Th e o r e t ic a l Lin g u is t ic s Pr o g r a m m e, Bu d a p e s t Un iv e r s it y (E L T E )
P H R A S E S T R U C T U R E A N D D E P E N D E N C E
Mic h a e l Br o d y
Re s e a r c h In s t it u t e fo r Lin g u is t ic s, Hu n g a r ia n Ac a d e m y of Sc ie n c e s Wo r k in g Taters in t h e Theo ry of Gr a m m a r, Vo l. 1, No. 1
Re c e iv e d: Se p t e m b e r 1994
P H R A S E S T R U C T U R E A N D D E P E N D E N C E
Mic h a e l Br o d y
Th e o r e t ic a l Lin g u is t ic s Pr o g r a m m e, Bu d a p e s t Un iv e r s it y (ELTE) Resear c h In s t it u t e for Lin g u is t ic s, HAS, Ro o m 119
Bu d a p e s t L , P.O. Box 19. H-1250 Hu n g a r y
E -m a i l: b ro d y6 n ytu d .h u
Th e o r e t ic a l Lin g u is t ic s Pr o g r a m m e, Bu d a p e s t Un iv e r s it y (E L T E ) Re s e a r c h In s t it u t e f o r Lin g u is t ic s, Hu n g a r ia n Ac a d e m y of Sc ie n c e s
Bu d a p e s t I., P.O. B ox 19. H-1250 Hu n g a r y
Te l e p h o n e: (36-1) 175 8285; Fa x: (36-1) 212 2050
S yn ta ctic structures have been analysed in term s o f c o n stitu e n t hierarchies and also in term s o f dependency relations. While these tw o traditions have som etim es been presented as com peting w ith each other, there is no reason in principle w h y s y n ta c tic dependencies and co n stitu e n t hierarchies should n o t both be p a rt o f the gram m ar, com plem enting each other. Indeed, w ith in transform ational gram m ar and more spe cifically w ith in the Principles and Parameters theory, various dependencies are postulated alongside the basic constituent hierarchy analysis o f sentences. The m ost salient o f these are the various types o f anaphoric dependencies, b u t o th e r relations, like for exam ple th e m a tic, governm ent or chain relations, have also fro m tim e to tim e been th o u g h t o f as constituting dependencies. W ith in this fra m e w o rk it remains true how ever, th a t the basic structural organizing principles are couched in term s o f co n stitu e n t hierarchies, dependency relations are no t taken to play a role here. In w h a t fo llo w s I shall propose that dependency relations are more ce n tra l to the gram m ar. I shall postulate a new module: dependency theory. The basic s y n ta c tic analysis o f a sentence w ill not sim ply consist o f a set o f elem ents arranged in a c o n s titu e n t hierarchy. It w ill include also a dependency s tru c tu re defined on these elem ents: all syntactic elements participate in both a hierarchical and a dependency stru ctu re .
In section I o f th is paper I shall sho w th a t dependency theory m akes it possible to radically s im p lify the theory of phrase stru cture . In section II, I shall argue th a t the evidence th a t has standardly been taken to m o tivate binary rig h tw a rd branching analyses is be tter treated in term s o f dependency relations b e tw e e n co n stitu e n ts.
I. Phrase S tructure and Dependency 1.1. S tructura l Dependencies
Let me sta rt w ith sketch ing the dependency structures I have in m ind. I shall propose th a t the concepts o f specifier-head and head-com plem ent relations on one hand and dependency relations on the other should be b ro u g h t tog eth er: a t least some specifier-head and head-com plem ent relations are in fa c t dependencies.
Suppose th a t phrasal categories can be in a structural dependency re la tion w ith some head X. It w ill fo llo w from the principles o f the th e o ry (to be presented im m ediately below) th a t there is a single category on w h ic h the head X d ire c tly depends, the spec o f X, and a single category w hich d ire ctly depends on th e head, the comp o f X. (Notice th a t spec and comp are not defined hierarchically here, but in term s o f dependencies.) I shall call the spec-head and the head-com p dependencies direct structural dependencies. Direct stru ctu ra l dependencies are then relations betw een a head and a phrase.
Spec-head dependencies can be taken to be instantiated by ch e ckin g
relations, -ta k in g the head or its checking features to depend on the spec or on the relevant features o f th e spec. A typ ica l comp-head dependency is object theta assignm ent/selection.
Turning now to th e principles o f dependency th e o ry I assume firs t th a t direct s tru c tu ra l dependencies interact tra n sitive ly creating indirect structural dependencies. I assum e also an adapted version o f Kayne's (1 993 ) Linear Correspondence A xio m . I shall take the principles o f Precedence and T o ta lity to regulate the relationship between the linear ordering o f term inals and the dependency relations holding between nodes dom inating them . I give a som ew hat inform al statem ent in (1) and (2):
(1) Precedence
if x structurally depends on y then the term inals dom inated by y precede the terminals dom inated by x
(2) Totality
all terminals m ust be ordered by the Precedence condition on stru ctura l dependencies
W illiam s (1992) has independently argued fo r a precedence condition on anaphoric dependencies. Precedence in (1) can also be thought o f as an application of his c o n d itio n to structural dependencies.
In addition I assum e th a t no redundant direct stru ctu ra l (spec-head or head- com p) dependencies are allow ed: there can be no direct dependencies beyond those th a t fu lly determine th e precedence relations between term inals.
(3) Non-redundancy
there are no d ire c t structural dependencies th a t are not required to satisfy T o ta lity
The principles o f dependency th e o ry entail tha t d ire ct structural dependencies alw ays link adjacent catego ries, they ca n n o t skip over them . This is not d iffic u lt to see. C onsider the case w h e re there is some element, oc, intervening betw een the tw o m em bers of a d ire c t dependency, as in (4):
(4) YP oc X b.
Such a configuration v io la te s either T o ta lity or Non-redundancy. T o ta lity requires th a t the re be some dependencies th a t force oc to precede the term inals dominated by X and fo llo w the te rm in a ls dominated by YP in (4). Suppose th a t such structural dependencies are present as in (4b). Then X indirectly depends on YP by virtue of these and the fact th a t the terminal dominated by X fo llo w s the term inals dom inated by YP w ill be ensured. Thus the structural dependency o f X on YP is unnecessary to order th e term inals dom inated by them and therefore it w ill violate N on-redundacy.
So spec-head and head-com p dependencies presuppose adjacency between
the linked elem ents. This has tw o immediate consequences. First a head can have only a single spec and a single com p and second the spec and the com p m ust be on d iffe re n t sides o f the head. Both consequences fo llo w since only one ca te go ry can be adjacent to a head on each side.
There is a diffe ren ce standardly made betw een specifier-head and head- com plem ent relations th a t I shall assume and exploit. H ead-com plem ent relations are lexically determ ined, -th e s e are taken to alw ays hold betw een le xica lly related sister co n s titu e n ts . I take head-comp relations to be a special case o f head- com plem ent relations covered by the same generalization. Spec-head relations/dependencies on the other hand are much freer. They can hold betw een lexically, and (as I shall assume follo w in g essentially Pesetsky 1992) even sem antically, unrelated categories. I shall therefore take spec-head dependencies to be in principle fre e ly assignable, -a lth o u g h subject to certain re s tric tio n s to be discussed below . The axioms o f dependency the ory w ill ensure th a t th is does not result in overgeneration.
