3.1 Agree as a vehicle for binding

As indicated in section 1, our approach is based on the Agree operation. In developing our analysis we elaborate on the specific properties of Agree-based dependencies. In this sense, our proposal is part of a more general debate about the proper characterization of Agree, concerning whether one probe can interact with multiple goals (Hiraiwa 2001; 2005), whether Agree proceeds upward or downward (Preminger 2013; Bjorkman and Zeijlstra 2019), the valuation by Agree in relation to spell-out (Preminger 2019; Reuland 2020), what the locality constraints on Agree are (see below), and whether Agree involves full feature bundles (Kratzer 2009; Reuland 2001; 2011a) or rather individual features. Reflexive binding in Russian serves as a further probe into these issues, contributing to a more general understanding of Agree.

The rationale behind using Agree to account for certain types of syntactic binding dependencies, is that there is a class of anaphors that lack the feature specification for person, number or gender (briefly, ϕ-features) a corresponding pronominal would bear. This class includes SE-anaphors (Reinhart and Reuland 1993), with or without a further component (Dutch zich(zelf), Norwegian seg (selv), and also Russian sebja and svoj.⁷ This deficiency should not be treated as accidental. It is explained if the deficiency allows the anaphor to be valued by an Agree operation copying the relevant values from the source NP and pasting them onto the anaphor. This effects unification of these occurrences and the formation of a feature chain, which is interpreted as binding by the interpretation system. Thus, Agree is involved in the interpretation of such underspecified pronominal elements, but since it crucially makes use of this deficiency, it follows that, as stated in fn. 2, no such claim is made for binding in general (see Reuland 2020 in response to Preminger 2019).

Unlike other Agree-based approaches to binding (e.g. Hicks 2009, or, specifically for Russian, Antonenko 2012, among others), we rely on independently existing ϕ-features rather than special features of reflexives such as <VAR>, or [+reflexive] and their kin, that play no other role in the grammar and possess no locally specified interpretation. In our approach we adopt the feature valuation system of Pesetsky and Torrego (2007), where interpretability and valuation are assumed to be independent properties, and take reflexives to enter the derivation with interpretable unvalued ϕ-features,⁸ building on the insight of Reuland (2005; 2011a: 5.9) that feature sharing between interpretable feature instances brings about an interpretive dependency. The reverse is of course not necessarily true, anaphoric dependencies can be established through other means too, but this is immaterial here. It is important to stress that feature sharing creates a single syntactic and semantic object occupying several positions rather than mere repetitions (cf. copy theory of movement/internal Merge). Different instantiations of the shared feature thus cannot make independent contributions to interpretation, including their perspectival properties, interaction with distributive operators etc.

Along with others (e.g. Chomsky 2000) we assume that probes (understood as the feature instances initiating the Agree operation) are properties of heads and can only see into their complements,⁹ i.e. essentially our approach employs a version of downward Agree, albeit with a sufficiently different view on the limits and effects of this operation. Many studies of agreement have solely concentrated on the valuation of the probe. The element initiating the operation has often been the only one that is considered to change its value as a result, without a clear distinction between the two independent properties. This, however, is not necessary. Goals may be affected too. Though the probe, which initiates the operation, c-commands its goals, we assume, again following Pesetsky and Torrego (2007), that the directionality of valuation is not determined intrinsically, and goals can get valued as well.

Anaphor binding typically allows patterns where one antecedent binds anaphors in different positions that cannot be related in succession. Therefore, assuming downward Agree, it must be a one-to-many operation that underlies binding, in line with Multiple Agree (Hiraiwa 2001; 2005; Boeckx 2003; Chomsky 2004; 2008), which is not limited to a single goal. Furthermore, following Zubkov (2018) we assume that valuation is carried out by the closest suitable feature instance. As structural conditions on the valuation of probes and goals are different, a probe is valued by the closest suitable goal, a goal is valued by the closest suitable probe.¹⁰ This is illustrated in (5). The unvalued occurrence of ϕ probes and finds two other occurrences of ϕ in its domain, one valued, which values the probe, and a (lower) unvalued occurrence, which is simultaneously valued as well, as there is no closer probe that could have valued it here.¹¹

Unlike for Hiraiwa (2001, 2005) and Chomsky (2004: 115, 2008: 142), where goals intervene unless inactivated by the probe itself, in this model probing continues all the way down to the next lower probe. This pattern of intervention essentially reduces to minimality.

