Ever since Lees & Klima (1963) inaugurated the study of English reflexives in Generative Grammar, a popular thread of theoretical development, culminating in Chomsky (1986), has emphasized the role of narrow syntax in the interpretation of reflexive pronouns. The common wisdom expressed in the Binding Theory (Chomsky 1981; 1986) says that reflexives should be structurally c-commanded by their antecedents in a local domain.

However, early work on the topic took notice of a class of phenomena which appear to resist simple syntactic accounts:

I follow Jackendoff (1969; 1972) in referring to data like (1) – originally discussed by Postal (1968) and Lakoff (1968) – as “unexpectedly bad cases of reflexivization”. Those reflexives are unexpectedly bad precisely because they are not ruled out by any obvious syntactic condition, given that they are fine in (apparently) parallel syntactic environments, such as (2):

These unexpected reflexivization contrasts (URCs), as l’ll call them, inspired Jackendoff (1969; 1972) and others (Wilkins 1988; Dalrymple 1993; Williams 1994) to pursue semantically based solutions, stated in terms of a thematic hierarchy. Their idea is that there must be an independent principle that ensures reflexives cannot be higher than their antecedents in a suitably arranged hierarchy of thematic roles.

Partly motivated by the assumption that thematic roles are transparently encoded in phrase-structure configurations (Baker 1997), proponents of purely syntactic accounts of anaphora appealed to more abstract properties of narrow syntax to distinguish (1) and (2). This was the strategy adopted by Postal (1968) and Lakoff (1968) themselves, who hypothesized that the strings in (1) were transformationally derived from structures like (2), in violation of the Crossover Principle. More recently, Belletti & Rizzi (1988), Larson (1988; 1990) and Pesetsky (1995) all argued in favor of distinct underlying structures for cases analogous to (1) and (2). These were arranged so as to entail that, at the level where the Binding Theory (Chomsky 1981; 1986) applies, only the reflexives in (2) would be c-commanded by their antecedents.

In this article, I will provide an explanation for URCs like (1)–(2) by pursuing a different kind of approach. The work of Reinhart & Reuland (1993) (R&R), insofar as it does not essentially rely on a thematic hierarchy or on phrase-structure-related notions of licensing, opens the possibility of greatly simplifying the structures that a theory of English reflexives requires from syntax. When augmented with an enriched conception of compositionality (Nunberg 1979; Pustejovsky 1995; Jackendoff 1997; Recanati 2010) and a theory of logophoric discourse roles (Sells 1987; Stirling 1993; Minkoff 2004; Oshima 2007; Charnavel 2019), this version of R&R’s reflexivity-and-chains approach makes it possible to reduce the problem of accounting for URCs to the more tractable problem of logophoric reflexives. The sensitivity of reflexives to perspective-based properties is a typologically robust and reasonably well understood phenomenon. It is, thus, methodologically convenient if URCs can be explained in similar terms.2

The main assumption needed will be a modified version of R&R’s Condition A (CA). This modified CA states that reflexive markers signal that a predicate is reflexive (i.e. that two of its arguments are coindexed at the level of semantic interpretation). This is essentially a correspondence rule – in the sense of Jackendoff (1997) – which encodes the lexical meaning of reflexive pronouns like himself.

My approach to these issues is inspired by Menuzzi (1999; 2004), who reformulates R&R’s theory in important respects. Following him, I argue that logophoric reflexives should not be treated as exempt from (my version of) CA, but as tolerable violations of it. Under adequate circumstances – e.g., when they occur within anti-reflexive predicates (predicates which cannot be interpreted as reflexive) – these violations give rise to logophoric interpretations. If (1)–(2) contain anti-reflexive predicates (as I’ll show), the reflexives therein count as logophoric.