N otice n o w th a t if spec and comp were understood in both hierarchical and dependency term s then dependency theory w ou ld force hierarchical s tru ctu re s to be s tric tly binary branching: there can be only one spec and one com p associated w ith any head. But since I understand spec and com p of a head X as categories in a direct stru ctu ra l dependency relation w ith X, these considerations do no t exclude stru cture s in w h ich a head has more than one specifiers or com plem ents. As long as the additional m axim al projections do not intervene betw een the spec or com p and the head, there can be dependency structures tha t do not violate the axiom s.
For exam ple in (5a) the head X has its spec and comp dependency (to QP and from RP respectively) and in addition it has another com plem ent:
As just noted, spec-head dependencies are assigned freely, hence unless ruled o u t by the dependency axioms there w ill be spec-head dependencies th a t do not correspond to the configurational spec-head relations o f standard stru ctu ra l analyses. Such a spec-head dependency from the head o f YP to RP, the com p o f X can provide the required ordering for the term inals dom inated by YP in (5a).
W e can define the com plem ent o f X as any sister ca te g o ry th a t ( + im properly) contains some element (directly or indirectly) s tru ctu ra lly dependent on X. In a parallel w a y any sister category o f X ' tha t ( + /- properly) co n tains some elem ent on w h ich X depends w ill be a specifier o f X. The fo llo w in g fo u r points need to be m entioned in connection w ith these definitions. First I shall argue be lo w for the elim ination o f the X '-level, and accordingly I should su b stitu te X fo r X ' in the d e finition o f the specifier. Secondly note th a t a spec is not necessarily a specifier in the present th e o ry : specifiers but not spec's are defined as sister nodes. Thus for example RP in (5a) is the spec o f Y, since Y d ire ctly stru ctu ra lly depends on it, but RP is n o t the sp e cifie r o f Y: it is outside YP. Thirdly, I assume th a t in a stru ctu re like (5a) if RP is a PP, the head P of this PP can always reanalyse w ith the head X form ing a com plex head w hose comp is the com p o f P. Thus if the head o f the
(5)a. YP
siste r phrase (corresponding to YP in (5a)) fo llo w in g this PP stru ctu ra lly depends on the com p o f P, this second sister w ill still s a tisfy the definition o f com plem ent o f X: it in d ire ctly depends on the reanalysed X + P. Fourthly note also th a t the system a llo w s not only for m u ltip le com plem ents bu t also for m ultiple specifiers. For exam ple in a structure like (5b), where YP is the spec o f X and both QP and YP are sp e cifiers o f X, the axiom s of dependency th e o ry can be satisfied. Sim ilarly to the m u ltip le com plem ent case o f (5a), a spec-head dependency between QP and Y can ord e r the term inals in th e tw o specifiers.
(5)b. Q P ^ X T ^ V ^ Z P ] ^ *X
Consider next th e choice between spec-head-com p and com p-head-spec w o rd orders. The assum ption th a t the dependent element must fo llo w the one it depends on is a natural one given general processing considerations. Thus if spec depends on the head and the head on comp then Precedence entails a com p-head- spec w o rd order, w h ile if the comp depends on the head and the head on the spec then Precedence forces the spec-head-com p order. Spec-head relations give no s tra ig h tfo rw a rd clue as to w hich is the co rre ct choice, but com p clearly is a fu n c tio n o f the head ra ther than the other w ay around. Hence com p depends on, and th e re fo re by Precedence fo llo w s, the head. Since, as we have seen, the spec m ust be on the opposite side o f the head the w o rd order must be spec-head-com p.
It th e n fo llo w s also by Precedence th a t the head must depend on the spec. If com plem e nts and specifiers are also defined in dependency term s as just suggested, then all specifiers w ill precede and com plem ents fo llo w the head, -a g a in as a consequence o f Precedence.
1.2. A M inim al Theory o f Phrase Structure
Let us next turn to som e o f those predictions o f Kayne's LCA for phrase structure th a t th e proposed dependency theory does not reproduce. Kayne's LCA is violated by X P 's th a t have no X° head as in (6a) and ones th a t have more than one such head as fo r example in (6b). ((6b) can also be th o u g h t o f as a stru ctu re w here the co m p le m e n t o f a head is not a phrase but another head.)
(6) XP
ZP
z
/X '
l IX
X 1 YP X 2
D ependency theory on th e other hand has no such consequence. The dependency
stru ctu re in (6b), where YP to X 1 is a head-comp and X 2 to YP is a spec-head dependency, does not violate the principles o f dependency th e o ry. As fo r (6a), if ZP has no specifiers then a spec-head dependency from Z to YP w ill o rd e r the term inals appropriately. Nothing forces the sim ultaneous presence o f a dependency betw een Y and ZP th a t w ould result in a violation. It seems to me h o w ever th a t the fa c t th a t dependency theory does not entail th a t a phrase m ust im m ediately dom inate a unique head is in fa c t an advantage, since it elim inates a potential redundancy.
Since phrases and their heads share properties, the assum ption th a t phrases are in som e sense projected by th e ir heads seems inelim inable. The fa c t th a t a phrase m ust have a unique head w ill fo llo w if we assume th a t phrases can only arise th ro u g h projection and th a t furtherm ore all heads m ust project phrases. Since the latter statem e nt is false for w ord-internal heads, a m inim al sta te m e n t o f this w ould have to be along the lines o f (7), w hich I shall call the Principle o f Phrasal Projection.
(7) Principle of Phrasal Projection (PPP)
a. Every non w ord-internal head must project a d is tin c t phrase b. Every phrase is projected by a (non w ord-internal) head
The PPP expresses the idea th a t syn ta ctic categorial stru ctu re is projected fro m the lexicon. This involves tw o assum ptions. Clause (a) o f the PPP states th a t a precondition fo r a lexical element to enter the syn ta ctic s tru c tu re is for it to project some nonlexical category, ie. a phrase. Clause (b) states th a t all s y n ta c tic categories are related to the lexicon: they must either com e fro m the le xico n or be projected by categories w hich do. That a phrase m ust have a head fo llo w s now from clause (b) o f the PPP, th a t is from the fa c t th a t all phrases are projecte d by their heads. That a phrase m ust not have more than one head w ill fo llo w from clause (a).
A basic property o f projection is tha t it is local: a phrasal node projecte d by a head is alw ays near to this head. Assume th a t the relevant notion o f lo ca lity is im m ediate dom ination: the phrase projected by a head m ust im m ediately dom inate this head. Then a phrase HP w ill immediately dom inate a unique head H, —its own.
A ny o th e r head m ust be im m ediately dominated by the phrase th a t this o th e r head projected, necessarily d istin ct fro m HP as we have seen.
C onsider next the converse statem ent: th a t a head projects a unique phrase.
