Theory and Application

Site  sponsored by Linguistic Technologies, Inc.




  Eldon G. Lytle

 Invited Lecture
Presented at the Annual Forum of the
Linguistic Association of Canada and the United States (LACUS)

 August 25, 1979


 Congregational fishing is an amusing sport to watch, but frustrating to participate in. A throng of fishermen attack the same expanse of water with vigor and straightway their lines become entangled. The participants then either tug angrily, exchanging accusations and unpleasantries, thereby creating a hopeless snarl, or they work together in a cooperative effort to discover the focal points of entanglement and free themselves.

If I understand the purpose of the LACUS Forum correctly, it is to provide a more productive and dignified method of disentangling a rather bewildering array of linguistic snarls, so that we may get on with the task of meeting the information and communication needs of a world society. 

Although I am new to LACUS, I get the impression from reading its publications that its members are communicating with one another honestly and constructively.  For this I congratulate you and your leaders, and express my sincere appreciation to them for this opportunity to participate with you.


The purpose of this paper is to introduce you to Junction Grammar(JG), and relate it to other more familiar approaches to linguistics.[1] The name itself is suggested by the emphasis of the approach on the need to formalize such traditional notions of joining as conjunction, modification, and predication at a level of non-lexical representation not recognized heretofore. The JG model as a whole, however, shaped by a number of other theoretically significant assumptions. 

An important initial assumption of JG is that the scientific study of natural language presupposes a holistic concern with the physical bases of language (the human body and its environment), as well as the sociological and psychological implementations and manifestations of language. With regards to human anatomy, I assume further that the physiological specialization of diverse organs and tissues involved in linguistic activity presupposes a number of interacting information systems, each with its own distinctive code or data type (I shall use these terms interchangeably).  

 Permit me to clarify the JG notion of data type by drawing an analogy.  In nature, we find that animals and plants group together as distinct species, and that the boundaries between them are rather firmly drawn. Thus, animals cross-breed within species and undergo transformation to some extent therein, but not between them.



Linguistic data types appear to be similarly constrained by virtue of the functional diversity of the tissues and organs which constitute the human capacity for speech. For example, code designed to activate the musculature of the vocal tract could not be used to drive the writing hand, nor to stimulate the neurological tissues of any "semantic" tract. From the point of view of junction theory, then, it would be unnatural to fuse these codes as though they belonged to a single representational system. Rather, it would be necessary to maintain them separately, in order to satisfy the unique content and formalism of each, and then, additionally, to provide a means for them to interact if necessary. 

The failure of linguistic science to do this systematically is largely responsible, in my opinion, for the embarrassing absence, at this late date, of the at least relatively unified base which other more prestigious areas of scientific inquiry enjoy. Some will counter, perhaps, that linguistics is a young science, and that the present diversity is healthy. 

Excuses aside, I am firmly convinced that the crux of the problem is rooted in the naive practice of "cross-breeding" data types in unnatural ways. While the resulting offspring are plentiful, and, to be sure, captivating in their infancy, and promising in their youth, they have a way of aborting prematurely, or developing fatal post-natal deformities. Still others mature as impotent mules, as it were, or as "linguicorns" which have no counterparts in reality, but exist only in the world of what might be termed "mythological linguistics." 

  Assuming, however, that human language is a natural phenomenon, and that we wish to describe it in terms of what it Is, rather than in terms of certain similarities it may bear to other phenomena, the following general objectives seem appropriate for linguistic science:

     (1)    To identify the data types operative in natural language;

     (2)    To establish their association with specific organs and/or tissues of the human body;

     (3)    To identify their information content;

     (4)    To devise appropriate formal notations for each of them;

     (5)    To discover and describe the encoding/decoding linkages which intermediate between them;

     (6)    To study the socio-psychological-communicative manifestations and implementations of the various interacting data systems.

  Junction Grammar looks to these objectives for general guidance. Of course, they constitute a gargantuan task and infringe upon sister disciplines in obvious ways. However, it is essential to approach science from a broad perspective, if we expect the contribution of individual researchers to accumulate systematically and form a cohesive whole. 


  A sketch of the current JG working model is given In Figure 5. It corresponds to the following conception of interacting data systems: 

      (1) Physical stimuli contact the various human sensors, which by virtue of their physiological design reduce the input data in a human-specific way to a distinctive subset of the input. We may refer to this in the aggregate as SENSORY DATA, noting that acoustic data, visual data, olfactory data, etc, constitute distinct data systems within it.  The linguistic relevance of these data types is amply demonstrated by expressions such as THIS WHOLE AFFAIR STINKS; YOUR GENEROSITY TOUCHES ME DEEPLY; IT LOOKS LIKE TROUBLE; etc.[2] Dreams and other hallucinatory phenomena, as well as experimentation with brain probes, indicate that sensory data is stored in a form compatible with the neural structure of the respective sensors. The task then is to identify each sensor and devise a notation which describes the information content and structure of the data type which activates it.

       (2) Since we do not necessarily hallucinate as we speak, it is postulated that sensory data is converted into a second data type in which individual entities are reduced to numbers, as it were, and natural classes are constructed by abstraction of distinctive attributes from the various sensory codes. Let us refer to this second general data type as BASIC DATA. 

 Whereas sensory data, constituting, as it were, raw information, is presumed to be neutral with regard to such notions as cause-effect and condition-consequence, basic data is not. It is subjective data, having been coded in conjunction with human interpretive processes by each separate individual in the context of his/her unique experience. It seems probable that an assortment of logical processors complement the basic data system, and that subsets of this data are coded out as input to them with the particular content and form each requires. I assume further that basic data is central to the construction of information nets which document general world knowledge and experience. Although the basic data system of JG remains for the most part an uncharted wilderness, initial studies in this area have been made.[3]

       (3) PRAGMATIC DATA is a non-lexical data type central to the socio-communicative functions of natural 1anguage. It is also the immediate basis from which the outward forms of language are coded. This data system has been investigated extensively over the past ten years with exciting results, to which we return later in the discussion.  I have termed this data PRAGMATIC because it incorporates information relative to the immediate practical needs of a discourse situation.

