I I I
1. MEANING AND PHILOSOPHY
1.1 Although often neglected in conventional linguistics, meaning has long been an object of dispute in philosophy. Since antiquity, philosophers have envisioned the construction of a mode of LOGICAL EXPRESSION. The mode was expected to be exact, non-ambiguous, and concise. Strict rules should make it decidable if any statement was true or false, and whether any statement could correctly be proven from another. All statements had obligatory symbolic formats that could be translated into declarative sentences of natural language: the subject/ predicate positions corresponded to the symbols or slots for argument/predicate, object/function, etc., depending on the type of logic. To connect statements, JUNCTIVES were defined according to their effects on TRUTH VALUE. If two statements were true by themselves, their conjunction with ‘and’ was also true; if either was false, the whole conjunction was false. A disjunction with ‘or’, on the other hand, was true provided only one of the statements was true.). The junctives ‘if - then’ and ‘if and only if’ usually written 'iff’ were also defined regarding truth value (for further discussion, cf. van Dijk 1977a, 1977b). n conjunction and disjunction, cf. I.2.15’ V.y7) regarding truth value (for further discussion, cf. van Dijk 1977a, 1977b).
1.2 The tendency to identify meaning with truth value has been
widespread. Rudolf Carnap (1942: 22), for example, remarks:
Semantic rules determine truth-conditions for every sentence of the
object language [...] To formulate it another way: the rules determine the
meaning or sense of the sentences. [emphasis added]
This conflation has several consequences. First, philosophers have
expended great energy on debating unresolvable paradoxes about truth, such as
Strawson’s (1949: 90) example
(28) What 1 am now saying is false
where the statement is true only if it is false. Second, the issue of REFERENCE assumed a disproportionately prominent role in theories of meaning. Third, statements whose truth value cannot be decided are to be considered meaningless; yet undecidable statements are produced and understood constantly in everyday communication (Miller& Johnson-Laird 1976).
1.3 REFERENCE is usually defined as the relationship between expressions and those objects or situations in the world the expressions designate. Among the very diverse and intricate forms of reference, logicians are concerned with very few, notably with ‘quantificational status.’ If one unique object is referred to, an ‘existential quantifier’ marks it as an existing object in the real world. The most obvious case is names of persons, as we can see by their frequency in logicians’ examples (and inherited over into a linguistics of ‘John and Mary’ sentences). However, the human activities of using proper names are not at all straightforward, to say nothing of descriptive expressions (cf. J. Anderson & Bower 1973; Ortony & R. Anderson 1977; J. Anderson 1978: Kalverkimper 1978). If a whole set of objects is referred to, a ‘universal quantifier’ signals that any statements must be true of every single object having that name. These two quantifiers allow one to make ASSERTIONS about objects and to construct proofs, which yield values of either true or false (cf. sample (87) in V.3.12).
1.4 Although logics of this kind are in themselves unobjectionable, they create vast confusion if taken as a model for human language communication. The following difficulties must be confronted.
1.4.1 ASSERTION is a HUMAN ACTION of entering a statement into a textual world. Logic misses the important factors of CONTROL (Levesque & Mylopoulos 1978:2) and of the speaker’s INTENTION (Cohen 1978: 18). REFERRING is also a human action and not a property of noun phrases (Morgan 197Sa: 109).
1.4.2 Human knowledge of the world creates a rich background of defaults, preferences, contingencies, and interactions for any assertion someone might make. Communicative situations are sensorially accessible and related to a wealth of past experience. All of these outside materials are usually allowed no place in logic.
1.4.3 The strict rules of logics render the assertions they permit obvious or even tautological. Human communication thrives on uncertainties, exceptions, variables, and unexpected events — all of which render a statement interesting, whether its truth can be determined or not.
1.5 If logics are to be useful in theories of natural language, their flexibility and scope will have to be enormously increased. Methods will have to be found for making logical procedures operational (se Simmons& Bruce 1971; Kowalski 1974; Cercone & Schubert 1975; Warren & Pereira 1977; Levesque & Mylopoulos 1978). The notions of truth and existence could be treated as DEFAULTS assumed in otherwise non-committal contexts. For example, people can be expected to believe in the truth of their statements (Grice 1975) except when signals to the contrary are provided (cf. Weinrich 1966a). This belief would yield not CORRECT ASSERTION (exact correspondence with the world), but JUSTIFIED ASSERTION; in many cases, however, we find MOTIVATED ASSERTION of materials whose truth is undecidable or even known to be false (Beaugrande 1978b: 7).
1.6 Due to an interest in quantification, theories of reference have often made use of SET THEORY. Whereas a CLASS is constituted according to some identifiable characteristic of its members and is thus indispensable for the organization of knowledge (cf. 111.3.19), the SET is constituted simply by the fact that some elements belong to it. I have misgivings about the usefulness of set theory in a model of human communication. To claim that by uttering:
(29) Macbeth doth murder sleep, sleep that knits up the ravelled sleeve of care. (Macbeth, Act 11, scene ii, 36 ff.)
the speaker is intersecting the (thankfully single-member) set ‘Macbeth’ with the set of people who murder sleep, sleep being itself intersected with the set of things that knit the ravelled sleeve of care, certainly doesn’t resolve the issue of meaning; it only restates it in more pompous terms. Moreover, set intersection is operationally cumbersome,1 [1. Smith, Shoben, and Rips (1974) propose a set-theoretical mode1 of meaning in which concept figures as an ordered set of features. But, as Hollan (1975) contends, their model can, in fact, be formulated as a network model with a gain rather than a loss of representative power. I would add that the ordering of pairs in sets would encourage an atomistic outlook on the task of modeling the meaning of whole texts.] since for a given statement, one often has to look at all members of at least one set, and in the worst case (e.g. disproving false statements about one member of a set) at all members of both sets (but see now Fahiman 1977: 31).
1.7 Future revisions of logic may amend the shortcomings I note here. However, it is difficult to imagine how a logical system could be devised without MODULARITY: independence not only of system components, but also of every statement and expression, from contextual influences (cf. I.2.7). The whole enterprise of formal logic seems to disregard the continuities that people experience through their senses (cf. Shepard & Metzler 197 1; Cooper & Shepard 1973; Kosslyn 1975). Perhaps a system for extremely fast computation of discrete symbolic descriptions, as envisioned by Marvin Minsky (1975), may yet approach logical rigor.
2. MEANING AS FEATURE CLUSTERS
2.1 When meaning entered into American linguistics after a long exile, it was approached with methods similar to those that had been successful in descriptive phonology and morphology. The meaning of all expressions in a language was treated like the sound substances decomposable into a repertory of minimal units (e.g. Katz & Fodor 1963; Pottier 1963; Prieto 1964; Bierwisch 1966; Greimas 1966; Coseriu 1967; Nida 1975).2 [2 I wonder whether this transfer of phonological methods to other levels of language did not undermine, in a meta-perspective, the proclaimed independence of levels from each other.] The minimal units were variously called ‘semes’ or ‘sememes’ (analogy to ‘phonemes’), or semantic ‘features’ or ‘markers’ (on the last two terms, cf. Hörmann 1976: 78). The status of these constructs was interpreted variously, for example:
2.1.1 as the ‘linguistic image of properties, relations, and objects in the real world’ (Albrecht 1967: 179; compare Pottier 1963);
2.1.2 as distinctive elements arising from the ‘apperceptive constitution’ of ‘human beings in regard to their environment’ (Bierwisch 1966: 98);
2.1.3 as elements for building up a semantic theory (Katz & Fodor 1963);
2.1.4 as conceptual elements into which a ‘reading’ decomposes a ‘sense’ (Katz 1966);
2.1.5 as constituents of a metalanguage for discussing meaning (Greimas 1966).
2.2 There are two general perspectives here: (1) psychological reality (Albrecht, Bierwisch, to some extent Katz), versus (2) linguistic theorizing (Katz & Fodor, Greimas). If we adopt the psychological perspective, the substance of meaning becomes an empirical issue (Winograd 1978: 30). In the linguistic perspective, the creation of theories of meaning is entirely the responsibility of introspection and systemization. Whichever approach is adopted, the following questions present inordinate difficulties:
2.2.1 How can the briefest, yet most universally applicable catalogue of units be set up for an entire natural language?
2.2.2 How many minimal units must a human store in order to communicate, and in what format?
2.2.3 How can these units reflect the fact that all domains of meaning cannot look the same (cf. Meehan 1976: 225; 111.2.4)?
2.2.4 How can we deal with RESIDUAL MEANING: idiosyncratic meaning in words and expressions that is not covered by usual units? If we convert all residue into units, we explode the system beyond all proportion with elements that might (in the worst case) be needed for only a single word.
2.2.5 Will the set of postulated units also apply to every new
expression that could ever be added to the language?
2.2.6 How can the units themselves be expressed without using natural language expressions that could be decomposed in their turn (cf. Wilks 1977a)?
2.2.7 How can we deal with the adaptation of expressions and their content to contexts: are there different unit configurations here, of the same units with different values (cf. Hormann 1976: 141)?
2.2.8 Where should decomposition stop without going into INFINITE REGRESS: the continual subdivision into ever smaller components (cf. Winograd 1978: 28)?
2.2.9 How could decomposition operate in real time without a dangerous explosion of content (Wilks 1975a: 22)?
2.2.10 How are word meanings acquired, given that miminal units are not encountered in everyday communication?
2.3 In a processing model, minimal units figure as PRIMITIVES: irreducible units for processing all comparable content in the same terms. Although they would be desirable for procedural considerations such as formatting and storage (cf. Winston 1977: 198), systems of primitives would have to meet formidable requirements: (1) the entire range of language expressions would have to be covered by a finite set of primitives; (2) primitives should not be explained in terms of each other; and (3) primitives should not be capable of further decomposition (Wilks 1977a; Winograd 1978). The question arises whether such thoroughness and completeness is even necessary for everyday comprehension (Rieger 1975: 204). Many utterances would present fearsome intricacies resulting out of unconventionality or vagueness of usage (on dealing with vagueness, cf. Eikmeyer & Rieser 1978).
2.4 There are clear differences in the internal structuredness of knowledge domains. The proponents of minimal units invariably select well-structured domains, such as kinship terminology (e.g. A. Wallace & Atkins 1960; Lounsbury 1964). Here, concepts are almost entirely relational themselves and hence perfectly suited for non-residual decomposition: ‘male/female’, ‘parent/ child’, and so on (Kintsch 1979b: 20). Speakers of English would be hard put to supply the components of concepts like ‘intelligence’, ‘beauty’, ‘absurdity’, ‘essence’, and so forth with any wide agreement. A model of meaning must make a distinction between concepts whose function is to represent relations, and concepts with more diverse and intricate functions of representing content (Shapiro 1971).
2.5 There appears to be a TRADE-OFF in the usefulness of minimal units. The larger the store of knowledge becomes and the more diversified the domains, the less we have to gain by reducing everything to minimal units. I would accordingly conclude that decomposition of meaning has the same human psychological status as that assigned to transformations in II.1.9: the operations involved can be performed if a task and a domain make it worthwhile, but they are not done routinely (see Kintsch 1974: ch. II for a survey of tasks). The question will have to be solved empirically rather than by linguists’ debates (Kintsch 1974: 242), and the evidence for decomposition is slight so far (J. Anderson 1976: 74).
2.6 The questions involving the featural approach will not be resolved very soon. Perhaps it would be useful to look in the opposite direction: not at segmentation but at continuity. While there is little evidence yet that humans break meaning into tiny units when they communicate (barring discussions among linguists), there is good evidence that people must build large configurations of meaning in order to utilize whole texts (e.g. when planning, learning, recalling, or summarizing textual content). I shall follow up some PROCESSES which could plausibly contribute to this continuity of meaning in communication via texts.
