0% found this document useful (0 votes)

36 views56 pages

A Dynamic Usagebased Model

Uploaded by

Yi Qi

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

36 views56 pages

A Dynamic Usagebased Model

Uploaded by

Yi Qi

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 56

Chapter 4

A dynamic usage-based model

This chapter and the next will address some foundational theoretical
issues.* The present chapter focuses on the very nature of linguistic
structure. It shows how a small set of basic psychological phenomena
operate in all domains, giving rise to patterns exhibiting any degree of
regularity. Chapter 5 will then address constituency, considering both
its cognitive status and its role in grammar.

1. The usage-based conception

In a usage-based model, "substantial importance is given to the actual
use of the linguistic system and a speaker's knowledge of this use; the
grammar is held responsible for a speaker's knowledge of the full
range of linguistic conventions, regardless of whether these con-
ventions can be subsumed under more general statements. [It is a]
nonreductive approach to linguistic structure that employs fully articu-
lated schematic networks and emphasizes the importance of low-level
Schemas"(FCG1: 494). The "maximalist", "non-reductive", "bottom-
up" nature of CG contrasts with the "minimalist", "reductive", "top-
down" spirit of generative theory, at least in its original (archetypal)
formulation. Let me start by briefly describing each property.
Generative theory has always tried to minimize what a speaker
has to learn and mentally represent in acquiring a language. Its mini-
malism was originally based on economy: the best grammar was the
one that did the job with the fewest symbols. In recent years, the em-
phasis has shifted to positing a richly specified universal grammar, so
that the role of experience in learning a language involves little more
than the setting of parameters. By contrast, CG accepts that becoming
a fluent speaker involves a prodigious amount of actual learning, and
tries to minimize the postulation of innate structures specific to
language. I consider these to be empirical issues. If one aims for psy-
chological reality, it cannot be maintained on purely methodological
grounds that the most parsimonious grammar is the best one. Should
it prove that the cognitive representation of language is in fact massive

and highly redundant, the most accurate description of it (as a psycho-

logical entity) will reflect that size and redundancy.1
The issue of reductionism pertains to the relation between
general statements and more specific statements that amount to special
cases of them. Suppose a speaker has learned both a general "rule"
(such as the pattern for combining prepositions with their objects) and
certain specific expressions which instantiate the pattern (e.g. for me,
on the floor, in the garage). Traditionally, in generative accounts, the
instantiating expressions would be excluded from the grammar on
grounds of economy. Since they are regularly derivable by rule, to list
them individually would be to miss a generalization. This reasoning
however rests on the spurious assumption that rules and lists are
mutually exclusive (the rule/list fallacy). There is a viable alternative:
to include in the grammar both rules and instantiating expressions.
This option allows any valid generalizations to be captured (by means
of rules), and while the descriptions it affords may not be maximally
economical, they have to be preferred on grounds of psychological
accuracy to the extent that specific expressions do in fact become esta-
blished as well-rehearsed units. Such units are cognitive entities in
their own right whose existence is not reducible to mat of the general
patterns they instantiate.
The "top-down" spirit of generative grammar is evident in its
emphasis on general rules and universal principles, as well as its
historical neglect of lexicon, low-level subpatterns, and the patient
enumeration of idiosyncrasies. Less-than-fully-general phenomena
were in fact embarrassing and problematic from the outset, handled
by a series of ad hoc devices (e.g. the "rule features" proposed in
Lakoff 1970a) appended to the rule-based system. Now certainly an
objective in CG is to capture whatever generalizations the data will
support. There are nonetheless several respects in which the frame-
work manifests a "bottom-up" orientation. For one thing, it recog-
nizes that linguistic patterns occupy the entire spectrum ranging from
the wholly idiosyncratic to the maximally general. In a complete
account of language structure, fully general rules stand out as being
atypical rather than paradigmatic. Another facet of CG's bottom-up
orientation is the claim that "rules" can only arise as schematizations
of overtly occurring expressions. However far this abstraction may
proceed, the Schemas that emerge spring from the soil of actual usage.
Finally, there is reason to believe that lower-level Schemas, expres-
sing regularities of only limited scope, may on balance be more essen-

tial to language structure than high-level Schemas representing the

broadest generalizations.
The assumptions made about mental abilities and cognitive
processing are both minimal and relatively non-controversial. A few
basic mechanisms are operative in all domains of language structure
and afford a unified account of phenomena traditionally handled sep-
arately and in very different ways. Together with the austerity they
entail in the positing of both psychological and linguistic entities,
these factors render a usage-based framework intrinsically desirable.

2. Psychological phenomena
I start by recognizing a number of basic and very general psychologi-
cal phenomena that are essential to language but certainly not limited
to it. The first of these, which I refer to as entrenchment, has also
borne such labels as "routinization", "automatization", and "habit
formation". The occurrence of psychological events leaves some kind
of trace that facilitates their re-occurrence. Through repetition, even a
highly complex event can coalesce into a well-rehearsed routine that is
easily elicited and reliably executed. When a complex structure comes
to be manipulable as a "pre-packaged" assembly, no longer requiring
conscious attention to its parts or their arrangement, I say that it has
the status of a unit. It is convenient notationally to indicate unit status
by means of boxes or square brackets, enclosing non-unit structures
with closed curves or parentheses: [A] vs. (A).
A second basic phenomenon, abstraction, is the emergence of
a structure through reinforcement of the commonality inherent in
multiple experiences. By its very nature, this abstractive process
"filters out" those facets of the individual experiences which do not
recur. We will mostly be concerned with a special case of abstraction,
namely schematization, involving our capacity to operate at varying
levels of "granularity" (or "resolution"). Structures that appear very
different when examined in fine-grained detail may nonetheless be
quite comparable in a coarse-grained view. A schema is the common-
ality that emerges from distinct structures when one abstracts away
from their points of difference by portraying them with lesser preci-
sion and specificity. I use a solid arrow for the relationship between a
schema and a more specific structure that instantiates or elaborates it:
A-»B. The formula indicates that B conforms to the specifications of
A but is characterized in finer-grained detail.

Also fundamental to cognition is the ability to compare two

structures and detect any discrepancy between them. This operation
involves an inherent asymmetry, whereby one structure functions as a
standard of comparison, the other as its target. We can reasonably
consider categorization to be a special case of comparison, obtaining
when the standard represents an established unit and the target (at
least originally) is novel. Categorization is most straightforward when
there is no discrepancy, i.e. when the standard can be recognized in
the target because the latter fully satisfies its specifications. In this
case the two structures stand in an elaborative relationship: [A]-»(B).
An act of categorization may also register some disparity between the
categorizing structure and the target. In this case I speak of extension,
indicated with a dashed arrow: [A]—>(B).
Yet another basic phenomenon is the combination of simpler
structures to yield a more complex structure. Let us call this com-
position. It involves the integration of two or more component struc-
tures to form a composite structure. If [A] and [B] are units, not pre-
viously combined, their integration to produce the novel composite
structure (C) can be given as follows: ([A][B])c· The formula should
not however be taken as implying that (C) is merely the union of [A]
and [B], nor that [A] and [B] occur unmodified in (C). When motor
routines are chained together into a complex action, their coordination
entails that no component routine is manifested in precisely the form it
would have in isolation; typing kl, for instance, is not just the same as
typing k then typing /. The same is clearly true of speech sounds, and
(I would argue) of most any kind of conceptual integration. A com-
posite structure has to be regarded as an entity in its own right, not
strictly reducible to it components. For this reason I speak of partial
compositionality.
Let us mention, finally, the well-known phenomenon of
association, in which one kind of experience is able to evoke another.
The particular kind of association that concerns us is symbolization:
the association of conceptualizations with the mental representations
of observable entities such as sounds, gestures, and written marks.
An established symbolic relationship—a symbolic unit—is conven-
iently given as [[A]/[a]], where upper and lower case stand respec-
tively for a conceptualization and a symbolizing structure. A symbolic
structure is said to be bipolar: [A] is the semantic pole, and [a] the
phonological pole (in the case of sounds).
While there may be differences in approach and terminology, I
consider it self-evident that something akin to each phenomenon has
to be ascribed to cognition generally and to language in particular. It

should also be evident that these operations occur in various combi-

nations, some applying to the results of others. Composition, for
example, is applicable to its own output—composite structures can in
turn function as components integrated to form a more elaborate com-
posite structure. Repeated episodes of composition yield constituency
hierarchies having indefinitely many levels of organization. Here is
another plausible sequence of operations: (Ai), (A2), (As) > [A] >
([A]-*(A4» > [[A]-*[A4]]. From a series of similar experiences,
represented as (Αι), (Α2), and (As), a schema emerges that embodies
their commonality and achieves the status of a unit, [A]. This struc-
ture is subsequently used to categorize a new experience, (A4), which
instantiates it. If (A4) recurs and continues to be recognized as an
instance of [A], both it and the categorizing relationship undergo en-
trenchment and gain unit status. [[A]-* [A4]] then constitutes an esta-
blished categorization.
I suggest that repeated applications of such processes, occur-
ring in different combinations at many levels of organization, result in
cognitive assemblies of enormous complexity. The vision that emer-
ges is one of massive networks in which structures with varying
degrees of entrenchment, and representing different levels of abstrac-
tion, are linked together in relationships of categorization, composi-
tion, and symbolization. I believe that all facets of linguistic structure
can be reasonably described in these terms.

3. Processing interpretation
The network model just presented deserves to be handled with cau-
tion, for like any metaphor it has the potential to mislead. In particu-
lar, the network metaphor encourages us to think of linguistic struc-
tures as discrete, object-like entities forming a static assembly observ-
able as a simultaneously available whole. All of these features are
problematic in regard to the neural implementationof language. From
the processing standpoint, language must ultimately reside in patterns
of neurological activity. It does not consist of discrete objects lodged
in the brain, and it cannot all be manifested at any one time. An im-
portant question, then, is whether and how these two perspectives can
be reconciled.
As a general orientation, I incline to the connectionist style of
computation based on parallel distributed processing (McClelland and
Rumelhart 1986; Rumelhart and McClelland 1986). This mode of
processing has the advantage of resembling what the brain actually

does, at least at a basic level, and I believe it to be both realistic and

revelatory with respect to language. I realize how enormous the gap is
between existing PDF models and a system that would approximate
the actual complexity of linguistic structure, even in limited domains.
This huge discrepancy reflects the fact that the PDF-style processing
constitutive of human language occurs in the context of a highly-
structured brain of unfathomable complexity, and draws upon the
structures that progressively emerge over the course of many years
through continuous and multifaceted interaction with a rich environ-
ment. Since actual connectionist systems are not embedded in such a
matrix, there are severe limitations on how close they can come to
realistically modeling linguistic structure. That is not per se an argu-
ment against connectionist-type processing, however.
Here I will merely try to indicate that the psychological pheno-
mena discussed in the previous section can all be given a connection-
ist interpretation. For analytical purposes, it is helpful (if not neces-
sary) to think in terms of discrete structures represented by distinct
symbols enclosed in brackets or boxes. Such reifications are not too
harmful so long as we do not lose sight of the dynamic reality they
conceal. In the final analysis, linguistic structures and relationships
reside in cognitive processing, identified as neurological activity.
Entrenchment is straightforwardly identifiable as an adjust-
ment in connection weights, brought about by the occurrence of a
pattern of activation, which renders more likely the re-occurrence of
the same or a comparable pattern. With respect to the system's move-
ment through state space, entrenchment amounts to the emergence of
an attractor. The term fits quite well with the topographical conception
of state space, where attractors of different strength are thought of as
wells or valleys of varying depth that the system tends to settle into as
it relaxes into a state of minimum energy.
Connectionist systems are well known for their ability to ex-
tract whatever regularities are inherent in their input. I have previously
discussed the extraction of schemas in connectionist terms (1990b;
FCG2: 12.3). Let me first reiterate that discrete representations such
as A-*B are not to be taken as implying that a schema and its in-
stantiation are wholly distinct and separately stored. Rather, I think of
schemas as being immanent in their instantiations, and while a schema
may in some cases be independently accessible (e.g. the schematic
notion common to apple, orange, banana, etc. is individually symbol-
ized by fruit), there is no supposition that this is true in general. In
saying that a schema is extracted, what is necessarily being claimed is
actually fairly minimal: that the commonality inherent in multiple ex-

periences is reinforced and attains some kind of cognitive status, so

that it has the potential to influence further processing.2
In offering a connectionist interpretation of schema extraction,
we can first equate a particular experience (or "structure") with either
a point in state space or a trajectory through it (cf. Elman 1990). To
the extent that two experiences are similar, their constitutive patterns
of neural activation will be neighbors in state space: either points in
close proximity, or trajectories following roughly parallel courses.
We can focus on the simpler case of points, since a trajectory reduces
to a series of points ordered in processing time. Suppose, then, that
the patterns representing a number of similar structures all cluster in
the same general region of state space. Call this region R. The occur-
rence of a given pattern will impact connection weights in such a way
that the occurrence of any pattern close to it in state space will tend to
be facilitated. The repeated use of similar structures will thus facilitate
the occurrence of any pattern within the general region R. This
amounts to the extraction of a schema. With reference to state space, a
schema is describable as a basin of attraction (R) which subsumes a
number of more point-like locations corresponding to its instantia-
tions. A schema is immanent in its instantiations in the sense that
being located in a point-like region of state space entails being located
in a broader region that encompasses it.
Categorization is then interpretable as capture by an attractor.
Presenting the system with a certain input tends to activate a variety of
previously established patterns, some of which may be mutually inhi-
bitory. When an input (B) results in the full activation of pattern
[A]—which may have won out over numerous competitors—we can
reasonably say that [A] is used to categorize (B). Of course, if the
input is only fragmentary, categorization via the activation of [A] may
serve to reconstitute the full, familiar experience it represents. The
categorizing experience will also be qualitatively different depending
on whether (B) is compatible with [A] or succeeds in eliciting [A]
despite some discrepancy between them. We will return to these mat-
ters in the following section.
With respect to composition, it can plausibly be suggested
that component structures are to some degree activated in the process
of carrying out the more elaborate pattern of activation constitutive of
the composite conception. Ultimately I would like to argue that com-
position reduces to categorization (FCG1: 12.2). I have already noted
that component structures are not appropriately conceived as "building
blocks" stacked together to form the composite structure. The latter is

an entity in its own right, facets of which correspond to the compo-

nents but either elaborate or diverge from their specifications. I
believe the component structures are properly thought of as categor-
izing those facets of the composite structure to which they corre-
spond. Like a categorizing structure, the components are in some
sense prior, most obviously in cases where they represent established
units while the composite structure is novel. Moreover, a composite
structure resembles the target of categorization in being the structure
of concern, the one being assessed in relation to others having some
kind of prior standing. Yet another resemblance between composition
and categorization is that in each case the quality of the target exper-
ience is partially shaped and (re)constituted by the structures activated
for its assessment.
Finally, symbolization is readily interpretable as one pattern of
activation reliably serving to elicit another. If semantic structures are
represented as patterns occurring in one bank of processing units, and
phonological structures in another, a connectionist system can easily
be trained to establish the proper correlations.

