[Editor's note: Due to the font limitations of HTML, a barred g will be used to represent a g with a breve diacritic, and an underlined s will be used to represent an s with a subscripted right half-ring.]

1 Introduction

Hankamer (1972) presents a number of facts from Turkish coordination constructions, with the aim of showing that forward and backward Gapping (or, to use a more neutral term, Coordination Deletion [CD]) have different (surface) structures. (In this squib, I limit attention to coordinations with two conjuncts and identical constituents at the right-hand periphery of respective conjuncts.) Forward CD genuinely "gaps" into the right conjunct, leaving a gap in the position of the identical element; backward CD, in contrast, erases both instances of the identical element and adjoins a copy to the top IP (or CP). A Right-Node-Raised structure results, and not a backward Gapped one. As a matter of fact, the larger claim of Hankamer (1972) is that UG has no (syntactic) mirror-image rules; therefore, there cannot be backward Gapping, strictly speaking.

In this squib, I present what I think is the factually most clear-cut and conceptually most convincing type of data from Turkish which is discussed in illustration of the hypothesis in Hankamer (1972) as well as Hankamer (1971) that apparent backward Gapping is Right Node Raising [RNR]. After showing some problems in the way in which these data are claimed to argue in favor of the "RNR Hypothesis" (and thus against backward Gapping), I suggest that Hankamer's claim that forward CD results in a different structure than backward CD is correct nevertheless. I base my view on additional Turkish facts.

2 The original facts and the claim

Turkish is a verb-final language (insofar as its basic word order is concerned). Given that I shall limit attention to right-peripheral constituents in conjuncts, I shall devote attention to identical verbs only.

In contrast to some other verb-final languages like Japanese, Turkish allows both forward and backward CD, as shown in the following two examples. Hankamer's claim forces different phrase-structural analyses for these examples (the positions of the omitted identical verbs are indicated by underlined spaces; the English translations are intended to reflect the properties of the Turkish examples as closely as possible and are not marked for their grammaticality):

(1)	[[Hasan	karides-i	ye-di],	[Mehmet	te	istiridye-yi	__ ]]
	Hasan	shrimp-ACC	eat-PAST	Mehmet	and	oyster-ACC
	'Hasan ate the shrimp, and Mehmet (ate) the oyster.'

(2)	[[Hasan	karides-i	__ ],	[Mehmet	te	istiridye-yi	ye-di]]
	Hasan	shrimp-ACC		Mehmet	and	oyster-ACC	eat-PAST
	'Hasan (ate) the shrimp, and Mehmet ate the oyster.'

According to Hankamer (1971) and (1972), the structure of (1) would be a genuinely gapped structure, essentially as indicated above. The structure of (2), however, would be different and as indicated in (3):

(3)	[[Hasan karides-i __ ], [Mehmet te istiridye -yi __ ]] ye-di

Hankamer makes the important observation that, in contrast to matrix coordination, where CD appears to be able to apply in both directions, only (apparent) backward CD is possible in (nominalized) embedded coordinate structures. Hankamer further claims that this root/embedded asymmetry follows from the different structures he posits for the two directions of "deletion". A brief discussion of the facts and their relevance for this posited asymmetry follows. The most general type of sentential complementation in Turkish involves "nominalized" embedded clauses, quite similar to English gerundives. There are two basic types of such "nominalizations" for argument clauses, and they are morphologically distinguished by exhibiting different "nominalization" morphemes on the predicate:

(4)	Ben	[Ali-nin	su-yu	iç-tig-in-i]	duy-du-m
	I	Ali-GEN	water-ACC	drink-FNOM-3.SG-ACC	hear-PAST-1.SG
	'I heard that Ali drank the water'

(5)	Ben	[Ali-nin	su-yu	iç-me-sin-i]	iste-di-m
	I	Ali-GEN	water-ACC	drink-ANOM-3.SG-ACC	want-PAST-1.SG
	'I wanted that Ali (should) drink the water'