O ptim ally th is should also hold, since it would radically sim p lify the th e o ry o f phrase stru ctu re . Suppose th a t it does. Let us first ask w h y this should be so. W e have already seen th a t a head m ust project a phrase, so we only need an a c c o u n t o f w hy it can not project more than one. Uniqueness o f the projected phrase will im m ediately fo llo w if the locality relation between the head and the projected phrase is understood s trictly, say as immediate dom ination defined as in (8) (where x,y,z are categories and dom ination is irreflexive): 8
(8) x im m ediately dom inates y iff x dominates y and fo r all z d is tin c t fro m x
if z dominates y then z dom inates x
Locality as in (8) rules out m ultiple projection by one head not only in the standard con figuration o f (9a) but also in the "inversely branch ing" (9b).
(9) a. XP XP XP
It th u s fo llo w s also that a head can project only a single phrase, since only a single phrase can im m ediately dom inate the head that projected it.
I thus assume th a t in the co n te xt o f the theory th a t assumes presyntactic projection o f phrases, the biunique relation between phrases and non w ord internal heads projecting them fo llo w s from the PPP and the loca lity requirem ent on projectio n. O ptim ally, the PPP and the locality requirem ent should exhaust the th e o ry of phrase stru cture . Suppose th a t they do. Then the the ory o f phrase s tru ctu re requires that e ve ry XP consists of a unique head X and an arbitrary num ber o f o th e r phrases. A s we have seen, dependency th e o ry ensures th a t only one o f the included phrases can serve as a spec and only one as comp, the designated fir s t com plem ent. The spec and the specifiers w ill precede and the comp and the o th e r com plem ents w ill fo llo w the head for the reasons given.
C urrent theories o f phrase stru ctu re diverge from this simple picture w hich o n ly contains the co n fig u ra tio n where an XP dominates a head and a number of o th e r XP's in tw o major b u t related respects. An interm ediate X ' level is assumed betw een th e head and th e phrasal node and the co n fig u ra tio n o f adjunction is a llow e d in addition. These tw o additions can be reduced to one if, as proposed by Kayne (1 9 9 3 ), the interm ediate X '-le ve l is treated as the low er segment o f ad ju n ctio n . But the ad ho c segm ent-category d istin ctio n and the attendant co m p lica tio n s in the d e fin itio n of c-com m and necessary in this the ory make a stip u la tive configuration o u t o f adjunction. Given such com plications, adjunction is clearly no t a notion "d ra w n from the dom ain of (virtual) conceptual necessity" to use C ho m sky's (1993) te rm s.
W ithin the present fra m e w o rk w ord-external adjunction is both impossible and unnecessary. The fa c t th a t adjunction is not a possible option w ord-externally, fo llo w s im m ediately fro m the PPP. Since no segm ent-category distinction is postulated, adjunction w o u ld by definition create a new category w ith o u t a head projecting it, violating the PPP. There are at least tw o alternative structures fo r phrases standardly treated as adjoined. First since m ultiple specifiers are allowed, an adjoined category can be analysed as an additional specifier. This would be a na tura l approach for exam ple to multiple w h-elem ents in phrases headed by a + WH ca te g o ry. (N otice that th e "head o f COMP" (Lasnik and Saito 1984) ie. the su b stitu te d w h-phrase in th e specifer o f the + WH C under standard treatm ents can
still be distinguished from the other "adjoined" w h-phrases: th is elem ent is the spec o f the + W H head u n d e rth e dependency module, w hile the o th e r w h -ca te g o rie s are only specifiers o f this head but they do not serve as its spec.) A lte rn a tiv e ly an adjoined category can also be taken to be the specifier o f some higher head. Under this op tion, instead o f left-adju nction of XP to YP as in (10a), we can have the co n fig u ra tio n in (10b) w ith the higher head Z w hich is either invisible fo r selection (selectional requirem ents are satisfied by the low er head Y) or it has the a b ility to sa tisfy the same selectional requirem ents as the low er head Y. (A pp are ntly fronted w h-phrases in some m ultiple w h -fro n tin g languages in sta ntiate also this p o ssibility,
—cf. Rudin 1988).
(10) a.
b.
YP l Y
Thus w e can dispense w ith phrasal adjunction in general, - a w elcom e re su lt for the reasons noted. See also Sportiche (1994) w h o reaches the conclusion independently, th a t adjunction does not exist in (presum ably w o rd -e xte rn a l) syntax.
C hom sky 1 9 94 develops a theory th a t radically restricts w o rd -e xte rn a l ad junction to cases w here the ta rg e t has no theta role (expletive-associate chains) or w here in his derivational system the adjunct is not present at LF (interm ediate traces deleted by LF and "sem antically va cu ous" scram bling w h e re LF re co nstruction elim inates the scrambled element). These cases do n o t seem to provide strong m o tivation for retaining this co n fig u ra tio n . LF a d ju n ctio n o f the associate to its expletive chain-m ate is a problem atic and probably unnecessary operation cf. eg. Brody 1993. The necessity o f adjoined interm ediate tra ce s in non- uniform chains is equally m oot (cf. eg. Manzini 19 92). As fo r scram bling, C hom sky suggests th a t LF reconstruction w ill provide an account o f th e con trast he finds betw een (1 1a) and (1 1b). Here the expectation is th a t (1 1a), th e adjunct case is w orse, since forced reconstruction in this exam ple w ill create a co n fig u ra tio n th a t violates principle C. Since the fronted phrase "w h ic h pictures of Jo h n 's b ro th e r" is not an adjunct in (11b), this exam ple w ill not be sim ilarly excluded:
(11) a. Pictures o f Joh nx's brother, hex never expected th a t I w ou ld buy b. W hich pictures o f Joh nx's brother did hex exp ect th a t I w o u ld buy c. Near Johnx, hex saw a snake
Note firs t th a t if topicalization is treated as ad junction then in th e m inim alist
fra m e w o rk it is quite unclear how the PF presence o f the topicalized element in the fro n te d position gets trig g e re d . (For a su b stitu tio n analysis o f this co n structio n cf.
W atanabe 1 992.) S econdly, it is not clear if there really is a co n tra st between (11a) and (11b), especially in v ie w of well kn o w n examples like (11c), tha t appear to a llo w coreference. T h ird ly forced reconstruction o f adjuncts w ould loose the explana tion o f the co n tra sts like those betw een (1 2a) and (12b) or (12c) and (12d), w h ic h depend precisely on forced reconstruction o f the selected argum ent but not o f the unselected ca te g o ry internal to the fronted phrase (cf. Lebaux 1989, and C hom sky 19 93, Brody to appear, for d iffe re n t w ays o f instantiating this idea). If to p ica liza tio n is ad junction and adjuncts are reconstructed then (12a,b) should c o n tra s t w ith (12c,d) ra th e r than (1 2a) and (12c) w ith (12b) and (1 2d):
(12) a. W hich claim th a t Johnx made did hex deny
b . ?*W hose claim th a t Johnx w as asleep did hex deny c. The claim th a t J o h n x made about this, hex later denied d . ?*The claim th a t J o h n x was asleep all day hex later denied
Thus the evidence fo r an account involving forced reconstruction o f elem ents adjoined to sem antically nonvacuous categories seems unconvincing. It therefore c a n n o t support the more general claim th a t w ord-external adjunction —although un der restricted c irc u m s ta n c e s - exists in syntax.
