representations in memory’ ‘are constructed step by step by a hearer or reader, and what strategies are used to understand a discourse’ (SD 61). Also, ‘previous models have seriously underestimated the complexity’ of ‘discourse comprehension’, which ‘involves processing a large amount of data’ (SD 95, 188; cf. 11.6, 10, 17, 20, 24, 26, 38, 41, 53, 57, 78, 82f, 91, 98).

11.5 Van Dijk and Kintsch now undertake to ‘present a broadly based, general, coherent approach to the investigation of discourse phenomena’, following the precept that ‘contextual information’ applies to ‘the whole range of communicative behaviour’ (SD ix, 238). Their ‘programmatic statements’ look ahead to ‘the future development of an interdisciplinary cognitive science’ (SD 19). Though their ‘theoretical outline’ is not ‘a worked-out information processing model’, ‘fully formalized and explicit’, they offer a ‘reasonably complete’ ‘framework for a theory’ within which ‘such models can be constructed eventually’ ‘given a particular comprehension situation’ (SD x, 95, 346, 351, 383, 385; cf. 11.21, 44, 90ff; 13.63). Their ‘model is general and flexible enough’ to be ‘later specified’, or ‘embedded’ ‘into a broader model of strategic verbal interaction in the social context’ (SD 9). This prospect befits the precept that a ‘social model should’ ‘have a cognitive basis’ and expound ‘strategies’ for ‘understanding, planning’, and ‘participating in interaction’. e.g., in ‘interpreting discourse’ (SD 19) (13.35). We might thus bridge the ‘gap between linguistic theory’ and ‘theory of social interaction’ (cf. 9.2, 6f). ‘Translating abstract textual structures into more concrete on-line cognitive processes’ can suggest how to do the same with ‘abstract structures of interaction and social situations’.

11.6 A ‘theory’ cannot be ‘at once specified and general’ because ‘comprehension’ is not a ‘unitary process’ but ‘differs’ according to ‘situations’, ‘language users’, and ‘discourse types’ (SD 383f, 9, 26, 259, 364). ‘New situations require new and different models’, as do particular ‘theoretical purposes’ (SD 383f) (cf. 9.1; 13.58). So we need ‘a framework’ for ‘discourse comprehension’, ‘a set of principles’ or ‘instructions for building specific comprehension models’ to fit ‘concrete cases’ and ‘a variety of behaviours’ (SD 383, 364, 346f). ‘Applications’ using ‘the same building blocks’ lead ‘beyond ad hoc, arbitrary miniature models’ that, however ‘simple’ and ‘elegant’, ‘deceive us about the real complexity of comprehension processes’ (SD 383). Also, it is easier to ‘agree’ about ‘the outlines of process model’, and ‘simplicity’ enables ‘testable empirical predictions at early stages’ of a ‘model’ (SD 293, 46; cf 11.90ff; 13.25, 57, 61). When we cannot ‘deal with the problem’ in its ‘full’ ‘complexity’, ‘a general framework’ keeps us aware of ‘where and what’ we are ‘simplifying’ (SD 384).

11.7 At the outset, van Dijk and Kintsch present a list of ‘cognitive’ and ‘contextual assumptions’ that ‘inspire the major theoretical notions and components of the model’ and indicate its ‘relationships with other models’ (SD 4ff). ‘The constructivist assumption’ is that ‘understanding’, whether of an ‘observed event’ or a ‘speech event’, ‘involves the construction of a mental representation’ (SD 4f; cf. 11.10, 20, 22, 25, 39, 51ff, 72, 100). ‘The interpretive assumption’ is that this ‘representation’ entails ‘not merely’ ‘visual and verbal data’ but ‘an interpretation’ of them (SD 5; cf. 11.19, 31, 36, 51). ‘The on-line assumption’ is that ‘the construction’ ‘takes place more or less at the same time as the processing of the input data’, not after the latter have been ‘first processed and stored’ (cf. 11.29, 36, 50, 101). ‘The presuppositional assumption’ is that ‘understanding’ entails ‘the activation and use of internal cognitive information’ about ‘general knowledge’ or ‘previous experiences’ (cf. 11.51). ‘The strategic assumption’ is that ‘processing’ is ‘flexible’ about the ‘kind’, ‘order’, or ‘completeness’ of ‘information’, and has ‘the overall goal’ of ‘being as effective as possible’ (SD 6; cf. 11.10). ‘The functionality assumption is that ‘discourse’ and ‘understanding’ are ‘functional’ in ‘a wider sociocultural context’, so that `processing’ is both a ‘cognitive’ and ‘a social event’ and the ‘representation’ covers ‘the social context’ as well as ‘the text’, which are ‘intertwined’ ‘at all levels’ (SD 6f, 221).³ ‘The pragmatic assumption’ is that ‘discourse’ is ‘social action’ consisting of ‘speech acts’, these too affecting ‘interpretation’ and ‘representation’ (SD 7; cf. 11.8f, 56f, 83f). ‘The interactionist assumption’ is that ‘discourse’ is ‘interpreted’ within ‘the whole interaction process’ among ‘speech participants’, including ‘verbal and nonverbal interaction’ (cf. 11.1, 5, 11, 17, 56, 83). ‘The situational assumption’ is that this ‘interaction’ is ‘part of a social situation’ wherein ‘participants’ may have ‘functions or roles’, and special ‘strategies’ and ‘conventions’ may apply’ (SD 7f; cf. 11.45, 51, 56f, 66, 74, 76).

11.8 Most importantly, ‘cultural information’ and its ‘communicative features’ ‘affect’ ‘all aspects of discourse understanding’ (SD 81) (cf. 3.1f; 13.63). ‘Cultural strategies have a very wide scope’, involving ‘knowledge’ about ‘geographical areas and locations’, ‘social structures, institutions, and events’, ‘speech acts’, ‘symbolic or ritual values’, ‘beliefs, opinions, attitudes, ideologies, and norms’ -- plus a whole ‘conceptual ordering of the world and society’ (cf. 11.20, 83). Such ‘cultural strategies’ may be ‘speaker or hearer oriented’, though ‘especially in everyday conversation, the two perspectives will coincide’ (SD 80). The ‘culture’ decides what people ‘believe to be important, relevant, interesting’, or ‘prominent’ ‘in discourse’ -- for example, whether ‘telling a story’ is intended to ‘amuse’, ‘reproach, give advice’, ‘reaffirm’ ‘norms’, or ‘teach history’ (SD 81, 239; cf. 11.60ff). For an unfamiliar ‘culture’, a ‘hearer or reader’ can apply ‘marked strategies’ and rely on ‘partial understanding’, ‘limited knowledge’, and ‘guesses’ (SD 81).

11.9 ‘Cultural strategies provide the basic background’ for ‘more specific social strategies’ relating to ‘context’ and ‘occasion': the ‘social structure of a group’ or ‘institution’, and the ‘roles or functions of participants’, who may be ‘young or old’, ‘rich or poor’, have ‘more or less power or status’, and so on (SD 82f).⁴ People know what ‘speech acts’ should be ‘performed’ in the ‘discourse’ of a ‘government, a bank, a judge in a courtroom, a student in a class, a friend in a bar, or a child at the breakfast table’. The ‘strategies’ applied here ‘limit the interpretation of many aspects of the discourse to rather restricted sets’ and help decide how ‘a discourse is ‘understood’ as ‘aggressive, helping, cooperative’, ‘obstructive’ etc., and how it ‘is meant to affect further verbal or nonverbal actions’ or ‘knowledge, beliefs, opinions, or motivations of the hearer’ (cf. 11.8, 20). Indeed, the ‘intention of the speech act may be inferred even before we hear’ it.

11.10 Van Dijk and Kintsch's ‘model’ centres ‘on the assumption that discourse processing, just like other complex information processing, is a strategic process’ ‘using both external and internal information’ in ‘understanding’ (SD 6, ix). ‘Strategies are flexible and operate on many kinds of input’ and ‘information’, even when these are ‘incomplete and partial’; they can ‘operate in parallel on several levels’ and collate the ‘results’; and they are ‘nondeterministic, often producing a large number of alternative outcomes varying in plausibility’ (SD 96f, 6, 10, 15f, 28, 73, 76, 98, 106, 127, 135, 151, 264, 308, 382; cf. 11.7; 13.52f). ‘A strategy’ can also be seen as ‘a cognitive representation’ of ‘the means of reaching a goal’ or ‘of a style’ for doing so ‘in the most effective way’ (SD 65). ‘Strategies’ themselves are ‘cognitive’ in that ‘they operate on’ ‘represented information': ‘things’, ‘events, or facts’ ‘in the world’ ‘are relevant for a cognitive model only’ as they are ‘distinguished, understood, and talked about through’ their ‘representation as concepts in memory’ and not as they ‘exist in some biophysical’ way (SD 80, 88; cf. 5.68; 11.43, 52f, 61). Still, we should ‘make a distinction’ whether ‘a meaning representation’ is ‘tied to language’ or to a fragment of the world’ (SD 88).

11.11 ‘A strategy involves human action': ‘goal-oriented, intentional, conscious, and controlled behaviour’ that ‘establishes’ or ‘prevents’ ‘changes in the world’ and its ‘states of affairs’ (SD 62, 264f). ‘If the results’ in ‘the final state’ fit ‘the intentions of the agent’, ‘the action is weakly successful’, but ‘strongly’ so if the action ‘brings about some goal’ or ‘far-reaching purpose’ (SD 62f, 264). ‘Cognitively’, ‘intentions are representations of doings plus their result’, whereas ‘purposes’ are those of ‘wanted consequences’; both ‘allow us to monitor’ our ‘actions’ as well as the ‘state of the environment (the action domain)’ (SD 63). ‘Actions are usually complex’, composed of ‘sequences’ in which some may be ‘automatized, that is, not governed by conscious intent nor individually’ aimed at the ‘general purpose’ (cf. 11.13, 15, 75, 77, 79, 83, 92, 95). In ‘interactions’, ‘several agents are involved’ with their own ‘intentions and purposes’, though ‘goals’ can be ‘coordinated’.⁵

11.12 The ‘notion of strategy’ can be ‘applied to actions in a strict sense: overt intended doings’ of a ‘bodily’ nature (SD 68, 62; cf. 5.21ff; 8.24f). But ‘overt action strategies’ also ‘presuppose thinking’, e.g., when ‘desires’ are ‘compared’ to ‘abilities’ and ‘possible or probable outcomes’ (SD 68f). So the ‘notion’ can apply also to ‘cognitive behaviour’ and ‘mental acts’ like ‘thinking and problem-solving’, which can ‘process much information’ and can be ‘conscious, orderly, and controlled’, each ‘mental step yielding the information necessary for the next’ (cf. 11.25, 51). Even in ‘cognitive activities that do not seem’ to work this way, such as ‘looking at a landscape or at a movie, or reading a text’, people have ‘the overall goal of comprehending’ and ‘follow a strategy of good’ or ‘fast understanding’ (SD 69; cf. SD 6, 18, 107).

11.13 These issues bear on how far ‘the notion of strategy is appropriate’ for ‘language use’ (SD 70). More than ‘problem-solving, the production and comprehension of verbal utterances’ is ‘automatized’ and ‘not monitored’ unless ‘difficult problematic, or unusual properties’ arise, e.g. an ‘unknown meaning of a word’, or a ‘complex’ ‘sentence structure’ (11.11). ‘Language production and comprehension’ are ‘continuous tasks’, made perhaps ‘of small scale problems’ but differing from ‘problem-solving’ in having ‘no single’ ‘well-defined’ ‘goal’ as ‘a final state’; and the ‘strategies’ are seldom ‘preprogrammed, intended, conscious, or verbalizable’ (SD 71) (but. cf. 11.51).⁶ Nevertheless, van Dijk and Kintsch postulate ‘strategies of language use’ that entail an ‘understanding of an action’ ‘step by step’, ‘a rather well-defined’ ‘starting point’, ‘alternative routes’, and at least a ‘fuzzy’ ‘goal’ (SD 70f). These ‘strategies’ belong to ‘the cognitive system’ and ‘apply to sequences of mental steps’ for various ‘tasks': ‘identifying sounds or letters, constructing words, analysing syntactic structures’, or ‘interpreting sentences and whole texts’. ‘Bottom-up’ strategies are ‘data-driven’, i.e., based on input, whereas top-down’ ones are ‘knowledge-driven’, i.e., based on the processor's predictions and notions about what is going on.⁷

11.14 So we should appreciate how ‘strategic processes contrast with algorithmic, rule-governed’ ones (SD 11, 67) (13.52). The latter ‘may be complex, long, and tedious, but guarantee success’ if ‘the rules are correct and are applied correctly’ (SD 11, 28, 67). ‘Rules’ form ‘a closed logical system’ which operates by ‘blind methodological application’ (SD 28, 67). ‘An algorithm always works but only in principle, not in real situations’ or for ‘practical purposes’, due to ‘human limits on time and resources’ (SD 67). In another sense,⁸ ‘rules’ are ‘general conventions of a social community, regulating behaviour in a standard way; strategies are particular, often personal ways of using rules’ and ‘making choices’ to suit ‘one's goals’. So ‘rules’ are ‘norms for possible or correct action’, and ‘sanctions’ follow if they are ‘broken’, e.g. in ‘games’ (‘chess’) or ‘traffic’. ‘Similarly, rules of language determine which utterances are correct’ in the ‘system’, e.g., the ‘syntactic parsing rules’ whereby a ‘generative grammar produces a structural description of a sentence’ (SD 67, 11) (cf. 7.49). The rules ‘represent’ in ‘idealistic terms what language users in general do or what they implicitly or explicitly think they do or should do’ (SD 72) (cf. 9.6). ‘Uses of the rules’, however, ‘depend on ‘variable’ ‘contexts’, ‘users’, and ‘goals’ (SD 72, 94).

