General Constraints on Process Models of Language Comprehension

General Constraints on Process Models of Language Comprehension

LANGUAGE AND COMPREHENSION J.-F. Le Ny. W. Kintsch (editors) © North-Holland Publishing Company. 1982 179 GENERAL CONSTRAINTS ON PROCESS MOOELS OF ...

1MB Sizes 0 Downloads 3 Views

LANGUAGE AND COMPREHENSION J.-F. Le Ny. W. Kintsch (editors)

© North-Holland Publishing Company. 1982

179

GENERAL CONSTRAINTS ON PROCESS MOOELS OF LANGUAGE COMPREHENSION Robert de Beaugrande English Oepartment Uniyersity of Florida Gainesville, Florida U.S.A. It is suggested that psychological research on language has adopted peculiar priorities by extrapolating from structural linguistics, rather than specifying general theories of cognition and communication to fit the issues of language processing. A framework is presented that situates text processing in such general human constraints as resources, memory, and linear capacities. A set of questions for future research is drawn from this foundation. 1. THE SENTENCE COMPREHENSION PARADIGM 1.1 To date, the most extensive and enduring achievement of linguistics has been the development of methodologies for ~tnuctuna{ analy~~: the classification of items or of relations between items into a sorting scheme often called a taxonomy. The means for doing this task varied considerably from one linguistic approach to another; but the task itself has remained the central concern 'throughout. It was assumed that the classification procedures were objective and formal, so that all linguists would necessarily arrive at the same results for a given set of samples (e.g. Bloomfield, 1933; Harris, 1951; Fries 1952; Chomsky, 1957) -- at least within a particular approach. 1.2 For many years, the major concern of "psycholinguistics" was defined as: to test the psychological reality of current linguistic theories or models. This outlook encouraged researchers to construe structural analysis (as performed by professional linguists) as the model of language comprehension (as performed by human language users). This tactic was adopted variously for immediate constituent analysis (e.g. Johnson, 1965) and transformational grammar (e.g. Mehler, 1963), while some researchers (e.g. Clark and Clark, 1977) experimented with eclectic combinations of many kinds of approaches. The usual assumption was that people comprehend language by sorting words and phrases according to much the same criteria as those used by linguistics itself. 1.3 Given the priorities of lingui.stics and psycholinguistics, this assumption was only to be expected. But I shall argue that the assumption is at best unproven, and at worst, incoherent. In the first place, the structural analysis performed by linguistic scholars is a specialized activity based on training that normal language users do not receive. In the second place, the taxonomies developed for language, as Sapir (1921) already pointed out, will never be totally complete or formalized, even for the specialist. How, for example, is the normal language user to decide if 'bright' is an adjective or an adverb in (1) The moon shone bright. when even linguists do not all agree? Research shows that people are not uniformly able to decide what does or does not constitute a permissable sentence of their language (Greenbaum [ed.], 1977); yet in all linguistic approaches so far, the sentence is the basic category from which all other analysis proceeds. 1.4 Consequently, structural analysis of linguists inspires a highly implausible picture of language comprehension. Quantitatively, it suggests an unduly lengthy treatment of surface structure, whereas normal language users may well be doing only a partial, approximative analysis (cf. Burton, 1976; Schank, Liebowitz, and

180

R. de Beaugrande

Birnbaum, 1978; Masson, 1979). q~vety, it incorporates distinctions and formalities normal lan9uage users could scarcely meet. People obviously do not delay all comprehension processes until they have been given a complete sentence; they start as soon as they have a few words (Schank et al., 1975). Moreover, the boundaries of sentences are often indistinct in speech, as in this recorded sample (a "I" indicates a pause): (2) this time everyone was really paranoid because everyone was pretty wasted I and I luckily I we all made it through without getting checked I got on the bus and started partying and just kept mixing drinks and drinking and drinking I and some people were taking drugs and things I um Even if the listeners waited for a sentence boundary, they often would be unable to identify it. There is a substantial literature (reviewed in Beaugrande, 1982a) showing that pauses do not necessarily indicate the boundaries of constituents or sentences. Yet linguistic analysis, whether based on immediate constituents or on with the complete sentence. Though transformational structures, always beg~n6 transformational grammar at least has detailed rules, it still analyze~ a sentence only by running through all the ways it could gene4ate (structurally describe) the sentence, a method that is enormously explosive and inefficient (Woods, 1970). the structural 1.5 Further qualitative problems arise in regard to ~nteg~ng analysis into the total activity of language comprehension. Most approaches foresee a relatively independent syntactic stage that is internally well-ordered, but externally poorly integrated with meaning and purpose in real communicative situations. The general trend had been to try forcing semantics and pragmatics into the same categories of analysis as those for syntax. But the limited success of those trends suggests that the categories of different levels are probably not of the same character. For example, there is no good evidence that a proposition or a speech act must fill the boundaries of a sentence. 1.6 In recent years, psycholinguistics has moved away from a literal transposition of linguistics onto comprehension. But the sentence continues to be viewed as the basic unit. The following quotations are symptomatic: The problem of when and how a sentence is understood is, in my view, the central problem of experimental psycholinguistics. (Gough, 1971: 109) The fundamental problem in psycholinguistics is simple to formulate: What happens when we understand sentences? (Johnson-Laird, 1974: 143) How do we understand the relevant meaning of sentences used in ordinary contexts? This is a central problem in psychology and a primary preoccupation of the psycholinguist. (Tanenhaus, Carroll, and Bever, 1976: 244) The agreement here is all the more striking because these researchers represent totally different approaches to language research. It is also rather ominous, because normal language users do not have a well-defined notion of the sentence or ~ not of its constituent word classes (1.8f). What ~ ~n 6act comp~ehended ~entenc~, but conceptual content; at most, the sentence is one surface unit among many that can be used as an aid for that task (O'Connell, 1977). But to presuppose the centrality of the sentence in everyday comprehension is to construe a special case as the general one. 1.7 In a neutral perspective, the sentence might be viewed as follows. It is a 60~at which is imposed on ~~etch~ 06 ~en ~COun4e to suggest several sorts of processing instructions. First, the inner organization of ~he sentence signals the conceptual ~eiatedn~~ of the content underlying its bouhded constituents (cf. Miller and Kintsch, 1980). In one common sentence type, the subject is mapped onto an agent, and the verb onto an action; it is then expected that all parts of the subject are conceptually related to the agent, and all parts of the predicate to the action. But this kind of expectation is only a preferred hypothesis that will be modified in many contexts. 1.8 Second, the inner organization of the sentence signals the pe4ceptual