Consider next the question of the interm ediate X '-leve l. If adjunction does not e x is t, then clearly this c a te g o ry type cannot be treated as a segm ent o f adjunction.
Suppose th a t there is no interm ediate X ' level. Then the question arises, h o w spec(ifiers) and com p(lem ents) can be distinguished. We could do this w ith o u t p o stu la tin g either a d junctio n structures or the existence o f categories th a t are n e ith e r w ord-level nor m axim al projections by an analysis partly in the spirit o f La rso n 's (1988) w ork. Suppose that we take a phrase to consist o f an internal XP th a t includes the head and its com plem ents and an external XP-shell th a t contains an e m p ty head and the sp e cifie r or specifiers of X as in (13). The em pty head X 1 and th e lexical head X 2 are the n taken to fo rm a unit, — a head-chain. (The tree in (13) is only partly Larsonian, since although it involves an em pty shell, it is not b in a ry branching.)
W e could then take the sp e cifie r to be th a t sister o f the higher head th a t does not c o n ta in the low er head, w h ile the com plem ent(s) w ould be sim ply the sister(s) o f the lo w e r head.
I do not th in k th a t th is solution is the correct one how ever. Larson's em pty shell approach to the evidence th a t is generally taken to m otivate binary branching s tru c tu re s is incom patible w ith the generalization th a t categorial projection and the (13)
spec X 1 X P 2
X 2 comp
selectional properties o f a head must be satisfied in the ro ot position o f its chain (Brody to appear). This problem w ould carry over to the analysis o f the phrase in (1 3 ) . In this case the subject is not in the same phrase (XP2) th a t contains th e root position o f the head-chain. The spec in (13) w ould therefore have to be selected from the position o f X 1, not the root position X 2o f the [X 1, X 2 ] chain. Furtherm ore the higher head X 1 projects an XP, again in spite o f not being in the ro o t position o f its chain.
The generalization tha t Larsonian em pty shell analyses o f m ultiple com plem entation violate was captured in the pre-m inim alist fra m e w o rk by the concept o f deep stru ctu re . Although the existence o f deep stru cture as a d is tin c t level of representation is quite dubious there are not many reasons to d o u b t the existence o f the generalization it expressed (Brody 1993, to appear, see also Chom sky 1993, 1 9 94 fo r relevant discussion). This generalization is a m ajor and a pervasive one. Not only categorial projection and them atic selectional requirem ents, but sy n ta c tic and sem antic selection in general hold in va riably in the root positions o f chains. Thus fo r example a verb V raised to som e higher fun ction al projection, say C, never forces the spec and com plem ents o f th is head to satisfy the selectional requirements o f V. I argued in earlier w o rk th a t an appropriately form ulated projection principle is com patible w ith a m inim alist fra m e w o rk and is in fa c t necessary to ensure the generalization th a t se le ctio n and categorial projection holds invariably in the root positions o f chains. T h in kin g o f the relevant features as "p ro je ctio n a l" ie. as being projected from the lexicon , I proposed th a t the projection principle requires th a t projectional features m u s t hold in and be satisfied by the root positions o f chains:
(14) Generalized Projection Principle (GPP)
Projectional requirem ents can involve only the root positions o f chains (ie. the y can hold in, and be satisfied by root positions only)
Thus I take the projection principle to require not only th a t the relevant features hold only in root positions but also that they m ust be satisfied by root positions.
Given this fu rth e r natural generalization the projection principle entails th a t on ly the root positions o f (XP-)chains can be them atic (ie. th a t m ovem ent ca n n o t land in a theta position) - th e M ain Them atic Condition.
The projectional features include centrally the categorial features o f a head.
Phrasal nodes are projected by heads, so categorial features are un der the ju risd ictio n o f the projection principle. As expected, these also take e ffe c t invariably in the ro ot position o f the relevant head chain: a verb forces the presence o f a phrasal node VP in the ro ot position of its chain, but not in any higher p o s itio n . A V raised to C for exam ple w ill not turn the CP into a VP. On the o th e r hand a categorial feature can apparently be assigned to a phrase in non-root p o sitio n s in XP-chains. I assume th a t this is because categorial features, in co n trast to selectional features, are assigned to categories and not to positions. B oth types o f features are assigned though through the position occupied by the re le va n t head, hence by the GPP only in root positions of chains. The generalization expressed by the GPP in (14) thus captures the behaviour o f categorial and various selectional
fe a tu re s and in addition it e n ta ils also the Main Them atic Condition. Since Larsonian e m p ty shells violate this pervasive generalization, the a tte m p t to find an alternative approach seems w e ll-m o tiva te d .
One possibility is to assume tha t the higher head creating the "e m p ty shell"
is in fact n o t em pty but is its e lf an ab stra ct lexical elem ent, one th a t carries the appropriate categorial fe a tu re s and selectional requirem ents o f the head w hose fe a tu re s are shared b e tw e e n a number o f head positions (This consequence o f La rso n 's approach is noted in Brody to appear.) M ultiple com plem ent verbs under a Larsonian analysis w o u ld all require such a decom position tre a tm e n t. Consider a p plying this treatm ent to th e present problem of elim inating the interm ediate X '- leve l in term s o f a s tru c tu re like (13). If X is decom posed into X 1 and X 2 and categories standardly take n as sisters o f X ' and sisters o f X are distinguished as s is te rs of X 1 and sisters o f X 2then also sim ple transitive and intransitive heads m ust decom pose in to tw o heads. The verb see w ould have to be com posed o f an agent se le cting segm ent and a non-agentive SEE, som ething like the passive was seen.
W h ile this is logically possible, there appears to be little independent evidence fo r proceeding along these lines. Furthermore w ith heads th a t assign no theta role to th e ir subjects, specifier and com plem ents could be distinguished only at the price o f po stulatin g a fully em pty head. For exam ple seem w o u ld have to decom pose into a higher head that does n o t select its subject and w h ich does not appear to c o n trib u te in any other w a y and a lower one w hich is exa ctly like seem. This seems to reduce the approach to v a cu ity.
But the re is no im m ediate need to explore this avenue fu rth e r. As far as the interm ediate X '-level is concerned, we in fa c t already have the alternative solution.
G iven dependency theory, th e asym m etry between spec and com p is ensured w ith o u t a difference in th e ir c-com m and relations to each other. Thus according to th e analysis th a t results fro m the interaction o f dependency theory and the minimal th e o ry of phrase structure based on the PPP, the basic stru cture o f an XP w ill be (1 5 ):
In (1 5) YP, th e comp asym m etrically depends on the head X, and X asym m etrically depends on ZP, the spec. S ince, more generally, com plem ents and specifiers were also defined in dependency te rm s, it is unnecessary to express this asym m etry also in th e hierarchical structure. In sum the m inim al, "v irtu a lly conceptually necessary"
th e o ry of constituency can be assumed w h ic h requires nothing beyond th a t heads p ro je c t phrases and phrases be projected by heads, —and w hich entails th a t a phrase contains its head and optionally an arbitrary num ber o f other phrases.
