5 Chapter Context Models in Discourse Processing Teun A. van Dijk University of Amsterdam THE NEGLECT OF CONTEXT IN PSYCHOLOGY Linguists, discourse analysts, and psychologists generally agree that context crucially influences the structures and processing of text and talk. However, whereas they have developed sophisticated theories of discourse structure and comprehension, the detailed structures of context and how these constrain language use have received much less explicit attention. If context is taken into account in the psychology of text processing at all, it is usually reduced to one or more independent variables that are assumed to affect text understanding, such as goals, task demands, previous knowledge, gender, age, or different types of readers. Although interest in contextual constraints is increasing in psychology, contextual analysis itself remains marginal when compared to the attention to the role of variable text structures and genres, inferences, knowledge, and their mental processing (Graesser & Bower, 1990; van Oostendorp & Zwaan, 1994; Weaver, Mannes, & Fletcher, 1995). Thus, the notion of context may be wholly absent in the subject index of representative recent books on text understanding (Britton & Graesser, 1996). Linguists and discourse analysts have paid a great deal of attention to the role of context, but have failed to develop explicit theories of text- con text relationships. As is the case in psychology, most sociolinguistic accounts tend to examine such relationships in terms of simple covariation, instead of analyzing the precise nature and strategies of contextual influ- 124 VAN DIJK ence. Following the early work of Dell Hymes and his SPEAKING model of context (Hymes, 1962), ethnographic approaches have so far been most interesting (Auer & Di Luzio, 1992; Duranti & Goodwin, 1992; Gumperz, 1982). In another development influenced by anthropological linguistics, functional-systemic linguistics and social semiotics show how the structures of discourse are to be defined in terms of the main dimensions of the context of situation or register, such as ongoing action, participant roles, channel, and symbolic purpose (Halliday, 1978; Martin, 1992). The most extensive work on context has been carried out in the social psychology of language (Brown & Fraser, 1979; Giles & Coupland, 1991), following various approaches to the social psychology of situations (Argyle, Furnham, & Graham, 1981; Furnham & Argyle, 1981; Forgas, 1979, 1985). Thus, Brown and Fraser (1979) presented a situation schema consisting of components such as Scene—consisting of Setting (Bystanders, Locale, Time) and Purpose (goals, tasks, topic)—and Participants and their various properties and relationships. Wish and Kaplan (1977), using multidimen- sional scaling, identified five basic dimensions people use in the interpretation of social situations: (a) co-operative—competitive, (b) intense—superficial, formal—informal, (d) dominant—equal, and (e) task-oriented—nontask- oriented (see also Forgas, 1985; Giles & Coupland, 1991). Against this background of theory formation in psychology, linguistics, and discourse analysis, the present chapter first argues that, strictly speak- ing, contexts do not directly influence discourse or language use at all. Rather, it is the subjective interpretation of the context by discourse participants that constrains discourse production, structuration, and understanding (see also Giles & Coupland, 1991). That is, given a communicative event in some social situation, its participants actively and ongoingly construct a mental representation of only those properties of this situation that are currently relevant to them. Herbert Clark (1996) recently developed a the- ory of some elements of such represented situations in terms of the com- mon ground participants share and extend during joint discursive and other action (see also Barwise, 1989; Cohen & Siegel, 1991). Second, extending earlier work on mental models, it is argued that such subjective interpretations of contexts are to be represented in specific models stored in episodic memory (viz., context models). Such context models are assumed to exercise the crucial overall and local control over all processes of discourse production and comprehension. A detailed analy- sis of these control strategies does not merely show that context (indirectly) shapes text and talk, but also how this happens exactly. In other words, context models are the necessary cognitive interface between text and context. Thirdly, I show that context models are a special case of a more general kind of model (viz., experience models). Such experience models represent 5. CONTEXT MODELS IN DISCOURSE PROCESSING 125 the ongoing, subjective interpretation of everyday episodes in the lives of social actors. They are discussed against the background of earlier work on event interpretation and on episodic and autobiographical memory. It is stressed that both context models and experience models should not be confused with, or reduced to, the familiar situation models of current theories of text processing. Whereas the latter provide the cognitive base for the semantics of the text, the former control its pragmatic, stylistic, and other properties that vary as a