Click here to go to Ch IV

 

Click here to go to Table of Contents

Informativity

1. MODIFYING INFORMATION THEORY  

1.1 Despite some diffuseness in its usage over the years, the term INFORMATION can be taken to designate not the knowledge that provides the content of communication, but rather the aspect of newness or variability that knowledge has in some context (cf. Loftus & Loftus 1976; Gröben 1978). If the actualization of a text system is constituted by a configuration of OCCURRENCES (cf. 1.1.3.; 1.4.1), then the INFORMATIVITY of a particular occurrence is its relative PROBABILITY (likelihood and predictability) as compared to other ALTERNATIVES. The lower the probability of the occurrence, the higher the informativity (cf. I.4.II.7).

1.2 In classical informational theory (Shannon & Weaver 1949), informativity (information value) was formalized by statistical methods. Suppose we had a language with a precisely enumerated set of possible elements (a 'finite-state language'). We could select an element, say X, and look at every occurrence of X in any chain. If we had a chain like W-X-Y etc., we could compare all these occurrences and compute the TRANSITION PROBABILITY between W and X, that is, the likelihood of X following W. A chain constituted according to this simple computation of transition probabilities between immediately adjacent elements is called a MARKOV CHAIN. It is questionable, however, whether Markov chains are a useable model for natural language utterances. Natural languages do not have a finite number of states, and the probability of any occurrence does not depend solely on the immediately preceding occurrence.

1.3 I hold a flexible, modified version of information theory to be valuable for theories of human communication via texts. The augmented transition networks I have proposed as operational representations for processes of sequential connectivity (II.2.12) and conceptual connectivity (III.4.7) bear a distant resemblance to the old Markov chains, because the main task is predicting the next link to a new node. Experiments inspired by the model of the augmented transition network did show that language users have fairly uniform expectations about how a sentence sequence will proceed from a given point (cf. II.2.14). A large quantity of learning experiments were based on Markov models, due to their mathematical simplicity (Kintsch 1977a: 82). But purely statistical models in general, and Markov models in particular, would lead to COMBINATORIAL EXPLOSION (II.1.2) for processes as intricate and varied as the utilization of texts: the decision about an impending occurrence rests less upon frequencies between adjacent items than upon the MOTIVATIONS the overall context supplies. Lcon Brillouin (1956) suggests that the statistical approach ignores the whole aspect of meaningfulness.

1.4 Psychologically, statistics might be applicable to the aggregate of EPISODES a person has in stored knowledge. Yet as episodic memory gradually feeds over into CONCEPTUAL memory (III.3.16), exact frequencies would tend to become blurred and unreliable for building expectations. To select an option at a given point during the production or prediction of a text sequence, people presumably consult all available CUES (signals for performing a processing action). The availability of cues depends upon the FOCUS of ATTENTION, where “attention” is defined as an expenditure of processing resources that limits the potential for another task at the same time (Keele 1973). Cues would be especially helpful if people were working with the various language systems in PARALLEL and merging shared parts of hypotheses about those systems (cf. Woods 1978b: I 1; III.4.14).

1.5 Due to the various exigencies of communication, the occurrence of an element could have quite different probabilities in different systems; it might, for example, be syntactically probable and semantically improbable, or vice versa. If we had PROBABILITY OPERATORS for the links of the grammatical and conceptual/relational network — a feature I hope to include as soon as sufficient empirical research makes it possible and reliable — the operators on the same link would be opposed in the two networks. I suspect that the PROBLEMATIC transition to the improbable element in one system (cf. 1.6.7) is eased by a comparatively unproblematic transition in the other. Probable content in a probable format would be uniformly easy to process and not informative. Improbable content in an improbable format would be uniformly difficult to process and intensely

problematic. But improbable content in a probable format, or probable content in an improbable format would be challenging and yet not unreasonably problematic. Literary and poetic texts (cf. VII.I.8.4-5) often manifest these last two combinations (cf. Beaugrande 1978b, 1979e; Koch 1978, 1979). We should bear in mind that the probabilities in virtual systems can beoverridden by those in actual systems (cf.IV.1.23.4). People seem to be quite skillful in adapting their expectations to an intricate pattern of actual episodic occurrences (cf. Friedman, Burke, Cole, Estes, Keller, & Millward 1963). While it was found that, when taken as abstract sentence patterns, the passive is harder to process than the active (Coleman 1964), a text with nothing but passive constructions removes the difficulty (Wright 1968).¹ [1. The difficulty of passive sentences is exaggerated in many experiments with samples where the roles are reversible (i.e. the agent might reasonably also be the affected entity and vice-versa) and no determinate contexts are given. Slobin (1966) demonstrated the importance of reversibility in such measurements.]