As a result, the probe spreads (or attempts to spread, depending on whether or not the potential goal is already valued) the value of the closest suitable goal to the farther goals in its domain, subject to intervention by other probes, as illustrated in (7) and (8) below. Assuming the feature-sharing view of Agree outlined above, this allows us to capture binding. Both the reflexive and its antecedent correspond to goals (as in the analysis of binding in Mandarin presented in Giblin 2016; see also Wong 2021), neither of them to the probe (unlike e.g. in Hicks 2009 or Rooryck and Vanden Wyngaerd 2011). Thus, even though the unvalued reflexive is merged lower than the valued antecedent, a resort to upward Agree is not warranted. Moreover, with upward Agree the prospects for deriving the complementarity patterns discussed below and correctly constraining the structural positions of possible antecedents would appear unpromising. From below the antecedents cannot be adequately characterized as the closest goals, as there may be other goals intervening, including features of lower arguments and of phrases embedding the reflexives.

We will start elucidating details of our approach on the basis of a simplex sentence.

3.2 Local binding – simplex clauses

Consider, then, a sentence as in (6), where a reflexive object is bound by the local external argument that bears nominative case and triggers verb agreement:

The relevant aspects of its structure and a possible derivation that we propose are demonstrated in (7), where Vanja is taken to be first merged as the specifier of vP and moved into Spec TP.¹² The structure contains two probing heads, to be discussed in turn.

In line with Chomsky (2008) and much preceding and subsequent work, we assume a high ϕ-complete probe with unvalued person and number (#) in the left periphery of the finite clause. As indicated, the structure also contains a low #-probe, whose contribution will be sketched in (8) below. The high probe is valued by the nominative goal Vanja.¹³ As it includes the closest person probe c-commanding sebe, once it gets valued, it simultaneously values the unvalued interpretable person feature of the reflexive as well. As a result, Vanja and sebe share a person feature. However, the high probe is unable to value the number feature of sebe by minimality, as there is a closer number probe that intervenes, which leaves sebe valued for person, but unvalued for number for the derivation in (7). As in this configuration the low number probe is valued by the same goal, an alternative derivation that results in number sharing between Vanja and sebe is available:

On the surface, in such cases the alternative derivations make almost no difference, arguably except for a variation in subtle interpretive effects. The different contribution of the probes becomes more conspicuous in other configurations to be discussed.

The ϕ-complete probe in the left periphery of the finite clause gets reflected in the verb agreement.¹⁴ Many phenomena amenable to Agree-based analyses may be better viewed as indirect dependencies between goals, mediated by probes that may be hardly ever visible on their own. Accounts of downward Agree are rarely explicit about how inflection ends up on the verb, and it is sometimes assumed that the verb enters the derivation without any features at all. Our model of Agree opens up a possibility to implement verb agreement as goal valuation within the narrow syntax, though it doesn’t depend on this.¹⁵

As evident in morphology and as is common across languages, in Russian this ϕ-complete probe can only be valued by a nominative goal, even if it is not the highest argument, for reasons that we don’t fully understand yet. It is conceivable that EPP-satisfying movement, which does not apply to inherently case-marked phrases in Russian and perhaps bleeds vP-internal structural Case assignment, may always result in the nominative being close enough to the probe to value it, especially if one allows for the possibility of lower copy spell-out for nominative internal arguments pronounced in situ.¹⁶ This probe can only account for binding by nominative antecedents, which, as already discussed, is not always the case.

To account for binding by non-nominatives, a #-probe, insensitive to case and considerably lower in the structure, will be shown to be necessary. Its low position captures several important patterns. First, the ability of non-nominatives to bind into PPs varies with the latter’s height, as was demonstrated in (3), repeated here:

Here, the event-external modifier iz-za POSS vkusov appears to be adjoined too high for the possessive svoj to be accessible for binding by the non-nominative antecedent, which is mediated by the low #-probe, although it is still accessible for binding by the nominative, which is mediated by the high probe.

Furthermore, constituents that are merged above the non-nominative antecedents (such as the non-thematic possessors in NPs, see section 4.1) are unable to bind or interrupt binding. Finally, as binding by non-nominatives is possible within extended projections of different categories, both nominal and verbal, it is reasonable to expect that the probe responsible for it is situated rather low, where the projections may still be maximally uniform, perhaps even category-neutral. In all relevant respects discussed in this paper NP-internal binding patterns together with binding through the low clausal probe against the dependency established through the high left peripheral probe. As non-nominative antecedents are normally the highest on their theta-grids, in the verbal domain the low probe should be situated immediately above vP, assumed to be the projection completing the argument structure.¹⁷

Note that our analysis assumes that the vP phase does not restrict Agree, following e.g. Bošković (2007) and Keine and Zeijlstra (2020). Cf. also Richards (2012). The head introducing the external argument (conventional v⁰) is unlikely to be phasal for the purpose of punctuating movement paths either, for otherwise intermediate copies of scrambled constituents at the edge of vP would probably be able to value the low probe and bring about various unattested binding possibilities. However, the head carrying the probe itself may well define a phase for that purpose in the same region. That would be in line with ideas associating phasehood with ϕ-feature probes (Chomsky 2008; Richards 2012 and others). As we don’t assume accusative Case assignment parasitic on ϕ-feature Agree and there is no independent justification for direct object agreement in Russian, unlike Chomsky we don’t postulate any ϕ-features on v⁰ proper. That said, it should be noted that there is no clear evidence for a phase boundary anywhere in the v region in Russian.¹⁸