1.3. Some Further Consequences of the PPP and the GPP
(1 5) XP
ZP X YP
Chom sky (1 9 9 4 ) rejects the assum ption tha t "certain features (categorial features) project fro m a term inal element to form a head, then on to fo rm higher categories w ith d iffe re n t bar levels". There are at least tw o assum ptions involved here: (a) categorial features project to label higher categories and (b) higher cate go ries have d iffe re n t bar levels. But he assumes at the same tim e th a t the operation o f Merge:
"p ro je c t[s j one o f the objects to w hich it applies, its head becom ing the label o f the com plex fo rm e d ". Thus assum ption (a) th a t categorial features project is in fa c t not rejected, categorial labeling is accomplished by Merge, w h ich in addition carries out the task o f joining (projected and unprojected) elements. It is assum ption (b) that projection creates d iffe re n t bar levels tha t is rejected. Follow ing M uysken (1982), Chom sky assumes th a t bar level status (minimal, m axim al, both or neither) is c o n te xtu a lly determ ined: a maximal projection is one th a t does not p ro je ct further, a minimal projection is the lexical element itself, an interm ediate p ro je ctio n is one th a t is neither maximal nor minimal.
W ith in the minimal theory o f phrase structure proposed here, th e re are no interm ediate projection levels, and thus a category is either a head or a phrase.
Under a set-the ore tical form alization these w ould correspond to a non-set element and a set respectively. Thus there remains no 'bar level s ta tu s ' and th e re fo re no question as to ho w this is to be determ ined, - a n unprojected elem ent is a head and a projected one is a phrase. Note also th a t since branching is taken to be n-ary there is no need to re strict (directly or indirectly) the principles building syn ta ctic stru cture s to binary operations. Given the sim plest assum ption about branching, namely th a t it is in principle unrestricted, there w ill be both unary and m ore than binary branching structures.
The PPP allow s a furthe r sim plification. As we have seen clause (b) o f the PPP ensures th a t every phrase m ust im m ediately dom inate a head th a t projects it and clause (a) entails in the co n te xt of presyntactic projection th a t th is head is unique. This means th a t there is no need to label phrases at all, phrasal labels can be th o u g h t o f as only inform al notation. ocP becomes shorthand fo r a phrase headed by °c. The PPP thus eliminates the indeterm inacy o f labeling present for example in C hom sky's (1994) system , where the operation o f Merge applied to tw o categories has to specify w hich o f the tw o projected. The phrase structure becomes genuinely "ba re": phrases correspond to sets w hose elem ents are either other phrases/sets or non-set elem ents/heads. The additional and quite unnatural com plication o f allow ing some o f the elements in these sets to be labels can be dispensed w ith .
Recall th a t w ord-internal heads do not need to project phrases, hence such heads m u st be exem pted from this requirem ent. Clause (a) o f th e PPP was form ulated accordingly. Clause (b) was stated sym m etrically: all phrases m ust be projected by word-external heads. This has the im po rta nt e ffe c t o f excluding phrases adjoined to heads: such phrases w ould necessarily be projected by w ord- internal heads. (I suggested in an earlier version o f this paper th a t phrasal adjunction is ruled ou t by a PF-requirement: w ords cannot co n ta in phrasal boundaries at PF, - c f . also Chomsky 1994 fo r essentially the same proposal. But Gugliemo Cinque points out th a t examples like "his out o f this w o rld a ttitu d e ",
w he re the stress pattern indicates a w ord -inte rna l phrase, may cou nte rexe m plify th is analysis. Such exam ples are in fa c t stro ngly suggestive th a t a PF condition is n o t the re le va n t one, since they seem to result from a marginal or ungram m atical c o n s tru c tio n being saved by the device o f creating a PF phrasal w o rd . The condition prohib itin g such configuration s would th e refore appear to be s tric tly synta ctic and n o t phonological in nature.)
Since the PPP is n o t sensitive to w h e th e r a head is in the root position of its chain, it provides a stra ig h tfo rw a rd a n sw e r also to w h y "m o ved" heads (ie. heads in non-root positions) c a n n o t substitute in to or adjoin to maximal projections. This is because th e resulting co n figuration s equally violate the PPP: such heads are not dom inated by a phrase projected by the m .
As w e have seen, th e GPP ensures th a t a head in a non cha in-ro ot position ca n n o t p ro je ct, and thus in particular it ca n n o t project a phrase. Suppose th a t such a "m o ve d " head H adjoined or substituted to some phrase avoids exclusion by the PPP by projecting a phrasal node HP. There are tw o options to exclude: the projected phrase HP may be either internal to the phrase to w hich H su b stitu te d / adjoined or it may force th e category label o f the ta rg e t phrase or segm ent to be HP. Both o p tio n s are o f cou rse excluded by the GPP: projection is restricted to root positions. T h u s the GPP and the PPP to g e th e r ensure th a t "m o ve d " non chain-root heads m ust invariably be head-internal. It is im portant to see th a t there is no direct c o n tra d ic tio n between the PPP and the GPP. They only create a co n tra d ictio n for w o rd -e x te rn a l heads in non chain root positions. The PPP requires all w ord-external heads to p ro je c t a phrase and the GPP restricts all projection to ro ot positions of chains. Hence w ord-external heads th a t are not in root positions can neither project n o r not p ro je ct: they ca n n o t exist.
Let us turn finally to the question o f the explanation o f the tw o non
dependency principles in vo lve d in this a cco unt, the PPP and the GPP. As for the PPP there does not seem to be much to explain. This principle states th a t all and o n ly w ord -extern al heads have a phrase projecting property and th a t there is no le x ic o n -in d e p e n d e n tsyn ta ctic category: phrases must be projected by heads. These seem to be natural and m inim al assum ptions.
In B rody (to appear) an explanation o f the GPP is provided along the fo llo w in g lines. The question is w h y checking o f projectional features has to hold and (at lea st in the case of selectional features also be satisfied in) the m ost deeply em bedded po sition in the cha in. Take a con figuration in w hich this is not the case, w h e re a head is in a chain in w hich it projects (categorially and/or selects) from a n o n -ro o t p o sitio n or w here an XP is selected in a non-root position o f its chain.
Suppose th a t a head can o n ly project and an XP can only be selected in a single p o sitio n in a chain. (This m ay be a sim plifica tion, cf. Brody to appear). Suppose h o w e ve r also th a t all p o sitions in a chain must be projectionally identified. An assignee p o sitio n is so id e ntified if it has th e appropriate feature w h ile the assigner p o sitio n is ide ntified if it has some feature indicating th a t proper assignm ent has ta ke n place. Suppose fin a lly th a t feature percolation in chains can only take place b o tto m to to p , it is s tric tly upward dire ction al. It fo llo w s th a t the projectional
12
feature m ust be assigned to the m ost deeply embedded position in the assignee chain, o th e rw ise low er positions in this chain w ill not be projectio nally id e n tifie d . Sim ilarly the projectional feature m ust be assigned from the m ost deeply em bedded position o f the assigner chain, otherw ise the feature indicating the s a tis fa c tio n of the projectional requirem ent cannot percolate to all members o f the assigner chain.
