The Rule Controversy: A Cognitive Grammar Approach

   THE RULE CONTROVERSY:  A COGNITIVE GRAMMAR PERSPECTIVE

                    Ronald W. Langacker
             University of California, San Diego

     Few issues are  more  consequential  for  our  view  of
language  and  mind than the nature and status of linguistic
rules.  The connectionist challenge to the  classic  genera-
tive conception of rules should therefore be welcomed by all
concerned, whether because it ultimately shows that explicit
rules  are  superfluous, or because it stimulates generative
theorists to refine their notion of rules and the basis  for
postulating  them.   My own stand on this issue represents a
kind of middle ground.  While cognitive grammar has  natural
affinities to connectionism, it does posit something compar-
able to rules.  At the same  time,  it  conceives  of  these
entities in a way that makes them amenable to an interpreta-
tion in terms of connectionist  processing.   My  objectives
here are thus to sketch this conception of linguistic rules,
to portray it as the logical culmination of  trends  observ-
able  within the generative tradition itself, and thereby to
point the way to a possible eventual convergence.


           1. Basic Features of Cognitive Grammar

     While I can offer here only the briefest of characteri-
zations,  I maintain that the theory of cognitive grammar is
intrinsically desirable on grounds of  naturalness,  concep-
tual unification,and theoretical austerity. [1] One aspect of
its naturalness resides in the  fact  that  it  posits  only
semantic  structures,  phonological structures, and symbolic
links between the two.  This  basic  organizational  feature
correlates  directly  with the primary function of language,
that of permitting meanings to be symbolized by phonological
sequences.   The  theory  achieves conceptual unification by
claiming that grammar is fully reducible to  symbolic  rela-
tionships:  lexicon, morphology, and syntax form a continuum
exhaustively describable in terms  of  symbolic  structures,
each comprising both a semantic and a phonological structure
together with the symbolic  relationship  that  links  them.
Grammar  is  therefore  said  to be symbolic in the specific
sense that it reduces to form-meaning pairings. [2]
     Cognitive grammar is theoretically austere by virtue of
the  content  requirement,  which stipulates that only three
kinds of elements are ascribable  to  a  linguistic  system:
(i)  semantic,  phonological,  and  symbolic structures that
occur overtly as (parts of)  expressions;  (ii)  schematiza-
tions  of permitted structures; and (iii) categorizing rela-
tionships between permitted structures.  To take a phonolog-
ical  example,  sim,  lep, and tich are syllables that occur
overtly in English; [CVC]  is  a  schematization  over  such
structures;     and    [[CVC]===>[sim]]    represents    the
categorization of [sim] as an instance of [CVC].  Similarly,
the  words  head,  heart,  hand,  hip,  and heel are overtly
occurring symbolic structures; the schema [BODY PART/hV(C)C]
expresses  their  commonality and thereby defines a class of
symbolic elements; and  [[BODY  PART/hV(C)C]===>[HEAD/head]]
indicates that head is a member of that class. [3] The effect
of the content requirement is to rule  out  any  descriptive
construct  not  straightforwardly  derivable,  by  the well-
attested processes  of  schematization  and  categorization,
from structures experienced directly (i.e. meanings and pho-
nological sequences).
     In this framework, therefore,  linguistic  regularities
take  the  form  of  schemas.   What do I mean by that term?
Importantly, a schema is not the same as a list of criterial
attributes or a bundle of separate features.  It is rather a
coherent, integrated structure comparable in  most  respects
to  those which support its extraction.  A schema's internal
organization is precisely parallel to that of the  semantic,
phonological,  or  symbolic  structures it schematizes, thus
reflecting whatever commonality they exhibit.  It does  how-
ever  abstract  away  from their points of divergence, being
neutral or less specific in regard to each;  overall,  then,
it  is  characterized  at  a  lower  degree of precision and
detail.  The essential component  of  this  relationship  is
granularity:  relative to the fine-grained specifications of
the structures it categorizes,  a  schema  provides  only  a
coarse-grained  characterization  in  which certain detailed
features fail to appear owing to insufficient resolution. [4]
A difference in levels of schematicity (or specificity) is the
basis for hyponymy in lexical semantics, and more generally,
for  the relation between superordinate and subordinate lev-
els in a taxonomic hierarchy.
     I will indicate a relationship of schematicity by means
of  a double arrow.  Thus, A===>B states that A is schematic
for B, or conversely, that B elaborates (or instantiates) A.
Another  basic type of categorization, indicated by a single
arrow, is extension from a prototype.  Hence  A--->B  states
that  A  categorizes  B  despite some conflict in specifica-
tions. Linguistic categories are typically complex,  in  the
sense that their proper characterization takes the form of a
network whose nodes are structures  linked  by  categorizing
relationships  of these two sorts.  The structures that par-
ticipate in these relations (the nodes in the  network)  can
be  semantic, phonological, or symbolic, and may exhibit any
degree of internal complexity.  The network may,  for  exam-
ple,  comprise  alternate  senses  of a lexical item, or the
various allophones of a phoneme.  The  individual  nodes  of
the network may even consist of entire grammatical construc-
tions, analyzed here as configurations  of  symbolic  struc-
tures (cf. Langacker 1988c; Lakoff 1987, case study 3).
     The possibility of reducing grammar  to  configurations
of  symbolic  structures  presupposes an appropriate view of
linguistic  semantics,  one  that   equates   meaning   with
conceptualization  and properly accommodates construal, i.e.
our capacity for conceiving the same situation in  alternate
ways.  One aspect of construal is our ability to conceptual-
ize an entity at various levels of specificity, as reflected
in  schematic hierarchies such as [THING] ===> [ANIMAL] ===>
[MAMMAL] ===> [DOG] ===> [BEAGLE] or  [SEGMENT]  ===>  [CON-
SONANT]  ===>  [STOP]  ===>  [VELAR  STOP] ===> [k].  We are
further able to conceptualize  a  situation  from  different
perspectives (consider come vs. go), to construe it in rela-
tion to different background  assumptions  and  expectations
(e.g. half-empty vs. half-full), and to render certain enti-
ties more prominent than others.  One type of prominence  is
figure/ground  organization.   Normally a relational expres-
sion accords one of its participants the  status  of  figure
within the scene; I call this the trajector, and refer to an
entity of secondary salience as a landmark.  Thus the seman-
tic  contrast between X is near Y and Y is near X resides in
whether X is construed as the trajector (relational  figure)
and  is being located with reference to Y, or conversely.  I
use the term profiling for  a  second  type  of  prominence,
whereby  every expression--within the conception it evokes--
singles out a particular substructure as  a  kind  of  focal
point;  this  substructure  (the profile) is the entity that
the expression designates.  For instance, knuckle evokes the
conception  of a finger and profiles (designates) one of its
joints.  Near profiles the  relationship  between  two  con-
ceived entities (its trajector and landmark).
     Given a conceptualist semantics based on construal, the
reduction of grammatical structure to symbolic relationships
becomes quite feasible.  Grammar can, I  believe,  be  fully
