the fall of COMP? (long posting)

[Joan Bresnan]

In his recent posting to this list, Chris Pinon paints an amusing
picture of adherents of a rival theory smiling "discreetly to
themselves" over the recent discussion of the COMP(lement) function and
thinking it is only a matter of time before lfg finally
adopts subcategorization for syntactic categories, which (he implies)
would solve all these problems.  He concludes with some rueful
thoughts about the nature of progress in the field of syntax.

Chris' faith in the explanatory power of syntactic categories like
"NP, PP, S' (S)" seems a bit naive when we consider the history of
generative syntactic analyses of English complementation and some of
the typological problems of categorization.

In the early days of transformational grammar (circa Rosenbaum 1967),
syntacticians exuberantly multiplied subcategorization frames and
syntactic categories for different complement types.  For example,
Rosenbaum had subcategorization for S, for NP over S, for PP over NP
over S, and for each of these cross-classified by complementizer type
(that, for-to, and poss-ing), and for transitive/intransitivity.  The
work was descriptively careful and exacting, but was felt to be
unexplanatory: why, for example, did the verb subcategorization
classes show semantic homogeneity in some cases and a complete lack of
it in others?  [Ideas of this period may be compared to the early use of
grammatical functions in LFG (e.g. OBJ, OBL, SCOMP) to describe
different syntactic classes at a level of f-structure.]

Ross and other proponents of the Abstract Syntax/Generative Semantics
movement proposed to reduce the inventory of distinct categories for
different sentential complement types to one: NP over S, and to
explain variations in the distribution of different complement types
in terms of independent constraints.  For example, to explain why S's
do not have full nominal distribution in English, he proposed a
constraint, perhaps having a perceptual basis, against NP immediately
dominating S in clause-internal positions; this constraint would force
S arguments to the periphery of clauses (e.g. in subject and
extraposed positions) while capturing semantic uniformities between
sentential and other arguments, allowing for the passivizability of
sentential complements, etc.  This approach was certainly more
insightful in its aims, at least, but was unsatisfying in its
execution, which often involved rule features or triggers to adjust
the simpler, semantically motivated base to the messy complexities of
surface distribution.  [Ideas of this period may be compared with the
proposals by Mohanan, Mohanan, and Alsina in lfg to simplify the set
of complement functions--e.g. taking COMP really to be just OBJ--and
appealing vaguely to features of other levels to account for
nonuniformities between sentential and nonsentential OBJs, leaving the
complex details that have been developed in a long literature without
a satisfactory explanation.]

Emonds' work (e.g. 1975) stood Ross' analysis on its head.  Emonds
gave evidence and careful argumentation showing that sentential (that)
complements in English do *not* have NP distribution, pace Ross.  They
are based generated at the periphery (nonnominal complements to
certain adjectives, nouns, verbs and prepositions), but they never
occupy NP positions by themselves; hence they cannot be OBJs.  Why
then do they appear as subjects of passives?  Showing that this
phenomenon is severely limited (to matrix-like contexts) he proposed
that this is the result of a non-structure-preserving rule
("intraposition").  [In lfg terms, these stylistically marked
deformations of root trees (or matrix-like embedded clauses) might be
thought of as related to nonargument discourse functions (TOP, FOC,
etc.) rather than argument functions (SUBJ, OBJ), etc.  Hence, Emonds'
analysis implies that sentential that-complements (at least in English and
some Germanic languages) are complements that are *not* in NP
positions and do *not* have NP functions such as SUBJ, OBJ.  This
approach is paralleled in lfg by Bresnan 1994, 1995, Berman 1996 in her
contribution to the Grenoble colloquium--which triggered the posting
by Mohanan, Mohanan, and Alsina and this whole discussion of COMP, and
by Helge Lodrup in the works he cites on Norwegian.]

When rules went out of fashion in favor of Rule ("move alpha"),
Emonds' view of sentential complements was reworked by Stowell (1983?)
and others.  Stowell tried to derive the distributional properties of
sentential complements from his Case Resistance Principle: categories
whose heads are case assigners resist case assignment, while
categories whose heads are not case assigners allow it.  Thus
sentential complements, which are headed by a case assigner, according
to Stowell, *are* generated in NP positions, where they receive a
theta role, but must vacate these positions by the surface to avoid
case assignment.  Phrases headed by nouns, which do not assign case,
can happily receive case; and so NPs appear in case positions.
Stowell claimed that this theory eliminated any reference to
underlying categories at all; everything could be predicted from theta
theory and case theory.  However, crosslinguistic evidence against
Stowell's theory of CP distribution pretty convincingly argues against
it, in my view (e.g. Bresnan (1991, 1994, 1995), Hoekstra 1984, Plann 1986,
and Luj\'an 1994. Bresnan also 1994 points out unexplanatory aspects
of Stowell's theory of English sentential complements: e.g., he simply
replaced certain categorial differences at D-structure by otherwise
unmotivated features requiring obligatory reconstruction at LF.)