MEANING AS PROCESS
3.1 The identification of meaning with usage was proposed especially
by Ludwig Wittgenstein (1953; cf. also Schmidt 1968b). I adopted a similar
outlook on Harris’s distributional approach (see I.2.3). However, we are hardly
likely to ever compile an exhaustive record of all uses of even one word, let
alone the whole lexical repertory of a language. We can at best seek to discover
processes that operate generally on usage as an activity of building up meanings
3.2 For that
undertaking, a PROCEDURAL SEMANTICS would be productive (Miller &
Johnson-Laird 1976; Winograd 1976; Bobrow & Winograd 1977; Johnson-Laird
1977; Levesque 1977; Havens 1978; Levesque & Mylopoulos 1978; Schneider
1978). Many approaches that do not expressly call themselves by that term share
the outlook that meaning results from actions in an intelligent processor (e.g.
Schank et al. 1975; Woods 1975; Fahlman 1977; Hayes 1977; Brachman 1978a; Cohen
1978). The formatting of knowledge for optimal processing has been in debate.
DECLARATIVE knowledge is formatted as statements that might be used in many
different and possibly unforeseen ways. PROCEDURAL knowledge, in contrast, is
formatted as programs designed to run in specifically anticipated ways.
Declarative knowledge is thus more versatile in its applications, but its actual
uses are less efficient. Debates stressing the opposition of these standpoints
(sample in Winston 1977: 390ff.) are misleading, however. The question is one of
different PERSPECTIVES taken on what is in essence the same knowledge (cf.
discussions in Winograd 1975; Scragg 1976; Bobrow & Winograd 1977; Goldstein
& Papert 1977). In a very small knowledge-world, only a few facts are known
and the processor is not yet very intelligent, creating a need for explicit
programs. But in an extensive and richly interconnected world, the declarative
and procedural aspects begin to converge: the structuring of knowledge is
simultaneously a statement of how it can be accessed and applied. Only if
meaning and use are taken as independent — as ‘inonisms’ that deny each other
(R. Posner 1979b) — do we have to content.
3.3 The basic unity for a procedural semantics would be the PROPOSITION as a RELATION obtaining between at least two CONCEPTS (cf. Kintsch 1972,1974; Rumelhart, Lindsay, & Norman 1972; J. Anderson & Bower 1973; B. Meyer 1975,1977; Frederiksen 1975,1977; J. Anderson 1977). These entities depend on the degree of detail required for a processing task. Many concepts can be analyzed into propositions (cf.III.4.4), and in a task like summarizing, propositions might be subsumed into single concepts (cf. Ausubel 1963). Searle (1971: 141) argues that REFERENCE can only be accomplished via propositions, because if someone merely expressed a concept, there would be no way to identify what was meant. Leonard Linsky (1971: 77) supports this view in suggesting that ‘referring expressions’ cannot be treated without their context. It seems to me that referring is in fact accomplished via the entire TEXT-WORLD MODEL as outlined in 1.6 and further depicted in the following section. If people do match the content of texts with their notion of the real world, then the completed text-world model should give the clearest indications of what to look for. There is probably a THRESHOLD OF TERMINATION, both for the degree to which concepts are broken into propositions (or propositions subsumed under concepts), and for the extent to which text content is actually matched with whatever is taken to be the ‘real world.’
3.4 A traditional example of a proposition would be something like:
(30) Socrates is Greek.
Where ‘Socrates’ is the ARGUMENT and ‘Greek’ is the PREDICATE. Since sentences are not propositions, however, many researchers prefer a format such as this:
(31) (GREEK, SOCRATES)
The conventional viewpoint in logic is that predicates are
‘designations for the properties and relations predicated of individuals’
(Carnap 1958: 4). My use of the notion of ‘proposition’ will be kept informal so
as to cover a very wide variety of content (cf. III.4.7ff.). \
3.5 WORDS or WORD GROUP UNITS are EXPRESSIONS: SURFACE names for UNDERLYING concepts and relations. The use of expressions in communication ACTIVATES these concepts and relations, that is, enters their content into ACTIVE STORAGE in the mind. The transition between expressions and their content is an aspect of MAPPING (cf. I.2.10). A given concept may have alternative names which are SYNONYMS to a greater or lesser extent, depending on how much conceptual relational substance they activate. Although synonymity is probably rare in the virtual system of the LEXICON cf. 18.104.22.168), it is may be accepted in the actual systems of textual worlds where the interaction of concepts controls the amount of substance being activated. In return, a single expression may be able to activate various concepts according to its use; the expression can then be said to have several SENSES (cf. P. Hayes 1977: Rieger 1977b; Small 1978). The existence of synonyms and multiple senses are evidence of the ASYMMETRY between expressions and their meanings (cf. 1.6.12). This asymmetry assumes different proportions in various languages (cf. Wandruszka 1976), so that concepts must be in part language- independent (cf. Schank 197Sa: 256, 1975b: 7). The borderline between expressions and concepts is not clear-cut (Wilks 1975a), and is presumably a matter of the DEPTH OF PROCESSING applied to communicative and cognitive operations (cf. S. Bobrow & Bower 1969; Craik & Lockhart 1972; Mistler-Lachman 1974): the degree to which an entity or configuration of entities is removed from the outward surface text. In general, conceptual connectivity is ‘deeper’ than sequential, and planning connectivity deeper than conceptual (cf. 1.2.12).
3.6 Concepts have FUZZY BOUNDARIES (Rosch 1973; Hobbs 1976.44;
Kintsch 1977a: 292ff.). They consist of a CONTROL CENTER in a KNOWLEDGE SPACE
around which are organized whatever more basic components the concept subsumes
(cf. Scragg 1976: 104). The center is the point where activation of the
concept’s content begins, but not necessarily where knowledge is concentrated
(cf. the ‘superatoms’ in Rieger 1975: 166f.). Though often assumed in
traditional philosophy (Hartmann 1963b: 104), the unity of a concept is probably
not guaranteed by strict identity of substance. Instead, unity emerges from the
unifying function of the concept in organizational procedures for managing
knowledge. The concept might be described as a block of INSTRUCTIONS for
cognitive and communicative operations (cf. Schmidt 1973: 86).
3.7 The constitution of concepts can be explored in regard to three processes: ACQUISITION, STORAGE, and UTILIZATION (Hörmann 1976: 485). A unified representation for all these processes would be desirable. If we assume that CONTINUITY, ACCESS, and ECONOMY are reasonable postulates for processing, the SEMANTIC NETWORK appears attractive (e.g. Quillian 1966, 1968; Collins & Quillian 1969, 1972; Carbonell Sr. 1970; Simmons & Bruce 1971; Simrnons & Slocum 1971; Rumelhart, Lindsay, & Norman 1972; Collins & Loftus 1975; Norman & Rumelhart 1975a; Shapiro 1975; Woods 1975; Fahlman 1977; Brachman 1978a, 1978b; Levesque & Mylopoulos 1978; Beaugrande 1979d, 1979e, 1979j; Findler [ed.] 1979).3 [3. The term ‘semantic network’ is somewhat misleading, as these nets do not actually analyse the meanings of concepts; hence, I prefer the term ‘conceptual-relational network’ (cf. Hendrix 1978: 1).] These various networks have a variety of uses, but they all consist of NODES and LINKS, Similar to the grammatical networks we saw in Chapter II. Whereas those networks were composed of GRAMMATICAL STATES, these are made up of KNOWLEDGE STATES.
3.8 If the network is a valid format for knowledge, it would follow that the total meaning of a concept is experienced by standing at its control center in a network and looking outward along all of its relational links in that knowledge space (Havens 1978: 7; cf. Quillian 1966, 1968; Collins & Quillian 1972: 314; Rieger 1975: 169; Fahlman 1977: 12; Brachman 1978a: 44). The interactions among surface words arise from precisely this connectivity: words in contexts (Kintsch 1974: 36), word associations (Deese 1962), the coherence of word senses (P. Hayes 1977; Rieger 1977b), and the preferences for utilizing some word senses over others in context (Wilks 1975b, 1978). Indeed, without this deeper connectivity, the selection and comprehension of words would be explosively unmanageable (see II. 1.3). Moreover, conceptual connectivity drastically constrains the utilization of syntactic options (Schank 1975b: 14) (cf. III.4.16ff.).
3.9 The human implications of networks are distinct from those of TAXONOMIES and LISTS. The usual decomposition proposed by linguists results in taxonomies, often with lists for many categories. In more recent research, lists of properties have been proposed for concepts (Collins & Quillian 1972: 313), and lists of propositions for the meaning of texts (Kintsch 1972, 1974; Meyer 1975; Frederiksen 1977; Turner & Greene 1977). For computer simulation of language processing, networks must be put in list format (cf. Simmons & Slocum 1971: 8; Riesbeck 1975: 103f.; Woods 1975: 51; a detailed presentation of the operations involved is given by Simmons & Chester 1979). But this requirement is an artefact of using serial processing (single operations in sequences), whereas human cognitive activities presumably function via parallel processing (multiple operations upon the same material simultaneously (Collins & Quillian 1972: 314). Scott Fahlman (1977) has shown how parallel processing can be simulated on serial computers.
3.10 The network is suited for an immense
variety of representational tasks (cf. Shapiro 1971; Woods 1978b:24),e.g.:
associative memory (Quillian 1966, 1968; J. Anderson & Bower 1973; Collins
& Loftus 1975); word disambiguation (P. Hayes 1977); dialogue understanding
(Grosz 1977); sensory apperception (Havens 1978); nominal compounds (Brachman
1978a); creativity processes (Beaugrande 1979c); and much more. This diversity
strongly recommends the network as a formalism for integrative and interactive
models of communication. There may even be purely formal benefits derivable from
notions in graph theory, such as ‘circuit,’ ‘separable and non-separable
graphs,’ and so on (cf. Chan 1969: 5ff.). The relevance of graph theory is not
obvious (J. Anderson 1976:147), but could lie in analogies and inspirations for
models of communication (cf. Taylor 1974 on abstracting; Dooley 1976 on
repartee).4 [.4 Taylor (1974) proposes that automatic summarizing could
be done with techniques like these: (1) removing the network nodes with the
densest linkage as probable topic nodes (cf. III.3.1 1.9; III.4.27); and (2)
assigning various strengths to the electric signal that each link type can
transmit, then doing a signal flow graph analysis. Hollan (1975) suggests that
graph theory offers the benefits of: (1) a substantial literature in abstract
mathematics (e.g. on traversal and search algorithms; cf. Ahlswede & Wegener
1979); and (2) the ease of implementing graph models as computer programs. I
might add that it would be worth considering whether the notion of ‘circuit’ and
‘separable/non-separable graphs’ could be helpful in modeling the coherence of
topic flow within textual worlds.]
3.11 The spatial organization of the network implies certain EPISTEMOLOGICAL tendencies (cf. Brachman 1979), such as the convictions that:
3.11.1 Entities of knowledge enter into multiple, interlocking, and configurational dependencies rather than sequences or lists.
3.11.2 An active point in a knowledge space can act as a control
center from which new impulses can connect on further material as processing
3.11.3 A knowledge space, such as in a textual world, has a characteristic TOPOGRAPHY that people can survey as a gestalt or walk through mentally in performing operations like integrating new knowledge, searching storage, deciding common references, and maintaining coherence. The more complex the topography, the longer the time needed to select the proper point for an addition or modification (cf. Kintsch & Keenan 1973).5 [3. However, this ratio would surely be affected by the expectedness of the new material as well (cf. Chapter IV).]
3.11.4 The notion of ‘semantic distance’ between concepts might have a graphic correlate: the total number of transition links for moving from one node to another (with caution: see Collins & Quillian 1972).
3.11.5 Cognitive processes work not on words or sentences alone, but more decisively on PATTERNS.
3.11.6 The notion of SPACES can be captured in diagrams in which routes of access are depicted. These spaces might function as CHUNKS, that is, integrated units that fit a great deal of content into ACTIVE STORAGE (cf. Miller 1956; Ortony 1978a) (cf. 111.3.16).