Adapting scattered suggestions by Jackendoff (1992), Safir (2004) and others, I will claim that reflexives in the URCs are not extensionally equivalent to their antecedents: one of them always comes to denote an inanimate proxy (an image or a more abstract representation). A minimal condition on logophoric reflexives is that they should refer back to animate discourse antecedents (cf. Sells 1987; Charnavel & Zlogar 2015; Charnavel & Sportiche 2016). Crucially, that condition is not met for (1) – as Kuno & Kaburaki (1977) first hinted at. The unexpectedly bad cases of reflexivization amount to failures in finding relevant logophoric antecedents. This provides a simple extra-grammatical explanation for URCs.

In Section 2, I present a broader range of URCs and argue against accounts based on the Thematic Hierarchy Condition (THC). In Section 3, I explore some of the semantic peculiarities of URCs and formulate a provisional descriptive generalization that captures the phenomena in a way that bypasses most of the counterexamples to the THC: the unexpectedly bad cases are the ones in which antecedents are interpreted as proxies. In Section 4, I discuss exceptions to this which suggest (along with other data) that reflexives in URCs exhibit logophoric properties. Finally, in Section 5, I present a revised version of R&R’s theory from which the logophoric properties of these reflexives comes out as a natural consequence.

In this section, I present a wide range of unexpected reflexivization contrasts (URCs) like (1)–(2) (repeated below) and argue against the Thematic Hierarchy Condition (THC) as an adequate explanation for them. I do this because, even though most would not accept the THC as a theoretical primitive (as Jackendoff (1969) did), the generalization it embodies is still taken to be a descriptively adequate account of URCs. These URCs fall into two groups: syntactic alternations (3)–(8) and semantically related predicates with different linking properties (9)–(13).

Dative alternations (Barss & Lasnik 1986; Larson 1988; Jackendoff 1990a):

Freely ordered PP complements (Postal 1968; Jackendoff 1990a):

Passive alternations (Jackendoff 1969; Reinhart & Siloni 2005):

Verbs of communication (Levin 1993; Culicover & Jackendoff 2005):

Psych-predicates: asymmetries between non-stimulus and stimulus subjects (Postal 1968; Belletti & Rizzi 1988; Grimshaw 1990; Pesetsky 1995):

It is not a priori clear what relevant causal factor on reflexivization co-varies with the shifting judgments in the a–b pairs above. As Reinhart (1983) and Büring (2005) note, contrasts like these are unexpected for theories which assume that licensing conditions for reflexives should be formulated exclusively in terms of phrase-structure relations, since they do not seem to correlate with any obvious syntactic distinction. For instance, all of the reflexives in (3) and (8)–(13) are locally c-commanded by their antecedents – at least insofar as the constituent structure for these sentences is determined according to traditional constituency tests, independently of the binding facts. The reflexives in (4)–(7), on the other hand, are not c-commanded at all, due to the branching structure of PPs. This obviously makes the wrong prediction for (4a), (5a), (6a) and (7a).4

Based on this kind of reasoning, Jackendoff (1969; 1972) and Wilkins (1988) propose that reflexives in English are subject to the following semantic condition:

Given usual assumptions about the assignment of thematic roles, the THC adequately describes URCs like (3)–(13): e.g. (3b), (4b), (9b), (10b) are all cases where the reflexives are Goals and their antecedents are Themes, in violation of (14).

The only thing purely syntactic accounts of reflexivization can do in light of these facts is to say the pairs in the URCs are, despite appearances, not syntactically alike after all (Belletti & Rizzi 1988; Pesetsky 1995; Larson 1988; 1990). These approaches try to motivate, at least for a subset of URCs, more abstract underlying structures where the antecedents do not c-command the reflexives in the bad cases of URCs. As Belletti & Rizzi (1988) and Larson (1990) acknowledge, this can be done by syntacticizing the thematic hierarchy, in accordance to some version of the Uniformity of Theta Assignment Hypothesis (UTAH) (Baker 1997).