The requirem ent th a t feature percolation in chains is s tric tly u p w a rd is in e ffe c t the representational equivalent o f the principle excluding lo w e rin g . In a derivational fra m e w o rk a representation th a t is in violation o f the GPP co u ld have arisen in tw o w ays. Either through raising in violation o f the derivational e q u iva le n t o f the GPP prohibiting m ovem ent into a position th a t involves projectional fea tures or through low ering from this position. D ow nw ard spreading o f the p rojectio nal features in a representational the ory appears to correspond to a lo w e rin g in a derivational system . This needs to be excluded. But since projectional feature checking like all feature assignm ent by heads can only involve a single p o s itio n , the GPP no w reduces to the principle that all positions in a chain need to be projectio nally identified.
(N otice th a t w hile the GPP follo w s from simple assum ptions o n ce the equivalent o f low ering is excluded in a representational th e o ry, the same is n o t true in a derivational system . Excluding lowering rules w ould no t help to e xp la in w hy raising into a projectional position is impossible. See the appendix for a discussion o f C ho m sky's (1994) alternative explanation o f the effects o f the GPP.)
II. Dependency and Binary Branching 11.1. Dual D erivations and Dependency
A sym m e tric behaviour o f anaphoric relations in the VP can be analysed in te rm s of precedence in addition to c-com m and. For example the paradigm in (16) through (19) (from Pesetsky 1992) can be accounted fo r if we assume th a t the antece den t, in addition to the c-com m and and locality requirem ents, m ust in these cases also precede the anaphor.
(16) Sue spoke to these peoplex about each o th e r 's friends in Bill's house (17) a. John spoke to Mary about these peoplex in each o th e r 's houses
b. M ary danced w ith these peoplex in each o th e r 's hom etow ns (18) *Sue spoke to each o th e r 's friends about these peoplex
(19) a. *Sue spoke to Mary about each o th e r 's flaw s in these houses b. *M a ry danced in each o th e r 's cities w ith these m ayorsx.
The curious conjunctio n of precede and c-com m and can be elim inated in te rm s of c-com m and defined on binary rightw ard branching stru ctu re s, (eg. Larson 1988, Pesetsky 1992, Kayne 1993) Given the evidence discussed above a g a in s t the Larsonian em pty shell approach from the projection principle, I shall a d o p t as a basis fo r discussion the analysis in Pesetsky 1992. Given binary branching trees like those in (20), the antecedent w ill c-com m and the anaphor in all and o n ly the gram m atical examples in (16) through (19): precedence is elim inated.
(20) VP
these people P PP
t S
ab o u t DP th e ir friends P
/
in DP
the garden PP
P DP
\ ^
on Tuesday As Pesetsky em phasizes, although such structures appear to be m otivated in various w ays, they do n o t provide the appropriate co n stitu e n t stru ctu re fo r m o vem e nt/ch ain relations. A s (21) sh o w s, on many o f the co n stitue nts in (20) no non-trivial chain can be fo rm e d . The m embers of the chains in (22) and (23) are not c o n s titu e n ts o f the b in a ry branching tree, but only o f the more traditional re pre senta tion in (24), —the examples are again Pesetsky's, w ith minor m o d ifica tio n s:
(21) a. * [T o John about him self] M ary s p o k e _
b. * l w onder [to w h o m on Tuesday] Mary s p o k e __
(22) a. [T o which ad ults] Sue showed the k id s _on each o th e r's birthdays?
b. [T o none of the o ffic ia ls ] did Sue send m o n e y _on each o th e r's birthdays c. [O n which table] did John put the book _during its construction?
d. [T o the children] w e re given b o o k s _on each o th e r's birthdays.
(23) a. ... and [give the b o o k to them in the garden] he did [_on each oth e r's birthdays]
b. ... and [give the b o o k to them ] he did [_in each o th e r's garden]
Thus there appears to be system atic evidence fo r both types o f phrasal organization. A ccord ingly, Pesetsky postulates th a t sentences have tw o parallel derivations, one involving binary branching and the other more tra d itio n a l tre e s, his
"cascade" and "laye red " representations. He suggests th a t selectional and other related re strictio n s hold in both derivations and postulates a set o f correspondence principles to relate them . But this means th a t the approaches in term s o f binary branching stru ctu re s to the problem o f anaphora are not fully successful e ith e r since the y create another unexplained duplication.
D ependency the ory captures the operative relations o f binary branching analyses and thu s it makes such analyses dispensable. In discussing th e relation betw een dependency theory and the theory o f phrase stru ctu re I have been co n centratin g on structures where a phrase contains only tw o m axim al p rojectio ns, a spec and a com p. Let us now consider more closely the situa tion w h e re an XP im m ediately dom inates more than one com plem ent: VPs w ith m ultiple com plem ents and non-subcategorized elements.
(25) X
give the book to them in the garden
In order to ensure the correct co n stitu e n t structure, I analyse th e non subcategorized PP in (25) as a sister o f a higher head. (Recall th a t I assum e the nonexistence o f phrasal adjunction.) Single line arrow s indicate th e lexically determ ined head-com p dependencies. These relations leave the tw o PP's " to them "
and "in the garden" unordered w ith respect to each other and the re st o f the stru cture . Let us assume tha t structural dependency relations c a n n o t cross argum ent (CP and DP) category boundaries. One of the w ays in w h ic h (25) can satisfy the axiom s o f dependency theory is if the tw o spec-head dependencies (indicated by double line arrows) are present in the stru cture . This w ill ensure that the stru ctu re expresses both the con stitue ncy facts and the a sym m etry sh o w n by the behaviour o f anaphoric elements.
Given the assum ption that anaphoric relations are defined over dependency stru ctu re s, it is then necessary to form ulate the dependency equivalent o f principle A. For th is a dependency equivalent o f c-com m and is necessary. This relation, c- dependence, is defined in a parallel w a y to c-command in (26). C-command is restated fo r comparison in (27) and the dependency version o f principle A is given in (28) :
(26) y c-depends on x if f y depends on x or z contains y and z depends on x (27) x c-com m ands y if f x, y are sisters or
z contains y and x , z are sisters (28) Principle A
An anaphor must c-depend on an antecedent in its local domain
Since all dependencies are right to le ft by the Precedence requirem ent, it is easy to see th a t the p rin cip le A in (28) w ill be violated by the ungram m atical examples in (18) and (19). In contrast, those in (16) and (1 7) w ill sa tisfy (28), since in each o f these the a n a p h o r is in a DP th a t indirectly depends on the antecedent o f this anaphor. It is clear the n that the analysis in (25) can account fo r both the co n s titu e n c y and the a n aph oric relations exhibited by m ultiple com plem ent VPs.