and revealingly characterized using only symbolic structures
(form-meaning pairings); on this view grammar is  indissoci-
able  from  meaning, and all grammatical elements are attri-
buted some kind of  conceptual  import.   For  instance,  an
expression's  grammatical  category  is  determined  by  the
nature of its profile (Langacker 1987a,  1987b).   Thus,  an
expression is categorized as a noun by virtue of profiling a
thing (under a highly abstract  definition  of  that  term).
Rather than being marked as such by a diacritic or syntactic
feature (devices not permitted by the content  requirement),
a  word  like cat is inherently classed as a noun because it
instantiates   the    schema    defining    the    category:
[[THING/...]===>[CAT/cat]].[5] Likewise, every verb is claimed
to profile a process (defined as a relation scanned  sequen-
tially  in  its  evolution  through  conceived  time),  e.g.
[[PROCESS/...]===>[SPRAY/spray]].   Semantic   characteriza-
tions have also been proposed for many other basic grammati-
cal notions, including noun  phrase,  finite  clause,  head,
complement, modifier, coordination, subordination, auxiliary
verb, subject, object, transitivity, unaccusative, and erga-
tivity (Langacker 1986b, 1986c, 1987a, 1989, 1990, in press,
to appear; Rice 1987b).
     Our focus here is on grammatical rules,  which  pertain
to  the  combination  of simpler symbolic structures to form
more complex  ones.   I  say  that  a  symbolically  complex
expression  is formed by the integration of two or more com-
ponent structures to yield a composite structure.  For exam-
ple,  the  two  component  structures,  jar  and lid, can be
integrated to form the composite structure jar lid.  Each is
symbolic,  comprising  a  semantic structure symbolized by a
phonological structure:  [JAR/jar], [LID/lid], [JAR  LID/jar
lid].   Semantically,  jar  designates  a particular type of
container, while lid  profiles  the  cover  to  a  container
characterized  only  schematically.  Phonologically, jar and
lid  are  each  characterized  as  words.   Integration   is
effected  by correspondences established between subparts of
the component semantic and phonological  structures.   Thus,
the  specific container profiled by jar is put in correspon-
dence with the schematic container evoked by lid, and lid is
identified as the word that directly follows jar in temporal
order.  The composite structure is formed  by  superimposing
corresponding  entities  and  merging  their specifications.
The composite structure [JAR LID/jar lid] inherits the  pro-
filing of lid, which is therefore the head.
     Jar lid instantiates a general pattern of compound for-
mation  in  English.   The  regularity exhibited by jar lid,
door knob, pencil eraser, milk carton,  sea  captain,  salad
oil,  school bus, garbage man, train station, delivery boy,
and countless others is captured by an  appropriate  schema,
which  expresses  their  commonality  while abstracting away
from their points of contrast.  Since one  aspect  of  their
commonality lies in their symbolic complexity and the nature
of the relationships holding among the component and  compo-
site structures, those organizational features are reflected
in the  schema.   This  constructional  schema  is  directly
analogous  to  the  complex  expressions  from  which  it is
extracted:  the only difference is that  the  component  and
composite  structures  are  themselves schematic rather than
specific.  Semantically, for instance, the  first  component
is  merely characterized as profiling a thing (this makes it
a noun), and the second, as profiling  a  thing  that  bears
some   unspecified  relationship  to  another  thing,  which
remains unprofiled.  However, the profile of the first  com-
ponent corresponds to the unprofiled thing within the second
(just as in jar lid), and the profile of the second prevails
at  the  composite-structure  level.   This  latter  feature
represents the generalization that  the  second  element  of
English noun-noun compounds functions as the head.
     In principle, any regularity observable across a set of
complex  expressions  can be captured by an appropriate con-
structional schema.  Such schemas  are  cognitive  grammar's
equivalent of grammatical rules.  As complex symbolic struc-
tures  derived  by  schematization  from   overtly-occurring
expressions,  they  clearly  conform to the content require-
ment.   Moreover,  they   represent   established   patterns
potentially   available  for  the  categorization  of  novel
expressions; they can also be viewed as  templates  employed
in  the  assembly or evaluation of such expressions.  A par-
ticular expression is simultaneously categorized by numerous
schemas, each pertaining to a different aspect of its struc-
ture.  The full set of categorizing relationships  in  which
an expression thusly participates constitutes its structural
description and determines its status vis-a-vis the linguis-
tic  system.   An  expression  is  conventional  (i.e. well-
formed) with respect to a given aspect of its structure just
in  case it elaborates the categorizing schema (i.e. A===>B)
rather than conflicting with its specifications (A--->B).
     Hence the entire complex structure represented  by  jar
lid  is categorized by the constructional schema, and to the
extent that jar lid constitutes a  fixed,  familiar  expres-
sion,  both  it  and the categorizing relationship belong to
the grammar of English (i.e. they are facets of a  speaker's
grasp  of  established convention).  These elements are part
of a substantial network describing the various patterns and
subpatterns  of  English  compounding.   The  nodes  of this
network--linked by categorizing relationships of elaboration
or  extension--include specific compounds with the status of
fixed expressions,  subschemas  characterized  at  different
levels of specificity, and higher-level schemas representing
broad  generalizations.   These  structures   have   varying
degrees  of  cognitive  salience, in the sense of "entrench-
ment" or ease of activation; thus accommodated is  the  dis-
tinction between "major" and "minor" patterns, or productive
vs. non-productive rules.  It can be seen, therefore, that a
network of this sort both captures any discernible regulari-
ties and specifies how general patterns are actually  imple-
mented  in conventional usage.  Its configuration (including
salience) determines which particular instantiations--out of
the  vast  range  that a higher-level schema would theoreti-
cally permit--actually tend to be used, and with what degree
of likelihood.  Distributional restrictions are thus handled
in a cognitively plausible way, without resorting to devices
that  would  violate  the  content  requirement  (e.g.  rule
features or diacritics).
     Though I can hardly prove it here, I believe this  view
of  linguistic  structure  to be both viable and revelatory,
and will assume its workability for purposes of the  follow-
ing  discussion.   Its  basic  affinity  with  connectionism
should be readily apparent.  First, it  reduces  grammar  to
form-meaning  pairings,  which PDP models are well suited to
deal with.  Second,  it  makes  no  qualitative  distinction
between  rules  and  data;  schemas and their instantiations
differ only in level of specificity, which is  a  matter  of
degree.   Third,  it relies on the extraction of generaliza-
tions from positive instances, through the reinforcement  of
common organizational features.  Finally, it considers local
regularities to be at least as significant in language  pro-
cessing  as  high-level generalizations.  Having noted these
similarities, let us now consider  how  cognitive  grammar's
conception  of  linguistic  rules  relates  to  developments
within the generative tradition.