However, besides being ultimately stipulative and inheriting all of the woes of taxonomic theories of thematic relations (Dowty 1991), the THC – either as a primitive condition on semantic structure or as a descriptive corollary of UTAH coupled with a syntactic Binding Theory – is also empirically misguided. This suggests that neither the abstract structures invoked in UTAH-based approaches nor the independent semantic condition proposed by Jackendoff (1969; 1972) can be motivated by their role in accounting for URCs.5

In this section I will begin to sketch a unified account for the patterns observed above. Specifically, the goal will be to formulate a tentative generalization that captures unexpected reflexivization contrasts (URCs) (including (25)–(26)) in a way that bypasses most of the problems posed by (15)–(27). My goal is not to provide a detailed analysis for any of these cases in particular, but simply to state a descriptive regularity which will then motivate a theoretical solution for URCs as a whole. The first step involves recognizing a property that is shared by URCs: all of these cases involve, in their semantic representations, the intervention of proxy functions.7

Proxy functions are functions from entities to their contextually salient proxies – i.e. entities that are “representationally related” to their function’s input (Lidz 2001: 130) (see Jackendoff 1992; Hornstein 2001). For example, picture′ is a proxy function that, when applied to ⟦John⟧, yields picture′(john): the contextually salient picture of John.8 The function story′, when applied to ⟦Sue⟧, yields story′(sue): the story about Sue, an abstract kind of proxy. Proxies can, accordingly, be concrete or abstract, depending on the representational relation they are parasitic on.

A way to detect whether proxy functions have been applied is to exploit their effects on inference – it is mainly because of these that we need to incorporate them into our semantic representations. Given that there is an intimate connection between representations and intensionality (Jackendoff 1975; 1980; Searle 1983), proxy functions should be capable of introducing referentially opaque contexts. This is indeed what we observe in scenarios that give rise to URCs.

Jackendoff (1975: 60–62) offers examples like (28), which are only possible because the two mutually inconsistent properties (having blue vs. brown eyes) are, in fact, attributed to two different entities: one is predicated of a girl and the other of her representational proxy. This duality is a telltale sign of referential opacity.

Similar effects can be found in cases analogous to the implicit statue readings in (25)–(26). The sentence in (29) has a reading in which it is true, even if we take it to be a fact that Ringo Starr has blue eyes. This is so because the property of having brown eyes is attributed to a proxy (i.e. the value of a proxy function) which is distinct from plain Ringo.

The sentences in (30), which are reminiscent of examples in Russell (1905), are also not contradictory – as they would be if show and paint could not have intensional Themes. This suggests a proxy function (namely, image′) lurking in the semantic representations of URCs (3) and (8).

Moreover, about-complements – but not to, from or by-complements – have de dicto interpretations for which substitution salva veritate and existential generalization are disallowed. These are typical diagnostics for opaque contexts.

Consider the substitution test first. If we accept that Zak Starkey’s father (i.e. Ringo) is, in fact, the sole blue-eyed Beatle, all of the equivalences in (31) follow with the exception of (31a). One simply cannot infer from Don told Sue about Zak Starkey’s father that Don told Sue anything whatsoever regarding the eye-color of a Beatle. That sentence (unlike the one below it) could be true in a scenario where Don does not know that Zak’s dad has blue eyes, or even that he was in The Beatles.

Note that it is the presence of a de dicto reading (and not the absence of a de re one) which is relevant as a diagnostic for intensionality. The de dicto reading for (31a) asserts a relation between Don, Sue and information about Zak Starkey’s father. Therefore, the particular mode of presentation of the about-complement is truth-conditionally relevant: it matters whether Don described Ringo as Zak’s father or as the blue-eyed Beatle. There is no analogous de dicto reading for (31b)–(31d).

Existential generalization also sets about-complements apart from other P-complements. The only phrase in (32) which could be coherently uttered by an atheist is (32a), as it merely entails the existence of information about a god. (32b)–(32d) all imply that the speaker believes (or has come to believe) that a god exists.