In the present approach Agree is assumed to proceed independently for different features and is subject to intervention on a feature-by-feature basis. For the more distant probe to be able to search across the low probe, they must differ in their feature composition, otherwise binding would have been blocked by minimality. This is what we have already suggested to account for the pattern in (4). As we have illustrated above, local nominative antecedents highest on their theta-grid can establish a dependency through either of the probes. However, the binding dependency that is established exclusively through the left peripheral probe across the low probe (which is the case when the antecedent is not strictly local) and the binding dependency established exclusively through the low probe (which is the case when the antecedent is not nominative) turn out to have different properties. Only the latter allows for inanimate antecedents,¹⁹ the former doesn’t; moreover, it yields an awareness effect (see section 4.2).²⁰ On the other hand, only the former can be collective, whereas the latter can only be distributive with plural antecedents (see section 4.4).²¹ Neither of them can result in a dependency that is collective and doesn’t imply awareness at the same time.

To account for this we propose that the lower probe is an incomplete ϕ-feature probe involving number but lacking person. As a result the dependency via the high probe across the low # probe is based on person sharing only, because of intervention, and the dependency via the low probe is based on number sharing.²² Although gender piggy-backs on the dependency that is established it doesn’t appear to play an independent role.

On this view, inanimates, which prima facie invariably appear to be 3rd person, do not in fact carry a person feature (cf. Richards 2015; Lochbihler, Oxford & Welch 2021) or at least do not endow it with an interpretation; therefore they cannot take part in an interpretive dependency based on person. Thus, with inanimate antecedents only one route is available:²³

The configuration is given in (11):

Since èta kniga is inanimate it cannot take part in an interpretable person-based dependency. However, it can provide interpretable number. Though the high # probe cannot value goals across the low one, the latter can value sebe’s unvalued # feature. Accordingly, binding by plural inanimates is always interpreted distributively.

For reasons of space in this paper we avoid discussing which copy gets valued if the goal moves between domains of several probes during the derivation. Scrambling doesn’t bleed reflexive binding in Russian, though some other kinds of movement appear to do so. Given that languages vary widely with respect to movement possibilities this may account for some part of cross-linguistic variation. Note, however, that, as far as variation is concerned, we don’t imply that all expressions that are used as anaphors across languages form a natural class and take part in the same syntactic derivations as in Russian.

With the essential properties of our approach in place, in the next section we will show how it works in more intricate cases.

4.1 NP-internal binding: Possessor versus author

NP-internal binding shows an interesting difference between two types of possessor phrases, namely possessors that are in a relation with the possessee that is merely contextually specified (ownership, or more generally the R-relation in Higginbotham 1983) and possessor phrases that stand in a thematic relation such as authorship. The former are unable to antecede reflexives; only the latter can, as in (12):

The possessor in (12a) can only be construed as the author, not a person who owns the letter. The latter interpretation is fine if the possessor doesn’t bind a reflexive, as in (12b). The same pattern was reported for Polish (Rozwadowska 1995). This contrast is derived by the position of the NP-internal # probe.²⁴ Thematically linked dependents are first merged within the domain of the Number probe, presumably because they need to receive theta-roles from the nominal in a local configuration. The non-thematic possessor with no such requirement is merged outside it. This is illustrated in (13), with N and the possessors in their base positions:

Here only the author Poss is accessible for probing by the # probe. The latter is then valued and simultaneously values the # feature of sebe, resulting in binding.

Thus, the author’s number feature can be probed by the NP-internal #-probe, which results in binding. This is not available for the possessor merged outside the domain of the #-probe.

4.2 Binding into NPs: the role of animacy/awareness

Binding into NPs shows an interesting animacy effect, illustrated in (14):

In (14b,c) the inanimate subject èta kniga cannot bind the anaphor across the possessor, whereas in (14a) in a similar position the animate Vanja can. Let (15) illustrate the derivation of (14a):

Here the high person probe is valued by the nominative Vanja and values the person feature of sebe, since there are no other person probes in the domain.

Consider next the inanimate subject in (14b,c). The structure is as in (16):

The low # probe on the verbal spine probes for number, finds a NP goal to be valued by, but cannot see beyond the NP-internal # probe by minimality. Hence, it cannot value the number feature of the reflexive. As discussed in section 3, the person-based dependency is unavailable for inanimates. Therefore there is no route available to èta kniga to bind the reflexive here. Thus, the animacy requirement on the antecedent is explained. As already illustrated in section 3.2, with local binding no such restriction applies.