1.6 It would be reasonable to distinguish various ranges on a scale of informativity. I shall propose three ORDERS, with “order” used in the mathematical sense: a higher-numbered order automatically subsumes the lower-numbered ones. The order results from the extent of PROCESSING RESOURCES that are expended upon input. The lower-order occurrences allow PROCESSING EASE, that is, the linkage of the occurrences to previous ones is non-problematic. The higher-order occurrences call for PROCESSING DEPTH (cf. III.3.5), because the linkage is problematic, perhaps seriously so (cf. I.6.7 on “serious problems'). The THRESHOLD OF TERMINATION where processing is considered satisfactory and discontinued (III.3.24) therefore moves along with the order of informativity.

1.7 The complexity of probabilities suggests that people could rely not only on prediction, but on “postdiction” as well (Kintsch 1979a). The understander would then notice an occurrence and seek some justification after the fact. Reliance on postdiction would increase either (1) if there were a wide spread of equally probable alternatives and a scarcity of determinate cues for selecting any; or (2) if an occurrence seems quite outside the predicted range, so that no cues are readily at hand. The second case doubtless requires a stronger focus of attention, and can be strategically induced for that motive (see note 14 to Chapter I).

1.8 The mere selection of one available option in a context — an option provided by any participating system — results in at least FIRST-ORDER INFORMATIVITY. In the simplest instance (a rare one) where there seems to be only one option, there are still two alternatives: occurrence versus non-occurrence. In a restricted sequence where only two options are possible (as in many learning experiments), there are the trivial alternatives of any occurrence being the same as or different from its predecessors (a principle of the “text-score” developed in Weinrich 1972).² [2. Intriguingly people expect a long series of the same occurrence to be broken for the sake of mere variety, even when probabilities remain constant — a phenomenon called “gambler’s fallacy’ (cf. Kintsch 1977a: 91f.). In more realistic worlds with multiple alternatives, first-order informativity applies when an option in the upper range of probability is selected. In all of these domains, we have a low INTERESTINGNESS value: the degree of cognitive involvement resulting from uncertainty (as well as from such factors as emotivity and salience — see section IV.2).

1.9 Many selections required for the production of any text are of this trivial first order. Given a conceptual configuration and the preferences for mapping it onto surface expression (III.4.16), many decisions regarding surface structure are made efficiently (cf. I.4.14). The effectiveness of certain formulations, notably in poetry, arises from low probability in mapping. In its attempts to set up a categorical, context-free grammar that stipulates what sentences can and cannot occur, generative grammar implied the postulate that all potential occurrences in a language system are of the first order, because specified by categorical rules (cf. I.3.4.7). People’s difficulties with judging unusual sentences (I.1.16) show that the variability of information orders should not be ignored when constructing a grammar for sentences. My proposal to include the notions of DEFAULT and PREFERENCE in a grammar for texts (cf. I.3.4.3) might help resolve this matter.

1.10 A normal reaction to triviality would be to reduce one’s ATTENTION, i.e., the concentration of processing resources on one object at the expense of others. In any case, humans in communication are not likely to perform a thorough analysis of all occurrences in all systems, such as a linguist might accomplish. I suggested in III.4.15 that the intense utilization of surface structure would be needed if there were numerous or evenly matched hypotheses about the underlying conceptual/ relational structure. If the latter were immediately obvious, on the other hand, people might do only “fuzzy parsing” on the surface. The processor would leave sonic nodes or links unlabled (cf. Burton 1976: 80), working along via approximative problem-solving. If it later emerges that the unlabeled states are needed after all, but are no longer available in active storage, problem-solving could become more detailed and rigorous to reconstruct the lost material.³ [3. Some successful computer simulation of the processing of indistinct or partial input uses precisely this approach (cf. Woods et al. 1976).] If this outlook is plausible, then low-order informativity is a reliable signal that fuzzy parsing is adequate in a given context.