11.15 In contrast, a ‘strategy’ is ‘simpler’, ‘intelligent but risky’, has no ‘guarantee’ and ‘no unique representation’, and produces ‘effective working hypotheses’ and ‘fast but effective guesses about the most likely structure or meaning of the incoming data’ within ‘available’ ‘resources’ in ‘real time’ (SD 11, 28, 67, 73f). Like ‘uses of rules’, ‘strategies’ ‘depend on ‘characteristics of the language user’ (‘goals or world knowledge’) as well as of the ‘text’ (SD 72, 11, 7). ‘Strategies’ are ‘part of an open set’ and ‘need to be learned and overlearned before’ being ‘automatized’; some, like ‘gist inferring, are acquired rather late’ or through ‘training’ with ‘new types of discourse’ like ‘psychological articles’ (SD 11). The ‘processing features of natural language utterances’ make ‘strategies’ ‘necessary': ‘language users have limited memory’, especially ‘short-term’; they ‘cannot process many different kinds of information at same time’; ‘production and understanding of utterances is linear, whereas most structures the rules pertain to are hierarchical’ (5.69); and ‘production and understanding require’ more than ‘linguistic or grammatical information’ (SD 72f) (13.44). ‘Whereas rules are abstract’ and ‘formulated a posteri for complete structures’ of ‘categories and units’, ‘strategies allow’ for ‘production or understanding linearly at several levels simultaneously’, using ‘different kinds of information’ and ‘limited knowledge’ (SD 73; cf. 11.7, 19, 26, 32, 35, 38, 58, 77f; 13.53, 57).

11.16 ‘Although strategic systems are nondeterministic’, ‘probabilistic’, ‘open-ended, and highly context-sensitive’, ‘scientific’ ‘theories’ about them can ‘be stated with precision and objectivity’ (SD 31, 74). ‘Evidence has been compiled showing that people really do operate that way’, whereas the ‘rule systems that linguists were using to parse sentences were implausible’ (SD 28; cf. 11.3). ‘Even if we accept the hypothesis that grammar is a theoretical’, ‘general, abstract, and idealized reconstruction of the language rules known by language users’, we still need ‘strategies’ for producing or understanding structures’ by using the various ‘levels’ such as ‘grammar, morphology, or syntax’, along with ‘the communicative context’ (SD 73; cf. 4.71; 5.34f; 7.45f; 8.51f; 9.30; 11.35, 56; 12.82; 13.29). On the other hand, it would be ‘uneconomical for the cognitive system’ if ‘strategies and rules’ were ‘independent’ and ‘did not make use of the same units’ and ‘categories’, ‘at least in part’ (SD 73f; cf. SD 91). When ‘strategic’ ‘guesses’ are ‘wrong’, ‘grammatical rules will establish, on second analysis, the correct structure or meaning’. Also, appropriate ‘schemas’ enable ‘interaction’ ‘between rules and strategies’ by ‘applying’ ‘patterns’ when ‘input data appear to be standard’.⁹ Some ‘strategies have their counterparts in rules of grammar’, though ‘other kinds’ do not, e.g., those applying to ‘the schematic structures of narrative’ (SD 91; cf. 11.60ff).

11.17 ‘The complexity of action or interaction’ requires ‘higher organization’ by a ‘global plan’, i.e., ‘a cognitive macrostructure of intentions, purposes’, ‘actions’, ‘consequences’, ‘goals’, and ‘strategies’ (SD 63, 265). ‘A course of action’ can be ‘represented’ by ‘a tree diagram’ of ‘alternative’ ‘paths’ among ‘changing’ ‘states’ in ‘possible worlds’ (SD 63f, 265). ‘Paths’ differ in ‘effort’ and ‘cost’, and may ‘involve unwanted intermediary states’ (SD 64). ‘A rational agent will try to reach an optimal goal along the lowest-cost path’, e.g. by ‘means-end analysis’ (‘comparing costs and goals’) (SD 64f). Though ‘in everyday life, we perform many actions without much of a strategy’, ‘strategies become necessary’ when ‘goals’ are ‘important or the means very costly or risky’ (SD 66). ‘A heuristic’ is ‘a system of discovery procedures’ to ‘acquire knowledge about conditions’ for ‘reaching a goal’, especially on ‘higher levels’ where we cannot ‘plan in advance each detailed action’ (SD 68) (9.15, 17). ‘A classic example is scientific investigation: to formulate some regularity’, we may ‘systematically observe’ some facts, or ‘we may first derive it’ and then check it ‘with the facts’, or we may try both ways (SD 70) (cf. 13.44).

11.18 A ‘plan’ is ‘dominated by a macroaction': ‘the global conceptual structure organizing and monitoring the actual action sequence’ and ‘defining global’ ‘goals’ (SD 63, 265). ‘Together’, ‘plans and strategies’ make up ‘the content and style of a global action’, with the ‘strategy dominating the moves’, that is, the ‘functional’ (‘bound’) ‘actions’ ‘in a sequence’ (SD 65ff). A ‘tactic’ is ‘an organized’ ‘system of strategies’ applying to ‘large segments or periods of lives and actions’ and influencing ‘the personality of the agent’; ‘bad tactics typically involve conflicting strategies’.

11.19 Therefore, ‘linguistic and cognitive theories of discourse’ entail ‘two sets of related strategies, local and global’ (SD 89; cf. 11.30, 32, 38, 47, 66, 82, 85). ‘The local strategies establish the meanings of clauses and sentences’ and of ‘relations between sentences’. The ‘global’ ones ‘determine’ the ‘meanings of fragments of discourse’ or of the ‘whole’. The ‘two kinds of strategy must of course interact’ in ‘text comprehension’, possibly in ‘hierarchical relations’ of ‘dominance’ (SD 89, 106). ‘Global information acts in top-down processing strategies’ for the ‘local’; and ‘local ‘strategies’ provide ‘constraints for specific meanings’ by looking ‘forward’ for ‘meanings to come’ or ‘backward’ for ‘meanings’ only ‘partially interpreted’ (SD 106f). In such ways, ‘knowledge’ can be ‘called’ by ‘all interpretation strategies’ to ‘provide precisely the relevant information at each point’ (cf. Winograd 1972). ‘These preparatory, communicative, and contextual strategies’ ‘specify’ ‘the overall goal of the reading act’ and ‘determine the choice’ of ‘local or global textual strategies of comprehension’.

11.20 The ‘role’ of ‘world knowledge in production and comprehension’ of ‘discourse’ has been strikingly ‘demonstrated’ by ‘psychology and artificial intelligence’ (SD 303, 307) (cf. Winograd 1972; 11.23). ‘Large amounts of knowledge’ are ‘not provided’ or ‘expressed in the text’ but must be ‘accessed’ and ‘retrieved’ to ‘provide a framework for the text’, ‘organize’ it, ‘understand’ it, and ‘construct’ a ‘mental representation’ in ‘memory’ (SD 6, 13, 46, 106, 188, 191, 303f, i.r.). Moreover, all this may be ‘formed or transformed’ during ‘discourse-related tasks’ themselves (SD 191). Of course van Dijk and Kintsch cannot ‘present a complete representation format for the knowledge’ and ‘cognitive’ and ‘contextual information’ ‘necessary’ for the ‘semantic operations of discourse understanding’ (SD 13, 8f). But we are continually reminded that their strategies and constructs involve or depend on ‘knowledge’, ‘beliefs’, ‘opinions’, ‘attitudes’, ‘ideologies’, ‘norms’, ‘conventions’, ‘evaluations’, ‘emotions’, ‘wishes’, ‘intentions’, ‘motivations’, ‘goals’, and ‘tasks’.¹⁰ Indeed, ‘knowledge is everything we know’ (SD 312) (cf. 4.14; 5.28).

11.21 Therefore, van Dijk and Kintsch only ‘sketch the overall outlines of a knowledge system’ with ‘many levels’ and ‘nodes’ ‘forming overlapping chunks’ (SD 311) (cf. 11.75f). Evidently, ‘knowledge is well organized’ in ‘flexible’ ways suitable for ‘the strategies of knowledge use’ (SD 13, i.r.). Instead of ‘blindly activating all possible knowledge’, these ‘strategies’ work from ‘the goals of the language user’, the ‘available knowledge from text and context, the level of processing, or the degree of coherence needed for comprehension’. ‘Knowledge’ can be broken down into (a) ‘episodic’, i.e. ‘construed’ or ‘inferred’ from ‘previous experience’, versus (b) ‘conceptual’ or ‘semantic’, i.e., ‘derived’ through ‘abstraction, generalization, decontextualization, and recombination’, and therefore ‘general, stable’, and ‘useful’ for many ‘cognitive tasks’ (SD 303, 13, 308, 312; cf. SD 11f, 106, 135, 151, 160, 273, 337, 344; 11.31, 51, 58, 74ff).¹¹ Thus, the `”knowledge system”‘ runs both on ‘context-embedded unique personal experience’ and on ‘decontextualized generalized information’, and uses them ‘in comprehension’ in ‘multilevel’ ways (SD 312; cf. 11.10, 13, 39f). One prominent way is ‘spreading activation’, which travels ‘automatically’ among ‘nodes’ associated in a ‘network’ (SD 24, 96, 167, 316) (Collins & Loftus 1975). A more controlled way is making ‘inferences’, i.e., adding ‘necessary, plausible, or possible’ ‘information’ to the ‘discourse’ (Rieger 1977) (SD 49) (11.25). ‘Bridging inferences’ are ‘required for coherence’, while ‘elaborative’ ones only ‘fill in additional detail’ (Kintsch 1974; Kintsch & van Dijk 1978) (SD 49, 51).¹²

11.22 In the past, most researchers in ‘philosophy’, ‘psychology’, and ‘linguistics’ have designed ‘associative networks’ or considered ‘how general concepts are abstracted from concrete instances’, e.g., via a ‘summary description’ stating ‘necessary and sufficient properties for class membership’ (SD 305, 307, 310). This approach works all right for ‘artificial concepts’, but not for ‘natural’ ones (cf. Bruner, Goodnow, & Austin 1956; Smith & Medin 1981) (SD 305). Today, ‘psychologists’ are ‘developing models’ providing for ‘nonessential features or dimensions’ or even for ‘concepts entirely characterized by exemplars’ (SD 305, 310). Or, ‘concepts’ are ‘defined’ ‘by their position in the semantic network and their mutual relations’, which ‘vary’ in ‘quality’ as well as ‘strength’ (SD 307; cf. 11.69). But for ‘a model’ of a ‘knowledge system’ in ‘discourse comprehension’, all this is still ‘too narrow’, too preoccupied with ‘categorizing’ and ‘classifying objects’ (e.g. ‘animals’, ‘kinship’). Using ‘concepts’ for ‘constructing text representations’ during ‘language use’ entails much ‘fuzziness’ of the kind usually ‘ignored or ruled outside linguistics’ (SD 306; cf. 5.47; 9.29; 11.26, 34, 39; 13.22, 59).

11.23 Recent research has turned to entire ‘knowledge structures’ -- termed ‘schemas’, ‘frames, or scripts'¹³ -- for ‘information in memory’, having ‘a label’ and ‘slots’ (‘variables’) within a ‘prearranged relation’, and ‘accepting information of a given type’ via ‘instantiation’ (SD 307, 47, 13). Here, ‘classifying knowledge structures’ is done not just ‘by content area’ but by ‘packets’ that ‘can function as wholes’ (SD 47; cf. 11.27). Such ‘schemas are descriptions, not definitions’, and vary from ‘concrete’ to ‘abstract’ (SD 47). Their ‘information’ ‘is normally valid’, but specifies ‘no necessary and sufficient conditions’ (SD 47f). ‘Instead, normal conditions from many different content areas are combined’, including ‘goals, consequences’, ‘implications’, and so on. ‘Although knowledge’ is ‘socioculturally variable’, its ‘generality’ evidently suffices ‘for intersubjective language use and communication’ (SD 303; cf. 11.16, 37; 13.58). ‘Without this general picture of the world’ no one could ‘understand words’ in ‘meaningful combinations’ within or among sentences’ or in ‘a discourse as a whole’, or ‘make sense of the facts’ (cf. 11.20).

