This paper was presented at the 1993 Mid-America
Linguistics Conference, and appears as pp. 1-22
of the Proceedings of that conference, edited by
Jule Gomez de Garcia and David Rood, available from
the Department of Linguistics, University of Colorado.

 

            Grammaticalization and Linguistic Theory
                         Scott DeLancey
                      University of Oregon

     My main purpose in this paper is to summarize various ways
in which our current understanding of the phenomena of
grammaticalization should inform linguistic theory more than is
generally the case.  I will focus on phenomena which I will argue
that, in the light of our current understanding of diachrony,
modern theory takes far too much account of.  What I intend to do
here is:  1) quickly outline the central phenomenon of
grammaticalization, 2) show how much of the stuff of modern
theories of phrase structure and word order typology follows
pretty directly from an understanding of the workings of
grammaticalization, and 3) argue that this diachronic explanation
constitutes a true explanation for the actual facts, and renders
many aspects of a theoretical account of heads and X' syntax
unnecessary and irrelevant.  My purpose is not primarily to
attack any particular conception of the overall task of
linguistic theory, but to suggest ways that various types of
empirical data should constrain any theory.

1. Introduction

I don't intend to address the specifics of any particular current
theory here; my interest is in the broad set of phenomena which
are widely accepted among contemporary syntacticians as the stuff
of which syntactic theory must be constructed.  In particular, I
will be looking at questions of constituent structure and
ordering and head-modifier relations of the kind dealt with in X'
theory.
     One obvious question to ask about any set of constructs
posited as primitives for linguistic theory is, where does it all
come from?  Why should language be this way rather than some
other?  Specifically, a problem that arises in connection with
any a priori set of categories (e.g. Spec, Head, Adposition,
Adj), or structures (e.g. XP) is: Why are these the categories? 
Where do they come from?  How do we know?  Questions of this sort
have been rather out of fashion for several generations now--the
Structuralist answer can be parodied as "we don't even know that
it is this way, only that to our eyes it looks like it", while
that of their Generativist inheritors is essentially "it's that
way because that's the way it is".
     I will suggest in this paper that substantive explanations
for many syntactic phenomena follow from an understanding of
various diachronic processes, particularly those currently
discussed under the general heading of grammaticalization.  The
argument can be framed in metatheoretical terms, i.e. that my
explanations are better than many others because they are
substantive rather than formalistic.  For those who find this
argument beyond comprehension, however, there is a more modest
argumment here--that the diachronic explanations that I will
present to explain the nature of various syntactic categories and
their patterns of combination are sufficient to account for the
data, and that formalistic explanations of the kind current in X'
theory and similar approaches are therefore superfluous and
unnecessary.  In other words, even if you believe in formal
theory as explanation, I hope to convince you that the correct
theory should not be concerned with accounting for the linguistic
patterns which I will be dealing with.

2. Grammaticalization

The nature and significance of grammaticalization have been
widely discussed in recent years (see Traugott and Heine 1991,
Heine et. al. 1991, Hopper and Traugott 1993, and references
therein).  A useful preliminary definition was given considerably
earlier by Kurylowicz:

     Grammaticalization consists in the increase of the
     range of a morpheme advancing from a lexical to a
     grammatical or from a less grammatical to a more
     grammatical status, e.g. from a derivative formant to
     an inflectional one. (Kurylowicz 1965:69)

Obvious examples are such things as the development of
auxiliaries from complement-taking verbs (as in the case of the
English modals), but the same general process is involved in, and
the term also applied to, the development of adpositions from
serial verbs or relational nouns, or of tense markers from
already-grammaticalized aspectual constructions.  An inadequacy
in this provisional formulation is that it restricts its
attention to lexical items; in current work the term
"grammaticalization" is used in a broader sense, referring also
to shifts in function of syntactic constructions.  For example,
shifts from paratactic to hypotactic structure, e.g. relative
constructions arising from conjunction, or the presumed
development of English that-complements from a bisentential structure:

     I saw that.  He came.  >  I saw that he came.

represent examples of grammaticalization in the broad sense which
I will be using here.

2.1. The process of grammaticalization

The starting point of the process of grammaticalization is a
productive construction:  NP with genitive dependent, matrix with
complement clause, conjoined or chained clauses, etc.  The
precondition for grammaticalization is that there be some lexeme
or lexemes which occur frequently in this construction for some
semantic/pragmatic reason, e.g. phasal or modal complement-taking
verbs like 'finish' or 'want', or semantically nonspecific
transitive verbs like 'use' or 'hold' conjoined or serialized
with more specific verbs.  This usually involves a lexeme with a
very general meaning, which can therefore be used in a wide range
of contexts.
     This situation, in which a particular construction--a
productive syntactic structure with a specific lexeme in a
specific slot--is a useful and regularly-used locution in the
language, is the initial point of grammaticalization.  The next
step is for the lexeme to undergo a certain amount of "semantic
bleaching", or, put another way, for the locution to be used in
an extended set of contexts, including some in which the literal
meaning of the lexeme is not applicable:

