Form Copy, Agree and Clitics

1 Introduction

Chomsky et al. (2023: 24) define Form Copy (FC henceforth) as follows:

The authors go on to show that FC operates under conditions of c-command (X and Y must be in a c-command relation), minimality (enforced by Minimal Search) and internally to a given phase (enforced by the Phase Impenetrability Condition). For example, FC applies in the passive in (2):

Here, the two occurrences of {many, people} are copies by FC, since the first copy c-commands the second, there is no intervening copy, and, since passive vP is defective, they are in the same phase. FC is relevant at both interfaces: at the Conceptual-Intentional (CI) interface copies may be interpreted, accounting for reconstruction and other effects as has been known since Chomsky (1993); in (2), the lower copy determines the thematic role of {many, people}, in that this argument is interpreted as bearing the thematic role appropriate to the internal argument. At the Sensory-Motor (SM) interface, only one copy is interpreted; this is almost always the highest one, as in (2).

FC is active in the derivation of A-movement (as in (2)), control (see Chomsky et al. 2023: 34–36) and other relations where X and Y are XPs. The purpose of this paper is to investigate the consequences of the possibility that FC can apply to features. We will suggest that FC can derive a very simple and natural version of Agree (this is hinted at by Chomsky et al. 2023: 58) and, following Roberts (2010), that this is turn can derive the effects of head movement and cliticisation. Furthermore, Kallulli and Roberts (2023) extend the approach to A’-movement.

2 FC and Agree

In this section, we begin by showing how Agree can be derived from Form Copy. We then look at standard cases of subject agreement, where the subject (i.e. the external argument in standard cases) raises to SpecTP. We show how the label that results in exocentric structures like the ɸP formed by raising the external argument and labelling TP with the ɸ-features shared between the external argument and T can only involve the features copied by Form Copy; this is a welcome result. We then go on to consider objects and object agreement. Here, we discuss the typological generalisation that object agreement is only found where there is also subject agreement; in order to explain this generalisation in terms of cyclicity and labelling, we propose a general difference between Case and agreement: Case-assignment/licensing is a simpler relation than agreement, canonically involving fewer features and a smaller structural context. For these reasons, Case licensing precedes ɸ-agreement in the derivation; objects are canonically case-marked and do not agree, and subjects canonically agree and show unmarked case. This result ultimately follows from cyclicity, as we show.

Consider a standard case of Agree between T (the Probe) and the external argument DP (EA), the Goal:

Here, all the conditions for Agree are met: T asymmetrically c-commands D, T has uninterpretable features matching a subset of those of D, D has an active Case feature and there is no closer Goal to T than D.

In terms of FC, we can simplify this structure as follows:

As long as T c-commands D, T and D are in the same phase and there are no intervening ɸ-bearing heads (minimal search), FC can apply to the two occurrences of the ɸ-features. More precisely, FC applies separately to the Person, Number and Gender features that make up the ɸ-set. These features are structurally identical since they have no structure: they are simply features (or perhaps Attribute-Value pairs). In that case, they are ‘interpreted in exactly the same way’ at CI (but see below for an important proviso). There is, therefore, no need for the distinction between interpretable and uninterpretable features, the Activity Condition or the other extra assumptions associated with Agree. Agree reduces to FC.

At PF, the highest copy is typically the only one realised. For subject-verb agreement, the reflex of T, D Agree (or, now, FC(T, D)), there is a good deal of cross-linguistic variation on this point. In most languages, D raises to SpecTP (we assume, following Chomsky 2013, 2015 that this is for labelling reasons) and so is the higher copy at PF. It is usually the case that more ɸ-features are realised on the subject than on the finite verb (i.e. V/T): subject pronouns always instantiate Person and Number features in fully differentiated form while verbal inflection frequently does not, even in null-subject languages (‘rich agreement’ in those languages may be the result of D-to-T incorporation; see Roberts 2019 and below).