1.11 The selection of an option in the middle or lower-middle degrees of probability results in SECOND-ORDER INFORMATIVITY. Here, the strongest defaults and preferences are noticeably overridden. The presence of at least some second-order occurrences is presumably the usual standard for textual communication, so that first-order occurrences could be UPGRADED (unless they are accorded no further attention) and third-order occurrences could be DOWNGRADED. The demands people make for informativity vary among types of texts and situations. Conversations between married couples appear (in my view) to function with very low informativity, while contemporary art works strive for very high.

1.12 Occurrences construed as outside the range of more or less probable options convey THIRD-ORDER INFORMATIVITY. These are unusual and extremely interesting occurrences, and correspondingly hard to understand and control. A SERIOUS PROBLEM in the sense of I.6.7 is present, because the linkage of the new occurrence to what went before is endangered in an unexpected way, and the probability of FAILURE is great. Major DISCONTINUITIES, GAPS, and DISCREPANCIES as defined in I.6.9 are the usual types of third-order occurrences and activate a MOTIVATION SEARCH to find out a source for the unexpected material. The search returns some pathway which makes the third-order occurrence accessible to its context and hence within the range of probable options after all (cf. Lenat 1977: 1097). This process in effect DOWNGRADES the third-order occurrence into the second order. Downgrading could have different DIRECTIONALITY: (1) if people regress to occurrences of a considerably earlier time to find the motivating pathway, they are doing BACKWARD downgrading; (2) if they wait and look ahead to further occurrences, they are doing FORWARD downgrading; (3) if they go outside the current context, they are doing OUTWARD downgrading. A text producer who deliberately supplies third-order occurrences may anticipate the directionality and results of the downgrading as part of the plan toward a goal (cf. Beaugrande 1978b; VII.2.33). The assumption that downgrading will be done is reliable (Berlyne [19601 suggests that “cognitive conflict” creates “epistemic curiosity” to obtain knowledge).

1.13 The directionality of downgrading suggests the control flow for processing third-order occurrences. In II.2.34, we considered what might ensue if a sentence structure were so misleading that an unaccountable element was left over at the end of parsing. The subsequent relabeling of the structure (Figures 9a and 9b) was an illustration of backward downgrading in the syntactic system. For a structure that cannot be downgraded via syntax alone, such as Simmons’ sample of ‘The old man the boats’ (see (22) in II.2.32), a processor could go outside to consult intonation or conceptual context (outward downgrading); or could leave the structure temporarily uniabeled until the context became more determinate later on (forward downgrading). If the processor interrupted the speaker with a demand for explanation, we would have a convergence of outward and forward downgrading.

1.14 These reasoning procedures doubtless extend far beyond the utilization of texts. If we are arrested with no warning and for no visible motive, we have encountered a third-order experience. We will be prone to react in the following ways: (1) mentally retracing our recent actions to see if any of them could be the ‘reason-of’ the arrest (backward downgrading); (2) waiting to be told the reason by an officer of the law (forward); (3) trying to remember cases where someone was arrested because of mistaken identity (outward). If successful, these activities downgrade the arrest-event, and if not, we will be unable to understand it. Meaninglessness, I would argue, results from the lack of continuity and connectivity, and not from the undecidability of truth values (cf. III.1.2).

1.15 STRENGTH OF LINKAGE (III.3.15) in world-knowledge is relevant to informativity orders. If a textual world asserts relations known to be DETERMINATE already, we have the first order only. The assertion of TYPICAL relations brings more informativity as typicalness decreases. The assertion of ACCIDENTAL relations is by itself neutral for informativity, since accidents may range from the trivial to the unique. The assertion of non-typical relations results in at least second-order, and the contradiction of determinate relations results in third-order informativity. If a tree in a textual world is assigned a trunk, little interest is aroused, that being a stored determinate ‘part-of’ link. If the tree has multiple trunks we are more interested, though not disoriented (non-typical but allowed, hence second-order). If the tree has no trunk at all, its branches hovering in mid-air, we are alarmed by a conflict with a determinate link (third order) and expect an explanation or assume we are dealing with a highly fictional text-world (downgrading).