     ... without an external change of its exponent a
     category may undergo important internal (functional)
     changes due simply to an extension of a limitation of
     its range.  The logical principle of the mutual rela-
     tion of range and content has to be applied in such a
     case:  the increase of the range of a given category
     entails the impoverishment of its content, and vice
     versa. (Kurylowicz 1965:57-8)

This can often be observed even in forms which have not undergone
any formal grammaticalization.  For example, finish in English is
often used to mean simply 'stop', so that it makes perfectly good
sense to say I've finished writing for today, even when the
project on which I am working is far from completed.  A more
grammaticalized example is the Tibetan verb sdad 'sit', which is
currently developing into a progressive auxiliary.  It can now
occur in sentences such as:

     kho las=ka byas(-byas) sdad-pa red
     he  work   did  (NF)   sit  PERF
     'He was running, kept running, was always running.'

The semantic incompatibility of 'run' and 'sit' is such that sdad
cannot possibly be interpreted with its original lexical sense
here.  Therefore the structure cannot be a simple case of verb
serialization, and thus represents an early stage of gram-
maticalization.
     The first perceptible stage of grammaticalization then
occurs when this lexeme begins to "decategorialize" (cp. Heine
et. al.), i.e. to lose the morphosyntactic behaviors
characteristic of its original category.  For example, in the
English construction on top of NP, top, while clearly a noun in
origin, is un-nounlike in several respects.  It lacks an article,
and it cannot pluralize:  we can say on top of all the houses,
with top as a relational noun, or on the tops of all the houses,
with top as an ordinary noun, but we cannot pluralize the
relational noun:  *on tops of all the houses.  Let us take the
standard X' characterization of the difference between Noun and
Preposition as being the value for the feature N, so that a Noun
such as top is  +N, -V, and a Preposition such as atop -N, -V. 
What is the status of the semi-grammaticalized top of (4)? 
According to the feature theory it must be one or the other, i.e.
either +N or -N; in fact it is somewhere in between.
     The most essential point for my subsequent argument is the
fact that, in grammaticalization, functional shift leads, and
structural change follows.  Note that top in its relational noun
use is already semantically bleached, in that the top of NP
necessarily refers to a specific part of the object, while on top
of NP simply refers to whatever side of it is uppermost at the
moment.  (I.e. if a refrigerator is lying on its back, something
resting on the door, which is the uppermost surface, can be des-
cribed as on top of the  refrigerator, but not on the top of the
refrigerator).  This is the normal pattern; it is easy to find
lexemes, like English finish, which have undergone some semantic
bleaching without any morphosyntactic decategorialization, but
decategorialization does not occur without some prior and causal
functional shift, in the direction either of grammaticalization
or lexicalization.  A lexical form or construction will always
begin to specialize in a potentially grammatical function, and in
this connection to show functional shift such as semantic
bleaching, while still manifesting its original syntactic
behavior.  Ensuing changes in syntactic behavior are motivated by
the functional shift more than (probably rather than) by any
structural considerations, and a grammaticalized form may retain
syntactic features associated with its original category
indefinitely.

2.2. Syntactic categories

One implication of the gradual nature of grammaticalization is
that we should expect to find in any language a substantial
number of minor morphosyntactic categories and individual forms
with unique and idiosyncratic morphosyntactic behavior.  This is
indeed the case; we regularly find cases of grammaticalizing
forms occupying intermediate categorial status.  We have seen
already the example of semi-grammaticalized English top (a semi-
productive phenomenon; cp. in place of, in back of, etc.), which
is not yet a preposition but is losing its noun behaviors.  As
another example, in modern English we find a range of more and
less verb-like characteristics among grammaticalizing "quasi-
modals" such as used to, want to, ought to, etc.  (cf. Bolinger
1980).  Compare the lexical verb-like behavior of have with the
auxiliary-like behavior of ought with respect to question
formation and negation:

     Do you wanna go along?

     *Want you to go along?

     *Do you ought to go along?

     Had(n't) you ought to go along?

     (#) Ought you to go along?

And with these compare the equally anomalous behavior of usta,
with its defective tense distribution:

     Did you usta go along?

     *Do you usta go along?

     *Used you to go along?
     
Such data were accommodated without serious difficulty in
structuralist descriptive models, but are awkward, at best, for
any theoretical framework which assumes that any language should
be describable by a grammar written in terms of a predefined set
of categories.  This assumption is implicit in X' theory, as
noted by Kornai and Pullum (1990:27), who point out the profusion
of undefined minor categories to which working X' grammarians are
nevertheless often forced to resort.  It is sometimes made
explicit; cf. the vehemence with which Lightfoot (1991) denies
the possibility of gradual recategorization.  He is willing to
countenance graduality across a set of forms, such that for
example some of the English modals may have recategorized earlier
than others, but absolutely not with respect to specific
constructions.  That is, it is inconceivable within Lightfoot's
view of the world that there could have been a time in the
history of English when--or ever any speaker of English for whom-
-will, say, was neither a verb nor an Aux, but an idiosyncratic
category; likewise usta and oughta must be, at least for any
given speaker, unambiguously members of one or another recognized
syntactic category.
     I am not going to pursue here the question of the most
appropriate way of dealing with this problem in linguistic
theory, beyond saying that in the face of data like Bolinger's
(and similar data can be easily adduced from any language that I
know anything about), acceptance of any theory based on strict
categoriality can only be accomplished as an act of faith.  But
it will become important for my subsequent argument to keep in
mind the fact that as forms or constructions grammaticalize, they
will often show mixed categorial behavior.