According to labelling theory, the EA raises to SpecTP to allow the structure formed by External Merge of the EA to a v-projection to be labelled. The EA and T share ɸ-features, which then label the category formed by Internal Merge of the EA to T. The relevant parts of the derivation are shown in (5):

(5b) is derived from (5a) by External Merge of T to α, Internal Merge of DP, labelling of α as vP and labelling of the root node as ɸP as a consequence of the ɸ-features shared by DP and T. Here, DP and T agree in ɸ-features. By FC, the ɸ-features are copies. We thus obtain the result that the label determined in ‘exocentric’ structures like ɸP in (5b) is in fact the feature (or feature set) that undergoes FC. This puts a welcome constraint on the features which can label in this way. The informal notion of ‘Spec-head agreement’ that we see here results from the interaction of labelling-driven Internal Merge of DP and FC applying to the ɸ-set.

For direct objects, we seem to have a structure of the kind in (6):

However, if direct objects are first-merged in SpecVP, as is widely assumed, then we have a structure more comparable to (5b), shown in (7):

Here, we have a further unstable structure for labelling, as in the case of the first-merged position of the external argument in (5a). This can give a motivation for object shift, assumed to hold in general by Chomsky et al. (2023). Object shift raises the object to SpecvP, where, as in (5b), ɸ-agreement between the object DP and v labels the category immediately dominating DP and vP (the category formed by first-merge of DP and V is now labelled as VP). The external-argument DP is merged to vP, giving the structure in (5a’) rather than (5a):

Assuming the external-argument DP has distinct ɸ-features from the internal argument, there is no possibility of labelling α with the same ɸ-set as the label of the ɸP immediately dominated by α; hence, the external-argument must raise to TP as proposed by Chomsky and described above. We will further refine (5a’) and the associated analysis below.

Many languages (including most Indo-European ones, even a language as morphologically impoverished as English) show subject-object asymmetries in agreement, in that subjects overtly agree with the verb but objects do not. In fact, (8) is a reasonably well-established typological generalisation:

The generalisation in (8) implies that (9a–c) are attested, but not (9d):

(9a) obtains in languages with both subject- and object-agreement, e.g. Georgian, Basque, etc.; (9b) is the familiar typical Indo-European situation of subject-agreement but no object agreement; (9c) is widespread in East Asian languages and elsewhere: here there is no overt agreement at all. Finally, (9d) is the case ruled out by the implicational statement in (8).

Consider in this connection the schema in (10), which represents a cross-linguistically common situation of subject-agreement with a possibly nominative-marked argument and accusative-case assignment to a non-agreeing direct object:

(This structure is simplified in that it does not take object shift into account, but the point concerning subject-agreement and case-licensing of direct objects is the central one here; see (5a”) below). The pattern in (10) is prevalent in Indo-European languages, but by no means restricted to them. Let us take it to be a canonical situation (at least for nominative-accusative languages) with departures from it in the directions of greater ‘richness’ (case on SU, agreement with DO) or ‘poverty’ (lack of case/agreement on one or both arguments) as departures from the basic pattern.

We can understand (10) in the following terms:

In these terms, given the discussion of object-shift and object-agreement above, we should recast (5a) once again, this time as (5a”), where K stands for (Accusative) Case:

The external argument has ɸ-features but no Case features (assuming Nominative to be a default Case); hence, the structure is unstable: α cannot be labelled unless the external-argument DP raises. When this happens, it can be labelled KP, instantiating the idea that this is the Case domain.

The statements in (11) are somewhat stipulative in themselves. However, they can be understood in the light of Chomsky’s (2013, 2015 labelling algorithm, summarised in (12):

As we have seen, (12b) is the relevant relation for licensing both subjects and direct objects, although (12a) may be relevant for licensing of oblique Cases; we return to (12c) in the Section 3. Arguments in the TP domain have more features for labelling, i.e. ɸ-features as opposed to Case features. This accounts for the observation that there is generally more inflection in TP than in vP.