       (4) Pragmatic data and its linkages with what may be termed SIGNATIVE DATA TYPES have been the focal point of our research to date.[4] Signative data types include those necessary to stimulate body language, as well as those required to activate the vocal tract, the writing hand, etc.

   In summary, then, we have thus far identified four general types of linguistic code:

     (1)    SENSORY DATA

     (2)    BASIC DATA

     (3)    PRAGMATIC DATA

     (4)    SIGNATIVE DATA

  Corresponding to these we either acknowledge or postulate the existence of specialized physiological mechanisms which may be referred to in the abstract as tracts or nets, e. g. the "basic net," the "discourse net," the "pragmatic tract," etc. While it is beyond the scope of this paper to discuss in detail the approach to discourse simulation associated with our research, it will be helpful to bear the following sketch in mind. Refer to Figure 5. 

  In the vernacular of JG, a pragmatic J-tree (formally a graph), is coded up in response to stimuli from a discourse monitor. This pragmatic graph in turn activates orthographic, articulatory, and/or kinesic coders, which produce their respective types of signative data.  Their outputs, when executed by the appropriate physical devices, yield external acoustic, visual, or tactile signals which are perceived and decoded by recipient(s) to recover the pragmatic graph (J-Tree). The manner In which this graph is structured then stimulates what we refer to as its COMPILATION, a process which brings about modifications in a mini net of basic data which has been activated to serve the needs of the immediate discourse. The immediate discourse net in turn interacts with a general net of basic data. 

  It is important to understand that within the JG framework data modification/transformation may occur within data types, but not between them.  Thus, except, for example, net modification, transformation is supplanted by coding. 

 Notice also that the model described does not include, among other things, the customary syntactic, and semantic components. Rather, the modules are simply identified with respect to data type and coding interfaces. How, then, does such a model relate to traditionally recognized areas of linguistic inquiry? 


 There is no semantic component per se in the model.  The "meaning" of an utterance is considered to be what a person experiences throughout the data systems as a consequence of what it stimulates in them collectively. If purely acoustic phenomena such as rhythm, rhyme, cadence, sound -symbolism, etc. evoke responses independently of any conventional symbolic associations, then those responses are considered to be part of the meaning of an utterance. This point of view is supported by the homage rendered such phenomena in literature, especially poetry, not to mention the role of chants, verbal charms, etc. 

Then, of course, there is the pragmatic J-tree which, by virtue of the explicit information it contains impacts the immediate discourse net and the basic net (which documents the semantics of reference and our encyclopedic accumulation of information about our experience and the world), augmenting or otherwise altering information there and setting into motion the calculation of truth values, probabilities, and so on. Linkages to sensory data also occur, eliciting intermittent hallucinatory flashes in our minds. 

Thus, JG makes the tacit assumption that there are multiple semantically relevant information systems operative in language. This point of view explains an otherwise paradoxical aspect of human linguistic behavior.  On the one hand, in the role of author, poet, debater, etc., we emphatically insist on the significance of every detail of a given text, suggesting that there are actually no true synonyms, that a particular arrangement of grammatical  relations has been structured precisely for its unique texture and semantic effect which it alone can evoke. But, on the other hand, are we not equally inclined following such displays of semantic sensitivity to justify some loose paraphrase containing totally different vocabulary and grammatical structure, asserting that "it means the same thing" as some other sentence? The fact is that we are "multi-bivalent" with regard to semantic judgments.  


 Such behavior is not paradoxical, however, in the context of the JG model.  Because every utterance has linkages to multiple systems, one may very well tend to focus on different systems under different circumstances.  And since the criteria for distinctiveness differ from system to system, semantic judgments shift with each change of focus. 

 For example, if one considers sensory data as the stimulus for a descriptive utterance, and attends to its great detail, a wide range of paraphrases may seem essentially equivalent, because each is a valid interpretation of that data. Similarly, if one attends only to basic data where the criteria of distinctiveness are not pragmatically oriented, then a wide variance in terms of pragmatic information is tolerated. On the other hand, however, if one is attending to the pragmatic precision of an utterance necessary to communicate with persons x, y, and z, in a given discourse environment q, then the margin of pragmatic tolerance shrinks considerably. 

 Of the various data types postulated, sensory data may be said to be the most objective, it being provided by common hardware, so to speak, in a form least dependent upon individual experience. Basic data is yielded via an interpretive coding process influenced both by the socio-linguo-cultural environment and personal experience. Thus it may be said that basic data contains more human-specific information than does sensory data.  But pragmatic data, drawing as it does upon the content of basic data, plus an additional increment of person-specific information from a particular communicative point of view, would appear to impose the most discriminating criteria of distinctiveness of all the data types. 

Now observe that this conclusion, involving as it does pragmatic data, which is presumed to govern the external forms of language, and which through its reverse linkage to basic and sensory data is in a sense "semantically accumulative, is not in harmony with approaches which minimize the semantic significance of "surface structure' while searching elsewhere for something to call semantics. On the other hand, however, it explains perfectly the tendency of those who have recently become sensitized to the semantic relevance of external form, and are designing models which take this into account. 

  Having considered semantics within the context of the JG model, and seen that there are multiple semantically relevant data types, let us next consider the question of syntax within the JG framework. 


 The assumptions set forth above regarding the need to maintain data types independently, while at the same time providing for coding linkages between them, lead to some rather disturbing conclusions regarding syntax and the troubles it has experienced.