N otice tha t there are alternative dependency structures fo r (25). The spec- head dependency from th e preposition in could lead not to the DP them but to the PP or VP dom inating th is DP. In sentences like (16) and (17) sim ilar dependency stru cture s w o u ld violate prin ciple A: under such analyses the anaphor w ou ld not c- depend on its antecedent. If we assumed th a t only argum ent categories (DPs and CPs) can serve as non-derived (chain root) spec nodes (Kayne 1984, Pesetsky 1992) then o n ly the dependency analysis indicated in (25) w ould be legitim ate, this principle w o u ld rule out th e alternatives w ith PP or VP spec's. But anaphoric relations appear to provide no m otivation fo r adopting the restriction: principle A for example can be satisfied if there is at least one dependency analysis o f the stru ctu re th a t satisfies it. Sim ilarly for o th e r anaphoric requirem ents.
For another illu s tra tio n of the inte ractio n o f dependency th e o ry w ith the m ultiple branching phrase stru cture , consider the examples usually taken to involve Heavy S h ift in (29):
(29) John offended t [by n o t recognizing pg] [his favourite uncle fro m India]
(30) We gave t to them on Friday [copies o f the reports on each other]
This operation exhibits b o th raising and low ering effects. The fa c t th a t it appears to license parasitic gaps as in (29) show s th a t the shifted phrase is higher than its tra ce. On th e other hand th e fa c t that an anaphoric elements is licensed inside the shifted phrase whose a n te ce d e n t is not higher than the trace o f the shifted phrase suggests low ering.
W illiam s (1990, 1 9 9 2 ) has denied th a t Heavy Shift has properties o f A '- chains, arguing against th e claim that it licenses parasitic gaps. He takes examples
like (29) to be derived by the operation of Right node raising. His evidence is th a t co n stru ctio n s w ith parasitic gaps like (29) allow stranded prepositions, as in (31).
As he points ou t the option of stranding prepositions is a prope rty ch a ra cte ristic o f Right node raising but not of Heavy Shift:
(31) John yelled at t, before punishing pg, all of those campers (32) John talked to, and I yelled at, Peter
(33) *l talked to, about Mary, his favourite uncle
A lthou gh this data m ight suggest a w ay o f avoiding the conclusion th a t Heavy S hift can license parasitic gaps, it clearly does not force this conclusion. This evidence is equally com patible w ith the claim th a t parasitic gaps are licensed both by Heavy S hift and by Right node raising, although prepositions are strandable o n ly by this latter operation. Examples like (34) show th a t the conclusion th a t the re is some operation (presumably Heavy Shift) th a t has both raising and lo w e rin g properties cannot be avoided in the w ay Williams suggests. These examples e x h ib it both types o f properties (leftw ard parasitic gap licensing and rig h tw a rd anaphor binding) sim ultaneously:
(34) a. I read to them , w ith o u t carefully checking pg, several reports on each other
b. We read to each boy, a fte r discussing pg in private, a report on his activitie s.
If anaphora is a dependency relation and the licencing o f (syntactic) variables -w h e th e r ordinary traces or parasitic g a p s - is not, then the m odular analysis th a t postulates a dependency structure again predicts exa ctly the observed facts. As the stru ctu re o f (34a) in (35) show s, the trace and the parasitic gaps are c-com m anded by the shifted phrase but the anaphoric element inside the shifted phrase can c- depend on its antecedent (I assume w ith Pesetsky th a t the shifted phrase is a PP):
Thus the m odular interaction o f dependency the ory w ith the m inim al th e o ry o f phrase stru ctu re can reconstruct Pesetsky's dual derivation analysis, w ith o u t the duplicate selectional restrictions and correspondence principles o f his analysis. But perhaps even more im po rta ntly, the interaction o f dependency and phrase s tru c tu re th e o ry is naturally taken to provide a classification o f syn ta ctic relations. Given the Precedence principle, we expect precedence or binary branching e ffe cts b e tw e e n
17
tw o elem ents just in case that there is a dependency relation betw een them.
11.2. Dependency and Precedence
Typical such cases in v o lv e anaphoric dependencies: anaphors under principle A, polarity item licensing o r pronominal bound anaphora. Consider pronom inal bound anaphora. Williams (1 9 9 2 , 1994) proposed an account o f WCO and other related relations in terms o f a restriction th a t anaphoric dependencies are governed by a precedence requirem ent, --an approach partly sim ilar to C ho m sky's "le ftn e ss"
c o n d itio n . Thus in the case of the W CO con figuration in (37), the pronoun cannot be de p e n d e n to n its p u ta tiv e antecedent everyone because this antecedent does not precede it:
(37) His m other likes everyone
W illiam s account o f W C O w ill fo llo w from the Precedence requirem ent proposed here if it is generalized in the ob vious w ay: the restriction o f Precedence to s tru ctu ra l dependencies needs to be removed so th a t it constrains all synta ctic dependencies. (Notice th a t to preserve earlier results, T o ta lity and Non-redundancy m ust n o t be sim ilarly generalized, th e y must remain conditions on structural dependencies.)
(38) Precedence (Generalized)
if x depends on y then the term inals dom inated by y precede th e terminals dom inated by x
G iven a copy th e o ry of chains, the more com plex, re construction cases o f WCO w ill also fall o u t on the assum ption tha t the bound pronoun m ust depend on an an tece den t in A -p o s itio n . (See Brody to appear, for argum ents fo r a theory where all copies in a ch a in are present sim ultaneously at LF).
(39) W ho does his m o th e r like (who)
(4 0 ) ? W h ic h picture o f w h ic h artist does his m other like (w hich picture o f w hich artist)
(41) W h ich o f his p ictu re s does every painter like best (w hich o f his pictures) In (40), w he re the a n te ce d e n t is "re c o n s tru c te d ", no V-chain and no accessibility co n d itio n (H igginbotham 1983) is necessary. The antecedent in the copy position is to the rig h t of the pronoun. (41), w here the pronom inal is "re co n stru cte d " is co rre ctly allowed on th e assum ption th a t an anaphoric element in a chain needs to satisfy its dependency conditions in only one of its positions in the chain (Barrs 1986, C hom sky 1993), -h e re the co p y position.
W illiam s generalizes his tre a tm e n t o f WCO as involving a dependency governed by precedence to other anaphoric relations including the relation between
a pronoun and its nonquantificational antecedent. Since the generalized Precedence principle o f the dependency the ory proposed here is identical to W illia m s' assum ption about the restriction governing these cases it should be clear th a t the present acco unt can be sim ilarly generalized.
Consider in co n tra st principle C of the binding theory. This is not a re s tric tio n on antecedence/dependence but rather a disjointness requirem ent. For exam ple in (42) coreference betw een he and the object John is prohibited w h e ther or n o t there is an alternative antecedent for the pronoun.
(42) *(J o h n x's m other said tha t )hex thought I liked Joh nx.