function of the communicative situation. In order to focus our theoretical discussion, we use (the processing of) news discourse in the press as the specific genre for which more specific observations may be made. We use news discourse as an example because, besides everyday conversation and professional discourse, it is the kind of discourse most of us are confronted with most frequently. Also it is un- doubtedly the kind of discourse from which we learn most about the world. Moreover, it is the discourse genre I have worked on most extensively, both theoretically and empirically, so that I have some insights into its contextual constraints (van Dijk, 1988a, 1988b, 1991). It should be stressed that it cannot be the aim of this chapter to examine in de tail all the context categories or properties that have been discussed (or ignored) in the literature, nor to propose an exhaustive list or a foun- dational theory of the discursive relevance of these categories. Rather, our much more modest objective is to stress that contexts are discursively relevant for language users only through their mental modeling, and to examine how such context models influence discourse processing. Also, our contribution is theoretical. Furthermore, empirical (experimental and other) research needs to be carried out to test and elaborate the various assumptions of the theoretical framework. EVENT MODELS Earlier work on mental models in episodic memory was limited to mental representations of what events, episodes, or situations discourses are about (Garnham, 1987; Johnson-Laird, 1982; Oakhill & Garnham, 1996; van Dijk & Kintsch, 1983). Such situation models (which I now prefer to call event models to avoid confusion with the communicative situation represented by context models) account for reference, co-reference, coherence, infer- ences, and other semantic aspects of discourse processing. Event models represent the subjective interpretation of discourse, the mental starting point of production. and what people later (correctly or falsely) remember of a discourse. Through generalization and abstraction, the information represented in event models provides the basis of socially shared knowl- edge. And conversely, during understanding, these models are constructed 126 VAN DIJK from information derived from discourse and from such instantiated so- ciocultural knowledge. Much experimental work has confirmed and ex- tended the basic tenets of this mental model theory (Morrow, 1990, 1994; Morrow, Bower, & Greenspan, 1989; see also Britton & Graesser, 1996; Lorch & O'Brien, 1995; van Oostendorp & Zwaan, 1994). Instead of further detailing this theory of situation or event models, it may be pointed out that most current approaches disregard such models' embodiment of evaluative beliefs about events, that is, opinions (but see, e.g., Graesser, Singer, & Trabasso, 1994). The subjectivity of mental models is most typically represented not only by how people selectively interpret and represent events about which they communicate, but also by what opinions they have about the events. The same is true for emotion. Thus, while reading a news report about genocide in Bosnia, we combine (a) new factual beliefs about historical events with (b) applied general information about genocide, (c) opinions about (failing) international intervention, as well as (d) emotions of sym- pathy with the victims. Each of these types of information may later act as a search and retrieval cue in recall of such complex event models. For obvious contextual reasons that we spell out later, event models are typically richer in information than the discourses that express them: Most known information about an event may be uninteresting, irrelevant, inap- propriate, or already known to the recipients and should therefore remain implicit. This means that we need categories in a context model that can handle such criteria of interestingness, relevance, and mutual knowledge and that can act as the communicative interface between event models and discourse structures. That is, pragmatic context models not only moni- tor how discourses are structured to make them appropriate to the context, but also regulate the relations between semantic event models and dis- course (Robinson & Swanson, 1990). EXPERIEN CE MODELS Context models are a special case of a more general type of model, which I call experience models. Experience models ongoingly represent, and make sense of the many episodes of our everyday life. Communicative events are not only functionally embedded in such episodes, they are themselves such daily episodes. The context models language users build to understand and manage communicative situations may therefore be expected to have the overall structure of such general experience models. Before we list some properties of experience models (EMs), it should he emphasized that they should not be confused with event models con- strued for discourse processing, which may be about any event (e.g., the 5. CONTEXT MODELS IN DISCOURSE PROCESSING 127 news events we read about in the press). EMs and event models only are the same for autobiographical discourse, such as personal stories about past experiences. However, because of the primacy of personal experiences and daily routines of building EMs, the structure of event models may well be built in analogy with EMs. Also, it is plausible that event models that are similar to EMs are more accessible (Larsen & Plunkett, 1987). Where no comparable EMs are available to help understand discourse, instantia- tions of more generally socially shared knowledge, such as scripts (e.g., about wars, catastrophes, etc.) will be used to construct event models. Let us now summarize some of the properties of experience models: 1. Experience models are subjective, unique interpretations of the spe- cific episodes in which particular people participate daily. 