  As a point of entry to this discussion, consider the nature of a cryptographic sequence.  The form and content of such a sequence is a function of the form and content of some primary data system, i.e., a base, plus a coding algorithm.  It follows, then, that the explication of any cryptographic string presupposes knowledge of both the primary system and the algorithm.

 The first grammarians, desiring to give an account of the outward form of their languages, were, in a sense, cryptographers. They didn't recognize their task as such, of course, because, though endowed with the ability to code and decode language, they were not overtly aware of the existence of the base from which language as they knew it was coded. The situation, then, was for them doubly difficult - In order to succeed at their conscious task, they needed first to DISCOVER the primary data system, and then BREAK THE CODE, that is explicate the relationship between that system and the surface form of their languages.  

  Unfortunately, the actual base went largely undiscovered, although certain truths about its structure were verbalized which modern linguistic science failed to include in its formulations (the various forms of joining previously alluded to and to which we shall return directly).  Consequently, grammarians, rather than viewing their task as one of discovering a covert coding base and then relating the surface forms of language to it, dealt with information both from the base and its external manifestations as a single data system. 

 Failure to discover the base is attributable, I believe, to a flaw in human linguistic perception, which for want of a better term I shall term the "presence” illusion. This phenomenon is analogous to a familiar optical illusion - objects at considerable depth in a deep pool or stream appear to be immediately present in the water.  The linguistic counterpart of this is simply that meaning appears to be immediately present at the surface of language also.  In fact, the sense of a word is such an integral part of it perceptually, that we experience something akin to the visual "death" of a word from time to time, when for some unknown reason the graph loses it symbolic link and suddenly appears strange and foreign to the eye. 

 Early on, of course, the presence illusion was partially defeated. People came to realize that the outward forms of language were symbols for internal units of meaning, which existed independently of external forms. With respect to another vitally important aspect of meaning, however, the presence illusion persisted – relational senses (subject, object, modifier, etc.) were still considered to be “present” among the external forms of language. This is evidenced by traditional methods of sentence diagramming, not to mention those famous descriptive devices known as phrase markers. 

  We now come directly to the point: The presence illusion has been perpetuated in our day under the name of syntax, While, as noted, it is no longer the case that the sense of external forms is considered to be present in the forms themselves, syntacticians still proceed as though the relational senses perceived in language are actually present among lexical symbols, rather than among the non-lexical units of meaning to which they are linked. Or to put it differently, modern syntax has inherited from traditional grammar the practice of stating with reference to a given lexical unit in a sentence not only that it belongs to a particular inflectional category and has certain properties relating of co-articulation when incorporated into the articulation stream, but also that it has the pragmatic function of BEING a subject, predicate, or a modifier.

 Now it may SEEM TO BE a matter of insignificance to say that an external form IS a subject rather than that it corresponds to a unit with this function in the base code, but that seeming insignificance is also an illusion - Our research in JG has taught us thatthe failure to remove relational senses from the domain of lexical representation is a theoretical error the same magnitude as (and corollary to) the failure to differentiate between a word and its sense(s). 

Consider also the following specific points:

     (1) Aside from the remark already made with regard to the semantic relevance of multiple data types, the question of whether syntax is independent of semantics (autonomous) becomes a non question if one acknowledges that what is called "syntax" includes within the domain of its notation, depending on the particular formulation, a sprinkling of basic data, and healthy portions of pragmatic data, the data type which not only includes person-specific information, but also situation-specific information. 

    (2) Further, the realization that "syntax" does not correspond to any natural linguistic data type, but is in fact a mixture of several of them, explains its "mule-ish" impotency to produce (for any of the several data types it attempts to combine) the natural notational systems actually required. We have found that pragmatic data requires a different form and content than graphic or articulatory data, and that notations that attempt to combine them tend to conceal vital aspects of each while ultimate failing to do justice to either.  

 Now observe that given a situation in which a "congregational effort" is devoted to the elaboration of such a "syntactic" notation, it is predictable that some participants would argue in favor of devices needed to describe one of the data types or coding linkages involved, while others would counter with arguments for whatever might be required to support other ingredients of the mixture. Hence the heterogeneous and fragmentary nature of the grammars thus far produced by syntactic models. 


 In addition to the problem of mixing data types, there are other equally serious entanglements to be dealt with in the area of linguistic notation.  A summary statement of additional aspects of the rationale applied in JG is provided in the following paragraphs.

 Given some entity, one intuitively categorizes it on the basis of the properties it exhibits, these being determined by how it is internally constituted.  Assuming that linguistic notation should reflect the basis of this intuitive capability, it follows that any category label used in a notation is an abbreviation for the internal content which determines that category. It further follows that if the specifications of content which determine the category are notationally present, then its label is a redundancy and may in principle be omitted, although it is often useful to retain labels as a means of improving notational "visibility."  

 However, given a structure of entities of known category, com~ binatorial patterning alone is not sufficient in many cases to determine the respective functions of the entities with regard to each other. Thus, the functional roles of the nominal constituents in the structure N -> N N are not specified by a rewrite rule in which only concatenation is expressed: Either UNCLE JOHN or DOORKNOB fits the pattern, but the relational senses (functions) perceived in the pragmatic base code are certainly not the same. Moreover, it is often the case that constituents are used in a way which is incompatible with their categorial content. Thus, a proper nominal can function as though it were common, if it is functionally related in the required way to other elements e. g. THE HENRY JAMES THAT I KNOW. The fact (1) that categorial patterns are often functionally ambiguous, plus the fact(2) that is a common phenomenon for constituents to function abnormally, makes the notation of functional relations, or, as we refer to them, junctions, indispensable.

 Translating this into more familiar terms, if tree-oriented notation is used, it must show more than mother-daughter relations (compositionality) - sibling relations must also be shown. In other words, the usual branching pattern must be augmented with the equivalent of a distinctive line joining the constituents of each branch at the base, as it were. 