3.11.7 A knowledge space could appear in different PERSPECTIVES, depending on the LINK TYPES and UTILIZATIONS being pursued (cf. VI. 1.2).
3.11.8 The procedures for acquiring, storing, and utilizing knowledge and meaning can be represented as operations that build, organize, rearrange, develop, simplify, specify, or generalize conceptual-relational structures.
3.11.9 The dominant TOPIC or TOPICS of a textual world should be discoverable from the density of linkages around nodes in an interconnected space (cf. III.4.27).
3.11.10 The relationship of a text to alternative versions, such as a paraphrase, summary, or recall protocol, is not a match of words and phrases, but of underlying conceptual-relational patterns (cf. VII.3.3Iff.).
3.11.11 Entities of knowledge hardly every occur in actual human experience as isolated elements. Instead, for any entity, there are always potential contexts to impose order and efficient recognition on the encounter, especially via SPREADING ACTIVATION (cf. III.3.24). Should the context not be apparent, PROBLEM-SOLVING can be employed (cf. I.6.7).
3.12 The ACQUISITION of concepts has for many years been an object of psychological investigation, though with distinct and disquieting limitations (survey in Kintsch 1977a: ch. 7). The tasks posed were in general designed as classification of ‘stimulus’ items according to some arbitrary feature or aspect selected by the experimenter, such as size, colour, shape or numerousness. The test subject learns what aspect is relevant by trying out hypotheses (Bruner, Goodnow, & Austin 1956; Restle 1962). The most decisive learning takes place when the subject makes an error and must revise the hypothesis being applied (Bower & Trabasso 1964; Levine 1966).
3.13 Great care was expended on excluding relevant world-knowledge in such studies (Kintsch 1977a: 428). Yet the number of real situations in which people must learn arbitrary distinctions without contexts is surely small in comparison to integrative learning situations. Indeed, an encounter with entities that stand in no recoverable relation to what the experiencer already knows is likely to be profoundly disturbing. It follows that the formation of hypotheses normally draws on previously acquired concepts (Freuder 1978: 234). Even visual apperception depends crucially on what humans expect to see (Neisser 1967, 1976; Kuipers 1975; Minsky 1975; Mackworth 1976; Rumelhart 1977a; Havens 1978).
3.14 Concept acquisition might plausibly be accomplished as follows. A human would first encounter some entity and NOTICE it, i.e. expend processing resources on its presence and characteristics. Attempts would be made to determine what relations obtain between the entity and elements or previously stored knowledge. Let us assume here that it happens to be a new type of entity, so that a new entry must be made for it in storage. As the entity is encountered again or subjected to further mental contemplation, the need to integrate it into knowledge stores becomes more acute. The processor must eventually decide what aspects of the entity should be used to characterize it. The aspect of SALIENCE rests upon the intensity of intrusions upon sensory apperception (cf. Kintsch 1977a: 397ff.). FREQUENCY seems to affect processing also (Ekstrand, W. Wallace, & Underwood 1966), i.e. how often an entity is encountered or a characteristic is noticed. TYPICALITY would concern the number of instances that share some characteristic. Stimulus-response theories of learning might be salvaged in part if we postulate internal cognitive operations that focus discerningly on these different aspects, rather than simple ‘all-or-none’ learning (Hilgard 1951) that reacts mechanically to the environment. Taken in isolation, any single aspect might be irrelevant or misleading. For example, a bright, salient colon would be construed as useful for identifying a kind of tropical fruit, but not a kind of automobile (Freuder 1978).
3.15 Since there are staggering numbers of entities and occurrences
to conceptualize in order to talk about even that portion of the world that an
individual speaker knows about, humans must have powerful techniques for
imposing organization upon knowledge to be acquired. CONCEPTUALIZATION
(conversion of input knowledge into concepts) must entail extracting relevant
aspects. The raw input might leave some direct sensory “traces,’’ 6
[6. We return to the notion of ‘trace abstraction’ later (V1.3.16, VII.3.11,
VIII.2.48).] but the conceptualization of the input surely involves conversion
into a SYMBOLIC format which is not a sensory copy (Miller & Johnson- Laird
1976: Ch. 4; Kintsch 1977a: 234). This format is suitable for the
PATTERN-MATCHING that so many processes demand (I.6.6). In particular, patterns
should be tagged regarding what portions are crucial or probable for most
instances. I accordingly use tagging operators for three relative STRENGTHS of
conceptual content: (1) DETERMINATE aspects are essential to the identity of any
instance in order to belong to the concept (e.g. humans are mortal); (2) TYPICAL
aspects are frequent and useful, but not essential to the identity of an
instance for its concept (e.g. humans usually live in communities); and (3)
ACCIDENTAL aspects concern the inherently unstable or variable traits of
particular instances (e.g. some humans are blond).7 [7. After introducing this design
feature, I found out that Hollan (1975:154) had also proposed to ‘represent
defining and characteristic features within a digraph by labeling the
appropriate edges as defining or characteristic.’] These strengths are probably fuzzy, so
that a gradation (‘more or less determinate,’ etc.) should be postulated (Loftus
& Loftus 1976: 134). Still, people must agree reasonably well on this
gradation if they want to communicate efficiently and informatively.
3.16 The acquisition, storage, and utilization of knowledge require concerted interaction between EPISODIC MEMORY and CONCEPTUAL MEMORY (I prefer the latter term to ‘semantic memory’) (cf. Tuiving 1972; Ortony 1975; Abelson 1975: 306f.; Schank 1975a: 225f.; Kintsch 1977a: 283f., 1979b; Rumelhart 1977a: 222-36). Episodic memory contains storage of specific incidents in the person’s own experience (‘what happened to me’); conceptual memory contains systemized knowledge (‘what I know about the world at large and how it all fits together’). When the person encounters a configuration of input, relevant contents of episodic and/or conceptual memory are brought into ACTIVE STORAGE (III.3.5) to be matched. The dominance of the one or the other type of memory varies according to the familiarity of the input and the person’s store of experience and expertise. The acquisition of concepts as sketched in III.3.14 could be described as a gradual feeding of episodic memory into conceptual memory. Of course, many items are lost along the way, since relevant, important aspects must be filtered out from among incidental, idiosyncratic ones. If intense processing is not expended because input is familiar, frequent, unimportant, or uninformative, that input would probably decay before it enters the conceptual store. On the other hand, unfamiliar, rare, or highly informative input might be considered beyond the normal organization of the world and hence in opposition to the contents of the conceputal store. I argue in VII.3.29ff. that the interaction of prior storage (and its organization) with current input is substantially affected by the outcome of matching in both active and long-term storage.
3.17 The utilization of texts is a special case in the utilization of knowledge as outlined in III.3.16. The selection of specific lexical and grammatical options tends to remain largely episodic and not enter conceptual storage; the same is true of accidental relations inside the textual world (cf. VII.3.29.5). But these surface options still have a function in concept activation (III.3.5). By applying these activation strategies in the reverse direction, a person might succeed in reconstructing a good deal of the original surface text. This possibility makes it hard to determine experimentally how much seemingly accurate recall is in fact a reproduction rather than a reconstruction (cf. VII.3.Iff.; VII.3.16).
3.18 For a theory dealing with the tremendous volume of knowledge people can handle, ECONOMY of cognitive processing is a major consideration. Stated in extremely strong terms, cognitive economy stipulates that all knowledge is organized in storage as a unified, heavily interconnected, and non-redundant network; in a weaker version, some redundancy would be allowed (cf. Collins & Loftus 1975). Presumably, there could be a compromise: frequently used patterns would constitute fixed entries of stable knowledge; infrequently used ones would have to be assembled by drawing on various storage addresses. There would be a TRADE-OFF between redundant storage consuming much space but allowing rapid search and matching, and non-redundant storage consuming little space but demanding lengthy search to assemble any needed configuration. Here, compactness is balanced against access (cf. Kintsch 1977a: 290). The human mind seems to have vast storage and slow search, while the computer has rapid search but limited, expensive storage (Loftus & Loftus 1976: 128). Economy also suggests that the distinction between linguistic knowledge and world knowledge cannot be very great or clear-cut (cf. Oller 1972: 48; Goldman 1975:307; Riesbeck 1975:83; Rieger 1975: 158f., 1978: 44; Wilks 1977b: 390). The issue is rather one of COMPATIBLE MODES of knowledge, such as language versus vision (Minsky 1975; Jackendoff 1978; Waltz 1978). Language ABILITIES should also be analogous to other human abilities (Chomsky 1975:41ff.; Miller &Johnson- Laird 1976; Winograd 1976: 24; G. Lakoff 1977).
3.19 The INHERITANCE of content among knowledge entries is essential for economy (Falhman 1977; Hayes 1977; Brachman 1978a; Levesque & Mylopoulos 1978). In a hierarchy of classes, each SUBCLASS inherits some knowledge from its SUPERCLASS; and each INSTANCE inherits from its CLASS. For example, if we know that the superclass ‘mammals’ has the attribute ‘warm-blooded’, we would not need to store that knowledge again for the subclasses of people and cows; nor for specified groups like Pavlov’s dogs, Thorndike’s cats, and Skinner’s rats; nor for individuals like Clyde the piano-playing elephant and his master, Scott Fahlman. [Insider joke: an imaginary animal in Fahlman's dissertation asking: how do we know so easily such facts as that elephants do not play piano, never having thought about them?] Depending on the context, inheritance is more or less inclusive. Subclasses inherit from superclasses via SPECIFICATION: a statement in which the traits of the narrower subclass are set forth. For example, people share many traits with mammals, but have atypically inefficient mating practices. INSTANCES inherit all properties of a class unless there are signals to the contrary. Because Napoleon was a human being, be presumably had toes, though we have probably never read such a fact in history books.8 [8.This has been a long-standing example used by Waiter Kintsch.] When a context demands it, any trait can be CANCELLED by an explicit statement that inheritance does not apply to a subclass or instance, e.g.: unlike other elephants, Fahiman’s pet was not born, but cloned in a stupendous test tube (Fahlman 1977: 70). We assume in absence of cancellation that inheritance is valid: if Napoleon had not had toes, we would have many historical anecdotes about it (this would be a ‘lack-of knowledge inference’ [cf. Collins 1978; III.3.211).
3.20 Inheritance could also function via METACLASS inclusion. The classes are ‘meta-classes’ because they are brought together by conscious consideration of their respective natures; class/instance or superclass/subclass relationships are based on subsumption vs. specification. Original metaphoring often entails metaclass assignment, e.g. when Shakespeare’s Marullus addresses people a ‘biocks’, ‘stones’, and ‘worse than senseless things’ (sample (134) in V.5.4.1). The people are not, of course, included in those classes; there is at most some overlap of their characteristics with the characteristics that define those classes. The inheritance would thus function via that overlap. As a general principle, inheritance via metaclass inclusion requires more explicit signaling than that via class and superclass inclusion.
3.21 The degree of CERTAINTY with which inheritance occurs among classes and instances is variable. Communication entails frequent occasions when people must reason from incomplete knowledge never stored or acquired by direct experience or explicit statement. In the simplest cases, people can reason by ANALOGY of the unknown domain to a known one (cf. D. Bobrow & Norman 1975). For example, experience with Ohio drivers is likely to engender the expectation that any new instance one could encounter is probably incompetent. A variant would be NEGATIVE ANALOGY (Collins 1978): assuming different traits because the unknown domain contrasts with the known. For example, a highly skilled driver encountered in Ohio could be assumed to be a tourist from another state. Certainty also depends upon IMPORTANCE of a trait for a particular context. In industry, Ohio is known for rubber products; in sports, for football players; in politics, for obscure U. S. presidents; and in fashion, for its many pie-faced Miss Americas. Conversely, people make negative inferences by assuming that they ought to know about important traits if they did apply: here, LACK OF KNOWLEDGE is a significant means of making predictions. For example, it would be generally known if Ohio had high mountains; hence, we are safe in assuming that it does not, even if we have never been there (see also Collins 1978).