4. Basic linguistic problems

In CG, a language is described as a structured inventory of con-
ventional linguistic units. The units (cognitive "routines") comprising
a speaker's linguistic knowledge are limited (by a restriction called the
content requirement) to semantic, phonological, and symbolic struc-
tures which are either directly manifested as parts of actual expres-
sions, or else emerge from such structures by the processes of
abstraction (schematization) and categorization. In describing these
units as an inventory, I am indicating the non-generative and non-
constructive nature of a linguistic system. Linguistic knowledge is not
conceived or modeled as an algorithmic device enumerating a well-
defined set of formal objects, but simply as an extensive collection of
semantic, phonological, and symbolic resources that can be brought
to bear in language processing. This inventory of resources is struc-
tured in the sense that some units are incorporated as parts of larger
units, specifically in relationships of categorization, composition, and
symbolization: [[A]—>[B]]; [[A][B]]c; [[A]/[a]]. For some purposes
it may be helpful to reify linguistic knowledge or ability as something
called the "grammar" of a language. We must however resist the
temptation to think of it as a separate or sharply bounded cognitive
entity, if only because a structure's characterization as being conven-

tional, linguistic, or a unit is inherently a matter of degree (FCG1:

2.1).
Can a grammar of this sort actually do the job? We somehow
have to deal with a number of basic problems: the creation and un-
derstanding of novel expressions; the ascription of particular struc-
tures to such expressions; judgments of well- and ill-formedness;
distributional restrictions; and the varying degrees of compositional-
ity, productivity, and generality exhibited by linguistic structures and
patterns.

4.1. Categorization of usage events

It is not the linguistic system per se that constructs and understands
novel expressions, but rather the language user, who marshals for
this purpose the full panoply of available resources. In addition to
linguistic units, these resources include such factors as memory,
planning, problem-solving ability, general knowledge, short- and
longer-term goals, as well as full apprehension of the physical, social,
cultural, and linguistic context. An actual instance of language use,
resulting from all these factors, constitutes what I call a usage event.
the pairing of a vocalization, in all its specificity, with a conceptuali-
zation representing its full contextual understanding. A usage event is
thus an utterance characterized in all the phonetic and conceptual detail
a language user is capable of apprehending. For immediate purposes
it makes no difference whether we consider the speaker or the addres-
see, since each has to establish some connection between the lin-
guistic system and a usage event that supposedly manifests it. In
comprehension, the hearer has to interpret the event as the intended
realization of particular linguistic structures. In production, the
speaker has to select linguistic structures capable of evoking the
desired contextual understanding, and has to then ensure that the
event can indeed be so interpreted.
In both production and comprehension, therefore, facets of a
usage event must somehow be assessed in relation to linguistic units.
I take this to be a matter of categorization. Let us consider the minimal
case, in which a single linguistic unit, [A], is used to categorize a
particular facet, (B), of a usage event. There are two basic possibi-
lities, depicted in Figure 4.1, where the box labeled L represents the
linguistic system (i.e. the inventory of conventional units), and the
overall usage event is given as the circle labeled U. On the one hand,
[A] can be recognized in (B), which thus instantiates [A], as seen in

Figure 4.1 (a). This amounts to the judgment that (B) is well-formed
with respect to [A]. We can also describe it as being conventional, in
the sense of conforming to the linguistic convention embodied in [A].
On the other hand, there may be some discrepancy between the two
structures. In this case (B) is not perceived as an elaboration of [A],
but rather as an extension from it, as shown in Figure 4. l(b). This
amounts to the judgment that (B) is ill-formed, or non-conventional,
with respect to [A].

Figure 4.1

We have already noted the oversimplification inherent in rep-

resenting the linguistic system, L, as a discrete box. For example,
since entrenchment is a matter of degree, there is no clear line of de-
marcation between novel structures and those with the status of units.
Moreover, a boundary imposed at any particular threshold will con-
tinually fluctuate, since every use of a structure reinforces it and
entrenches it more deeply, whereas non-use has the opposite effect.
Even the first occurrence of a novel structure constitutes an initial step
along the path of progressive entrenchment and conventionalization,
for it must leave some kind of trace in at least one member of the
speech community. Suppose, then, that structure (B) begins to occur
with some frequency in a speech community, and that speakers
consistently invoke unit [A] to categorize it. For example, [A] might
be the basic meaning of a lexical item, and (B) a semantic extension,
as when mouse first started being applied to a piece of computer
equipment. With frequent recurrence, both (B) and ([A]—>(B)), i.e.
(B)'s categorization as an extension from [A], will become progres-
sively entrenched and eventually achieve the status of units. To the
extent that this happens, both [B] and the categorizing relationship
[[A]—>[B]] become members of L as a matter of definition. Figure
4.2 diagrams the expansion of L to incorporate these new conven-
tional units.

Figure 42

Actually, a slight refinement of Figure 4.2 should be noted.

Since (B) is part of a usage event, it represents a conceptual or
phonetic structure in the full detail of its contextual apprehension.
Numerous fine-grained details, as well as contingent features of the
context, are bound to vary from one usage event to the next. On
successive occasions, for example, the referents of mouse may be
slightly different shades of gray and occupy different positions vis-a-
vis the computer. Failing to recur consistently, such details will not be
reinforced and hence will not be included in the new conventional unit
that emerges. The categorizing judgments that occur on particular
occasions can thus be given as ([A]—>(Bi)), ([A]— >(B2»,
([A]—>(B3)), etc., where the subscripts indicate divergence in the
targets. The structures that undergo entrenchment and achieve the
status of conventional units are schematic relative to those which
figure in any actual usage event: [B] is schematic with respect to (Bi),
(82), and (63), and [[A]—>[B]] with respect to ([A]—>(Bi)),
([A]—>(B2)), and ([A]—>(B3>). The point is a general one-
linguistic units are always schematic in relation to their instantiations
in actual usage events.
Repeated occurrences of the processes sketched in Figures 4.1
and 4.2 can naturally be expected, and a unit added to L at any point
is then eligible to serve as a categorizing structure in subsequent
occurrences. Thus, from a single structure [A], there may eventually
develop an elaborate network comprising any number of conventional
units linked by categorizing relationships. These structures and rela-
tionships are said to form a complex category. To the extent that the
network consists of chains of extensions radiating outward from [A]
(thereby identifiable as the prototype), it constitutes a "radial cate-
gory" (Lakoff 1987).
While accepting the insight and basic validity of the radial
model based on extension (i.e. A—>B), I also emphasize schemati-
zation and relationships of instantiation (A-»B), if only because the
latter correspond to essential linguistic phenomena: the extraction of
generalizations, and judgments of well-formedness (conventionality).

The two kinds of categorization are in any case very intimately related
(and may in practice be hard to distinguish). Both involve an act of
comparison in which a standard (S) is matched against a target (T).
Instantiation can then be regarded as the special, limiting case of ex-
tension that arises when the discrepancy registered between S and T
happens to be zero. Conversely, if categorization is interpreted as the
attempt to "recognize" S in T, instantiation represents the privileged
case where this happens unproblematically, and extension constitutes
recognition accomplished only with a certain amount of "strain". The
source of the strain is that, for S to be recognized in a target which
does not fully conform to its specifications, the conflicting features of
S somehow have to be suppressed or abstracted away from.
Extension can thus be thought of as recognition achieved at
the cost of invoking a schematized version of the categorizing struc-
ture, one whose coarser-grained specifications are satisfied by the
target. For this reason I suggest that extension tends to be accompa-
nied by schematization, that the "outward" growth of a network by
extensions from a prototype tends to induce its "upward" growth via
the extraction of higher-level Schemas. The general mechanism is dia-
grammed in Figure 4.3.

Figure 4.3

I presume that extension does not occur at random; there is

always some basis for it. The categorization of (B) as an extension
from [A] implies some abstract commonality—however limited or
tenuous—which enables [A] to be evoked for that purpose in the first
place and to successfully categorize (B) despite a conflict in their
properties. By definition, that commonality amounts to a schema,
labeled (A') in the diagram, having both [A] and (B) as instantiations.
It is not necessary that (A1) be salient or separately apprehended, or
that it endure beyond its fleeting occurrence as an implicit facet of the
categorizing event ([A]—>(B)). Still, the very fact that [A] and (B)
occur together tends to reinforce their commonality and thus facilitates
(A1)'s emergence as an established cognitive entity. Should (A1) attain

the status of a unit, it is validly describable as both a schema instan-

tiated by [A] and (since the latter is prior) as an extension from it:

I assume, then, that linguistic categories are usually complex,

developing from prototypical structures via such processes as
extension, the extraction of Schemas, and the articulation of coarse-
grained units into more specific ones (as finer discriminations are
made and particular instantiations gain unit status). Bearing in mind
the limitations of the metaphor, we can view complex categories as
networks in which linguistic structures of any kind and any size are
linked in pairwise fashion by categorizing relationships (FCG1: ch.
10). These structures— the "nodes" or vertices of the network— might
consist, for example, of the allophones of a phoneme, the alternate
senses of a lexical item, a family of related metaphors, or variant
forms of an elaborate grammatical construction. There is more to such
a network, however, than just a set of nodes and a set of arcs
connecting them. Additionally, each structure and each categorizing
relationship has some degree of entrenchment and ease of activation.
Moreover, the target of categorization in each case lies at a certain
"distance" from the standard, depending on how far T elaborates S or
how many features of S it violates. Entrenchment and distance are
respectively indicated in Figure 4.4 by the thickness of boxes and the
length of arrows. In general, though, my diagrams will not attempt to
represent these parameters.

Figure 4.4

4.2. Selection of categorizing structures

At this juncture a basic problem arises. I have suggested that the units
constitutive of linguistic knowledge are related to actual expressions
by means of categorization, as shown in Figure 4.1. Yet even a single
category may well contain a large number of units all of which are in
principle available to categorize some particular facet of a usage event
They cannot all do so at once, for chaos would then ensue. A given
target is well-formed with respect to certain potential categorizing
units and ill-formed with respect to others. Unless these units are
recruited in some specific way to assess the target (on any one occa-
sion), there will be no basis for the clear judgments of well- and ill-
formedness that commonly occur. I assume, in fact, that primary cate-
gorization is effected by just one unit at any given moment. How,
then, does a particular target manage to be categorized by a single unit
selected from a large network of potential categorizing structures?
Let me first point out that this is a general problem in cogni-
tion, not specifically a linguistic one. Consider the recognition of a fa-
miliar face. Among others, I possess schematized representations of
both Suzanne Kemmer's face and Sydney Lamb's. When Suzanne
walks into the room, I usually manage to correctly recognize her as
Suzanne and not confuse her with Syd. To do so, I have to activate
the Kemmer schema for the categorization of my visual experience,
not the Lamb schema—otherwise I would see her as Sydney Lamb
and marvel at how much he had changed. Suzanne is, after all, a very
good instance of Suzanne, but a rather poor instance of Syd. Depen-
ding on which schema I activate, therefore, the episode of facial re-
cognition will yield a judgment of either well-formedness or deviance,
as sketched in Figure 4.5.

(b)

KEMMER

|LAMB| '

Figure 4.5

For both linguistic and non-linguistic input, I assume we have

to tell a certain kind of story. It has no claim to novelty, being
basically what is envisaged in the interactive activation model (Elman
and McClelland 1984; McClelland and Elman 1986a; McClelland and
Elman 1986b), the competition model (MacWhinney 1987), and for
that matter in connectionist processing generally. The story runs more
or less as follows. A particular target of categorization tends to
activate a variety of established units, any one of which could in
principle serve to categorize it. Let us call this set of units (which may
belong to a single complex category or to multiple categories) the
activation set of the target. Initially, as shown in Figure 4.6(a), the
members of the activation set are all activated by T to some degree.3
Only one member can actually categorize T, however, so in effect
they compete for this privilege; some are no doubt mutually inhibitory
and tend to suppress one another. One member of the activation set
eventually wins the competition in the sense of becoming highly
active relative to all the others. It is this unit—termed the active
structure—which serves to categorize the target, as seen in 4.6(b).

D D

Figure 4.6

Several factors interact to determine which particular member

of the activation set wins the competition and emerges as the active
structure evoked to categorize the target. The first is level of entrench-
ment, or inherent likelihood of activation. In a neutral context, for
example, Bali would more easily be misheard as belly than con-
versely. A second factor is contextual priming, which can override the
effects of familiarity. Suppose we are discussing our upcoming trip to
Bali and (my mind wandering) I happen to say—out of the blue—that
I would like to see some belly dancers. The context might well lead
you to misinterpret me as saying that I want to see some Bali dancers.

A third factor is the amount of overlap between the target and a

potential categorizing structure. We can reasonably assume that the
sharing of features is what enables the target to stimulate members of
the activation set in the first place, and that the degree of stimulation is
roughly proportional to the number of features shared. This has the
important consequence that lower-level Schemas, i.e. structures with
greater specificity, have a built-in advantage in the competition with
respect to higher-level Schemas. Other things being equal, the finer-
grained detail of a low-level schema affords it a larger number of
features potentially shared by the target.
We can now describe, in very general terms, how expressions
are evaluated with respect to the linguistic system. A usage event has
many facets susceptible to categorization by conventional linguistic
units, which in CG are limited to semantic, phonological, and sym-
bolic structures. The potential categorizing units are entrenched to
varying degrees and form a vast, structured inventory through relat-
ionships of symbolization, composition, and categorization. These
units can be of any size, represent any dimension of linguistic struc-
ture, and are characterized at all levels of specificity. By virtue of
overlapping content, each facet of the usage event serves to activate a
set of potential categorizing structures—its activation set—which then
compete for the right to categorize it. The winner of each competition
is determined by a dynamic interactive process on the basis of degree
of overlap, inherent ease of activation (correlated with entrenchment),
contextual priming, and mutual inhibition.
The resulting set of categorizations, in which the winners (the
active structures) categorize the facets of the usage event which elicit
them, constitute the expression's structural description (i.e. its inter-
pretation with respect to the linguistic system). The expression is fully
well-formed (conventional) provided that all of these categorizations
are elaborative in nature. It is ill-formed (non-conventional) to the ex-
tent that any of them involve extension rather than elaboration. We
can expect many extensions to pass unnoticed in normal language
use. It is only when a conflict is egregious, or when small conflicts
have a cumulative effect, that the strain they produce rises to the level
of conscious awareness (cf. Ross 1987).
In principle, then, a linguistic system conceived as being non-
generative and non-constructive can nonetheless support the ascrip-
tion of structural descriptions to expressions and provide the basis for
judgments of well- and ill-formedness. One need only envisage this
inventory of semantic, phonological, and symbolic resources as being

embedded in a dynamic processing system which operates in accor-

dance with minimal and highly plausible assumptions.