 Similarly, if rule- or formula- oriented notation is used, a function/junction/operation must be given to specify how its elements are joined. In the absence of junctional specification, or if a constituent is cited out of context, then functional node labels are necessary. Such labels are unnecessary and redundant, however, if junctional specifications are properly provided. 

 If net notation is used, an adequate variety of connections (junctions) must be utilized, and care must be taken to distinguish notationally not only between compositionality within a data type and coding realization between data types, but also between specification of variables within a data type and realization between data types. In other words, in networks, one must be concerned not with just two phenomena (compositionality and coding realization), but with three (constituent compositionality, variable specification, and coding realization). 

  In general, it is my observation that the junctional specificity of linguistic descriptions (as opposed to compositional specificity) is often neglected and can’t in fact be determined at all unless unambiguous categorial patterns happen to be given. In other instances, one or two junctions are used for everything, or constituents are simply concatenated, or compositionality, specification, and realization are confused and/or not notationally distinguished.


 Thus far I have emphasized the JG postulate that data types having distinct form and content should be maintained separately while attending to the coding interfaces required to intermediate between them. We noted that from this point of view, semantics is an accumulative, multi-system phenomenon rather than something that fits neatly within a single component, and that syntactic formulations currently extant constitute an unnatural combination of pragmatic data and other data types.

 We also likened the task of early grammarians to the task of the cryptographer. It was noted that their task was, in effect, to discover the coding base of natural language, and then to relate external form to that base.  While the illusion that word and relational senses were present in external form itself interfered with the successful accomplishment of that task, these folks did, as previously suggested, discover and discuss certain fundamentals of structure which modern linguistic approaches have failed to elaborate. In the following paragraphs I will attempt to demonstrate how, by elaborating an appropriate notation for these traditional ways of joining pragmatic constituents, powerful generalizations can be stated which simultaneously explain both matters of external form and a range of interesting semantic phenomena. 

 Let us now consider how such traditional notions as predication, subordination, and coordination, which we shall refer to as ADJUNCTION, SUBJUNCTION, AND CONJUNCTION, respectively, can be incorporated into a non-lexical notation for pragmatic data that makes possible the statement of generalizations relevant to both external form and meaning which seem to have to been overlooked by other approaches. 


  In accordance with the principles of notation previously set forth, we advance as the basic junction rule (J-Rule) format and corresponding graph the expressions 

    X o Y = Z             Z → X o Y

where X, Y, and Z are category labels for the internal content of nodes; "o" is a junction which specifies the functional roles of X and Y with regard to each other; and "=" assigns a category label to the constituent (X o Y) in terms of its internal content. It is presumed that the entities X and Y entering into the junction will belong to the same data type, and yield Z of the same data type. Thus, if X and Y are pragmatic units, then Z is likewise a pragmatic unit. On the other hand, if X and Y were articulatory units then Z would likewise be an articulatory constituent. 

 For example, the articulatory representation for /hway3 iyt/ would appear as an aggregate of junctions specifying the various modes of co-articulation to be used and the articulatory categories of the units to be articulated. 


 The pragmatic representation corresponding to the articulation graph, on the other hand, would be given as an aggregate of junctions specifying the various modes of pragmatic junction in force, and the pragmo-semantic categories of the constituent units involved. The conventions of JG would require that any pragmatic information appearing in the articulation graph be considered strictly as a coding gloss rather than an integral part of the articulatory notation system. By the same token, any lexical information added to the pragmatic graph would be given only as a coding addendum, and not part of the pragmatic representation proper.  


Compare the representations of Figures 10 and ll.   Observe that the pragmatic constituents WHY/DID/YOU, which are not directly joined in Figure 11, are fused into a single word in Figure 10. DID and YOU, in fact, correspond to the consonant /3/ and the vowel  /ð/,   respectively,  in  the final syllable of the word.  Articulatory reduction, of course, while characteristic of live speech, occurs within the model without any alteration of pragmatic data.[5] 

Notice that each representation is sufficient in its own domain. There is neither reason nor opportunity to combine them, since to do so would compromise the integrity of both. To insert articulatory word-graphs at the terminal nodes of a pragmatic graph would on the one hand destroy its sememic references to the data nets, and on the other hand deprive the word graphs of their supra-segmental connections with each other. We shall consider directly the nature of the coding interface between pragmatic and articulatory data. For the moment, however, let us return to the question of adjoining, conjoining, and subjoining within the domain of pragmatic data. 

  In general, adjunction joins units from distinct categories, and yields a composite unit whose category does not correspond to that of either unit entering into the junction. Thus, the PREDICATE is a different category than either the VERB or NOUN entering into its composition.

  The individual junction patterns group together into distinctive combinations to form cohesive functional units of various kinds. These are referred to in the vernacular of JG as junction templates. Thus, the adjunctions "+" and "#" occur in tandem to form the STATEMENT template (SX). It is the particular kind of adjunction used which results in our perception of some nouns as objects and others as subjects. The SX template is depicted in Figure 12. 

 In junction grammar, however, the nucleus of the SX template (given as X), is a variable ranging over the categories V(verb); ADJ(adjective); ADV(adverb); P(preposition); and, in some formulations , N(noun). This makes it possible to give pragmatic graphs directly for "telegraphic" statements which have no verbal nucleus, e. g. YOU IDIOT; WITHOUT SCRUPLES, THAT GUY; ME SAD?; etc.  The node labels of the SX template are specialized to reflect the category of the nucleus, as illustrated in the Figure. 

 Conjunction requires that the constituents entering into it be equal or at least very proximate categorially. Thus, one can conjoin nouns to nouns, verbs to verbs, and statements to statements; but not nouns to verbs, verbs to statements, etc.  Of course, this form of junction is iterative, yielding both open and closed constituents, e. g. JOHN AND HENRY; JOHN, HENRY, MARK, ETC. There are numerous specializations of conjunction, including the AND's, OR's, BUT's, and other more specialized forms utilized in the verbal rendition of, for example, mathematical expressions, e. g, FIVE PLUS TWO. 