                 2. A Spectrum of Positions

     To put the issue in proper perspective, let us  outline
two  extreme  positions concerning the cognitive representa-
tion  of  linguistic   structure.    These   positions   are
caricatures--no  serious scholar subscribes to either.  They
do however define the endpoints in a  spectrum  of  possible
views  and are therefore useful as reference points.  We can
think of them as corresponding to the most simplistic notion
that  an uninformed generativist or connectionist might con-
ceivably entertain concerning what the other  believes.   In
this  spirit of misconception and caricature, we can imagine
a connectionist referring to the position he wrongly imputes
to  a  generativist as "empty symbol pushing".  We can simi-
larly imagine a generativist offering for his misinterpreta-
tion  of connectionism the catchy descriptive label "mind as
mush".
     "Empty symbol pushing" is the  (hypothetical)  position
that  a  language is fully describable by a set of rewriting
rules very much  like  the  grammars  of  simple  artificial
languages  found  in  Chomsky's  early  writings (e.g. 1957,
1965; Chomsky and Miller 1963).  Except for a  vast  differ-
ence in complexity, a natural-language grammar is thought of
as being precisely analogous, say, to  that  comprising  the
two  rewriting  rules  S--->aSb and S--->ab, which generates
all and only the sentences consisting of  a  string  of  a's
followed  by  a  string of b's of equal length.  Grammars of
this kind have four  essential  properties:   (i)  They  are
explicit  in the strong sense of being generative, i.e. they
offer a full and precise  mathematical  characterization  of
all  and only the well-formed sentences of a language.  (ii)
They are constructive, consisting of  rules  for  assembling
grammatical  expressions.  Observe that these rules are for-
mally distinct from the expressions  they   generate (e.g.   
S--->aSb  is formally dissimilar  to aaaabbbb).  (iii) Every
rule is fully general; it applies in the  derivation  of  an
open-ended  set  of  sentences,  and  is  applicable without
exception to any structure that  meets  its  specifications.
(iv)  The  rules  manipulate  contentless  symbols.   A non-
terminal symbol (such as S) has neither semantic nor  phono-
logical  content.  And while terminal symbols (such as a and
b) might be thought of as vocabulary items, their meaning is
irrelevant  to  grammatical derivation--a string's grammati-
cality can be determined from its form alone (e.g. by count-
ing the number of a's and b's).
     At the other extreme,  "mind  as  mush"  describes  the
(hypothetical)  view that mental processing has no interest-
ing structure  whatever--it  merely  involves  an  array  of
faceless  units  behaving in a squishy fashion.  With suffi-
cient training, a PDP system comes to function in a way that
mirrors  the statistical regularities inherent in its input,
and that is all the investigator is  concerned  with  accom-
plishing.  Since learning is confined to adjustments in con-
nection weights, and since no individual weight  or  adjust-
ment can be identified with a specific linguistic structure,
there is no point  in  searching  for  anything  that  might
correspond  to the discrete constructs posited by linguists.
In particular, the explicit rules they  formulate  are  con-
sidered  superfluous and fictitious--after all, a PDP system
accommodates linguistic regularities  without  resorting  to
such entities.
     Thus the "empty symbol  pushing"  and  "mind  as  mush"
positions   represent   polar  opposites  that  are  clearly
irreconcilable.  They do not  represent  positions  actually
espoused,  however,  and  as we move away from these carica-
tures to more accurate descriptions  of  what  generativists
and  connectionists  currently believe, the contrast appears
considerably less stark.
     Now as I understand the connectionist  enterprise,  its
objective  is a realistic model of cognitive processing that
accurately reflects observed behavior displaying all degrees
of  regularity, from the idiosyncratic to the exceptionless.
Importantly, PDP systems are  quite  capable  of  structured
behavior which is "rule-governed" at least in the sense that
specific patterns of activation  are  crisply  and  reliably
elicited  by  particular kinds of input.  Connectionists are
very much concerned, moreover, with  finding  out  just  how
their systems work.  It is for this reason that they monitor
the activation of hidden units to  see  what  features  they
serve  to  detect; perform cluster analyses to determine, on
the basis of response similarity, the  implicit  categoriza-
tion  the system has imposed on the input data; consider its
behavior in terms of locations  and  trajectories  in  state
space;  and so on.  The functioning of a connectionist model
is therefore regarded as organized activity  susceptible  to
being  studied  and  understood, and it is quite conceivable
that certain aspects of this processing might be  identified
with particular linguistic constructs.
     Likewise, contemporary generative theory  bears  little
resemblance  to the "empty symbol pushing" caricature.  With
respect to all  four  properties  noted  previously,  trends
within  the  generative tradition have moved it closer to an
outlook having a certain amount of commonality with the con-
nectionist  perspective.  (i)  Although  explicit  rules and
representations are still considered de rigueur,  no  longer
is a grammar universally conceived as a strictly generative,
algorithmic device.  In fact, theories (notably  government-
binding  theory)  are now proposed and developed without any
serious attempt at formalization.  (ii) It  is  increasingly
less  common for theorists to use rewriting rules or to con-
ceive of grammars as constructive devices.   There  is  more