3. Grammaticalization, order, and constituency

It has been noted for over a century now that there are cross-
linguistic patterns in constituent ordering, such that, for
example, OV languages tend to have postpositions, while OV
languages have prepositions.  The first discussion of word order
typology that I am aware of is in the work of the English
Sinologist Terrien de Lacouperie (1887); the first systematic
typological classification scheme in terms of word order patterns
is in Schmidt 1926.  TesniŠre (1959) offered an explanatory
account of the patterns in terms very compatible with more modern
work, categorizing languages as "centrifugal" or "centripetal"
according as they prefer head-dependent or dependent-head
ordering patterns.  But these phenomena attracted little
attention in American linguistics until they were discussed (in
English) by Greenberg in the mid-Sixties.

3.1.  The typological data

The basic observation is that within a language we find, with
greater than chance frequency, cross-categorial correlations like
the following:

          OV                       VO
          N Adp                    Adp N
          V Aux                    Aux V
          S COMP                   COMP S
=====================================================
          Adj N                    N Adj
          Gen N                    N Gen
          RC N                     N RC

Not all of these correlations have stood up to subsequent
reexamination; in particular, convincing evidence for a
correlation between order within NP and within V''--i.e. the
claimed correlation between verb-object order and the ordering of
modifiers and their head nouns--is lacking (Dryer 1988).  But
robust data provide prima facie evidence that at least the
correlations which I have listed above the line are valid, which
implies that the ordering of verb and object and of adposition
and object, for example, are at some level of description the
same phenomenon.  But what phenomenon, and at what level of
description?
     There was considerable discussion of this problem in the
typological literature during the 70's, most of it based on the
assumption that the commoner, "consistent" patterns must be
unmarked and "normal", so that linguistic theory must be
constructed so as to predict them.  Many arguments were presented
in terms of "operator-operand" or "functor-argument" relations,
claiming that languages tend to have a consistent ordering of
operator and operand, or whatever, or in terms of a principle of
consistent ordering of head and dependent (first suggested by
TesniŠre 1959, and more recently by Hawkins 1984).
     There is a major empirical problem with such explanations,
in that, while these correlations are common enough to be
undoubtedly significant, they are not universal--not all
languages conform to all the patterns.  There has thus been much
discussion of "consistent" or "harmonic" and "non-harmonic" lan-
guage types, all deriving from the fundamental assumption that
these cross-categorial correlations must reflect something about
the nature of linguistic competence.  (This distinction is also
made by TesniŠre, who distinguishes "accus‚e" and "mitig‚e"
examples of each of his types).  This is the essential problem
with any theoretical explanation of these facts:  the
correlations are far too consistent to be ignored, but simply not
reliable enough to support an account that builds them directly
into the theory.  So we saw in the 70's a certain amount of
misguided historical work arguing that various languages were
"transitional" or in some way carrying traces of an earlier word
order type.  The line of reasoning here is, if "consistent"
patterns are more common, it must be because they are somehow
favored, which presumably means easier to learn, and we need to
build this into our theory of linguistic competence.

3.2. X' theory

One claimed advantage of X' theory is the ease with which it
captures intralinguistic cross-categorial parallelisms of this
sort, and many scholars now talk about "parameter setting" with
respect to X' theory as a way of accounting for cross-linguistic
variation in constituent order patterns (e.g. Travis 1989, Lieber
1992).  I would argue that while these generalizations do require
and are susceptible of explanation by the field of linguistics,
they are not "linguistically significant" in Jackendoff's (1977)
sense, i.e. are not reflections of basic linguistic patterns that
are somehow wired into the language learner.  Rather, they are
explainable in terms of what Greenberg (1969) has called
"processual generalizations", i.e. cross-linguistically recurrent
patterns of diachronic change.  If this is correct, of course, it
means that the attempt to construct a linguistic theory in such a
way as to predict the prevalence of harmonic patterns without
reference to diachrony is misguided.
     It is evident that, at least as far as surface patterns go,
linguistic theory can at best treat these consistent patterns as
unmarked, since "inconsistent" patterns are widely and robustly
attested.  There are several strategies for accounting for the
marked patterns.  One is to use movement rules, i.e. to argue
that underlyingly our anomalous language is really consistent,
but it has additional rules (hence a more marked grammar) which
then change things around in surface structure.  This is one
(seriously entertained!) explanation for why English relative
clauses occur on the opposite side of their head noun from
adjectival or genitive modifiers.  An alternative approach is to
relax theoretical restrictions, and "set parameters" separately
for different domains, somehow building into our theory the claim
that having the same settings for each domain is more natural
than setting them independently.  The most popular in current
work involves trying to predict some aspects of ordering from
some other module of GB theory, which then has priority over the
basic ordering parameters.