3.22 It is disputable whether people use the general classes and superclasses in routine processing of specific instances. If the general class is the actual storage address of the shared knowledge, people might mentally shift up the scale of generality during understanding tasks. In an experiment by Stephen Paimer (reported in Rumelhart 1977a: 234), people were presented with fragments that differed along this dimension, such as:
(32a) The boy noticed the flowers in the park.
(32b) The boy noticed the tulips in the park.
In subsequent recognition tests, people were far more inclined to mistakenly remember seeing the general class after seeing the specific subclass than vice-versa (compare de Villiers 1974).
3.23 The issue of class inclusion is a further demonstration of the TRADE-OFF between compactness of storage and length of access in search (III.3.18). Although non-redundant storing of all detailed classes under the headings of the most general classes would conserve storage space, the activities needed to access a relatively specific class or instance would have to travel much longer, more intricate pathways. Rosch, C. Simpson, and S. Miller (1976) suggest that people normally use a BASIC degree of generality as a compromise between extremely general superclasses and extremely specific subclasses. People would not want to process every object by running up the hierarchy to ‘object’, ‘thing’, ‘entity’, or the like: such computation would be explosive, and these general superclasses are too indeterminate to be of much use. At the other extreme, only experts could be expected to possess detailed knowledge of the most specific subclasses in a domain. Presumably, people would prefer the ‘basic’ degree of generality and would have recourse to other degrees according to the demands of the context for DIFFERENTIATION (cf. IV.2.6.5). Here also, there would be a THRESHOLD OF TERMINATION where processing is sufficiently general or detailed for current needs. In the ‘rocket’ experiments I discuss in following sections (V1.3; VII.3), our test persons often did not specify a ‘V-2 rocket’, but they all used ‘rocket’ as opposed to the more general classes of aircraft’ or ‘flying object’.
3.24 I cited the UTILIZATION OF CONCEPTS as a third issue besides acquisition and storage (III.3.7). I suggested that concepts are ACTIVATED in the mind and MAPPED onto EXPRESSIONS in text production and mapped back again in text reception (cf. III.3.5). Due to SPREADING ACTIVATION, more material becomes active than just the immediate content covered by the expressions of the text (cf. 1.6.4) (Collins & Loftus 1975; Ortony 1978a). The original point from which spreading proceeds would be a special case of the CONTROL CENTERS which I consider essential in text processing (cf. 11.2.9; III.3.6; VI.3.5; VII. I.Sff.; VII.3.34). The extent of spreading would be regulated by the THRESHOLD OF TERMINATION that I have also postulated for many processes (cf. I.3.4.3; I.6.1; I.6.4; III.3.3; III.3.23; IV.1.6; VII.2.7; VII.2.10). The controls upon spreading need not be conscious (cf. J. Anderson & Bower 1973; Rieger 1974; M. Posner & Snyder 1975). The spreading would normally proceed from several points at once, so that INTERSECTIONS of activated paths support coherence and engender predictions about how the concepts in a text world fit together (Rieger 1974, 1975; cf. ‘coincidence detection’ in Woods 1978b). Certain types of paths are presumably suited as spreading routes: (1) TYPICAL and DETERMINATE links in CONCEPTUAL memory (cf. III.3.15); and (2) strong associative links of personal experience in EPISODIC memory (cf. III.3.16). However, the activity of daydreaming shows that spreading can on occasion follow paths whose motivation and directionality is not readily evident.
3.25 In an experiment conducted with students of various ages in Gainesville, Florida, I attempted to study some types of activation for familiar concepts.9 [9 I am most indebted to Carolyn Cook, Reba Dean, Gail Kanipe, Mamie Kelsey, Mary Morgan, and Mary Sharp of the Gainesville Public Schools for their participation in running these tests.] We simply asked our test subjects, ranging from fourth grade to tenth grade, to name the ‘typical parts of a house, in any order.’ I observed a small set of strategies at work across most of the population, indicative of a corresponding set of SEARCH TYPES: ‘part-of,’ ‘substance- of,’ ‘locationally-proximate-to,’ and ‘containment-of’ searches. The ‘part-of’search recovered a listing of major rooms (‘living room’ ‘kitchen’, etc.), or of structural Components (‘roof’, ‘floor’, ‘walls etc.). The ‘substance-of’ search netted building materials (‘nails’ ‘bricks’, ‘paint’ ‘glue’ etc.). The ‘locationally-proximate-to’, search brought together itemss that a person could notice by standing at a given location inside a house. Unlike the adults interviewed by Linde and Laboy (1975), our subjects did not often perform a continuous mental walk-through of a floor plan, perhaps because in our tests, they were not asked to describe their own houses. In one group, 15 out of 28 subjects began with the ‘front door’, and only 5 made it to the’back door’ . The tendency was rather to switch without mediation from one location to another and begin a new listing of nearby objects.
3.26 The degree of consistency and organization varied according to the children’s age. The youngest children did not choose and pursue a given search type with the same concentration as the older ones. Whereas older children preferred a constructivist outlook that stressed parts and substances, the young children took an episodic approach by regarding their own personal homes as typical. They had a corresponding inclination toward ‘containment-of’ searches assembling many objects that houses might well not encompass. They stipulated that houses should have ‘three telephones’, a ‘walnut table’ and a ‘glass what-not shelf.’ They mentioned domestic animals (‘bird’, ‘fish’, ‘kittens’ ‘mouse’,), items of food (‘cake’, ‘ham’,’coke’, ’tea’), and of course ‘people’ — all considered typical parts of a house. Evidently, even familiar concepts have fuzzy boundaries (cf. III.3.6); indeed, familiar ones might have especially fuzzy boundaries because of the richness of personal experience with them (Peter Hartmann, personal communication). The processes of acquiring and stabilizing a concept seem to evolve over considerable time spans, e.g. between the ages of fourth to tenth grade. And the concept looks different within its knowledge space depending upon the PERSPECTIVE of the current utilization (cf. III.3.2; III.3.11.7; VI.I.2).
3.27 Early attempts to systemize the notion of conceptual memory often appealed to ‘semantic memory’ (cf. Collins & Quillian 1969). The main relation in these models was either superclass/sub,class (the link type of ‘specification-of’ in III.4.7.25) or class/ instance (the link type ‘instance-of’ in III.4.7.24). It was reasoned that the verification of (33a) would take longer than that of (33b) because the processor would have to run through at least one more class layer.
(33a) A chicken is an animal.
(33b) A chicken is a bird.
But experiments did not verify this prediction very consistently (Collins & Quillian 1972). Smith, Shoben, and Rips (1974) proposed to account for the distance between concepts in terms of FEATURE OVERLAP (e.g. how many features of a ‘bird’ a ‘chicken’ has). High overlap would allow rapid verification of statements about class membership; a low overlap would have the opposite effect (e.g. a ‘chicken’is not judged to be a ‘bird’ as quickly as a ‘robin’ because the former cannot fly and the latter can). The subclass having the highest overlap with the superclass would be the PROTOTYPE of the latter (cf. Rosch & Mervis 1975; Rosch 1977; V.3. 10).
3.28 Principled objections can be raised against such models of human memory. The hierarchical approach is unduly restricted to the relation of class inclusion (cf. Kintsch 1979b). There are doubtless many other relations that hold stored knowledge together (cf. III.4.7ff.). Moreover, in domains less structured than the classification of animals, it might not be clear if a subclass belongs to one or many superclasses, or to no obvious ones at all. A subclass might efficiently be treated via ANALOGY to a superclass that does not in fact include it (e.g. treating a ‘whale’ as an odd kind of ‘fish). The featural approach is saddled with all of the problems for such theories that I raised in III.2.2. And both the hierarchical and featural approaches leave human memory looking rather static. Perhaps we could reinterpret them both in terms of SPREADING ACTIVATION. In the hierarchical aspect, the intersections of pathways spreading out from the control center of two concepts (e.g. ‘chicken’ ‘bird) occur on ‘specification-of’ links. In the featural approach, the intersections occur on such links as ‘attribute-of,’ ‘form-of’, ‘part-of,’ ‘agent-of,’ and so on. The most rapid and certain judgments about the truth of statements like (33a/ b) would arise if these links are DETERMINATE; TYPICAL links would be next best, and ACCIDENTAL links would work the least well.
3.29 In retrospect, the distrust of some researchers (e.g., Schank 1975a; Kintsch 1979b) regarding ‘semantic memory’ certainly seems justified. A more flexible and inclusive model of CONCEPTUAL memory must deal with many more types of relations and with the effects of contexts of utilization upon stored knowledge configurations. In absence of such a model, the differences in times needed to verify the content of isolated sentences may not be telling us much about the organization of memory (Kintsch 1979b). I would submit that the study of textual processing might be a more productive means of gaining insights into knowledge and memory in realistic human situations.
4. BUILDING THE TEXT-WORLD MODEL
4.1 A TEXTUAL WORLD is the cognitive correlate in the mind of a text
user for the configuration of concepts activated in regard to a text (I.6.1).
Although I occasionally use this term for the configuration of concepts and
relations, which I have designed, I am in fact only dealing with TEXT- WORLD
MODELS that are idealizations of the actual cognitive entities involved. My
models include at least some materials not explicitly signaled in the text as
such; but the textual worlds of participants in communication probably include
far more. The text functions via the activation of concepts and relations
signaled by expressions (III.3.5). Spreading activation, inferencing, and
updating perform substantial modifications upon this basic material (1.6.4). The
interaction of text-presented knowledge and previously stored knowledge can be
depicted in terms of PROCEDURAL ATTACHMENT: the currently active knowledge
stores specify and control what is done to build a textual world, so that
operations are reasonably efficient (D. Bobrow & Winograd 1977). However, if
the text is informative in the sense of I.4.11.7, the textual world will not be
a perfect match for stored knowledge. In this section, I explore more
conventional aspects of procedural attachment in model-building, and look into
questions of informativity in Chapter IV.
4.2 If procedural attachment is to function efficiently for a wide
range of occurrences, its categories cannot be unduly diffuse or detailed. I
shall propose a TYPOLOGY of concepts and relations, whose task is not to capture
the exhaustive meaning of textual occurrences, but only to constrain meanings to
the point where the RESIDUE can be picked up as far as the language user desires
to do so (cf. 1.5.6). Obviously, my typology could hardly contain rare,
idiosyncratic concepts like Leskov’s (1961) ‘left-handed Tula craftsman’ or
relations like Charniak’s (1975a: 21) ‘up-to-the-third-floor-of’ ‘which only
applies when the action takes the object up to the third floor of a building.’
My typology will be reasonably small, and designed along comparable lines to
that for sequencing: the relational labels for the network links will
characterize the concepts in the nodes. Further detail can be obtainable by
combining types (cf. III.4.4).10
[10. Upon occasion, I provide labels with arrows for secondary
concepts at both ends of a link. As yet, I have no hypothesis about the
directionality of control flow in such cases.]
4.3 There are several domains that should be covered by such a semantics, notably: (1) the structures of events, actions, objects, and situations (e.g. attributes, states, times, locations, parts, substances, etc.); (2) general logical notions like class inclusion, quantity, modality, causality, etc.; (3) human experience (apperception, emotion, cognition, etc.); and (4) contingencies of language communication via a symbolic intersystem (e.g. significance, value, equivalence, opposition, etc.). I make no claims that my typology is definitive or exhaustive. It has been sufficient for the text-world models of numerous samples I have studied. And by means of type combining, it is able to handle nearly all of the one hundred primitives developed by Yorick Wilks (1977a) over a ten-year period. Those familiar with Roget’s famous Thesaurus may perceive some resemblance to that classification also. Nonetheless, there are concepts whose residual content (III.2.2.4) does not-and indeed, not — be captured by such a typology. Residual content is a matter of what is stored in the conceptual LEXICON. My typology merely constrains concepts to the extent needed for intelligent utilization (cf. 1.5.6).