11.24 ‘Many unsolved problems’ remain in ‘building a knowledge structure’ and getting ‘a knowledge base to deliver nicely packaged schemas’ yet to ‘retain flexibility and context sensitivity’ (SD 48, 311). ‘In each new context’, ‘a subtly different complex of information’ may be ‘relevant’ (SD 48; cf. 4.16; 5.76). ‘The meaning of a concept cannot be specified for once and for all by some small set of semantic elements’ but ‘requires’ ‘a large, open set of complex statements’ (SD 311) (cf. 5.76; 7.77; 13.59). Hence, ‘problems of schema use’ may arise for both ‘identification and application’ (SD 48). Also, ‘misrepresentation’ and ‘distortion’ can arise when ‘readers’ ‘supply’ ‘knowledge’ left ‘implicit’ by ‘a text’ about ‘causal relations in the physical world and the goals, plans, and intentions of human actors’ (cf. Stevens, Collins & Goldin 1979; Graesser 1981) (SD 46, 304). The ‘naive action theory’ and the ‘causal model people use’ is not ‘the unambiguous, contradiction-free system of science’; ‘even experts’ may ‘reason at multiple, mutually inconsistent levels’ (SD 46f) (cf. 13.24).

11.25 Moreover, ‘most discourses’ and the ‘actions and events’ they refer to are ‘new’ and ‘interesting’ ‘in some respects’, and ‘preestablished knowledge’ may ‘not fit’ ‘precisely’ (SD 304). To deal with ‘new’ material, ‘background information’ must ‘accommodate many variations’ and ‘contextual demands’ by adjusting, combining etc. ‘Schematic structures often occur in a transformed way’ in ‘actual discourse’, and the ‘reader’ must ‘determine’ the current ‘schematic function’ ‘from the global content’ (SD 92).¹⁴ For such reasons, van Dijk and Kintsch do not equate the ‘instantiated frame or script’ or schema ‘with the textual representation’ (SD 307f). Instead, the ‘use of general knowledge’ involves ‘two steps': (1) ‘activation’ and ‘instantiation’ of a ‘schema, ‘frame or script’ via ‘some input’; and (2) ‘construction’ of ‘the knowledge base for understanding the text’. ‘Once selected, a schema’ ‘provides readers with a basis for interpreting the text’, and a ‘conceptual skeleton’ to which they can ‘bind the semantic units derived from the textual input’ (SD 48). ‘schemas’ ‘also provide a basis for more active, top-down processes’, such as ‘inferences’ that supply ‘missing information’ or ‘assign default values’ (cf. 11.21). ‘Deviations’ may be ‘registered and accepted’ or may trigger ‘problem-solving’ ‘to account for them’.

11.26 Since these ‘knowledge systems’, like other ‘concepts’, are ‘fuzzy’, ‘flexible, and context dependent’, we encounter ‘difficulties in designing’ ‘representations’ for them (SD 310, 71). ‘Neither concepts nor schemas can be defined in the strict sense’,’ and ‘dynamic, flexible systems’ are much harder to envision than ‘definitional’ ones (SD 311). There may be ‘no end to special tracks’, and special versions ‘can be generated on demand’ (SD 310). We must ‘work with complex, messy interactions’ in a ‘multileveled system’ of ‘features, concepts, propositions, and schemas’ (SD 311). We must inquire if ‘knowledge representations are abstract and propositional or if they involve imagery’; ‘how we can identify the internal structure of a knowledge system from behavioural data’; and so on.

11.27 Despite such worries, ‘the schema notion’ now figures in ‘theories’ ranging from ‘letter perception’ to ‘macrostructure formation’ (SD 48). This accord may lead to ‘a truly general, comprehensive theory of discourse perception and comprehension’ (cf. Adams & Collins 1979). ‘Good evidence’ indicates ‘schema-based knowledge systems are real or at least psychologically plausible’, i.e., able to ‘function as psychological units’ or ‘chunks in memory’ (SD 309f) (cf. 11.75). Experiments show that people ‘cluster’ or ‘list the actions of a script together or make recognition errors among them’; if ‘presented out of order’, ‘the actions’ get ‘reordered’ (Black, Turner, & Bower 1979) (SD 309f). A ‘script is retrieved as a unit’, the ‘speed’ of retrieval depending not on how many ‘actions’ it has but on ‘how close the actions are to each other and how central they are to the script’ (Anderson 1980; Smith, Adams, & Schorr 1978; Galambos & Rips 1982). Apparently, ‘scripts’ serve ‘both as cognitive cueing structures and as guides for the allocation of attentional resources’ (SD 310). ‘Evidence’ also reveals ‘substructures in scripts': ‘subjects’ ‘distinguish fixed scenes’ and ‘mark them linguistically with a single word’ (i.r.). And ‘hierarchical’ ‘structures’ appear when ‘actions’ ‘in a narrative activate their superordinates’ (Abbott & Black 1980; cf. 11.62).

11.28 ‘Linguistics’ too has ‘widely’ postulated knowledge structures, often called ‘verb frames’ with ‘case roles’ for ‘agent, patient, instrument’ ‘goal’ ‘source’, etc. (cf. Fillmore 1968; J.M. Anderson 1971; Dik 1978) (SD 308, 114) (cf. 7.63; 11.48, 61). These ‘frames’ can form a ‘hierarchy’ and ‘inherit properties’ from ‘superordinate’ ‘frames’, e.g., a ‘transitive act’ being assigned ‘agent and patient slots’ (SD 309). We need not decide ‘how many cases there are’; beside ‘a few general’ ones, many ‘specialized cases’ can appear with certain ‘verbs’ and do not form ‘a closed set’, just as ‘a schema’ need have no ‘finite, fixed set of slots’ but may add ‘special-purpose’ ones -- yet another obstacle to ‘formal theories’ (cf. 11.26).

11.29 Knowledge patterns are managed through ‘a system of strategies as used by speakers and hearers to establish, construct, discover, or recognize’ ‘coherence’ (SD 79, 151). ‘Extensive work in text linguistics and psychology’ has already explored ‘the conditions for discourse coherence’ ‘in terms of semantics, pragmatics’, and ‘world knowledge’, but largely with a ‘structural approach’ looking for ‘abstract relations between sentences’ or ‘propositions’ ‘relative to some possible world’ (SD 79, 150f) (11.40). In contrast, ‘language users establish coherence as soon as possible, without waiting for the rest of the clause’, ‘sentence’, ‘sentence sequence’, ‘paragraph’, or ‘discourse’ (SD 15, 154, 205, 44, 237, 285; cf. cf. 11.7, 50, 101). They must do so ‘in real time and with a limited short-term memory capacity’, so ‘propositions are constructed on line’ when ‘information is available’ (SD 44, 373, 186; cf. SD 19, 134, 138, 166, 245, 351, 143). Hence, we need to find out how the ‘strategies’ ‘handle the information involved’ in ‘textual coherence’, ‘what memory resources and mechanisms are involved’, and so on (SD 151).

11.30 For ‘language users’, ‘coherence intuitively means’ a ‘unity’ and ‘a normal, possible, understandable, or correct continuation’ for the ‘ongoing discourse’ (SD 79) (cf. 3.25; 9.93). ‘These intuitive notions can be theoretically represented’ via ‘local and global semantic properties of a discourse’, and ‘reformulated as strategies’ for handling ‘surface structure’ and using ‘knowledge’ and ‘contextual information’ (SD 80). ‘Whereas an abstract linguistic semantics will formulate’ ‘a ‘general and abstract definition of coherence’, ‘a cognitive model’ should deal with ‘cultural, cognitive, and personal’ ‘contents’ of ‘coherence’ (SD 150). Yet this mix of ‘objective’ and ‘(inter)subjective’ ‘does not mean that ‘coherence is arbitrary’; some ‘properties’ ‘remain constant’, e.g, ‘relations between denoted facts’ and ‘fact elements’ (cf. 11.16, 23; 13.58).

11.31 Though ‘coherence’ can also be ‘syntactic’, ‘stylistic’, and ‘pragmatic’, van Dijk and Kintsch focus on ‘semantic coherence’ (SD 149).¹⁵ They see ‘two fundamental types': ‘conditional’ (or ‘extensional’, i.e. ‘referential’) based on ‘cause’, ‘consequence’ and ‘temporality’, versus ‘functional’ (or ‘intensional’) based on ‘example’, ‘specification, explication’, ‘contrast’, ‘comparison’, ‘generalization’, ‘conclusion’, and so on (SD 149f, 159, 182, 184f, 204).¹⁶ ‘Functional’ ‘links’ dominate in ‘typical expository’ ‘texts’, and conditional links’ in ‘narrative ones’ (SD 183, i.r.). A ‘distinction’ is also made between ‘three levels of coherence’ gauged by ‘depth of interpretation: superficial’ if two ‘propositions’ are ‘in the same frame or script’; ‘normal’ if the two also ‘instantiate a direct conditional or functional connection’, and ‘full’ ‘if further information is inferred from semantic or episodic memory’ (SD 160; cf. 11.21). ‘The reader’ pursues one or more of these levels ‘depending on the type of text’ and ‘context (tasks, goals, interests, time, etc.)’.

11.32 But by far the most crucial distinction falls between ‘local’ and ‘global coherence’ (SD 11f, 13, 80, 150, 308, 337; cf. 11.19, 47, 56, 66, 77, 83, 85).¹⁷ ‘Local coherence strategies’ ‘establish meaningful connections between successive sentences in a discourse’ ‘or between constituents of sentences’ (SD 14f, 150, 189). ‘Global coherence ‘organizes’ and ‘orders’ ‘predicates’, ‘referents’, ‘properties’, and so on, around the ‘central’ ones, and imposes ‘unity’ and ‘sequence’ (SD 151). ‘Schematic structures’ (as in 11.27) apply to ‘the organization of discourse’ both ‘locally’ to the ‘morphological, syntactic, and semantic levels’ and ‘globally’ to ‘the macrolevel’ (SD 92, 204f, 308). Against much of linguistics, van Dijk and Kintsch assert that ‘the strategy types of the largest scope’ are the ‘most fundamental to understanding’ ‘language’ and ‘semiotic practices’, as well as ‘interactions, events, and objects’ (SD 80) (cf. 13.57). ‘Local coherence strategies’ need ‘guidance’ and ‘constraints’ from the ‘global’ to relate to the ‘discourse as a whole’, to surmount ‘discontinuities’, and so on (SD 188f, 233; cf. 5.19, 38; 11.19, 35; 13.32). ‘Local coherence strategies operate both bottom-up’ with ‘words and phrases’ and ‘top-down’ with a ‘schema, frame, script, or macroproposition’ (SD 159; cf. 11.13). Even when the ‘local is minimal’ or ‘degenerate’, ‘adequate macrostructures are formed’, e.g. in ‘skimming newspaper reports’ (Masson 1979) (SD 233). Therefore, we should ‘investigate’ the ‘interaction of local and global’ for ‘easy and difficult texts, stories and essays, skimming and memorizing’, and so on.

11.33 These precepts lead to a special view of ‘linguistic parsing’ (cf. SD 8, 19, 27, 59, 134, 385; cf. 7.49; 11.14, 16, 77, 79). In that view, ‘phrases’ and ‘sentences’ are addressed not because they are the central units of an abstract grammar, but because ‘psychological evidence’ indicates ‘readers and listeners are sensitive’ to them as ‘functional psychological units’ for ‘processing’ and ‘chunking’ (SD 28, 37; cf. 13.31). Evidently, ‘readers segment at phrase boundaries’ (Garrett, Bever, & Fodor 1966) and ‘hold the final phrase in short-term memory, dumping it’ ‘at a clause boundary’ (Jarvella 1971); also, ‘most errors in learning a sentence occur at major clause boundaries’ (Johnson 1965) (SD 28).¹⁸ So ‘clause boundaries are important’ because many ‘strategies deal with constituents no larger than the clause’ when ‘local information’ is ‘sufficient’ (SD 36). But ‘whether the clause boundary’ actually is ‘a decision point’ depends on what ‘information’ is needed for a ‘semantically complete’ ‘unit’.

11.34 Some theories hold that ‘people’ ‘rely on linguistic rules’ ‘applied when parsing a sentence’, e.g., those for ‘phrase structure or ‘transformations’, within a ‘closed system’ (SD 28) (cf. 7.49; 11.3, 16, 81, 92; Winograd 1983). In contrast, ‘strategy theories of sentence comprehension’ hold that ‘parsing’ runs on an ‘open nondeterministic fuzzy system’ (11.14f). Sample ‘strategies’ might be: ‘whenever you find a function word, begin a new constituent’ (e.g. a ‘determiner’ to start a ‘noun phrase’, or ‘a relative pronoun’ to ‘begin a new clause’);¹⁹ or ‘attach each word to the constituent that came just before’ (again, as with ‘relative clauses’) (SD 29f). For longer stretches, we might have: ‘select the grammatical subject of the previous sentence as the preferred referent for a pronoun’ in the next ‘sentence’; this ‘strategy’ makes ‘reading times faster’ (Frederiksen 1981), but is less ‘dominant’ than assigning ‘role’ (e.g. ‘agent’ of an action), ‘recency’, and ‘topicality’ (cf. 11.28, 45, 63f, 68, 79, 86). Even less rule-bound (in Chomsky's sense) is the ‘strategy’ of ‘using semantic constraints to identify syntactic function’, which ‘in extreme cases allows the construction of propositional representations directly from the sentence, bypassing syntactic analysis’; young ‘children’ seem to do this (SD 30) (cf. 9.11; 11.3; 13.53).