3.3. The explanation from diachrony

The essential problem with the X' account of ordering facts, as
with other formalistic "explanations", is that it is not an
explanation.  Without some story about why there should be such
categories as SPEC, e.g., in language, and why determiners and
complementizers should belong to it, the X' account amounts to
nothing more than the statement that language is the way it is
because God made it that way.  This is not only unsatisfying, it
is empirically inadequate as well, since there are well-document-
ed exceptions to all such generalizations.  If "because that's
the way it is" is unsatisfactory as an explanation, "because
that's the way it is, except sometimes" is even worse.
     I want to propose here a radically different approach. 
Rather than assuming a priori that any and all attested patterns
must somehow be licensed by some formal mechanism, let us return,
at least for the sake of argument, to an older way of thinking,
and assume that any attested pattern reflects a particular course
of historical development.  Commoner patterns, then, are commoner
not because something in the wiring of the left hemisphere, or
the mathematics of formal languages, makes them easier to learn
or simpler to generate, but because they reflect commoner
historical processes.
     Several scholars (Givon 1979, Aristar 1991) have already
suggested for at least some of these patterns a historical
explanation which obviates the need to build these correlations
into a theory.  The most obvious case is that of the correlations
between adpositional phrases and VP's and NP's.  The essential
question is, what do verbs and adpositions have in common? 
Grammaticalization theory provides an answer--they behave
similarly because many adpositions were once verbs.  Consider the
status of the coverb ?aw 'take, take up, pick up' in Thai; a
situation with parallels in other Southeast Asian and West
African languages.  This functions as an instrumental marker in
sentences like:

          ?aw  takiab    kin kwaytiaw
          take chopstick eat noodles
          '[] eat noodles with chopsticks'

Such a sentence is exactly parallel syntactically to a clause
chain referring to a sequence of events, such as:

          ?aw  nangsyy paj rongrian
          take book    go  school
          '[] pick up []'s books and go to school'

Adpositions with this origin will necessarily occur on the same
side of their argument as does a verb--what imaginable diachronic
process would occur to change the order?
     Thus in a language in which adpositions have this origin, a
mismatch between these two orders could come about only as a
result of a secondary shift in verb-object order.  Then there is
no longer a need for a special explanation of why verbs and
adpositions occur on the same side of their argument; this is a
natural consequence of the history of adpositions.  If we find
them in differing ordering patterns, we need a special
explanation, since it appears as though reordering took place,
and this would need some explanatory mechanism.  Similarly, a
correlation between adposition-argument order and the ordering of
head nouns and genitive modifiers is unproblematic--those
adpositions that do not originate in serial verbs originate in
relator noun constructions like on the top of the NP, in which
the erstwhile head noun becomes an adposition, and remains on the
same side of its argument as other head nouns are of their
genitive modifiers.
     Conspicuously missing from this account is any explanation
for the claimed correlation between OV and Gen N order.  If such
a correlation actually exists, it is not susceptible to this kind
of diachronic explanation.   In fact, however, there may be
nothing to explain here; see Dryer 1988 for some suggestive work
showing that the putative correlation between the ordering of
adjective and noun and of object and verb is illusory.

4. Tibetan N'' -- A speculative case study

What I am suggesting here is that the only reason for the cross-
categorial ordering patterns which have been observed is that
they fall out from commoner diachronic processes.  Thus the
Greenbergian data do not provide any evidence that some ordering
patterns are more natural or learnable than others.  Now I want
to look at a specific problematic example, that of order within
NP in modern Tibetan.  Tibetan NP shows a pattern which within
the word-order typology framework must be considered highly
marked, and which in fact would be impossible to accommodate in
an un-kludged version of X' theory; I will suggest that data such
as these support a claim that, in accounting for ordering
patterns, diachronic explanation is both necessary and
sufficient.

4.1.  Ordering generalizations within NP

The explanation for some of the cross-linguistic correlations of
ordering within NP is fairly straightforward.  It has been
remarked at least since Jespersen (1924), and recently
elaborately by Dixon (1977), that adjectives tend to be related
(both by synchronic behavior and by history) to either nouns or
verbs.  The existence of languages in which adjectives show the
same ordering behavior with respect to N as genitives suggests
noun-like adjectives, which came into their position as genitive
modifiers.  The existence of languages in which adjectives
pattern with relative clauses suggest verb-like adjectives; we
would hypothesize such a pattern as having originated in
something like Chinese, where the functional equivalent of
English adjectives are a subclass of verbs, and typically modify
nouns in the same relative construction as verbs do:

     zhong shuiguo-de nongren
     grow fruit-GEN   farmer
     'farmers who grow fruit'