2. EMs are stored in episodic memory (Tulving, 1983; but see McKoon, Ratcliff, & Dell, 1986). Together they define people's personal, autobio- graphical memory (Robinson & Swanson, 1990; Rubin, 1986; Thompson, Skowronski, Larsem, & Betz, 1996; Trafimow & Wyer, 1993). 3! EMs are the experiential basis of, but are distinct from, more general, context-free personal knowledge stored in episodic memory (Nelson, 1993). Such personal knowledge may be relevant for the construction of many different EMs at various moments of one's life, and may, for example, include personal scripts. Thus, "My shopping of this morning" represents an EM, whereas "My shopping" (or "My neighbor") would represent per- sonal knowledge. Thus, personal scripts are typically derived from personal routines, that is, repeated mundane EMs. 4. As long as people are awake and conscious, they continuously are engaged in the construction of EMs. However, EMs themselves are discrete, and segment the activities of everyday life in a sequence of separate, mean- ingful episodes of different levels and sizes (Newtson, 1973). This process may be compared to the meaningful segmentation and interpretation of ongoing discourse as different units at various levels. 5. EMs consist of various kinds of propositional and analogical infor- mation organized by a limited number of categories defining an efficient model schema (Barclay & Subramaniam, 1987). Typical categories are Setting (Time, Location, Circumstances), Participants in various roles, Goals, and various types of Activities, as well as their properties. 6. Although EMs are different from scripts (which represent general, socially shared knowledge and not unique personal experiences) their schematic structure may be similar to the structure of scripts for routine activities (Graesser & Nakamura, 1982; Schank & Abelson, 1977). As epi- sodic structures, therefore, EMs are closer to so-called MOPS (Schank, 1982; see also the earlier work of Schank on the representation of episodes, e.g., Schank, 1975). 128 VAN DIJK 7. Social scripts may be acquired through processes of generalization, abstraction, decontextualization, and social normalization of EMs. Con- versely, once acquired, scripts will typically be applied and instantiated in order to construct routine EMs. It is still a matter of debate when, how, and how much social knowledge (and inferences derived from it) are integrated into EMs and other episodic models (Graesser & Bower, 1990; Trafimow & Wyer, 1993). Instead of integrating applied social knowledge in the models themselves, one might assume that the models merely feature pointers to such general knowledge. I assume, however, that EMs feature specific, situation-bound, that is, adapted instantiations of social knowl- edge—those and only those that are relevant for the current interpretation of ongoing episodes. 8. EMs are ongoingly construed in an effective strategic way, for example, online, tentatively, and hence possibly erroneously, using different infor- mation of various levels at the same time (van Dijk & Kintsch, 1983). 9. EM construction strategically uses and combines the following kinds of information: (a) interpreted sense data, (b) personal knowledge and scripts, including Self, (c) old EMs (personal memories of previous epi- sodes), and (d) instantiated and adapted social knowledge and attitudes. 10. Segmentation of EM sequences is based on changes in the information stored under one of the category nodes (e.g., a change of place, time period, participants, or (overall) activity type). 11. EMs are typically segmented, understood, and recalled at (or aboye) some prototypical middle-range level, such as "my taking a shower" or "my having breakfast" rather than "my opening the door" or "my starting the car." Lower-level and basic actions are only attended to, separately stored, remembered, and talked about later in situations of trouble or when they otherwise become interesting or salient. 12. Model schema categories, together with higher-level macrorepresen- tations of activities, may also be used in the overall organization of EMs in episodic memory (e.g., "My time as a student," "My vacation in Spain," or "My life with Claudia," etc.; Anderson & Conway, 1993; Seifert, Abelson, McKoon, & Ratcliff, 1986). That is, EMs may be further organized at various levels into compound, complex, and higher-level EMs (Hanson & Hirst, 1989; Neisser, 1986). 