General Constraints on Process Models

181

ehanact~ 06 a ~ee~e, object, event, etc. being mentioned or described (Osgood and Bock, 1977). Charles E. Osgood and his associates have consistently argued the order in which things are perceived influences the order in which words are arranged in syntactic formats. One reason that the agent is a common candidate for the subject slot in English sentences is that agents are perceived readily as the "figure" before the "ground," i.e. as the most active element in an environment; thus the subject slot, being toward the start of English sentences, is a "natural" position. 1.9 Third, the inner organization of the sentence signals the ~n60kmat{onal ~~ehy of the underlying content. For the famous examples: (3) I smeared paint on the wall. (4) I smeared the wall with paint. Fillmore (1977) showed that the item placed in the direct object slot is taken as more important to the event; only in (4) do people assume that the entire wall was covered. For the sequence: (5) Witness A: Before the accused emptied the safe, he shot the watchman. Witness B: That's not true! Posner (1980) points out that the denial of Witness B only attacks the content underlying the main clause (the "shooting"), while that underlying the subordinate clause ("emptying the safe") remains uncontested, because it is assumed to be outside the main focus. at which ehu~~ 1.10 Fourth, the boundaries of sentences help to control the ~ate of discourse content get constructed, and also the ~~ze of those chunks. The end of a sentence is a usual, though not obligatory point to consolidate building blocks. Of course, conceptual relatedness extends beyond sentence boundaries, if the discourse is to be coherent. But normally, a reader can afford to pause briefly at the end of the sentence in order to integrate the current chunk with the content that has accrued so far. This procedure is sometimes described in terms of "clearing a buffer" that has been filled by the current sentence (cf. Miller and Kintsch, 1980). 1.11 These four functions of the sentence in discourse seem reasonably secure. Others might be added, for instance, that sentence formats respond to degree of "ego involvement" (Ertel, 1977). But the point here is that in each function, the sentence is never the only entity capable of being so used. Thus, it hardly seems justified to uphold the sentence (or a taxonomy or sentence structures) as a theoretical framework for comprehension research. For example, though numerous experiments have been designed to contrast active sentence against passives (cf. surveysin Osgood and Bock, 1977: 99-102; Levelt, 1978: 24-25), the human decision to use an active or a passives depend on non-syntactic factors such as salience, imagery, vividness, activeness, etc. in the conceptual content. A purely syntactic definition of "passive sentence" thus describes no empirically unified phenomenon. 1.12 The strategy should rather be to group experiments according to p~oe~~~g typ~. The sentence would then be distributed over its various functions, such as the four enumerated in 1.7-10. In section 2, I propose a framework for one such approach, grouping together samples from the experimental literature as a way of illustration. This framework is'ntmeant as a testable "theory," but as an outline model to coordinate issues and define their mutual relevance.

2. LANGUAGE WITHIN COGNITION AND ACTION 2.1 Abandoning the traditional contention of linguists that language is an isolated faculty, let us define language processes as ~pe~zat{o~ of more general process types. Syntax would then be a special case of ~ean ~~e~genee (cf. Beaugrande, 1981b, 1981c, 1982a); semantics a special case of the aeq~~on and ~zat{on 06 knowledge (Hormann, 1976; Kintsch, 1977; Beau9rande, 1980a);