emphasis on surface constructions (as opposed to derivations
from underlying structures), and characterizations involving
the  simultaneous  satisfaction  of multiple constraints are
envisaged both in "unification-based" approaches and in dis-
cussions  of "modularity".  (iii) That a linguistic descrip-
tion requires statements at  all  levels  of  generality  is
fully  recognized, and a variety of formal devices have been
proposed for exceptions, irregularity, and rules of  limited
productivity.   These  are often consigned to "the lexicon",
which is now regarded as being of  prime  importance.   (iv)
The intimate association of grammar and meaning is coming to
be appreciated.  It has long been acknowledged that  gramma-
ticality  judgments  cannot be based on strings of words per
se, but pertain to particular  structural  descriptions  (if
not  specific  interpretations).   There  is  also a growing
realization that differences in grammatical behavior  corre-
late  with  differences in meaning, and that semantic (or at
least "pragmatic") consequences follow from the choice among
alternative constructions.
     Despite these developments, the  distance  between  the
generativist and connectionist outlooks remains substantial,
and differences in underlying philosophy may  well  preclude
an eventual convergence. [6] This is where cognitive grammar
enters the picture, for in a  sense  it  provides  a  bridge
between  the  two.  Certain affinities to connectionism have
already been pointed out. Now I will not assert that  cogni-
tive  grammar  shows  comparable  affinities with generative
grammar; some  radical  adjustments  in  generative  thought
would be needed to bring them into alignment.  Nevertheless,
cognitive grammar is not unreasonably viewed as representing
the  logical  culmination  of all the aforementioned trends,
the kind of natural, unified, and  restrictive  theory  that
might emerge if they ran their full course and certain basic
but erroneous assumptions were abandoned. [7]
     Cognitive grammar's position on  these  issues  can  be
summarized  as  follows:   (i) A grammar is specifically not
conceived as a generative device.  Since  meaning  is  open-
ended  ("encyclopedic"--cf.   Haiman 1980) and based on con-
strual, one cannot envisage  the  recursive  enumeration  of
"all  and only the well-formed sentences [form-meaning pair-
ings] of a language", for this is not  a  well-defined  set.
Moreover,  there is no expectation that any single formalism
or representational format will prove  uniquely  appropriate
for  describing a particular aspect of linguistic structure,
or capable of providing an  exhaustive  account. [8] (ii) The
grammar  of  a  language  is  not thought of as constructing
expressions (giving them as  "output"),  but  simply  as  an
inventory  of  conventional  structures  available for their
categorization.   An  expression's  structural   description
resides  in simultaneous categorization by numerous schemas,
each amounting to a constraint pertaining to some aspect  of
its    organization.    Hence   cognitive   grammar   is   a
"unification-based" model par excellence.  (iii) It is  also
a  usage-based model (Langacker 1988c), by which I mean that
considerable emphasis is placed on specific expressions  and
low-level generalizations.  A speaker's linguistic knowledge
subsumes a vast set of fixed expressions--not  just  lexical
items  in  the  usual sense, but also standard collocations,
formulaic expressions, and all manner of  complex  locutions
representing  the  normal  way  of  phrasing  things  in the
language.  Rules are merely schematizations of  expressions;
they  represent all levels of generality, and coexist in the
grammar with any instantiating expressions that are  learned
and  familiar.   Moreover, since schemas compete for activa-
tion (i.e. for the privilege of  categorization  and  struc-
tural  description)  on  the basis of specificity as well as
entrenchment, lower-level schemas are essential to  linguis-
tic  structure,  serving  as  the primary locus of distribu-
tional information.  (iv)  In  this  framework,  grammatical
structure  reduces  to  the structuring and symbolization of
conceptual content.  Meaning and grammar are not just  inti-
mately associated, therefore, but inherently indissociable.
     To generative theorists  the  non-generative  and  non-
constructive  nature  of  cognitive grammar should no longer
seem exotic or unduly bothersome, and its  greater  emphasis
on  low-level  generalizations  is  primarily  a  matter  of
degree.  What about the idea that rules are just schematized
expressions?  Though more likely to be resisted, this too is
based on notions also encountered in the  generative  tradi-
tion  (structural  templates;  multiple constraint satisfac-
tion); adopting it would simply be a matter  of  recognizing
their universal applicability.  It is much harder to imagine
generativists ever accepting the claim that grammar  reduces
to  symbolic  relationships,  so  fundamental to their world
view is the autonomy thesis.  I would argue,  however,  that
consideration  of  the autonomy thesis has been clouded by a
certain amount of conceptual unclarity and  the  confounding
of  distinct issues.  When these are properly sorted out, it
becomes possible to accommodate the valid observations  that
have been taken as sustaining that thesis, while at the same
time achieving the reduction of grammar to configurations of
symbolic structures. In sum, even though these basic proper-
ties of cognitive grammar may seem quite  radical  from  the
standpoint of generative theory, they have some precedent in
that tradition and are not entirely unresponsive to its con-
cerns.   And it is these properties that offer the realistic
prospect of a connectionist interpretation.