13. Self is a central category in EMs. However, the unique hic et nunc nature of EMs requires that the actually constructed Self in an EM is also a unique construct. That is, it is a specific instantiation of a more general, abstract, and more permanent Self represented in episodic memory (Bar- clay & Subramaniam, 1987; Kihlstróm, 1993; Markus, 1977; Srull & Wyer, 1993). Again, this distinction shows the difference between EMs and more permanent episodic knowledge. The Self category organizes many of the 5. CONTEXT MODELS IN DISCOURSE PROCESSING 129 other categories of the EM schema, such as relations between participants, perspective, and so on. 14. As is the case for all models, EMs feature opinions and emotions, especially because of the personal relevance or involvement of the Self in these experiences (Neisser & Fivush, 1994). In the same way that scripts may be derived from abstracted, generalized, and socially normalized EMs, socially shared altitudes may be derived from EMs that feature personal opinions. And vice versa, attitudes may be instantiated in the construction of opinions in EMs (e.g., "My opinion now about this abortion" from "My group's opinion about abortion"). Obviously, as is the case for all personal instantiations of socially shared cognitions, EMs will always be unique and adapted to the current circumstances. Hence the individual variation of EMs. 15. EMs not only define the details of our personal past and present, they also represent overall designs of future actions, such as plan, tasks, and goals, which may also be used to retrieve EMs (Anderson & Conway, 1993; Wyer & Bodenhausen, 1985). 16. As forms of concurrent thought, EMs may represent unfinished business that gives rise to our everyday, involuntary ruminations (worries, regrets, anxiety, anticipation, etc.), especially when their goals have not yet been realized (Martin & Tesser, 1996; Singer, 1993; Wyer, 1996). 17. When being recalled, EMs may become the typical stuff of everyday storytelling, especially if they are relatively exceptional or otherwise inter- esting for recipiente. However, stories are not only shaped by EMs but also by relevant context models that define the specific communicative goals and circumstances of storytelling. That is, for contextual reasons, storytellers may transform their EMs in many ways (Brewer, 1982; Bruner, 1987, 1994; Edwards & Middleton, 1986; Kerby, 1991; Labov & Waletzky, 1967; Lieblich & Josselson, 1994; Loftus, 1979; Means & Loftus, 1991; Nelson, 1989; Po- lanyi, 1985). 18. Changes of episodes in EMs may be represented in discourse by a change of underlying semantic episodes, each governed by a topic or macroproposition. Such a change, for example, of participants, setting, overall action, or perspective, is typically expressed by beginning a new paragraph in written texts (van Dijk, 1982). These summarizing features of experience models each need to be developed in detailed theories. However, the idea of experience models is persuasive and nicely occupies a theoretical niche left open between such earlier notions as situation model, script, autobiographical memory, personal knowledge, Self, and the like. Indeed, experience models explain some of the relationships between these notions, while providing the basis for a more explicit theory of episodic and personal memory. 130 VAN DIJK CONTEXT MODELS As suggested, context models are special kinds of experience models. They represent communicative episodes in which we participate, often as part of other everyday episodes (conversation at breakfast, meeting at work, etc.). Because, among other elements of the situation, context models represent ongoing action, they are of course dynamic: They will be con- tinuously updated during the processing of text or talk. Different participants in a communicative event each have their own, personal context model, defining their personal interpretation of the cur- rent situation. However, discursive interaction and communication is pos- sible only when such models are at least partly shared, synchronized, or negotiated. Indeed, participants may jointly produce and ongoingly update each other's models. Speakers may have partial models of the context models of recipients and vice versa, especially about the knowledge they share. Such mutual beliefs about each other's models are• theoretically infinite, but in practice are constrained by contextual relevance (for details about mutual knowledge in language users, see Clark, 1996). Context models have the same overall schematic structure as other expe- rience models, but with specific categories tuned to communicative events. So far, these categories have only partly been made explicit in discourse analysis (e.g., age, gender, ethnicity, class, roles, power, goals, or beliefs of participants, as well as setting characteristics, such as time, location, and circumstances). To distinguish explicitly between contexts and the full complexity of social situations (Argyle et al., 1981; Furnham & Argyle, 1981), we define contexts as the structure of all properties of the social situation that are systematically relevant for the production, comprehension, or functions of discourse and its structures. Relevance may be both personal and social in this case, and is defined by the current context model (Sperber & Wilson, 1986). That is, it is not objective age, ethnicity, sex, or similar social features that constitute the context, but their socially based and mentally represented