2.1. The productivity of Greek TCs

Previous studies of Greek TCs tend to state that these constructions have very little productivity and are severely constrained. For instance, according to Joseph (Reference Joseph1990: 107) ‘although sentences of this type are of rather low frequency (speakers generally opt for a paraphrase), they nonetheless do occur and can be elicited from speakers and heard spontaneously under the proper conditions.’ Even more strongly, Tsimpli (Reference Tsimpli, Alexiadou, Horrocks and Stavrou1999: fn.4) notes that ‘Greek does not have the equivalent of easy-to-please constructions. Predicates like efkolo are always impersonal in that the subject position cannot be occupied by a referential NP’. In a similar vein, Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020: 20–22) claim that only two adjectives can productively appear in TCs and have an impersonal counterpart (namely, efkolos ‘easy’ and dhiskolos ‘tough’).

To start with, the list of adjectives licensing TCs does not appear to be so limited – consider for instance a range of adjectives in (8) (cf. the similar examples in Tsimpli Reference Tsimpli, Alexiadou, Horrocks and Stavrou1999: 244).

However, it is true that only some of these naturally alternate with an impersonal option:Footnote ³

Furthermore, as noticed by Roussou (Reference Roussou, Stavrou and Terzi2005), the verb thelo in its non-volitional use allows a configuration very similar to TCs, with an impersonal counterpart, as shown in (10):

Moreover, there are indeed some constraints on what DPs can be subjects of TCs, but these do not seem to be as clear-cut as assumed by Tsimpli (Reference Tsimpli, Alexiadou, Horrocks and Stavrou1999): for instance, referential DPs (i.e., proper names and 1st/2nd person subjects) are not always banned (cf. Dimitrakopoulou Reference Dimitrakopoulou2012: 94–102), as shown by the acceptability of (11b, c), although previous literature has reported similar cases as ungrammatical (11a):Footnote ⁴

2.2. TCs are not purpose clauses

Tsimpli (Reference Tsimpli, Alexiadou, Horrocks and Stavrou1999) associates TCs with purpose clauses (PCs), noting the optional presence of the prepositional complementiser ja (‘for’) introducing the na-clause in TCs (as in 12a), which is usually found in PCs, like (12b, c).

However, some evidence suggests that TCs cannot in fact be reduced to PCs: in other words, the two options (with and without ja) represent two distinct configurations, with different properties. First, the addition of ja is not always possible in TCs (13a–c):

Secondly, na- and ja na- clauses cannot be coordinated (14), which again suggests they have a different status:

Thirdly, in genuine PCs the clitic can (more or less freely) alternate with a lexical DP, unlike in TCs where this is categorically excluded (cf. Section 3.1). This is why (12b, c) are grammatical even if there is no clitic resumption (as in 15a, b). This seems to be possible only if ja is present, as shown by (15c): in fact, this is the most revealing difference between TCs and PCs.

2.3. TCs are not Null Operator Constructions/Clitic Left Dislocation

In terms of the syntactic analysis, both Tsimpli (Reference Tsimpli, Alexiadou, Horrocks and Stavrou1999) and Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020) treat TCs as A’-dependencies: namely, a null A’-operator structure and a Clitic Left Dislocation (CLLD) structure, respectively. More precisely, Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020: 20 and passim) claim that all sentences in (16) involve topicalisation, in contrast with their impersonal non-dislocated counterpart in (17).

The three options in (16) differ in terms of agreement and case: in (16a) the DP is nominative and triggers agreement with the matrix predicate (+Agr), whereas in (16b) and (16c) the DP is respectively nominative and accusative but does not agree with the adjective (-Agr). In all three cases, the object clitic is obligatory and coreferential with the sentence-initial DP.