I. Primary Concepts
II. Secondary concepts
A. Defining events, actions, objects, and situations
B. Defining human experience
C. Defining class inclusion
D. Defining relations
E. Defining contingencies of symbolic communication
It is divided into PRIMARY and SECONDARY concepts. The PRIMARY concepts include: (1) OBJECTS (conceptual entities with a stable constitution or identity); (2) SITUATIONS (configurations of objects present and their current states); (3) EVENTS (occurrences that change a situation or a state within a situation); and (4) ACTIONS (events intentionally brought about by an agent).11 [11. ‘One might argue that STATE and AGENT should also be included as primary concepts; but these two types seem to be derivative from objects/ situations and actions, respectively. I note in VI.3.13 that states do not seem to receive processing focus as much as events.] These primary concepts are the usual CONTROL CENTERS for building textual worlds, i.e., the points of orientation from which a processor sets up the relationships to the secondary concepts. For example, spreading activation tends to work outward from the primary concepts (unless one had an odd task like ‘think of all the yellow objects you know’, etc.). To keep my models less crowded, I do not label the primary concepts (though it might be expedient to do so if I had a decent graphics program), but only the relations connecting them to secondary ones. The detailedness of secondary concepts utilized in text processing depends in part on the concept names (expressions) selected by the text producer; and in part upon the demands of the context. By combining concepts, we can derive many more specific ones: ‘quantity of substance’ could yield ‘weight’or ‘size’, ‘quantity of motion’ could yield ‘speed’, ‘initiation of cognition’ could yield ‘ideation’, and so forth (some highly complex combinations are found in VI.4.33).11a [11a. I use the division sign ‘.-’ to combine labels in the diagrams.] As I noted in III.3.3, concepts can often be restated as propositions on a plane of greater detail.
4.5 It might be argued that concepts are themselves not unitary, but
possess internal structuration or limits. 12 [12. I noted in III.3.3 that all
concepts may in principle be decomposable — possibly, Frederiksen’s (1977)
application of ‘rank-shifting’ would be useful here, provided that the ‘ranks’
are treated as relative, not absolute.] Leonard Talmy (1978) cites several such
considerations: (1) PLEXITY, being the capacity for having discernible component
parts (e.g. uniplex vs. multiplex); (2) BOUNDEDNESS, being the presence or
absence of defined liniits; (3) DIVIDEDNESS, being a lack of internal
continuity; and (4) DISTRIBUTION, being the pattern of matter arranged in space
or of action arranged in time. Michael Halliday (1967a) calls attention to such
distinctions as ‘action vs. ascription’ and ‘directed action vs. non-directed
action.’ Both Talmy and Halliday appear to suppose that these issues are
grammatical in nature. I would view them rather as an interaction between the
grammatical, lexical, and conceptual aspects of language. The question then
concerns the extent to which grammatical and lexical options create PREFERENCES
for activating certain aspects of a conceptual-relational space (cf. Fillmore
1977). One can distinguish between expressions such as ‘he sighed’
(uniplex) and ‘he kept
sighing’(multiplex), or ‘the prisoners marched’ (non-directed action) and ‘they
marched the prisoners’ (directed); but surely the structures appealed to by
these terms belong to the events, not just to surface grammar. The same might be
said for such categories as ‘count nouns’ (e.g., ‘bottles) and ‘mass nouns’
(e.g. ‘water’) (cf. Leech & Svartvik 1975: 49ff.)13 [13.
Christian Rohrer (1979) suggests a correlation in French between count nouns and
the simple past, and mass nouns and the imperfect. The boundedness and
dividedness of events appear to affect the perspectives adopted on the objects
4.6 The primitives developed by Roger Schank, on the other hand, show less detail than the concept types I am using. Schank focuses on human ACTIONS, and his set of ‘primitive acts,’ though more detailed than in his early work since 1970, are still very global, such as ‘physical transportation,’ ‘mental transportation,’ and the like (cf. Schank et al. 1975; Schank & Abelson 1977). The focus on actions is justified by their status as OCCURRENCES ON MULTIPLE LEVELS: they are control centers which often have linkage to numerous secondary concepts; they map onto grammatical nodes which are control centers on the level of sequential connectivity (cf. III.4.14); they appear prominently in chainings of cause, reason, enablement, and purpose; they update the situation in a textual world; they update their agents’ outlook on the situation in the textual world; they correspond to steps in a plan and are relevant to a goal; and they direct the flow of a narrative. If the maintenance of connectivity and continuity is as crucial for cognitive operations as I claim, it follows that more processing resources are required for actions than for any other concept type. Schank’s treatment of other concepts is in fact much closer to the surface, as the appearance of entries like ‘cheese’, ‘mushrooms’, ‘saliva’, ‘money’, ‘fist’, ‘bullet’, or ‘poison’ in his networks (Schank 1975c: 49-66) attests.
4.7 My typology of relations is designed especially for labelling connections between the secondary concepts and the primary concepts (cf. Table 1; III.4.4). The traversal of any link in the direction indicated by the arrow will thus arrive at a node characterized by the link label. This DIRECTIONALITY is intended to suggest the flow of control in a manner comparable to the operations of the AUGMENTED TRANSITION NETWORKS described in II.2.12ff. (cf. III.4.16). The relation types are as follows (the mnemonic labels are the first two letters of the word except where avoidance of duplication leads to using the first and third letters):
4.7.1 STATE-OF [st] signals the current condition of some entity, rather than its characteristic one (e.g., ‘sea-stormy).14 [14. It follows that there would be no determinate state-of linkages.]
4.7.2 AGENT-OF [ag] is the force-possessing entity that performs an action and brings about a change in the situation that would not have occurred otherwise (cf von Wright 1967) (e.g. ‘general-attack).
4.7.3 AFFECTED ENTITY [ae] is that entity whose situation is changed by an event or action in which it figures neither as agent or instrument. In demonstration sentences, ‘Mary’ usually gets stuck with this role, e.g. ‘John shot Mary’ (Schank 1975c: 52).
4.7.4 RELATION-OF [rl] subsumes a range of detailed relations not
worth assigning to a separate link, e.g. ’father-of’, ‘husband-of’, ‘boss-of’
etc.15 [15. This
label was overused in many early networks, at least in practice (cf. survey in
Brachman 1978a). I have not had to use it much myself so far.]
4.7.5 ATTRIBUTE-OF [at] signals the characteristic or inherent condition of some entity (e.g. ‘sea-saline).
4.7.6. LOCATION-OF [lo] links an entity with concepts of spatial position, and is often tagged with prepositions (e.g. ‘at’, ‘in’). Entry (e.g. ‘into’, ‘onto’), exit (‘out of, ‘off of’), and proximity operators (‘next to’, ‘near’, ‘above’) are very common for this link (cf. III.4.12; VII.3.22).
4.7.7 TIME-OF [ti] links all specifications of time, such as absolute
(e.g. dates) or relative (‘soon’, ‘then), often with proximity (‘before’,
4.7.8 MOTION-OF [rno] is used when entities change their location, whether or not the places of origin and destination are given (e.g., ‘run’, ‘rise’). Entry (‘arrive’) and exit (‘leave’) are common operators.
4.7.9 INSTRUMENT-OF [it] applies when a non-intentional object provides the means for some event or action (e.g. ‘fuel-propulsion’, ‘scissors- cut’). Instruments thus differ from agents in lacking intention; and from causes and enablements in that instruments are objects, while causes and enablements are events.
4.7.10 FORM-OF [fo] connects entities to concepts of form, shape, and
contour (e.g. ‘block-lumpy’).
4.7.11 PART-OF [pa] connects an entity with a component or segment (e.g. ‘automobile-engine’, or ‘Fred-Fred’s arm’) (cf. Hayes 1977).
4.7.12 SUBSTANCE-OF [su] signals relations between an entity and the materials of which it is composed (cf. ‘source’ and ‘stuff’ in Wilks 1977a) (e.g. ‘automobile-metal’ or ‘Fred-tissue’).
4.7.13 CONTAINMENT-OF [co] signals relations between entities of which one contains the other (cf. Wilks 1977a) (e.g. ‘automobile-Fred’, ‘Fred-six pack of Budweiser’).
4.7.14 CAUSE-OF [ca]: an event E1 is the cause of an event E2 if El creates the necessary conditions for E2 (e.g., ‘injury-pain’, ‘theft-loss).
4.7.15 ENABLEMENT-OF [en]: an event E1 is the enablement
of an event E2 if E1 creates the sufficient, but not
necessary conditions for E2 (e.g. ‘negligence-injury’, ‘owner’s
4.7.16 REASON-OF [re]: an event E1 is the reason for an event E2 if the agent or initiator of E2 is reacting rationally to E1 (on cause vs. reason, cf. Rieger 1975; Schank 1975a; Wilks 1977c) (e.g. ‘injury-anxiety’, ‘luck- happiness).
4.7.17 PURPOSE-OF [pu]: an event E2 is the purpose of
E1 if the agent of E1 has a plan in which E1 is
expected to enable E2 (Cf. ‘goal’ and’purpose’ in Wilks 1977a) (e.g.
‘warning-escape’, ‘theft-being rich). Whereas cause, enablement, and reason look
forward in time from an earlier event to a later one, purpose looks backward
from the later to the earlier event (Beaugrande B. N. Colby 1979; Beaugrande
& G. Miller 1980).
4.7.18 APPERCEPTION-OF [ap] relates sensorially endowed entities with the operations whereby knowledge is integrated directly via sensory organs (e.g. ‘scientist-observe’). Simulations can fall under this heading as well (e.g. ‘radar-track).
4.7.19 COGNITION-OF [cg] links sensorially endowed entities with cognitive operations (e.g. ‘Einstein-imagine’, ‘Roger Schank-think’). Simulation would be possible here also (e.g. ‘Shrdlu the robot-comput’).
4.7.20 EMOTION-OF [cm] links sensorially endowed entities with experientially or evaluatively non-neutral states of excitation or depression (e.g. ‘Fred-ticked off, ‘Mary-enraptured’). Simulation has been undertaken here also, as in K. Colby & Parkinson’s (1974) paranoid computer PARRY.
4.7.21 VOLITION-OF [vo] links sensorially endowed entities with activities of will or desire (e.g. ‘population-want’, ‘Jimmy Carter-vainly hope’).
4.7.22 COMMUNICATION-OF [cm] links sensorially endowed entities with
activities of expressing or transmitting cognitions (e.g. ‘Fred-say’,
4.7.23 POSSESSION-OF [po] signals relations where a sensorially endowed entity is believed to own any entity (e.g.’Fred-have). Initiation (e.g., give), entry (e.g. ‘buy), termination (e.g. ‘take away), and exit (e.g.’sell’) are all common operators.
4.7.24 INSTANCE-OF [in] obtains between a class and one of its
members (e.g. ‘automobiles-Fred’s clunker). A member inherits all of the traits
of the class that are not cancelled (cf. III.3.19).
4.7.25 SPECIFICATION-OF [sp] obtains between the superclass and its subclass (e.g. ‘automobiles-convertibles). Inheritance is restricted according to the distinguishing traits of the classes (cf. III.3.19).
4.7.26 QUANTITY-OF [qu] labels all links between an entity and a concept of number, extent, scale, or measurement (e.g. in the multiple series ‘Clyde-weight-kilograms-3000’ in Fahiman 1977: 102). One might want to subdivide quantity into groupings like ‘measurements’ (e.g.’kilograms’) and ‘numericals’ (e.g. ‘3,000). Because empirical tests do show some differences in processing such groupings (cf. VII.3.29.5), I shall introduce such a scheme in the future. But I do not mark logical quantification (cf III. 1.3), assuming existence as a default (III. 1.5), and set inclusion as relevant only if enumerated (cf. III.1.6).
