November 1998

1 Introductory remarks

Almost twenty years ago, Riny Huijbregts and I conducted a series of tape-recorded discussions with Noam Chomsky which later appeared as The Generative Enterprise (Chomsky (1982)). One of the topics addressed rather extensively was the issue of tree-representability of syntactic objects. Lasnik & Kupin (1977) had appeared a few years earlier and Edwin Williams had just published his formalization of Ross' (1967/1986) notion of across-the-board (ATB) rule application in Williams (1978). Here are some passages about the analysis of coordination, the Lasnik & Kupin formalism and related matters:

In a field which has seen an enormous explosion of work in a great variety of areas, coordination and what Chomsky calls "the whole mass of phenomena that go along with it" have attracted remarkably little attention, and furthermore Chomsky's idea's of which way to go have been followed up by only a few researchers. In the domain of coordination proper, the main exceptions are Goodall (1987), Muadz (1991) and Moltmann (1992). One example for the reluctance to touch this difficult topic is that Riny Huijbregts, who I am sure could have taken this idea further than most anyone (although, whether he would have published it is a different matter), dropped this line of inquiry because, as he put it, "nobody believes in this stuff anyway." Another telling example in this respect is Progovac (1998), a state-of-the-art article on coordination, in which multi-dimensional or non-tree-representable approaches are only briefly mentioned and quite summarily dismissed. Apart from the problem of the occasional asymmetry among conjuncts, which Moltmann addresses extensively and adequately, in my view, the basic ground for disapproval seems to be the second half of the following sentence:  

Chomsky points out in the quote above that formalisms which define syntactic objects that are not tree-representable are not ipso facto richer and that in fact they can be more restrictive in certain ways. As in the case of the standard tree-formalization, the real issue is how such a formalism is substantively constrained. The matter is reminiscent of the discussion about representing reanalysis in terms of three- or multi-dimensional structures, again an option readily available under a Lasnik & Kupin type formalization. When we suggested to analyze verb clusters in the Continental West-Germanic languages in Haegeman & Van Riemsdijk (1986) in terms of reanalysis instead of using Evers' original proposal in terms of head adjunction, we were confronted with massive criticism, among others from Koster (1987): using multi-dimensional structures amounts to increasing the richness of the formalism. In this particular case, however, it is easy to see that reanalysis is in actual fact more restrictive in one crucial sense: reanalysis presupposes adjacency, while adjacency must be stipulated under any kind of head raising approach, cf. Van Riemsdijk (1998a).  

It is true, of course, that in most cases alternative analyses can be constructed in terms of the standard formalization, and often they are. The Progovac article cited above, for example, implicitly concludes that a tree-representable analysis of coordination must be correct. Similarly, Wilder (1997. 1998) argues in favor of analyses of ellipsis under coordination and transparent free relative clauses in terms of tree-representable structures in combination with backward deletion processes. What is most remarkable in the few works dealing with the matter is that the fundamental choice of the formalism in question is generally hardly argued at all. To cite just one typical example: "A third approach not considered here utilizes multi-dimensional phrase-structures […]" (Wilder (1997: note 2 p. 103).

I will not pretend to make up for this lack of argumentation here. My modest purpose in the present contribution is to present an overview of constructions which are potential candidates for an analysis in terms of multi-dimensional or non-tree-representable structures. And the main reason for doing that is that, in my opinion, the range of constructions that avail themselves of these extended formal means is far broader than has hitherto been suspected. The relevant constructions, and I will deal with so-called transparent free relatives in some detail, will be called grafts in the sense that one tree structure, the scion, is grafted onto another tree, the stock. Thereby a structure is created in which two (or more) tree structures share a portion of terminal and non-terminal material.

2 The logical possibilities

There are two main ways in which the standard tree-representation formalism has been thought to be insufficient. For one thing, a single terminal string has sometimes been assumed to be simultaneously structured by two or even more trees. (In what follows, whenever I say two, what I really mean is two or more.) This is the case of reanalysis. Take the classical case of verb-preposition combinations which allow (pseudo-)passives.