Crucially, Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020) offer an analysis of Greek TCs without delving into the deeper distinctions between the three alternatives in (16) nor explaining their optionality; instead, they propose that (16) are all instances of CLLD rather than raising/control. Their argument goes as follows: Greek CLLD is associated with a base-generated DP in an initial topic position and a resumptive clitic/verbal agreement marker in the corresponding argument slot. These requirements are indeed met by TCs, as the subject DP is a sentence topic obligatorily resumed by an object clitic. At the same time, though, the DP can also be an argument of the matrix predicate and trigger agreement with it (as in 16a). They claim that the possibility of agreement in TCs cannot be a consequence of raising or control since these configurations in Greek necessarily involve long distance agreement (LDA) (also referred to as backward control (BC)), even when the subject DP surfaces in the embedded clause (see the contrast in 18).

They then show how long distance agreement is impossible in TCs (19b), explaining the contrast with A-dependencies like (18) by proposing that adjectives block $ \unicode{x03D5} $ -feature agreement between the matrix and the embedded predicates – in Section 4 we will show that this is not a real problem for a control analysis.

Conversely, they suggest (Alexiadou & Anagnostopoulou Reference Alexiadou and Anagnostopoulou2020: 33) that TCs resemble English copy-raising (20a), analysed as null A’-operator configurations, as discussed by Landau (Reference Landau2011) (20b):

So, they propose that the DP (to arthro ‘the article’) is base-generated in the matrix Topic position, and that it is at the same time an argument of the matrix predicate as well as a sentence topic, while the object clitic is inserted in the embedded clause through CLLD.

On the one hand, we agree with both Tsimpli (Reference Tsimpli, Alexiadou, Horrocks and Stavrou1999) and Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020) that TCs cannot be an instance of (hyper)raising of the DP from the embedded object to the matrix subject position – see Section 3.2.

On the other hand, Alexiadou & Anagnostopoulou’s (Reference Alexiadou and Anagnostopoulou2020) reasoning and empirical coverage in support of an A’-dependency derivation for TCs do not seem to be fully correct in a number of respects. In particular, we argue that +Agr TCs cannot be subsumed under the same CLLD analysis as -Agr for at least three reasons (21a–c):

1. (21)

   1. a. There is no other case in Greek in which a topic left dislocated from an embedded clause triggers agreement with and receives case from the matrix verb/predicate;

   2. b. There are more restrictions on what DPs are allowed in +Agr TCs than in -Agr;

   3. c. +Agr TCs are bounded, whereas -Agr configurations are not.

As for (21a), CLLD chains can only fill one argument slot, and this must be the one associated with the position where the clitic is found – see a canonical instance of CLLD (22a) and compare with its non-dislocated version in (22b). Here, the topic ton Jorgo is clearly not an argument of the matrix verb and cannot agree with it nor can it receive a case from the matrix clause.

In Greek TCs like (16a), repeated here for convenience as (23a), conversely, the dependency at issue seems to hold between two A-positions (although it might of course be extended to an additional A’-position as is trivially the case for any other preverbal subject), namely: the object position of the embedded verb (to which the clitic attaches) and the subject position of the matrix verb (as shown by agreement between the nominative DP and the A). In fact, (23a) is the only case where the DP behaves like a genuine argument of the adjective. This is shown by the fact that, when substituting the na -clause with a nominal periphrasis, +Agr and nominative case are the only grammatical options (23b):

By a similar token, a CLLD account cannot explain why the DP receives nominative case in the matrix clause and triggers agreement with the matrix predicate. In fact, the expected outcome of CLLD from the impersonal option (17) is (16b), repeated below as (24):

Here, the case of the fronted constituent is preserved: the DP is accusative because the only argument slot it is associated with is the object of matho. Even if the DP had nominative case and did not match case on the resumptive clitic, we would still not expect agreement with the adjective under an A’-dependency analysis: this is what we get in (16c), which can be analysed as a Hanging Topic, repeated here as (25).