4.3. Categorization vs. construction

I have posed the question of whether the linguistic system per se

specifies in full detail how expressions are constructed (thus
characterizing a well-defined set of expressions as its "output"), or
whether it is merely an inventory of units invoked for the categori-
zation of usage events. Should responsibility for constructing
expressions be assigned to the "grammar", or to the language user
drawing upon a full range of psychological and contextual resources?
Though it may seem slight, this distinction has important conse-
quences for how we think about linguistic problems. At stake are
basic questions such as the scope of linguistic meaning, how rules are
related to instantiating expressions, and whether the linguistic system
constitutes a discrete, well-delimited cognitive entity.
The difference between construction and categorization can
first be illustrated by a simple case of semantic extension, e.g. the
aforementioned extension of mouse to indicate a piece of computer
equipment. Prior to the first occurrence of this usage, the linguistic
system (for a representative speaker of English) contained the sym-
bolic unit [[MOUSE]/[mouse]j, where the semantic structure given as
[MOUSE] designates a type of rodent, and [mouse] stands for the
phonological structure that symbolizes it. Consider now a speaker
who—for the very first time—faced a usage event in which the same
term was used in reference to a computer device. We can represent
this novel expression as ((MOUSE')/[mouse]), where (MOUSE1) is
the conception of the new referent. Now, in either producing or com-
prehending this novel usage, the speaker must somehow relate it to
the conventional unit [[MOUSE]/[mouse]], from which it derives via
the metaphorical extension ([MOUSE]—>(MOUSE')). It is evident
that the linguistic system per se cannot be responsible for constructing
the new expression ((MOUSE')/[mouse]), if only because the concept
(MOUSE1) is (by assumption) a novel one. It is clearly the speaker
who, from the context (e.g. seeing the device on a desk) and by
means of abilities that are not specifically linguistic, entertains the new
conception and apprehends the resemblance that motivates the exten-
sion. The role of the conventional unit [[MOUSE]/[mouse]] is not to
construct but simply to categorize the new expression, which it moti-
vates by serving as the basis for metaphorical extension. Of course,

once the usage becomes familiar and conventionalized, it is incor-

porated in the language as the new symbolic unit [[MOUSE1]/
[mouse]], in the manner of Figure 4.2.
In terms of this scenario, what can we identify as the meaning
of mouse at different stages? According to standard doctrine, its
linguistic meaning was simply [MOUSE] when first applied to the
novel conception (MOUSE1), since the latter was not yet a conven-
tional semantic value; its metaphorical understanding in the context of
the initial usage event lies beyond the scope of linguistic semantics.
Yet standard practice would accept [MOUSE1] as a conventional
meaning of mouse at the present time. When did this change in status
occur? When did (MOUSE1) go from being an extra-linguistic under-
standing of mouse to being one of its linguistic semantic values? Does
such a transition occur after one usage event? After seven? After m
usage events involving each of n speakers? We could certainly adopt a
threshold number to determine when new senses will be described as
"linguistic" and admitted to the mental lexicon. This would allow us
to maintain a strict dichotomy between linguistic and non-linguistic
meanings, consistent with the notion that a language is a discrete and
well-delimited cognitive entity. I submit, however, that any particular
threshold would be arbitrary, in which case the claim that linguistic
meanings are clearly distinguishable from contextual understandings
is vacuous, rendered true just as a matter of definition.
I prefer to view things in a rather different manner. The very
first time the term mouse is used in regard to a piece of computer
equipment, it is contextually understood by the interlocutors as refer-
ring to that device. It is also so understood (in appropriate contexts)
once the new sense is fully established as a conventional semantic
value. Since mouse is understood with the value (MOUSE1) from the
very outset, and winds up having [MOUSE'] as an indisputably
linguistic meaning, it seems pointless to say that (MOUSE1) was ever
non-linguistic (though it did start out being non-conventional). Stated
more precisely, and in positive terms, I would want to say the fol-
lowing. On the occasion of the initial usage event, mouse has the
meaning ([MOUSE]—>(MOUSE')), i.e. the conception of the com-
puter device construed metaphorically as a kind of rodent. Through
continued usage, this complex meaning undergoes progressive en-
trenchment and conventionalization, and eventually the metaphorical
value [[MOUSE]—>[MOUSE']] emerges as a fully conventional
meaning of mouse with unit status. The only thing special about the
initial usage event is that the linguistic meaning's prior entrenchment
and conventionality lie at the zero end of the scale. However, the very

first use starts to move it away from the endpoint, and to the extent
this happens it becomes part of the linguistic system.
The same holds for complex expressions involving gramma-
tical composition. Consider the use of printer to indicate a computer
output device. Its compositional meaning, i.e. the one predictable on
the basis of the V-er morphological pattern, is simply 'something that
prints'. Its conventional semantic value is far more elaborate: printer
designates an electronic device, of a certain approximate size, run by a
computer to record its output on paper, etc. These extra-compositional
specifications correspond to facets of its contextual meaning that were
no doubt present from the outset, eventually becoming entrenched and
conventional through their recurrence in usage events. They are un-
problematic because the V-er compositional pattern is not at any stage
responsible for constructing the semantically enriched expression, but
merely for its categorization. In this way an expression analyzed as
belonging to a particular grammatical construction can nonetheless
diverge from its specifications, by either elaboration or extension. Let
me note just in passing that the point is equally valid for syntax. What
we intuitively accept as the meaning of a clause or a sentence is
usually more elaborate than its compositional value, if not in conflict
with it. The contrast with morphology is simply that syntactic
expressions are less likely than single words to recur with non-
compositional meanings and establish themselves as conventional
units.

4.4. Composition
Let us take a closer look at how a non-constructive model deals with
complex novel expressions and their relation to established grammati-
cal patterns. In CG, complex expressions are described as assemblies
of symbolic structures. These assemblies consist primarily of compo-
sitional relationships, wherein two or more component symbolic
structures are integrated—semantically and phonologically—to form a
composite symbolic structure. For example, the component symbolic
units [[JAR]/{jar]] and [[LID]/[lid]] are integrated to form the com-
posite symbolic structure [[JAR LID]/[jar lid]]. An assembly of this
kind, involving composition at one level of organization, is a minimal
construction. Larger assemblies arise when the composite structure of
one minimal construction functions in turn as a component structure
in another, representing a higher level of organization (e.g. jar lid
might be pluralized to form jar lids). Naturally this can happen

repeatedly, at progressively higher levels, yielding composite sym-

bolic structures of ever greater semantic and phonological complexity.
Expressions of any size can thus be assembled.
Grammar consists of patterns for creating symbolic assem-
blies. In accordance with basic principles of CG (in particular the con-
tent requirement), these patterns can only assume the form of
schematized expressions: templates abstracted from a set of complex
expressions to embody whatever commonality is inherent in them.
Hence grammatical patterns are themselves assemblies of symbolic
structures comprising compositional relationships at various levels of
organization. These assemblies are directly analogous to their
instantiating expressions, except that the symbolic structures which
form them are more schematic. In particular, the schematic template
corresponding to a construction—a constructional schema— itself
resides in component symbolic structures integrated to form a com-
posite symbolic structure. For example, the compositional pattern
instantiated by jar lid, garage door, and countless other noun-noun
compounds is merely a symbolic assembly in which two schematic
nouns, [[A]/[a]] and [[B]/[b]], are integrated in a certain manner to
yield the composite symbolic structure [[AB]/[ab]] (Figures 1.5-6).
This entire symbolic assembly serves to categorize either a fixed or a
novel expression that instantiates it. Moreover, the global categorizing
relationship is resolvable into local categorizing relationships between
particular substructures. The categorization ofyar lid as an instance of
the noun-noun compounding pattern therefore subsumes the local
categorizations [[[A]/[a]]-[[JAR]/[jar]]], [[[B]/[b]]-[[LID]/[lid]]],
and [[[AB]/[ab]]-»[[JAR LID]/[jar lid]]].
Consider, then, the construction of a novel expression in
accordance with an established grammatical pattern. We may suppose
that at the time of the utterance the linguistic system L contains the
various conventional units indicated in Figure 4.7(a). One such unit is
a constructional schema comprising two component symbolic struc-
tures, [A] and [B], as well as the composite symbolic structure, [C].
We can further suppose the existence of two lexical items, [A1] and
[B1], which respectively instantiate the schematic units [A] and [B].
Assuming that these instantiating relationships represent established
categorizations, they constitute the categorizing units [[A]-»[Ar|] and

(a)

Figure 4.7

Together, these units implicitly define a complex expression,

assumed to be novel, which in Figure 4.7(b) is thus surrounded by a
closed curve (rather than a box) to indicate its non-unit status. The
latent composite structure, (C), represents the expression's composi-
tional value: the structure that emerges if [A'] and [B1] are integrated
precisely as the constructional schema specifies, at both the semantic
and the phonological poles. In this case (C) elaborates [C], and the
complex expression ( [A '][B'])c' elaborates the constructional schema

In referring to (C1) as merely "latent", and saying that the con-

ventional units "implicitly" define the complex expression, I am once
more emphasizing the non-constructive nature of a grammar. Novel
expressions are not created by the linguistic system per se, but rather
by the speaker, drawing on all available resources. Of course, the
distinction is of little moment provided that we confine our attention to
the (possibly hypothetical) situation of full compositionality. On the
assumption of full compositionality, both (C1) and the entire assembly
([A'][B'])c· are wholly prefigured by conventional units of L: [A1]
and [B!] are established units, in which the schematic units [A] and
[B] are respectively immanent; the constructional schema [[A][B]]c
represents an established pattern for integrating [A] and [B] to form
[C]; therefore, constructing ([A'][B'])c'— with composite structure
(C)— is simply a matter of carrying out the established pattern of
integration when [A] and [B] are embedded in the more elaborate
structures [A1] and [B1]. In effect, the potential structure ([A'][B'])c'
is rendered actual just by (»activating the conventional units in ques-
tion. I call this actualization (FCG1: 11.3).

I have argued, however, that full compositionality is unchar-

acteristic of normal language use. In a typical usage event, the con-
textual understanding and phonetic rendition of a complex expression
diverge from its compositional value if only by being more specific,
and often in more drastic ways. This is shown in Figure 4.8, where
the expression's contextual value, (C"), is depicted as an extension
vis-ä-vis its compositional value, (C1). To be sure, those facets of the
expression that are extra-compositional (i.e. the discrepancies be-
tween (C1) and (C")) start out as being non-conventional, but it is
only by terminological fiat that they are also considered non-
linguistic, for the usage event begins their conventionalization. In-
deed, prior to actualization in a usage event, (C) itself lies beyond the
scope of established convention, a non-unit prefigured by L but
unexploited and unfamiliar to speakers.

Figure 4.8

Because (C") represents the way the expression is actually

understood (and may eventually become its conventional meaning, as
we saw in the case of printer), it is only (C1) whose status is reason-
ably called into question. I believe it does have both linguistic status
and some kind of cognitive presence. Since it represents the latent
potential inherent in an assembly of conventional units, we can
reasonably suppose that (C) is activated when those units are
invoked in a usage event. It thus embodies whatever motivation
conventional units provide for (C11) and serves as a kind of stepping-
stone on the way to it. The expression's actual value, then, is neither
(C1) alone nor (C"), but (C") construed in relation to (C), which
categorizes it: ((C)—>(C")). I would say, for example, that the
compositional value 'something that prints' does figure in the mean-
ing of printer as a computer term. More obviously, a metaphorical

expression like chopper (for 'helicopter') retains the compositional

value 'something that chops' as a secondary facet of its meaning; an
expression is metaphorical just by virtue of construing the target
domain against the background of the source domain. If (C) and the
categorizing relationship ((C')—>(C")) do indeed have some cogni-
tive presence, they might well retain it as the expression coalesces into
an established unit, as shown in Figure 4.9.

Figure 4.9

4.5. Degrees of regularity

If all grammatical patterns reside in constructional Schemas, they
nonetheless vary considerably in the nature and extent of their regular-
ity. At least three parameters need to be distinguished: composition-
ality, generality, and productivity. While there is some tendency for
these properties to be associated with syntactic patterns, and their
opposites with morphological ones, I see no empirical grounds for
believing that position along these scales correlates in any absolute
way with whether a pattern obtains above or below the word level.
The lack of absolute correlations is one reason for a basic claim of
cognitive grammar, namely that morphology and syntax form a con-
tinuum (fully describable as assemblies of symbolic structures). Nor
is grammar distinct from lexicon, defined in the theory as the set of
"fixed" expressions in a language (i.e. expressions with unit status),
regardless of size or type.
Compositionality pertains to how closely an expression ap-
proximates the result predicted for the integration of particular compo-
nent structures in the manner specified by a constructional schema. To
be completely compositional, therefore, an expression can exhibit no

discrepancy between its predicted value and its actual composite struc-
ture. In terms of Figures 4.7-9, the complex expression ([A'][B'])c'
is compositional with respect to pattern [[A][B]]c just in case the
predicted value (C) and the contextual value (C") precisely coincide;
their identity entails that the composite structure ((C1)—>(C")) col-
lapses onto (C). Previous discussion raises the question of whether
an expression is ever completely compositional. One can plausibly
argue that an expression's contextual understanding always diverges
to some extent (however minimally) from its predicted value, and that
a residue of such divergence is retained even when the expression
coalesces into an established lexical unit. However, in this non-
constructive framework nothing much hinges on whether the limiting
case of zero divergence is ever actually attained.
As defined, degree of compositionality is not a property of
grammatical patterns per se, but rather of particular expressions that
they categorize. The other two parameters do pertain to the patterns
themselves and are thus reflected in the constructional Schemas which
embody them. Generality relates to the level of specificity at which
such Schemas are characterized. A given pattern has greater or lesser
generality depending on whether it is potentially applicable to a wider
or a narrower range of elements. In English, for example, tense
marking is applicable to essentially any verb, whereas only "perfec-
tive" verbs enter into the progressive construction (Langacker 1987c).
Constructions can also be limited to smaller classes that I likewise
consider to be semantically definable, e.g. change-of-state verbs,
verbs of transfer, "unaccusative" verbs, etc. Limitations of this sort
are readily accommodated by the proper formulation of constructional
Schemas. Thus, whereas the Schemas describing tense marking iden-
tify one component structure just as a verb, the progressive schema is
more specific by requiring a perfective verb in particular. There is no
inherent limit to the level of specificity at which a constructional
schema can characterize its components. Indeed, a schema can incor-
porate a particular lexical item, even a specific variant of a lexical
item, as one of its component structures.
Productivity is a matter of how available a pattern is for the
sanction of novel expressions. Though productivity tends to correlate
with generality, they are often dissociated and have to be distin-
guished. Patterns of comparable generality can easily differ in their
degree of productivity. For example, a causative construction applic-
able (say) to intransitive verbs might be fully productive, applying to
any such verb if the result is semantically coherent, or it might be
limited to particular lexical combinations and unavailable for the crea-