The reanalysis proposal, going back to Chomsky's (1974) Amherst Lectures, is to say that (2) has two structures associated with it, one a VP with a V and a PP, the other a VP with a V and an NP. To represent this linearly, two different types of brackets can be used, […] for one dimension and {…} for the other dimension, resulting in the following abbreviated structure.

The second major type of non-tree-representable structure is that of two strings (or substrings) associated with a single tree structure. This is the classic case of fully parallel coordinated structures. On such a view, a sentence like (4) would have a structure like (5).

Structure (5) would then be linearized to (4) Also, omitting the adverb respectively, (5) could be linearized to (6).

This, then, would be one way of dealing with gapping.

Once we turn to constructions like Right Node Raising, things become more complicated. In the RNR construction, we have a shared part of the string and a non-shared part. While the shared parts may be thought to not only share the terminals but also the relevant part of the structure, the non-shared part can diverge greatly in structure, as is know from such examples as (7).

While the common part of the string is a direct object in both conjuncts, the strings that each end in this direct object are greatly different in their own structure. The fact that the shared string does not have to be a constituent is by now well-established, witness examples of the following type from German.

Structures such as (7) or (8) could be characterized as two full-fledged trees, each with its own string of terminals and its own tree structure, which stand in a conjoined relationship with one another and which share the rightmost parts of their strings (with associated structure). This, then, leaves us with the following overall classification of syntactic representations.
 

3 Grafts

3.1. Right Node Raising

In the case of Right Node Raising (RNR), it is well-known that while the structure is coordinated, the two conjuncts need not be identical in structure. The only parallelism requirement is that the common part of the string (which need not be a constituent, cf. i.a. Wilder (1997)) be a structurally and semantically well-formed continuation of each of the conjuncts. Consider, for example, (9).

The common part of the string is in italics while the non-identical parts of the strings are indicated as […}. The structure of the latter is clearly quite different. This is what suggested the metaphor of 'graft' to me: two tree structures with some similarity but by no means identical in structure which are joined to form a single organism. Another metaphor that comes to mind is that of siamese twins. Siamese (or conjoined) twins can be symmetric or asymmetric. In the latter case, one of the two is called the host and the other is the parasite. But then, the same asymmetry is found in grafts, where we distinguish the stock and the scion. And furthermore, grafting can actually be applied to trees, which is why I prefer this terminology.

I should emphasize again, that the notion of grafting is rejected by some researchers. Wilder (1997) proposes a backward deletion solution, not in the least because he wants to preserve tree-representability. He also suggests that backward deletion is intimately linked to another property of RNR which is well-established, viz. the fact that the deleted string (or the common part in a graft) is right peripheral. This is not, however, a logical connection: deletion, just like grafting, can in principle be initial, medial and final. What I will argue in the remainder of this paper is that the number of constructions that lend themselves to an analysis in terms of grafting (or backward deletion) is much larger than has hitherto been assumed and that indeed left peripheral, medial and right peripheral grafts exist. The terminology is somewhat tricky here. I will refer to RNR as 'lefthand grafting' because if the shared part of the string is on the right, the graft is on the left. The opposite case, in which the shared part is on the left I will call righthand grafting for the same reason. The remaining cases, in which the shared part is somewhere in the middle I will call medial or 'saddle' grafting. I will first briefly discuss a number of cases which receive more extensive treatment elsewhere in the literature and then turn to transparent free relatives at greater length.

3.2. Wh-prefixes, matching free relatives, and Dutch/German parasitic gap constructions

A first case of grafting, lefthand grafting as it happens, is found in certain constructions which I have referred to as (syntactic) wh-prefixes. These are also discussed in Lakoff (1974), McCawley (1988) and Wilder (1998) as 'sluice parentheticals'. The following exemplify this construction:

Particularly revealing is a case which I have studied in some detail in Van Riemsdijk (to appear). The Swiss German form wäisch 'do you know' can be used as a regular main clause introducing an embedded wh-question, as shown in (12a). That this is the correct analysis for (12a) is shown by the fact that the second clause actually has the verb in the final position, hence it is an embedded clause. However, with this particular form (as well as a number of others, depending on various dialects etc.), main clause word order in the second clause is also possible, as shown in (12b). (13) illustrates the fact that we have a case of a sluice which is, in a sense, prefixed, grafted, onto the wh-word. In this case there is a special meaning optionally associated with examples like (12b), a reading which could be described as a 'rhetorical exclamative.'