The approach provides an account for a further person-related contrast, which is illustrated in (17):

The reflexives cannot be bound by the matrix subject Maša in (17) if Masha doesn’t realize that the book she bought is about herself or her own mother. Here too, the dependency is built on the person feature, and follows the pattern in (15). No such restriction applies to the interpretation of the pronominals.

This indicates that person sharing does not just require animacy but also implies awareness, which may follow if, as we assume, it is person features that carry perspectival properties, and feature sharing by Agree effects their unification.

The patterns we discussed show two things. One is that animacy and awareness requirements are not limited to traditional ‘long-distance binding’ across clause boundaries.²⁵ The other is that such requirements can be captured by a straightforward Agree-based analysis. In the next section we will discuss intricate binding patterns presented by high non-nominative experiencers.

4.3 Experiencer predicates

Where the highest argument is not nominative, valuation of the probes is split between two goals. The nominative values the probe associated with finite agreement, and the highest argument (dative) values the number probe. This results in two possible antecedents, as in (2), repeated here as (18):

As is standard with this type of verb, and according to the mapping rules of the theta-system (Reinhart 2002), we assume that the dative Vane is merged higher than the base position of the nominative Anja but is unable to satisfy the EPP (see e.g. Slioussar 2007 on this).

The derivations are illustrated in (19–20). The high person probe is valued by the privileged nominative goal, the low # probe is valued by the nearby dative goal:

Note that some PPs appear to be merged too high to be accessible to the number probe. In this case (repeated here from (3)) only the nominative antecedent remains available:

4.4 The number feature and distributivity

Whereas the animacy effect reflects the nature of the person feature, we will now briefly discuss a contrast that may well reflect the nature of the #-feature. To see this, consider (22):

Example (22a) shows that where for semantic reasons the reflexive cannot range over singular individuals it can nevertheless be bound by a plural nominative antecedent in Russian even if the predicate is interpreted as distributive on the antecedent. By contrast, as (22b) shows, this is different for plural non-nominative antecedents. These are not appropriate if the reflexive cannot range over singular individuals. Note that binding by the dative (22c) is otherwise straightforward.²⁶

As we saw in the preceding discussion of experiencer verbs, non-nominative antecedents can only bind through the #-probe, which we propose to be taken to be the source of the contrast. Likewise binding by plural inanimates is always interpreted distributively.²⁷

4.5 Binding into infinitives: the role of animacy/awareness

Russian also allows non-local binding into infinitives. However, there is varying acceptability with different control verbs (Timberlake 1979, Klenin 1974) and some degree of inter-speaker variation documented by Klenin (1974: 40, 57). For instance, as reported by Klenin, in the following sentence some of her informants accepted only the (k) indexing, some others both (k) and (i), but not (j), still others all three options (i, j, k):²⁸

Like other cases of non-local binding we discussed, binding into infinitives also shows sensitivity to animacy:

In (24a,b) the inanimate èta kniga ‘this book’ is unable to bind the possessive anaphor svoj, although its structural position is the same as the animate Vanja’s in (24c), which can.

In order for binding into infinitives to be possible at all, their left periphery should differ from that of finite clauses. Specifically, the high probe should be absent or at least lack the person feature. That is, infinitival clauses may show the effect of restructuring.²⁹

The crucial assumption is that the differences reside in the high probe. This provides a straightforward means to account for the variability: control verbs may differ in the richness of the CP they select, and speakers may differ in whether or not they allow CPs without a high probe. The three options in (23) observed by Klenin follow if restructuring is unavailable, obligatory and optional respectively.³⁰

Assuming that the low #-probe is always present the animacy effect follows. Just as in the cases of non-local binding into NPs discussed in section 4.2, the low #-probe may be crossed by a person probe in the matrix clause, but not by a #-probe in the matrix clause. The relevant patterns are graphically represented in (25)–(26):

Interestingly, there is an environment where binding into infinitives is not allowed, namely in the presence of an adjunct clause, see the contrast in (27):

The infinitive in (27a), unlike (27b), is modified by a clause. It seems to be a reasonable assumption that a high modifier clause goes with a richer structure in the left periphery, and hence is incompatible with restructuring.³¹

Summarizing, we have seen that infinitival clauses undergo optional re-structuring. Restructuring entails the absence of a high person probe in the C-domain, but the low #probe above vP is retained. Consequently, animate matrix nominatives can bind anaphors in infinitival complement clauses, since the person probe in the matrix clause – valued by the nominative – can probe across intervening #-probes. Anaphor binding by inanimate matrix nominatives is still blocked, since these can only participate in a dependency based on number, which cannot cross intervening #-probes. Though their exact positions currently remain unclear, some evidence suggests there may be further #-probes in this configuration, but the one indicated is sufficient to guarantee that the dependency is blocked. This explains the animacy requirement on antecedents in non-local binding, without an appeal to any other process but Agree-chain formation.