tion of new expressions. Conversely, a pattern representing any level

of generality has the potential to be fully productive. For instance, a
certain plural marker might be restricted to a small class of nouns
(such as animal names) but be freely applicable to any noun in that
class.
In a usage-based model with dynamic processing (of the sort
described previously), productivity amounts to likelihood of being
selected as the active structure used to categorize a novel expression.
The constructional schema representing a highly productive pattern
must be well-entrenched and easily activated for this purpose. The
schema representing a non-productive pattern presumably expresses a
valid generalization, but if it cannot compete successfully for selection
as an active structure, this pattern cannot extend beyond the range of
data from which it is extracted. For example, English past-tense
formation subsumes both the productive, default-case pattern with the
conditioned variants [-d], [-t], and [-ad] (as in rowed, kicked, and
goaded), and also a variety of largely non-productive patterns restric-
ted to fixed sets of verbs, such as the ablaut pattern which changes
everything after the initial consonant cluster to [Dt] (as in bought,
brought, caught, and a number of others). We must therefore posit a
deeply entrenched, easily elicited past-tense schema that we can
abbreviate as [[V/...][PAST/-D]], whose phonological pole specifies
the suffixation of an alveolar stop, as well as various non-salient
Schemas, among them [[V/C(C)...][PAST/C(C)ot]]. We must further
assume their difference in salience to be such that, even though they
are both quite non-specific in their characterization of the verb stem,
the former will virtually always win the competition for the privilege
of categorizing a novel form.
Suppose, then, that leached is offered as the past-tense form
of leach. Its interpretation as a past-tense form tends to activate both
[[V/...][PAST/-D]] and [[V/C(C)...][PAST/C(C)Dt]] as possible cate-
gorizing Schemas. By assumption, f[V/...][PAST/-D]] will win the
competition and be selected as the active structure, so the expression
is judged well-formed (since leached conforms to its specifications).
On the other hand, suppose that lought is offered as the past-tense
form. While this does conform to a pattern in the language, namely
[[V/C(C)...][PAST/C(C)3t]], that alone is not sufficient to guarantee
its well-formedness. A judgment of well-formedness additionally re-
quires that the schema an expression instantiates be in fact selected as
the active structure with respect to which the assessment is made. The
situation in English is such that usually the default-case pattern sup-

presses less salient patterns even when the latter have a greater degree
of overlap with the target.4 As a consequence, lought will not be cate-
gorized by [[V/C(C)...][PAST/C(C)Dt]], but rather by [[VI...]
[PAST/-D]], whose specifications it violates. Only established past-
tense forms like bought, brought, caught, etc. are accepted as well-
formed. Since they are themselves entrenched conventional units,
they require no sanction from constructional Schemas.
I conclude that a usage-based model with dynamic processing
is able in principle to accommodate the full range of regularity
encountered in natural language. Degree of compositionality is free to
vary owing to the non-constructive nature of constructional Schemas
(whose role is merely to categorize target expressions), while gener-
ality and productivity are respectively determined by the level of
specificity at which such Schemas are characterized and their ease of
selection as an active (categorizing) structure. It should be emphasized
that nothing precludes the emergence of patterns that are highly
general and fully productive. For example, a schema that we can
abbreviate as [[V][NP]] might describe the semantic integration of a
verb with an object noun phrase and specify phonologically that this
NP immediately follows V in the temporal sequence. The pattern has
full generality: since V and NP are schematic characterizations, it
refers to the combination of any verb with any noun phrase. The
pattern is productive to the extent that entrenchment assures its activa-
tion in preference to any lower-level constructional Schemas making
conflicting specifications. A dynamic usage-based model is therefore
perfectly capable of handling productive general rules whose appli-
cation is exceptionless for all intents and purposes.
At the same time, CG agrees with construction grammar (e.g.
Fillmore, Kay, and O'Connor 1988; Goldberg 1995) in viewing such
rules as special and actually rather atypical cases in the overall spec-
trum of linguistic patterns, most of which exhibit some lesser degree
of generality and/or productivity. Even with respect to word order,
there will usually be alternatives to the basic pattern that are able to
preempt it in specific circumstances. A language might have, for ex-
ample, both the general constructional schema f[V][NP]] and also the
more specific schema [[PRON][V]], which describes the verb's sem-
antic integration with an object pronoun and specifies phonologically
that a pronominal NP precedes V in the temporal sequence rather than
following it. Assuming that [[V][NP]] and [[PRON][V]] are compar-
able in their degree of entrenchment, it is [[PRON][V]] that will be
elicited to categorize an expression with a verb and an object pronoun,

by virtue of its greater overlap with the target. The sequence PRON V
will thus be judged grammatical, and V PRON ill-formed, despite the
fact that the latter conforms to the higher-level schema [[V][NP]]. The
pattern described by this high-level schema is rendered less produc-
tive by the existence of a more specific pattern that preempts it.

4.6. Distribution

Like most linguistic phenomena, grammatical patterns usually arrange

themselves in complex categories comprising numerous related vari-
ants (see, for example, Lakoff 1987: case study 3). Such families of
patterns are describable as networks, as in Figure 4.4, where each
node in a network consists of an entire constructional schema. The
patterns are thus characterized at different levels of specificity, some
are special cases of others (constructional subschemas), some consti-
tute extensions relative to more prototypical variants, and so on. At
the extreme, the lowest-level subschemas in such a network incorpo-
rate particular lexical items (even lexical variants) as component
structures. If, for example, the schema [[V][NP][NP]] describes the
English ditransitive pattern in general terms, the constructional sub-
schema [[SEND/send][NP][NP]] represents the lower-level general-
ization that send in particular conventionally occurs in this construc-
tion (e.g. I sent my mother a birthday card).
It is perhaps more obvious that particular instantiations of
morphological patterns have the status of conventional units. The
English past tense, for instance, requires a family of constructional
Schemas including the category prototype, [[V/...][PAST/-D]], as
well as various Schemas representing minor patterns, e.g.
[[V/C(C)...][PAST/C(C)Dt]], which can be regarded as extensions
from the prototype. Clearly, it is part of a speaker's conventional
knowledge of the language that particular verbs like buy, bring, and
catch occur in this latter pattern. This knowledge takes the form of
lower-level subschemas in which these specific stems function as
component structures: [[BUY/buy][PAST/C(CX>t]]; [[BRING/bring]
[PAST/C(C)Dt]]; [[CATCH/catch][PAST/C(C)jt]]. In fact, although it
might not be apparent from this notation (which omits the composite
structures), these subschemas are nothing other than the specific
forms bought, brought, and caught. Experimental evidence suggests
that instantiations of even productive morphological patterns are

stored as units provided that they occur with sufficient frequency

(Stemberger and MacWhinney 1988).
It is in this manner that a usage-based framework accommo-
dates distributional restrictions. The fact that send participates in the
ditransitive construction is not indicated by means of an arbitrary
device such as a "rule feature" or a diacritic, but merely by the inclu-
sion of the constructional subschema [[SEND/send][NP][NP]] in the
conventional units comprising the linguistic system. The fact that buy
occurs in the morphological pattern [[V/C(C)...] [PAST/C(C)Dt]] is
likewise given by the inclusion in L of the instantiating subschema
[[BUY/buy][PAST/C(C)Dt]] (i.e. bought). I conclude that idiosyncra-
sies such as these are readily described in a theory that posits only
assemblies of symbolic structures for the characterization of lexical
and grammatical structure.5
Moreover, the examples illustrate the "bottom-up" orientation
of CG and the observation that lower-level Schemas, expressing
regularities of only limited scope, may on balance be more essential to
language structure than high-level Schemas representing the broadest
generalizations. A higher-level schema implicitly defines a large
"space" of potential instantiations. Often, however, its actual instan-
tiations cluster in certain regions of that space, leaving other regions
sparsely inhabited or uninhabited altogether. An adequate description
of linguistic convention must therefore provide the details of how the
space has actually been colonized. Providing this information is an
elaborate network of conventional units including both constructional
subschemas at various levels and instantiating expressions with unit
status. For many constructions, the essential distributional informa-
tion is supplied by lower-level Schemas and specific instantiations.
High-level Schemas may either not exist or not be accessible for the
sanction of novel expressions.
A simple example is provided by postpositional phrases in
Luiseno (a Uto-Aztecan language of southern California). In this
language postpositions occur suffixed to either inanimate nouns or
pronouns, as in ki-yk 'to the house' and po-yk 'to him', but not to
animate nouns: *hunwu-yk 'to the bear'. For the latter we instead find
expressions like hunwut po-yk (bear it-to), where the postposition
attaches to a coreferential pronoun. Forms like ki-yk 'to the house'
give rise to the constructional schema [Ninan-P], and those like po-yk
'to him', to the schema [PRON-P]. From these two patterns, more-
over, the higher-level schema [N-P] is presumably capable of emerg-
ing to embody the generalization that postpositions attach to nouns of

any sort: they occur on both pronouns and non-pronouns, and they
are not limited to inanimates (since the pronouns are usually animate
in reference). [N-P] is thus an expected outcome of the usual process
of abstraction, whereby commonalities are reinforced and points of
divergence effectively cancel out. Additionally, forms like hunwut
po-yk 'to the bear' permit the extraction of the more complex con-
structional schema [Nan [PRON-P]], which incorporates [PRON-P]
as a component structure.
It is readily seen that the crucial distributional information
resides in the lower-level Schemas [Ninan-P], [PRON-P], and [Nan
[PRON-P]]. If the high-level schema [N-P] were accessible for the
categorization of novel forms, expressions like *hunwu-yk 'to the
bear', which conform to its abstract specifications, would be accepted
as conventional. We must therefore suppose that [N-P] always loses
the competition to be selected as the active structure; it is consistently
superseded by the lower-level Schemas as a function of its own non-
salience and the inherent advantage accruing to more specific struc-
tures through their greater overlap with the target. Hence a form like
hunwu-yk 'to the bear' would not be categorized by [N-P], but rather
by either [Ninan-P], [PRON-P], or [Nan [PRON-P]], all of whose
specifications it violates.
We can say that the space of potential structures defined by the
high-level generalization [N-P] is only partially inhabited. In particu-
lar, the region corresponding to expressions with non-pronominal
animate nouns is completely unoccupied; the notions potentially coded
by forms in this region are instead handled by another, more complex
construction, namely [Nan [PRON-P]]. A constructive model might
account for this unexpected "gap" in the general pattern by positing a
rule which transforms the non-occurring "underlying" forms into
those which actually surface in their stead: [Nan-P]==>[Nan [PRON-
P]]. Alternatively, one could remove the non-occurrent forms from
the grammar's output by means of a filter: *[Nan-P].6 We have just
seen, however, that a dynamic usage-based model straightforwardly
accommodates the data without resorting to either filters or underlying
structures. The distributional gap simply results from the existence of
[Nan [PRON-P]] as a possible sanctioning unit, and the non-existence
of [Nan-P]· That in turn reflects the respective occurrence and non-
occurrence in the input data of expressions like hunwut po-yk and
*hunwu-yk. The Schemas speakers extract are those supported by the
expressions they are exposed to.

If it is workable, a theory that does not posit filters or

derivations from underlying structures should definitely be preferred.
Their avoidance simplifies the problem of language acquisition, which
in essence then reduces to reinforcement of the commonality inherent
in expressions that actually occur. I emphasize in particular that des-
criptions comprising only positive statements of what does occur are
in principle able to account for distributional gaps. To be sure, a
systematic attempt has not yet been made in CG to show in precise
detail how every known type of distributional restriction could be
dealt with, and every proposed filter eliminated. The working hypo-
thesis that only positive specifications are needed could be weakened
if necessary without undermining the essential claims of the theory.
Yet I see little reason to doubt that appropriate arrays of constructional
Schemas, varying in their degree of specificity and ease of activation,
are capable of handling actual distributional phenomena
This expectation extends to general constraints, such as those
advanced for movement rules (Ross 1967b [1986]; Chomsky 1973)
and pronoun-antecedent relationships.7 Such constraints pertain to
constructions in which corresponding entities lie at some distance, so
that in their most general form the patterns describe the intervening
material only in schematic terms. When the corresponding entities
occur in certain "structural" configurations—e.g. when one occurs
inside a tensed clause, or when they are separated by more than one
boundary of a certain kind—deviance ensues, the details being in
some measure language-specific. It is reasonably supposed that the
restrictions partially reflect processing difficulties associated with
simultaneously activating particular sorts of complex structures and
selectively accessing their substructures (Deane 1991; Kluender
1992). Still, the processing limitations are not absolute, and a given
language manifests them in conventionally determined ways.
In CG, the constructions in question are described by families
of constructional Schemas characterized at varying levels of specifi-
city. A relative clause construction, for example, will have multiple
variants differing as to whether the argument corresponding to the
head noun is a subject, an object, or has some other grammatical rela-
tion within its clause, whether that clause combines directly with the
head or is part of a larger clause which does so, whether the clause is
finite or non-finite, and so on. The highest-level constructional sche-
ma may define a vast space of structural possibilities, but occurring
expressions will not be distributed evenly within it. As with the other
kinds of patterns described above, constructional subschemas specify

which regions of that space are actually used, and with what degree of
likelihood. If well-entrenched subschemas sanction particular confi-
gurational relationships between the corresponding entities, they can
consistently win out over higher-level Schemas for the privilege of
categorizing novel expressions. Configurations not covered by the
subschemas will consequently result in judgments of ill-formedness.
It can even happen that comparable sets of configurational
relationships become conventionally established for multiple construc-
tions (e.g. for multiple "extraction rules"). If, in one construction,
speakers learn to effect a dependency between two elements in a
particular kind of structural configuration, that itself constitutes a
pattern which might be extended to other constructions. For instance,
once a speaker learns to make a correspondence between the object
argument in a finite clause and a nominal in the clause containing it
(say for relative clauses), it might subsequently be easier to make an
analogous correspondence in another type of construction (e.g. in
clefting). Conventionalized dependencies of this sort can themselves
be represented as constructional Schemas which abstract away from
the differences between the types of constructions involved. Thus,
although a detailed study has not yet been undertaken, I believe that
even such "parameter setting" is susceptible to characterization in a
dynamic usage-based model.

5. Structural applications
The usage-based model described above is applicable to all domains
of language structure: semantics, phonology, lexicon, morphology,
syntax. A linguistic system comprises large numbers of conventional
units in each domain, and a target expression is simultaneously
categorized by numerous active units, each assessing a particular facet
of its structure. A few basic psychological phenomena (listed in
section 2), applying repeatedly in all domains and at many levels of
organization, give rise to structures of indefinite complexity, which
categorizing relationships—each pertaining to a particular structural
dimension—link into cross-cutting networks. A description of this
sort is further unified in that seemingly diverse phenomena are seen as
residing in different aspects of the same or comparable structural
assemblies, or the same aspects "viewed" in alternate ways.