182

R. de Beaugrande

and pra~matics a special case of the eOn6tnuetion and ~pteme~on 06 ptan6 and goaRA (Schank and Abelson, 1977; Cohen, 1978; McCalla, 1978; Allen, 1979; Beaugrande, 1979, 1980b). We can easily see that these three domains are themselves highly interdependent. For example, linearity applies to sequences of events in stored knowledge, and to the arrangement of steps in a plan. Thus, domains like syntax, semantics, and pragmatics are themselves distinctive only as f>pecUalization6 06 otheJ1l'.l.U,e anatogoUf> pfweeMing eapacA;Uef>. In th i s sense, we could envision a particular process being assigned to a f>peei~t, e.g., a "semantic specialist" that keeps track of referents in the world of a discourse (cf. Winograd, 1972). The processor would be an agglomerate of ~outinef> ptUf> f>pe~t eommu~ef>: these two components interact for the performance of a given task. 2.2 It follows that theories of language comprehension should be stated within those constraints which are known to apply to human processes in general. A theory that deliberately excludes limitations upon resources and memory (e.g. Chomsky, 1965) merits no further consideration, because it is humanly unreasonable. Of course, general processing constraints will not uniquely determine one theory as "correct" above all others they will only narrow down the class of theories that are worth pursuing, and will indicate how language is situated within larger human contexts. text is by necessity tin~y o~~ed, either as a 2.3 To begin with, the f>~6aee int~genee would stream of sound or as a succession of printed images. LLn~ be the capacity to process language, or any other task depending upon a linear mode. At least seven principles should apply to any linear activity (cf. Beaugrande, 1981b, 1982a). The eo~e-and-adjunet principle distinguishes between central versus peripheral entities; the central ones are operationally defined as the control centers in ongoing processing. The pauf>e principle allows for suspending the linear sequence in accord with limitations or demarcations of processing. The correlation of the current items in the sequence with previous or subsequent ones is managed by the principles of took-baek and took-ahead, respectively. The heavinef>f> principle concerns the gradations of importance, involvement, informativity, focus, emphasis, and salience -- all of which aspects seem to create a "heavier load" on processing as they rise. The fuambiguation principle is responsible for resolving any cases where two or more patterns, formats, interpretations, etc., compete. The ~ting principle applies whenever comparable items or patterns are enumerated in a series. 2.4 These principles would apply to any tevet or f>eate of processing, or to any combination of these. For example, the core-and-adjunct principle would apply to syntactic distinctions like content words vs. function words, heads vs. modifiers, or main clauses vs. dependent clauses; to semantic distinctions like primary concepts (event, action, object, situation) vs. secondary concepts (location, time, cause, quantity, etc.) (cf. Beaugrande, 1980a: 79-84); to pragmatic distinctions like main goal vs. sub-goal; and so forth. The pause principle would respond to various pressures for interrupting or at least retarding the processing of a sequence: breathing, marking boundaries between surface units, consolidating retrospective operations, planning predicted operations, and so on (cf. Goldman-Eisler, 1968; Rosenberg, 1977; Beaugrande, 1982a: section IV.2). 2.5 The functions of the sentence described in 1.7-10 fall under these principles as well. The use of sentence constituents as indicators of conceptual relatedness (1.7) is chiefly an application of the principles of core-and-adjunct, look-back, and look-ahead. The relationship between the sentence formats and perception illustrates mainly the heaviness and disambiguation principles. The correspondence between sentence formats and informativity falls mostly under the heaviness and core-and-adjunct principles. The correlation between sentences or sentence constituents and the rate or size of processing chunks falls largely under the core-and-adjunct, pause, and heaviness principles. Of course, the dominance of any one principle over others will be sensitive to contexts, and will reflect other processing needs besides building sentences.

General Constraints on Process Models

183

2.6 Linear processing is further defined by the nature of cognitive resources like attention and memory. Language is sufficiently complex that a processor can't hope hope to attend to all levels or features at once. Older models assumed that prostages, like a relay from one black box to the next cessing was done in ~equential (e.g., Fromkin, 1971; Gibson, 1971). The current trend is to favor p~altel stages that run in concurrent or overlapping ways, consulting each other freely as needed (e.g. Marslen-Wilson, 1975; Rumelhart, 1977; Woods, 1978), e.g. using semantics However, the notion of "specialist to eliminate syntactic ambiguities. communities" (2.1) allows for flexible scheduling to meet varying contingencies, and thereby suspends the opposition between serial vs. parallel models. 2.7 Resource limitations suggest that parallel processing involves a shifting of dom~nce from level to level. Dominance would entail an assignment of resources above a thn~ho£d that is set according to the overall resources available for the occasion. If processing must be done at great speed, these thresholds would be lowered all throughout the system's operations, producing what Norman and Bobrow (1975) call a "graceful degradation" of performance (cf. 2.15). The same would apply to the thresholds of ~~~on (the criteria where a process is selected and run) and of tenmination (the criteria where results are considered satisfactory and processing stops). The lower the available resources, the more inclined the processor would be to accept fuzzy, provisional criteria for these thresholds. Even so, processing does reasonably well, as shown for instance in studies of skim-reading (Masson, 1979). 2.8 One important factor in ongoing processing is of course memony. The principles of look-back and look-ahead require both a netno~pect~ve and a pne~~ve representation of the text. The scheme I have worked out (Beaugrande, 1981c, 1982a) is given in Figure 1. The various levels of processing are shown with the "shallower 1eve1s on the top and the "deeper" ones on the bottom. The levels are factored according to the materials they process, either from the text or the processor's memory: sounds/letters, words, syntactic phrasing, concepts/relations, ideas, and goals. The text is factored (left to right) into: retrospective representation of prior text, perception of current text, and predictive representation of subsequent text. The direct contract with the surface text, according to this scheme, is only a very brief interval of perception; everything else depends upon mental representations that differ from the surface text (though including some surface traces) in being a complex, multi-dimensional array, not a simple word-sequence (cf. 2.12). 2.9 The exact time point of perception is indicated by the zero on the time can be distinguished along this continuum by continuum. The various memony ~tan~ reference to their ~pan (time of perseverence): ~hont-tenm ~e~ony ~tonage (one memony (some twenty seconds), and long-tenm memany or two seconds), ~hont-tenm (unlimited) (cf. Atkinson and Shiffrin, 1968; Keele, 1973; Loftus and Loftus, 1976; Kintsch, 1977; Posner, 1978; Shall ice, 1979). In early work, it was implied that each store was exclusively devoted to one sort of materials, e.g. short-term memory to syntax and long-term memory to conceptual meaning, because these materials decay at different rates (cf. Sachs, 1967; Begg, 1971; Garrod and Trabasso, 1973; Anderson, 1974). However, in the approach I am outlining, these Short-term sensory tendencies are viewed as a further instance of ~pec~z~on. storage is specialized for sounds and letters; short-term memory for syntactic phrasings and local chunks of concepts and relations; and long-term memory for ideas and goals. However, sensory perception can have at least some access to deeper levels such as conceptual meaning (Raser, 1972; Shulman, 1972). And long-term memory can retain some traces of the surface, e.g. the recollection of whether an item was presented visually or acoustically (Bray and Batchelder, 1972; Hintzman, Block, and Inskeep, 1972). Intriguingly, the latter two papers indicate that the trace of visual vs. acoustic mode has no profound effect on long-term retention of content. 2.10 Finally, there is a store of working memory defined not according to its range of decay, but according to its capacity of about seven chunks (cf. Miller,

seconds I

I

I

I

I

I 20

I

I II

··.12

I

I

I

101 STSS

.~

I

. .