4.7.27 MODALITY-OF [md] labels relations between an entity and a concept of modality (probability, possibility, etc.) (e.g. ‘departure- impossible). Modality subsumes negation, and is often conveyed via modal auxiliary verbs (e.g. ‘should’. ‘can’t’ ‘mus’t).
4.7.28 SIGNIFICANCE-OF [si] applies when two concepts are expressly
stated to stand in a symbolic relation (e.g. ‘gesture-mean).16 [16. This label would be
frequent in ‘metalanguage’ used to assign or explain the meaning of symbolic
4.7.29 VALUE-OF [val applies to relations between a concept and some
assignment of value (e.g. ‘diamond-precious). Value relations can also be
comparative (e.g. ‘brand X-better than-brand Y).
[eq] applies to relations of equality, similarity, correspondence, and so on
(e.g. ‘high-lofty’,’dark-sombe’). These relations, which are crucial to the
internal organization of knowledge in texts, frequently involve proximity (e.g.
4.7.31 OPPOSED-TO [op] is the converse relation to equivalence, and also figures strongly in knowledge organization (e.g. ‘high-low’, ‘dark-light’).
4.7.32 CO-REFERENTIAL-WITH [cr] is the relation between concepts whose inherent content is different, but which happen to be used to refer to the same entity in a textual world (e.g. ‘morning star-evening star). CO- reference often entails the use of pro-forms (cf. V.4).
4.7.33 RECURRENCE-OF [rc] is there relation between two occurrences
of the same concept in a textual world, but without necessarily having reference
to the same entity (as in ‘it fell to the earth near mounds of earth).17
[17. It is not always decidable whether or not recurrence converges
with co-reference. Where I feel that such convergence is given, I map the
recurrences onto the same node (e.g. ‘rocket’ in sample (35)). But the
positioning of the recurrences may have psychological consequences that should
be explored. It may prove expedient to subdivide recurrence and co-reference
into a more detailed typology, such as that outlined in Chapter V.]
4.8 Many of these relations are familiar from various attempts to explicate the uses of grammatical structures in terms of conceptual ones. The verb-centered grammars such as the so-called ‘valence theory’ (cf. Tesnière 1959; Brinkmann 1962; Erben 1964; Helbig [ed.] 1971) sought to classify verbs according to the number of elements that were conventionally dependent on them in a sentence. All of these attempts failed to the extent that the grammatical environment of verbs is in part a matter of the conceptual environments of the concepts which verbs can be used to activate. A listing of verbs with ‘valences’ of 1, 2, 3, etc. (i.e. according to connected surface elements) fails to capture these variable and diverse factors.
4.9 In the grammars of some languages, the roles of elements respective to the verb are marked by surface inflections often termed ‘cases,’ for example, in Latin. Charles Fillmore’s (1968) ‘case grammar’ (proposed in order to quietly introduce some aspects of meaning into transformational grammar) was inspired by this tradition. he naturally tended to focus on the cases that were explicit in languages like Latin. Fillmore’s framework of orientation encouraged the assumption that cases are building blocks of abstract sentences, rather than of conceptual dependencies. In recent work (Fillmore 1977), be has migrated away from his original position by taking the structure of cognitive ‘scenes’ into account.
4.10 The notion of ‘case’ has had a profound effect on theory of language. Cases are now generally viewed as conceptual, not grammatical, with a range of compromises and intermediary positions (compare and contrast Chafe 1970; Bruce 1974; Kintsch 1974; Charniak 1975a; Grimes 1975; Nilsen & Nilsen 1975; Schank et al. 1975; Longaere 1976; Minsky 1977; Turner & Greene 1977). Conceptual cases must be MAPPED onto grammatical structures via relevant decisions and controls. Some constraints apply to structures that can be connected to individual verbs, but constraints on situations, events, and actions are more basic (cf. Goldman 1975: 317; Grimes 1975: 52; Schank 1975c: 82). The PREFERENCES for selecting a certain verb arise from the preferences regarding how to connect concepts and relations (cf. Wilks 1978). Although these preference types are not symmetrical, they impose major controls on use of verbs and verb complements (cf. Fillmore 1977).
4.11 There is no clear justification for insisting on the sentence as the framework of conceptual ‘cases.’ Language processing ought to be more concerned with the similarities between (34a) and (34b) than with the sentence boundaries (suggested by Robert F. Simmons, personal communication):
(34a) There was a knock at the door. It was John. be was using his cane.
(34b) John knocked at the door with his cane.
The ‘conceptual dependency’ understander at Yale, for, example, would pick up the relations for (34a) just as if it had been presented with (34b) (Roger Schank, personal communication). Efficient processing obviously needs to extend its predictions about the organization of events and situations beyond the boundaries of single sentences; otherwise, the production and reception of texts would lack continuity. Indeed, Bransford & Franks (1971) found that test persons who saw chopped-up sentences like (34a) were later quite confident in believing they had seen the fluent versions like (34b).
I complete my set of link labels with the OPERATORS which specify the status of
relations as needed. These operators are concerned with: (1) beginnings and
endings; (2) fuzziness; (3) counter-factuality; and (4) strength of linkage. To
make the operators visually distinctive in the diagrams, I use Greek letters,
for example ‘π
would be ‘proximity of time, ‘ca + ε
+ lo’ would be ‘cause of entry into location,’ and so on. The operators
4.12.1 The INITIATION operator [ι] signals that the relation is just being brought about by some applied force or agency (e.g. ‘takeoff is an initiated motion, while ‘fly’ is not).
4.12.2 The TERMINATION oprerator [†] signals that the relation is ended by some force or agency (e.g. ‘land’as compared to ‘descend).
4.12.3 The ENTRY operator (ε] signals that an entity is entering into a relation rather than bringing it about (e.g. ‘sicken’ as entry into state, in comparison to ‘sick’ as state).
4.12.4 The EXIT operator [χ] signals that an entity is leaving a relation (e.g. ‘recover from illness’ as exit from state, as opposed to ‘health’ as the new state).
4.12.5 The PROXIMITY operator [ρ] signals somer mediation or distance in a relation (e.g. ‘nearby’ as proximity of location, ‘sooner proximity of time).
4.12.6 The PROJECTION operator [p) signals that a relation is possible and under consideration, but not yet realized in the textual world (e.g. ‘if he arrives’ as projected entry into location).
4.12.7 The DETERMINATENESS operator [δ] is used in world- knowledge for relations required by the identity of a concept (III.3. 15) (e.g. ‘house-walls’ as a determinate ‘part-of’ link).
The TYPICALNESS operator [τ]
applies to world-knowledge relations that are usual, but not obligatory, among
representatives of a concept (e.g.’house-cement’ as a typical ‘substance-of’
link). The operators for determinateness and typicality are used only in the
configuration I term the ‘world-knowledge correlate’ (cf. III.4.36), as they are
not aspects of the text-world itself (unless we had a text-world which was a
whole microcosm, e.g., in an extensive novel .
4.13 Here also, one could argue in favor of additional classifications, such as a ‘cancellation operator’ for links that cease to obtain when a textual world is UPDATED by events and actions. 19 [19. 0n ‘cancel links’ see III.3.19; VI.3.4. On updating, see I.6.4.] However, this operator would make sense only if one wished to take the status of the textual world phase by phase. Eventually, all links would be cancelled by updating, except perhaps conventional stabilizations like ‘they lived happily ever after’. Also, one might want to introduce operators to signal the issues raised by Halliday (1967a) and Talmy (1978) (see III.4.5).
4.14 In chapter II, I demonstrated how a grammatical dependency network could be constructed for a sentence-length fragment. I stressed that such a network can serve as a useful indicator of the CONTROL CENTERS for a given stretch of text (cf. II.2.9). The preference strategy would be to postulate that the heads of grammatical macro-states (nouns in noun phrases or prepositional phrases, verbs in verb phrases or participial phrases) are expressions of primary concepts (cf. III.4.4). The operational consequences of this strategy might work at least two ways. In a serial procedure, an understander would run the syntactic analysis forward through a phrase until the head is found. then the conceptual analysis backtracks and incorporates elements into a semantic network (e.g., if the syntactic analysis found a noun head, it could backtrack and pick up the adjectives as attributes or whatever). This is essentially the approach of Rusty Bobrow’s RUS system (R. Bobrow 1978). In a parallel procedure, an understander runs various kinds of analysis simultaneously and combines all structure-building operations that have the same configurations as an outcome (e.g., a hypothesis about a noun head with adjectives can be tested along with a hypothesis about an object with attributes). This is essentially the approach of William Woods’ cascading network system (Woods & Brachman 1978 b; cf. II. 2.13). In both procedures, the sharing of structural configurations is an important contributor to accuracy and efficiency, especially with regard to refining probabilities. Woods’s system, however, is better equipped to deal with missing or indistinct elements, since disconnectivity in one cascading network could be overcome by the connectivity of the others (see Woods, Brown, Bruce, Cook, Klovstad, Makhoul, Nash-Webber, Schwartz, Wolf, & Zue 1976).
It is conceivable that under certain conditions humans might BYPASS surface
syntax during text comprehension. This question has not been pursued in
linguistics very far; a sentence linguist who suggested such a thing would have
risked being banned in Boston as a threat to public decency. Yet the ‘key word’
systems which pick up particular words here and there (e.g. Weizenbaum 1966) and
the ‘conceptual parsers’ such as Riesbeck (1 974) did in fact make only limited
use of surface syntax. Perhaps humans perform something more like ‘fuzzy
parsing’ (Burton 1976), that is, classifying word categories, inflections, and
grammatical dependencies only as far as is needed to uncover the conceptual/
relational constitution of the textual world. When the hypotheses about the
text-world structure are numerous or evenly matched, syntactic parsing would be
more thorough — a question of degree of informativity (cf. IV.1.10). In one
respect at least, syntax is always relevant to text processes: it determines the
temporal order of occurrences. That factor may be peripheral in an abstract
theory of well-formed sentences, but it is central for a realistic theory of
4.16 There ought to be preferences not only between phrase heads and primary concepts, but also between grammatical dependencies and conceptual links. Possibly, a network could be built up by AUGMENTING the TRANSITIONS between nodes with a combined grammatical and conceptual search (cf. II.2.12ff.; III.4.7). The results of the one aspect of the search could thus be applied to aid the other (cf. Burton 1976; Woods 1978c) — bearing in mind, however, that grammatical units and structures are not always of the same size as conceptual ones. All the detailed cues provided by the actual material at hand would be handled by augmenting transitions still further. The following are some reasonable (though certainly not verified) candidates for preferential correlations between the grammatical and the conceptual level (the three dots indicate that other hypotheses would be tested if these fail):
4.16.1 For ‘subject-to-yerb,’ prefer ‘agent-to-action,’ ‘object-to-state’...
4.16.2 For ‘verb-to-object,’ prefer ‘action-to-affected entity’...
4.16.3 For ‘verb-to-indirect object,’ prefer ‘action-to-affected entity- entering into-state’, ‘action-to-affected entity-entering into-possession’...
4.16.4 For’verb-to-modifier,’prefer’state-to-state,’ ‘state-to-attribute,’ ‘,state-to-location’...
4.16.5 For ‘verb-to-auxiliary,’ prefer ‘action-to-time,’ ‘action-to-modality’...
4.16.6 For ‘verb-to-dummy,’ withhold predictions and continue.
4.16.7 For ‘modifier-to-head,’ prefer: (1) for adjectives, ‘state-to-object,’ ‘attribute-to-object,’ ‘attribute-to-agent,’ ‘attribute-to-affected entity’... (2) for adverbials with verb heads, ‘attribute-to-action’ (cf. ‘manner’), ‘location-to-action,’ ‘time-to-action,’ ‘instrument-to-action’...