5.1. Lexicon and grammar

CG itself offers conceptual unification. It posits only semantic, pho-

nological, and symbolic structures. Lexicon, morphology, and syntax
form a gradation claimed to be fully describable as assemblies of
symbolic structures. The distinction between grammatical rules and
symbolically complex expressions is only a matter of whether (or the
degree to which) the symbolic assemblies constituting them are
schematic rather than specific. While there is some tendency for
morphological and syntactic rules to differ in terms of generality and
productivity, the only consistent basis for distinguishing them is
whether the phonological composition they specify takes place within
a word or involves word sequences. Expressions constructed in
accordance with grammatical Schemas can also be of any size. With
repeated use, an expression of any size or degree of compositionality
can be entrenched and conventionalized. The lexicon of a language is
then definable as the set of expressions with the status of conventional
units.
Constructional Schemas and complex lexical items both con-
sist of symbolic assemblies with unit status, often comprising com-
ponent and composite symbolic structures at multiple levels of organi-
zation. The reason for referring to such an assembly as a rule or con-
structional schema, rather than as a lexical item, is the incorporation
of one or more symbolic components too schematic—especially
phonologically—to actually be expressed as such. There are however
degrees of schematicity, even at the phonological pole, and in a com-
plex structure different numbers of components can be characterized
schematically. The fixed expression crane one's neck would generally
be considered a lexical item, yet the possessive element is actually
schematic: one's is just a placeholder for my, your, his, etc., all of
which are monosyllabic. Does crane one's neck count as a gramma-
tical pattern instead of a lexical item by virtue of this schematic
component? What about X take Υ over X's knee and spank Y, which
is schematic in several positions? If these are still considered lexical
rather than grammatical, there is no evident reason why a construc-
tional schema that incorporates a specific element, e.g. [[send][NP]
[NP]], should not also be a lexical item. That in turn is only one step
away from according lexical status to assemblies like [[V][NP][NP]],
all of whose components are schematic. My point, of course, is that
lexicon and grammar grade into one another so that any specific line
of demarcation would be arbitrary.

To make the same point in another way, let us consider more

carefully the status of the ditransitive pattern [[send][NP][NP]], i.e.
the commonality inherent in complex expressions like send me a
package, send your mother an eviction notice, send Washington a
message, etc. On the left in Figure 4.10, enclosed in an ellipse, is a
fragment of the network of constructional Schemas and subschemas
constituting conventional knowledge of the English ditransitive con-
struction. Verbs of transfer function as a prototype giving rise to
various extensions (Goldberg 1992). Subschemas specify the occur-
rence of particular verbs in this pattern, give and send of course being
common and well-entrenched. Special cases of these subpatterns may
themselves be established as familiar units, e.g. [[give] [me][NP]]
(note the contraction gimme).

TRANSFER||NP||NP

Figure 4.10

At the same time, however, the subschema [[send][NP][NP]]

represents a lexical property of send and belongs to a network of
constructional Schemas describing its grammatical behavior. Some of

these are shown in the ellipse on the right in Figure 4.10. In both
construction grammar and CG, a lexical item's characterization in-
cludes a set of "structural frames" in which it conventionally occurs.
While comparable in function to the "syntactic features" used in gen-
erative theory to specify the permitted contexts of lexical insertion,
these frames are actually just partially schematic symbolic assemblies
representing the commonality of certain complex expressions. They
are, moreover, inherent and essential to a lexeme's value. Lexical
items arise through a process of progressive decontextualization,
where non-recurring aspects of usage events are filtered out through
lack of reinforcement. Part of the relevant context is their occurrence
in larger symbolic assemblies. To the extent that a form like send has
any cognitive status independently of the structural frames in which it
appears, it emerges by abstraction from these larger assemblies.
Figure 4.10 should not be read as indicating that send is a distinct ele-
ment which merely happens to be incorporated in a set of construc-
tional subschemas. Rather, it is immanent in these assemblies and
apprehended as a separate entity only by suppressing them.
What, then, is the status of [[send][NP][NP]]? Does it belong
to the ditransitive construction or to the lexical item sendet The
answer, of course, is that the question is wrong: it is simultaneously
part of both. Viewed in relation to the construction, it constitutes a
subschema helping to specify the conventional distribution of a more
general grammatical pattern. Viewed in relation to the lexical item, it
specifies one grammatical environment in which the form occurs. In
the present model, it is unproblematic (and certainly usual) for the
same element to participate in multiple networks, which thereby
intersect.

5.2. Lexical semantics

A lexeme is not precisely the same in all its environments. Since

elements are always shaped by the contexts in which they occur, it is
only by abstracting away from contextual variation that a constant
representation emerges. Send is thus shown in Figure 4.10 as having
the contextual variants [sendi], [senc^], and [send3]. In particular,
[sendi] chooses the recipient as its direct object—defined in CG as a
participant receiving a secondary degree of "focal prominence"—and
further highlights the resultant possessive relationship. On the other
hand, [send2] confers object status on the mover and highlights the
path it follows, whereas [sends] downplays both the mover and the

recipient, focusing instead on the entity the sender hopes to obtain. It

is I think pointless to ask whether these differences in relative promi-
nence are responsible for the variants occurring in distinct structural
frames, or whether the frames themselves induce the differences by
virtue of what they explicitly encode. In any case a variant enters into
a kind of "ecological system" with its structural context and does not
necessarily exist outside that habitat. I am suggesting that these
context-dependent variants may be more fundamental than the
context-neutral schematization we tend to regard as primary.
In a "bottom up" account of this sort, the polysemy of lexical
items should be expected as the normal state of affairs. Whether the
contexts are structural, collocational, or pragmatic, they inevitably
shape the construal of symbolic structures and thus give rise to
semantic variants. Polysemy results when multiple variants become
entrenched as units, provided of course that some connection is
established between them (otherwise we speak of homonymy). Often
a particular variant is both prior and sufficiently salient to serve as the
basis for extension to other contexts, in which case we anoint it as the
lexical item's prototypical semantic value. Through the reinforcement
of common features, Schemas emerge at different levels of abstraction
to represent the commonality inherent in sets of variants. These
alternate semantic values constitute a complex category describable as
a network (as in Figure 4.4).
A classic problem of lexical semantics is whether an expres-
sion is truly "ambiguous", so that we must indeed posit two senses,
or whether it is only "vague", in which case there may be just one. In
practice the line is often hard to draw, with standard tests (Zwicky
and Sadock 1975) failing to produce a clear-cut distinction. Consider
the question of whether the verb paint is ambiguous between desig-
nating an artistic endeavor and a utilitarian one, or whether it is merely
vague in this regard. One way to test this is by ascertaining whether a
sentence like Bill has been painting and so has Jane is semantically
coherent or anomalous when the two clauses are construed as differ-
ing on this point. Thus, if Bill is painting a portrait while Jane is
putting lines on a highway, the sentence feels zeugmatic. This sug-
gests two distinct meanings, whereas the anaphoric expression so has
requires that the two clauses be semantically parallel. However,
Tuggy (1993) has argued convincingly that judgments like these are
often graded, making the test results indeterminate. The above exam-
ple is less zeugmatic if Jane is instead painting a wall, and virtually
normal if Bill is covering a wall for artistic purposes (say with a mu-
ral) and Jane for purely utilitarian purposes.

Tuggy shows that graded judgments like these, as well as

clear-cut assessments of vagueness or ambiguity, are expected and
readily accommodated in a dynamic usage-based model using net-
works for the description of complex categories. The nodes in such a
network vary in their entrenchment and ease of activation, hence in
the extent of their accessibility for specific grammatical purposes. The
issue of vagueness vs. ambiguity hinges on the relative status of three
structures: the putative specific senses (e.g. 'paint for artistic pur-
poses' and 'paint for utilitarian purposes'), and also the schematic
meaning representing their commonality ('paint'). Each structure is
established to some degree as a conventional meaning of the lexical
item in question. The possibilities range from being relatively un-
familiar and lacking unit status, at one extreme, to being very well-
entrenched and easily elicited, at the other.

A
AMBIGUITY VAGUENESS

Figure 4.11

As shown in Figure 4.11, clear-cut cases of ambiguity are

those where only the specific senses are entrenched and accessible. If
paint were like this, a neutral sense would be unavailable for ana-
phora, and one could not say Bill has been painting and so has Jane in
mixed circumstances comparing artistic and utilitarian intent. Con-
versely, definite cases of vagueness are those where only the sche-
matic meaning is entrenched and accessible. If this were true of paint,
it should always be felicitous to use the sentence in mixed circum-
stances (even with portraits and lines on a highway). The actual
situation appears to fall somewhere in between. When the two in-
stances of painting are similar enough, the neutral value is able to
emerge for anaphoric purposes. However, egregious differences call
attention to themselves and make it harder to suppress the specific
senses in favor of the neutral one. Thus, if Bill has been painting a

portrait, Jane's having done so with lines on a highway can only be

zeugmatic.
Comparable differences in salience and likelihood of activation
are responsible for another important dimension of lexical semantics,
namely degree of analyzability. By analyzability I mean the extent to
which speakers are cognizant of the presence and the semantic contri-
bution of component symbolic elements. A novel combination is by
definition fully analyzable, since a speaker has to actively manipulate
the components in constructing it. Its meaning then is not just the
novel composite conception (C), but (C) construed in relation to the
component meanings [A] and [B], which categorize different facets of
it. If I coin the term /linger to describe a new device, I necessarily
recognize the contribution of fling and -er, thereby understanding it to
mean 'something that flings'. That is the situation shown on the left in
Figure 4.12.

Figure 4.12

A composite expression's entrenchment as a conventional unit

does not entail any immediate loss of analyzability. The composite
meaning [C] will then have unit status, but it is still construed against
the background of the component meanings [A] and [B], as depicted
in the second panel of Figure 4.12. Thus sprinter is still understood
as 'something that prints', and a scanner as 'something that scans'.
Once [C] is established as a unit, however, it has at least the potential
to be activated independently of [A] and [B]. There is an overall long-
term tendency for the analyzability of composite lexical units to grad-
ually decline: I do not always think of a computer as 'something that
computes', I am less likely to think of a. freezer as 'something that
freezes', and a propeller is hardly ever thought of as 'something that
propels'. We can interpret this lessening analyzability as a reduction
in the likelihood of [A] and [B] being activated along with [C] and/or
as a decrease in their level of activation. This is shown in the third
panel of the diagram. The final step is for [A] and [B] to remain com-

pletely inert when [C] is activated. Linguistically we can then describe

the expression as having lost its original symbolic complexity via
reanalysis.
One factor facilitating a decline in analyzability is the usual
discrepancy between an expression's expected compositional meaning
and the actual contextual meaning that eventually becomes its
conventional value. In accordance with the discussion of Figures 4.7-
9, the structure given as [C] in Figure 4.12 is resolvable into
[[C1]— >[C"]], in which the compositional value [C1] is merely a
stepping-stone for arriving at the contextual value [C"]. Thus sprinter
is not just 'something that prints' but a specific kind of computer
equipment. More drastically, a ruler is less commonly understood as a
device used for ruling lines than as an instrument of measurement.
The more [C"] diverges from [C], the less it is motivated by [C1],
and the easier it becomes for [C11] to be activated autonomously.
Hence a decline in analyzability involves [C1] gradually fading out the
picture along with [A] and [B].
Because mainstream theories have little to say about analy-
zability, it has largely been ignored by linguistic theorists despite its
omnipresence as a significant dimension of meaning and a recurring
problem in lexical and grammatical description. The one situation
where degrees of analyzability are often noted is in the discussion of
"fading" metaphors. In yachting, for example, there is a crew member
who turns a crank in a manner resembling that involved in grinding
meat. This person is metaphorically called a grinder, but for those
immersed in the sport the term's familiarity has no doubt substantially
reduced the salience of its metaphorical basis. We can describe this
fading of the metaphor as a gradual decrease in the likelihood of the
"literal" meaning [C1] being activated along with the "figurative" mea-
ning [C11], as well as in its level of activation.

5.5. Phonology

Let us now turn to phonology and to the phonological aspects of

morphology. Phonological structures are one of the three kinds of
units permitted by the content requirement, the others being semantic
and symbolic units. The difference between phonology and morpho-
logy resides in whether we consider phonological organization in its
own terms or with respect to its role in symbolic structures.8

Purely phonological structures include such elements as

segments, syllables, words, feet, and intonation groups. These can all
be represented at various levels of specificity. A form like pot, for
example, can be characterized phonologically in precise phonetic de-
tail as a sequence of "phones" ([phat]), more abstractly as a series of
phonemes (/pat/), or schematically as a syllabic template ([CVC]).
From the usage-based perspective, we can reasonably anticipate that
structures of any size and any level of abstraction are capable of being
learned and represented as conventional units. A schema that des-
cribes a general pattern, thus defining a space of possible structures,
coexists with instantiating units that specify which regions of the
space are actually used. For instance, the syllable schema [CVC]
might coexist with subschemas such as [SVC] (where S=STOP),
[NVC] (N=NASAL), [SVN], [SVS], [pVC], [NVt], etc. (but not-
for English-either [CVh] or [r)VC]).
Phonological units are also organized into complex categories
describable as networks. A phoneme, for example, is a complex cate-
gory whose prototype corresponds to what was traditionally regarded
as its basic allophone, the one occurring in the greatest variety of con-
texts. Since every context induces some phonetic adjustment (if only
very minor), the prototype must in some measure be schematic. Its
manifestations in particular contexts constitute either instantiations of
the schema, which may themselves have unit status, or extensions re-
cognized as secondary allophones owing to their divergent specifica-
tions. Higher-level Schemas may also be abstracted to represent what
is common to the prototype and different sets of extensions.
A variety of basic phonological entities can be seen as
naturally arising via the same process of abstraction that we have been
discussing throughout. Consider sound segments. At the phonetic
level, segments have no independent existence. To the extent that we
need to posit them, we can regard them as being abstracted from
syllables—perhaps to be modeled as constellations of articulatory
gestures, as proposed by Browman and Goldstein (e.g. 1992)—
which have some claim to being the minimal units of speech. It is
only through phonological decontextualization that a segmental
phoneme like /p/ emerges as a distinct cognitive entity. From actual
syllables, an array of Schemas are presumably extracted representing a
p-like sound in various syllabic contexts: syllable initial, syllable
final, before or after particular vowels, as part of certain consonant
clusters, etc. We can identify these p-like sounds with phones or
allophones of the complex category defining /p/, and the syllabic