I

I

STM

I

I

I

I

I

I I I

......................""11:::·'1

I I I

.................. ................ . . . . .. .. .. .. .. .. .. .. .. .. .. .. . . . .

• • • • I • • • • • •• · · · . · . · · · . . : . . · ·

21 ••••••.•••••. 20

II

. .

.. ................ .. .. .. .. .. . .. • • •WORDS

• • • • • . .. :-'. -: -: -: ..: :-:- ·SyNTI~X

.... + + + . · .

1'-...

::: : : : : :: ::: :: : ~:.

look-ahead ...

Figure 1. lTM: long-term memory; STM: short-term memory; STSS: short-term sensory storage; 11111: working memory

STM

I

:-:-:'~".=1

............... ...............

PERCEPTION OF CURRENT TEXT

'"

now

••• + + + + + b J M +

look-back

II I

I::"..·.· · · . . . ·.·

SYNTAX· .' ••••• ' • ' ••••••••• ...................... .. .. .. .. . .. .. .. . .. .. ..

WORDS •

+

~ ~

'§l

i

~

;:0

~

General Constraints on Process Models

185

1956; Kintsch and Polson, 1979). Often, working memory has been simply equated with short term memory, perhaps due to the tendency to think of memory stores as distinct "locations" where materials were "sent." This static, spatial viewpoint now seems inappropriate (Kintsch, 1977). If the stores are instead differential modes of activation and specialization, then working memory can take its materials from any of the three stores defined by their spans. 2.11 The specializations of the memory stores are suggested by dot shadings in Figure 1. The different spans of the stores are indicated by squeezing the graph toward its middle. So far, there is little experimental support for the right-hand side of predictive memory. I am provisionally assuming that predictive memory should be more or less symmetrical to retrospective memory, though probably shorter in spans. Intuitively, this assumption seems reasonable. A language processor can look ahead to main ideas or goals much further than to detailed concepts or syntactic phrasings; anticipation of exact words or letters is possible only over very slight distances. 2.12 The nature of the mental representation that serves in place of the text in retrospective and predictive operations is as yet hardly explored. It should have a format suitable for the tasks performed upon it. It could scarcely be a linear string of words, simply because that format is not efficient: it would unrealistically demand traces of every word (out of the question in view of the experimental literature); it would give undue prominence to the surface and thereby neglect the underlying meaning and purpose of discourse; it would mediate against those processes that require a hierarchical arrangement to organize text content on a large scale (cf. Meyer, 1975; Mandler and Johnson, 1977; Thorndyke, 1977; Kintsch and van Dijk, 197B; Kintsch and Vipond, 1979). On the other hand, the representation must be highly compatible with the surface text, because it evidently guides the impressive ability of understanders to anticipate future events. For example, eye movement studies reveal high skills in fixating those words that will contribute best to the gist, especially during skim-reading (Just, 19B1). It would seem reasonable to envision the mental representation as a multi-directional network constructed and annotated to serve all the levels of processing, including words and phases (Beaugrande, 19B1c). This format would be economical in enabling operations of same or similar patterning to be consolidated in operation (Woods, 197B). One illustration could be a general look-back that searches for materials required by the various levels (e.g. a co-referent for a pronoun, a subject for a predicate, a concept for a relation, etc.) at the same time, rather like a bus whose passengers are watching for their personal destinations (compare the search procedures in Fahlman, 1979). 2.13 The factors outlined so far -- linearity, processing level, memory store specialization, distribution of resources -- cannot stipulate exactly what must happen at a given instant of processing. The interaction between ~outin~ and ~pe~~ (2.1,8) allows flexible, adaptive processing for specific contexts, occasions and human beings. For example, the amount of processing expended on an unfamiliar word is lower if the word appears in a familiar story (Wittrock, Marks, and Doctorow, 1975). Therefore, there cannot be one uniform procedure that treats all words in all settings. 2.14 Figure 2 indicates how resource loads would react to demands of context. We assume a complex, interactive package of processing routines moving along through the task (left-to-right). This package is accompanied by (or has access to) a community of processing specialists, many of which need not be concerned with each other. The package traverses a constellation of parameters that affect what the specialists will do to influence, complement, or rearrange the routines. These parameters include: (a) moto~ ~o~ol (e.g. the formation of sensory traces of visual patterns); (b) 6eedba~R ~eq~eme~ (the processor's need to monitor what it's doing); (c) attention (resource expenditure on one task that interferes with others); (d) no~e (competing, extraneous events); (e) th~~hol~ 06 ~~~on and t~~n~on (cf. 2.7); (f) p~eptua1 ~me~~y (extent to which a language