A further point concerns Case-realisation. In (10), ACC is typically not realised either on v or on DP. We can account for this in terms of two ideas: first, Case features can never be realised on verbal categories; hence, v cannot mark ACC. Second, the ACC-feature on DP is a property of the deleted lower copy of object shift. In this connection, a further typological generalisation, Greenberg (1963)’s Universal 41, is relevant:^[1]

If OV order is always derived by some form of object shift without V-movement, or with V-movement to a position lower than the landing site of object shift, then we have the derived structure in (14):

Here, DP’s ACC-feature, the highest copy, is realised. If object shift is general, then VO orders are realised when the verb raises past the object. In that case, the object no longer bears the highest copy of the Case-feature, but that feature cannot be realised on the verb, hence, canonically there is no overt case-marking.

The statements in (11), and the general differences between Case and agreement, ultimately derive from the theoretical generalisation in (15) (this is discussed in much more detail in Roberts Forthcoming):

This generalisation in turn derives from the Strict Cycle Condition, which can be formulated as in (16):

In other words, in the configuration in (17), any rule which can apply in domain A must apply there before applying in domain B:

In the case at hand, domain A corresponds to vP and domain B to TP:

Furthermore, case-licensing is an intrinsically simpler relation than agreement. First, agreement, at the clausal level, typically involves more than one feature: person, number and gender/noun-class being the most common and very frequently found together. Case, on the other hand, can be seen as a simple, atomic feature (accusative, ergative, genitive, etc.)^[2] and unmarked, non-dependent cases (nominative, absolutive) are frequently analysed as absence of case, especially given their strong cross-linguistic tendency to be morphologically unmarked. Second, as we have seen, agreement seems to require a richer goal than case-marking. Thus, case-licensing is a simpler relation than agreement, in terms of both the features shared and the structure of the goal.

Put simply, there’s more going on in agreement-licensing than case-licensing and this follows from the difference in the labelling features. Furthermore, Case-labelling as in (5a”) operates in a smaller structural domain than ɸ-labelling for subject agreement, as can be seen in (5a”). It follows that the Strict Cycle predicts that we apply Case-labelling first (in vP before TP), i.e. we can deduce (11), which in turn is ultimately derived from the Strict Cycle.

3 FC and head movement

Roberts (2010) develops an approach to head movement which, although this could not have been foreseen at the time, is compatible with the Labelling Algorithm in (12). One consequence of the Labelling Algorithm is that when two heads are combined, giving the structure in (19), the complex head cannot be labelled if X and Y are distinct categories:

(19)

Chomsky (2015:8) concludes that the configuration in (19) is only possible where one of X and Y is an acategorial root (i.e. a lexical root with no category feature) and the other is a categoriser, e.g. v, n, etc. However, as Roberts (2019:157) points out, as long as X and Y are non-distinct in categorial features, the shared features can label the derived head α. Complex heads can be identified in terms of the definition of minimal category in (20):

This approach allows for head-movement in a highly restricted set of cases, of the general form in (21) (the numerical subscripts on Y1 and Y2 merely serve to identify the two occurrences of Y for expository purposes; they are not part of the label):

(21)

By the definition in (20), Y₂ can be minimal, but only if X is minimal and has a label non-distinct from Y. This approach derives the following generalisation:^[3]

In the light of this, consider the familiar cases of verb-movement to ‘T’. This arises in the following configuration:

The complex v, containing the Root and v, is labelled as v, in line with Chomsky (2013, 2015. At this point in the derivation, neither α nor β have been labelled. FC causes the V-feature of v to be interpreted as a copy of T’s V-feature (v is the head of the lower phase and so accessible to T in line with the Phase Impenetrability Condition). Let us now adopt the following postulate:

(24) is a natural condition in that features form categories rather than segments. The notion of copying a segment is not coherent: a segment is merely a part of the structure of a category, more precisely a part of the label of a structure formed by Pair-Merge. In general, parts of labels cannot participate in formal operations. Therefore, segments cannot participate in operations such as FC, only features and categories (labelled by features) can.