Therefore the present th e o ry predicts that principle C m ust be stated in te rm s o f c- com m and and th a t this principle w ill show no binary branching or precedence e ffects. Thus the fa c t th a t it does not require disjointness in the w ell kn o w n cases o f (43a,b) is as expected if the adjunct and the extraposed clause are in a high enough position in w h ich the pronoun does not c-com m and the name th a t the clause includes:
(43) a. We sent him x there in order to please Jo h n x's m other b. Someone had phoned herx who M aryx m et at the party
Again, as is w ell know n, there is independent evidence from c o n s titu e n c y tests th a t there exists a VP in (43) th a t does not include the sentence fina l clauses, hence also fo r the claim th a t the pronoun indeed fails to c-com m and the elem ents contained in these clauses. Final clauses like the ones in (43) can be le ft stranded by VP-deletion as (44) and (45) exemplifies, and sentential adjuncts can also be stranded by V P -fronting, as in (46):
(44) A lthou gh none o f the MEN did who were visiting from NEW YORK, several o f the WOMEN w e n t to the concert w ho were visiting from BOSTON
(Culicover and R ochem ont 1990)
(45) a. MARY sent him there in order to PLEASE Joh n 's m other and KLARA did in order to UPSET her.
b. A lthou gh MARY did in order to UPSET J o h n 's m other, KLARA sent him there in order to PLEASE her.
(46) ...and send him there Klara did, in order to please J o h n 's m other
Sentential adjuncts like the in order to clause in (43a) behave like th e locative and tem poral adjuncts and the Heavy Shifted elem ents considered earlier: under an anaphoric dependency like WCO they act as if the y were embedded more deeply in the tree:
(47) W h o x did you criticize t x in order to please him x
As usual w ith con figuration s where an element on the right is in a higher position examples can be constructed th a t exhibit both configuration al and dependency relations sim ultaneously, creating an apparent co n tra d ictio n . (48) fo r exam ple
vio la te s neither WCO (a dependency e ffe ct) nor principle C:
(48) a. W h o x did you talk to t x about hery before M a ryy told him x not to listen b. I sent each boyx to hery in order to make M a ryy meet him x
It has been suggested th a t principle C can be stated on binary branching tre e s. But judgem ents on exam ples like (49) are quite equivocal:
(49) a. Sue spoke to him x about Billx's mother
b. Sue spoke to Mary about him x in Billx's house
The slight unaccep tab ility o f these could be due to the infelicitio us repetition o f an R-expression presupposed in the co n te xt. The difference between a violation o f this w e a ke r prohibition against repeating an R-expression and genuine principle C effe cts is illustrated by (50) and (51). (50), w hich violates the w eaker, presum ably pragm atic, prohib itio n co n trasts w ith the more stro n g ly ungram m atical (51) th a t vio la te s principle C:
(50) ?Bill's m other spoke to Bill (51) * Bill spoke to Bill's m other
(52) a. Sue spoke to Billx about Billx's mother
b. Sue spoke to Mary ab out Billx in Billx's house
The examples in (52) appear to be on a par w ith (50) rather than (51), reinforcing the conclusion th a t principle C is not sensitive to precedence in the same w ay as p rin ciple A or WCO is. Thus it is not a principle stateable on binary branching trees.
As noted, the present approach predicts this result: since principle C is not a dependency principle it w ill not constrain dependencies and should therefore be se n sitive to standard c-com m and.
Notice th a t the present approach entails also th a t w herever there is subject o b je c t asym m etry w ith respect to principle C, the subject m ust be outside the (lo w e st) VP th a t contains the object: otherw ise there can be no asym m etry e ffe cts.
Haider (1 9 9 3 , 1994) w h o proposes a version o f the binary branching hyp oth esis notes problems w ith stating principle C on such trees. He then argues th a t principle C is not a stru ctu ra l principle at all, he claims th a t it could not be sta te d on non-binary branching trees either. He questions the account o f (43b) th a t assum es th a t the relative clause extraposed from subject is in a high enough p o sitio n in the tree where the object does not c-com m and it. Many o f his argum ents do not distinguish betw een various types o f e xtrapo sition con structio ns.
I shall look brie fly below at some o f those th a t are relevant to the question o f the p o sitio n of the extraposed clause in extraposition from DP co n structio ns. His main a rg u m e n t is th a t extraposed relatives alw ays precede an extraposed argum ent clause and ob ject pronouns m ust be disjoint from any R-expression in the argum ent clause:
(53) Someone has told herx [w h o M aryx met] [that M a ry „x w ill inherit the castle]
If the pronoun in (53) must c-com m and the extraposed a rg um ent clause th e n there is no w a y to avoid it c-com m anding the extraposed relative. Since no disjoin tne ss e ffe c t obtains w ith the R-expression in the relative this appears to q u estion the possibility o f a structural account o f principle C.
There is strong evidence how ever tha t extraposed o b je ct argum ent clauses originate under the VP, as originally proposed by S tow ell 1 9 8 2 (see also Postal 1986, Pesetsky 1992). If w e take this extraposition to be A '-m o v e m e n t, then principle C should hold fo r the VP-internal position, ju st like in the case o f (54) it holds fo r the source position:
(54) *[W h ic h claim th a t Joh nx was asleep] do you th in k hex denied t
Hence (53) in fa c t creates no problems. It's structure is like (55), w h e re the pronoun c-com m ands the R-expression in the appropriate p o sition in the ch a in of the extraposed argum ent clause, but fails to c-com m and th e R-expression in the extraposed relative whose stru ctura l position is higher:
(55) Someone has told [VP herx [th a t M a ry.x w ill...] ][w h o M a ry x m et] [that M a ry w ill inherit th e castle]
(M artina W iltschko points out th a t when the R-expression is in a relative clause inside the extraposed object argum ent clause, the exam ple im proves. If co rre ct, this provides fu rth e r evidence for the analysis suggested here. Since non- selected argum ents (adjuncts) like the relative clause re c o n s tru c t only o p tio n a lly (Lebaux 19 89), the contrast between this case and (53) w o u ld be predicted on the re co n stru ctio n analysis proposed -cf. Brody to appear.)
A n o th e r set of argum ents against adjoining the extraposed relative clause higher than the VP and thus against the possibility o f a stru ctura l a c c o u n t of principle C predicated on non-binary branching trees involves binding fa c ts .
(56) I w o u ld not tell everyonex all the details at once [th a t hex m ight be interested in]
Haider argues th a t in (56) the trace o f the quantifier m ust c-com m and the pronoun, o th e rw ise the structure w ould give rise to a WCO violatio n. It would fo llo w from this th a t the extraposed clause m ust be not higher in the tree tha n the com plem ents o f the verb, making a stru ctura l explanation o f the fa ct th a t an R-expression in an extraposed relative can be object bound impossible. But as w e have seen, W CO is a dependency principle, hence it provides no evidence fo r the c-com m and relations invo lvin g the extraposed clause.
A rguing against C ulicover and Rochem ont's IP -adjunction analysis, Haider show s th a t re strictive relatives cannot take antecedents s p lit between elem ents in tw o co-ord in ate clauses:
(57) * l'll in te rv ie w everyone and tape every man here w h o kn o w each o th e r