Turning to a second case now, consider again the long debate in the late 70s and early 80s about the position of the relative pronoun in free relatives (cf. Bresnan & Grimshaw (1978) and Groos & Van Riemsdijk (1981) along with a large number of later publications). Two main facts were established. First, the relative pronoun must be assumed to be in the Spec,CP position of the relative clause because when the relative clause is extraposed, as is possible in Dutch and German, the relative pronoun moves along with the rest of the clause. Second, the relative pronoun must nevertheless satisfy contextual requirements (selection, case) in the matrix clause as well. This was called the matching phenomenon and is exemplified by the following type of examples.

In view of our current discussion about shared strings and grafting, we may hypothesize that free relatives of the matching kind are grafts in which the relative pronoun is directly dominated by whatever its category is in the matrix clause while it is dominated by Spec,CP in the graft. In other words, the question as to whether the relative pronoun is in Spec,CP position or in the position of the head of the relative clause may well be fundamentally misguided and imposed by the wrong kind of theory about syntactic representations, since in a multidimensional theory such as the one under discussion, the relative pronoun could be simultaneously in both positions. The syncretism found in examples like (15) is then interpretable as the kind of purely phonological identity requirement familiar from RNR as in (16), which is from Wilder (1997: (122)). Cf. also Van Riemsdijk (1998b) for discussion of further cases of morphophonological effects in free relatives.

The final case I would like to briefly discuss in the present section is that of parasitic gaps of the type found in Dutch and German. The main difference, as has been noted many times, is that parasitic gaps in these languages need not be licensed by some A-bar binder. Instead, it appears as if in situ direct objects can license parasitic gaps. This is shown in (17).

One way of dealing with this is to assume that the direct object has been scrambled across the adjunct, as in (18).

However, it is clear that the chain created by scrambling must have properties of an A-chain, in view of the absence of any weak crossover effects in examples like the following.

In Huijbregts & Van Riemsdijk (1985), an alternative is explored and argued for in terms of what we then suggested might be called Left Node Raising, where, as in the case of RNR, there is no raising at all but just some constituent in situ which is shared by two strings, the matrix and the adjunct clause. The main arguments, which I will not review here, had to do with (a) phonological identity effects and (b) a leftward peripherality condition (the shared constituent has to be left peripheral in each of the separate parts of the string, a fact entirely unexplained on a scrambling analysis.

This concludes my brief overview of three plausible candidates for grafts. In the next section I turn to a more detailed examination of transparent free relatives and will argue that they instantiate another construction that is profitably treated in terms of grafts.

4. Transparent Free Relatives

4.1. Main properties

The most complete and interesting description and analysis of transparent free relatives, to my knowledge, is found in Wilder (1998). The following summary follows his article in its essentials, even though I will suggest a number of alternative interpretations below.

In the normal case, free relatives come in two semantic varieties: with a definite reading or with a universally quantified reading, as in the following examples. This corresponds to what Grosu & Landman (1998) call 'third type relatives.'

It is a characteristic of Transparent Free Relatives (TFRs) that they can be and often are indefinite. The following are typical examples of TFRs.

(21a) means that I ate a steak (or something that they referred to as a steak). Similarly, (21b) means that a meteorite has fallen on the roof, and (21d) that he drives a flashy car. The adjectives in (21c/d') cannot be literally indefinite, of course, but they are similar in that they introduce a new quality, not one already given in the discourse.

In addition to this, Wilder (1998) gives the following further properties as characteristic of TFRs.