(24) has the consequence that in (23) both segments of v must be copied, with the result that the Root is also copied to T, giving the derived structure for T in (25):

In copying v’s V-feature to T in this way, it may appear that the criterion of structural identity for FC is not met, but, given (24), the condition is in fact met in the most economical way: the identical feature V, which labels the complex v-Root head is copied (given this, it might be more perspicuous to define FC in terms simply of ‘identity’ rather than ‘structural identity’). In (25), v is Pair-Merged to T; this is the only possibility, since ‘head substitution’ is impossible in bare phrase structure (see also Chomsky 2020: 56). Note further that ‘T’ is labelled as T-V. As Chomsky (2015) points out, this is natural since the combination of the verb and Tense is a tensed verb, not simply Tense. Finally, T’s ɸ-features will contribute to labelling the structure formed by Internal Merge of the EA, as we saw in (5). The resulting structure is (26):

We can clearly see that this structure is exocentric.

Now consider cyclic head movement. For example, C arguably has a T-V feature in a verb-second structure, as shown in (27):

Here, C’s T-V features form copies with the T and V features of ‘T’ (T’s ɸ-features are irrelevant to this; FC applies to individual features), so the T-V complex is raised to C. Given the constraint in (24), this is the required result of FC. I leave aside here the related questions of labelling γ and raising an XP to SpecC_T-V (see Blümel 2017; Roberts 2019, Chapter Five, for proposals).

What of the three conditions on Form Copy? As we have seen, according to Chomsky et al. (2023), these are (i) compliance with the Phase Impenetrability Condition (PIC); (ii) compliance with Minimal Search, giving rise to minimality effects and (iii) c-command. Since heads of phases are accessible to categories outside the phase by the PIC, there is never a problem with (i) for head-movement. Similarly, Minimal Search applied to FC will derive the Head Movement Constraint; in a structure like the following FC cannot apply between positions X and Z without also involving Y:

Finally, the c-command condition on FC also derives from Minimal Search: ‘[t]he smallest set in which the copy relation can hold between X and Y is {X, Z} where Y is a term of Z’ (Chomsky et al. 2023: 26). In (27), for example, ‘T’ and its features are contained in ɸP, the sister of C_T-V.

Next, let us compare the case of standard Agree between transitive v* and a direct object with head-movement of an object clitic. Since we now see Agree as FC applied to features, we have (29) in the case of Agree between v* and the direct object (compare Agree between the subject and T shown in (3)):

Since Agree reduces to FC, the copy of D’s ɸ-set is deleted at the SM interface. Roberts assumes that the clitic consists only of ɸ-features, which are then deleted at the SM interface in line with FC. Thus, the effect of head-movement can be derived from FC in exactly the case where the moved head has a subset of the features of the target head. It also follows that features shared between the two heads label the complex head.

More generally, we can define a clitic host as Fɸ, i.e. a head which has ɸ-features and possibly other features. If F is a left-peripheral head, we have some form of ‘second-position’ clitic (which may not always be strictly in second position, as in (30), but is intrinsically associated with the left-periphery). Such clitics are arguably found in European Portuguese (Madeira 1993; Uriagereka 1995, etc), giving rise to enclisis in many main-clause contexts as in (30):

Here, the subject is in a left-dislocated position (see Barbosa 1995 for extensive arguments for this), the verb has moved into the left periphery (a natural extension of V-to-T movement, as we saw above) and the clitic a is attracted to a host Fɸ in the left periphery below the landing site of verb-movement, giving rise to enclisis (see Kayne 1991).

We saw above that FC is subject to three conditions: (i) the higher copy must c-command the lower copy, (ii) there is no intervening copy and (iii) the copies must be in the same phase. The first condition is clearly not a problem for left-peripheral clitics as the left-peripheral head Fɸ c-commands the first-merged position of the clitic (which we take to be VP-internal). Concerning (ii), in a system with left-peripheral clitics, v cannot be a clitic host or it would count as an intervener for FC. This conclusion is consistent with the widely accepted idea that there are two types of clitic systems; left-peripheral (or ‘second-position’) systems and adverbal systems (see Roberts 2010 for discussion and references). In the former, the clitic hosts are in the left periphery as we have just seen for European Portuguese; in the latter, v (or the ‘low’ left periphery associated with vP; see Belletti 2004) is the clitic host, as we briefly discussed for object cliticisation above. What follows from our approach is that the two systems are distinct: if a language has left-peripheral clitics, it cannot have adverbal clitics and vice versa. As far as I am aware, this is the consistent cross-linguistic pattern (although an anonymous reviewer points out that Macedonian may be a counterexample, a point I leave for future research).