Schemas as their conditioning environments. If these phonological

variants have enough in common, and occur in enough distinct
environments, a schematized segment arises which embodies their
commonality but makes no specific reference to syllabic position or
the surrounding context.
Likewise, classificatory phonological features constitute
abstractions from sounds (or sound sequences). To the extent that we
need to posit them, they are merely the schematic characterizations of
"natural classes" of sounds. Representing the feature [STOP], for
example, would be a stop consonant schematic in regard to such
properties as voicing and place of articulation. Of course, sounds vary
in the nature and the degree of their commonality, so alternate sets
give rise to Schemas with different numbers and combinations of
specific properties: [STOP], [ALVEOLAR STOP], [HIGH FRONT
VOWEL], [CONSONANT], [NASAL], etc. Standard features can
then be described as Schemas that are specific in regard to just a single
phonological parameter. In the same vein, the "tiers" employed in
contemporary phonological description amount to sequences of seg-
mental Schemas all of which are specific in regard to certain para-
meters while abstracting away from the others.
What about phonological rules? As in grammar, rules are
limited to Schemas abstracted from actual expressions. This concep-
tion of phonological rules is quite straightforward in the case of
phonotactics. Consider the constraints a language imposes on the
form of permitted syllables. An array of syllable Schemas—[nVt],
[NVt], [NVS], [CVS], [CVC], etc.—represent the patterns inherent
in occurring syllables, described at various levels of generality. When
embedded in a dynamic processing model based on interactive
activation, these Schemas specify the actual syllabic distribution of
segments, certain combinations being assessed as non-conventional.
For example, we can posit for English the low-level Schemas [mVC]
and [nVC], as well as the higher-level generalization [NVC], but not
the non-occurring pattern [r)VC]. To account for the judgment that a
syllable like (rjaek) is un-English, we need only assume that a lower-
level schema wins out over [NVC] for the privilege of categorizing it,
e.g. ([nVC]->(r,a*)).
Less obvious is the treatment of phonological rules tradition-
ally formulated in process terms as derivations from underlying
representations. Although the content requirement proscribes deriva-
tions from underlying structures, it does permit relationships of cate-
gorization, including extension.9 Moreover, from other domains of

linguistic structure we know that chains of extensions often occur,

that extensions are sometimes limited to particular contexts, and that
analogous categorizing relationships can themselves give rise to
Schemas describing their abstract commonality. These properties
suggest an analysis of derivational phonological rules as patterns of
phonological extension.10
Consider, for example, a rule voicing [t] to [d] intervocali-
cally. We can posit such a rule when there is evidence that speakers
pronounce certain forms with a [d] which they nonetheless categorize
as instantiating /t/. For instance, variants with [t] and [d] might co-
exist as the careful and fast-speech pronunciations of numerous
lexical items: [fita]~[fida], [oti]~[odi], [ketul]~[kedul], etc. I follow
Bybee (1994) in supposing that each habitual pronunciation is
mentally represented in considerable phonetic detail. Assuming, then,
that [t] is felt to be "basic", phonological characterizations of the
lexical items in question include categorizing relationships between
their prototypical and extended phonetic variants: [[fita]—>[fida]],
[[oti]—>[odi]], and [[ketul]—>[kedul]]. Like any other regularity,
the commonality inherent in these alternations can be extracted as a
schema: [[...VtV...]—>[...VdV...]]. We can describe this schema in
a number of mutually consistent ways. First, it is immanent in the
networks describing the phonological variants of individual lexical
items. Second, it is part of the complex category representing the
phoneme /t/. It specifies that the basic allophone [t] is extended to [d]
in the context [...V_V...]. Finally, the schema can be regarded as a
phonological rule (which may or may not be productive, i.e.
accessible for the sanction of new instances). These are not competing
analyses, but a matter of the same cognitive entity being considered
from alternate perspectives.

5.4. Morphology

Let us turn now to morphology.11 A number of classic problems

pertain to the notion "morpheme". We present the concept to students
by means of data sets like {fast, faster, fastest, cool, cooler, coolest,
red, redder, reddest}, where words are exhaustively decomposable
into discrete chunks from which they derive in a transparently regular
way. However, linguists are well aware that this archetypal concep-
tion of morphemes as building blocks has severe limitations—with
any representative array of data, the metaphor breaks down imme-

diately. The difficulties lie with the metaphor itself. When the same
phenomena are examined from the usage-based perspective, the prob-
lems simply fail to arise.
Just as segments are abstracted from syllables, morphemes are
abstracted from words. Though some stand alone (just as vowels can
stand alone as syllables), there are many morphemes—in some
languages the vast majority—which only occur as part of larger
words. By and large, it seems fair to say that speakers are more
intuitively aware of words than of their parts, and that large numbers
of complex forms are initially learned as wholes and analyzed only
subsequently (if at all). Words, then, have some claim to primacy.
In the usage-based perspective, morphemes are naturally seen
as arising by the usual process of abstraction. The interpretation of
abstraction as the reinforcement of recurring commonalities echoes the
basic technique of classic morphemic analysis, where the objective is
to identify recurrent pairings between particular conceptual and
phonological structures. The pairing observed in fast is also inherent
in faster and fastest. Analysts therefore posit the symbolic unit
[FAST/fast], just as speakers abstract it from usage events. From
forms like fastest, coolest, and reddest, both linguists and speakers
extract the morpheme [MOST/-est] to represent the systematic co-
occurrence of the concept 'most' (with respect to a property) and the
phonological sequence -est. In straightforward cases like fastest, the
symbolic units thus extracted are exhaustive of the word and readily
taken as yielding it compositionally.
We have seen, however, that complex words are not in gen-
eral fully compositional, whether we look at their initial use or their
established conventional value. The morphemic analysis of printer
into [PRINT/print] and [ER/-er] does not (in conjunction with com-
positional patterns) provide a full characterization of its linguistic
meaning (where it specifically indicates a piece of computer equip-
ment). We saw earlier how this is a natural consequence of learning
via schematization based on contextual understanding (Figures 4.7-
9). From the standpoint of morphemic analysis, this typical situation
is nonetheless problematic if one thinks of words as being built out of
morphemes (where does the extra material come from?). On the other
hand, it is unproblematic if words have a status of their own and
morphemes are abstracted from them. While [PRINT] and [ER] do
not exhaust the specialized meaning of printer, they are discernible in
that meaning. [PRINT/print] can thus be extracted from printer,
printing, printed, etc. by reinforcing their commonality, and [ERAer]
from printer, freezer, eraser, etc., regardless of whether any particular

word is fully compositional. This is the morphemic consequence of

the distinction previously discussed between construction and catego-
rization.
Once the abstractive nature of morphemes is recognized, a raft
of other classic problems evaporate. Degrees of analyzability are
readily accommodated. If forms \ikepropeller, ruler, and stretcher are
originally learned as unanalyzed wholes, it makes little difference to
their efficacious use whether speakers ever make a connection with
propel, rule, and stretch, acquired from other contexts. If they do
establish a connection, the extent to which the composite expressions
activate these components is likewise inessential and no doubt
variable. Should we then say that words like propeller, ruler, and
stretcher are poly morphemic? Either a positive or a negative answer
would, I think, be simplistic. In the present framework it is both
reasonable and coherent to say instead that their analyzability into
component morphemes is a matter of degree.
The extraction of morphemic components need not be uniform
for all portions of a word. For example, since -er occurs in so many
nouns, particularly those indicating agentive or instrumental partici-
pants in actions, it is sufficiently entrenched and salient that appro-
priate forms strongly tend to be analyzed in terms of it. Intuitively, I
would judge that -er is more clearly apparent in propeller, ruler, and
stretcher than are propel, rule, and stretch. The disparity is even more
evident in pliers and plumber, where ply and plumb cast a faint
shadow at best. The limiting case of such disparity is when morphe-
mic analysis touches only part of a word, leaving the remainder as an
unanalyzed residue. There are of course numerous examples invol-
ving -er: father, mother, brother, sister, hammer, roster, sliver,
master, geyser, crater, miser, and so on. Bybee (1988: 128) notes the
residues Mon, Tues, Wednes, etc. left when an obvious commonality
is observed in the names for days of the week. An account of
morphemes based on the reinforcing of common features renders it
unproblematic for only portions of words to participate in such
associations.
Bybee correctly notes that this account extends to phonaes-
themes, e.g. the sir of strip, stripe, strap, strop, street, stretch,
strand, string, etc., which indicates length and thinness. Whether it
has iconic motivation or comes about merely by historical accident, a
recurring sound/meaning association of this kind allows the extraction
of a schema with the potential to be used in coining or analyzing other
expressions. Elaborate systems of sound symbolism might arise in
this fashion (cf. FCG1: 399-401). On the other hand, the schema may

have little salience if only because so many forms containing the

phonological sequence lack the meaning component, and conversely.
Still, the basic process is the same one operative in canonical mor-
phemic analysis. What varies is how far the intrinsic organization of
the data allows this process to proceed—how close it comes to the
"ideal" situation where the morphemic components are salient and ex-
haust the content of the words from which they are extracted.
In her description of modern Aramaic verb morphology,
Rubba (1993) shows that with this kind of approach no special pro-
blems are posed by non-concatenative morphology. A simple illustra-
tion is given in Figure 4.13.

Figure 4Λ3

The symbolic units shown at the bottom are specific words

and stems: the infinitival, third feminine singularjussive, and agentive
forms of 'work', and the agentive forms of 'lie' and 'study' (the
agentive stems are further inflected for gender and number). If mor-
phemes are abstracted from words by a process of schematization, the
data straightforwardly yields the symbolic units shown at the top. The
stem [WORK/ρ...Ι...χ...] comprises the consonants [p], [1], and [x],
occurring in a particular linear sequence, but abstracts away from the
placement (and also the identity) of vowels—since the vocalism is

variable, it is filtered out as common features are reinforced. Con-

versely, the agentive morpheme [AG/CaCaC] is specific in regard to
vocalism but schematic with respect to the surrounding consonants.
Such morphemes are non-prototypical in the sense that their specified
segments are not all contiguous in the temporal sequence. However,
the process of schematization does not itself require any particular
distribution of shared and canceled properties. While there is often a
partitioning between specific and schematic substructures, their inter-
digitation is not at all unusual (cf. X take Υ over X's knee and spank
n. Figure 4.13 further illustrates the point that the same network
of structures can be viewed in different ways, corresponding to
different linguistic constructs. I noted earlier that part of the overall
characterization of a lexical item is a set of structural frames in which
it conventionally occurs. While these frames were identified earlier as
constructional subschemas (as in Figure 4.10), they also include
specific complex expressions with unit status (constructional sub-
schemas and complex expressions form a gradation in any case).
Hence the structures in the ellipse on the left are part of the overall
characterization of the lexical item [WORK/ρ...Ι...χ...], and those in
the ellipse on the right help define the agentive morpheme. At the
same time, many specific forms like palax- 'worker' have the status
of conventional units and thus constitute lexical items (fixed expres-
sions). A form like palax- is of course polymorphemic, comprising
the two morphemes under discussion. The structures in the middle
ellipse are thus interpretable as a construction, in which two compo-
nent structures categorize and motivate the composite expression.
These in turn instantiate (or have immanent in them) a constructional
schema describing the formation of agentive noun stems.

5.5. Morphophonemics
Rubba (1993) has also examined in preliminary terms the treatment of
phonological rules in a network account of morphology. Recall that
phonotactic rules are simply schematized representations of occurring
phonological sequences, whereas "derivational" rules are Schemas
representing patterns of phonological extension. The examples given
previously were purely phonological in the sense that the Schemas
made no reference to any particular morphological context. Some-
times patterns of extension are however limited to certain morpho-
logical contexts, in which case the rules are considered "morpho-

phonemic". Their description in CG remains the same except that

appropriate reference to the context is incorporated in the schema
characterizing the extension (for illustration, see Langacker 1988:
143-145). The example to be considered here involves a phonotactic
constraint and a rule that is morphophonemic in the sense that exten-
sion affects a segment's phonemic categorization.
Rubba documents for modern Aramaic a phonotactic con-
straint to the effect that obstruent clusters agree in voicing. For our
purposes, it is sufficient to posit the cluster template [TT], reflecting
the frequent co-occurrence of voiceless obstruents, as well as the
absence of the schema [DT], which speakers do not extract since
clusters consisting of a voiced and a voiceless obstruent do not occur.
There are however verb roots, such as [HEAL/b...s...m...], where
voiced and voiceless obstruents occur in consecutive consonantal
slots. These are abstracted from specific occurrent forms in which a
vowel appears between the two obstruents, so that the phonotactic
constraint is not violated. Now certain stem forms, including the
infinitive, involve patterns of vocalization that leave the first two root
consonants adjacent to one another. The composite expression is then
pronounced with a voiceless initial obstruent, e.g. psama 'to heal'
(not *bsama). In process terms, one could say that a rule changes /b/
to /p/ to agree in voicing with the following consonant. Of course, it
has sometimes been taken as problematic that the assimilation rule
effectively duplicates the phonotactic constraint, which has to be
posited for independent reasons. Much has also been made of the fact
that such a rule can either be phonological or morphophonemic depen-
ding on whether the phoneme it applies to happens to have a
counterpart with the opposite voicing (Halle 1959).
A partial description of the situation is given in Figure 4.14,
adapted from Rubba (1993: 499). At the top in this diagram are three
symbolic structures: the root [HEAL/b...s...m...], the infinitival
morpheme [INF/CCaCa], and the composite infinitival expression
[HEAL INF/psama]. Collectively these make up a construction, in
which the two component morphemes categorize the composite struc-
ture. The box labeled (i) is part of the overall characterization of the
lexeme b-s-m 'heal'. Its full description comprises not only the root
morpheme, but also an array of structural frames, including various
specific composite expressions in which it conventionally appears.
One of these is the infinitive psama 'to heal'. At the phonological
pole, the relation between [b...s...m...] and [ps...m...] (inherent in
[psama]) is a fragment of the complex category representing the
variant phonological shapes of the root morpheme. It is in particular a

relationship of extension from the presumed prototype [b...s...m...]

to the variant that appears in the context of the infinitival construction.

Figure 4.14

Also apparent in the diagram is the schema [TT], which

specifies the conventionality of a cluster of voiceless obstruents. This
is instantiated by [pT], representing a special case of the general
pattern, in turn elaborated by the specific cluster [ps] (as part of
[psama]). These units and categorizing relationships are a fragment of
the network describing permissible consonant clusters in the lang-
uage. Moreover, these elements also function as components of the
structures in boxes (ii) and (iii), which belong to another network
describing a family of phonological extensions. Extracted from nume-
rous specific extensions like [[b...s...m...]—>[ps...m...]], schema
(ii) represents the pattern of [b] extending to [p] before a voiceless
obstruent. Since the analogous extension occurs with other voiced
obstruents, the higher-level schema (iii) is extracted to embody the
generalization that any such obstruent devoices in this context. Ob-
serve that (ii) is a subschema with respect to (iii), and that the relevant
portions of (i) are a specific instantiationof (ii).

Yet other facets of the diagram pertain to the phonemes /b/ and
/p/. The elements [b] and [p] in (ii) are more or less identifiable with
the basic allophones of these respective categories. The relationship
[[D]-*[b]] shows [b] as a member of the more inclusive class of
voiced obstruents, while [[Γ]-*·[ρ]] relates [p] to the class of voiceless
obstruents. The existence of both voiced and voiceless obstruent
phonemes entails that the systematic extension [[b]—>[p]] (before T)
effects a change in category membership: one variant of /b/ coincides
with the basic (or at least a central) allophone of /p/. That, however, is
a contingent matter which depends on the specific inventory of
phonemes the language happens to have—the extension in (ii), and its
generalized version in (iii), are not intrinsically either phonological or
morphophonemic. Another instantiation of (iii) in the same language
could perfectly well be purely phonological (this is in fact the case in
Aramaic).
How, then, do we characterize the [p] or [psama]? It is
shown in Figure 4.14 as both an instantiation of the phoneme /p/
(note the arrow labeled (iv)) and also an extension vis- -vis the
phoneme /b/ (arrow (v)). The former categorization has a phonetic
basis, and the latter a morphological one, [psama] being understood
as a manifestation of [b...s...m...] (arrow (vi)). While phonologists
will have to determine the relative salience and the consequences of
the alternate categorizations, it seems to me that the framework
portrays the complex situation in a realistic way.