     shufu-de        yizi
     comfortable-GEN chair
     'a comfortable chair'

     hong(-de) hua
     red       flower

As the third example shows, stative (or "adjectival") verbs in
Chinese can in some circumstances modify a noun without relative
marking.  This represents a further development of the relative
clause construction in the second example, and illustrates how
the same kind of diachronic argument can account for this
correlation.
     So far, this argument suggests that we would be unlikely to
find a language in which adjectives pattern with neither
genitives nor relative clauses.  X' theory makes slightly
different predictions.  Since adjectives and relative clauses are
both modifiers, they should always appear on the same side of the
head.  Genitives, if they are under the SPEC node, should show
the same ordering in a restrictive version of X' theory in which
XP must be either head-initial or head-final.  In a more relaxed
version of X' theory in which head-initial or head-final ordering
can be separately specified for each level, there would be no
reason to expect any correlation between the position of
genitives and that of adjectives or relative clauses:

     Since it is possible in a language for specifiers and
     modifiers to appear on one side of the head and
     complements on the other ... or for specifiers to occur
     on one side of the head and both complements and
     modifiers on the other ... it appears that in fact we
     need three separate Head Initial/Final parameters, one
     concerning the position of complements, a second
     concerning the position of specifiers, and a third
     concerning the position of modifiers.  (Lieber 1992:35)

But languages do exist in which not all modifiers appear on the
same side of their head, even within N''.  English, of course, is
one, where adjectives and genitive modifiers precede their heads
but relative clauses follow.  In the rather scant literature on
this problem within an X' framework this is usually dealt with by
a movement rule, i.e. it is assumed that in underlying structure
all modifiers within N'' are generated on the same side of their
head, and then either adjectives (Stowell 1981) or relative
clauses (Lieber 1992) are moved to their actual surface position. 
Such analyses are, to put it mildly, somewhat ad hoc.
     In a language in which relative clauses are nominalized
genitive modifiers of their head I would expect them to order
with genitive modifiers, as they do e.g. in Tibetan, Chinese, and
Japanese.  Where they are not, there is no reason to expect that
particular correlation--look at English, for example.  In a
language like Chinese, where "adjectives" are a subset of verbs,
and adjectival modification thus a subspecies of relative clause,
we would expect adjectives and relative clauses to pattern
together, as they do.  Where adjectives are a distinct syntactic
category, as in English, there is no particular reason to expect
this correlation.

4.2.  Tibetan modifiers

Fudgeless X' theory can account only for languages in which
adjectives and relative clauses pattern together; a strict
version of X' theory would predict that genitives should also
pattern with the modifiers, while a more relaxed version would
predict no correlation between the position of specifiers and
modifiers.  Diachronic explanations are readily available for
languages in which adjectives pattern with relative clauses, and
for those in which they pattern with genitives (e.g. English). 
Since both patterns (and others; see below) are widely attested,
X' theory is empirically inadequate some tinkering;  a language
like English, with Adj=GenitiveRC, provides prima facie evidence
in favor of my account, and can be accommodated within X' theory
only through some ad hoc kludge.
     Similar problems arise in Tibetan, where we find a pattern
not predicted either by X' theory or by the normal diachronic
paths of development of adjectives: relative clauses and
genitives pattern one way, and adjectives the other.  Tibetan is
a verb-final language, and outside of NP the ordering principles
are consistent with SOV typology: postpositions, post verbal AUX,
sentence-final COMP, etc.  In other words, except for NP, Tibetan
is unambiguously head-final.  Within NP, genitive modifiers and
relative clauses precede the N; other categories--adjectives,
numerals and quantifiers, and determiners--follow:

     stong=thung dkar=po cig
     shirt       white   a
     'a white shirt'

     blo=bzang-gi stong=thung de
     Lobsang-GEN   shirt      that
     'that shirt of Lobsang's'

     blo=bzang-gis nyos-pa-'i     stong=thung
     Lobsang-ERG   bought-NOM-GEN shirt
     'the shirt that Lobsang bought'


We see here one reason for a statistical correlation between the
position of genitive modifiers and relative clauses:  in a great
many languages (those I know best are from East Asia, but the
pattern exists elsewhere as well), relative clauses are genitive
modifiers--the relative clause construction consists of a
nominalized clause which then modifies a head noun.  Since it is
a noun phrase, it modifies its head in a genitive construction,
and so behaves like a genitive modifier because that is exactly
what it is.  Again, these typologically robust facts directly
contradict the predictions of un-kludged X' theory.
     Now, from the standpoint of X' theory demonstratives, and
perhaps numerals, are specifiers rather than modifiers, and thus
need not necessarily follow the same ordering with respect to
their heads as modifiers.  Genitive possessors should also be
Specifiers, but in Tibetan they do not pattern with determiners
or numerals.  If we want to treat determiners as heads of DP,
then DP, like other categories, is head-final, but the N-Adj
order remains anomalous--there is no apparent way around the fact
that N' is head-initial with respect to adjectives, but head-
final with respect to genitives and relative clauses.
     Accommodating facts such as these requires either ad hoc
analyses (e.g. in terms of a movement rule) to allow for an
underlying structure which is consistent with the theory (thus
entailing the claim that this aspect of Tibetan syntax is marked,
i.e. less natural than the corresponding aspect of French or Thai
grammar), or kludging of the theory to the point of vacuity.  But
there is another alternative--we can abandon the theoretical
account altogether.  I suggest that Tibetan NP is the way it is
because of its history, and that no theoretical account of these
data is necessary at all.
     There is no question that a distinct class of adjectives
exists in modern Tibetan, which cannot be synchronically analyzed
away as either nouns or verbs.  An adjective occurs in several
forms.  The bare adjective stem occurs only in the comparative
construction, where it is inflected as a verb:

     NP1 NP2-las yag=gi red
             ABL good-IMPF
     'NP1 is better than NP2.'

In predicate or modifying position, the adjective stem requires a
nominal suffix, generally -po:

     'di  deb  yag=po red
     this book good   be
     'This is a good book.'

     deb  'di  yag=po red
     book this good   be
     'This book is good.'

A handful of adjectives take nominal suffixes other than =po in
this construction, e.g. gsar(=pa) 'new'.  This pattern has been
regularizing over the attested history of Tibetan; earlier texts
show a larger range of nominal suffixes here, with the feminine
forms -ma and -mo attested as well.
     These facts allow an internal reconstruction of the source
and development of the Tibetan adjectival construction.  The
nominal suffixes, po/bo, pa/ba, mo, ma, have no other productive
use in modern Tibetan (except for -pa, which is the general
nominalizer), but they occur in a large number of nouns, e.g.
bu=mo 'girl', spo=bo 'grandfather', rgyal=po 'king', rgyal=mo
'queen'.  There is good evidence from earlier stages of the
language to suggest that these were once a productive
construction, being used widely to mark gender on nouns (Francke
1929:111) and to derive nominalizations meaning 'N which is/does
V, V one'.
     This suggests a history for the modern adjective
construction in which the -po form is a nominalization, of which
the suffix is then the head.  Thus we can make perfect sense of
the predicate adjective construction:

     deb  'di  yag=po red
     book this good   be
     'This book is good.'

which would have originated in a copular construction: 'this book
is a good one'.  The analysis of the modifying construction is
less obvious, however.  If the adjective were originally a
nominal modifying the head, then it would be marked as genitive
and precede its head.  Thus the construction exemplified by deb
yag=po must have a different origin.  At this point I have to
speculate somewhat, but there is an obvious candidate.  Tibetan,
especially older forms, is quite fond of appositive constructions
in which a second noun provides further specification or
quantification of a preceding NP.  I have gathered here a few
examples from a single literary text (Norbu and Ekvall 1969). 
(The translations are taken directly from Norbu and Ekvall's
English rendering):


bla=ma yab=sras   gnyis-ka-s smon=lam    yang  
lama   father-son two-ERG    wish-prayer CONTR 

dag-pa-r mdzad
sincere  make 
"The lama--father-son--the two made effectual wish-prayers." 

de-r rgyal=bu-s yab=rje     bla=ma-la zhus-pa
then prince-ERG father=lord lama-LOC  speak-NOM
"The prince then spoke to the lord-father--the lama, ..." 

rgyal=po bla=ma sras=mo  lha-gcig
king     lama   daughter god-one
"the king, the lama, and the daughter, the goddess one" 

yab=yum       gsum
father=mother three
"father-mother--the three" (referring to two brothers and their
wife in a polyandrous marriage)

Since the typical form of Tibetan adjectives is etymologically
nominal, and they occur in the same position as these
appositives, I would suggest that this construction is a
grammaticalization of an original appositive construction: 
'book, a good one'.  If the history is in fact something like
this, as I think it must be, then we have an explanation for the
"disharmonious" N-Adj order in Tibetan--adjectives do not precede
their heads because they originate in a construction in which
they were not dependent, and the N was not their head.
     Now, one standard line of argument at this point is
empirical:  here are these data, I can explain them with my
theory, other theories can't explain them, therefore my theory
wins.  But I want to focus not only on the empirical problems,
but also on deeper objections to the X' account.  The empirical
problems can always be solved; we can always invent new
parameters to set.  There is a more fundamental metatheoretical
problem with the claim that the parametric theory provides an
explanation for the facts; I would argue that history provides a
substantive explanation of a sort that formal theory is in
principle incapable of.  But my main point here is that a
diachronic explanation for data such as these is neceesary no
matter what, and that if the data are entirely explained by the
diachronic facts, then it is not only unnecessary but incorrect
to attempt to build an explanation into synchronic theory.  That
is, I suggest that the X' parameters are irrelevant and
unnecessary to the task of accounting for these data, that
instead the facts I am discussing can be entirely explained by
reference to certain characteristic structures of Tibetan consti-
tuting the initial state of the grammaticalization process which
has produced the modern Tibetan N' structure.