4.16.8 For ‘modifier-to-modifier,’ prefer ‘attribute-to-attribute,’ ‘attribute-to-location’...
4.16.9 For ‘determiner-to-head,’ prefer ‘quantity-to-object’ or test hypotheses about knownness and definiteness (cf. V.3).
4.16.10 For ‘component-to-component,’ prefer ‘possessor-to-object,’ ‘superclass-to-subclass,’ ‘class-to-instance,’ ‘object-to-part,’ ‘substance- to-object,’ ‘form-to-object’...
4.16.11 For ‘conjunction,’ ‘disjunction,’ and ‘contrajunction,’ try to reapply to the second of the joined configurations those hypotheses that were successful for the first.
4.16.12 For ‘subordination’ , prefer ‘cause-of,’ ‘reason-of,’
‘enablement-of’, ‘proximate-in-time-to"... (cf. V.7.6ff.).
4.17 The real ordering of such preferences will have to be discovered by empirical investigation. For the time being, I only suggest some plausible candidates. The preferences would be a major support in the PROBLEM- SOLVING activities of maintaining both sequential and conceptual connectivity: in essence, problems in the one subsystem are solved via hypotheses drawn out of the other. The immediate application of the preferences to actual texts would require considerable PROCEDURAL ATTACHMENT (11.2.19). Many surface expressions, such as classes of nouns, verbs, prepositions, and junctives, would tip the balance toward specific hypotheses. For example, individual prepositions would narrow down the range of conceptual links: ‘in’ would suggest ‘location-of’, ‘time- of’, ‘containment-of’...; ‘of’ would indicate ‘possession-of’, ‘part-of’, ‘substance-of’...; and so on. Individual conjunctions would have the same effect :’because’for ‘cause-of,’ ‘reason-of...;'when’ for ‘proximate-in-time- to’...; ‘beside’ for ‘proximate-in-location-to’...; and so on. Procedural attachment would be maximally efficient if it focused on the most reliable indicators and tested the most constraining hypotheses first (cf. P. Hayes 1977: 8).
4.18 Although the matter is far from worked out, I surmise that tense, voice, and mood can also be utilized as cues for building hypotheses about the arrangement of textual worlds. Tense is responsible both for the time organization of a textual world and for the relationship of the communicative situation to that world. Mood indicators signal the modality of text-world events and situations, e.g. as projected or counterfactual (cf. Goldman 1975: 360). Voice helps to distribute focus on the participants in events and actions (e.g. agent, affected entity, instrument, etc.) (cf. Beaugrande 1977a, 1977b).
4.19 The preferences I have proposed would operate in the other
direction during the PRODUCTION of texts. Here, the organization of concepts and
relations would give rise to preferences about mapping onto surface structure.
There would of course be ASYMMETRY in production just as much as in
comprehension, but the problem-solving for sequential continuity of the surface
text would still be greatly simplified. The partial non-determinacy that arises
from asymmetry would affect production in the form of occasionally competing
strategies of expression, i.e., several ways of saying much the same thing are
trying to assert themselves at the same time — a major source of errors or
inconsistencies in speaking and writing (cf. IX.4.3). I shall postpone a more
developed treatment of text production for section VII.2.
Equipped with the typologies of concepts, relations, and operators presented so
far, we can observe how a text world model could be built for the ‘rocket’
sample that has already supplied some fragments for discussion. I use this text,
especially because it has been investigated before (e.g. McCall &
’Reprinted by permission of the publisher from McCall-Crabbs Standard Test
Lessons in Reading, Book C, p. 8. (New York: Teachers College Press, 10
1926, 1950, 1961, by Teachers College, Columbia University.) The original does
not have a paragraph break after ‘fire the rocket’, as I found out after the
tests were run; and ‘miles per hour’ was used rather tha ‘mph’. Aquino (1969:
353) notes that this text received relatively low scores on cloze tests — a
finding which may be related to the inexact match with the schema (cf.
Miller & Coleman 1967; Aquino 1969; Kintsch & Vipond 1979). The text runs like this:
(35.1.1) A great black and yellow V-2 rocket 46 feet long stood in a New Mexico desert. (35.1.2) Empty, it weighed five tons. (35.1.3) For fuel it carried eight tons of alcohol and liquid oxygen.
(35.2.1) Everything was ready. (35.2.2) Scientists and generals withdrew to some distance and crouched behind earth mounds. (35.2.3) Two red flares rose as a signal to fire the rocket.
(35.3.1) With a great roar and burst of flame the giant rocket rose slowly and then faster and faster. (35.3.2) Behind it trailed sixty feet of yellow flame. (35.3.3) Soon the flame looked like a yellow star. (35.3.4) In a few seconds it was too high to be seen, (35.3.5) but radar tracked it as it sped upward to 3,000 mph.
(35.4.1) A few minutes after it was fired, (35.4.2) the pilot of a
watching plane saw it (35.4.3) return at a speed of 2,400 mph and plunge into
earth forty miles from the starting point.
4.21 In chapter II, we worked through a fragment of the opening stretch of this text, ending up with a labeled grammatical dependency network shown as Figure 6 back in II.2.18. If the preference strategy cited in III.4.14 were applied, the nodes of ‘rocket’ and ‘stood’ would be taken as the control centers: the primary concepts from which the processor can work outwards to identify the other nodes. The ‘rocket’ is thus an ‘object’ node, and the connected nodes are not difficult to characterize: ‘great’, ‘black’, ‘yellow’, and ‘long’ are all ‘attributes: ‘V-2’ is a ‘specification’ of ‘rocket’, being a subclass; and ‘46’ and ‘feet’ are both ‘quantities’ hanging on ‘long’. In moving from ’rocket’ to ‘stood’, the preference that ‘subject-to-verb’ should correspond to ‘agent-to- action’ (III.4.16.1) is not tested, because ‘rocket’ was already taken as an ‘object’ concept; the second preference for ‘object-to-state’ is tested and confirmed. The preposition ‘in’ and the two place names ‘New Mexico’ and ‘desert’ offer sufficient evidence that ‘locations’ should be connected to the ‘state.’
4.22 The outcome of this processing is the labeled conceptual/ relational network shown in Figure 11. The arrows show the DIRECTIONALITY of the control flow outward from the central
points. The arrows are aimed toward the concept node whose type the label describes, e.g., ‘great - rocket’. can be read off as ‘great is an attribute of rocket.20 [20. On the use of arrows in both directions, see Footnote 10.] I use the English words of the text not as words per se but as concept names privileged by their actual occurrence. The creation of such a network is not intended to explicate the meaning of the individual concepts (e.g., what ‘yellow’ means), but only to show how the concepts are interconnected. This task is a simple case of problem-solving as depicted in I.6.7ff. Notice that the configuration could still be recovered if the surface structure were not fully perceived, as my tests with indistinctly pronounced function words proved (II.2.18). Even a disjointed fragment like ‘rocket.…desert’ would not be hard to label as object-to-location.’
4.23 As we can see, the determiner ‘a’ was suppressed in the conceptual/relational network as a non-concept; it is, however, a useful signal that a new node should be created for its head, since the indefinite article usually precedes items just being introduced (cf. V.3.13). As processing continues to the next sentence-stretch, the pronoun ‘it’ is also suppressed as soon as it can be identified with a concept already introduced. From the standpoint of grammar alone, this ‘it’ might be applicable to’rocket’, ‘desert’, or even ‘New Mexico’. If the criterion of greatest proximity in surface structure were used, the proper referent would not be found. If processing consisted of looking up words in a mental lexicon, there would still be no resolution. No lexicon stipulates what a rocket, a desert, or, Lord knows, what the state of New Mexico ought to weigh. A lexicon of ‘markers’ in the style of Katz and Fodor (1963) wouldn’t be helpful either, since all three candidates are (+ physical object) and (+ mass), and thus have weight. However, in all of the tests run with this text (see sections VI.3 and VII.3), nobody mistook this referent. People were simply using world knowledge that the weight of flying objects is relevant and problematic in a world where gravity can cause a flight to fail.21 [21. I argue that this problematic access also impels readers to recall the ’take-off’ especially well (V1.3.1 1). In VIII.1.11, I further suggest that problematic linkage is favored in continuation utterances in conversation.] In contrast, geographical regions probably won’t be moved, so a rational language user would not expect their weight to be relevant, or indeed even calculable. Along the same lines, the referent for ‘it’ was found in (35.3.4),.(35.4. 1), and (35.4.2) to be ‘rocket’ despite some other candidates in the vicinity (‘flame’, ‘star’, ‘radar) because of expectations that ‘rocket’ is the most likely object to be ‘fired’ to ‘return’, and to ‘plunge’.
4.24 We see that even ostensibly straightforward usage demands inferences from world-knowledge for efficient processing. The knowledge activated when the ‘rocket’ concept is initially encountered precludes the need for lengthy searching and weighing of common referents throughout the text. The heavy use of ‘it’ may be a sign of unskilled writing, yet it does not constitute an obstacle to understanding. A linguistic theory which would see syntax and grammar as autonomous of meaning, and linguistic meaning as distinct from world knowledge, would lead to very intricate and possibly unresolvable computations over issues as simple as these.
4.25 In models of language comprehension in both cognitive psychology
and artificial intelligence-even models whose creators are quite hostile to
conventional linguistics-the SENTENCE is routinely construed as the standard
unit of processing. Though I have used a sentence myself in the foregoing
demonstration, I have misgivings about such an a priori assumption. Strictly
speaking, a sentence is composed of expressions rather than of concepts and
relations, so that its use in building networks like mine is somewhat
inconsistent. For instance, when I combine all occurrences of a concept onto one
node, no matter how many sentences contain the corresponding expression(s), I
seem to be moving in a domain in which ‘sentencehood’ is a disturbing
notion.22 [22. I suggest in VII.2.18ff. that sentence
boundaries arise during text production from the partitioning of
conceptual-relational networks according to criteria of motivation,
informativity, and focus.]
4.26 The heavy use of sentences in comprehension models keeps us from addressing the question of how long a stretch of text people actually process at one time. The units of surface syntax cannot be the only determining factor for marking off a workable section of material. Other factors might be: (1) the span of active storage for maintaining conceptual connectivity of input; (2) the internal compactness or diffuseness of a knowledge configuration; (3) the number and relative probability of competing hypotheses; (4) noise, i.e. non- useable occurrences in the environment of actualization. The sentence could at most be one convenient and well-structured processing unit alongside others (O’Connell 1977). Other units could be: the PHRASE (a grammatical configuration with a head and at least one dependent element); the CLAUSE (a sentence component with its own subject-verb dependency); the TONE GROUP (a sequence of language items spoken as a unit with an apperceivable beginning and end) (cf. Halliday 1967c); the UTTERANCE (the action of producing spoken language items); the DISCOURSE ACTION (a text- producing action constituting a step in a plan to attain a goal via communication) (cf. VI.4.2); and the CONVERSATIONAL TURN (the text that a participant in communication utters before another participant begins to speak) (cf. VIII.1.18). Future research is needed to sort out the role of these units in the utilization of real texts.
4.27 As each stretch of text (of whatever length and nature) is processed and added on to the material already done, a MODEL SPACE within the text-world model is gradually formed (cf. the ‘activated subgraph’ in Ortony 1975a: 57). The model space serves to integrate text-world knowledge into a CHUNK (cf. III.3.11.6) for use in further processing and for both active and long-term storage. I illustrate the model space for the first paragraph of our sample, as viewed in two ways. Figure 12a shows the content in sentence- length fragments; 12b shows the model space fully assembled. The integrating is a straightforward procedure here, because the fragments all share a
for ‘rocket’ in a central position. This node-sharing is a graphic correlate of TOPIC (cf. III.3.11.9). The shared node survives best in storage because of frequent utilization and re-activation during processing. A topic node is thus a privileged CONTROL CENTER attracting material whose status is otherwise vague, e.g. the material introduced with a careless use of ‘it’ throughout the ‘rocket’ sample (cf. III.4.23). If only the topic nodes were connected together when the whole text-world model is complete, we would have a MACRO-STRUCTURE (cf. van Dijk 1979b) that could be mapped onto the surface as a SUMMARY (cf. Taylor 1974; van Dijk 1977a: 157). In accordance with this view, model space can be considered a CONCEPTUAL MACRO-STATE analogous to the grammatical macro-states I postulated in 11.2.9. The summary rests on linking together the control centers of all macro- states.