186

R. de Beaugrande

lower

+

+

lesser

less

more

more

~

PERCEPTUAL IMMEDIACY

higher

less

greater

greater

FAMILIARITY

lesser

lower

INFORMATIONALITY

worse

GOODNESS OF FIT

stronger

MOTIVATION

more

TIME ALLOTMENT

+-

~

~

COMPLEXITY

+

+-

more

lesser

+

+

NOISE

~

greater

THRESHOLDS OF INITIATION AND TERMINATION

lower

+

ATTENTION

less

+

+

higher

FEEDBACK REQUIREMENTS

+

+

MOTOR CONTROL

~

~

~

~

higher

~

better

~

weaker~

less

~

Figure 2. Processing variables in task operations

General Constraints on Process Models

187

correlates with input from sensory modalities); (g) Qomptexity (configuration part-whole relationships); (h) 6am~y (degree to which tasks have been encountered and performed by the same processor before); (i) ~n60~ationatity (proportion of unpredictable to predictable); (j) goodne¢~ 06 6~ (extent to which entities should match a feature configuration in order to be accepted as meeting current specifications); (k) mo~vation (why the text is being processed, e.g. to learn the content, to receive a command, to advance a goal, etc.); and (1) time ateotment (how long the processor has to get through the task). 2.15 These twelve parameters can easily be related to each other, but there will be cases where they are also distinct. The Figure suggests that whenever one of these parameters rises, these should be some sort of compensation along one or more other parameters. Ideally, a strain from one parameter would be strategically distributed throughout the system, in the manner of cybernetic regulation. There would then be no important degradations concentrated in one area. However, the wide differences in the quality of the results in both production and comprehension of texts indicates that this ideal is often not upheld. For example, if noise obtrudes too much one may not even get the gist of a discourse, but only a few incoherent fragments. Educational training might be developed to enable more balanced redistribution of processing load (Beaugrande, 1982a). 2.16 These parameters might have some comparatively neutral degree that creates no strain within the system. However, anything but the most practised, conventionalized discourses probably call for some adaptation of loads. Yet routine discourse is likely to lower interest and thereby decrease the overall allotment of resources throughout the system, so that task performance is degraded. Very little content gets remembered from a boring lecture; workmanship declines on tedious motor activities; and so forth. Thus, discourse processing is driven by the dynamic tension to reduce load and to increase it again, to overcome challenges and yet to seek them out. 2.17 Whatever tasks are imposed in language experiments may load these parameters in diverse ways. Usually, researchers prefer to assume that any factor besides the experimental variab1e(s) is frozen at its most neutral degree. For example, most "sentence comprehension" studies don'tdea1 with the motivations people normally have when comprehending discourse, nor with the many criteria set outside the experimental setting (familiarity, goodness of fit, etc.). Hence, we can't tell if laboratory tasks are ~~pt~ than natural ones (because factors are excluded) or mo~e comptex (because people don't have skilled routines to apply). The thorough processing of syntax that many experiments indicate may merely result from re-channe1ing resources that are normally devoted to the meaning and purpose of discourse. In natural settings the use of syntax could be quite fuzzy and non-essential. Clearly, broader experiments are needed to manipulate resource allotments and compare the relative contributions of the several levels of processing (ct 3.5.14). 3. QUESTIONS, QUESTIONS 3.1 Theories and models of language processing have had a long and diffuse history, ranging from "associative chaining" (e.g. Washburn, 1916) to "behavioral conditioning" (e.g. Skinner, 1957) to "grammatical transformations" (e.g. Mehler, 1963). At present, new directions are emerging, but there is little agreement about what research should be done for what motives. At most, there is a consensus that the analysis of strings of words, the favorite approach of past decades is no longer very rewarding. had surfaced as a framework for organizing memory 3.2 The notion of ~Qhema (Bartlett, 1932), word order (Lashley, 1951), and actions (e.g. Oldfield, 1954; Evans, 1967). It remai ned on the periphery of experimental psychology, however, because it complicates the design and interpretation of experiments. Recent research has returned to the schema with new emphasis, especially as a memory pattern (Kintsch, 1977; Rume1hart, 1980). Language processing is seen as the