                   3. The Nature of Rules

     If the generativist position is that explicit rules are
needed  for the proper characterization of linguistic struc-
ture, while the connectionist position is that they are not,
I can summarize my own view by saying that I agree with them
both.  The apparent contradiction is resolvable once  it  is
realized that the generative and PDP programs stem from dif-
ferent initial concerns, and that certain statements made in
regard to rules pertain to distinct issues and are therefore
incommensurate.
     Despite its concern with  psychological  questions,  we
should  not  forget  that generative grammar grew out of the
tradition of descriptive linguistics,  whose  goal  was  the
recording  and analysis of a language, followed by an expli-
cit description of its structure for the  benefit  of  other
investigators.   At least in the early days, writing a (par-
tial) transformational grammar of language X was  considered
a reasonable goal, and for some languages such a grammar was
the only substantial source of information.  There was noth-
ing  inherently  odd  about  the  notion  of  dispatching  a
transformational grammarian to the  field  for  purposes  of
describing  an  otherwise unknown language (though fieldwork
tended not to be a primary emphasis).   By  contrast,  faced
with  the  imperative of recording and describing an unknown
language on the verge of extinction, one would hardly  think
of sending a connectionist.
     The task of describing linguistic structure in a usable
form,  for  purposes  of documentation and further analysis,
thus provided the original context in which  the  generative
notion  of explicit rules was conceived and developed.  This
basic descriptive goal has no counterpart in the connection-
ist  program,  which  is solely concerned with the nature of
cognitive processing.  Now if the converse were also  true--
that   is,  if  generativists  were  solely  concerned  with
description, and not at all with cognition--there  would  be
no grounds for conflict.  But of course there are, since one
of  Chomsky's  major  innovations  was  the  proposal   that
linguistic  descriptions be considered hypotheses about cer-
tain aspects of cognitive structure.  With purposeful  ambi-
guity,  he used the term grammar for the cognitive represen-
tation of linguistic structure, as well for  the  linguist's
attempt  to  describe  it.  Our interest here, though, is in
sorting out the issues to see just where the conflict  lies.
I  will  therefore  distinguish  between  an internal and an
external grammar, i.e. between the mental representation  of
language (whatever its nature) and what linguists produce by
way of its characterization.
     The descriptive legacy is, I believe, one factor in the
generative  commitment to explicit rules.  An external gram-
mar has to be reasonably explicit in order  to  fulfill  its
purpose, whether this be practical or intellectual.  In par-
ticular, explicit statements of linguistic regularities pro-
vide  a  characterization  of  what  it is that a processing
model has to account for (irrespective of whether it  incor-
porates any direct analogs of those statements).  One can of
course argue about what form a description ought  to  take--
how  formal  it needs to be, what kinds of constructs should
be posited, how much regularity the data actually  exhibits,
etc.   But  if we confine our attention to external grammars
serving a descriptive function, the validity  and  even  the
necessity  of  formulating  explicit  rules  (of one sort or
another) seems readily apparent.  The controversial issue is
whether  such  rules should also be ascribed to the internal
grammar.
     A second factor in the generative commitment to  expli-
cit rules is the autonomy thesis, the claim that grammar (or
syntax in particular) constitutes a distinct level  or  com-
ponent of linguistic structure with its own representations,
primitives, and organizational principles.  To  sustain  the
autonomy  thesis, it is argued that grammatical patterns and
restrictions cannot be derived as automatic consequences  of
meaning or other independent factors--consequently they have
to be stated explicitly as part of a linguistic description,
and   specifically  learned  in  language  acquisition  (cf.
Newmeyer 1983).
     Let me first point out  that  this  argument  fails  to
establish the autonomy thesis, for it harbors a fallacy. The
basic observation is certainly correct:  while  grammar  can
usually  be seen as motivated on grounds of meaning or func-
tion, its specific detail is not in general predictable  and
must  therefore  be  described  by  linguists and learned by
speakers.  But from this observation one cannot legitimately
draw the further conclusion that grammar (or syntax) consti-
tutes a distinct and autonomous component of the  linguistic
system.   This  further step confuses two issues that are in
principle quite distinct, namely  the  kinds of  structures
that  must  be  posited  and  the  predictability  of  their
behavior. Cognitive grammar is thus coherent and  consistent
in accepting the non-predictability of grammatical structure
while nevertheless denying its  autonomy.   It  acknowledges
that  grammatical  patterns  and restrictions must indeed be
learned and explicitly  described,  but  claims  that  their
proper  characterization  requires  only symbolic structures
(networks of constructional  schemas).   Rather  than  being
autonomous  vis-a-vis  semantics,  grammar  reduces to form-
meaning pairings.
     I thus consider the generative conception  of  explicit
rules  to  be  valid  in  certain  respects  but not all.  A
language does  exhibit  structural  (including  grammatical)
regularities,  many of which are "autonomous" in the limited
sense that they do not follow inexorably as wholly  predict-
able  consequences of other factors.  These regularities can
reasonably be  referred  to  as  rules,  and  to  serve  its
descriptive  function  an  external grammar must state these
rules explicitly.  Moreover, since a speaker  has  to  learn
the  patterns  and  restrictions of his language, comparable
information must somehow be provided by the  internal  gram-
mar,  i.e.   it  must have some kind of cognitive instantia-
tion. The form this knowledge takes, however, may  be  quite
different from what generativists tend to assume; my central
point is that cognitive grammar affords  a  new  and  useful
perspective  on  this  question.  In particular, it offers a
distinct  conception  of  linguistic  rules  that   is   not
inherently incompatible with either the letter or the spirit
of connectionism.
     Cognitive grammar recognizes the importance of explicit
description,  both for practical reasons and as an essential
step  toward  determining  the  mental   representation   of
linguistic structure.  Various formats have been adopted for
descriptive purposes, each revelatory in its own way, but no
single  format is considered uniquely privileged or presumed
capable  of  capturing  every  significant   aspect   of   a
phenomenon  (cf.  fn. 8).  There is also no supposition that
any particular notation  or  descriptive  device  translates
directly  into  claims  about  the basic nature of cognitive
processing.[9] Descriptions are attributed the more  limited
role of elucidating certain regularities that we can reason-
ably expect to  be  reflected  (and  hopefully  discernible)
somewhere  within  the  mental  processing  constitutive  of
linguistic ability.  To the extent that processing regulari-
ties  correspond  to the kinds of patterns linguists seek to
discover and describe, they can be regarded as the cognitive
embodiment  of  linguistic  rules.  Two basic questions then
arise:  What in fact is their  nature?   And  where  can  we
expect to find them?