1. In accordance with their potential indefiniteness, TFRs can appear in positions in which only indefinites can occur – FRs cannot:

2. When used as a subject, the bare what of a TFR can trigger plural agreement on the verb – in FRs it cannot:

3. In TFRs, what can refer to humans, which it cannot do in FRs:

4. -ever cannot be attached to what in TFRs, in FRs it can:

5. TFRs do not induce so-called island effects (constraints on the structural distance between a moved element and the place it came from), while FRs, like ordinary relative clauses, do:

As a test of this, notice that the definite/indefinite ambiguity, i.e. the ambiguity between an FR-reading and a TFR-reading tends to disappear under apparent island violations:

6. The TFR must contain a predicate nominal or adjective which could/can replace the relative clause construction in the matrix clause – a FR is not subject to such a constraint:

Among these, the first, fourth and sixth properties might be called definitional, but the others are quite surprising and will provide crucial clues as to the analysis of these constructions. Another way of looking at TFRs is, indeed, in terms of 'shared constituents.' Where in regular FRs of the matching type the wh-word itself may well be a shared constituent, in TFRs it seems as if the predicate constituent is shared between the matrix clause and the free relative. And this is what makes the construction interesting in the present context of grafts or multidimensional representations.

4.2. Backward deletion vs. grafts

As already pointed out above, alternative analyses which might preserve tree-representability can be and have been proposed. In the case of TFRs, Wilder (1998) has suggested that these, along with RNR-constructions, should be handled in terms of backward deletion. Under such an approach, a TFR like (32a) would be analyzed as (32b).

Wilder offers two arguments for this proposal:

A. String identity

There are indeed cases with TFRs in which phonological identity seems to be a stronger condition than full categorial identity. Observe the following contrast.

It is the ing-form which can show up as the small clause predicate (the category labels are Wilder's):

Note now that (33a) we have a category clash, and still the sentence is (reasonably) good, while in (33b), where there is fully categorial but no morphophonological identity is bad.

This argument is neutral because it suggests that string identity is crucial and not categorial/syntactic identity, but this is exactly what the graft approach says. A property (phonological string identity) is attributed to backward deletion which is not independently established. Furthermore, this property does not seem to be applicable to Right Node Raising (RNR) which Wilder takes to be the ‘classical’ case of backward deletion.

B. The right edge property

It is well-known that RNR is subject to the right edge property, in other words that the shared part of the two conjuncts must be right-peripheral in each of the two conjuncts. Wilder furthermore assumes that TFRs are subject to the same right edge property. This is also a non-argument. If the right edge property is a real property, then again right peripherality is simply a property stipulated for backward deletion. There is no reason why it could not be stipulated as a property of string sharing in a theory of grafts. But I will argue below that the right edge phenomenon is spurious and that internal string sharing (saddle grafting) does occur with TFRs, not only in German and Dutch but also in English. Note also that the right edge property plays a crucial role in an incremental left-to-right structure building theory such as the one proposed in Phillips (1996, 1998). The arguments for internal (or 'saddle') grafts to be presented below argue against such an approach as well.

The alternative is the ‘graft’ approach, which follows (roughly) Lakoff (1974), Goodall (1987), McCawley (1988), Muadz (1991), and Moltmann (1992), and in which the constituent is truly shared by two tree structures and two strings. The consequences are far-reaching and cannot be developed here beyond some initial suggestions and, in particular, the establishment of the fact that there may very well be grafts of every positional type:

• lefthand grafts (i.e. shared string on the right): RNR;
• righthand grafts (i.e. shared string on the left): matching FRs, Dutch and German parasitic gap constructions, some TFRs;
• 'saddle' grafts (i.e. shared string in the middle): some TFRs and possibly (see below) internally headed relative clauses.

Let us now turn to some evidence that saddle grafts of the TFR variety exist.

4.3. Internal Transparent Free Relatives: "Saddle" Grafts

It is true, of course, that the predicate XP (a DP or an AP in the cases considered so far) is at the right edge of the free relative it is in. In order to check the extent to which this is an essential property, the obvious next move is to check whether SOV languages such as Dutch or German, in which the predicate XP precedes the verb, have TFRs. Wilder (1998) does precisely this and concludes, erroneously as I will argue, that German does not have any TFRs, while Dutch does, but only by virtue of extraposing the predicate XP into the right peripheral position. Here are the examples Wilder gives (his (46-49), with his grammaticality judgments). 

