representations start at the bottom with the Wrst merge operation, and then progresses upwards. (‘‘a’’ stands for a, ‘‘p’’ stands for portrait,‘‘s’’ stands for of someone, and ‘‘t’’ stands for the relative clause that Rivera painted. (45) {a, {{p, {p, s}}, a}} 4 " {hp, pi, {{p, {p, s}}, t} 5 " a Merge !{p, {p, s}} Pair merge ! t 3 " 2 p Merge ! s 1 Simple merge combines p and s (portrait of someone) into the set {p, {p, s}} in step 1 . This set is the input into two distinct operations which are not ordered with respect to each other, nor interact with each other. In 2 , the adjunct that Rivera painted (t) is merged with {p, {p, s}} resulting in {hp, pi, {{p, {p, s}}, t}} portrait of someone that Rivera painted. Simultaneously, {p, {p, s}} is identiWed in 3 as a member of the set {a, {{p {p, s}}, a}} a portrait of someone. It is this later set that merges into the derivation with the potential c-commander that should trigger (but fails to trigger) a condition-C eVect. The set containing t (that Rivera painted) 5 intersects with 4 , accounting for the compositional meaning, but is not part of it—explaining the Lebeaux eVects. The set- theoretic notation then amounts to a tiered three-dimensional repre- sentation, much like that found in autosegmental phonology, except that the tiers are not linked geometrically but through set intersection. We return to multidimensional representations in Chapter 10, but it is worth noting that this explanation fails in one crucial regard—it does not explain why sentences such as (44d) are ungrammatical. If adjuncts are on a diVerent dimension and do not interact with the c-command relationships in the non-adjunct portions of the representation, then (44d) is predicted to be fully grammatical. One possible solution, which lies beyond the scope of this book, can be found in Stepanov (2001). 8.4.3 Bottom-to-top and top-to-bottom derivations Another consequence of the BPS system is the strictly upwardly cyclic nature of the phrase structure derivation. The merge operation works from the terminals to the root. This is of course true of all projection- based theories of phrase structure, but is particularly acute in BPS where there is no plausible top-to-bottom or purely representational 160 controversies equivalent. This is even more the case in the latest ‘‘phase-theoretic’’ versions of BPS, where the bottom-to-top derivation is crucial to the mechanics of the Move operation. 8.4.3.1 Bottom-to-top derivations It has long been observed that both syntactic and syntactico-phonological operations are frequently limited in their scope by the hierarchical structure of the sentence (see for example the notion of a ‘‘kernal sentence’’ found in Chomsky 1955, 1957; and the notion of cycle found throughout the 1960s and 1970s, Fillmore 1963; the notion of Government Domain in the 1980s; the notion of minimality in the 1990; and phases in the 2000s). In particular, various phenomena from stress placement to binding theory to displacement operations seem to function as if they work on the smallest, most embedded structure outwards to the largest least embedded structure. If BPS is a strictly bottom-to-top model, we can see syntactic cycles as a natural consequence of the derivation procedure. This was Wrst pointed out by Epstein (1999) and Epstein et al.(1998), who argue that all locality constraints, especially those based on c-command relationships, follow from this basic organizational principle (as you will recall from our discussion in Ch. 3, section 3.5.4).22 8.4.3.2 Top-to-bottom derivations A consequence of bottom-to-top derivations—especially ones that are subject to the LCA—is that they are also typically right to left (or more saliently, from the end of the sentence to the beginning). There is something singularly counter- intuitive about this, especially since sentences are obviously spoken (and processed) from the beginning to the end. We might use the old competence–performance distinction to get out of this, but it still feels like a cop-out in the end. Phillips (2003) presents some evidence that a top-to-bottom tree derivation23 elegantly explains why some constituency tests appear to 22 Uriagereka (1999) converges on a similar view, but suggests further that units formed by merge (command