Thus, genuine TCs like (16a/23a) seem to be crucially different from other resumptive clitic configurations like CLLD in that the overt DP must have received case and triggered agreement in an A-position, resulting in a case mismatch between the DP and the clitic. This is reminiscent of Greek quasi-ECM constructions (as in 26b): they cannot be the product of applying CLLD to sentences like (26a). On the other hand, (26c), which preserves the case assigned to the DP by the embedded clause, can be considered to be CLLD to the embedded left periphery (Kotzoglou Reference Kotzoglou2002, Reference Kotzoglou2017, Kotzoglou & Papangeli Reference Kotzoglou, Papangeli, Davies and Dubinsky2007).

The above-mentioned studies (Kotzoglou Reference Kotzoglou2002, Reference Kotzoglou2017, Kotzoglou & Papangeli Reference Kotzoglou, Papangeli, Davies and Dubinsky2007) show that, contrary to what Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020: 25) assume, in (26b) the accusative DP is a genuine argument of the matrix verb throughout the derivation and that it controls the subject of the embedded verb, whereas in (26c) the nominative DP is in the (left periphery of) the embedded clause.

So, there really seems to exist no evidence in Greek for CLLDed embedded arguments getting case from or triggering agreement with the matrix verb.

Similarly, if the DP undergoes contrastive focalisation, which does not require clitic insertion in Greek, the clitic is nevertheless necessary in +Agr TCs (27a) but not in the -Agr option (27b): this shows that in the former case the clitic is not normally inserted by CLLD, but it is needed independently of topicalisation.

Furthermore, the CLLD-based analysis is also counter-evidenced by the behaviour of negative quantifiers in TCs.Footnote ⁵ Negative quantifiers like kanenas ‘nobody’ cannot be CLLD-ed:

However, in TCs they can be resumed by the clitic; in fact, they must be, like in (27):

As for our second argument (21b), there are other significant syntactic differences between the +Agr and -Agr options. First, non-resumed indefinite singular DPs are ungrammatical in the former (30a) but not in the latter (30b), which patterns like CLLD structures (30c) (cf. Alexiadou & Anagnostopoulou Reference Alexiadou and Anagnostopoulou2020: 24).

Second, +Agr TCs also seem to be more sensitive to the animacy of the DP and the felicity of its interpretation as object of the embedded verb (31a, b, cf. Tsimpli, Reference Tsimpli, Alexiadou, Horrocks and Stavrou1999:fn.4):

Our third, and most important, argument (21c) is that +Agr TCs are bounded dependencies, in that they do not allow a second level of embedding under a no-control verb like thelo (‘want’) (32a) or mandevo (‘guess’) (33a) in contrast to -Agr configurations (32b, 33b) which generally allow further embedding regardless of whether the verb allows control or not:

Similarly, the -Agr configurations do not care what argument (subject vs object) is fronted (34a), unlike the +Agr ones (34b):

This straightforwardly shows that +Agr TCs are not A’-dependencies like the -Agr options. Then, as expected, only the -Agr, but not genuine TCs, can be captured by a CLLD/null-operator analysis, which does not explain the ungrammaticality of (30a, 31a, 32a, 33a, 34a). For this reason, we will only be concerned with +Agr TCs in the remainder of this paper.

3.1. Argument 1: Obligatory coreference between the subject and the clitic

From a strictly syntactic point of view, TCs exhibit the property of having a clitic that is obligatorily coreferential with the subject (36a–c):

The existence of a strict coreference requirement is corroborated by the fact that the embedded clause cannot have a lexical DP object instead of the clitic:

This is in plain contrast with other configurations where the object clitic does not have to refer to the matrix subject and can, instead, freely alternate with lexical DPs (cf. Roussou Reference Roussou, Stavrou and Terzi2005): for instance, na-clauses selected by volitional thelo (38a, b) or purpose clauses (39a, b).