I take that the predicate responsible for these opacity effects in about-complements (which appear in (4)–(7), (9a), (10a) and (12a)) is the intensional proxy function information′, which maps any entity x into a piece of information involving x.

Statue readings also block existential generalization over the depicted objects: (29) does not entail that a non-proxy brown-eyed Ringo exists in reality. This is part of what makes fiction possible: one can succeed in referring to a representation of an entity without being committed to the existence of that entity in the real world.9

Though judgements are less clear, stimuli arguments in psych-predicates like (11)–(13) seem to fail existential generalization as well. Contrast (35), where the inference to (33) clearly does go through, with (34):

This is expected if stimuli arguments (but not cause arguments such as the one in (35)) are interpreted under a proxy function like information′.10 This is roughly what Grimshaw (1990: Chapter 5) proposes in her analysis of these cases.

Proxy functions have various sources. Three of them will be relevant for URCs: the lexicon, type coercion and (enriched) composition. The first of these consists in cases where functions from entities to proxies are lexically encoded as the meanings of particular words. A likely example is the preposition about that appears in (4)–(7), (9a), (10a) and (12a). Montague (1974: 238), Dowty et al. (1981: 244) and Dalrymple et al. (1997: 264) suggest that this preposition is not semantically empty, but denotes an intensional operator of some sort. I will regard this operator as the aforementioned proxy function information′. If we take lexical items to be tripartite curryesque signs (Curry 1963; Jackendoff 1997; Culicover to appear), a lexical entry for about can be represented in the following attribute-value format (where alphabetical indices indicate correspondences):

Assuming, with Heim & Kratzer (1998) and Büring (2005), that reflexives correspond to individual variables in semantic representation, the logical syntax of a PP like about himself in (4a) is, thus, information′(x).

Proxy functions can also emerge as a result of type coercion: i.e. “a semantic operation that converts an argument to the type which is expected by a function, where it would otherwise result in a type error.” (Pustejovsky 1995: 111). A possible circumstance for type coercion occurs when predicates place specific contextual restrictions on their arguments. This arguably happens in some of the URCs above.

The predicate denoted by paint in (8), for example, requires that one of its arguments be an animate Agent and, the other, a kind of representation. Following Pustejovsky (2012: 18–19), I will tentatively represent these selectional features as lexically specified typing restrictions on arguments – a proposal reminiscent of McCawley (1968; 1971), who treated selectional restrictions as presuppositions. This requires a lattice of expanded types, where rep and anim are both subtypes of e. The following is a partial lexical entry for the active form of non-locative paint (as it appears in (8a)), employing the same attribute-value format as (36):

Since a reflexive pronoun does not lexically satisfy the subtyping restriction for being the Theme argument of paint′ (because its semantics is that of a mere e-typed variable), coercion is needed to avoid type underspecification in painted herself (cf. (8a)). The proxy function image′ comes to the rescue and maps the variable to the proper subtype, acting, thus, essentially, as an expression of type ⟨e, rep⟩. Likewise, if paint receives as its Theme argument a non-image individual like Sue, a proxy function has to intervene, yielding λy.paint′ (y, (image′(sue)). I assume something similar applies to the psych-predicates in (11)–(13), which also select instances of the representational subtype as their stimuli arguments.