units) are spelled out cyclically. This in turn leads to the Chomskyan Phase Theory (Chomsky 2001) and the Minimalist/Tree Adjoining Grammar fusion of Frank (2002). Interestingly, Uriagereka’s command unit hypothesis also explains why structures like [ A [ B DE][ C F G]] are not violations of the LCA, since B and C would be independently spelled out. 23 Richards (1999) has also argued for a top-to-bottom derivation using data from island eVects (and the occasional lack thereof). Baltin (2006) provides an alternative analysis of some of Phillips facts using VP-remnant movement. set-theoretic constituency 161 be contradictory. Consider the following sentences taken from Phillips, which show straightforward evidence for a rightward-branching struc- ture. Coordination (46a), ellipsis/replacement (46b), movement (46c), binding/c-command (46d, e) all seem to point to a Verb þ Object constituent (all examples taken from Phillips). (46) (a) Gromit [likes cheese] and [hates cats]. VO constituent (b) Gromit [likes cheese] and Wallace does too. VO constituent (c) [Like cheese] though Gromit does t, he can not stand Brie. VO (d) Wallace and Gromit like each other. (e) *Each other like Wallace and Gromit. But when we look at more complex constructions like ditransitve verbs, the results seem mixed. We see behaviors where strings that we normally identify as non-constituents obey some constituency tests (47a–c), but the same strings do not obey others (47d–e). (47) (a) Wallace gave [Gromit a biscuit] and [Shawn some cheese] for breakfast. (b) [Wallace designed] and [Gromit built] an enormous tin moon rocket. (SV) (c) Alice [knew that Fred wanted to talk] and [hoped that he wanted to argue] with the president. (d) *[Gromit a biscuit] Wallace gave t for Breakfast. (e) *Wallace gave t at breakfast [his favorite pet beagle an enor- mous chewy dog biscuit]. Phillips claims that an explanation for these kinds of phenomenon follow if we build trees incrementally from left to right (and conse- quently in a right-branching language, top to bottom.) The basic rule of tree construction is given in (48). ()X X ABAY BC The behaviors seen in (47) follow if constituency tests apply to those strings that are constituents at the point in the incremental derivation the phenomenon is exhibited. This can be seen in the various 162 controversies applications of coordination at various left to right portions in the following sentences. (49) (a) [Wallace will] [Wallace will] and [Wedonlene probably won’t] [Wallace will] and [Wedonlene probably won’t] give Grom- mit crackers. (b) [Wallace will give] [Wallace will give] and [Wendolene will send] [Wallace will give] and [Wendolene will send] some crackers. (c) [Wallace will give Gromit] [Wallace will give Gromit] and [Wendolene will give Preston] [Wallace will give Gromit] and [Wendolene will give Preston] a new collar. (d) [Wallace will give Gromit crackers] [Wallace will give Gromit crackers] and [Wendoline will give Preston dog food] [Wallace will give Gromit crackers] and [Wendoline will give Preston dog food] before breakfast. (e) [Wallace will give Gromit crackers before] [Wallace will give Gromit crackers before] and [Wendoline will give Preston dog food after] [Wallace will give Gromit crackers before] and [Wendoline will give Preston dog food after] breakfast. 8.4.4 Derived X-bar theory The particular version of derived X-bar theory found in BPS has some interesting empirical and theoretical consequences. Consider the case of a head that does not project. Such an element is simultaneously an X8 and an XP (as it is dominated by the projection of another cat- egory). Interestingly, such animals appear to exist. Clitics have the behavior of elements that are simultaneously X8s and XPs. They are theta-marked and allow violations of the head-movement constraint, yet they adjoin to other X8s; such behavior is expected of elements that are ambiguous in their phrasality. Chomsky limits the range of such ambiguity by proposing that morphology (or its LF analog ‘‘word interpretation’’) gives no output to an element that has complex internal structure (cf. Nun ˜ es 1998). So set-theoretic constituency 163 we expect to Wnd cases like the clitic case above, where a simplex element is simultaneously an XP and