188

R. de Beaugrande

application of interactive schemas on the various levels (e.g. Rumelhart and Ortony, 1977; Adams and Collins, 1979). 3.3 As yet, the notion of schema is still too vague (Thorndyke and Yekovich, 1980) to be much more than a £abel for organized patterns of action or knowledge, rather than an exp£anation. We need to know more about how schemas are formed; how they evolve during use; how strictly they control processing; what conditions trigger their selection; and how they can interact or compete with each other. Their ~tnuct~e is less crucial than their a~ve ~e. 3.4 The issues raised in section 2 constitute a challenge to schema theories. No doubt schemas are vital for linear activities, as Lashley (1951) proposed. But they must also serve in a multitude of operations such as search, pattern-matching, scheduling, classifying, and so forth. If the distinction between routines and specialists is valid (2.1), then schemas may be created on the spot, not just summoned from storage as fixed frames (cf. Kintsch, 1980). Also, a schema might itself be quite flexible about the materials that would trigger it and fill its variables (cf. Rumelhart, 1980). All these possibilities need exploration if we are to overcome the limitations imposed by a literal adherence to structural analysis. 3.5 I would accordingly like to conclude not with a set of data, but with a set of questions. Until there are at least some provisional answers, it will be hard to know how our data should be interpreted in the larger picture of language within cognition and action. 3.5.1 How and when ~e actual op~o~ ~eheduled? Given the large numbers of mathematically possible combinations of steps, how and when does a processor decide which ones to use in real ljfe? Can provisional selections be revised at the last minute? Is there a distinct group of "scheduling specialists" called in when routines fail to apply? Or does scheduling result as a consequence of the specialists that watch over resource allotment among the concurrently operative processing levels? 3.5.2 How ~e tim~ed ~~o~e~ aetualty allotted? Does a processor have routines for allotting resources? Can they be overridden by context, and when? How are decisions about text, type, audience, purpose, etc., translated into resource all otments ? 3.5.3 How ~ ovenload ~eated, notieed, o~ ~edueed? If there is a rise along several parameters of the kind indicated in Figure 2 (complexity, informativity, etc.), overload would be expected. Is the processor able to anticipate such cases and react before degradation becomes obvious? What indicators could be consulted, and what counter-measures taken? Options might include: slowing processing down; increasing redundancy; rescheduling simultaneous operations into successive ones; and so forth (Beaugrande, 1982a, 1982b). 3. 5. 4 To what extent ean p~el p~o e~~ing c.ombine the tas k~ 06 di6 6~ent !eve£4? It would be economical if redundant or similarly patterned operations could be done in a single run-through (2.12), but such may not be the case. Perhaps this economy would overtax the specialists needed to enact it. Possibly, too, the system is too diffuse for optimal co-ordination among all its subsystems. 3.5.5 To what extent ~ each act 06 c.ompltehe~ion unique, and to what extent ~ it c.omp~b!e to oth~? Uniqueness of processing events is a troubling prospect when researchers set out in search of general findings. However, the essential commonalities among classes of processing events may be on a much higher plane of operati ona 1 power than that of "noun phrases," "re 1ati ve c1auses ," and the other popular entities of early psycholinguistics. High-powered processes could easily deal with a unique event as a complex intersection of the subprocesses it controls within a stipulated context. weak .6pOU and bottieneeu in text pltoc.~.6iYl.fl? If increased 3.5.6 Me th~e

General Constraints on Process Models

189

limitations or overloading should arise, are there typical points in the overall system where degradation should be expected? Can increased resources or conscious training offset such cases? tightened o~ ~etaxed? 3.5.7 How ane ~~ If so many processes depend upon defining the conditions for an action (e.g., those for initiation, termination, or goodness of fit), how can such definitions be adjusted? Can search procedures be instructed to take less thorough samples, or to be satisfied with approximate candidates? Would such instructions be specific, or generally distributed throughout the whole system? 3.5.8 How ane autom~c p~oc~~~ co~a£idated? Automatic sub-routine packages would be an efficient way to operate. They might intersect one another and thereby create large complexes of activity without controls from the central feedback mechanism. These packages might make no demands on attention unless they had to be changed or suppressed. But if attention is a time-sharing mechanism, the packages would eventually interfere when enough of them were running at once to overload the system. Can attention be injected at trouble spots without degrading operations? ane the contingencie~ among p~oc~~~? At least some processes 3.5.9 How ~g~d should require the results of others in order to function. But it is absolutely necessary to obtain those results, or it is sufficient to estimate or predict them? Are the contingencies themselves routine, or do they arise sporadically? Is there some delicate timing of contingencies that might become disrupted, for example, if people had to drastically alter their rate of action? ~y~tem ~e ~n a child, and how do~ ~ evolve ~ 3.5.10 How do~ a p~oc~~~ng the child mat~~ and develop~? The human processing system must trave~se stages in order to deal with increased complexity. Steady accrual of experience should allow for refinement of both routines and specialists, and for generalizing across larger classes of events. But a non-strategic organization might prevent the evolution of the system and create learning blocks, e.g. the functional illiteracy some students retain throughout 12 years of pub1 ic schooling in America (cf. Beaugrande, 1982a). CM pMpMly d~~gned ;(:fuUyUng help to aptim~ze pMc~~~ng? More 3.5.11 elaborated processing models might make it apparent how schooling could promote comprehension skills in a more detailed, circumspect, and effective fashion than hitherto. Much language training is currently based upon very simplified or implausible models, encouraging such wasteful techniques as rote learning -- a brute force attempt to build surface traces without regard for cognitive organization. In consequence, children are usually left alone to devise their own methods of comprehending or learning texts. It would be a striking coincidence if many of them did not develop less than optimal systems (cf. Beaugrande, 1982b). me~~able time ~66Menc~ ~n exp~ental data be 3.5.12 How ~hould ~ntMp~eted? The classical experimental paradigm is to devise conditions, one of which includes some postulated process, and one that excludes it; if the inclusive task consumes more time, then the existence of that process appears supported or confirmed. Recent overviews (e.g. Clark and Clark, 1977; Levelt, 1978) indicate that this method has been extensively employed in the study of language comprehension as well. But if comprehension is composed of complex interactions of levels in parallel, and of detailed schedu1ings and contingencies, there is no clear reason to assume that all processes should be detected via increases in processing times. Automatized packages of sUb-routines (3.5.8) would not show up if they run simultaneously with other operations. Time would be linearly additive only for those processes demanding attention above a certain threshold. Moreover, time increments might indicate any of a great number of processes that participate in comprehension. dealwilh the multiple ~nt~p~etab~y 06 data? The time 3.5.13 How CM ~~eanch question (3.5.12) is just one illustration of a more general difficulty: the lack