     Their nature is indicated by two fundamental claims  of
the  theory:  that grammar reduces to configurations of sym-
bolic structures (form-meaning pairings); and that rules are
merely  schematizations of expressions (coarse-grained char-
acterizations representing the commonality that  emerges  at
an  appropriate  level  of  abstraction).   The  first claim
implies that the  rules  of  the  internal  grammar  neither
comprise nor manipulate contentless symbols.  All linguistic
structures are either semantic, phonological,  or  symbolic.
Being  symbolic,  grammatical  structures have both semantic
and phonological value (though it may  be  quite  abstract).
The second claim entails the absence of any qualitative dis-
tinction between rules  and  expressions--apart  from  their
level  of  specificity,  generalizations have the same basic
character as the data they account for.  Now  if  rules  are
conceived  in this fashion, they should pose no problems for
a PDP system (provided that it is  capable  of  representing
both semantic and phonological structures).  The generaliza-
tions extracted by such a system can, I suggest, be  identi-
fied with the schemas posited in cognitive grammar.
     Where can these rules  (schemas)  be  found?   How  can
their  postulation  be  reconciled  with  the  connectionist
assertion--waved like a red cape in front of  generativists-
-that  linguistic regularities can be handled by PDP systems
which make no use of explicit rules?  We must first be clear
about  what  is  actually  intended  by this assertion.  Its
import is twofold:  that a system's specific activity is not
directed  by a central program (a list of statements telling
each unit what to do at each step); and that information  is
stored  exclusively in connection weights (none of which can
be equated with any particular linguistic  construct).   Now
it  would seem that this abnegation of rules precludes their
ascription to the internal  grammar.   There  is  indeed  an
incompatibility  if  one  insists  that  rules  are directly
analogous to the instructions of a computer program, or that
rules  are  stored  as  such  at  the  most  basic  level of
representation.  Rules can, however, be accorded a very dif-
ferent  status,  in  which case they do not run afoul of the
connectionist prohibition but are simply incommensurate with
it.   In cognitive grammar, rules are conceived as regulari-
ties in the mental  processing  constitutive  of  linguistic
ability.   They are consequently emergent rather than funda-
mental; instead of being separately stored or represented in
the  form of instructions, they are inherent in the system's
processing activity.  Interpreted as recurrent  patterns  of
neural  activation, rules are wholly consistent with connec-
tionist principles.
     Let us consider this conception of rules and their  PDP
implementation  in  somewhat more detail.  At the most basic
level, linguistic knowledge is stored in connection weights.
Neither rules nor any other linguistic elements are directly
or individually discernible at that level, however.  To find
the  cognitive  correlates of linguistic constructs, we must
instead look at higher levels of organization, and  specifi-
cally  at  the  patterns  of activation constrained by those
weights.   The  evocation   of   a   particular   linguistic
structure--be   it  semantic,  phonological,  or  symbolic--
resides in the occurrence of a particular pattern of  neural
activation.   This  pattern  may be relatively simple, or it
may  be  extraordinarily  complex,  comprising   elaborately
architectured  cascades  of  activity involving many popula-
tions of units over a substantial span of processing time. [10]
But  regardless  of  complexity, a pattern is describable as
either a location in the state space defined by the  activa-
tion  levels  of  the  system's  units, or else a trajectory
through state space (i.e. a series of locations).
     The structures that concern us are  schemas  and  their
instantiating  expressions.  It is crucial that the notation
employed for their relationship, namely A===>B, not be  con-
strued  as  making  a specific claim about the nature of its
cognitive representation.  The practice  of  using  distinct
symbols  for a schema and its instantiation is helpful (even
necessary) for analytical and descriptive purposes,  but  it
is  not  meant  to imply that they are discrete and separate
psychological entities.  Rather, I conceive of a  schema  as
being  immanent  in  its instantiations, i.e. as inherent in
(and shared by) the activation patterns in which its instan-
tiations reside.  What does it mean, exactly, for one struc-
ture (or activation pattern) to be immanent in another?  One
way to interpret it in connectionist terms pertains to loca-
tions and trajectories in state space. [11] A location can be
characterized  with  varying  degrees  of  precision,  being
point-like or diffuse depending on whether activation levels
are  specified quite narrowly or only as falling within cer-
tain bands of values.  And as a series of locations, a  tra-
jectory  can similarly be characterized with varying degrees
of exactitude (as either line-like or swath-like).   We  can
thus  describe  the  relationship  between  a schema and its
instantiations as one of inclusion in state space:  a schema
corresponds  to  a diffuse region (or swath-like trajectory)
in state space,  and  each  instantiation  to  a  point-like
region (or line-like trajectory) contained within it.
     On this account, the extraction of schemas  is  a  non-
mysterious  process  which  results in essentially automatic
fashion from  the  use  of  instantiating  expressions.   An
expression's  occurrence  tends to strengthen the connection
weights responsible for the pattern of  activation  that  it
comprises,  and thus to facilitate the subsequent occurrence
of another pattern in  the  same  general  region  of  state
space.   Hence  the frequent use of expressions sufficiently
similar that they  cluster  in  such  a  region  induces  an
adjustment  of  the  responsible  weights  which renders the
occurrence of any pattern falling within  that  region  more
likely  or easily elicited than it would otherwise be.  That
adjustment constitutes the extraction of a schema.  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 diffuse region that encompasses it.  Moreover,
by virtue of facilitating a pattern falling anywhere  within
that  region,  a schema has an active, causal role in cogni-
tive processing--it is not epiphenomenal, unless one  wishes
to say that all linguistic entities are epiphenomenal.
     Let me conclude by noting certain challenges that  cog-
nitive  grammar poses for connectionist modeling.  First, to
be linguistically viable a PDP system  must  be  capable  of
representing  structured conceptualizations of extraordinary
intricacy (see especially Langacker to appear).  Second,  it
must allow distinct structures to be co-activated and linked
by correspondences while  to  some  degree  retaining  their
separate  identity.  This is needed for the characterization
of grammatical constructions (cf. Figs. 1 and 2), metaphori-
cal  structuring  (Lakoff  and Johnson 1980), and correspon-
dences between elements of different mental spaces  (Faucon-
nier  1985).   Finally, a linguistically adequate PDP system
will have to accommodate the many  dimensions  of  construal
(Langacker  1988b),  including  such  factors  as profiling,
figure-ground organization, and vantage point.  While I have
little doubt that these are all susceptible to connectionist
treatment, not much attention has  thus  far  been  accorded
them.   Serious consideration of these matters would greatly
facilitate a mutually instructive interaction  between  con-
nectionism and cognitive linguistics.