The obligatory coreference in TCs is maintained even if the embedded verb is passivised, although speakers report that these alternatives tend to be less acceptable sometimes (Dimitrakopoulou Reference Dimitrakopoulou2012: 99) – see (40):Footnote ⁷

In sum, the matrix subject needs to be resumed either by a clitic or by (passive) subject agreement in the na-clause. These facts show that subject agreement/the clitic does not just happen to be coreferential with the matrix subject but that TCs might actually rely on a type of obligatory control involving clitics.

3.2. Argument 2: The subject DP is base-generated in the matrix

The second argument, which we have in common with Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020), is that TCs cannot possibly involve (hyper)raising of the matrix subject from the embedded object position. For us too, a raising analysis is not the most appropriate one for various reasons.

First, as Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020) also show, reconstruction effects, which would be expected under raising, are not found in TCs. See for instance the absence of a narrow scope reading in (41a, b), which strongly suggests the DP is base-generated in the matrix.

(41)	a.	Liga	vivlia	ine	dhiskola	na	ta	dhiavasi
		few.nom.n.pl	books.nom.n.pl	be.3	hard.n.pl	na	cl.acc.n.pl	read.pnp.3sg
		kathe	fititis.
		each	student
		‘Few books are hard for each student to read.’
		(OK wide scope: few $ \rangle $ each; *narrow scope: each $ \Big\rangle $ few)
		(Alexiadou & Anagnostopoulou Reference Alexiadou and Anagnostopoulou2020: 29)
	b.	Dhio	askisis	mathimatikon	ine	dhiskoles	na	tis
		two	exercises.nom.f.pl	Maths.gen	be.3	hard.f.pl	na	cl.acc.f.pl
		lisis	bam ke kato.
		solve.pnp.2sg	bam and down
		‘Two maths exercises are hard to solve in the blink of an eye.’
		(OK wide scope: two $ \Big\rangle $ hard; *narrow scope: hard $ \Big\rangle $ two)

Interestingly, applying further tests with scope ambiguities proves tricky since there seems to be some inter-speaker variation with regard to the narrow scope reading with the quantifier kapji (‘some’) (hence the % rating in 42):Footnote ⁸

(42)	Kapjes	askisis	mathimatikon	ine	dhiskoles	na	tis
	some.nom.f.pl	exercises.nom.f.pl	Maths.gen	be.3	hard.f.pl	na	cl.acc.n.pl
	katalavi	kathe	fititis.
	understand.pnp.3sg	each	student
	‘Some Math exercises are hard for each student to read.’
	(OK wide scope: some $ \Big\rangle $ each; %narrow scope: each $ \Big\rangle $ some)

However, marginally allowing a narrow scope reading for some people even within well-established canonical cases of control (so, independently of TCs) is an independent quirk of this quantifier (kapji). As shown in (43), kapji can have narrow scope for some people, even if it was base-generated in the matrix. As a result, the acceptability of narrow scope in (42) for some speakers is not genuine counter-evidence to our claim that there is no real reconstruction of the subject DP, even if we do not have an explanation for the behaviour of this specific quantifier.

(43)	Kapji	fitites	kataferan	na	prosklithun	apo	kathe	kathigiti.
	Some	students.nom	manage.pst.3pl	na	invite.pnp.nact.3pl	by	every	professor
	‘Some students managed to be invited by each professor.’
	(OK wide scope: two $ \Big\rangle $ each; %narrow scope: each $ \Big\rangle $ two)

Even if it were argued that quantifier scope is not a good diagnostic for raising vs control in Greek,Footnote ⁹ we can, however, rely on additional evidence that movement has not occurred, due to the impossibility for the embedded subject to bind a pronoun in the matrix subject (44).

A reciprocal anaphor in the matrix subject position also cannot be bound by the embedded subject (as shown by 45a, in contrast to the impersonal counterpart 45b):Footnote ¹⁰

This evidence is complemented by the fact that Condition C effects cannot be found: a possessor in the matrix subject can be coreferential with the subject of the embedded verb.