190

R. de Beaugrande

of a univocal causality between a processing event and an observable, measurable action. The traditional method has been to design and test simplistic theories based upon such single causalities: if the variance was accounted for, all was well. Such a method is not sufficient for pursuing complex process models. Instead, we need to incorporate all plausible causes into a single model. We can then try to manipulate each factor independently and watch for variances in observed data. We can show that the several causes pote~y affect processing; but we will not normally be able to prove which cause aetua£1y affected an event occurring spontaneously in a natural setting (e.g. an observed pause in everyday reading). At best, complex process models can be correlated with their empirical domains through a steady elaboration of distributed probabilities. 3.6 The questions raised under 3.5.1-13 cannot be easily answered, yet they cannot be ignored either. They call for a fundamental reorganization both in theory design and in scientific methodology. We will have to learn to live with complexity not by factoring it out, but by enriching our theoretical models. In this paper, I have argued that progress will be more decisive if we concentrate less upon the structural analysis of brief artifacts, and more upon the exigencies of maintaining operations within complex, interactive systems. REFERENCES Adams, M., and Collins A. A schema-theoretic view of reading. In R. Freedle (ed.), New directions in discourse processing Norwood, N.J.: Ablex, 1979, 1-22. Allen, J. A plan-based approach to speech act recognition. Toronto: Univ. of Toronto diss, 1979. Atkinson, R. and Shiffrin, R. Human memory: A proposed system and its control processes. In K. Spence and J. Spence (eds.), The psychology of learning and motivation. New York: Academic, 1968, 89-195. Bartlett, F. Rememberin9. Cambridge, England: CUP, 1932. Beaugrande, R. Text and sentence in discourse planning. In J.S. Petofi (ed.), Text vs. sentence. Hamburg: Buske, 1979, 467-494. Beaugrande, R. de. Text, discourse, and process. Norwood, N.J.: Ablex, 1980. (a) Beaugrande, R. de. The pragmatics of discourse planning. Journal of Pragmatics, 1980, 4, 15-42. (b) Beaugrande, R. de. Design criteria for process models of reading. Reading Research Quarterly, 1981, 16, 261-315. (a) Fact, fiction, frontier? In J. Beaugrande, R. de. The linearity of reading: Flood (ed.), Issues in reading comprehension. Newark: IRA, 1981. (b) Beaugrande, R. de. Text, attention, and memory in reading research. In R. J. Tierney, J. Mitchell, and P. Anders (eds.), Understanding reader's understanding. Hillsdale, N.J.: Erlbaum, 1981. (c) Beaugrande, R. de. Text production. Norwood, N.J.: Ablex, 1982. (a) Beaugrande, R. de. Learning to read and reading to learn in the cognitive science approach. In H. Mandl, N. Stein, and T. Trabasso (eds.), Learning from text. Hillsdale, N.J.: Erlbaum,' 1982. (b) Begg, I. Recognition memory for sentence meaning and wording. Journal of Verbal Learning and Verbal Behavior, 1971, 10, 176-181. Bloomfield, L. Language. New York: Holt, 1933. Bray, N., and Batchelder, W. Effects of instructions and retention interval on memory of presentation mode. Journal of Verbal Learning and Verbal Behavior, 1972,11,367-374. Burton, R. Semantic grammar. Cambridge: BBN, 1976. Chomsky, N. Syntactic structures. The Hague: Mouton, 1957. Chomsky, N. Aspects of the theory of syntax. Cambridge: MIT, 1965. Clark, H., and Clark, E. Psychology and language. New York: Harcourt, Brace and Jovanovich, 1977. Cohen, P. On knowing what to say: Planning speech acts. Toronto: Univ. of Toronto di ss , 1978. Ertel, S. Where do the subjects of sentences come from? In Rosenberg (ed.), 1977, 141-167.

General Constraints on ProcessModels

191

Evans, S. A brief statement of schema theory. Psychonomic Science, 1967, 8, 87-88. Fahlman, S. NETL: A system for representing and using real-world knowledge. Cambridge: MIT Press, 1979. Fillmore, C. The case for case reopened. In P. Cole and J. Sadock (eds.), Grammatical relations. New York: Academic, 1977, 59-81. Fries, C. The structure of English. New York: Harcourt, Brace and Co., 1952. Fromkin, V. The non-anomalous nature of anomalous utterances. Language, 1971, 47, 27-52. Garrod, S., and Trabasso, T. A dual-memory information processing interpretation of sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 1973, 12, 155-167. Gibson, E. Perceptual learning and the theory of word perception. Cognitive Psycho logy, 1971, 2, 351-368. Goldman-Eisler, F. Psycholinguistics: Experiments in spontaneous speech. New York: Academic, 1968. Gough, P. Experimental psycholinguistics. In W. Dingwall (ed.), A survey of linguistic science. College Park: Univ. of Maryland, 1971. Greenbaum, S. (ed.). Language and acceptabil ity. The Hague: Mouton, 1977. Harris, Z. Methods in structural linguistics. Chicago: Univ. of Chicago Press, 1951. Hintzman, D., Block, R., and Inskeep N. Memory for mode of input. Journal of Verbal Learning and Verbal Behavior, 1972, 11, 741-749. Hormann, H. Meinen und Verstehen. Frankfurt: Suhrkamp, 1976. Johnson, N. The psychological reality of phrase structure rules. Journal of Verbal Learning and Verbal Behavior, 1965, 4, 469-475. Johnson-Laird, P.N. Experimental psycholin9uistics. Annual Review of psychology, 1974, 5, 135-160. Just, M. What your mind does when your eyes are reading. Paper at the AERA, Los Ange1es , 1981. Keele, S. Attention and human performance. Pacific Palisades: Goodyear, 1973. Kintsch, W. Memory and cognition. New York: Wiley, 1977. Kintsch, W. Psychological processes in discourse production. Paper at the Conference on Text Production, Kassel, June, 1980. Kintsch, W., and Dijk, T. van. Toward a model of text comprehension and production. Psychological Review, 1978, 85, 363-394. Kintsch, W., and Polson, P. On nominal and functional serial position: Implications for short-term memory models? Psychological Review, 1979, 4, 407-413. Kintsch, W., and Vipond, D. Reading comprehension and readability in educational practice. In L.G. Nilsson (ed.), Perspectives on memory research. Hillsdale, N.J.: Erlbaum, 1979,329-365. In L. Jeffress (ed.), Lashley, K.S. The problem of serial order in behavior. Cerebral mechanisms of behavior. New York: Wiley, 1951, 112-146. Levelt, W. A survey of studies in sentence perception: 1970-1976. In W. Levelt and G. Flores d'Arcais (eds.), Studies in the perception of language. New York: Wiley, 1978, 1-74. Loftus, G., and Loftus, E. Human memory: The processing of information. Hillsdale, N.J.: Erlbaum, 1976. Mandler, J. A code in the node. Discourse Processes, 1978, 1, 14-35. Mandler, J., and Johnson, N. Remembrance of things parsed. Cognitive Psychology, 1977, 9, 111-151. Marslen-Wilson, W. Sentence perception as an interactive parallel process. Science, 1975, 189, 226-228. Masson, M. Cognitive processes in skimming stories. Boulder: Univ. of Colorado diss, 1979. McCalla, G. An approach to the organization of knowledge for the modelling of conversation. Vancouver: Univ of British Columbia TR 78-4, 1978. Mehler, J. Some effects of grammatical transformations on the recall of English sentences. Journal of Verbal Learning and Verbal Behavior, 1963, 2, 250-262.