an X8, but complex elements are not allowed in this position. In Carnie (1995) and (2000), I argue that a more liberal deWnition of phrasality is in order. In particular, I suggest that the phrasal status of a constituent is not determined by its position in the projection struc- ture in the tree. Further, I claim that phrasal status does not predict outward behavior with respect to interface conditions and constraints (such as the morphology, thematic relations, case marking, and head- movement rules). Rather, the syntax can operate freely on any node in the tree, applying operations as needed. The output conditions, in particular the morphology, then determine what nodes are to be spelled out as words and which are to be spelled out as phrases. The cases that show this to be true are those where we have either an apparently X8 functioning like a phrase (as in the clitic case mentioned above) or more interestingly, an apparent XP functioning like an X8. We can identify such cases with the following criteria: if a particular p-marker behaves diVerently in terms of its phrasality with respect to two diVerent output conditions or two diVerent components of the grammar, then it must be the case that its phrasality is not determining its behavior, but rather the behavior determines the phrasality relative to the particular output condition or component of the grammar. The Wrst case I consider are copular constructions in Irish. Irish is a VSO language, and I assume that this order is derived via movement of the V around the subject into some functional head via head movement. There is a privileged position occupied by predicates in the language, which is between sentence initial complementizer particles and agreement (Particle < Predicate < Agreement < Subject < .). Following standard practice (Sproat 1985 and many others), I identify this position as being derived through head movement. In predicative copular constructions (ºx[NP’(x)]) and nominal and other non-verbal predicates appear in this position (a), even when they are phrasal (b): (50) (a) Is ollamh (e ´ ) Aindriu ´ . C professor (agr) Andrew ‘‘Andrew is a professor.’’ (b) Is amhra ´ na l bhuailWdh an pı ´ obaire ‘‘Yellow Submarine.’’ C song C play.fut the piper ‘‘‘Yellow Submarine’ is a song which the piper is going to play.’’ 164 controversies Here we have a surprise. If indeed the predicate initial order is derived through head movement then we do not expect phrases to appear in this position. Nevertheless there is evidence to suggest that these phrases function like X8s with respect to other parts of the grammar. I’m going to contrast the predicative cases to equative constructions (ºxºy [COP’ (x, y)]) in the language, which do not have head move- ment to the privileged position. (52)Ise ´ Aindriu ´ an t-ollamh. C agr Andrew the professor ‘‘Andrew is the Professor.’’ Irish does not, in general, seem to have the usual brand of islands. In particular, no nominal or wh-islands are observed as long as a special complementizer type and resumptive pronoun are used (53). The exception to this appears to be precisely in the head-moved XP in predicatives (54), where an island emerges. Contrast this with the equative construction where there is no predicate movement (55).24 (53) (a) Bı ´ onn Wos agat i gconaı ´ [ cp caide ´ i a L bhuailWdh be.hab know at.2.s always what i C play.fut an pı ´ obaire t i ]. the piper t i ‘‘You always know what the piper will play.’’ (b) Ce ´ nPı ´ obaire j [ cp a N mbı ´ onn Wos agat i gconaı ´ which piper C be.hab know at. 2.s always [ cp caide ´ i a L bhuailWdh se ´ j t i ]]? what i C play.fut. him ‘‘Which piper do you always know what he will play?’’ (54) *Ce ´ nPı ´ obaire j arb [ np amhra ´ n i [ cp a L bhuailfeadh se ´ j t i ]] (e ´ ) Which piper rel song C play.cond him agr ‘‘Yellow Submarine?’’ ‘‘*Which Piper is ‘Yellow Submarine’ a song which he/t i is going to play?’’ (55)Ce ´ nPı ´ obaire j arb e ´ ‘‘Yellow Submarine’’ [ np an t-amhra ´ n i [ cp Which piper re agr the song a L bhuailfeadh se ´ j t i ]]? C play.cond him ‘‘Which Piper is ‘Yellow Submarine’ the song which he/t i is going to play?’’ 24 These factsare similar to the facts of canonical and reverse constructions in Italian discussed in Moro (1997). However, as discussed in Carnie (1995),thesearequiteadiVerent set of facts. set-theoretic constituency 165 Relatedly, wh-in-situ is usually not allowed in Irish, but is required when the question word is inside a complex predicate (56b). (56) (a) Is [dochtu ´ ir ainmhı ´ ] Daibhı ´ . C doctor animals.gen Davey ‘‘Davey is a doctor of animals.’’ (b) Ø [Cen sort dochtu ´ ra] Daibhı ´ ? what kind doctor.gen Davey ‘‘What is Davey a doctor of ?’’ (lit. ‘‘Davey is what kind of doctor?’’) Other evidence comes from the responsive system of Irish. Standard Irish lacks words for yes and no. Instead, the verb is typically repeated in either positive or negative form. (57) (a) An bhfaca tu ´ an teangeolaı ´ ? Q saw you the linguist ‘‘Did you see the linguist?’’ (b) Nı ´ fhaca. (c) Chonaic. neg saw saw ‘‘No.’’ (lit. ‘‘not saw’’) ‘‘Yes.’’ (lit. ‘‘saw’’) McCloskey (1991) analyzes this as a kind of ellipsis, where everything but the verb in InX and any complementizer particle (i.e. the InX head and its adjoined C) is elided. It follows that any predicated in InX, even complex ones should remain in response to a yes/no question. This appears to be the case (although the complex cases show up as a pronoun). (58)Q:Anle Sea ´ n an Subaru? A: Is leis. ‘‘Yes.’’ Q with John the Subaru CþINFL ‘‘Does Joh own the Subaru?’’ C with.him (59) Q: An ceart mo chuimhne? A: Is ceart. Q right my memory CþINFL ‘‘Is my memory is right?’’ ‘‘yes’’ (from Doherty 1992) (60) Q: An dochtu ´ ir tu ´ ? A: Is ea. Q doctor you C it ‘‘Are you a doctor?’’ ‘‘Yes.’’ 166 controversies I presented other arguments from the construct state in Irish vs. Arabic, the Ezafe construction in Persian, and from CP-inside-N nouns in English. This analysis has been strongly criticized. Among other arguments, it has been claimed that ambiguous phrasality should show greater exponence in grammar. Many others have suggested that an alternative analysis should be given to the predicate initial orders exhibited in Irish and other VSO languages. In particular, several scholars have suggested that such orders arise through VP remnant movement and full NP movement. For discussion see Adger and Ramchand (2003), and many of the papers in Carnie and Guilfoyle (2000) and Carnie, Harley, and Dooley (2005), especially McCloskey (2005) and Massam (2005). 8.4.5 Label-free and projection-free structures In an inXuential but controversial paper, Collins (2002) also suggests that the BPS representation contains too much information. He sug- gests that labeling in general (including feature structure) can be dispensed with and reduced to mechanisms motivated elsewhere in the grammar. So a phrase {V, {V, X}} is actually {V, X}. In Collins’s system, instead of categories, operations target the featural properties of the heads of phrases. Without labels, the principles of phrase structure composition must target elements that are complete. For example, a verb such as see takes a DP as a complement, targeting the relevant features in a D head. But for reasons of compositionality, it is crucial that it not target a determiner such as the by itself. In versions of BPS that have labels, this is trivial since the verb looks for a DP. In label-free BPS this is accomplished by placing a restriction on the grammar (known as the Locus Principle), whereby two elements can only be merged if the non-head element is complete in terms of its featural requirements. So the must combine with an N before it can be merged with a head with unchecked requirements. Again a conse- quence of this is that constituency representations must be derived strictly bottom to top. As observed by Collins himself in later work, this proposal also naturally leads to the view that syntactic structure shouldn’t take constituency as the primary relation, but instead leads to a view where semantic dependencies lie at the heart of the matter. We turn to this claim in the next chapter. set-theoretic constituency 167 9 Dependency and Constituency 9.1 Introduction1 Up to this point in the book we’ve been assuming that constituency representations—usually trees or sets—are primitives of syntactic rep- resentation and the primary means of representing syntactic structure. Semantic relations (such as thematic relations) and semantico- syntactic relations (such as subjecthood, or