I follow Lees (1968) in referring to the "nominalization" morpheme -DIK as "Factive Nominal" [FNOM], and to the morpheme -mA as "Action Nominal" [ANOM], because in many (although not all) instances, these terms do reflect the semantics of the thus marked predicates. What's important here is that, as illustrated in (4) and (5), the choice of the appropriate form is due to the matrix predicate. Hankamer assumes (as does Lees) that the embedded predicates owe their respective shapes to a syntactic transformation. Given the dependence of the particular nominalization on the matrix verb, this transformation can take place only after the embedded cycle in the derivation. Hankamer further assumes that this transformation is subject to Ross's island constraints, and, in particular, to the Coordinate Structure Constraint [CSC]. The relevance of this for the posited asymmetry is explained as follows: If the nominalization transformation were to apply to a gapped structure, with a deleted verb in one conjunct, it would violate the CSC: by nominalizing the only surviving verb, it would be applying to only one conjunct and not the other. The prediction would be that gapped (nominalized) sentences cannot be embedded in Turkish. As just mentioned, this is indeed true for forward Gapping:

(6)	*	Zeynep	[[Hasan-1n	karides-i	ye-dig-in-i],
		Zeynep	Hasan-GEN	shrimp-ACC	eat-FNOM-3.SG-ACC

		[Mehmed-in	de	istiridye-yi	__ ]]	duy-du
		Mehmet-GEN	and	oyster-ACC		hear-PAST

'Zeynep heard that Hasan ate the shrimp, and Mehmet (ate) the oyster.'

If backward CD constructions had a corresponding "gapped" structure, they should also result in ungrammaticality when embedded; however, such constructions are perfectly grammatical:

(7)	Zeynep	[[Hasan-1n	karides-i	__ ],
	Zeynep	Hasan-GEN	shrimp-ACC

		[Mehmed-in	de	istiridye-yi	ye-dig-in-i]]	duy-du
		Mehmet-GEN	and	oyster-ACC	eat-FNOM-3.SG-ACC	hear-PAST

'Zeynep heard that Hasan (ate) the shrimp, and Mehmet ate the oyster.'

Because of this contrast, Hankamer concludes that at the point of the derivation when Nominalization applies to a structure underlying (7), the embedded verb is no longer part of the coordinate structure. Rather, this verb has been raised and adjoined to the right of the coordinate structure:

(8)	Zeynep [[[Hasan-1n karides-i __ ], [Mehmed-in de istiridye-yi __ ]] ye-dig-in-i] duy-du

In such instances, therefore, Nominalization does not violate the CSC, thus making the grammaticality of such examples an expected fact. The ill-formedness of (6) and the well-formedness of (7) are thus predicted and are taken to constitute evidence for a distinction between forward gapped versus right-node-raised output structures resulting from conjoined sentences with identical verbs.

3 Some problematic facts

Almost three decades have gone by since Hankamer's work. There has been a general change towards lexical analyses (and away from transformational ones) in theoretical syntax; today, probably very few syntacticians would be inclined to derive the "nominalized" predicates of embedded clauses in Turkish via a syntactic, island-sensitive transformation. The ill-formedness of embedded backward CD, in view of the well-formedness of matrix backward CD would therefore still need an explanation.

Furthermore, even in a theoretical framework that would recognize nominalization as an island-sensitive transformation, there was reason to suspect the proposal that the contrast between (6) and (7) was due to different input structures to nominalization. I shall disregard conceptual reservations and will concentrate on problematic empirical facts.