constructions as they are made or taken to be relevant by social members in interaction. This does not mean that anything goes. Despite personal and contextual vari- ation, the relevance criterion is socially based while grounded in social rules and strategies. Precisely in order to distinguish between the theoretically infinite complexity of the social situation and the context constructed out of this situation, language users have learned to focus on those properties of the social situation that aresystematically relevant for discourse in a given culture. For instance, they know that speakers may vary formal discourse properties (such as pronouns) as a function of their (represented) age or gender rather than as a function of hair color or height. Moreover, efficiency 5. CONTEXT MODELS IN DISCOURSE PROCESSING 131 and strategic processing demands also require that the number of systemati- cally relevant situation properties be relatively small. The Structure of Context Models Against the background of earlier work on context in discourse studies and psychology (see references given previously), I provisionally assume that context models are organized at least by the following schematic categories (definition and illustration of these categories are given later for news processing): I. DOMAIN II. SITUATION A. Setting A.1. Time A.2. Location A.2.1. Props A.3. Circumstances B. Events B.1. Participants B.1.1. Roles B.1.1.1. Social roles B.1.1.2. Interactional roles B.1.1.3. Communicative roles B.2.Action/Discourse B.2.1. Action types, Genres B.3. Cognition B.3.1. Aims, goals, or purposes B.3.2. Knowledge B.3.3. Opinions B.3.4. Emotions The point of this model schema is to organize and reduce the complexity of the social situation in such a way that language users have an efficient device to contextualize discourse production and comprehension. As sug- gested before, the criterion of inclusion of each category is defined in terms of systematic relevance for a given language or culture: Properties of discourse have to be able to vary according to the information stored under each category of the schema. The schema should be read as follows: A social situation is part of a social domain (such as politics, education, or law) and consists of a number of events in a spatiotemporal setting. These events themselves consist of participante with different roles and with different mental properties (e.g., 132 VAN DIJK goals and knowledge) engaging in various kinds of actions, of which the verbal action is the crucial one for the definition of a social situation as a context (for details, see following sections). The seeming simplicity of this schema might hide the fact that each category may itself cover fairly complex representations. Because context models are a special type of experience models, they might for instance feature possibly complex person models of participants. Such participant models might be constructed from the extensive general (lay) knowledge people have about themselves and other persons (Markus, 1977). However, it is here assumed that for the construction of efficient context models for language use, it is sufficient that the participants know the relevant (a) identities (roles), (b) ongoing actions, (c) current beliefs of themselves and other participants in the situation, and (d) various properties of the setting. Note that given the richness of the social situation in which people discursively interact, many other categories may be proposed for inclusion in the schema. For instance, participants may be aware of, and orient to, one or more objects in the situation, and such (possibly joint) awareness may be signaled by deictic expressions (Clark, 1996). However, for several theoretical reasons, we prefer to represent the world or situation talked about separately, as in the event model discussed previously. One of these reasons is that we do not want to make a fundamental theoretical distinc- tion between the representation of the referents (objects, people, etc.) that are part of the communicative situation itself and those that are not. However, this example does suggest that event models and context models may overlap. This is obviously necessary in order to account for all other expressions that refer to elements of the context. Context Modeis in Text Processing One of the first interesting implications of the context model schema just presented is that the mental representation of the ongoing discourse itself should be part of the context model. This is true because context models were assumed to represent ongoing action, and discourse is merely one specific type of such action and an inherent part of the whole communi- cative event and situation. The traditional distinction between text and context is therefore only an analytical one, based on a notion of a (com- pleted) discourse being abstracted from its context. Thus, especially in spoken discourse, the (representation of the) previous part of a dialogue becomes automatically part of the context model that influences what is currently being said and done. In other words, we here encounter a first link between the notion of Text Representation in earlier theories of text processing, and the notion of context model proposed here: Both are