Concerning the third condition, we take it that in left-peripheral clitic systems, the clitic A’-moves to the left periphery via the vP-edge. In this respect, the movement is a standard case of A’-movement transiting through the lower phase edge in order to comply with the Phase Impenetrability Condition.^[4]

The full derivation of left-peripheral clitic-movement proceeds as follows. First, the clitic, here indicated as ɸ, moves to the edge of the vP phase, giving the structure in (31):

The structure in (32) is then formed by Internal Merge of the EA to the projection of T, labelling of the category resulting from this as ɸP and labelling of α in (31) as vP. All of this follows the proposals in Chomsky (2013, 2015:

Next, the clitic host Fɸ is externally merged and the clitic internally merges to that position, giving (33):

Finally, left-dislocation of EA and V-movement to a left-peripheral head higher than F give rise to the word order seen in (30). Here, as we have pointed out, clitic-movement to F complies with the PIC; Tɸ is not an intervener for FC between the moved clitic and its copy because it has a superset of the clitic’s features, having at least T- and V-features in addition to (subject) ɸ-features.

It is well known that European Portuguese shows complex proclisis–enclisis alternations in main clauses. For example, enclisis is impossible in negative clauses. Instead, proclisis is required as shown in (34):

In (34b), we have left-dislocation of the subject as in enclisis cases like (30) and the same clitic-movement to a left-peripheral host Fɸ. However, negation blocks verb-movement (presumably because it lacks the features that give rise to Form Copy with the verb, given the discussion of verb-movement at the beginning of this section) to the higher left-peripheral position, so the verb does not move higher than T, giving rise to proclisis as seen in (34b). This general analysis can be extended to other cases of proclisis.

Consider finally null subjects. Here, the EA (or more precisely the argument that becomes the null subject) is a ɸ-set as shown in (35):

As such this argument counts as a copy of T (i.e. of T’s ɸ-features) and so is deleted at PF, with T’s features typically realised ‘richly’. In line with Chomsky (2015: 9), we assume that the external argument does not need to raise to SpecTP for labelling reasons. As Chomsky says, ‘[i]n terms of labelling theory, Italian T, with rich agreement, can label TP. For English, with weak agreement, it cannot’.

4 Conclusions

In this paper, I have proposed that Form Copy, applied to features, can derive a very simple form of Agree without the need for the distinction between interpretable and uninterpretable features or the Activity Condition. Furthermore, since Roberts (2010) derives head-movement from Agree, FC can be similarly extended in such a way as to derive the effects of head-movement. This conclusion holds independently of whether head-movement is a purely SM operation or whether there are narrow-syntactic cases; all that is required is that at least some stages of the derivation in SM have hierarchical structure, which is independently necessary both for linearisation and for aspects of the syntax-prosody mapping. I have also shown how an account of cliticisation, both proclisis and enclisis, can be derived, updating the proposals in Roberts (2010) in terms of FC.

The proposals made here illustrate the interplay of theory and observation. The theoretical motivation behind the proposals is straightforward: to show how the recent proposals for Form Copy as an operation independent of Internal Merge in Chomsky et al. (2023) can extend to Agree, head-movement and cliticisation (all cases of FC applying to features of heads). As Chomsky et al. emphasise, this is an example of the prosecution of the minimalist programme for linguistic theory, having as its ultimate goal a genuine explanation of the nature of the language faculty. This paper is intended as a small contribution to that overarching goal. The empirical domain of the paper is comprised of numerous empirical observations regarding agreement, case, verb-movement and cliticisation. Many further empirical issues can and should be treated in these terms: notably subject clitics, and the interaction of labelling, case and word order, extending Chomsky’s (2015) notion of ‘weak’ and ‘strong’ heads along the lines of Roberts (2019, Chapter Two). Furthermore, as noted in the Introduction, Kallulli and Roberts (2023) extend this approach to A’-movement.

I have barely scratched the surface of these rich empirical domains here, but I hope that the proposals made will open new and fruitful lines of research and continue progress towards genuine explanation of the apparent complexities of natural language.