3.1 Non-nominalized embedded coordinate structures

While the vast majority of verbs in Turkish require, as stated above, complements with nominalized predicates, there are a few verbs which can optionally admit complements with fully finite predicates. These verbs are: san, zannet, bil (all translate, roughly, as 'believe'), which admit complement clauses that have predicates of all tenses, and iste 'want', which admits complement clauses that have predicates in the subjunctive. Thus, the following examples are fully grammatical where the complement clauses have predicates that are not nominalized and exhibit forms found otherwise in main clauses:

(9)	Ahmet	[Hasan	karides-i	ye-di]	san-1yor
	Ahmet	Hasan	shrimp-ACC	eat-PAST	believe-PRPROG
	'Ahmet believes Hasan ate the shrimp'

(10)	Ahmet	[Hasan	karides-i	ye-sin]	isti-yor
	Ahmet	Hasan	shrimp-ACC	eat-SUBJUNCT	want-PRPROG
	'Ahmet wants (that) Hasan should eat the shrimp'

These fully finite, tensed complements can be coordinated, just as nominalized complements can be. Given that nominalization does not apply here, we might expect that omission of identical verbs behaves similarly to what we see in matrix clauses, i.e. that CD can apply both forward and backward.

This, however, is not the case. Embedded non-nominalized coordinate structures behave just like their nominalized conterparts. The following two examples show that "forward-gapped", non-nominalized embedded coordinations are ill-formed:

(11)	*	Ahmet	[[Hasan	karides-i	ye-di],
		Ahmet	Hasan	shrimp-ACC	eat-PAST

		[Mehmet	te	istiridye-yi	__ ]]	san-1yor
		Mehmet	and	oyster-ACC		believe-PRPROG

'Ahmet believes Hasan ate the shrimp and Mehmet (ate) the oyster'

(12)	*	Ahmet	[[Hasan	karides-i	ye-sin],
		Ahmet	Hasan	shrimp-ACC	eat-SUBJUNCT

		[Mehmet	te	istiridye-yi	__ ]]	isti-yor
		Mehmet	and	oyster-ACC		want-PRPROG

'Ahmet wants (that) Hasan should eat the shrimp and Mehmet (should eat) the oyster'

The next pair of examples show that (apparent?) backward CD in embedded non-nominalized coordiante structures is well-formed (I shall not indicate any structure beyond general embedding here, as I don't want to pre-judge the situation with respect to backward CD):

(13)	Ahmet	[[Hasan	karides-i	__ ],
	Ahmet	Hasan	shrimp-ACC

		[Mehmet	te	istiridye-yi	ye-di]]	san-1yor
		Mehmet	and	oyster-ACC	eat-PAST	believe-PRPROG

'Ahmet believes Hasan (ate) the shrimp and Mehmet ate the oyster'

(14)	Ahmet	[[Hasan	karides-i	__ ],
	Ahmet	Hasan	shrimp-ACC

		[Mehmet	te	istiridye-yi	ye-sin]]	isti-yor
		Mehmet	and	oyster-ACC	eat-SUBJUNCT	want-PRPROG

'Ahmet wants (that) Hasan (should eat) the shrimp and Mehmet should eat the oyster'

These contrasts challenge Hankamer's explanation for the corresponding contrasts in nominalized embedded coordiante constructions. It is possible, of course, that Hankamer is still correct about the explanation concerning the directionality of (apparent) CD in nominalized coordinations (and I think that he is), and that the facts we just saw concerning non-nominalized embedded coordinations have a different explanation. Nevertheless, it is clear that Hankamer's proposed structures for forward versus (apparent) backward CD do not explain the contrasts we just saw, and that it would be preferable to find a unified account for the similar directionality facts concerning nominalized and non-nominalized embedded coordinate structures.

3.2 Nominalized, forward-gapped coordinate structures

Turkish allows for non-verb-final sentences; post-verbal material is interpreted as presupposed, old information. Such scrambling is particularly felicitous when the post-verbal material consists of DPs. Given that nominalized embedded clauses are DPs (for example, they are overtly marked for Case), they can be scrambled to post-verbal positions; this holds for nominalized coordinate structures, as well.