5.6. Larger assemblies

Diagrams like Figures 4.13-14 are initially rather forbidding. To be

sure, they are no more so than a set of algebraic rules or formulas
providing comparable information, and probably less so once certain
notational conventions become familiar. Complex representations
such as these can only be avoided at the price of failing to be even
minimally explicit about essential aspects of linguistic organization.
These diagrams are in any case drastically oversimplified relative to
the actual complexity of the linguistic reality they seek to model. At
best they depict only small fragments of pertinent structures and
networks, selected for minimal illustration of some particular point.
These fragmentary representations afford at least a hint of the
large coalitions of structures and relationships that are brought to bear
in shaping even the smallest portions of expressions and determining
their linguistic interpretation. Coalitions of diverse character and

indefinite complexity have the potential to coalesce into higher-order

structures having some kind of cognitive status in their own right.
The most obvious examples are lexical items (fixed expressions),
which in principle have no upper bound on their possible size. While
X take Υ over X's knee and spank Υ is longer than lexical items are
traditionally thought of as being, it is certainly a conventionally esta-
blished unit, and by no means the largest conventional expression to
be found. Complex expressions can also give rise to constructional
Schemas spanning multiple levels of organization, i.e. not just a single
pattern for integrating two component structures to form a composite
structure, but multi-level assemblies of such patterns defining consti-
tuency hierarchies.
For example, English has both a morphological pattern deriv-
ing adjectives from nouns by means of -fill, and one deriving adverbs
from adjectives by means of -ly. We can represent the two construc-
tional Schemas as [[...]N-ful]ADJ and [[...]ADJ-ly]ADV- On the basis
of a large set of well-established forms—artfully, carefully, hope-
fully, sinfully, dutifully, shamefully, deceitfully, beautifully, play-
fully, cheerfully, successfully, lawfully, scornfully, zestfully, etc.—it
is evident that the combination of these two patterns in successive lay-
ers of morphological organization also represents a conventional pat-
tern. We can therefore posit the higher-order constructional schema
[[[...]N-ful]ADJ-ly]ADV to capture the generalization. Immanent in
particular forms like those cited, this structure is both a complex pat-
tern in its own right and a subschema of [[...]ADJ-ly]ADV in the net-
work spelling out its conventional exploitation. Even a morphologi-
cally impoverished language like English has still more elaborate as-
semblies of this kind, e.g. [[[[...]-al]ADJ-iz]v-ation]N, fr°m central-
ization, normalization, nationalization, radicalization, marginalization,
lexicalization, grammaticalization, and so on (cf. Chapin 1967). With-
out going into any detail, let me suggest that comparable assemblies
of constructional Schemas correspond to such traditional descriptive
devices as morpheme-order charts and templates specifying permitted
clitic sequences.
I should also mention higher-order coalitions such as para-
digms and conjugation classes. I make no apriori claims about the
proportion of specific inflected forms that are learned and stored as
units, nor about the nature and extent of their organization into
psychological assemblies analogous to the paradigms described by
grammarians. However, as Bybee has long maintained, we can rea-
sonably suppose that speakers learn and store large numbers of
specific forms, especially those that are idiosyncratic or represent

minor patterns, but no doubt also including high-frequency forms

instantiating major patterns. A particular stem is abstracted from an
array of inflected forms, many of which may have unit status, as
shown in Figure 4.13 for Aramaic p...L..x... 'work'. Through the
categorizing relationships thus established, a stem provides access to
a set of inflected forms which—if complete enough in relation to the
structural patterns of the language—we can recognize as a paradigm.
Like a grammarian's paradigm, moreover, it is a structured set in
which forms are connected to one another in myriad ways. Schema-
tized forms capture similarities observable with respect to various
parameters at different levels of abstraction. With a verb, for instance,
Schemas might be extracted to represent the commonalities of singular
forms, non-future forms, third-person plurals, etc.
At the same time, other Schemas are extracted to capture what
is shared by analogous forms in the paradigms of different lexemes.
These amount to constructional Schemas and subschemas. Consider
Figure 4.15.

V AG
CiaC2aC 3 ...

v
/ AG
Q...C2...C3... CaCaC

WORK AG LIE AG STUDY AG

palax... dagal... daräs...

f \ S X / \
WORK AG T.TF. AG STUDY AG
p...l...\... CaCaC d...g...l... CaCaC d...r...s... CaCaC

Figure 4.15

The diagram depicts a constructional subschema describing

one pattern of agentive-noun formation in modern Aramaic. It speci-
fies how a root is integrated with the agentive morpheme to yield the

composite agentive expression which interdigitates the root conso-

nants with the agentive morpheme's vocalism. Three instantiations are
shown, the same ones as in Figure 4.13, which however did not give
the constructional schema or the component structures for dagal...
'liar' or dar as... 'studier'. Hence the two diagrams offer different
partial views of the same elaborate web of structures and relation-
ships.
Figure 4.15 is a fragment of the network of constructional
Schemas, subschemas, and specific instantiations describing agentive-
noun formation in Aramaic. Comparable networks describe the pat-
terns for other forms appearing in verbal paradigms (see Rubba 1993
for details). Consider now a particular verb root, e.g. p...l...x...
'work'. A given root functions as a component structure in numerous
constructions, corresponding to the different inflected forms in its
paradigm. Just two of these are shown (enclosed in ellipses) in Figure
4.16: P...I...X... 'work' is a component of both the agentive palax...
and the infinitival form plaxa. We can think of this diagram as abbre-
viating a much larger array of constructions in which the root occurs.
This larger collection of symbolic assemblies constitutes a higher-
order coalition of the sort whose cognitive status is under discussion.

Figure 4.16

Let me now suggest that the entire complex configuration in

Figure 4.16 (and even the larger configuration that it abbreviates) may
have the status of a conventional linguistic unit. It may in any case
provide a basis for schematization. Suppose a number of other verb
roots participate in precisely the same inflectional patterns. For each
of them a higher-level symbolic assembly directly analogous to Figure

4.16 can thus be posited. The usual process of abstraction could then
apply, resulting in the schematized higher-order assembly depicted in
Figure 4.17. This is a coalition of particular constructional subsche-
mas, describing inflectional patterns all of which are conventionally
applicable to the same root. In other words, this higher-order schema
defines a conjugation class. It is a set of associated inflectional pat-
terns, which certain verbs plug into, and which might have sufficient
entrenchment and salience to exert an influence on others. If acces-
sible for the sanction of novel expressions, it can simultaneously
specify all the inflected forms of a newly minted root.

Figure 4.17

6. A final issue
I would like to conclude by comparing this model to two alternate
proposals with respect to the nature of linguistic "rules". One pro-
posal, recently advanced by Pinker and Prince (1991: 230-233), is
that regular and irregular expressions are handled by distinct systems
very different in nature:

Regular inflection (e.g., walk-walked) is perfectly rule-governed,

and thus looks like a paradigm case of a grammatical rule imple-
mented in human brains. Irregular inflection (e.g., sing-sang)
shows varying degrees of unpredictability and thus would seem
to involve brute-force memory...The regular process seems to be

the very essence of the symbol-manipulating, algorithmic ap-

proach to language underlying most theories of grammar, where-
as the irregular process seems to involve a quite different kind of
memory-driven processing...The conclusion we draw is that gen-
erative theories are fundamentally correct in their characterization
of productive rules and structures, but deficient in the way they
characterize memory of less predictable material, which must be
associative and dynamic, somewhat as connectionism portrays
it. It is necessary, then, to develop a new theory...which expli-
citly acknowledges the roles of rules on the one hand and of
associative memory on the other. From such a theory, it
follows that regular and irregular inflection must be computed
by two different systems. Regulars are computed by an
implementation of a classic symbolic rule of grammar, which
concatenates an affix with a variable that stands for the stem.
Irregulars are memorized pairs of words, but the linkages
between the pair members are stored in an associative memory
structure with certain connectionist-like properties.

Thus, something akin to a dynamic usage-based model is accepted for

the kinds of examples—involving less than full compositionality,
generality, and productivity—that have always been problematic in
generative grammar. At the same time, rules in the classic constructive
sense (algorithmic operations on strings of discrete symbols) are held
necessary for truly systematic phenomena.
In one respect Pinker and Prince are no doubt correct: there is
a difference in processing between regular and irregular forms. The
latter are indeed stored and retrieved from memory, whereas the for-
mer are assembled in accordance with productive patterns (an option
available even for high-frequency forms also learned as units). The
two modes of processing are qualitatively distinct in ways that might
very well explain the psycholinguistic and neurological evidence the
authors advance to support their dichotomous view. Still, I am mysti-
fied by their apparent inclination to seek a dichotomous account in
preference to a unified one. This inverts the usual scientific practice of
seeking a unified account for seemingly diverse phenomena. I should
think that positing two distinct cognitive systems would be done only
reluctantly and as a last resort, after all other options have been tho-
roughly explored.
The linguistic facts do not suggest a strictly dichotomous
organization, since sporadic exceptions are possible even for highly
productive general rules, and since minor patterns often show a
certain measure of productivity (Bybee and Slobin 1982). Research

from the usage-based perspective consistently reinforces the idea that

linguistic patterns run the full gamut in terms of systematicity, that
rules approximate full generality and productivity more commonly
than they actually reach it, and that such rules are at best a small
minority in any case. The question, then, is whether a unified account
can be given for the entire spectrum of patterns, including those
which approach full systematicity. I have tried to show that a dynamic
usage-based model offers a realistic prospect of achieving this. Under
the right circumstances, a dynamic system of the sort envisaged is
capable of crisp, reliable behavior representing any desired approxi-
mation to categorical rules. I see no reason to doubt that the observed
degrees and kinds of systematicity can all be accommodated by posit-
ing appropriately configured networks in which conventional units
vary in their specificity and inherent ease of activation.
The other alternate proposal is for linguistic regularities and
the creation of novel expressions to be described in terms of
"analogy" rather than "rules". This is, of course, a classic issue (see,
for example, Bloomfield 1933: 16.6; Householder 1971: ch. 5),
which appears to be reviving as alternatives to the generative
paradigm are increasingly being sought (e.g. Itkonen and Haukioja
1996). What the issue amounts to naturally depends on how the key
terms are defined. For example, if "rules" are equated with construc-
tive statements and "analogy" with Schemas or templates, then the
model proposed here is purely analogical. My own practice, however,
is to use the term "rule" for extracted regularities with some kind of
enduring cognitive presence, regardless of their specific nature. I will
understand "analogy" as referring to expressions being directly
formed on the model of others, not on the basis of stored abstracted
patterns. By these definitions, the dynamic usage-based model I pro-
pose is rule-based rather than analogical. The rules, though, are
templatic Schemas (as opposed to constructive statements), and are
immanent in their instantiations (as opposed to being represented as
distinct cognitive entities).
The question, then, is whether rules (schemas) can be wholly
dispensed with in favor of the direct modeling of novel expressions
on the basis of familiar ones. I think the answer is clearly negative.
An exclusively analogical account—one that posits no abstraction or
schematization—runs directly counter to the usage-based notion that
language exhibits patterns at all levels of generality. It is also inter-
nally inconsistent, since even the learning of specific expressions
(required as the basis for analogy) involves abstraction and schema-
tization from actual usage events. Moreover, the standard argument

against analogy is still a forceful one: if only analogy is posited, how

is the distinction drawn between those analogical formations we find
acceptable and those we do not? If the past tense of swim is swam,
why can I not analogize and use tram as the past tense of trim! One
cannot merely say that there are relatively few pairs like swim/swam
to analogize from, and many more like film/filmed, since nothing
would ensure that only the latter would be chosen as the model. If one
were to claim that the much more numerous pairs like film/filmed
reinforce one another and thus offer an irresistible model for analogy,
the analogical position becomes indistinguishable from one which
posits Schemas. As I have characterized them, Schemas are simply
reinforced commonalities with the potential to influence subsequent
processing.
As this example shows, the distinction between an analogical
and a schema-based account is not necessarily a drastic one (FCG1:
11.3.4). For one thing, Schemas can represent any level of abstrac-
tion, and low-level Schemas are preferentially invoked (other things
being equal) for the categorization of novel expressions. An expres-
sion sanctioned by a low-level schema rather than a higher-level
generalization is likely to be considered "analogical", but a schema
abstracted from just a handful of forms is a schema nonetheless.
Moreover, analogy itself presupposes structural parallelism. In solv-
ing a proportion—e.g. in computing trimmed as the value of X in the
formula film/filmed = trim/X—one must first determine thai film and
filmed are related in a specific way, and then find an X such that trim
and X are related in the same way. But what is "the same way"? It is
an abstract commonality, which the two pairs share. It is therefore a
schema which they both instantiate, and if made explicit it would
actually constitute a constructional schema of the sort proposed in
CG. I have no doubt that true analogies do occur, where new
expressions are modeled on others without the prior extraction and
enduring cognitive presence of any schema. However, the very
process of analogizing induces the apprehension of an abstract
commonality, at least as a fleeting occurrence. The distinction be-
tween rule and analogy then reduces to whether the operative schema
has already achieved the status of a unit. This is at most a matter of
timing and may well be one of degree.