One final argument against raising of the DP from the embedded clause comes from Negative Concord Item (NCI) licensing. In (47a), the NCI kamia is ungrammatical (compare the impersonal counterpart 47b), meaning that it cannot be licensed by the negation in the embedded clause (as would be expected under a raising analysis, cf. Kotzoglou Reference Kotzoglou2017: 303 on quasi-ECM as in 47c).

The above evidence, in addition to the semantic constraints on the matrix DP discussed in Section 2 – which would remain completely mysterious under a raising analysis since the latter is supposed to be completely oblivious to the $ \unicode{x03B8} $ -role/features of the moving argument – leads us to exclude a raising analysis for TCs, in accordance with Alexiadou & Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2020).

3.3. Argument 3: Nesting

To understand the mechanism involved in establishing the tough-dependency, we must assess the presence of locality constraints on TCs. In particular, if we try to embed the na-clause containing the clitic in one more na-clause, we obtain different results depending on the selecting verb. This test is based on the well-known (lexically encoded) distinction between obligatory control verbs (i.e. verbs that strictly require the subject of the embedded clause to be coreferential with the controller) and no-control verbs (i.e. verbs that can allow coreference, but are grammatical without it as well). As argued by Landau (Reference Landau2004, Reference Landau2015), Roussou (Reference Roussou2009), Grano (Reference Grano2015), such a distinction is reflected in the syntax; i.e. there is a syntactic difference between an obligatory control configuration and a no-control configuration (even if the latter allows coreference). In particular, following Landau (Reference Landau2015), there is an Agree relation between an element of the matrix clause and the controllee only in obligatory control contexts, whereas in a no-control environment the subject of the embedded clause may happen to be coreferential but does not Agree syntactically with an element of the matrix clause.

First, TCs reject ‘no control’ predicates (for the latter see Spyropoulos (Reference Spyropoulos, Davies, Dubinksy, Davies and Dubinsky2008), Grano (Reference Grano2015)) like thelo (‘want’) or elpizo (‘hope’), which do not impose any coreference constraints on the subject of their complement clause – compare (48a) with (48b) (cf. 32–33):

Note, in particular, that no-control verbs are bad even if they do co-refer with the most embedded subject (as shown in 48); therefore, it cannot be the case that ungrammaticality is just a consequence of the lack of co-reference, but it has to stem from the no-control nature of the intervening verb.

On the other hand, verbs that do force coreference (Spyropoulos Reference Spyropoulos, Davies, Dubinksy, Davies and Dubinsky2008, Grano Reference Grano2015) on the subject of their complement clause like boro (‘can’), matheno (‘learn’) or arxizo (‘start’) (i.e. obligatory control verbs) are fine in TCs, as shown by (49a, 50a, 51a), with the (b) examples offering their respective impersonal alternates:

Another class of verbs, prospatho (‘try’) and kataferno (‘manage’), which according to Grano (Reference Grano2015) force coreference only when the embedded verb has the same agreement features as the main one, are also fine with TCs (52) (again with (b) examples offering the impersonal option), provided there is no reference switch (cf. 53):

What do such nesting facts tell us about TCs? We have shown that only verbs triggering obligatory coreference between two subjects can embed the na-clause containing the clitic within a TC: so, the TC dependency is disrupted if the intervening clause does not control its complement. In other words, a control verb seems to extend the locality domain of the lower na-clause, thus allowing the dependency between the matrix subject and the clitic to cross the intervening clause. This is expected under theories of control which claim that Greek control configurations are monoclausal (Grano Reference Grano2015) or that the boundary between the two clauses is transparent (Roussou Reference Roussou2009) or that they involve a multiple-Agree chain between T heads (Alexiadou & Anagnostopoulou Reference Alexiadou, Anagnostopoulou, Franco, Moreno and Reeve2019, Reference Alexiadou, Anagnostopoulou, Mucha, Hartmann and Trawiński2021). As a consequence, the dependency between the matrix clause and the na-clause containing the clitic is blocked by a no-control verb (54a) (regardless of co-reference), while it is allowed if the intervening predicate obligatorily controls the verb of the lower na-clause (54b).