Note first that (37b) purports to show that predicate XPs cannot be extraposed. Actually, in my own judgment extraposed predicate XPs are not that bad (for me: '?') when, as in (37b), they are introduced by a preposition-like element such as als ('as'), für ('for'), etc. The corresponding example in Dutch (39a) is fully grammatical though examples lacking a predicate introducer such as als ('as') are entirely ungrammatical. The question then is first whether (37a) can be interpreted as a TFR, and second whether either of the examples in (38), which would have to be TFRs, are grammatical.

Starting with the latter question, the initial evidence appears to be conflicting. The Dutch example in (39b) is fully grammatical. In German, the ungrammaticality of (38a), with the predicate XP not extraposed, is unsurprising given the general requirement that prenominal adjectives be adjacent to the head noun. I personally find (38b) better than (38a), though still about '??' or '?*'. This would be more or less in line with my judgment on extraposed predicate XPs introduced by a preposition. Once we take predicate XPs which are not introduced by prepositions into consideration, things become more straightforward. Take the following examples from Dutch:

What these examples show is that the bare AP cannot extrapose in the case of a normal FR, but that it can in the case of a TFR. This is first of all an indication that the right peripheral AP in (40a) is possible only because it is shared with the matrix clause and is not purely inside the FR, as it is in the case of (41a). And this is confirmed by the corresponding examples where the shared adjective is prenominal:

The prenominal adjective has to inflect (-e), which it never does in predicative position, as (42b) shows. In order to be able to inflect at all, the adjective has to be adjacent to the noun, as in (42a). If the adjective remains in situ, which (40b) shows is normally possible, the sentence is out, whether the adjective is inflected or not, witness (42c). As Wilder correctly observes, the fact that the adjective inflects shows that it is a shared constituent. What Wilder fails to point out, however, is that if the (T)FR is a true and full free relative from which a right peripheral adjective is deleted, we have two serious problems. First, how does the adjective-to-be-deleted get into the final position inside the free relative in the first place? And second, how come a free relative is tolerated at all in what is presumably something like a premodifier position of the attributive adjective?

Turning now to the interpretation of (37a), I disagree with Wilder. I believe these examples can be shown to be TFRs by several of the criteria listed in 4.1. Criteria 2/3/4 are not applicable in German or Dutch for independent reasons, and 6 is satisfied by definition. Consider, therefore, criteria 1 and 5.

Criterion 1 says that, unlike a definite FR, TFRs can occur in positions in which only indefinites can occur. Consider  now an example like the following, in which an indefinite predicate DP (or NP) internal to the free relative apparently makes the whole free relative indefinite:

Observe that this does not mean 'I ate the thing that they were talking about and which they called a steak' but rather 'I ate an alleged steak.' Semantically, then, this is a TFR. Consider then a there-insertion type context.

By the indefiniteness criterion, then, these free relatives with an internal predicate XP are TFRs.

Let us now turn to criterion 5, the extractability of elements from the predicate XP. At first glance, the evidence seems to be somewhat conflicting in this case. In a German example like (45), a PP as a whole is apparently extracted from the predicate DP in the free relative. The same extraction from a regular FR yields a sharply ungrammatical result, as (45b) shows.

A Dutch example with extraction of waar from the predicate DP, exemplified in (46a) is ungrammatical, however. This seems to go against my argument that internal TFRs exist. There are reasons to assume, however, that even in the simple case, waar-extraction from the PP is only possible if the PP has been right-extracted out of (or detached from) the DP in the first place, cf. Van Riemsdijk (1997) for detailed argumentation. The question then is, of course, where the PP extraposes to in a case like this. The general fact is that the PP moves rightward but to a position preceding the verb cluster at the end of the clause. This is why the extraposition is usually string-vacuous. Suppose now that the apparently embedded predicate DP is in fact shared with the matrix. Then it would not be surprising to find that the extraposition is across the remainder of the TFR to the position preceding the final verb of the matrix clause. (46) shows that this is exactly what happens. Instead of being conflicting, the evidence is actually very much in favor of my position, which is that the internal predicate XP is a shared constituent and hence that these are true TFRs, which in turn argues in favor of the existence of saddle grafts.

Returning now from this excursus on German and Dutch to the issue of the right edge phenomenon in English, consider again the statement that in English the predicate XP is right peripheral in its free relative. Actually, this turns out not to be true either. The following examples show that a predicate XP can be followed by an experiencer PP. This is true in the normal case, as shown in (47), but it is equally true in (48), which exemplifies the case of TFRs.