11.35 ‘Many models of language’ in ‘linguistics and psychology’ postulate ‘levels of morphonology, syntax, semantics, and pragmatics’ (SD 10) (11.16). Yet such a ‘description’ is ‘not particularly relevant’ for ‘processing models’, where the ‘levels interact in an intricate way’ (cf. 11.7, 15, 19, 26, 32, 38, 58, 77f; 13.28). ‘The strategic approach’ stresses ‘close cooperation’ among ‘phonological, morphological, lexical, syntactic, and semantic strategies’ (SD 272, 282). This befits ‘functional approaches to grammar’, which explore the ‘dependence of surface structures upon underlying semantic and pragmatic representations and their cognitive and social processing’ (SD 283; cf. 9.22f).

11.36 Accordingly, ‘semantic interpretation does not simply follow full syntactic analysis but may already occur with an incomplete surface structure input’, and ‘syntactic analysis may use information from semantic and pragmatic levels’ (SD 10; cf. 11.7, 100). Nor must we uphold the ‘fundamental principle of linguistic and logical semantics that the interpretation of a unit’ rests on that of its ‘constituent parts’ (SD 190; cf. SD 126) (cf. 5.64, 67, 75ff; 6.47f; 7.82; 12.27, 93; 13.18, 59).²⁰ The principle was convenient when ‘linguistic semantics’ considered ‘the meaning of expressions’ ‘abstract, stable’, and ‘intersubjectively invariant’, ‘belonging to the language system as opposed to actual language use’; thus, meaning could be ‘specified independently of contextual and personal variations, which were left to psycho- and sociolinguistics’ (SD 192) (13.55). ‘Psychology’ too sought ‘abstract and generalized models of language understanding’ and ‘principles followed by all language users’ (SD 193; cf. 11.3).

11.37 Again, such ‘abstracts accounts’ are ‘insufficient for cognitive models’, which should ‘define the actual processes by which macrostructures are derived’, ‘the strategies’ for ‘handling’ the ‘information’ (‘macrostrategies’), ‘the memory constraints’ and ‘representations’ for ‘macrostructures’, the ‘knowledge types’ needed, the ‘retrieval and (re)production of discourse’, and the tasks (like ‘summarizing, question answering, problem-solving, or learning’) that involve ‘macrostructures’ (SD 191f, i.r.; cf. 11.29f, 82). Yet insofar as ‘the understanding of a discourse depends on variable features of language users and contexts’ (11.6, 14f, 21, 24, 47, 58, 66), ‘each language user assigns his or her own macrostructure’ and ‘finds different meanings prominent, important, relevant, or interesting’ (SD 193). Still, ‘individual differences presuppose’ ‘common information’, and ‘macrostructures’ cannot be ‘completely arbitrary or disparate’ (cf. 11.23; 13.58).²¹

11.38 Van Dijk and Kintsch's own ‘model operates’ on ‘complex chunks’, and works ‘from the word units on the lower levels up to the units of overall themes or macrostructures’, with each end helping to ‘construct’ or ‘understand’ the other (SD 10; cf. 11.19, 21, 26, 32, 35, 47, 58, 77f). The ‘model is not level-oriented but complexity-oriented’, with ‘understanding’ applying to ‘words’, ‘clauses’, ‘complex sentences, sequences of sentences, and overall text structures’, and sharing ‘feedback between less’ and ‘more complex units’. ‘The function of a word in a clause’ ‘depends on the functional structure of the clause as a whole’ -- a further reason to ‘operate with a strategic model’, not a ‘conventional structural’ one (cf. 11.33ff). Thus, van Dijk and Kintsch adopt a semantic approach for both local and global structuring.

11.39 ‘Ideally’, an ‘explicit processing model would take text as its input and derive a semantic representation’, as some ‘parsers’ do for rather ‘restricted domains’ of ‘English’ (SD 38). In ‘discourse comprehension models’ and ‘cognitive semantics’, ‘the proposition’ is the ‘fundamental’ ‘cognitive unit’ and the ‘intensional’ or ‘conceptual representation’ ‘assigned to sentential surface structures’ (SD 109, 112f, 124) (cf. 3.36, 44f; 8.55; 9.72, 9²⁴). Van Dijk and Kintsch also ‘take propositions for granted as theoretical units of a cognitive model’ and ‘formulate’ ‘typical psychological operations’ and ‘strategies for (re)constructing’ them (SD 125). The ‘theory assumes’ that during ‘comprehension’, ‘verbal input is decoded’ into ‘propositions, which are organized into larger units on the basis of knowledge structures to form a coherent textbase’ (SD x, 109; cf. 11.50). ‘Complex propositions’ ‘are expressed by clauses and sentences’ and ‘represent facts in some possible world’ (SD 109, 125). That is, ‘propositions’ ‘represent possible facts’ but during ‘understanding’ are ‘instantiated’ to ‘refer’ to ‘specific facts’; and a ‘structured but fuzzy set of categories may be associated with the proposition’ (SD 125). In this way, both the ‘general and specific meaning’ (or both ‘context-free and context-sensitive meaning’, or both ‘sentence meaning and language user's meaning’) ‘are cognitively relevant’ for ‘strategic processes of understanding’; and ‘a model of subjective understanding’ gains ‘a more objective, intersubjective component accounting for general abstract knowledge’ (cf. 11.21, 23; 13.58).

11.40 Some ‘milestones’ are reviewed in the early ‘literature on propositions’ (e.g. Ogden & Richards 1923; Carnap 1942, 1947; Russell 1940; Reichenbach 1947; Quine 1960) (SD 126, 110ff). Despite ‘intricacies’, ‘disagreement’, and confusion’, the main idea of a ‘proposition’ emerges as ‘the meaning of a declarative sentence’ (its ‘intension’) having some ‘truth value’ (its ‘extension’) (SD 110ff; cf. 3.35f; 6.22; 9.72; 11.31). Due to ‘positivism’, this meaning was claimed to be ‘not subjective’ (not ‘a “mental occurrence”’) but ‘an objective conceptual structure’ or even a property of ‘“eternal sentences”‘ free of all ‘contextual factors’ (SD 109f, 125) (cf. 7.73, 79; 11.36). But in ‘more recent theories’, this ‘truth value’ is made ‘relative’ to ‘possible worlds’ (cf. Cresswell 1973; Montague 1974). In ‘linguistics’, meanwhile, ‘the influence of behaviourism’ ‘precluded a systematic study of meaning’ ‘until the sixties’, when ‘sentence meanings’ and ‘semantic interpretations’ came under discussion, and ‘the seventies’, when ‘logical semantics’ was prominent (SD 111f) (cf. 13.17f). ‘Although it is wise in general not to introduce uncritically notions from philosophy, logic, or linguistics into psychological theories of language understanding’, we may, by using ‘propositions’, tap ‘a long tradition’ and formulate ‘constraints of surface structure expression’ as a ‘direct manifestation’ of ‘abstract or underlying theoretical units’ (SD 126). Just as we can ‘couple lexemes with words’, we can ‘couple’ ‘complex semantic units’ ‘with clauses or sentences’. Besides, the ‘proposals from philosophy and logic’ ‘have undergone serious revision in last the ten years from linguists and psychologists’ to accommodate more ‘intuitions about meaning’ (cf. 11.2ff).

11.41 For van Dijk and Kintsch, the ‘proposition’ is a ‘composite unit’ of ‘concepts': ‘a predicate’ for ‘properties or relations’ and ‘one or more arguments’ for ‘individuals such as things or persons’ (SD 113). ‘It would be nice if natural language would respect this distinction in surface structure’ with ‘predicates expressed by verbs and arguments by nouns’. Instead, ‘sentences are usually much more complex’, with ‘not only verbs’ and ‘nouns’ but ‘adjectives, adverbs, modal expressions’, ‘connectives’, and so on (SD 113, 125). The ‘logical analysis’ of these ‘categories’ and ‘structures’ ‘has met with extremely difficult problems’ and become too ‘complex’ to use for ‘representation formats’ ‘in a cognitive model’ (SD 113) (cf. 13.17). If we want to ‘account for the so-called semantic roles or cases’ in the ‘structure of a sentence’ (11.28), we find they are often ‘implicit in the ordering of the arguments’ and must be given ‘ad hoc labels’ in the absence of an ‘explicit formal semantics’ (SD 113f).

11.42 All the same, ‘psychological research in the last few years’ shows that ‘propositions’ as ‘semantic units devised’ for ‘linguistic considerations’ can indeed ‘function’ as ‘processing units’ (SD 38). ‘Lines of converging evidence’ include: ‘cued recall studies’, where ‘words from the same proposition are more effective’ in cueing memory ‘than words from different’ ones; ‘free recall studies’ (i.e. without cues), where ‘propositional units’ are ‘recalled as wholes’, even without the aid of ‘preformed associations, familiarity’, or ‘semantic plausibility’; ‘recognition time’, where ‘how fast people read’ and what they can recognize afterwards ‘depend’ ‘on the propositional structure of sentences’; and ‘priming’, where ‘recognition latencies’ between words are less when ‘two words come from the same proposition’, irrespective of ‘closeness’ ‘in the surface structure’ (SD 38-41).²² When ‘textual input’ seems ‘unrelated to the propositions’ ‘in the short-term buffer’, ‘the reader searches episodic memory’ to ‘reinstate’ some ‘proposition’ ‘sharing an argument’ with the ‘input’, or else makes ‘a bridging inference’ (SD 45; cf. 11.21, 48, 65 70, 76, 95, 100). That both ‘operations’ are ‘resource-consuming’ is shown by ‘experimental evidence’ for ‘reading difficulty’ (Kintsch & van Dijk 1978) (SD 45; cf. 11.95f). In sum, ‘the evidence for the psychological reality of proposition units is overwhelming’ (SD 41).

11.43 Though ‘problems’ and ‘arbitrariness’ beset any ‘system’ for ‘representing meaning in ‘propositions’, the ‘analyses have worked very well in practice’ ‘for many purposes, such as scoring recall data’ or ‘representing the semantic level’ in a ‘processing model’ (SD 37f). Researchers ‘learn to propositionalize texts quickly, and the interjudge reliability’ is ‘high’ (cf. 13.51). Of course, such ‘representations’ are not ‘all-purpose’ but only ‘rather primitive’ ‘tools’, and must be tailored to each ‘branch of science’. Since van Dijk and Kintsch ‘do not hold the view that “meanings” or “concepts” are inherently tied to natural language’, ‘propositions’ can also ‘figure more generally in models of comprehension’ for ‘real or pictorial images’, ‘scenes, sequences of events, pictures’ or ‘other semiotic systems’ (SD 113, 62; cf. 11.10, 52f, 61; 13.22). But the ‘discussion’ in SD is ‘limited’ to ‘natural language’ (SD 113).

11.44 ‘For simplicity’ at any rate, ‘a representation’ is adopted that is ‘far from complete’ or ‘adequate’ for ‘linguistics’ or ‘logical semantics’ ‘but is ‘cognitively relevant’ (SD 114, 116). It does not cover ‘all expressions in surface structure’ but ‘only semantic properties’, as compared to ‘pragmatic, stylistic, rhetorical, cognitive, interactional, or social’ ones; for example, no entry is made for the ‘definite article “the”‘ ‘expressing that an individual’ is ‘known or identifiable’ (SD 114; cf. 7³⁹; 11.86). The ‘representation’ is ‘a propositional schema’ in which ‘semantic categories of the meaning of a sentence are represented as the nodes in a tree-like structure’ made of ‘atomic propositions’ as ‘terminal elements’ (SD 113f, x). ‘Each category may have a subordinated modifier category’ (‘adjectives and adverbs’ etc.) for ‘circumstances’ and ‘modals’ that ‘localize the complex proposition’ (SD 116) (cf. 9.66, 79f).