Interestingly, such post-verbal nominalized coordinate structures can be "forward-gapped":

(15)	ZEYNEP	duy-du	[[Hasan-1n	karides-i	ye-dig-in-i],
	Zeynep	hear-PAST	Hasan-GEN	shrimp-ACC	eat-FNOM-3.SG-ACC

		[Mehmed-in	de	istiridye-yi	__ ]]
		Mehmet-GEN	and	oyster-ACC

'ZEYNEP heard that Hasan ate the shrimp, and Mehmet (ate) the oyster.'

While it is stylistically somewhat awkward to have very little pre-verbal material in the presence of heavy post-verbal material, focus on the pre-verbal material makes examples like (15) perfectly acceptable. In any event, the forward "Gapping" of the identical verb in the nominalized embedded coordinate structure does not lead to ill-formedness. What, then, is the difference between (15) and the ungrammatical (6)?

Obviously, the grammaticality of (15) poses a challenge to Hankamer's explanation of the ungrammaticality of (6). Why does nominalization of the verb in the left conjunct not lead to a violation of the CSC, given that nominalization does not apply to the right-hand conjunct? Within the syntactic model of Hankamer's work, there is a way out. One could say that (15) is not a scrambled version of (6), but rather a gapped version of a scrambled, full coordination of embedded nominalized clauses:

(16)	Zeynep	[[Hasan-1n	karides-i	ye-dig-in-i],
	Zeynep	Hasan-GEN	shrimp-ACC	eat-FNOM-3.SG-ACC

		[Mehmed-in	de	istiridye-yi	ye-dig-in-i]]	duy-du
		Mehmet-GEN	and	oyster-ACC	eat-FNOM-3.SG-ACC	hear-PAST

'Zeynep heard that Hasan ate the shrimp, and Mehmet ate the oyster.'

(17)	ZEYNEP	duy-du	[[Hasan-1n	karides-i	ye-dig-in-i],
	Zeynep	hear-PAST	Hasan-GEN	shrimp-ACC	eat-FNOM-3.SG-ACC

		[Mehmed-in	de	istiridye-yi	ye-dig-in-i]]
		Mehmet-GEN	and	oyster-ACC	eat-FNOM-3.SG-ACC

'ZEYNEP heard that Hasan ate the shrimp, and Mehmet ate the oyster.'

(16), then, would be the source, with the embedded coordination in its underlying position as the object of the matrix verb; the next step in the derivation would be (17), with the full, ungapped coordination being scrambled to a post-verbal position, and (15) would be the result of forward Gapping within the embedded coordination. Since Gapping applies after nominalization in this derivation, the CSC is not violated. (Gapping must, of course, be a kind of rule that is not sensitive to the CSC, and it would have to be non-cyclic.)

While this derivation solves the problem of the grammaticality of (15), it raises another question: Why is a similar derivation not possible for (6), i.e. a derivation where nominalization avoids violating the CSC? In other words, why can't forward Gapping apply to (16), without scrambling of the whole coordinate structure to post-verbal position applying afterwards? If the reason is that Gapping is cyclic, then the outlined derivation for (15) is impossible.

It is possible to find a way out from this dilemma within a framework that allows for crucial rule ordering. However, it is also clear that, just as with the previous problematic instance concerning non-nominalized embedded coordination, it would be much preferable to find a general and principled account for both the grammaticality of (17) and the ungrammaticality of (6). This would be all the more preferable nowadays, considering that we would like a solution which is independent of stipulations concerning rule ordering.

4 A possible solution

I would like to propose that there is a condition in Turkish syntax which precludes the generation of embedded clauses that are not verb-final and which are internal to a higher clause. This condition is probably perceptually motivated; the embedded verb is the perceptual clue for the right-hand boundary of an embedded clause—something which is important especially when there is material of a higher clause following, so that there is a clear-cut boundary between the embedded level and the higher one.