Most typically, however, proxy functions come from the compositional procedure itself, taking into account how meanings are built up in interaction with their linguistic (and non-linguistic) context. The proxy function picture′ in (38) does not come about to coerce the variable into a subtype required by the verb – it is simply the contribution of the string that picture of to the meaning of the sentence:

If semantic composition can be enriched by inferentially relevant unarticulated constituents (cf. Pustejovsky 1995; Jackendoff 1997; Recanati 2010), the same can be said for statue scenarios like (25)–(26). In (39), for instance, even though there is no specific string in narrow syntax which encodes a proxy function – nor is a proxy function required to satisfy the subtyping restrictions on the arguments of admire′ – an unarticulated constituent corresponding to the proxy function modifies the variable and maps its referent into a statue, yielding (39b):

This also seems to be an adequate description for (3) (i.e. Yoko showed John himself in the mirror) and other cases involving transitive perception verbs (e.g. John saw himself in the mirror). As a reviewer notes, the Themes in these perception predicates are only circumstantially interpreted as images (i.e. as image′ (john)). This suggests that proxy functions are not lexically required by these predicates, but, in fact, emerge as optional unarticulated constituents, giving rise to intensional effects such as the one observed in (30a).

These “unarticulated proxy functions” which appear in (3), (25)–(26) and (39) are licensed by a construction which shifts the usual reference of an NP to an entity that is related to the NP’s literal denotation (cf. Nunberg 1979; 1995). What this construction does, in effect, is to embed the independent semantics of an NP into the scope of a proxy function P \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ {\cal P} \] \end{document} . This general proxy construction, which licenses the proxy function in (39), is stated as follows:

On this account, the proxy construction is licensed for any entity-denoting NP. In this respect, I differ from Hornstein (2001) and Lidz (2001), who take proxy functions in cases like (39) to be explicitly encoded by the -self morpheme. In my view, these proposals have the very counter-intuitive consequence of dissociating proxy readings like (39) from virtually identical readings which are licensed for non-reflexive NPs, such as pronouns and proper names: e.g. I saw Ringo in the wax museum. He looked very lifelike (see also Safir 2004: 113).

One interesting piece of evidence for this treatment of unarticulated proxy functions as parts of an NP construction is that controlled subjects and null objects do not allow proxy readings, as the contrasts in (41) and (42) show (cf. Lidz & Idsardi 1998: 122). This is expected if we take proxy readings to be the result of a proxy construction which is only licensed for overtly realized NPs.

Since both reflexives and their antecedents are NPs whose semantic type is e, they can always, if necessary, supply their meaning as an argument to a proxy function.

Proxy functions are, thus, a common property of the sentence pairs that participate in URCs. Regardless of where these functions come from (the lexicon, type coercion or semantic composition), the resulting predicate is always an instantiation of the schema in (43), where R \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ {\cal R} \] \end{document} is a meta-variable for n-ary predicates (for any n ≥ 2), P \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ {\cal P} \] \end{document} is a meta-variable over proxy functions, n is a meta-variable over indices and τ and υ are meta-variables over terms (individual constants or variables):

Given what I said about (39), it is clear how this works for the URCs in (3) and (25)–(26), repeated below:

Though less obvious, I contend that the lexical URCs in (4)–(12) (leaving (13) aside from the moment) are also instantiations of (43). For these cases, the proxy functions are either driven by type coercion or by simple functional application of information′, which, as I argued, is a proxy function lexically encoded by about. The sample representations below are based on proposals by Kuno & Kaburaki (1977), Grimshaw (1990), Jackendoff (1992), Pustejovsky (1995) and Safir (2004):

As suggested by the opacity tests in (30)–(34), the verbs in these URCs require that their Theme arguments belong to representational subtypes of e – hence the pervasiveness of proxy functions. In (50), (51b) and (52) the proxy functions come from type-coercion. In the other cases, the proxy functions are simply the result of applying the semantics of about (i.e. information′) to the P-complements.

As for (13), note that even though (53a) does not contain a proxy function in its semantics (since its subject is an extensional cause), (53b) does, because stimuli arguments are, as suggested by (34), construed as abstract kinds of proxies.

If URCs all involve proxy functions, we can formulate a more adequate generalization to replace the THC (cf. Jackendoff (1992) and Safir (2004) for similar ideas):

This generalization fully accounts for (44)–(53) and avoids many of the counterexamples mentioned in Section 2. No contrast is predicted for (17)–(19) (repeated as (55)–(57)), simply because the predicates in question have no proxy functions in their semantic representations, given the absence of opacity effects.