11.45 ‘Interpreting the verb phrase as the proposition predicate’ ‘sets up the propositional schema’, with ‘the topic noun phrase’ being ‘assigned to the agent participant’ and other ‘roles’ being made ‘ready to receive their content’, e.g. ‘time and place’ (SD 158) (cf. 7.63; 9.57). Filling these roles to ‘bind’ the ‘free variables’ or to ‘substitute constants’ makes ‘the action part of the schema’ into a genuine ‘proposition’ which can be ‘true or false’ (SD 116) (cf. 9.72; 11.40). ‘Overall coherence’ among ‘propositions’ is established as ‘relevant information’ is picked via ‘the knowledge schemas activated by the first proposition interpretation’ ‘about possible facts in the world’ and ‘situation’ (SD 158). Insofar as ‘the possible links between facts’ and ‘propositions are limited’, ‘the language user can apply a ready-made strategy': ‘match the proposition’ with a ‘conditional or functional’ ‘category’ (SD 158f; cf. 11.31). ‘The language user searches’ for ‘potential links among facts’, e.g. via ‘identical referents’ (‘objects, persons’ etc.) or ‘related’ ‘predicates, participants, or circumstances’ (SD 15, 150 157, 183; cf. 11.39). Thus, a ‘proposition’ can ‘activate expectations’ and ‘hypotheses’ about the ‘continuation’ based on some ‘coherence link’, and can set up a ‘local coherence goal’ of ‘establishing a relation’ (SD 157). ‘Predicates belonging to the same semantic class’, for instance, yield ‘an obvious semantic link’ (cf. 9.93).

11.46 In ‘our earlier work’ (Kintsch & van Dijk 1978), the chief ‘strategy’ was to look for ‘repeated’, ‘shared’, or ‘coreferring’ ‘arguments among propositions’; but this is only an ‘attractive’ ‘oversimplification’ and ‘reduction’, and is just one ‘by-product’ or ‘example’ of the ‘more embracing strategy’ of ‘relating whole propositions or facts’ (SD 15, 43, 46, 154, 183).²³ Still, ‘relations’ based on ‘argument repetition’ are ‘quite predictive of recall’, particularly ‘in short paragraphs’ (Kintsch & Keenan 1973) (SD 43). And ‘the psychological importance of shared reference has been demonstrated’, e.g., allowing ‘sentences to be read more rapidly’ (Haviland & Clark 1974). But ‘readers’ also build ‘a hierarchical structure of coherence relations’ that ‘is not based on argument repetition’; and ‘hierarchical textbases’ with ‘superordinate’ and ‘subordinate propositions’ ‘predict free recall rather well’, the higher ones being heavily favoured (Kintsch, Kozminsky, Streby, McKoon, & Keenan 1975; Meyer 1975) (SD 58, 44).

11.47 We thus return to van Dijk and Kintsch's major concern, namely the ‘global coherence’ imposed by a ‘theme, topic’, ‘gist, upshot, or point’, all ‘theoretically reconstructed as macrostructures’ (SD 15, 52, 104, 150f, 170, 189f, 193f, 224, 237) (cf. 11.18f, 25, 30, 32). ‘A central component of the model is a set of macrostrategies’ for ‘inferring macropropositions’ ‘from the sequence of propositions expressed locally by the text’ (SD 15). The ‘macropropositions may’ in turn be ‘organized into sequences’ or ‘levels’, leading to ‘the macrostructure of the text’. The ‘macrostrategies’ too are ‘flexible and heuristic’, since ‘the language user’ does ‘not wait until the end’ of a ‘sequence of sentences’ or of ‘a paragraph, chapter, or discourse before inferring’ the ‘global’ content, but ‘guesses’ ‘with a minimum of textual information from the first propositions’ (SD 16, 205; cf. 11.29). ‘Titles, thematic words, first sentences’ ‘settings’, and ‘information from context’ can all contribute (SD 16, 54, 89f, 92, 107, 144, 203, 221f, 361). ‘In some discourse types’, however, e.g. ‘literary or everyday stories, rhetorical devices’ may ‘delay’ such ‘indications’ to ‘arouse interest or suspense’ (SD 221; cf. 11.58, 84).

11.48 ‘Macropropositions may be directly expressed’ and may have their own ‘connectives’ e.g., ‘conjunctions or adverbs’ (‘“however”, “moreover”’) for indicating ‘conditional’ or ‘functional’ ‘coherence structures’ (SD 204ff; cf. 9.87; 11.31). Or, they may be ‘inferred from underlying representations’, ‘organized world knowledge’, and ‘schematic or superstructural’ ‘information’, e.g., about the ‘normal’ ‘ordering’ in ‘a narrative’ (whereas ‘literary texts’ may present ‘propositions that are ‘abnormal and interesting’ or may use ‘abstractness’ to impede ‘the derivation of a macroproposition’) (SD 205f, 207f). In ‘general’, ‘if a sentence’ cannot be ‘subsumed under the current macroproposition’, several options are open: (a) ‘setting up a new’ one; (b) ‘reinstating’ one from ‘memory’; (c) using ‘a wait-and see strategy’; (d) being content with ‘only local coherence’; or (e) just ‘deleting’ the material (SD 204, 206, 208, 221f).

11.49 ‘In the ‘semantics of discourse, macrostructures are defined’ via the ‘macrostrategies’,²⁴ which ‘map’ ‘propositions’ or ‘sequences’ of them onto those of ‘a higher level’ and create a ‘hierarchical’ structure (SD 190, 236). These ‘macrostrategies’ include: ‘deletion’ of a ‘proposition’ that is not an interpretation condition for another’; generalization’ to ‘substitute’ ‘a proposition’ for a ‘sequence’, ‘each of whose propositions’ ‘entails’ it; and ‘construction’ of a ‘proposition’ ‘entailed’ by ‘the joint set’ of a sequence as a whole (SD 190). These ‘rules’ ‘reduce’ materials, but at ‘higher levels’ they may also ‘assign further organization to the meaning of a discourse’.

11.50 ‘The coherent sequence of propositions’ ‘formed’ ‘during comprehension’ is called ‘the textbase’ (SD 11, 44f, 51, 109, 342ff, 371). This ‘textbase’ too is ‘constructed’ ‘in real time’ (‘on line’), as ‘the reader accumulates semantic units’ and ‘adds’ them ‘level by level’ ‘to the fragment’ in ‘short-term memory’ (SD 44, 373; cf. 11.7). To stay within ‘limited short-term memory’, a ‘leading-edge strategy’ carries ‘superordinate propositions’ ‘from cycle to cycle’; ‘if none are available in short-term memory’, one is ‘chosen from the current input’ (SD 44). ‘Superordinate’ units are ‘processed more’ and therefore ‘recalled more’, as studies have shown: ‘the level of a semantic unit in the textbase hierarchy determines the likelihood of its recall’ (Kintsch & Keenan 1973; Kintsch & van Dijk 1978) (SD 44f, 226, 241). We have here an exemplary ‘processing explanation for a structural effect’ (SD 44) (13.31).

11.51 ‘In parallel’ with the ‘textbase’, ‘a situation model is elaborated’ -- a ‘cognitive representation of events, actions, persons’ -- which ‘integrates the comprehender's existing world knowledge with information derived from the text’ and thus supports ‘interpretation’ (SD x, 337f, 11f, 51, 163f, 308, 340ff, 348). ‘The main semantic and pragmatic function of a text is to enrich this model’; unless we ‘imagine a situation’, ‘we fail to understand’ (SD 337f). ‘The situation model’ subsumes ‘relevant’ ‘knowledge’ ‘left implicit’ or ‘presupposed’ by the ‘text’, both ‘general’ (‘semantic’) or ‘specific’ (‘episodic’), and ‘may incorporate previous experiences’ or ‘textbases’ (SD 337f, 344, 12; cf. 11.21). We may be ‘reminded of past situations’ and ‘experiences’ in ‘clusters’, which may offer some ‘analogy’ whereby ‘ill-fitting models are transformed’; ‘in this respect, discourse comprehension is a problem-solving task’ (SD 337f, 245, 346; cf. 11.12f, 25).

11.52 Numerous ‘linguistic and psychological arguments’ are given why the ‘situation model’ is ‘necessary to account for’ ‘discourse comprehension and memory’ (SD 338).²⁵ It ‘fills the gap’ ‘between “meaning” and “reference”‘ (cf. 11.10, 43, 61). It provides a ‘perspective’ or ‘point of view’ from which ‘the facts’ -- not ‘real facts’ but ‘representations of them’ -- are ‘seen, interpreted’, ‘talked about’, and ‘connected’ (SD 339). It handles ‘parameters’ of ‘possible world, time, and location in discourse’, often ‘inferred’. It supplies the ‘individuals’ to which ‘expressions in discourse refer’ in ‘co-reference’ (rather than to ‘other expressions’) (SD 338; cf. 9.89, 9⁴²). It ‘functions’ in ‘updating and relating’ ‘general knowledge and personal experiences’ in ‘memory’, e.g. when ‘an existing model is modified on the basis of a new text’ (SD 342). It can be ‘remembered’ without the ‘text representation’ (e.g. if the latter is ‘difficult to construct’ or entails ‘minimal distinctions’), whereas ‘the textbase’ is ‘rarely reactivated’ (SD 340f, 344).²⁶ It accounts for ‘individual differences in comprehension’ of ‘the same information’, whence the ‘debates about what a classical text “means”‘ (e.g. in ‘literary’ studies) (SD 339f). It ‘forms the basis for learning’ and for taking an ‘action’ upon ‘reading a text (as in ‘problem-solving’ and ‘formal reasoning’ in ‘mathematics and logic’) (SD 344, 341; cf. 11.98ff).²⁷ It is ‘reconstructed’ in ‘retelling a story’ and encourages people to put ‘events’ in the ‘canonical order’ (SD 341; cf. 11.27, 55, 94, 10¹⁴). It provides a ‘link’ for ‘crossmodality integration’ from ‘textual and nontextual sources’ (SD 341). It ‘relates text representations’ in the ‘source’ and ‘target language’ during ‘translation’, particularly when ‘the languages’ ‘differ widely’ in ‘cultural code’ (SD 339).

11.53 This many ‘reasons why a situation model is needed’ might suggest we ‘throw out’ ‘the text representation’ and have ‘just words on the one hand and the situation model on the other’ (SD 342). But ‘text representations’ are ‘necessary’ too, because ‘discourse expresses meanings or refers to facts’ ‘in a specifically linguistic way’, and may be ‘stored’ this way in ‘memory’ (SD 343). So we need the ‘intervening’ ‘text representation’, and theories which dispense with it ‘introduce some notational variant through the back door’. ‘Cognitive scientists’ should be ‘clear about what they attribute to text’ or ‘to the world’ and not ‘confuse the two’ (SD 344) (11.10). Van Dijk and Kintsch recommend ‘limiting the textbase to information expressed or implied by the text’, while other ‘activated knowledge’ goes into ‘the situation model with which the textbase is continuously compared’ (SD 12).

11.54 Like ‘scripts or frames’, the ‘situation model also has a schematic nature’ with ‘variable terminal categories’, which it ‘can instantiate’ and ‘fill’, or can ‘form’ by ‘learning’ from ‘one's own experiences’ and ‘abstracting’ out ‘details’ during ‘frequent use’ (SD 344f, 172; cf. 11.44). ‘The model’ may have ‘a structure’ of ‘propositions’ with ‘predicates’ and ‘participants’ ‘ordered’ by ‘recency’, ‘relevance’ etc. (SD 344f, 361). This ‘format’ ‘can be easily retrieved’ via ‘reminding’, and ‘information chunks from the current text’ can be ‘inserted’ into the ‘categories’ (SD 345f). As ‘a flexible schema’, the ‘situation model’ helps in ‘collecting’ and ‘grouping together’ ‘similar experiences’ and thus in ‘organizing’ ‘memory’.

11.55 Van Dijk and Kintsch further postulate ‘superstructures': ‘typical schemas’ for ‘conventional text forms’, which ‘consist of conventional categories, often hierarchically organized’, ‘assign further structures’ and ‘overall organization to discourse’, and ‘facilitate generating, remembering, and reproducing macrostructures’ (SD 16, 54, 57, 92, 104f, 189, 222, 236f, 242, 245, 275, 308, 336, 343).²⁸ We are assured that ‘superstructures are not merely theoretical constructs of linguistic or rhetorical models’ but also ‘feature in cognitive models’ as ‘relevant’ ‘units’ (SD 237). ‘During comprehension’, they are ‘strategically’ ‘assigned on the basis of textual’ ‘information, i.e. bottom-up’, yet also create ‘assumptions about the canonical structure’ and applicable ‘schema’, i.e. ‘top-down’ (SD 237, 105; cf. 11.13). The ‘superstructures provide the overall form of a discourse and may be made explicit’ as ‘categories defining’ the ‘type’ (SD 189, 235f). They are ‘acquired during socialization’ with ‘discourse types’; ‘language users know’ the ‘categories’ and ‘schemas’ ‘implicitly’ or even ‘explicitly’ and ‘make hypotheses’ about them ‘when we read’ (SD 57, 92).