1. (54)

   1. a. *Subj_i V_i [ na V_j [ na Cl_i V_j/k]]

   2. b. Subj_i V_i [ na V_j [ na Cl_i V_j]]

The behaviour we just described is expected from TCs under the view that they are bounded dependencies (unlike clitic left dislocation and null A’-operator constructions, see §2). A similar locality constraint is found on canonical obligatory control predicates themselves: a matrix control verb is no longer able to control an embedded verb if there is an intervening no-control verb (as in 55a–c, schematised in 56a). Conversely, an intervening control verb is unproblematic (as in 55d–f, and the schema in 56b). This common boundedness condition suggests once again that TCs are an instance of control.

1. (56)

   1. a. *V_i [ na V_j [ na V_i]]

   2. b. V_i [ na V_i [ na V_i]]

We return to how nesting can be accounted for by our control analysis in Section 5.2.

4.1. Other instances of control of clitics in Greek

Most theories of control (e.g. Chomsky Reference Chomsky1981, Chomsky & Lasnik Reference Chomsky, Lasnik, Jacobs, von Stechow, Sternefeld and Vennemann1993, cf. Polinsky Reference Polinsky and den Dikken2013) assume or predict that control is restricted to embedded subjects: all other arguments cannot be controlled. In Landau (Reference Landau2015), the subjecthood of PRO is derived as a consequence of independent locality constraints, so that only the highest argument of the clause is available for a higher head to control – compare (57a) with (57b) from English:

1. (57)

   1. a. Mary wants PRO to congratulate Jane.

   2. b. *Jane wants (Mary) to congratulate PRO.

However, it is well known that control does not work in the same way in Greek as in other languages due to the lack of non-finite verbal forms in this environment. In addition to this, there have been independent observations in the literature that, whereas a gap corresponding to an argument other than a subject would be ungrammatical with a control verb, the judgement significantly improves if the antecedent is resumed by a clitic, especially if the clitic acts as a quirky subject. Both with obligatory control verbs and verbs that do not always trigger control but strongly prefer coreference, control of quirky subjects via clitics is completely grammatical (58a–d), as reported by Alexiadou & Anagnostopoulou (Reference Alexiadou, Anagnostopoulou, Franco, Moreno and Reeve2019: 64):

Moreover, Roussou (Reference Roussou, Stavrou and Terzi2005: 337) reports some examples, originally from Ingria (Reference Ingria1981), where a clitic is controlled by the matrix subject, showing that control of objects is indeed an independent syntactic possibility in Greek (59a, b):

The same is true with object control and ‘quasi-ECM’ (Philippaki-Warburton & Catsimali Reference Philippaki-Warburton, Catsimali, Alexiadou, Horrocks and Stavrou1999: 155; Kotzoglou & Papangeli Reference Kotzoglou, Papangeli, Davies and Dubinsky2007: 199, Spyropoulos Reference Spyropoulos, Davies, Dubinksy, Davies and Dubinsky2008: 174), even when the clitic is a genuine object:

Thus, the clitic seems to be controlled by the main subject (58–59), as in TCs, or the main object (60), depending on whether the matrix verb is a subject or an object control verb.