5.  Heads

The old notion of "head of a construction", developed by TesniŠre
and Bloomfield, is currently once again central in virtually
every formal approach to syntactic structure, whether
constituency or dependency-based.  There is, however, a fair
amount of disagreement across theories, and even within
particular theoretical camps, about the identification of heads
of particular constructions, as well as about the definition--and
the extent to which there needs to be a motivated definition--of
the basic concept.

5.1.  Heads of constituents

The concept of nouns and verbs as heads of their phrases is
intuitively clear, and can be given an explicit structural
characterization--in Bloomfieldian terms, these are the lexical
heads of endocentric constructions, i.e. those which have the
same privileges of distribution as their heads.  (Recent work in
X' theory has been concerned with arguing that all constructions
are endocentric.  There are grounds for seeing a degree of
circularity in this argument, but to sidestep it for the moment
let us use "endocentric" to mean transparently endocentric in
Bloomfield's sense).  There is no particular difficulty, either
conceptual or formal, in extending the notion to other
endocentric constructions, e.g. adjective phrases or adverbial
phrases (really ugly, truly amazingly fast).  But in recent
theorizing, with its desire for cross-categorial parallelism of
structure, the notion of head has become central and must be
extended to exocentric constructions as well.
     Thus any constituent must have a head, and our theory must
tell us what it is.  This is more problematic in some cases than
in others.  Adpositional phrases, for example, are fairly
straightforward--although the NP, because of its lexical content,
seems to be on center stage, it is clear that the syntactic
behavior of the AdpP (its privileges of occurrence,
subcategorization etc.) is determined by the adposition.  Thus,
though the naive intuition that looks for heads in lexical
material might briefly suggest a different answer, it is in fact
clear that, assuming that we need to find a head in an AdpP, it
is the adposition.
     Still, even if we can satisfy ourselves that the adposition
is the head of its phrase, and ignore any nagging doubts about
the extent of the parallelism between this construction and N and
V phrases, the theory does not offer us any direct explanation
for why this parallelism should obtain.  The modern concern with
cross-categorial parallelism is motivated by a desire for formal
economy and simplicity which has its roots in obsession with the
"logical problem of language acquisition".  But this has no
explanatory potential; it is interesting, and very satisfying
from a formal point of view, to find evidence for parallelism of
structure, but we would hope that any substantive theory would
offer some reason why things are this way.
     The answer here should be obvious by now, since it is simply
a slight extension of our explanation for cross-categorial
parallelism of constituent order.  Since I am trying to argue as
atheoretically as possible, I do not presume to explain why ad-
positions "are" heads of their phrases, which is a theory-bound
proposition, but in terms of grammaticalization it is easy to
explain why adpositions are to some extent like lexical nouns and
verbs in terms of their relation to their phrases--it is because
AdpP's originate in NP's or VP's, in which the soon-to-be
adposition is the lexical head.  There is, then--and this is the
essential point which I want to make here--no need to stipulate
this structural parallelism in linguistic theory, no need to
suppose that some version of X'-bar theory with a predetermined
set of categories that are heads of their phrases is wired into
the dominant cerebral hemisphere of human beings, thus
guaranteeing that AdpP's will have a head-dependent structure
like NP's and VP's.

5.2.  Heads of words

An influential development in the study of morphology and word-
formation has been the attempt to extend notions of phrase-
structure from syntax into the study of word structure:

     ... the most attractive theory for dealing with phrasal
     compound data and the like would be one in which all
     principles of grammar apply both above and below word
     level.  (Lieber 1992:33)

Given the suspicion which I have attempted to cast on some work
in syntactic phrase-structure, you will not be surprised if I
express some doubt about aspects of this program.  Again, the
thrust of my argument is that many of the phenomena which some
linguists are busily trying to build into the framework of
linguistic theory fall out naturally from the nature of
diachronic processes, and need not--and thus should not--be
stipulated in a general theory of linguistic competence.
     To take a simple example (not directly a question of
grammaticalization, but it will do), Lieber and others have noted
a parallelism in head-modifier ordering within NP's and compound
nouns.  Why should that be?  Obviously, because compounds are
lexicalized NP's.  White house and White House have the same
head-dependent structure because the latter is simply the former
with a new stress pattern.
     For a more complicated example, let us return to Tibetan. 
Compounds in Tibetan are not consistent in terms of their
ordering.  Compounds of two nouns are modifier-head:

     lag=shubs  "hand-covering" 'glove'

but compounds of noun and adjective are head-modifer:

     glang=chen  "ox-large" 'elephant'