11.56 These ‘additional organizational patterns’ may apply to ‘the discourse as a whole’, e.g. ‘narrative’ or ‘argumentation’, or to ‘segmented paragraphs’, or to ‘specific’ ‘levels’, e.g. the ‘morphological, syntactic, and semantic’ (SD 235f, 241, 105, 92). ‘Participants in a given situation may expect a range’ of ‘discourse types’ and make ‘strategic guesses’ about a ‘probable superstructure’ ‘according to the culture’, as ‘experiments’ and ‘ethnographic’ ‘studies’ show (cf. Bartlett 1932) (SD 238).²⁹ People can use a ‘discourse as a whole’ to ‘perform a global speech act’, or can use the ‘interactional context’ to make ‘inferences about possible speech acts being performed’ (SD 239). These ‘acts’ and their ‘sequencing’ have ‘systematic links’ to ‘global semantic content’ and to ‘schematic categories’ with a certain ‘ordering’. Hence, ‘text types’ are ‘defined in pragmatic terms’, not merely by ‘surface structure style or semantic content and schemas’. In ‘argumentative discourse’, for example, ‘premises and conclusions’ ‘are linked through a semantic chain of implication, entailment’, and ‘inference’, and through ‘speech acts of asserting, assuming, drawing conclusions’, and so on. A ‘global request’ or ‘recommendation’ might appear not ‘in the introduction category’ but in a later ‘evaluation or coda’.

11.57 ‘Superstructures’ also include ‘metrical or prosodic patterns’ in ‘literary, aesthetic’, or ‘ritual’ texts’, e.g., ‘meter’, ‘rhyme, alliteration, repetition, and figures of speech’ like ‘metaphor and irony’ (SD 92f, 241f). Thus, van Dijk and Kintsch's model is much concerned with ‘stylistic and rhetorical’ aspects (SD 18, 57, 81, 83, 92, 94, 104, 114, 197, 221, 235ff, 241f, 254, 275f, 278, 282, 285, 292, 343). ‘The style of a discourse’ is defined as its ‘variation of grammatical’, ‘schematic, or rhetorical rules or devices’ (SD 94) (cf. 3.69; 5.82; 6.52; 8.83; 9.102; 11.10). ‘In principle, stylistic variation’ correlates ‘alternate ways of expression’ with an ‘underlying identity or similarity’ of ‘theme’, ‘semantic representation’ (‘meaning, referent’ etc.), or ‘speech act’, ‘under the controlling scope of text type and context’ (SD 94, 17).³⁰ This ‘variation’ has ‘highly complex effects’, such as ‘signalling’ ‘the relationship of speaker to hearer’ or of ‘discourse’ to ‘social context’ (as ‘formal, friendly’, etc.), or regulating ‘ease of decoding’ and ‘understanding’ (SD 94, 18). The ‘language user has the task’ of ‘selecting words’ from a certain ‘register’ and providing ‘indicators’ of the ‘personal or social situation’ by ‘strategic use of style markers’ (cf. Sandell 1977) (SD 17f) (9.105).

11.58 ‘Rhetorical operations’ are ‘communicative devices to make the discourse more effective’ (SD 343) (cf. 11.47, 86, 94f). ‘Rhetoric in classical times’ studied ‘effective’ or ‘correct manners of speaking’, especially for ‘persuasion’ (SD 92). But ‘in principle, any kind of discourse’ ‘exhibits’ ‘rhetorical structures, even everyday conversation’ (SD 93). So ‘understanding discourse implies’ some ‘recognition of rhetorical devices’, and a ‘processing model’ needs to ‘specify what strategies a language user applies’ to do this and how they ‘interact with the semantic and pragmatic representation of the discourse’ (SD 92f). We should examine the ‘additional processing’ whereby the devices attain ‘effectiveness’, ‘assign’ ‘additional structure’, and ‘facilitate semantic comprehension’, ‘organization’, and ‘recall’ (SD 93, 18, 241). Or, ‘rhetorical devices’ may ‘relate the semantic representation to personal experiences, or to episodically or emotionally relevant information’, e.g. by ‘vividness’; or may ‘signal the macrostructures of a text’ by ‘pointing to what is important’ and ‘highlighting the theme’ (SD 93, 18; cf. SD 254-59). Similarly, ‘representation’ may be ‘connected’ ‘with an evaluation’ by ‘an assignment of additional structures’ leading to an ‘aesthetic effect’, e.g. in ‘literature’ (cf. Dillon 1978; Groeben 1982) (cf. 3.68f).

11.59 Hence, ‘rhetorical form’ gets used in SD alongside ‘superstructure’ to designate types like ‘argument, definition, classification, illustration, and procedural description’ (SD 254). Although ‘forms’ ‘rarely’ appear in ‘pure examples’ and may be ‘combined’ ‘in multiple, unpredictable ways’, they help ‘readers’ to ‘organize the text’ and to apply ‘top-down processing’. By ‘using rhetorical forms’ in the normal ‘order’ or ‘signalling’ the ‘categories’ ‘clearly’, ‘writers’ can convey their ‘intentions’, so that ‘the right rhetorical schema is triggered’ for ‘the reader's’ ‘organization’. If ‘the rhetorical structure’ is ‘hidden’, however, ‘the reader’ may ‘still comprehend’, but ‘miss’ the ‘point’ or ‘intention’. This aspect has in fact been demonstrated by ‘experiments’ with ‘texts’ (for ‘classification, illustration, comparison-contrast, and procedural description’) in which ‘content’ was ‘identical’ but ‘rhetorical organization’ either did or did not ‘conform’ to the proper ‘schema’ (Kintsch & Yarbrough 1982) (SD 254f, 259). ‘Effects’ showed up ‘at the macrolevel’ (probed by questions about ‘main ideas’), not in ‘local processing’ (probed by ‘cloze test’, cf. 11.94) (SD 254f, 257). Moreover, ‘rhetorical form’ did not seem to ‘interact’ with ‘complexity’, being ‘just as helpful with simple texts as with complex ones’ (SD 257f). When the ‘rhetorical form’ was ‘concealed’, however, the ‘complex’ versions were ‘almost unintelligible to our college student subjects’, who either ‘did not form macropropositions’ or formed ‘inappropriate ones’ based on ‘some salient detail’ instead of ‘the main idea’ (SD 259). Data on ‘free recall’ also reveal a major ‘dependence on macrostructure’ (Meyer, Brandt, & Bluth 1980), though ‘micro- and macroprocesses are confounded’ there, as are ‘textual structure’ and ‘the structure of the content itself’ (SD 259f).

11.60 ‘The form’ ‘most widely explored’ so far is ‘the story’ or ‘narrative’, with a ‘schema’ or ‘superstructure’ for ‘forming macrostructures’ whose ‘categories’ are ‘the main events’ (SD 55, 92, 235f, 251). ‘A story’ centres on ‘actors’ and ‘major actions’ that ‘change’ the ‘states’; the ‘goals and actions’ fill ‘the story schema’ (SD 55; cf. 11.11).³¹ Each ‘episode’ consists of ‘actions falling into the categories of exposition’, which ‘introduces the actors and the situation’; ‘complication’, which ‘brings in some remarkable, interesting event’; ‘and ‘resolution’, which ‘returns’ ‘to a new stable state’ (SD 55, 57f, 16, 55, 236, 240, 275). This ‘form can be elaborated’ by ‘embedding’ or ‘concatenating episodes’, or ‘overlapping’ the ‘categories’ (SD 55).

11.61 ‘Recently’, a ‘fierce debate’ arose whether ‘story grammars’ (inspired by Chomskyan notions) are merely ‘theoretical artifacts for factors better ‘explained’ or ‘modelled in terms of the structure of actions’, e.g., ‘motivation, purpose, intention, and goal’ (SD 55).³² Following ‘available data’, van Dijk and Kintsch ‘compromise’ by ‘arguing that narrative schemas and action structures are both necessary for story processing’; ‘not all superstructures can be reduced to action-theoretical categories’ (SD 56f). ‘Stories’ are just ‘a subset of action discourses’ dealing with ‘plans’, ‘purposes’, and ‘goals’, and thus cannot be the only concern of a ‘general’ ‘cognitive account’ for ‘a variety of tasks’. Also, ‘stories’ need to be modelled not within ‘a theory of action’ but within a theory of the ‘cognitive representation’ and ‘description’ of ‘action’, taking account of ‘completeness, level’, ‘ordering, style, perspective or point of view etc.’ (SD 57, 264). ‘Semantic and pragmatic constraints’ ‘conventionalized’ in the ‘culture’ decide which ‘aspects of actions’ should be ‘told’, e.g., the ‘unknown, interesting’, ‘funny, dangerous, unexpected, uncommon’ ones; and ‘the actions’ may be told out of their sequential ‘ordering’ (SD 56f). Moreover, ‘not all action discourses are stories, e.g. police protocols, ethnographic studies, or manuals for repair’.

11.62 ‘Evidence’ has accrued that ‘episodes function as psychological units in story comprehension’ and ‘recall’ (SD 57). ‘The hierarchical structures’ foreseen in ‘story grammars predict recall': ‘superordinate nodes’ fare ‘better than subordinate’ ones, but ‘semantic content’ may ‘override’ this effect, e.g., ‘actions’ being ‘more salient than states’ (SD 58; cf. 11.27f, 46, 50). Also, ‘beginning, attempt, and outcome are usually recalled better than goal and ending’.³³ When ‘the same sentence’ was put ‘in different parts of the story’, ‘subjects took longer to read it’ if it was situated to fit an ‘important narrative function’ (Cirilo & Foss 1980) or to fall at an ‘episode boundary’ (Haberlandt, Berian, & Sandson 1980).³⁴ Still, since ‘narrative categories tend to be confounded with action schemas’, we need also to ‘investigate texts whose semantic content and rhetorical form are less interwoven’, i.e. ‘nonnarrative’ ones (SD 59; cf. 11.97-100).

11.63 The ‘“topic”‘ is another key factor ‘in the cognitive processing of textual information at the semantic level’ (SD 182) (cf. 5.34, 59; 11.34, 45, 47, 86f, 89, 96). ‘Topics function both as instructions to search the text representation’ and as ‘indicators of how and where to connect propositions of the textbase’ (SD 156, 171). If we had a ‘theory of the internal relevance structure of sentences’, we could specify ‘degrees of topicality and focus’ on a ‘schematic’ basis (SD 171, i.r.). This could capture ‘the general cognitive (and hence universal) property that some semantic information is linked with the previous’ and is ‘more relevant for the continuation of the discourse’ and its ‘coherence’ (cf.11.30). The ‘relevance structure’ could ‘assign functions such as “topic” or “focus” to nodes in the semantic representation’, could be ‘scanned’ for ‘antecedents’ to be ‘retrieved’, and so on. Presumably, ‘the favoured positions for relevant antecedents’ are (1) ‘last occurring, (2) main clause/main proposition, (3) first position, (4) subject, (5) agent/person, and (6) topical noun phrases, in this order of increasing importance’ (cf. 11.28, 34, 45, 64, 68, 79, 86).

11.64 Several ‘functions’ or ‘levels of topicality’ are ‘differentiated’ for ‘information’ ‘“in focus”‘ (SD 169f, 181). ‘The sentential topic’ in the ‘vast’ ‘linguistic and psychological literature’ is ‘a function assigned to a part of the semantic representation’, ‘often marked in surface structure’, e.g., by ‘initial position in English’ (as in ‘the first noun phrase’, ‘especially a definite’ one or a ‘pronoun’, 11.34, 68) (SD 169f, 156) (cf. 7.63; 9.46, 57). So far, though, no ‘explicit representation format’ or ‘adequate formal definition’ has been given for ‘topic functions’ based on ‘intuitions and linguistic data from various languages’, e.g., ‘word order phenomena and topic markers’ given by ‘morphemes’ (SD 155, 167, 171, 182). As ‘a cognitive definition’ for this level of ‘topicality’, van Dijk and Kintsch consider ‘the “topic”‘ ‘a discourse function of the sentence’, ‘exhibiting partial coherence with the (con)textual representation of the previous part’ (SD 155). This ‘function’ ‘selects an element (a subtree)’ to use in ‘constructing the next propositional schema’ -- an ‘account for the overlap defining semantic relatedness’ and ‘continuation’, as found in ‘the stereotypical manner of discourse production’ (SD 155, 170; cf. 11.87ff). Thus, ‘sentential topic’ flows into ‘sequential topic’, which ‘represents a participant’ ‘for a sequence of sentences’, even ‘discontinuous’ ones (SD 169f, 181).

11.65 Just as ‘macropropositions control processing in short-term memory’ (11.32, 47f), ‘macrotopics lead to expectations’ and ‘interpretations’ for ‘sentence topics’ (SD 170). Here, ‘topic functions’ are assigned to ‘complex semantic elements’ in the ‘cognitive process of expanding and linking information in discourse representations’ and of ‘keeping or reinstating concepts in short-term memory’ while integrating ‘new information’ (SD 181). ‘Readers’ may ‘maintain macroparticipants’ and ‘sequential topics’ as ‘central referents’ with ‘the strongest claim for local topicality’. Experiments where readers had to ‘write a likely continuation’ for a short ‘paragraph’ showed them ‘basing their expectations’ on ‘topic’ rather than on ‘local sentence properties’, but ‘reverting’ to the latter if no topic was ‘available’ (Kintsch & Yarbrough 1982) (SD 325, 328).