4.28 The model space seems a likely correlate of the PARAGRAPH in the surface text. Paragraph boundaries are prone to appear when there is a transition in conceptual material (but cf. IV.4.2). These transitions are not left as gaps, as we shall see, but bridged by inferences as necessary. Our first sample paragraph is conventional in providing a topic node for the entire text (cf. Jones 1977: 32). In traditional school instruction, it was suggested that paragraphs should have ‘topic sentences.’ (Of course, it is not the sentence that is topical, but its underlying conceptual content.) The efficacy of beginning with topical content lies in making obvious control centers available right away for the material to be later connected. Yet topic sentences have been found to be less common than is claimed in schools (Braddock 1974). We shall see later that topic postponement can also be effective (cf. VII.3.7ff.).
4.29 The model space for the second paragraph is more difficult to build. The three sentence-length fragments appear to activate no shared concepts. We are not told why ‘scientists and generals’ are on the scene, nor what their motions have to do with ‘red flares’. However, INFERENCING readily overcomes these potential discontinuities. The state of ‘readiness’ can be taken to be the ‘reason-of’ the motions toward shelter, and for the ‘rising’ of the ‘flares’ as a ‘signal’. Figure 13 shows how this minimal inferencing produces an internally connected model space.
More inferencing must be done to connect this space with that for the first paragraph: that ‘everything’ refers to whatever was required to ‘enable’ the rocket’s take-off, and that the ‘scientists and generals’ were there to observe the rocket. The empirical tests we conducted with this text showed that these inferences were indeed made by a substantial number of readers (cf VI.3.9; VII.3.26). In Figure 14 we have the merger of the two spaces, with inference nodes in square brackets
4.30 An individual reader of the text might well do much more inferencing than I have shown here (cf. III.4. 1). For example, one might reason that the ‘fuel’ is about to ignite, so heat will impel personnel to hide behind non- flammable earth mounds. Later on, I shall illustrate a matching knowledge configuration I call the WORLD-KNOWLEDGE CORRELATE (cf. III.4.36) in which these additional pieces of knowledge are included. As far as the text-world model is concerned, I suggest that inferencing be postulated whenever necessary to establish at least one connection between all nodes of the model. In other words, a gap in connectivity is construed as a problem (possibility for failure of transition, cf. 1.6.7), and a ‘problem-occasioned’ inference must be done (cf. 1.6.9). Empirical research with whole texts will be needed to determine how many additional inferences are made by representative groups of language users.
In a different perspective, inferencing from world-knowledge could be addressed
to the evolution of the textual world. As events are added on, the processor
would know that earlier situations have become UPDATED in at least some
respects (cf. 1.6.4). I pointed out in III.4.13 that this fate overtakes virtually ail
of the textual world eventually, especially when the events are in past time, as
in our sample. Further experiments may show that by interrupting the
understanding process at strategic points, we can observe the effects of partial
updating along the way. Certainly, computer simulation of understanding has a
great updating task to manage, because the knowledge base otherwise stays
constant. Roger Schank (1975e: 42) even suggests that the "true meaning" of an
action is the set of inferences and updatings it elicits (cf. III.4.6).
4.32 The model space for the third paragraph resembles that for the first in having a prominent shared 'rocket' node. Figure 15 presents the whole model space with its topic node.
combining of relation types: speed (‘slowly’) as ‘quantity of motion,’ or
direction (‘upward') as 'location of motion’. I use a division sign ‘÷’ for
combining. We also see some uses of the proximity operator ‘π’, e.g. the
‘proximity of cause’ between the ‘roeket's rising’ and the ‘roar’and ‘burst’; or
the ‘proximity of time’ between the ‘apperception’ ‘tracked’ and the ‘motion’
‘sped’. Proximity of cause flows in one direction (hence one arrow); proximity
of time could flow in both directions, depending on viewpoint (hence two
4.33 In order to connect this model space to the previous textual world, we need only merge the topical 'rocket' nodes to attain sufficient connectivity. Figure 16 displays the outcome of the merger.
I include the inference that the people who could ‘not see’ the rocket when it was 'too high' were (or included) the 'scientists and generals'. This inference was also made by our test subjects, and is plausible because it re-uses available material instead of creating new nodes such as 'everyone on earth' or whatever.
4.34 In the model space for the final paragraph, we have to assign ‘quantities’ not to nodes, but to links, in order to represent ‘a few minutes after’ and ‘forty miles from’. I employ pointer links, as in Figure 17 shown below. A further issue is the linkage between 'plane' and 'pilot'. The pilot is in ‘containment-of’ the plane, while the plane is the ‘affected entity-of' the pilot's ‘agency’. I illustrate this double linkage in Figure 17 too. The extent to which it is advisable to work with multiple linkages throughout the text-world model depends on the detailedness and differentiation one desires to attain. If one breaks a concept down into components and creates links between components— along the lines of "feature overlap" mentioned in 111.3.27 — multiple links would become the rule rather than the exception. Dedre Gentner (1978) reports evidence that degrees of linkedness among the components of concepts affect ease and frequency of exact recall. I shall content myself here with single links as the minimum for coherence.
The complete text-world model for the ‘rocket’ sample is thus diagrammed in Figure 17.
The vertical arrangement corresponds to the progression from the
initial to the final stage of processing. This model is undeniably an
idealization. It suggests complete and accurate recovery of all relations. It
shows none of the decay that a human processor would experience in real time
(cf. discussions in VII.3). It makes no provision for the time organization in
the textual world — I always use the basic form of verbs, irrespective of the
tenses in the surface text — but portrays all the relations at once. No attempt
is made to capture factors of value, emotion, or mental imagery. Nonetheless,
such a text-world model can be a helpful starting point for exploring the
processes applied in such tasks as: (1) forming a ‘gist’ of the text; (2)
storing the text content and recalling it at a subsequent time; and (3)
controlling and compensating for decayed or confused components.
4.36 To suggest how an understander would MATCH the textual content against prior knowledge of the world, I have designed a format I designate the WORLD-KNOWLEDGE CORRELATE. This entity is drawn with much the same proportioning as the text-world model. It contains only the nodes that people might reasonably know to be linked before they ever encountered the ‘rocket’ text. I attempt to distinguish the STRENGTH of linkage with the operators for determinateness and typicalness as expounded in III.4.12, although some cases might be disputed. Figure 18 illustrates the results in a world-knowledge correlate for 'rocket'.
For example, it is essential to the identity of the concept ‘fuel’ that fuel ‘burns’ and enables some vehicle such as a rocket to move. An object cannot ‘rise’ except with an ‘upward’ motion. It is a requisite that ‘radar’ be able to ‘track’. Barring bizzare counter-examples, relations like these can be labeled determinate. Others are merely typical, such as those between ‘scientists’ and ’explore’or ‘generals’ and ‘attack’; scientists might give up research upon getting tenure and dump it on their assistants, and generals might only march up and down like popinjays or doze through interminable staff meetings. But the typicalness of exploring and attacking as their respective agencies is the presumed reason they would want to make use of rockets.
4.37 As we can see, these world-knowledge links hold together many
elements whose relatedness is not asserted or mentioned in the text itself.
These links would be available by SPREADING ACTIVATION of the pertinent concepts
(III.3.24), and would make the recovery of relations that are asserted in the
text efficient. This utilization of knowledge is a form of PROCEDURAL
ATTACHMENT: modification and specification of stored procedures for an immediate
task (cf. III.4.1). The COHERENCE of the text seen in isolation is only partial,
since its continuity as a processing object comes from prior knowledge as well
as presented knowledge. Without such interaction, processing would be explosive,
requiring an unmanageable number of alternatives to be considered (cf.
II.1.2f.). A comparable outlook on coherence is suggested in many of the
“substitution” types discussed by Roland Harweg (1968a), as well as by notions
like “lexical solidarity” used by Eugenio Coseriu (1967; cf. Dressier 1970a:
194) and “preference semantics” used by Yorick Wilks (1975b, 1978).
4.38 The elusive nature of conceptual relations out of context is manifested in some of the linkages in Figure 18. ‘Fire’ and ‘flame’ could either one be thought of as the ‘substance-of’ the other, depending on usage; ‘seeing’and ‘watching’ could be each other’s ‘enablements’ in appropriate settings. I employ two arrows for such instances. In actualization via a text, however, only one direction would normally be relevant, especially if structural processing is viewed as a directional flow of control.
4.39 The standard for stipulating world knowledge is an admittedly “naive psychology” (Rieger 1975: 187f.). A theory of human activities has no special motives for insisting on an exhaustive, precise, logically perfect base of knowledge. Instead, we want to explore COMMONSENSE REASONING (Wilks 1977c: 236), and COMMONSENSE KNOWLEDGE (Petöfi 1975a: 43) (cf. I.6.4). These domains correspond to what the average person in a given language group of society can plausibly be presumed to know and reason about. The same presumption underlies communicative processes at large, and if it were invalid, people simply could not understand each other much of the time. Moreover, an unduly exact knowledge base would entail very laborious procedures of utilization and matching, rather than the “fuzzy matching” that makes knowledge spaces so versatile and useful (Rieger 1977a: 277).
4.40 Commonsense knowledge easily imposes coherence on the newspaper advertisement cited back in II.2.36:
(26) PIZZAMAN EXPRESSES WE DELIVER
50¢ OFF ANY PIZZA
plus 2 free cokes
Open at 11:00 A. M.
I argued at the time that such a text is certainly not understood via conversion into complete sentences, but rather via inferencing with concepts and relations. The ‘Pizzaman Express’ is an instance of ‘pizza parlor’ whose ‘management’ is the agent of the expressed action ‘deliver’ as well as of the inferable action ‘offer’. The ‘’pizza’ and’ cokes’ are ‘specifications-of’ the ‘merchandise’ that the management ‘delivers’. Their ‘prices’ (the “instrument-of-entry-into-possession”) have the ‘quantities’ of being’ 50¢ less’ in the first case and ‘’free’ in the second. Some further nodes are readily attached as “times.” The requisite material is supplied from stored knowledge about business dealing and restaurants — a "restaurant script” of the kind cited in VI.1.3 and VI.4.13. The proof that such knowledge is really available is the advertisement itself. Business people do not waste money circulating ineffective or incoherent messages.
4.41 This chapter has been devoted to exploring meaning as a PROCESS, not as a property of grammarians’ “sentences.” These processes should apply to the acquisition, storage, and utilization of knowledge. The production and comprehension of texts was proposed as a profitable area for studying meaning from the standpoint of maintaining CONCEPTUAL CONNECTIVITY as the basis of COHERENCE. These criteria are vital to the stability of systems of meaning in which a CONTINUITY of OCCURRENCES allows a steady directional flow of control (I.4.4). Consequently, text-presented knowledge must interact heavily with previously stored knowledge of the world, so that possible discontinuities are overcome by problem- solving, pattern-matching, spreading activation, inferencing, and class inheritance. I outlined the procedures for building a model of a textual world.
4.42 In the remaining chapters, I explore a number of issues I hold to be vital for a science of texts. These issues offer a rigorous test of the usefulness of the basic theory presented so far, or of any theory which deals with texts in communication.
Click here to go to Ch IV
Click here to go to Table of Contents