1.16 The fictional text-world instructs the processor to relax the application of real-world expectations. In Alice's Adventures in Wonderland (Carroll 1960), the initial plunge down an impossible rabbit-hole filled with cupboards and bookshelves at once marks the textual world as not governed by the same organization as the reader’s. After a series of strange occurrences, the narrator remarks about a normal event (Carroll 1960: 33):

(36) Alice had got so much into the way of expecting nothing but out-of- the-way things to happen, that it seemed quite dull and stupid for life to go on in the common way.

Yet the Alice-world is by no means devoid of continuity and coherence. Many real-world expectations still apply: gravity makes things fall, water makes things wet, characters speak English, etc. Some domains are understandable via opposition to the real-world: assignment of human roles to animals or playing-cards, violation of politeness conventions, etc. The enduring interestingness of the Alice books arises from experiencing a text-world whose third-order occurrences are downgradable by discoverable principles (cf. IV.1.23.1). During the activities of downgrading, readers discover by analogy how the organization of the real-world is arbitrary and amusing (cf. ‘Vll. 1. 8.4).

1.17 Original METAPHORS can constitute third-order occurrences. The fragment of Dylan Thomas’(1971: 196) poem ‘In my craft or sullen art’ that runs:

(37) In the still night, when only the moon rages

presents a totally non-expected action or emotion of the moon that no reader would have in stored knowledge. To process the fragment, the reader must integrate the problematic element, for example, by reasoning: (1) that the moon’s surface resembles the face of a ‘raging’ person with staring eyes and open mouth; (2) that the moon is traditionally believed to cause lunacy and hence ‘raging’ in people; (3) that the moon’s casting light in all directions resembles a ‘raging’ person throwing things all around; and so forth. Hence, the original metaphor elicits a resolvable discrepancy between text-presented knowledge and previously stored knowledge. There need be no particular literal expression that accomplishes the same thing as the metaphoring (Ortony 1978c). The discrepancy is below surface structure, and its downgrading may be undecidable, as we saw with the Thomas fragment. A literalized restatement could be an impoverishment or even a mis- representation.

1.18 In recent times, literary texts are characterized by more numerous third-order occurrences that are increasingly resistant against downgrading. That trend is conspicuous even in the progression of James Joyce between Ulysses and Finnegan’s Wake. In the earlier novel, the selective principles applied to language options are periodically reorganized, calling forth an adaptation of expectations. In the later novel, the author applies the far more complex principle of simultaneous partial actualization of different options, many from other languages besides English, so that no comprehensive expectations about surface structure occurrences can be maintained, and even logical identities are blurred. Experiments of the latter kind (also in the poetry of Hans G. Helms) have an intrinsically limited acceptability as texts, because they run counter to human processing strategies. Constant blocks against downgrading third-order occurrences place an enormous strain on processing energy, which most readers see no reason to sustain. For some readers, an enriching awareness results about human reliance on expectations in ordinary communication. Yet the processing of a text or situation where continuity is steadily at the break-down point is internally paradoxical and is tolerable only for correspondingly pre-trained readers. It is noteworthy that some literary critics have undertaken instead to explicate Finnegan’s Wake in conventional language: perhaps the most colossal downgrading in history.

1.19 The procedures of UPGRADING are also intriguing. If something is well-known or even determined by standards of logic or science, people should have little reason to assert it by means of a text. Here again, a MOTIVATION SEARCH (IV.1.12) is likely to take place. Consider the example (in Beaugrande 1978b: 11) of a woman introducing her husband at a party with the utterance:

(38) My husband is a human being.