11.66 ‘Sets of possible topics’ are constrained by ‘discourse type’, ‘communicative context or situation’, ‘culture or subculture’, ‘social’ ‘roles’, or even by ‘sex, age, or personality of speakers’ (SD 197ff, 200). Such a topic ‘set may be ordered’ in ‘a hierarchy’ of likelihood or acceptability’ and may have ‘degrees of freedom or boundedness’. Thus, ‘contextual information’ for ‘possible topics’ can be ‘reduced to a manageable size’ (SD 200). Yet so far, ‘topic sets’ and their ‘precise forms, order’, and ‘constraints’ have received little ‘systematic research in linguistics, sociology, or anthropology’ (SD 197). We still need ‘a cognitive theory of discourse understanding’ that ‘incorporates a model of language users’ applying ‘macrostrategies’ to ‘decide which topics are functional’ in ‘the global or local context’ (SD 200f).

11.67 Although ‘macropropositions’ can readily be ‘inferred from semantic interpretations’, ‘topical expressions’ can also be indicated in ‘surface structure': they can ‘precede or follow a discourse’ (e.g. ‘titles’, ‘summaries’), or be ‘expressed in independent sentences’, or be signalled by ‘type styles’, ‘highlighting’, and ‘paragraph indentation’ in ‘written discourse’ or by ‘intonation, stress’, and ‘pausing’ in ‘spoken discourse’, and so on (SD 201-05; cf. 11.85). ‘Major cues’ for ‘macropropositions’ range from ‘purely grammatical features’ and individual ‘key words’ to ‘sequences of sentences’ (SD 205, 202f, 182). ‘Syntactic signalling’ can ‘indicate’ ‘local importance’ and focus’ with a ‘passive’ or ‘cleft sentence structure’; and can ‘foreground information’ by means of ‘super-’ versus ‘subordinate’ or ‘first’ versus ‘final’ ‘clauses’ (SD 203) cf. 9²², 9²⁴; 11.85). ‘In English’, ‘final, stressed position’ is ‘preferred’ for ‘newness’ and ‘focus’, but ‘deviations’ from this can ‘mark contrasts’ or ‘breaches of expectations’ (cf. 9.69f).

11.68 In all these ways, ‘the syntactic structure and meaning of the current sentence’ get ‘analysed’ for ‘topical function’ (SD 170). ‘If the discourse referent is a human being, first its role as an agent will be preferred': ‘hence sentence topics’ are often ‘subjects of sentences’ and ‘agents or causes of predicates’; if not, ‘a different role is specifically signalled’ (SD 281) (cf 9.46, 57). For instance, if a ‘first position pronoun’ triggers a ‘search for an antecedent’ with ‘topical function’, the ‘strategies’ ‘operate more reliably and faster’ when the ‘topic’ also has ‘agent function’ (SD 170f, 181f, 157; cf. 11.28, 34, 45, 63f). This ‘cotopicality strategy’ ‘operates whether or not a pronoun is structurally ambiguous’ (SD 170). Sometimes, the ‘strategy assigns only partial’ ‘provisional coherence’, pending a ‘definitive interpretation’ based on a ‘whole clause or sentence’ and on ‘links with previous sentences’ (SD 171). As usual, the most crucial ‘criterion’ ‘is the accessibility in short-term memory’ of such ‘information’ as ‘frames, scripts, situation models, and macropropositions’ (SD 172) (11.23).

11.69 As we see again, van Dijk and Kintsch believe ‘the process of comprehension cannot be understood’ without considering ‘current memory theory’ --- fortunately an ‘advanced’ ‘field’ of ‘research’ with a ‘consensus’ about ‘the major phenomena’ ‘studied in the laboratory’ (SD 60). ‘Memory’ has been found to depend both on ‘strength of encoding’ and on ‘retrieval operations’ (SD 357). The current ‘consensus model of memory’, proposed by Raaijmakers and Shiffrin (1981), ‘is sufficiently formalized’, ‘accounts for standard laboratory phenomena’, and ‘incorporates the major features of memory models of the past decade’ (SD 295, 297). ‘The model assumes an associative network with complex nodes containing sensory, semantic, and associative information’, e.g. ‘word concepts’ or ‘propositions’ (SD 298). ‘The probability’ of a ‘retrieval’ depends on ‘the relative strength of the association’. During ‘retrieval’, ‘a probe’ with ‘an array of cues’ for ‘context’, ‘task’, or ‘topic’ is ‘held in short-term memory’; ‘retrieved’ ‘items’ get ‘added to the probe’, possibly ‘displacing others’. So ‘retrieval dynamically changes the memory structure itself’ (the ‘cue’ or ‘probe’ or the ‘associative strength’), possibly creating ‘output interference’ (SD 296, 298). ‘Implicitly’, ‘the retrieval operation is always successful’ but ‘the item’ may not be actually ‘recovered and produced’ if ‘strength is too low’ (SD 298). The ‘primary’ concerns’ are ‘the number of retrievals’ or ‘failures’ ‘before stopping’ or else ‘purging the probe’, plus ‘the strength increment between a cue and a retrieved item’.

11.70 Memory constraints are of two types. First, ‘short-term memory capacity is limited to about four chunks’, or less when ‘resources’ are in heavy demand (SD 335; cf. 11.76, 94). Second, ‘retrievability’ is ‘limited’ because ‘the retrieval cue must match, at least partially, the encoded item’, which is then ‘reinstated in short-term active memory’ (SD 335f). ‘Effectiveness’ is raised by ‘operating within a retrieval system’ that supplies ‘integrated memory episodes’, not ‘isolated’ ‘traces’ -- an aspect in which ‘unorganized word lists as used in classical studies’ differ from ‘discourse’ (cf. 11.93f; Beaugrande 1985).

11.71 It would be ideal to ‘get at memory retrieval in its simplest, purest form’ and decide if this is ‘identical’ with the ‘retrieval studied in laboratories for the last two decades’ (SD 295). Provisionally, van Dijk and Kintsch suggest that ‘memory is a by-product of processing’ and ‘recovers’ things according to the ‘depth’ and ‘elaboration’ of this ‘processing’ (SD 335). ‘Memorability’ depends on ‘semantic, meaningful encoding, and embedding experiences in a rich accessible matrix’ -- just what occurs in ‘discourse comprehension’ (SD 335f). In the ‘usual episodic memory task, the subject is presented with some items’ and ‘later asked to recall them’, and ‘learning the items consists in associating them with an experimental context’ that serves as a ‘retrieval cue’; in discourse memory, however, the ‘cue’ is ‘an association with some topic’ (SD 295).

11.72 ‘Most discourse processing models assume that during comprehension a language user gradually constructs a representation of the text in episodic memory’, including ‘surface, semantic, and pragmatic information’, and ‘schematic superstructures’ (SD 336) (cf. 11.9, 29, 31, 35f, 58, 86). Of course, ‘all the information’ ‘processed’ in ‘discourse comprehension’ does not make it into ‘short-term memory’, nor is it ‘conscious’ (SD 335) (cf. 2.35, 2¹⁶; 13.72). For usual ‘purposes’, ‘comprehension’ aims at ‘memory not for the discourse’ but ‘for what the discourse is about’ (SD 336). So ‘the problem is': ‘how many’ ‘knowledge elements’ ‘become part of the text representation’ for ‘memory’? To keep it ‘relatively uncontaminated’, van Dijk and Kintsch allow only what's ‘necessary to establish coherence’, as opposed to ‘much richer text representations’ (e.g. Graesser 1981) (SD 336f).

11.73 ‘The propositional structure’ and ‘macrostructure’, by yielding a ‘coherent, interrelated network’, ‘form an effective retrieval system’ (SD 348). ‘Retrieval’ ‘follows’ the arrangement of ‘the textbase’ and ‘the situation model’, working from a given ‘text element’ to those ‘directly connected’, which in turn become ‘starting points’ for new ‘operations’ (SD 357; cf. 11.50f). In this way, many ‘paths’ among ‘nodes’ arise, and ‘if a textbase is fully coherent’, ‘all elements can be retrieved in principle’ by ‘starting anywhere’. In ‘top-down’ ‘recall’, though, ‘retrieval’ ‘starts at the top node and proceeds to lower nodes in the text representation’, favouring ‘propositions’ that ‘fill a slot in the schema’; and if ‘operation is probabilistic, retrieval failures accumulate as the number of nodes’ ‘traversed along a path increases’.

11.74 To run their whole model van Dijk and Kintsch postulate an ‘overall control system’ ‘fed’ by ‘information about the type of situation’, ‘discourse’, ‘plans’, ‘goals’, and ‘schematic superstructure’ or ‘macrostructure’ (SD 12). ‘This control system will supervise processing in short-term memory’, ‘guide effective search’ in ‘long-term memory’, ‘activate’ ‘episodic’ and ‘semantic knowledge’ and ‘situation models’, collate ‘higher’ and ‘lower order information’, ‘coordinate’ ‘strategies’, and so on (SD 12, 350). Thus, ‘the control system’ manages ‘strategies’ for ‘producing information’ and ‘representations’ that are ‘consistent with the overall goals of understanding’, and for ‘incorporating all the information’ which ‘the short-term buffer’ ‘cannot keep in store’ (SD 12). For example, ‘the most recently constructed macroproposition’ and ‘situation model’ are kept ‘directly available’ to ‘influence ongoing processing at other levels’ (SD 350).

11.75 The total scheme (Fig. 11.1) foresees three ‘interacting memory systems': ‘the sensory register, which briefly holds incoming perceptual information and makes it available to the central processer’; ‘text memory’, which includes the ‘surface memory, the propositional textbase’, ‘the macrostructure’ and ‘the situation model’; and ‘long-term memory’, which includes ‘general knowledge and personal experience’ (SD 347f; cf. 11.21, 31, 51, 58).

-- INSERT FIGURE 11.1 HERE --

‘Surrounded’ by these three ‘memory systems’ and linked to them via the ‘control system’ is ‘the central processor’, where ‘all cognitive operations take place’ (except ‘retrieval’) (SD 348). Here, ‘resource limits’ constrain ‘the amount of processing’, but can be ‘circumvented by automatizing’, whereas ‘data limits’ constrain the amount of ‘information’ and can be offset by ‘chunking’ (SD 349, 334f) (cf. 11.27, 33, 38, 70). And since the ‘control system’ itself is ‘not directly conscious’ or ‘limited’, ‘many more elements’ can ‘participate in discourse processing’ than would fit into the ‘active, conscious’, and ‘capacity-limited core’ and can thus be included ‘in the model’ when testing its ‘predictive power’ (SD 350f).

11.76 ‘Short-term memory’ ‘maintains’ ‘the current chunk’ (a ‘complex proposition’ often ‘corresponding to some phrase or sentence’ or to several ‘simple’ ones), ‘plus some carry-over from the previous chunk to establish coherence’ (SD 349). Thus, the buffer might hold ‘the surface representation of the most recent’ ‘sentence’ and ‘the atomic propositions derived from it’ as well as some ‘stripped down version’ and ‘main slots’ of ‘previous’ ones. ‘At this point’, ‘operations’ are ‘strategically controlled and differ’ by ‘situation’, so they are hard to stipulate ‘precisely’ (cf. 13.52). ‘Previous work’ suggests ‘the buffer is limited to three atomic propositions’, e.g. the ‘top slots in the complex proposition’, while ‘surface expressions’ like ‘modifiers’ for ‘time and location’ etc. are ‘discarded’ ‘from short-term memory’ as soon as ‘propositional information is computed’. When ‘needed’, ‘information’ can be gotten back by a ‘reinstatement search’ of ‘text memory’ (11.42, 48, 65, 70); the ‘resources’ for this would vary by how many ‘sentences away’ the search travels (SD 350).

11.77 Despite ‘the emphasis’ on ‘higher order’ ‘processes’, van Dijk and Kintsch do not overlook ‘lower order’ ones, such as the ‘graphic analysis’ of ‘letters’ and ‘words’ (SD 59, 22) (13.33). Like memory, the ‘perceptual processes’ involved also are addressed by ‘a well-developed field of research with a rich empirical data base and a history of instructive theoretical controversies’ (SD 59, 21). We may find ‘analogues at this level’ for the ‘higher level’ ‘processes’, and the ‘theoretical framework’ from the lower may ‘form the basis’ for the higher (SD 21).³⁵ ‘The most basic result’ from studies of ‘letter identification and word recognition’ is that ‘the perception of letters is influenced by knowledge about words’, and ‘the recognition of words’ by ‘the sentence context’ (SD 22; cf. 11.33ff). ‘Thus, recognition is not simply a sequence’ of ‘bottom-up processes’ ‘starting with feature detection and letter identification and continuing through word recognition and sentence parsing to more global discourse processing’. ‘Interacting’ with such activities is a ‘top-down’ ‘process’, with ‘higher levels’ ‘affecting the lower ones’ (SD 22f; cf. 11.13; 13.32). For example, ‘words are easier to retain than random strings of letters’, especially ‘if perception is fragmentary’; indeed ‘being part of a word makes the letter easier to see’ (Reicher 1969). Similarly, ‘words are easier to perceive’ ‘in a meaningful sentence’ or ‘text’ (Tulving & Gold 1963; Wittrock, Marks, & Doctorow 1975). Such ‘context effects’ are due both to ‘automatic facilitation of perception’ and to more ‘controlled hypothesis testing’ (Stanovich & West 1981).