  Subjunction, corresponding roughly to what grammarians have discussed variously as subordination and/or modification, occurs in great variety, and when properly understood and formulated creates order out of chaos for a remarkably large corpus of natural language data.

 We have noted that adjunctions join constituents of distinct category and yield constituents of distinct category, and that conjunctions, in general, join constituents of the same category. Subjunction, on the other hand, functions in such a way that one of the constituents entering into the junction passes its category to the resultant composite unit. Thus, in the subjunction corresponding to CENTURIES OLD, it is the adjective OLD that categorizes the structure as a whole, while for the expression THE VERB TO RUN, the subjunction uses the nominal constituent to categorize the whole. 

 Another distinctive property of subjunction is that elements so joined either overlap one another referentially, or interact in such a way that the reference of one defines that of the other. For example, in the expression, THE FACT THAT HE LEARNED SURPRISED US, the ambiguity is resolved by either subjoining the clause THAT HE LEARNED to FACT, where the overlap of reference occurs between FACT and the entire clause as its complement (N * SV 5 N), or by subjoining only THAT(WHICH) to FACT as a relative pronoun N * N/PV = N), which identifies some fact not explicitly mentioned. The respective pragmatic graphs are given in Figure 14.  Please bear in mind that the "words" appearing as terminal elements in pragmatic graphs represent pragmo-semantic values, not lexical entries. 

 The complement reading utilizes what we term a "full subjunction," as opposed to the "interjunction" of the relative clause reading. Notice that interjunction entails the use of the same node by two distinct J-rules (N * N = N and V + N = PV). This is reflected in the format of the respective J-rules (for N * SV = N read N subjoin SV yields N, and for N " N/PV = N read N subjoin N of PV yields N). 

 Now observe that if N * SV = N corresponds to noun complement constructions, then ADJ * SV = ADJ would require a clause (and here it comes) SUCH THAT IT IS EMBEDDED TO AN ADJECTIVE HEAD. 

 Similarly, the rule ADV * SV = ADV would require us to subjoin a clause (and here it comes) SO THAT IT FUNCTIONS AS THE COMPLEMENT OF AN ADVERB.


Observe further that if N * N/PV = N corresponds to the interjunction of an N in a superordinate construction to the object of a verb in a relative clause, then the rule N * N/SV = N would interjoin a noun to the subject of a relative clause, as in THE PERSON WHO WON THE PRIZE.

Similarly, the rule N " N/SP = N would interjoin a noun to the subject of a prepositional statement, as in THE MAN DY THE DOOR; and the rule N * N/SA = N would interjoin a noun to an adjective statement, as in THE RICH BOY. 


In each of these cases the antecedent represents a class of entities (MAN, BOY, etc.), and the N intersect of the subordinate construction subjoined to it represents an element, or in certain cases, a subset, of that class. This structure is given for phrases, as well as clauses, because this duality of reference is present in both types of construction, whether the N Intersect is lexically coded or not. 

  Aside from the important semantic generalization captured by this formulation, there is another equally important reason for not joining simple adjectives, adverbs, and prepositional phrase modifiers directly to a head independently of the statement template: These rules are needed for other constructions.  For example, you can't fully subjoin, say, an adjective to a noun (N * AD = N) to represent a modifying adjective (RICH BOY), because such a junction is required to account for the nominalization of adjectives (RICHNESS). 

 Interjunctions, in addition to occurring individually, may also occur iteratively or recursively. Thus, BIG LITTLE ELEPHANTS is a logical contradiction if the interjoined SA's iterate. If they are recursive, however, then no contradiction occurs. We are simply talking about a subset of little elephants which are relatively big within the set. In the latter instance, of course, BIG would be articulated with heavy stress.


 Thus, in pragmatic data, phrasal modifiers are symmetrical in their structure to relative clause modifiers. If, for example, a grammarian states that OFTEN in the sentence HE ATTENDS THE RACES OFTEN "modifies" the predicate ATTENDS IHE RACES, then, in terms of JG, the rule PV * PV/SAV = PV has been used to interjoin an adverbial statement to the PV node in the pragmatic graph.

 The rule V * V/SAV = V, on the other hand would yield a particle interjoined to the verb itself, as in HE THREW AWAY ALL HIS FOOD. And the rule SV * SV/SAV = SV would yield an adverb modifying at sentence level, as in FORTUNATELY, NO ONE WAS HURT.


Notice that there is a direct correspondence between classes of modifiers and the particular J-rules required to represent them. Thus, for THE OBVIOUSLY RICH BOY the rule SA * SA/SAV = SA would be required, since OBVIOUSLY is a statement-level modifier. 

 On the other hand, THE OSTENTATIOUSLY RICH BOY would require
PA * PA/SAv = PA, since OSTENTATIOUSLY is a predicate-level adverb of manner.


  An other junction template of prime significance is what we term the referment. See Figure



 This template implements subjunctions of various kinds for purposes of referential refinement and appears to occur for all the node categories. The functional labels of the various nodes in the referment, as shown in the illustration, are CLASSIFIER, QUANTIFIER, AGGREGATE, MODALIZER, and REFERMENT, and are held to be as semantically significant as those associated with the adjunction template, i.e., SUBJECT, OBJECT, etc. For example, WE TWO STUDENTS corresponds to the referment of Figure 26.


 Notice that lexical part-of-speech distinctions (personal pronoun/article/title, cardinal pronoun, and common noun) are correlated with positions in the referment. The junction *r corresponds to definacy, effecting a recovery from the net; *e enumerates the individual students in question; and *c classifies TWO as belonging to the set of students. 