Therefore, we find the same alternation between control of subjects (which are expressed as verbal agreement in Greek) (61a, 62a, 63a) and control of clitics (61b, 62b, 63b) for all of these control predicates (cf. Roussou Reference Roussou, Stavrou and Terzi2005):

1. (61) Subject control verbs

   1. a. (DP_i) V_i na V_i

      (e.g. (I Maria) emathe na ektimai tis operes ‘Maria learnt how to appreciate opera’)

   2. b. (DP_i) V_i na Cl_i V

      (e.g. (I Maria) emathe na tis aresun i operes ‘Maria learnt how to like opera’)

2. (62) Object control verbs

   1. a. V DP_i/Cl_i V na V_i

      (e.g. Vlepo ton Jani / ton vlepo na perni ta lefta ‘I see Janis taking the money’)

   2. b. V DP_i/Cl_i V na Cl_i V

      (e.g. Vlepo to Jani/ ton vlepo na ton sproxnun ‘I see Janis being pushed’)

3. (63) Tough-constructions

   1. a. (DP_i) A_i na V_i

      (e.g. Afti i trofi ine efkoli na paraxthi ‘This food is easy to produce’)

   2. b. (DP_i) A_i na Cl_i V

      (e.g. Afti i trofi ine efkoli na tin paragis ‘This food is easy to produce’)

The syntactic possibilities as to which argument of the embedded clause can be controlled seem to be similar across these three classes: in particular, the grammaticality of obligatory control verbs like matheno and vlepo, when there is a coreferential clitic (regardless of subject agreement on the embedded verb), suggests that control is not syntactically restricted to embedded subjects in Greek. The data reported in (58–60) show exactly this, namely that, in an environment where control is required by the nature of the matrix verb, such a requirement can be satisfied by coreference between the controller and a clitic. Thus, TCs might be just another instance of the same syntactic mechanism.

Having said this, many control verbs generally tend to control the embedded subject rather than a clitic – the most common exception being clitics acting as quirky subjects (Alexiadou & Anagnostopoulou Reference Alexiadou, Anagnostopoulou, Franco, Moreno and Reeve2019: 63–65). This property seems to be ascribable to a semantic requirement: most control verbs (but not all of them, as shown by (59–60) as well as by non-volitional thelo) ‘look for’ arguments whose semantic role requires some intentionality (agents) or mental state (experiencers), depending on the lexical semantics of a given verb. This is why clitics cannot be controlled at all times – see (64a–c), where the clitic corresponds to a patient/theme and its control by the matrix subject is incompatible with the semantics of these specific control verbs.

So, we propose that there is nothing in the syntax that prevents the clitics of (64a–c) from being controlled, but that the semantics of these specific control verbs prohibit control of internal arguments realised as clitics. We resort to this solution because we could not find any purely syntactic criterion to explain the distribution of control of clitics depending on the matrix predicate (and no other theory of control would fare better with this; i.e. this is completely independent of the technical implementation we will propose):

• • Some control (or quasi-ECM) verbs (perimeno, vlepo, … see 59–60) seem not to care about the semantic role of the controlled argument; so, they allow control of both mental-state-bearing and non-mental-state-bearing arguments, regardless of whether they surface as subjects or clitics.

• • Some control verbs (matheno, ksero, boro, … see 58, 64) strongly prefer control of mental-state-bearing arguments, so they usually control syntactic subjects and can only control clitics when they are quirky subjects.

Thus, Occam’s razor applies and the best explanation is that control of a clitic is syntactically possible in all cases, but is constrained by the semantics of the control verb.

At this point, then, we can hypothesise that there is a similar semantic requirement on TCs, modulo its orientation: the main predicate tends to control not mental-state-bearing arguments like agents or experiencers but themes/patients. Therefore, TCs will most often involve control of (i) clitics (regardless of their case, 65a, b) as well as (ii) subjects of passive verbs (65c), and (iii) subjects of some unaccusatives (65d), but not agents (see the ungrammaticality of 65e). In fact, arguments realised by clitics, passive subjects and unaccusative subjects are a natural class in that they usually receive internal $ \unicode{x03B8} $ -roles: so, it is more than reasonable to suggest that a semantic constraint imposed by the matrix predicate of TCs is what makes control of this class of arguments – rather than of agent arguments – the most natural option.