192

R. de Beaugrande

Meyer, B.J.F. The organization of prose and its effects on memory. Amsterdam: North Holland, 1975. Miller, G.A. The magic number seven, plus or minus two. Psychological Review, 1956, 63, 81-97. Miller, J., and Kintsch, W. Readability and recall of short prose passages. Journal of Experimental Psychology: Human Learning and Memory, 1980, 6. Norman, D., and Bobrow, D. On data-limited and resource-limited processes. Cognitive Psychology, 1975, 7, 44-64. O'Connell, D. One of many units: The sentence. In Rosenberg (ed.), 1977, 307-313. Oldfield, R. Memory mechanisms and the theory of schemata. British Journal of Psychology, 1954, 45, 14-23. Osgood, C., and Bock, K. Salience and sentencing: Some production principles. In Rosenberg (ed.), 1977, 89-140. Posner, M. Chronometric explorations of mind. Hillsdale, N.J.: Erlbaum, 1978. Posner, R. Types of dialogue: The functions of commenting. Discourse Processes, 1980, 3, 381-398. Raser, G.A. Recoding of semantic and acoustic information in memory. Journal of Verbal Learning and Verbal Behavior, 1972, 11, 692-697. Rosenberg, S. Semantic constraints on sentence production: An experimental approach. In Rosenberg (ed.), 1977, 195-228. Rosenberg, S. (ed.). Sentence production: Developments in research and theory. Hillsdale: Erlbaum, 1977. Rumelhart, D. Toward an interactive model of reading. In S. Dornic (ed.), Attention and performance VI. Hillsdale: Erlbaum, 1977. Rumelhart, D. Schemata: The building blocks of cognition. In R. Spiro, B. Bruce, and W. Brewer, (eds.) Theoretical issues in reading comprehension. Hillsdale, N.J.: Erlbaum, 1980,33-58. In R. Rumelhart, D. and Ortony, A. The representation of knowledge in memory. Anderson, R. Spiro, and W. Montague (Eds.), Schooling and the acquisition of knowledge. Hillsdale, N.J.: Erlbaum, 1977, 99-135. Sachs, J. S. Recognition memory for syntactic and semantic aspects of connected discourse. Perception and Psychophysics, 1967, 2, 437-442. Sapir, E. Language. New York: Harcourt, Brace, and Co., 1921. Schank, R., and Abelson, R. Scripts, plans, goals, and understanding. Hillsdale, N.J.: Erlbaum , 1977. Schank, R., Goldman, N., Rieger, C., and Riesbeck, C. Conceptual information processing. Amsterdam: North Holland, 1975. Schank, R., and Liebowitz, M., and Birnbaum, L. Integrated partial parsing. New Haven: Yale CS-TR 143, 1978. Shallice, T. Neuropsychological research and the fractionation of memory systems. In L.G. Nilsson (ed.), Perspectives on memory research. Hillsdale, N.J.: Erlbaum. 1979, 257-277. Shulman, H. Semantic comparison errors in short-term memory, Journal of Verbal Learning and Verbal Behavior, 1972, 11, 221-227. Skinner, B.F. Verbal behaVior. New York: Appleton-Century-Crofts, 1957. Tanenhaus, M., Carroll, J., and Bever, T. Sentence-picture verification models as theories of sentence comprehension. Psychological Review, 1976, 8, 310-317. Thorndyke, P. Cognitive structures in comprehension and memory of narrative discourse. Cognitive Psychology, 1977, 9, 77-110. Thorndyke, P., and Yekovich, F. Critique of schemata as a theory of human story memory. Poetics, 1980, 9, 23-49. Washburn, M.F. Movement and mental imagery. Boston: Houghton Mifflin, 1916. Winograd, T. Understanding natural language. New York: Academic, 1972. Wittrock, M., Marks, C., & Doctorow, Reading as a generative process. Journal of Educational Psychology, 1975, 67, 484-489. Woods, W. Generalizations of ATN grammars. In W. Woods and R. Brachman, Research in natural language understanding. Cambridge: BBN Technical Report 3963, 1978, 21-77.