 One can feel the relational sense of a word change as its pragmo-semantic counterpart is moved about in the referment. Thus, SOME in SOMEBODY reflects a modalizing function, but in HE NEEDS SOME HELP, it reflects a quantifier function. Similarly, cardinal numbers generally quantify, but not always: In Mormonese, there are elders, priests, teachers, deacons, and "seventies," so-called because there are (here it comes) SEVENTY SEVENTIES in each Quorum of Seventy (the first occurrence of the term reflects the quantifier function while the second reflects the classifier function). The modalizer function is reflected by none other than the same term in SEVENTY NEAL MAXWELL, because social roles (titles), as well as discourse roles, are given as modalizers. 

 We observed in connection with the basic adjunctive template that modifiers interjoined at V, PV, and SV had distinct semantic properties. The same principle holds for interjunctions to the referment template. Those at the classifier have generic reference, while those at the aggregate have specific reference. For example, depending upon context, THE YOUNG STUDENT could refer to young students generally, or to a particular individual. Graphs for both readings appear in Figure 27. 

 Interjunctions to the quantifier, by contrast, constitute the class of comparative/superlative modifiers plus others relating to the extension of the focal referent, as in MORE MONEY THAN HE NEEDS; ENOUGH FOOD TO EAT; etc. 

 Although the scope of this paper does not permit more than the brief introduction to pragmatic structure presented here, there is one additional property of referments of such importance that it deserves comment: They, like the STATEMENT (the basic adjunctive template), have the property of recursion, i.e., referments may be embedded at particular positions in other referments to yield a variety of complement construcitons. For example, the expression WE THE PEOPLE subjoins one referment to the aggregate of another, as illustrated in Figure 28.  

 It is possible, using the property of referment recursion to give plausible pragmatic graphs for such horrors as COME ON UP FROM OUT OF DOWN IN UNDER THERE.  We shall of necessity, however defer discussion of such constructions to a subsequent writing. 

 With the foregoing as background, let us now consider how one does linguistics within the domain of pragmatic data.  Our experience is that two complementary techniques produce effective results: (1) Generalization; and (2) Specialization. In the following paragraphs we illustrate how these are used. 

 There exists a rule schema for each of the major junction types. Thus, given the formulaic expressions,

    (1)    X + Y = Z   ADJUNCTION

    (2)    X & X = X   CONJUNCTION

    (3)    X * Y/D = Z  INTERJUNCTION

    (4)    X * Y = Z   FULL SUBJUNCTION


a table of pragmo-semantic node labels, and a set of appropriate formal constraints, computer algorithms can be devised which generate schemata for the general junction formulas. An excerpt from the output of such a computer program is depicted in Figure 29.  

Given such a list of joining possibilities obtained by deductive generalization, it is both productive and instructive to search natural language data for expressions that satisfy the predicted patterns. 

  Let us consider a real life case. Among the rules projected by the first deduction I attempted, was the Interjunction SV * N/SV = SV. The relational import of this junction is that a main clause overlaps (intersects) the subject nominal of a subordinate clause. This meant, in effect, that I was looking for a relative pronoun whose antecedent was an entire clause. That, of course, is precisely the situation which holds in sentence relative constructions, e. g. THE ATTEMPT SUCCEEDED, WHICH SURPRISED US. I was immediately struck by the fact that the rule for what many had considered to be a problematic construction fell out as effortless y as any other interjunctive pattern. 

  It next occurred to me that if WHICH could overlap with a clause, it could probably overlap with other constituents as well. The following rules from the schema predicted that such was true, and the confirmatory data was not hard to come by: 




 So far, so good, but back to the sentence relative construction: If an SV (a clause) could serve as the antecedent for a nominal constituent, then why not for, say, an adverb:


This was exciting:  SO in (4) was nothing more nor less than a "relative pro-adverb" with an SV antecedent. My attention was then drawn to another ridiculous looking rule: 


The prediction of this rule is that two clauses can overlap on the same predicate, i.e., that a subordinate clause could consist of a subject nominal adjoined to a “relative pro-PV”. At first I was convinced that this was a spurious prediction, but then the one cited jumped out of print at me. 

  These, then, are examples of the technique of deductive generalization made possible by the schematic nature of pragmatic junctions. The possible patterns number in the thousands. We have verified many, and expect to verify many more as we become sensitive to more and more of what natural languages have to offer. 

  Another technique, which we have successfully used to complement the first, is referred to in JG as junction specialization. JG assumes tacitly that while generalization in linguistic descriptions IS important, its importance has been emphasized to the detriment of the equally important requirement of specialization. A linguistic description that shows what constructions have in common, but fails to characterize their differences, only does half the job. 

Consider again the sentence relative rule SV * N/SV = SV and the example corresponding to it:


 Observe that the presence of the comma corresponds to the NONRESTRICTIVE property of the sentence relative modifier. Observe further that if the operands (nodes) of the J-Rule in question are inverted, the full subjunction rule is obtained which corresponds to the complement inversion of the sentence relative.


 The complement inversion, however, bears a RESTRICTIVE relation to its antecedent.  We are impelled to ask: Why is the sentence relative NON-RESTRICTIVE, while the corresponding complement construction is RESTRICTIVE, and how is this contrast to be explicated in terms of junction theory? 

 To answer the first question, we first return to the assertion that subjunctions entail the referential overlap of constituents. There are, to be precise, three obvious ways that the scope of operands can overlap: 

    1. X>Y (Y RESTRICTS X)


    3. X<Y (X RESTRICTS Y)

 All of the restrictive modifiers discussed thus far correspond to the first situation. Therefore, in order for the notation to be explicit, we define the specialized symbol "-*" to use for such junctions. 