even predicate–argument relations) can be viewed as derived from, or at least dependent on, constituent hierarchies. For example, in many of versions of Chom- skyan generative syntax (Extended Standard Theory, Government and Binding, Minimalism), grammatical relations such as Subject are arboreally deWned. In the EST, the ‘‘subject’’ of the sentence is the non-VP daughter of the S node. In GB and MP, the notion of subject is at least partly synonymous with speciWer (and in particular the spe- ciWer of IP/TP/AgrSP and NP/DP). Objects are also deWned in terms of trees with a variety of deWnitions. Hale and Keyser (1991) argue that it is not only grammatical relations that are deWned by argument position in a tree; for them, thematic relations (agent, theme, etc.) are deWned by the base position of the argument in the tree. In particular, the phenomenon whereby certain thematic relations such as agent are more prominent than others (such as theme or goal) corresponds to the position of those arguments in the tree in their base positions. Even such rich notions as quantiWer variable-scope relations can be viewed as dependent upon the hierarchy of the constituentstructure (May 1985; Diesing 1992). Diesing and Jelinek (1995) and Jelinek and Carnie (2003) take this idea and propose that in fact almost all hierarchy eVects (such as person rankings, deWniteness eVects, ergative splits, and animacy eVects) are in eVect side eVects of hierarchical constituent structure. 1 For excellent in-depth and textbook treatments of many of the questions touched on in this chapter, see Van Valin 2001) and Moravcsik (2006). However, such a constituent-centric view is by no means universal. The idea where the relationships are reversed (that is, constituentstructure is dependent upon semantic relations2) is much older (dating at least to Pa ¯ n _ ini or Aristotle) and is fairly pervasive outside the Chomskyan tradition. There are various approaches to non-constituency-based syntax (perhaps better referred to as ‘‘derived constituency’’ approaches). These vary along the following dimensions: (1) (a) The degree to which relational structure determines constitu- ent structure or vice versa. (b) Whether the relational structure is based on syntactic cat- egory, grammatical relations, or more traditional semantic relations (such as predication or thematic relations). (c) The mechanism by which constituent structures are derived: by dependency grammars, type-logical proofs, mapping rules, or construction schemata/templates. The variation in (1a) concerns the degree to which constituency is dependent on semantic or semantico-syntactic relations and to what degree it is independent. On one extreme we have the view articulated above and in previous chapters, where syntactic constituency is pri- mary and semantic relations are dependent on these. On the other extreme, we have views like those of Dependency Grammar in its various guises including, Word Grammar, Functional Grammar, Construction Grammar, and Cognitive Grammar surprisingly, some recent versions of the Minimalist Program, where constituentstructure is entirely or mostly derived from semantic relations. In between these two extremes we have views where we have simul- taneous and equal semantic and syntactic representations. This is represented in the views of Lexical-Functional Grammar and Role and Reference Grammar. The variation in (1b) refers to the type of semantic or syntacticose- mantic relationship serving as the primitive from which constituentstructure is derived. Theories such as Lexical-Functional Grammar, Relational Grammar and to a lesser degree Construction Grammar rely on primitive grammatical relations or functions such as ‘‘subject’’ and 2 Miller (1999) provides formal proofs that dependency grammars and (some) constituent-grammar types are equivalent in terms of their expressive power (strong and weak generative capacity). dependency and constituency 169 . in Chomsky 195 5, 195 7; and the notion of cycle found throughout the 196 0s and 197 0s, Fillmore 196 3; the notion of Government Domain in the 198 0s; the notion. variable-scope relations can be viewed as dependent upon the hierarchy of the constituent structure (May 198 5; Diesing 199 2). Diesing and Jelinek (199 5)