                         Footnotes


[1] I have been developing this theory since 1976. By now, it
has  been described in numerous works and applied to a broad
variety of representative linguistic  phenomena.   See,  for
example,  Casad  1982;  Casad and Langacker 1985; Cook 1988;
Hawkins 1984; Janda 1984, to appear; Langacker  1982,  1984,
1985, 1986a, 1986b, 1986c, 1987a, 1987b, 1987c, 1988a, 1990,
in press, to appear; Lindner  1981,  1982;  Maldonado  1988;
Poteet  1987;  Rice  1987a,  1987b, 1988; Rudzka-Ostyn 1988;
Smith 1985, 1987; Tuggy 1980, 1981, 1986,  1989;  Vandeloise
1984, 1985a, 1985b, 1986, 1987.

[2] Crucially, this sense must not  be  confused  with  that
implied  in  speaking of symbolic (as opposed to connection-
ist) accounts of cognitive processing.

[3] This particular class happens not to be structurally signi-
ficant,  but  a  comparable  class might very well be.  In a
given language, for instance, the class of  body-part  terms
conforming  to a certain phonological pattern might all form
their plurals in the same way.

[4] Another way to put it is that a schema allows a wider range
of  values along some or all parameters of the characteriza-
tion.

[5] That is, a noun is  characterized  schematically  as  an
expression  that designates a thing (defined abstractly) and
is manifested phonologically as  any  kind  of  phonological
sequence.

[6] I refer here to a true rapprochement,  not such half-way
measures as implementing a standard generative analysis in a
PDP model, or using rule-based  and  connectionist  accounts
for  different  components (e.g. syntax vs. lexicon, or com-
petence vs.  performance).

[7] As a historical note, I should point out that the  basic
framework  of  cognitive  grammar has been in place for well
over a decade--it has for the most part  anticipated  rather
than followed these trends.

[8] Though I would not want to push the metaphor too far, it is
useful  in  this regard to think of language as being analo-
gous to a biological organism--however thoroughly  it  might
be  described,  further characterization (in finer detail or
from another perspective) can still be both valid and  reve-
latory.  (Cf.  Langacker to appear, 12.1.)

[9] Hence a pictorial representation does not  imply  that  the
brain  stores  information  in  the form of pictures.  Simi-
larly, a formulaic representation does not imply that cogni-
tive  processing  involves  formulas  or the manipulation of
discrete symbols.

[10] A pattern of either sort is referred  to  as  a  cognitive
event in Langacker 1987a.

[11] This interpretation was suggested to me by Steve Poteet.


                        Bibliography


Casad, Eugene H.  1982.   Cora  Locationals  and  Structured
    Imagery.  San Diego:  UCSD doctoral dissertation.

-----, and Ronald W. Langacker. 1985.  '"Inside"  and  "Out-
    side" in Cora Grammar'.  IJAL 51.247-281.

Chomsky, Noam.  1957. Syntactic Structures. The Hague:  Mou-
    ton.

-----.   1965.  'Three  Models  for   the   Description   of
    Language',  in  R.  Duncan  Luce  et al. (eds.),  Readings 
    in Mathematical  Psychology,  vol.  2,  p.105-124.   New  
    York: Wiley.

-----, and George A. Miller.  1963.   'Introduction  to  the
    Formal  Analysis of Natural Languages', in R. Duncan Luce et
    al. (eds.), Handbook of  Mathematical  Psychology,  vol.  2,
    p.269-321.  New York and London:  Wiley.

Cook, Kenneth W.  1988.  A Cognitive Analysis of Grammatical
    Relations,  Case,  and  Transitivity  in Samoan.  San 
    Diego: UCSD doctoral dissertation.

Fauconnier, Gilles.  1985.  Mental Spaces:  Aspects of Mean-
    ing Construction in Natural Language.  Cambridge, Mass. and
    London:  MIT Press/Bradford.

Haiman,  John.   1980.  'Dictionaries  and   Encyclopedias'.
    Lingua 50.329-357.

Hawkins, Bruce W.  1984.  The Semantics of English  Spatial
    Prepositions.  San Diego:  UCSD doctoral dissertation.

Janda, Laura A.  1984.  A Semantic Analysis of  the  Russian
    Verbal Prefixes ZA-, PERE-, DO-, and OT-.  Los Angeles: UCLA
    doctoral dissertation.