Non-restriction corresponds to the second situation - both operands have exactly the same scope of reference. As to why the sentence relative is non-restrictive, while its complement inversion is not, we merely note that an SV is inherently an individual, as opposed to a set, and therefore is not in principle restrictable. Since WHICH refers precisely to the antecedent SV, the scope of the two operands is equal.  We therefore introduce the specialized junction "-*-" to join mutually non-restrictive operands. When the operands are inverted, however, an operand moves into primary position which is restrictable (FACT), and the SV, being a fact, does restrict the scope of its head. We predict, then that the successful inversion of other restrictive complement constructions will yield other non-restrictive sentence relatives: 



The third situation (X < Y) corresponds to a situation where the primary operand is restricting a secondary operand.  FIVE BOYS corresponds to this description, We observe also that there exists a certain class of interjunctions which appear to find an explanation in this pattern. In the literature of JG they are referred to as FRAME II MODIFIERS. We have in mind expressions such as: 





 Notice that each of the noun-modifying expressions in question has a non-restrictable head. Notice also that these expressions are not non-restrictive attributes (WHITE SNOW/ SNOW, WHICH IS WHITE, but not ONLY JOHN/*JOHN, WHO IS ONLY). For the X < Y situation we introduce the specialized junction "*-"and use it for Frame II modifiers. The term "Frame II" reflects the semantic fact that such interjunctions relate a constituent whose reference is already resolved to a second frame of reference involving other relevant entities.

The specialized subjunctions already discussed in conjunction with the referment were also introduced into the notational system of JG as a consequence of specialization. Thus, by using deductive generalization in conjunction with subsequent specialization of junctions in order to account for the differences as well as the similarities in related patterns, we attempt to arrive at an understanding of pragmatic structures in all their varieties. 

  Having provided an overview of the pragmatic data system, we now turn to its linkage with the outward forms of language. 


The form and content of pragmatic data, as described above, serves as the basis for stimulating signative code in other data systems. Since most of our work to date has dealt with pragmatic data and its interface with articulatory data and/or orthographic strings, I will limit-my comments in this section of the paper to a brief description of the methods we have employed to simulate the required coding interface. 

Since the data in pragmatic graphs is not chronologically ordered, it is necessary to determine the order in which the various constituents of the graph will be coded. We refer to the rules which accomplish this as Lexical Ordering Rules (LO-Rules). It is also necessary to determine which items in the graph will be "understood" implicitly rather than coded explicitly. Lexical Hiatus Rules (LH-Rules) are used for this purpose. Those items which are to be coded must stimulate the introduction of specific lexical units in the orthographic or articulatory coding space. This is accomplished by Lexical Matching Rules (LM-Rules). Patterns of concord, inflection, conjugation, and affixation in the lexical data must also be provided. This is accomplished by Lexical Agreement Rules (LA-Rules) which accomplish the required refinement of the lexical data by referring to the pragmatic graph. Finally, if the signative data is orthographic, Graphological Rules (G-Rules) operate to punctuate the output string. If the signative data is articulatory, Phrasing Rules (PH-Rules) are used to assemble the individual mini-graphs present in the coding space into the macro-graphs that are executed as phrases in the articulation stream.

 Although some work has been devoted to the integration of the various coding" rules into what may be termed coding grammars, these rules have been implemented algorithmically on computers. We find that it is possible, given the form and content of pragmatic data as described above, to encode from that base well-formed sentences in such diverse languages as English, Japanese, Spanish, Chinese, German, Portuguese, and French. It is likewise possible, given well-formed sentences in some language, to construct decoding algorithms which recover the appropriate pragmatic graphs. 

 Let us now consider briefly some past and on-going efforts to bridge the gap between theory and application with JG.


 Junction Grammar serves as the linguistic base for the well-known Brigham Young University computer translation project.[6] This project differs from others in that we take an interactive rather than automatic approach to machine translation; hence, the name Interactive Translation System (ITS). Because of the general descriptive power of pragmatic graphs and their applicability across languages, multiple translations can be generated from this base. ITS, therefore, has a one-to-many configuration, that is, one source text (English) is analyzed, reduced to pragmatic data, and then translated into Spanish, French, German, Portuguese, and Chinese. Figure 30 depicts the structure of the ITS system.

Source materials enter the analysis phase, where a graph is built for each sentence. Because of the inherently ambiguous nature of language, the machine is sometimes unable to build that representation without human assistance.  

Ambiguities are dealt with at two levels using human interaction. At the word sense level, the computer asks the human operator to resolve certain ambiguities which are present in the text being processed. For example, the word "temple" could refer to human anatomy or to a structure devoted to religious worship. The human determines from context which meaning was intended and responds to the computer so that automatic processing may continue. Interaction at word level is also required in the transfer phase of processing where specific adjustments are made for target languages. A word such as "rock" in English may have more than one corresponding translation in, say, French, depending on context. In a case such as this, the machine presents the various options for the operator's selection. Intervention at this point on word sense problems in the transfer phase has proven to be more economical than leaving the question to be resolved by post-editing in the output of five languages. In the analysis phase, interaction is also needed to resolve the abundance of structural ambiguities found in English.

  Consider this sentence:  THE BOY THREW THE BALL ON THE TABLE. Potential meanings of this sentence could be:




 Since these ambiguities are not directly translatable (for example, in German, meaning (a) would require the accusative case, and meanings (b) and (c) would require dative case), the operator must indicate to the machine which meaning was intended so that the unambiguous JG representation of the intended structure can be built. A pragmatic graph is required for each of the potential meanings of the ambiguous sentence. Once the correct graph has been built, accurate, high quality output can be produced. 

In the transfer phase, structural adjustments are made automatically to accommodate target language requirements. For example, the English sentence I LACK CONFIDENCE would require a structural adjustment in which the object of the English sentence becomes the subject of the target sentence (CONFIDENCE IS LACKING IN ME) to produce acceptable Portuguese output.