Even though (57b) shows that passives in general do not partake in URCs (and this militates against the THC), passives like (50b) do, since they involve proxy functions. The contrast here is between the passives of predicates with fully coextensional coarguments (cf. (58)), which are reasonably acceptable (perhaps requiring special accenting), and the passives of predicates whose Theme arguments are representational (cf. (59)), which are unequivocally bad:

This may be at the heart of what Lees & Klima (1963), Langacker (1969), Jackendoff (1969) and other early work perceived as problematic with reflexives within by-phrases. The problem is, however, more fine-grained than what the vocabulary of thematic roles can express: the passive alternation, which turns Themes into antecedents, does not yield universally bad results, given (57) and (58). URCs only emerge when these passivizable Themes are embedded within (intensional) proxy functions, as in (59). This is exactly what the generalization in (54) predicts.

But (54) is not without problems. First of all, the reciprocal and bound pronominals in (15)–(16) are still somewhat of a puzzle. Since (54) merely replaces the THC as a descriptive statement, it does not offer any compelling reason for why only reflexives fall under its domain. Moreover, it also fails to predict the acceptability profile of sentences parallel to URCs with first person and inanimate antecedents like (20)–(24). These problems can only be solved when (54) is dropped in favor of a more explanatory framework which is capable of retrieving the effects that (54) gets right. This is the task I’ll undertake in the remainder of this article.

In this section, reflexives in unexpected reflexivization contrasts (URCs) will be shown to share crucial properties with the class of so-called logophoric reflexives: namely, an animacy restriction, a generalized acceptability in the first person and an allowance of split antecedents. Recognizing these common traits, along with the role of proxy functions discussed in Section 3, will be decisive for motivating a general explanation for URCs that overcomes the flaws of generalization (54). The proposal in Section 5 entails that reflexives in URCs are, in fact, logophoric.

Recall that generalization (54) leaves out two of the scenarios that I raised, in Section 2, as objections against the Thematic Hierarchy Condition: one involving first person antecedents and the other involving inanimates. Contrasts between these problematic cases and the data that fall under (54) are rearranged below.

What is puzzling about (60)–(65) is that the acceptability contrasts are not due to proxy functions at all. Thus, (60)–(65) actually falsify generalization (54). In (60)–(63), the proxy functions all apply to antecedents, whereas in (64)–(65), they apply to reflexives. The only variable that changes within each pair is the nature of the antecedent: when the antecedent is the first person, the result is fine and, when it is inanimate, the result is bad, regardless of where the proxy function goes.13

The contrast between third and first person reflexives is reminiscent of data Ross (1970) presented as a motivation for his performative analysis. His examples explored contexts in which reflexives were only acceptable in the first person:

Many have also observed that these same contexts also yield contrasts between inanimate and animate antecedents for reflexives (Kuno & Kaburaki 1977; Golde 1999; Minkoff 2004; Postal 2006; Charnavel & Sportiche 2016):

Though judgments are less sharp, Postal (2006) also mentions focalized reflexives (72) and topicalized reflexives (73) as subject to the animacy constraint. The as for x-self reflexives in (67) and (70) are probably related to topics as well. Relevant contrasts are provided below (small caps indicate strong stress):

What has generally been assumed is that the reflexives in these contexts – like x-self, as for x-self, picture-NPs and focus/topic – are conditioned by the perspective or empathy assumed in the discourse (Lebeaux 1985; Kuno 1987; Zribi-Hertz 1989; Pollard & Sag 1992; Charnavel & Zlogar 2015 i.a.). This is why they are called, somewhat loosely, logophoric reflexives (Reinhart & Reuland 1993), in analogy to the logophoric pronouns of African languages (see Clements 1975). These reflexives must refer to entities whose viewpoints are represented within their discourse contexts. The following contrast, based on Pollard & Sag (1992: 274), makes this perspective-sensitive nature particularly clear.