The fact that (49b) is ungrammatical is again not attributable to the alleged right edge property of TFRs, but to the usual adjacency condition on prenominal adjectives which also excludes the ungrammatical examples in (50).

I conclude, then, that TFRs are constitute a prime candidate for medial and lefthand grafts.

4.4. Internally Headed Relative Clauses

Having gone this far, let me extend these ideas yet a little bit further. It is a well-known though ill-understood fact that some languages, including Japanese, Lakhota, Quechua, and Mojave, have relative clauses with so-called internal heads. (51) exemplifies such constructions as found in Japanese.

These have constituted a considerable problem for pretty much any theory about the structure of relative clauses. Cf. Grosu (to appear) and references cited there for a recent discussion. To the extent that TFRs have a predicate XP which is shared with the matrix, we might say that that XP is in a sense more of a head, semantically and syntactically, than the spurious empty head if there is one in the first place. And this conclusion immediately leads us to wonder if the internal head of internally-headed relative clauses could not be considered a shared constituent in the same way. Such internally-headed relative clause constructions would then constitute yet another source of evidence for saddle grafts. I will not pursue this idea, which is preferably studied by experts on the languages in question, any further here.

Before leaving this topic, however, let me briefly address the question of the extent to which the properties of these constructions (TFRs, internally headed relative clauses) must be stipulated and to what extent they can be derived from more general principles. Beyond the rather facile statement that as much as possible should be derived, it may nevertheless not be entirely wrong to assume that some stipulative elements are involved as well. Clearly, not just any constituent can be a shared element in English, German and Dutch TFRs. The fact that it is the predicative XP which lends itself particularly well to this status follows, I would venture, from some more general property. Nevertheless, the fact that English, German and Dutch are so restricted, while languages with internally headed relative clauses such as Japanese, Quechua, or Lakhota are not or at least not to this extent, may not be attributable to anything but a choice made at the level of the individual language. Assuming this to be so, where would we expect such a stipulation to be located in the grammar? Recall that some of the examples of grafts discussed above have a strong lexical (or idiomatic) flavor to them. This is certainly true of wh-prefixes like God knows what, who knows when, etc. as well as the case of wäisch in Swiss German. My guess, at this point, would indeed be that individual languages have a lexical subcomponent in which language specific conditions on grafting in that language are laid down.

5. Concluding Remarks

The quest for a restrictive theory of syntax is a noble one. And a theory of syntactic representations which admits only tree-representable structures may well be such a restrictive theory, though the reader is reminded of Chomsky's cautious words in this regard in the quote given at the outset of this article. It has been my purpose here to review the scattered and often bashfully hidden evidence in favor of a theory which admits (certain but not all) non-tree-representable structures and to add new materials to the evidence which has been accumulating at least since Lakoff (1974). I believe that a formal theory of syntactic representations along the lines of the one developed in Moltmann (1992) is probably best suited to deal with the whole variety of grafts as they have been discussed here and will undoubtedly yet come to light. That such a theory will then be in dire need of substantive restrictions goes without saying. We may lose the tree as we knew it, but grafting opens up new dimensions and exciting ranges of syntactic creatures. If biology is unrestrictive enough to permit certain (but not all) botanical organisms to enter into a grafting relationship, why shouldn't we expect the same thing to be true in the case of syntactic organisms.

6. References

Bresnan, J.W. & J. Grimshaw (1978) 'The syntax of free relatives in English' Linguistic Inquiry 9.3.

Chomsky, N. (1955/1975) The Logical Structure of Linguistic Theory. Ph.D. Dissertation, Harvard. Published in 1975 by Plenum.

Chomsky, N. (1974) The Amherst Lectures. Distributed by Documents linguistiques, UniversitJ de Paris VII.

Chomsky, N. (1982) The Generative Enterprise -- A discussion with Riny Huijbregts and Henk van Riemsdijk. Dordrecht: Foris.

Goodall, G. (1987) Parallel Structures in Syntax. Cambridge: CUP.