11.78 ‘Theories of word recognition’ are accordingly ‘interactive models’ with ‘bottom-up’ and ‘top-down processing’ on several ‘levels': even ‘word identification is not a single process or unitary skill’, and ‘overwhelming’ ‘evidence’ rules out ‘the traditional view of reading as extricating information from text’ (SD 25f; cf. 11.13). Because ‘perception works’ by ‘processing all kinds of information’ rather than ‘filtering out everything that is not relevant at the moment’,³⁶ ‘word identification’ ‘works as a parallel system that fully analyses the input for all possible interpretations and picks out what it needs’ (SD 34). ‘All the information’ is used at ‘the cost’ of ‘brute force calculation’ but in return, ‘decisions’ are ‘fully informed’ -- a `respect’ in which ‘perception’ starkly ‘differs’ from ‘higher mental processes’ (SD 35). The whole ‘system’ has ‘the form not of a strict hierarchy’ but of ‘a cascade’ in which ‘output’ from one ‘level feeds not only into adjacent levels up or down’ but also into ‘more distant’ ones (SD 25). Yet a ‘completely interactive system’ might attain ‘horrendous’ ‘complexity’; the actual degree of interaction must yet be determined by ‘empirical’ means or by ‘theoretical simulations’ on ‘feasibility and efficiency’ (cf. 11.88f).

11.79 If ‘understanding sentences’ operates like ‘perceiving words’ (and ‘there are no a priori reasons’ why it should), people might ‘work on many possible parsings in parallel’, contrary to what ‘introspection suggests’ (SD 34f). But the two operations ‘differ’ because words ‘are relatively fixed chunks in memory and their retrieval is highly automatized’, whereas people ‘do not automatically retrieve sentences, let alone discourse meanings; rather, we construct them’. Also, ‘irrelevant alternatives’ for a whole sentence could lead to ‘combinatorial explosion’ (cf. Woods 1970). All the same, ‘parsing’ may entail ‘extensive calculations’, as has been argued from ‘a computational standpoint’ (Woods 1980). Either ‘one alternative is explored, information’ about ‘others being carefully stored’ for possible ‘backtracking’; or ‘several’ are ‘followed up in parallel’ (‘virtual processing’) ‘until a choice’ can be made by a ‘higher-level plausibility judgment’. For instance, ‘evidence’ suggests that ‘people compute all possible referents for a pronoun': ‘sentences with an ambiguous pronoun’ (like ‘“The city council refused to grant the women a parade permit because they advocated violence”’) ‘take much longer to read’ (Frederiksen 1981). But ‘even if computations are cheap, there must be limits’, e.g., making ‘clause boundaries serve as decision points’ for ‘selecting’ some things and ‘discarding’ others (SD 35f; cf. 11.33ff). Experiments indicate that ‘completing ambiguous phrases’ is harder than ‘unambiguous’ only if no ‘clause boundary’ occurs before the point for making the completion’ (Bever Garrett, & Hurtig 1973). ‘Sharp breaks’ in ‘verbatim recall’ also ‘establish’ ‘the relevance of sentence and clause boundaries’ (Jarvella 1979) and ‘indicate that the syntactic structure of sentences is used’ in ‘scheduling’ ‘short-term maintenance’ of ‘discourse’ (SD 353).

11.80 In various ‘experiments’, ‘simple sentences’ ‘appeared to function much like’ ‘single words in traditional list-learning’ (SD 353f; but cf. 11.70; 13.28). The usual ‘interference’ ‘effects’ were found for ‘discourse’, e.g., that ‘reading’ takes ‘longer times’ if interrupted by ‘doing addition’ or ‘counting’ (Glanzer, Dorfman, & Kaplan 1981). These ‘laboratory’ findings were confirmed by ‘observations’ in a ‘naturalistic situation’, namely ‘how shopkeepers’ ‘answered simple questions’ ‘over the telephone': ‘the wording of the answer’ ‘reflected’ that of ‘the question’ unless ‘short-term memory’ encountered ‘interference’ (Levelt & Kelter 1982).³⁷

11.81 So far, ‘most work in psycholinguistics is about comprehension’, using ‘the structure of utterances’ as an ‘independent variable’ subject to ‘adequate control’ (SD 261ff).³⁸ For many years, ‘structuralism and behaviourism favoured the analysis of observable phenomena’, such as `surface structures’. Also, ‘generative transformational grammar’, ‘despite its claims’ to be ‘neutral’, was ‘biased toward analysis’ (cf. 7.83f). But ‘a complete discourse processing model should include production’ as well, and van Dijk and Kintsch treat at least ‘some problems’ and ‘strategic aspects’ (SD 16). One main ‘insight’ is that ‘production is not simply the reverse of comprehension’ (SD 261, 16) (cf. 7.83; 11.85; 12.47; 13.57). ‘The initial data and the goals differ’, as do ‘the strategies’ (SD 262, 16). Yet the ‘processes’ can hardly be ‘completely separate’, since ‘comprehension’ too is ‘constructive’ (SD 262, 17).

11.82 So we should ‘specify which structures and principles’ work ‘in either direction’ and which do not (SD 262). ‘Insights’ about ‘general or comparable cognitive mechanisms’, e.g. ‘episodic and short-term memory, help us to specify’ the ‘initial internal representation’ and ‘the constraints in production’ and thus to ‘develop’ ‘hypotheses and experimental techniques’ (SD 264). Since ‘a language user’ cannot ‘construct a long sequence of propositions’ as ‘input to the surface structure formulators’, ‘strategic’, ‘fast’, and ‘flexible’ processes are needed to handle an ‘enormous’ ‘amount of information’ for ‘constructing semantic representations, lexical expressions, and syntactic and phonological structures’, while ‘taking into account’ ‘goals, local and global constraints’, and ‘fluctuating’ ‘contextual information’ (SD 264, 267; cf. 11.10, 20, 53).

11.83 Clearly, ‘the production’ of ‘utterances’ is ‘a complex task which needs planning’ -- a factor ‘neglected in the sentence production literature’ (SD 263, 17, 265). ‘Whereas at the sentence level this planning may be’ ‘automatic’, it may be ‘conscious’ for a ‘complex’ ‘discourse’, especialy a ‘written’ one (cf. Miller, Galanter, & Pribram 1960; Clark & Clark 1977; Hayes & Flower 1980) (SD 263, 272; cf. 11.11).³⁹ For van Dijk and Kintsch, ‘the actual production of discourse begins’ with a ‘plan’ that has both ‘pragmatic and semantic representations’ (SD 265f, 17, 272, 276, 279, 289, 293) (11.17). The plan foresees ‘a series of preparatory, component, and terminating speech acts’, and ‘the ultimate goal’ is to ‘say something about reality’, give ‘new information’, or apply ‘persuasion’, in order to ‘change’ ‘the knowledge, beliefs, or opinions of the hearer’ (SD 266f, 269f, 277; cf. 11.20). So a ‘macro speech act’ has ‘a global propositional content’ (a ‘macrostructure’, ‘macroproposition’, ‘theme, or topic’) ‘derived from the interaction context and its cognitive counterparts’ (SD 266, 272ff, 279f, 284, 289ff). ‘The overall speech act may be indirect’, leaving ‘interpretation to the hearer’ who can ‘draw other conclusions’ without seeming ‘uncooperative’ (SD 269). Also, ‘monological discourse, such as a lecture, scholarly article, or a news story’, allows less ‘interaction’, though ‘implicitly taking it into account’; and ‘planning may be more conscious and explicit, and its execution better controlled’.

11.84 The ‘principles and strategies’ of ‘production’ must ‘leave a lot of freedom in the actual formation of a textbase’ (SD 291). ‘The basic strategy for textual meaning production consists in selection of one or more arguments’ (e.g., ‘discourse referents such as persons or objects’), ‘to which a series of predicates is systematically applied’, e.g., to supply the ‘properties and relations’ ‘associated’ in a ‘schema’ or ‘frame’ (SD 281f; cf. 11.23). Then, ‘production’ undergoes ‘linearization’ by following ‘an appropriate order’, e.g. ‘a natural order’ ‘parallel to the temporal or conditional order of facts’ or going from ‘general to particular’ (SD 275, 281) (cf. 7.3). Or, ‘deviations’ can arise from ‘cognitive reordering’ to fit ‘perceptions, understanding’ etc., or from ‘rhetorical’, ‘interactional’, and ‘pragmatic reordering’ for ‘effective execution of speech acts’ or for ‘aesthetic functions’ and ‘suspense’ (SD 276f, 282; cf. 11.47). To gain ‘cooperation’, we may ‘delay’ ‘information’ that would be ‘difficult’ to ‘accept’, or may ‘give’ ‘conclusions first’, and so on (SD 277; cf. 11.56).

11.85 To work down from ‘global information’, van Dijk and Kintsch propose ‘the inverses’ of their ‘macrostrategies’ for ‘comprehension’ (in 11.49): ‘specification, addition’, and ‘particularization’ (SD 267, 274, 278). The resulting ‘distribution’ depends on ‘complexity’, ‘importance’, and ‘relevance’, high degrees of which call for ‘independent’ units like ‘clauses’ or sentences’ rather than ‘dependent’ ones like ‘modifiers’ or ‘relative clauses’ (SD 282f, 202; cf. 9²⁴; 11.67). A common ‘strategy’ is to ‘start from given information’ (e.g. in ‘the first noun phrase functioning as subject and topic’ of a ‘sentence’) and go on to the ‘new’ (e.g. in ‘a predicate phrase, functioning as the comment’ (SD 279ff, 283; cf. 9.51, 57; 11.28, 34, 46, 63f, 68). In such ways, the ‘global plan’ or ‘representation’ again ‘controls the local, linear’ (‘lower’) ‘levels of discourse production’ in ‘sentential structure’ (SD 266, 283; cf. 11.19). The ‘plan’ may not be ‘conscious and orderly’ but ‘sketchy’ and ‘general’ (or, in ‘literary prose’, ‘well hidden’) (SD 266, 17; cf. 11.47, 84). Or, ‘control’ may be ‘data driven’ and ‘episodic’, following only ‘the plan’ to ‘keep up the conversation’; or, ‘plans may be changed during speaking’ (SD 266f, 269, 17, 273).

11.86 Finally, ‘propositions’ must be ‘given to the sentence formulation mechanism’, where ‘syntactic form is constructed’ from ‘semantic and pragmatic information’ plus ‘lexical and phonological expressions’ (SD 278f) (cf. 7.67; 9.20). ‘The process of lexicalization will select appropriate lexical items to express the concepts of the propositions’, keeping within the ‘bounds’ of ‘style, register’, ‘text type, and communicative context’ (e.g. ‘metaphor, irony’ etc.) (SD 292). As in comprehension, ‘coherence conditions’ can call for ‘explicit’ ‘surface structure signals’ that ‘depend on and determine textual structures’, such as ‘boundaries’ of ‘sentences, clauses’ and ‘episodes’, ‘word order’, ‘semantic roles (cases)’, ‘topic-comment’, ‘connectives, pronouns’, ‘adverbials’, ‘definite articles’, ‘demonstratives’, or ‘tense and location markings’ (SD 279f, 282-85, 292) (cf. 11.35, 40, 48, 64, 67, 75f). Or, ‘control’ may be exerted by ‘feedback from the surface structure’ or from concurrent ‘nonverbal’ events (‘gestures, facial expressions’ etc.) (SD 279, 266). So again, a ‘sentence not only expresses its own meaning but also the multiple links’ ‘with the whole text and communicative context’ (SD 285) (cf. 5.56f; 9.16; 11.91). ‘On full analysis’, ‘few surface structure items’ do not ‘signal a semantic, pragmatic, cognitive, social, rhetorical, or stylistic function’ (cf. 11.35). Hence, ‘little is left of the old Saussurian arbitrariness in the relation between expressions (signifiers) and their meanings (signifieds)’ (cf. 2.28; cf. 4.27; 9.13, 32, 36; 13.27).

11.87 In tests carried out by Donna Caccamise (1981) applying the ‘model of memory’ of Raaijkmakers and Shiffin (1981), ‘subjects’ were given ‘topics’ and told to ‘talk about’ them, ‘saying everything that came to mind, without regard to organization or repetition’ (SD 293). The ‘topics’ were varied between (a) ‘familiar’ or ‘unfamiliar’ (‘“education”‘ vs ‘“energy”’) and (b) ‘general’ or ‘specific’ (e.g. ‘“energy”‘ vs ‘“nuclear energy”’) (SD 296). Because ‘subjects’ ‘generalized the specific topics’, only the first variation led to ‘large and regular differences': ‘familiar topics produced twice as many ideas’ and so more ‘chunks’, while the ‘unfam