Such a condition would explain all the facts that we have seen so far: Internal nominalized and non-nominalized clauses would be treated alike, and forward gapping would be blocked (or its result thrown out) for both, given that the verb marking the rightmost periphery of the embedding, i.e. the verb of the second conjunct, would be missing for both types of embedding.

Furthermore, in examples like (17), the embedded clause is not internal to the higher clause; in other words, there is no material belonging to the higher clause that follows the embedded coordinate construction. Therefore, the condition just proposed is not enforced; there is no need for the embedding to be delimited by a rightmost verb. Consequently, forward gapping can apply in such instances.

The same condition would also be in force with respect to scrambling constituents to postverbal positions. Where an embedded clause, nominalized or not (and coordinated or not) is internal to a higher clause, no constituent of that embedding can scramble to the right of the embedded verb; however, such a constituent can, for most speakers, scramble to the right of the matrix verb. This is not surprising, given that the embedded verb is still the rightmost element of the embedded clause and can thus fulfill its role as a perceptual clue as the boundary between the embedded clause and the higher clause (the underlying position of the scrambled constituent is indicated by an underlined space, coindexed with that constituent):

(18)	*	Zeynep	[[Hasan-1n	__i	ye-dig-in-i]	karides-ii ]	duy-du
		Zeynep	Hasan-GEN		eat-FNOM-3.SG-ACC	shrimp-ACC	hear-PAST
		'Zeynep heard that Hasan ate the shrimp.'

(19)		ZEYNEP	duy-du	[Hasan-1n	__i	ye-dig-in-i]	karides-ii
		Zeynep	hear-PAST	Hasan-GEN		eat-FNOM-3.SG-ACC	shrimp-ACC
		'ZEYNEP heard that Hasan ate the shrimp.'

While the condition just proposed explains the contrast between these two examples, Hankamer's account of CD facts cannot be extended to capture these scrambling data. Of course it is not intended to; however, I would suggest that an account that can explain these particular scrambling data along with the previous facts about CD is preferable to one that does not do so.

Note that this solution is compatible with Hankamer's proposed structures, i.e. with a gapped structure for forward CD and a RNR-structure for backward CD. (It is, of course, also compatible with gapped structures in both directions.) As a matter of fact, Hankamer is probably correct about his structural analyses concerning all the problematic facts I discussed here. What I have shown is merely that the Turkish data used by Hankamer as evidence in favor of his claims do not, in fact, offer such support.

5 Additional facts to distinguish between forward versus backward CD

The observation has been made in the literature (most recently, by Bozsahin 2000) that in Turkish, coordination exhibiting forward CD does not need parallel conjuncts with respect to word order (again, I limit observation and discussion to V-final structures), while coordination exhibiting (apparent) backward CD does. (Since parallel versions for both directions were illustrated earlier, I offer examples with non-parallel examples.) Forward CD with SOV and OS:

(20)	a.		[[Hasan	karides-i	ye-di],	[istiridye-yi	de	Mehmet	__ ]]
			Hasan	shrimp-ACC	eat-PAST	oyster-ACC	and	Mehmet
			'Hasan ate the SHRIMP, and MEHMET (ate) the oyster.'

Forward CD with OSV and SO:

	b.		[[karides-i	Hasan	ye-di],	[Mehmet	te	istiridye-yi	__ ]]
			shrimp-ACC	Hasan	eat-PAST	Mehmet	and	oyster-ACC
			'HASAN ate the shrimp, and Mehmet (ate) the OYSTER.'

Backward CD with SO and OSV:

(21)	a.	*	Hasan	karides-i	__ ,	istiridye-yi	de	Mehmet	ye-di
			Hasan	shrimp-ACC		oyster-ACC	and	Mehmet	eat-PAST
			'Hasan (ate) the SHRIMP, and MEHMET ate the oyster.'

Backward CD with OS and SOV:

	b.	*	karides-i	Hasan	__ ,	Mehmet	te	istiridye-yi	ye-di
			shrimp-ACC	Hasan		Mehmet	and	oyster-ACC	eat-PAST
			'HASAN (ate) the shrimp, and Mehmet ate the OYSTER.'