Brought to you by | Columbia University Library The Burke Library New York (Columbia University Library The Burke Library New York)
Authenticated | 172.16.1.226
Download Date | 7/26/12 5:53 PM
Brought to you by | Columbia University Library The Burke Library New York (Columbia University Library The Burke Library New York)
Authenticated | 172.16.1.226
Download Date | 7/26/12 5:53 PM

Neubert and Shreve, Textuality
No ratings yet
Neubert and Shreve, Textuality
30 pages
1 3 5 6.mastering Advanced English Language
67% (3)
1 3 5 6.mastering Advanced English Language
21 pages
Development of Systems Engineering People To Support Major Transformation Plans in Thales
No ratings yet
Development of Systems Engineering People To Support Major Transformation Plans in Thales
15 pages
Frames and Semantics
No ratings yet
Frames and Semantics
8 pages
Chomsky Bare Phrase Structure
No ratings yet
Chomsky Bare Phrase Structure
15 pages
Catford JC 1975 Ergativity in Caucasian Languagues
No ratings yet
Catford JC 1975 Ergativity in Caucasian Languagues
59 pages
Second Language Assessment
No ratings yet
Second Language Assessment
33 pages
Language Planning
No ratings yet
Language Planning
34 pages
Discourse Analysis An Introduction
No ratings yet
Discourse Analysis An Introduction
25 pages
[Ebooks PDF] download Atoms Of Language The Mind's Hidden Rules Of Grammar by Mark C. Baker The full chapters
100% (3)
[Ebooks PDF] download Atoms Of Language The Mind's Hidden Rules Of Grammar by Mark C. Baker The full chapters
24 pages
08 Pragmatics
No ratings yet
08 Pragmatics
3 pages
Language and Thought - Carroll
No ratings yet
Language and Thought - Carroll
136 pages
Lexico Grammar
No ratings yet
Lexico Grammar
4 pages
M Lynne Murphy 2003 Semantic Relations A
No ratings yet
M Lynne Murphy 2003 Semantic Relations A
6 pages
Anaphoric Encapsulation: Maria-Elisabeth Conte
No ratings yet
Anaphoric Encapsulation: Maria-Elisabeth Conte
10 pages
Pragmatic Conditionals
No ratings yet
Pragmatic Conditionals
15 pages
English Studies Today: Prospects and Perspectives
No ratings yet
English Studies Today: Prospects and Perspectives
324 pages
Formulaic Language in Native and Second Language Speakers: Psycholinguistics, Corpus Linguistics, and TESOL
No ratings yet
Formulaic Language in Native and Second Language Speakers: Psycholinguistics, Corpus Linguistics, and TESOL
22 pages
Acquisition of Tense-Aspect Morpholgy - OCR
No ratings yet
Acquisition of Tense-Aspect Morpholgy - OCR
21 pages
Silverstein - "Distinctive Features" in Leonard Bloomfield's Phonology
100% (1)
Silverstein - "Distinctive Features" in Leonard Bloomfield's Phonology
37 pages
DIXON - Ergativity
No ratings yet
DIXON - Ergativity
81 pages
Usage Vs Use
No ratings yet
Usage Vs Use
1 page
Text Linguistics
No ratings yet
Text Linguistics
30 pages
Adding Sense Context and Interest in A Grammar of Multimodal Meaning by Mary Kalantzis Bill Cope
No ratings yet
Adding Sense Context and Interest in A Grammar of Multimodal Meaning by Mary Kalantzis Bill Cope
398 pages
Universal Grammar Theory and The Classroom
100% (1)
Universal Grammar Theory and The Classroom
16 pages
Inflection 1. Inflection 1. Inflection 1. Inflection
No ratings yet
Inflection 1. Inflection 1. Inflection 1. Inflection
23 pages
Ammar English
0% (1)
Ammar English
13 pages
Langacker 2015 How To Build An English Clause
No ratings yet
Langacker 2015 How To Build An English Clause
45 pages
Translation by Omission
0% (1)
Translation by Omission
6 pages
#2 Canale and Swain
100% (1)
#2 Canale and Swain
5 pages
Givon Logic Pragmatics Epistemology
No ratings yet
Givon Logic Pragmatics Epistemology
53 pages
Bridging The Gap Between Second Language Acquisition Research and Memory Science: The Case of Foreign Language Attrition
No ratings yet
Bridging The Gap Between Second Language Acquisition Research and Memory Science: The Case of Foreign Language Attrition
8 pages
03 - CLARK - Using Language
No ratings yet
03 - CLARK - Using Language
3 pages
Theoretical Grammar 5
No ratings yet
Theoretical Grammar 5
35 pages
Querying - Keywords - Paul Baker
No ratings yet
Querying - Keywords - Paul Baker
25 pages
01.givón (1983) - Introduction (Topic Continuity)
100% (2)
01.givón (1983) - Introduction (Topic Continuity)
22 pages
Interface Between Syntax and Morphology
No ratings yet
Interface Between Syntax and Morphology
5 pages
Didactic S
No ratings yet
Didactic S
8 pages
Collins Activities Using A Corpus
No ratings yet
Collins Activities Using A Corpus
63 pages
Lexicology Course Description Buja
100% (1)
Lexicology Course Description Buja
2 pages
HuangYan 2007 1introduction Pragmatics
No ratings yet
HuangYan 2007 1introduction Pragmatics
20 pages
COMPETENCE Versus Performance
No ratings yet
COMPETENCE Versus Performance
15 pages
Reflection No. 2 Psycholinguistics
No ratings yet
Reflection No. 2 Psycholinguistics
1 page
A Contrastive Study of English and Arabic Syntax
No ratings yet
A Contrastive Study of English and Arabic Syntax
30 pages
Complaints As Positive Strategies-What The Learner Needs To Know
No ratings yet
Complaints As Positive Strategies-What The Learner Needs To Know
24 pages
Fraser - 2005 - Towards A Theory of DM
No ratings yet
Fraser - 2005 - Towards A Theory of DM
33 pages
The Translation of English Tenses and Aspects in Narrative Mode Into Vietnamese
No ratings yet
The Translation of English Tenses and Aspects in Narrative Mode Into Vietnamese
12 pages
Sinclair and Coulthard
No ratings yet
Sinclair and Coulthard
6 pages
Cognitive Semantics: Prepared By: Maricar Talosig Angel Ann Masiddo
No ratings yet
Cognitive Semantics: Prepared By: Maricar Talosig Angel Ann Masiddo
6 pages
English As A Lingua Franca. Attitude and Identity
No ratings yet
English As A Lingua Franca. Attitude and Identity
38 pages
X(Bar) Syntax - A Study of Phrase Structure
100% (1)
X(Bar) Syntax - A Study of Phrase Structure
267 pages
Beaugrande R., Discourse Analysis and Literary Theory 1993
No ratings yet
Beaugrande R., Discourse Analysis and Literary Theory 1993
24 pages
Contrastive Linguistics Approaches PDF
No ratings yet
Contrastive Linguistics Approaches PDF
17 pages
Biber and Egbert 2023 What Is A Register
No ratings yet
Biber and Egbert 2023 What Is A Register
22 pages
Group 1 - Chapter 9 Critical Discourse Analysis
No ratings yet
Group 1 - Chapter 9 Critical Discourse Analysis
25 pages
Theories of Second-Language Acquisition - Wikipedia
No ratings yet
Theories of Second-Language Acquisition - Wikipedia
69 pages
Analysis of a Medical Research Corpus: A Prelude for Learners, Teachers, Readers and Beyond
From Everand
Analysis of a Medical Research Corpus: A Prelude for Learners, Teachers, Readers and Beyond
Georgette Nicolas Jabbour
No ratings yet
Meaning, Language, and Time: Toward a Consequentialist Philosophy of Discourse
From Everand
Meaning, Language, and Time: Toward a Consequentialist Philosophy of Discourse
Kevin J. Porter
No ratings yet
Introduction: A Rough Guide To Cognitive Linguistics
No ratings yet
Introduction: A Rough Guide To Cognitive Linguistics
1 page
Cognitive Semantics
No ratings yet
Cognitive Semantics
3 pages
Chapter 1 Generative Grammar
No ratings yet
Chapter 1 Generative Grammar
13 pages
Formulaicity Lundell
No ratings yet
Formulaicity Lundell
12 pages
Entrenchment From A Psycholinguistic and Neurolinguistic Perspective
No ratings yet
Entrenchment From A Psycholinguistic and Neurolinguistic Perspective
24 pages
Does Bilingualism Come With Linguistic Costs? A Meta Analytic Review of The Bilingual Lexical Defcit
No ratings yet
Does Bilingualism Come With Linguistic Costs? A Meta Analytic Review of The Bilingual Lexical Defcit
17 pages
Are Frequent Phrases Directly Retrieved Like Idioms? An Investigation With Self-Paced Reading and Language Models
No ratings yet
Are Frequent Phrases Directly Retrieved Like Idioms? An Investigation With Self-Paced Reading and Language Models
12 pages
Object Recognition System Design in Computer Vision: A Universal Approach
No ratings yet
Object Recognition System Design in Computer Vision: A Universal Approach
18 pages
Prelim Quiz 1: Question Text
No ratings yet
Prelim Quiz 1: Question Text
62 pages
Module 5 Chapter 6 Cross-Cultural Leadership Studies
No ratings yet
Module 5 Chapter 6 Cross-Cultural Leadership Studies
26 pages
Facts About The UPCAT and UP Applications
No ratings yet
Facts About The UPCAT and UP Applications
8 pages
No. 44 Role of Budgets in Educational Planning by Jacques Hallak
No ratings yet
No. 44 Role of Budgets in Educational Planning by Jacques Hallak
20 pages
TFN
No ratings yet
TFN
14 pages
rohini_36824404257
No ratings yet
rohini_36824404257
6 pages
EHealth DSI - System Architecture Specification - v2.1.0
No ratings yet
EHealth DSI - System Architecture Specification - v2.1.0
69 pages
Dmgt505 Management Information System
No ratings yet
Dmgt505 Management Information System
272 pages
Management Control Systems (302) BY: Pravin Tungare
No ratings yet
Management Control Systems (302) BY: Pravin Tungare
9 pages
Computers and Electronics in Agriculture
No ratings yet
Computers and Electronics in Agriculture
8 pages
OOSE-Chapter - 3-Part II (Structural Modeling)
No ratings yet
OOSE-Chapter - 3-Part II (Structural Modeling)
81 pages
SPS Creative Technologies G7 Curriculum
No ratings yet
SPS Creative Technologies G7 Curriculum
12 pages
Computational Mathematics Laboratory: Lab Manual
No ratings yet
Computational Mathematics Laboratory: Lab Manual
44 pages
Discrete-Event Simulation of Health
No ratings yet
Discrete-Event Simulation of Health
42 pages
Role of It in Business
No ratings yet
Role of It in Business
7 pages
MODELING, OPTIMIZATION AND DESIGN OF A SOLAR THERMAL (PDFDrive)
No ratings yet
MODELING, OPTIMIZATION AND DESIGN OF A SOLAR THERMAL (PDFDrive)
245 pages
Explain Each of The Functions of Applied Social Sciences
No ratings yet
Explain Each of The Functions of Applied Social Sciences
15 pages
CHN Reviewer Midterms Final
No ratings yet
CHN Reviewer Midterms Final
14 pages
CIRED2011 0457 Final
No ratings yet
CIRED2011 0457 Final
4 pages
Toyo Ito On Space
100% (1)
Toyo Ito On Space
8 pages
Fundamentals of Information System 2nd Year Unite 3
No ratings yet
Fundamentals of Information System 2nd Year Unite 3
16 pages
Slawomir Wycislak, Michael A Radin - The Fundamentals of Systems Thinking, Management & Effective Leadership-World Scientific Publishing Company (2024)
100% (1)
Slawomir Wycislak, Michael A Radin - The Fundamentals of Systems Thinking, Management & Effective Leadership-World Scientific Publishing Company (2024)
306 pages
Software Requirement Engineering Notes
33% (3)
Software Requirement Engineering Notes
15 pages
Mocking Test 6: A. LISTENING (50 Points)
No ratings yet
Mocking Test 6: A. LISTENING (50 Points)
9 pages
Content Module 1. Biostatistics
No ratings yet
Content Module 1. Biostatistics
103 pages
BRUCKMEIER. Global Environmental Governance
100% (2)
BRUCKMEIER. Global Environmental Governance
296 pages
10-Personnel Cost Planning
No ratings yet
10-Personnel Cost Planning
52 pages
Sen - Project 14
No ratings yet
Sen - Project 14
22 pages

A Dynamic Usagebased Model

Uploaded by

A Dynamic Usagebased Model

Uploaded by

Chapter 4

A dynamic usage-based model

1. The usage-based conception

and highly redundant, the most accurate description of it (as a psycho-

tial to language structure than high-level Schemas representing the

Also fundamental to cognition is the ability to compare two

should also be evident that these operations occur in various combi-

does, at least at a basic level, and I believe it to be both realistic and

periences is reinforced and attains some kind of cognitive status, so

an entity in its own right, facets of which correspond to the compo-

4. Basic linguistic problems

tional, linguistic, or a unit is inherently a matter of degree (FCG1:

4.1. Categorization of usage events

We have already noted the oversimplification inherent in rep-

Actually, a slight refinement of Figure 4.2 should be noted.

I presume that extension does not occur at random; there is

the status of a unit, it is validly describable as both a schema instan-

I assume, then, that linguistic categories are usually complex,

4.2. Selection of categorizing structures

For both linguistic and non-linguistic input, I assume we have

Several factors interact to determine which particular member

A third factor is the amount of overlap between the target and a

embedded in a dynamic processing system which operates in accor-

4.3. Categorization vs. construction

I have posed the question of whether the linguistic system per se

once the usage becomes familiar and conventionalized, it is incor-

repeatedly, at progressively higher levels, yielding composite sym-

Together, these units implicitly define a complex expression,

In referring to (C1) as merely "latent", and saying that the con-

I have argued, however, that full compositionality is unchar-

Because (C") represents the way the expression is actually

expression like chopper (for 'helicopter') retains the compositional

4.5. Degrees of regularity

tion of new expressions. Conversely, a pattern representing any level

Like most linguistic phenomena, grammatical patterns usually arrange

stored as units provided that they occur with sufficient frequency

If it is workable, a theory that does not posit filters or

5.1. Lexicon and grammar

CG itself offers conceptual unification. It posits only semantic, pho-

To make the same point in another way, let us consider more

At the same time, however, the subschema [[send][NP][NP]]

5.2. Lexical semantics

A lexeme is not precisely the same in all its environments. Since

recipient, focusing instead on the entity the sender hopes to obtain. It

Tuggy shows that graded judgments like these, as well as

As shown in Figure 4.11, clear-cut cases of ambiguity are

portrait, Jane's having done so with lines on a highway can only be

A composite expression's entrenchment as a conventional unit

pletely inert when [C] is activated. Linguistically we can then describe

Let us now turn to phonology and to the phonological aspects of

Purely phonological structures include such elements as

Schemas as their conditioning environments. If these phonological

linguistic structure we know that chains of extensions often occur,

Let us turn now to morphology.11 A number of classic problems

word is fully compositional. This is the morphemic consequence of

have little salience if only because so many forms containing the

The symbolic units shown at the bottom are specific words

variable, it is filtered out as common features are reinforced. Con-

phonemic". Their description in CG remains the same except that

relationship of extension from the presumed prototype [b...s...m...]

Also apparent in the diagram is the schema [TT], which

5.6. Larger assemblies

Diagrams like Figures 4.13-14 are initially rather forbidding. To be

indefinite complexity have the potential to coalesce into higher-order

minor patterns, but no doubt also including high-frequency forms

WORK AG LIE AG STUDY AG

The diagram depicts a constructional subschema describing

composite agentive expression which interdigitates the root conso-

Let me now suggest that the entire complex configuration in

Regular inflection (e.g., walk-walked) is perfectly rule-governed,

the very essence of the symbol-manipulating, algorithmic ap-

Thus, something akin to a dynamic usage-based model is accepted for

from the usage-based perspective consistently reinforces the idea that

against analogy is still a forceful one: if only analogy is posited, how

You might also like