She assigns to a person a relation that should be stored already as a determinate ‘instance-of’ link for all people. Hearers will want to discover why the woman makes the effort to say so, because communication is by default presumed to have a reason (cf. Rieger 1975: 160). They could recast the utterance into an expanded format with explicit motivation, such as:

(38a) My husband is so nondescript that one can’t say much about him except that he is a human being

(38a) overturns the expectation that one ought to be able to say more than (38). (38b) serves to signal that the ‘instance-of’ relation is in fact less probable than might be assumed. These replacements of (38) with assumed alternative versions illustrate outward upgrading of a first-order occurrence in the conceptual/relational system. A demonstration of forward upgrading in that system — not an uncommon procedure for the beginnings of texts — can be found in this opening passage from a science textbook (quoted in Beaugrande 1978a: 29f.):

(39) The sea is water only in the sense that water is the dominant substance present. Actually, it is a solution of gases and salts […]

The first-order informativity of the determinate “substance-of’ relation in ‘the sea is water’ is made upgradable by the subsequent assertion that this piece of common knowledge is ‘actually’ not accurate and is hence not so probable as it seems. The demands of informativity can even eliminate alternative readings, as is shown by this headline (Gainesville Sun, Dec. 20, 1978):

(40) San Juan Gunfire Kills One

The reading where ‘one’ is taken as an impersonal pronoun (hence: ‘San Juan gunfire kills people’) is ruled out as uninformative (unless, of course, gunfire in other Puerto Rican cities were not fatal) and hence not newsworthy.

(41) His life was gentle, and the elements

And say to all the world, ‘This was a man!’

the audience will attribute more to the utterance ‘this was a man’ than a first- order ‘instance-of’ relation. They will rather prefer an understanding such as that this ‘man’ is an infrequent class of humans who possessed the full range of ‘elements’ offered by ‘Nature’.

1.22 If communication is composed of LEARNING SYSTEMS that adapt to their environment (I.4.3), it follows that the immediate expectations of a context would override those based on general knowledge. Over time, special utilization of systems engenders evolution. For example, highly respected literary texts could serve to expand the possibilities for conventional expression, or to propagate alternative viewpoints about reality via the mode of fictionality (cf. VII.1.8.4f.). Wolfgang Iser (I975: 302) observes that the literary text both stabilizes and interferes with the operations of communicative systems.

1.23 To explore communicative probabilities in more detail, we need to classify expectations into a hierarchy such as the following:

1.23.1 Stored knowledge and episodic experience lead people to see the world in a certain way. The socially dominant model of the human situation and its environment evolves into the notion of the REAL WORLD and is henceforth privileged over all other models. Propositions judged to be true in this world are conventionally called FACTS (cf. Schmidt 1979), and are entered into socially shared BELIEF SYSTEMS (cf. Bruce 1975) as the most fundamental assumptions about the organization of knowledge and experience. Some facts and beliefs are so firmly established that they act as defaults pervading almost any textual world that might be created: that causes have effects; that time can move in only one direction; that matter cannot be totally destroyed; that entities cannot be both existent and non-existent, present and absent, or possible and impossible at the same time and under the same circumstances; and a great deal more. A textual world in which such basic facts and beliefs are countermanded, e.g., science fiction stories, must provide distinct cues in relevant contexts. These cues are instructions that the text receivers should make specified modifications in their expectations lest the textual world become inaccessible and its organization unbearably problematic. On the few occasions where Lewis Carroll does make use of the reversal principle derived from mirror imagery in Through the Looking-Glass (Carroll 1960: 205, 249f., 290), he is very emphatic. I suspect that strict adherence to such nonce facts in fictional worlds would soon lose informativity as a corresponding set of expectations is tailored to the occasion (cf. IV. 1. 5). The continuing interestingness of the Alice world is upheld by the variety of its principles for unconventional organization (cf. also IV. 1. 16).

1.23.2 People also have expectations about LANGUAGE, such as about sequencing (ch. II) and conceptual connectivity (ch. III). People rely on this knowledge to deal with predictable expressions. Users of English do not anticipate unpronounceable clusters of consonants (except in abbreviations), so that when asked to “read aloud’ a line on an eye-testing chart, such as:

(42) PDZTLF (Snellen eye chart)