The Interactive Translation System enters its production evaluation phase this fall and is scheduled to initiate full production work in January of 1980. Initial indications, are very encouraging, and the system promises to significantly reduce the time required to produce a finished translation.


 Because of the descriptive power of Junction Grammar, several studies have been made and others are underway to test its pedagogical viability in the areas of first and second language instruction. 

In studies conducted at the LanguageTrainingCenter, Prove, Utah, a simplified Junction Grammar representation was used to teach structural generalities about Spanish and Portuguese to second language learners. In these studies, significant gains were noted in the following areas: Number of subordinations per sentence generated in a creative setting; reduction of structural errors; and size of structures generated.[7] 

Informal studies were conducted in ESL programs in Brazil by the International Linguistic Foundation with similar results. Of particular note was the increased range of structures readily generated by participants in this program, including such non-textbook but idiomatic sentences as IT'S THE MAN WHO COMES TO TEACH US ENGLISH EVERY OTHER DAYS'S NECKTIE,

Pilot studies have been conducted in primary language instruction in Freshman English classes at BrighamYoungUniversity,  as well  as eighth and ninth grade junior high school classes in two separate school districts using simplified Junction Grammar workbooks to  present a variety of linguistic notions. Gains have been consistently noted in these studies in the areas of recognition of ambiguity, recognition of parts of speech, ability to specify modificational relationships, and increased use of subordinations in writing. Publication of the results of these studies is forthcoming.


In the course of our work with pragmatic graphs, significant correlations were noted between their structure and acoustic parameters present in the spoken utterances corresponding to them. Studies were therefore undertaken to determine the feasibility of coding up articulatory graphs from the pragmatic graphs, and then using the articulatory data thus obtained for computerized voice synthesis.[8]  

Essentially, semantic contrasts in the pragmatic data were recoded as distinctions in the prosodic super-structure of articulation graphs.  The latter were then used to generate pitch contours dynamically.  The pitch contour system produced output for each reading of a set of ambiguous sentences. Sentences were then evaluated in a formal test environment for naturalness and intelligibility of intonation.  Results of the testing showed that generated pitch contours were judged nearly as natural as human-produced contours. The generated contours were intelligible in the sense of causing the listener to perceive the intended reading of the sentence.


I stated initially that the purpose of this paper was to introduce JG and relate it to other more familiar approaches to linguistics. Although I have made no explicit mention of Transformational Grammar, Systemic Functional Grammar, Stratificational Grammar, Tagmemics, or any other specific approach, I have attempted to relate these as well as others to JG by stating forthrightly my position on crucial issues which relate to them as well as to JG: My remarks regarding the need to approach linguistic science from a broad perspective, rather than some narrow artificial point of view; my remarks regarding the need to recognize distinct data types (strata if you prefer), and a rationale for identifying them; my remarks regarding the multi-system property of semantics; my remarks regarding mixed data types, linguicorns, and the world of mythological linguistics; my remarks with regard to the presence illusion and the need to represent relational senses at a non-lexical, pragmo-semantic level; my remarks with regard to notational rationale - the significance of category and functional labels - the necessity of specifying junctional values; my remarks regarding the need to formalize our intuitive knowledge about conjunction, adjunction, subjunction; the demonstration of how one does linguistics in terms of the schemata which can be stated in terms of these fundamental notions of structure; the technique of balancing generalization with specialization; and finally the identification of JG activities in machine translation, speech synthesis, and pedagogy which relate to real world,  real-people linguistic applications …  

  Certainly this should suffice to initiate a productive and mutually beneficial dialogue between us.

[1] An  embryonic exposition  of JG  appeared in 1969 as a research monograph submitted  to the BYU Research  Division under the title  "A Grammar of  Subordinate Structures in  English."  A published revision of this work appeared much 1ater.  See Eldon G. Lytle, A Grammar of Subordinate Structures in English (The Hague: Mouton & Co., 1974).  An improved formulation of the model appears in dissertation form.  See Lytle, "Structural Derivation in Russian," University of Illinois/Champaign-Urbana, 1971.  For a developmental sketch of JG, see Lytle, "The evolution of Junction Grammar," Junction Theory and Application, Vol. 1, No. 1. 

[2] Marcel Danesi, "English Words Derived From Geometrical Terminology;" The Fifth LACUS Forum, 1978. 

[3] Lee Meador, "Information Processing Stimulated by Nouns," BYU Linguistics Symposium Proceedings, 1977.  See also Kenneth R. Lee, "The Information Net in JG – A Discovery Aid." BYU Linguistics Symposium Proceedings, 1978. 

[4] See Alan K. Melby and Jill E. Feterson, "An annotated Junction Grammar Bibliography." Junction Theory and Application, Vol. 1, No. 1. 

[5] For further discussion of the treatment of articulatory data in JG, see Lytle, "Junction Theory as a Base for Dynamic Phonological Description. "BYU Linguistics Symposium Proceedings, 1976. 

[6] For an outline description of this project see Alan K. Melby, “ITS - An Interactive Translation System”, BYU Linguistics Symposium Proceedings, 1979.  

[7]Olsen & Tuttle, "The Effect of a Graphic Representation on the Acquisition of Sentence Generation Skills in a foreign Language." Printed at 1978 AERA Annual Meeting in Toronto, Canada. Also see: Olsen & Tuttle, "The Effect of a Graphic and Explicit Representation on the Acquisition of Sentence Generation Skills in a Foreign Language", printed at 1974 AECT Convention. 

[8] Alan K. Melby et al, “Pitch Contour Generation in Speech Synthesis: A Junction Grammar Approach.” Ph.D. Dissertation published as AJCL microfiche #60, Spring, 1976.


[Home] [Origins] [Article Archive] [Foundations] [Formalizations] [Analyses] [Pedadogy] [Forum] [Guest Book] [BYU]

Copyright© 2004 Linguistic Technologies, Inc.