----.  To appear.  A Geography of Case Semantics:  The Czech
    Dative and the Russian Instrumental.

Lakoff, George.  1987.  Women, Fire, and  Dangerous Things:
    What  Categories Reveal about the Mind.  Chicago and London:
    University of Chicago Press.

----, and Mark Johnson. 1980. Metaphors We Live By.  Chicago
    and London: University of Chicago Press.

Langacker, Ronald W.  1982. 'Space  Grammar,  Analysability,
    and the English Passive'. Language 58.22-80.

-----.  1984.  'Active Zones'.  BLS 10.172-188.

-----.  l985.  'Observations  and  Speculations  on  Subjec-
    tivity',  in  John Haiman (ed.), Iconicity in Syntax, p.109-
    150.  Amsterdam and Philadelphia:  John Benjamins.

-----.  1986a.  'Abstract Motion'.  BLS 12.455-471.
    
-----.  1986b.   'An  Introduction  to  Cognitive  Grammar'.
    Cognitive Science 10.1-40.

-----.  1986c.   'Settings,  Participants,  and  Grammatical
    Relations'.   Proceedings  of  the  Annual  Meeting  of  
    the Pacific Linguistics Conference 2.1-31.

-----.  1987a.  Foundations of Cognitive  Grammar,  vol.  1,
    Theoretical  Prerequisites.   Stanford:  Stanford University
    Press.

-----.  1987b.  'Nouns and Verbs'.  Language 63.53-94.

-----.    1987c.    'Grammatical   Ramifications   of    the
    Setting/Participant Distinction'.  BLS 13.383-394.

-----. 1988a.  'Autonomy, Agreement, and Cognitive Grammar',
    in  Diane  Brentari  et al. (eds.), Agreement in Grammatical
    Theory, p.147-180.  Chicago:  Chicago Linguistic Society.

-----.   1988b.   'A  View  of  Linguistic  Semantics',   in
    Rudzka-Ostyn, p.49-90.

-----.  1988c.   'A  Usage-Based  Model',  in  Rudzka-Ostyn,
    p.127-161.

-----.  1989.  'Absolute Construal', in F.  J. Hevyaert  and
    F.  Steurs  (eds.), Worlds Behind Words: Essays in Honour of
    Prof. Dr. F. G. Droste  on  the  Occasion  of His  Sixtieth
    Birthday, p.65-75.  Leuven:  Leuven University Press.

-----.  1990.   'Subjectification'.   Cognitive  Linguistics
    1.5-38.

-----.  In press.  Concept, Image, and Symbol:   The  Cogni-
    tive Basis of Grammar.  Berlin:  Mouton de Gruyter.

-----.  To appear.  Foundations of Cognitive  Grammar,  vol.
    2,  Descriptive Application.  Stanford:  Stanford University
    Press.

Lindner, Susan.  1981. A Lexico-Semantic Analysis of English
    Verb-Particle Constructions with UP and OUT. San Diego: UCSD
    doctoral dissertation.

-----.  1982.  'What Goes Up doesn't Necessarily Come  Down:
    The Ins and Outs of Opposites'.  CLS 18.305-323.

Maldonado, Ricardo.  1988.  'Energetic Reflexives  in  Span-
    ish'.  BLS 14.153-165.

Newmeyer, Frederick J.  1983.  Grammatical Theory:  Its Lim-
    its and Its Possibilities.  Chicago and London:  University
    of Chicago Press.

Poteet, Stephen.  1987.  'Paths Through  Different  Domains:
    A  Cognitive Grammar Analysis of Mandarin Dao'.  BLS 13.408-
    421.

Rice, Sally. 1987a.  'Towards a Transitive  Prototype:  Evi-
    dence from Some Atypical English Passives'.  BLS 13.422-434.

-----.  1987b.  Towards a Cognitive Model  of  Transitivity.
    San Diego:  UCSD doctoral dissertation.

-----.  1988.  'Unlikely Lexical Entries'.  BLS 14.202-212.
    
Rudzka-Ostyn, Brygida  (ed.)   1988.   Topics  in  Cognitive
    Linguistics.  Amsterdam and Philadelphia:  John Benjamins.

Smith, Michael B.  1985.  'An Analysis of German Dummy  Sub-
    ject  Constructions'.   Proceedings of the Annual Meeting of
    the Pacific Linguistics Conference 1.412-425.

-----.  1987.  The Semantics of  Dative  and  Accusative in
    German:   An Investigation in Cognitive Grammar.  San Diego:
    UCSD doctoral dissertation.

Tuggy, David. 1980. '|Ethical Dative and Possessor  Omission
    Si,  Possessor  Ascension  No!'  Work  Papers  of the Summer
    Institute of Linguistics, University of North Dakota  24.97-
    141.

-----.  1981. The Transitivity-Related Morphology of  Tetel-
    cingo  Nahuatl:  An Exploration in Space Grammar. San Diego:
    UCSD doctoral dissertation.

-----.  1986.  'Noun Incorporations in  Nahuatl'.   Proceed-
    ings  of  the  Annual  Meeting  of  the  Pacific Linguistics
    Conference 2.455-469.

-----.   1989.   'The  Affix-Stem  Distinction  in   Orizaba
    Nahuatl'.  Linguistic Agency University of Duisburg.

Vandeloise, Claude.  1984.  Description of Space in  French.
    San Diego:  UCSD doctoral dissertation.

-----.  1985a.  'Au-dela des  Descriptions  Geometriques  et
    Logiques   de  l'Espace:   Une  Description  Fonctionnelle'.
    Lingvisticae Investigationes 9.109-129.

-----.  1985b.  'Les Prepositions Sur/Sous  et  la  Relation
    Porteur/Porte'.  Leuvense Bijdragen 74.457-481.

-----.  1986.  L'Espace en Francais.   Paris:   Editions  du
    Seuil.

-----.   1987.    'La   Preposition   a   et   le   Principe
    d'Anticipation'.  Langue Francaise 76.77-111.