With the advent of extended X' theory and functional projections, the
theory of categories changed again, and Grimshaw 1991 made a valiant
attempt to simplify and rationalize the theory of FPs, including our
friends the sentential complements, now called CP.  She hypothesized
that they are really just verbal extended projections.  Her elegant
and appealing approach, unfortunately, also has serious empirical
problems with CP (and PP).  As I showed in Bresnan 1995, Chichewa has
CPs headed by the complementizer kuti 'that' which show the typical
transparency to selection for subjunctive mood on the verb that
Grimshaw takes as evidence for verbal status for CP.  Nevertheless,
they have nominal (NP) distributional properties (such as appearing in
the position of prepositional objects), which I trace to the status of
'kuti' as an infinitival form of the verb 'say'.  The
grammaticalization of verbs of saying into sentential complementizers
is common and widespread.  The nominal properties of kuti come from
the fact that it is a class 15 infinitive/gerund form, and these are
nominal in Bantu, belong to the noun class system and inducing noun
class concord (Myers 1987, Bresnan and Mchombo 1995).

Thus we see that in the history of that part of generative syntax
which has always embraced subcategorization for category (in one form
or another), there has been a similar oscillation between theories
which multiply category distinctions for sentential complements
(Rosenbaum, Emonds) and theories which reduce category distinctions
(Ross, Stowell, Grimshaw).  There is very little agreement about what the
category of sentential complements actually is, and why it is that.

The fact that categorization issues have come into the discussion of
the function COMP is not surprising.  After all, functions in lfg
*are* the links between c-structure and a-structure, and of course
there are constraints on the correspondence mappings (such as the
proposal of Grimshaw/Pinker/Bresnan that only nominal categories
have direct functions (SUBJ/OBJ)).


Joan

[sorry that refs are omitted; this msg is too long already!]

...

>"COMP is a brainchild of LFG.  No other mainstream syntactic theory (GB,
>GPSG, HPSG, Construction Grammar, Categorial Grammar, RG, Arc-Pair Grammar,
>Word Grammar) has a special *grammatical function* for S' (S) syntactic
>categories.  That LFG syntacticians are now considering to get rid of COMP
>is surely a welcome move, but given that other syntacticians never believed
>in such a grammatical function to begin with, the move will have absolutely
.no practical effect on the work of other syntacticians.  The realization
>that COMP is unnecessary is not an empirical discovery---it is a purely
>LFG-internal development.  So what should other syntacticians get excited
>about?"
>
...

>"At the bottom of all this is the question of whether or not syntactic
>category selection (i.e., for NP, PP, S' (S)) is necessary.  In most
>(though not all) of the above-mentioned theories it has long since been
>claimed that it is, at the very least for those cases where syntactic
>selection isn't semantically predictable.  LFG has long since been
>officially against introducing syntactic subcategorization frames, although
>it has certainly acknowledged a vital role for c-structure.  The suggestion
>that syntactic category selection may be necessary has appeared a number of
>times in the debate on COMP, but typically in the "margins", with a mild
>discomfort.  But what was it that made subcategorizing for COMP (as opposed
>to S' (S)) so acceptable for so long?  Surely that was a *bias* of LFG, and
>not something that was demanded by an analysis of the data.  Or are we
>mistaken?"
...

>"And when LFG officially does decide to countenance syntactic
>subcategorization frames, and we sense that it's only a question of time,
>the move again will be welcome, but once again, what will other
>syntacticians have to get excited about?  Of course, we realize that these
>thoughts are a bit unfair, picking on LFG to the exclusion of the other
>theories.  But at the back of our minds we wonder about the nature of
>progress in the field of syntax, think about the fall of COMP, and ask
>whether that will be progress."







-----------------------------------------        ______     __o       __o
Joan Bresnan	bresnan at csli.stanford.edu      ______     _`\<,_    _`\<,_
-----------------------------------------       ______   (*)/ (*)  (*)/ (*)