This difference can be accounted for in a number of ways. One could say, for example, that forward and backward CD have different conditions of application, with the backward direction being stricter with respect to parallel word order in the conjuncts. This would be rather lacking in insight, however.

But if we adopted an approach like Hankamer's which posits structural differences for the two directions, and, more specifically, if we claimed, along with him, that forward CD is genuine Gapping, an operation that affects strings only, while backward CD involves RNR, an operation that involves structural hierarchy, we come closer to understanding the grammaticality differences just observed. We can now say that for the sequential operation, and especially one that deletes only one occurrence of identical constituents, all that matters is to find the identical verbs in the right periphery of the conjuncts and delete in a forward direction; word order in the remainder of the conjuncts is immaterial. the resulting structure is a coordination which is, by definition, not parallel, as one conjunct has a gap and the other doesn't. Therefore, the two conjuncts are allowed to be non-parallel in other respects, too.

In contrast, RNR, while affecting a structural change in adjoining the identical verb to the coordinate structure and deleting both occurrences of the identical verb in the right periphery of both conjuncts creates a coordination which is parallel with respect to having gaps in both conjuncts. Therefore, the remainder of the conjuncts must be similar, too, i.e. parallel.

Another contrast, also having to do with strict versus more relaxed identity in the two conjuncts, has to do with different inflection on the deleted verb(s). Here, too, CD is more permissive in the forward direction and strict in the backward direction. In the following examples, the first conjunct requires subject-verb agreement morphology for the second person singular, while the second conjunct requires corresponding morphology for the first person singular. Full coordination, without any deletion:

(22)		Karides-i	sen	ye-di-n,	istiridye-yi	de	ben	ye-di-m
		shrimp-ACC	you[sg.]	eat-PAST-2.SG	oyster-ACC	and	I	eat-PAST-1.SG
		'YOU ate the shrimp and I ate the oyster'

Forward CD:

(23)		Karides-i	sen	ye-di-n,	istiridye-yi	de	ben	__
		shrimp-ACC	you[sg.]	eat-PAST-2.SG	oyster-ACC	and	I
		'YOU ate the shrimp and I (ate) the oyster'

Backward CD:

(24)	*	Karides-i	sen	__ ,	istiridye-yi	de	ben	ye-di-m
		shrimp-ACC	you[sg.]		oyster-ACC	and	I	eat-PAST-1.SG
		'YOU (ate) the shrimp and I ate the oyster'

Similar facts obtain in embedded coordination, as well.

If forward and backward CD were the same operation, this difference would be unexpected. The fact that forward CD can apply to verbs whose stems are identical but whose agreement inflections don't have to be, while backward CD needs inflectional identity, as well, would have to be stipulated.

However, if RNR is formulated as a process which needs strict identity of both occurrences of the verbs it deletes and whose single copy it adjoins to the coordination, the difference between the two directions is explained. In other words, the single copy must be identical to both elements that are copied; otherwise, there is a clash of information for the copy. Forward CD, on the other hand, is in fact genuine Gapping, i.e. it operates on sequences and affects one occurrence of the "identical" verbs only. Therefore, it is sufficient for the deletion to recognize an identical stem.

References

Bozsahin, C. (2000) Directionality and the Lexicon: Evidence from Gapping. Unpublished ms.; Middle East Technical University, Ankara.

Hankamer, J. (1971) Constraints on Deletion in Syntax. Doctoral dissertation, Yale University, New Haven, CT.

Hankamer, J. (1972) "On the Nonexistence of Mirror Image Rules in Syntax"; in Syntax and Semantics, vol. 1; J. Kimball (ed.); New York and London: Seminar Press; 199–212.

Lees, R. B. (1968) The Grammar of English Nominalizations. The Hague: Mouton.