Despite this difference between TCs and canonical control configurations in the semantic constraints imposed on the controllee, we have shown that the syntactic mechanism yielding coreference in TCs is likely to be the same as the one in control verbs. Further research is needed to clarify the details of subject-orientation vs object-orientation in the semantics of control verbs and how it correlates with the syntax.

4.2. A potential issue for control of clitics

There is, however, one asymmetry between TCs and subject control configurations (as originally noticed by Alexiadou & Anagnostopoulou Reference Alexiadou and Anagnostopoulou2020) which must be explained if we are to advocate a control analysis: control verbs do not have an impersonal counterpart when the DP stays in the embedded clause, as shown in (66a, b).

In other words, if the subject does not appear in the main clause but in the embedded one, agreement on the matrix verb is still obligatory triggering what has been claimed to be an instance of Backward Control (BC) (Alexiadou et al. Reference Alexiadou, Anagnostopoulou, Iordachioaia, Marchis, Polinsky and Hornstein2010, Alexiadou & Anagnostopoulou Reference Alexiadou, Anagnostopoulou, Franco, Moreno and Reeve2019, Reference Alexiadou and Anagnostopoulou2020, Reference Alexiadou, Anagnostopoulou, Mucha, Hartmann and Trawiński2021). However, for TCs (67a) this is not the case: an impersonal counterpart (67b) is indeed available, whereas its equivalent with BC is impossible (67c):

The data in (67) may prima facie appear to be a problem for a control analysis of TCs, like the one proposed here, given that control in Greek seems to normally require agreement on both verbs regardless of the position of the controller DP, and in (67b, c) we clearly see that agreement is not found on the matrix verb if the DP is not in the matrix subject position. However, we argue that this is not a real counterargument to a control analysis of TCs since the lack of BC is not a unique feature of TCs but is common to many canonical control configurations. Below we present our arguments as to why the unavailability of BC is not unique to TCs.

Firstly, as shown by Alexiadou & Anagnostopoulou (Reference Alexiadou, Anagnostopoulou, Franco, Moreno and Reeve2019), BC is not possible with object control verbs either, not just with TCs. The grammaticality of BC with object control verbs does not improve even if an object clitic overtly realises object agreement on the matrix verb (68).

Secondly, quasi-ECM constructions (which are considered to be a sub-case of object control, see Kotzoglou Reference Kotzoglou2002, Reference Kotzoglou2017, Kotzoglou & Papangeli Reference Kotzoglou, Papangeli, Davies and Dubinsky2007) impose a similar restriction, namely they do not allow for BC either, again in parallel with TCs. In fact, quasi-ECM constructions match TCs even more closely than other object control verbs do because they allow a dual pattern: (a) the one in (69a, 70a) where the accusative object of the matrix verb is also interpreted as the subject of the embedded verb; and (b) the one in (69b, 70b) in which the matrix verb selects a na -clause with an independent subject (respectively, the nominative case-marked DPs o kapetanios ‘the captain’ and i Faranduri). Kotzoglou (Reference Kotzoglou2002, Reference Kotzoglou2017), Kotzoglou & Papangeli (Reference Kotzoglou, Papangeli, Davies and Dubinsky2007) argue extensively that these two options in quasi-ECM are not related by movement or CLLD, but instead represent two syntactically independent possibilities.

(70b, 69b) are not instances of BC, but are parallel to the impersonal counterpart of TCs: if they were BC, agreement (in the form of a clitic) between the matrix verb and the subject of the na-clause would be possible (see discussion in Alexiadou & Anagnostopoulou Reference Alexiadou, Anagnostopoulou, Franco, Moreno and Reeve2019: 76-80), but it is not (71a, b):

Furthermore, if a subject control verb refers to a non-nominative argument in the embedded clause, BC becomes ungrammatical (whereas the forward control counterpart is fine), as in the following pairs (72a, b) and (73a, b):