From the previous discussion of order within N' in Tibetan, we
can see that, however inconvenient the inconsistent behavior of
compounds may be from a theoretical point of view, the facts fall
out directly in a diachronic perspective:  N-N compounds are
derived from NP's in which a genitive modifier precedes the head,
N-Adj compounds from NP's in which an adjectival modifier follows
the head, and these orderings are, not surprisingly, preserved
unchanged in the lexicalized forms.
     The same argument can be extended to derivational
morphology.  The 80's brought us the discovery that derivational
affixes can usefully be considered heads of their words.  So a
good deal of contemporary theoretical work in morphology is
devoted to trying to ensure that linguistic theory will assign
similar structures below as above the word level (e.g. Lieber
1992).
     Again, if we are interested in substantive explanation, many
of these arguments seem suspect.  But, again, my main point here
is to suggest that in any case this effort is misguided.  As
Giv¢n (1971) pointed out quite some time ago, derivational mor-
phology is typically the remnant of older syntactic
constructions.  A nominalizer, for example, typically reflects an
older abstract noun, which is the head of its phrase.  If we have
reason to believe that it remains the head in its morphologized
stage of existence, this still does not mean that we must con-
struct our overall theory of synchronic structure so as to ensure
that in the unmarked case it will have the same ordering relation
to its dependent as does the head of N'.  It will have the same
ordering relation because it used to be a N', and has not changed
its order in the course of morphologization, not because some
abstract principle of linguistic structure demands parallelism of
ordering above and below the level of the word.


References

Aristar, Anthony. 1991. On diachronic sources and synchronic
     pattern:  An investigation into the origin of linguistic
     universals.  Lg. 67:1-33.

Bolinger, Dwight. 1980. Wanna and the gradience of auxiliaries. 
     pp. 292-299 in G. Brettschneider and C. Lehmann, eds., Wege
     zur universalien Forschung.  T�bingen:  Gunter Narr.

Dixon, R.M.W. 1977. Where have all the adjectives gone?  Studies
     in Language 1:19-80. (repr. in Dixon, Where have all the
     adjectives gone? and other essays in semantics and syntax. 
     NY: Mouton).

Dryer, Matthew.  1988.  Object-verb order and adjective-noun
     order:  Dispelling a myth.  Lingua 74:185-217.

Francke, A.H. 1929.  Addenda.  pp. 105-161 in H.A. J„schke,
     Tibetan grammar.  Berlin:  de Gruyter.

Givon, T. 1971.  Historical syntax and synchronic morphology:  An
     archeologist's field trip.  CLS 7:394-415.

     . 1979. On understanding grammar.  New York:  Academic
     Press.

Greenberg, Joseph.  1966. Some universals of grammar with
     particular reference to the order of meaningful elements. 
     pp. 73-113 in J. Greenberg, ed., Universals of language. 
     Cambridge, MA: M.I.T. Press.

     . 1969. Some methods of dynamic comparison in linguistics. 
     pp. 147-203 in J. Puhvel, ed., Substance and structure of
     language.  Berkeley and Los Angeles:  University of
     California Press.

Hawkins, Jack. 1984.  Modifier-head or function-argument
     relations in phrase structure?  The evidence of some word
     order universals.  Lingua 74:219-59.

Heine, Bernd, Ulrike Claudi, and Friederike H�nnemeyer. 1991.
     Grammaticalization:  A conceptual framework.  Chicago: 
     Unviersity of Chicago Press.

Hopper, Paul, and Elizabeth Traugott.  1993.  Grammaticalization. 
     Cambridge:  Cambridge University Press.

Jackendoff, Ray. 1977. X' Syntax:  A study of phrase structure. 
     Cambridge:  MIT Press.

Jespersen, Otto.  1924.  The philosophy of grammar.  New York: 
     Henry Holt.

Kornai, Andr s, and Geoffrey Pullum. 1990. The X-bar theory of
     phrase structure.  Lg. 66:24-50.

Kurylowicz, Jerzy.  1965.  The evolution of grammatical
     categories.  Diogenes 51:55-71.

Lieber, Rochelle.  1992. Deconstructing morphology:  Word
     formation in syntactic theory.  Chicago:  University of
     Chicago Press.

Lightfoot, David. 1991. How to set parameters.  Cambridge, MA:
     MIT Press.

Norbu, Thubten, and Robert Ekvall. 1969.  The younger brother Don
     Yod.  Bloomington:  Indiana University Press.

Schmidt, Wilhelm. 1926. Die Sprachfamilien und Sprachenkreise der
     Erde.  Heidelberg: Carl Winter.

Stowell, Timothy. 1981. Origins of phrase structure.  Ph.D.
     dissertation, MIT.

Terrien de Lacouperie. 1887.  The languages of China before the
     Chinese.  London:  David Nutt.  (Repr. Taipei: Ch'eng Wen,
     1970).

Tesniere, Lucien.  1959.  �lements de syntaxe structurale. 
     Paris:  Klincksieck.

Traugott, Elizabeth, and Bernd Heine. 1991.  Approaches to
     grammaticalization.  (Typological Studies in Language 19).
     Amsterdam:  Benjamins.

Travis, Lisa.  1989. Parameters of phrase structure.  in M.
     Baltin and A. Kroch, eds., Alternative conceptions of phrase
     structure, pp. 263-79.  Chicago: University of Chicago
     Press.