Groos, A. & H.C. van Riemsdijk (1981) 'Marching effects in free relatives' in A. Belletti et al. (eds.) Theory of Markedness in Generative Grammar. Pisa: Scuola Normale Superiore.

Grosu, A. (to appear) 'The semantic diversity of internally-headed relative clauses.'

Haegeman, L. & H.C. van Riemsdijk (1986) 'Verb projection raising, scope, and the typology of rules affecting verbs' Linguistic Inquiry 17, 417-466.

Huijbregts, M.A.C. & H.C. van Riemsdijk (1985) ‘Parasitic gaps and ATB’ in Proceedings of the NELS XV Conference, pp 168-187. GSLA, University of Massachusetts, Amherst.

Koster, J. (1987) Domains and Dynasties. Dordrecht: Foris.

Lakoff, G. (1974) 'Syntactic amalgams' Chicago Linguistic Society (CLS) 10 Proceedings, pp. 321-344.

Lasnik, H. & J. Kupin (1977) 'A restrictive theory of transformational grammar' Theoretical Linguistics 4.3.

McCawley, J.D. (1988) The Major Syntactic Phenomena of English. Chicago: Chicago University Press.

Moltmann, F. (1992) Coordination and Comparatives. Ph.D.Dissertation, MIT. Distributed by MIT Working Papers in Linguistics.

Muadz, H. (1991) Coordinate Structures: A Planar Representation. Ph.D. Dissertation, University of Arizona.

Phillips, C. (1996) Order and Structure, Ph. D. Dissertation, MIT. Distributed by MITWPL.

Phillips, C. (1998) 'Linear order and constituency,' unpublished manuscript, University of Delaware.

Progovac, L. (1998) 'Structure for coordination,' Parts I and II. GLOT International 3:7, pp 3-6, and 3:8, pp 3-9.

Riemsdijk, H.C. van (1997) 'Push chains, and drag chains: complex predicate split in Dutch' in S. Tonoike (ed.) Scrambling. Tokyo: Kurosio Publishers.

Riemsdijk, H.C. van (1998a) 'Head movement and adjacency' Natural Language and Linguistic Theory 16, 633-678.

Riemsdijk, H.C. van (1998b) 'Syntax driven (crazy) by morphology' in A. Bruyn & J. Arends (eds.) Mengelwerk voor Muysken. Publicaties van het Instituut voor Algemene Taalwetenschap #72. University of Amsterdam.

Riemsdijk, H.C. van (to appear) 'Wh-prefixes: the case of w@ isch in Swiss German.'

Ross, J.R. (1967) Constraints on Variables in Syntax. Ph.D. Dissertation, MIT. Published as Ross (1986).

Ross, J.R. (1986) Infinite Syntax! Norwood, NJ: Ablex.

Wilder, C. (1997) 'Some properties of ellipsis in coordination' in A. Alexiadou & T.A.Hall (eds.) Studies on Universal Grammar and Typological Variation. Amsterdam: John Benjamins.

Wilder, C. (1998) 'Transparent free relatives' ZAS (Berlin) Working Papers in Linguistics pp. 191-199.

Williams, E.S. (1978) 'Across-the-board rule application' Linguistic Inquiry 9, 31-43.

Note

I would like to thank the Collegium Budapest for offering me its hospitality, which was conducive to writing the present essay, my department and colleagues at Tilburg University for making this research leave possible, Misi Brody, László Kálmán, Edith Moravcsik and Kati Kiss for helpful discussion, Eörs Szathmáry for being a source of inspiration for the biological metaphor of grafting, Michael Gervers for helping with some of the data, and Tamás Szalay for helping out with a bunch of HTML tricks.

Author's address:

until February 1 1999:

Collegium Budapest; Szentháromság u. 2; H-1014-Budapest; Hungary
tel. +36 1 4577614/4577600; fax. +36 1 4577610; e-mail: riemsdyk@colbud.hu

after February 1 1999:

Tilburg University; Department of Language and Literature -- Grammar Models Group (CLS);
P.O. Box 90153; NL - 5000 LE - Tilburg, The Netherlands
tel. +31 13 4662642/4662773; fax. +36 1 4663110; e-mail: riemsdyk@kub.nl