The animacy constraint follows from the fact that bearing a perspective role in discourse – like John does in (74a) but not in (74b) – minimally entails animacy (see Minkoff 2004; Postal 2006; Charnavel 2019). The general acceptability in the first person follows from the speaker’s inalienable status as a possible – in fact, the default – perspective-bearer for every discourse (Sells 1987; Stirling 1993).

The two properties which were problematic for generalization (54) suggest, therefore, an observational correlation between logophoricity and the URCs data examined here. But the similarities between URCs and logophoric contexts do not stop there. Like the typical logophoric reflexives in (66)–(68), about complement reflexives are also (relatively) acceptable in the first person even in the absence of an overt antecedent (Reinhart & Reuland 1993: 715, Büring 2005: 233):14

Another relevant trait is the possibility of having split antecedents. It is known since Lebeaux (1985) that, unlike normal coargument reflexives, typical logophoric reflexives have this property:

Judgments like (77), also reported by Büring (2005: 233) and Charnavel & Sportiche (2016: 58), suggest that some of the reflexives in URCs might be similar to logophoric reflexives in this respect too – though Wasow (1979: 25) and Hornstein (2001: 155) report a case like (77b) as bad.

Reflexives in URCs share, thus, a number of crucial observable traits with logophoric reflexives. Since patterns in data invite the inference to a common natural kind, this already constitutes a weak motivation for saying that reflexives in URCs are logophoric. Charnavel (2019), however, argues for a stronger claim: given that animacy is a necessary condition for a reflexive to be logophoric (due to the nature of perspective), inanimate reflexives are always non-logophoric. From this it follows, conversely, that animate reflexives that cannot be replaced by inanimate ones are necessarily logophoric. What this means is that a restriction to animate antecedents can function effectively as test for identifying logophoric reflexives.

Therefore, according to Charnavel (2019), the properties mentioned above – in particular, the animacy restriction – are enough to count reflexives in URCs as logophoric. If this is true, a general solution to URCs comes for free: Since the antecedents of the bad cases in (44)–(53) are mapped into inanimate proxies (e.g. images, statues, information), they are not genuine perspective bearers. They can’t therefore, be linked to the types of discourse roles that license logophoric reflexives (see Sells 1987; Zribi-Hertz 1989; Golde 1999; Minkoff 2004; Charnavel 2019). These sentences would be bad for the same reason as (69)–(73).

The general acceptability of first person reflexives in URCs would also be predicted on these grounds, because the speaker is always independently available as a logophoric antecedent. This entails that the speaker does not lose its capacity to license a logophoric reflexive, even when a pronoun which refers to it is embedded within a proxy function which maps its referent into an inanimate proxy. The speaker’s omnipresence makes it immune to the effects of generalization (54).

But how can we make sense of this theoretically? Taking note of the phenomenal identity between URCs and logophoric contexts (as evidenced by their common properties discussed in this section) does not replace a theory of when and how logophoric readings emerge. In this respect, it seems puzzling to think of reflexives in URCs as instances of logophoric reflexives, because these are usually conceptualized as non-local, or as exempt from the locality embodied in CA (see Pollard & Sag 1992; Reinhart & Reuland 1993, i.a.).

However, this conception of logophoric reflexives is fraught with problems, as others have also argued (Dalrymple 1993; Menuzzi 1999; Charnavel & Sportiche 2016).15 In Section 5, we will see that these issues can be overcome if we treat logophoric reflexives not as exempt from CA but as but as tolerable violations of it. A similar theory was proposed, for independent reasons, by Menuzzi (1999). Besides accounting for the data, this view of logophoric reflexives enables a simplification of CA, which, in fact, turns it into a basic lexical property of reflexives. Insofar as Section 5 succeeds in presenting an independently motivated theoretical characterization of logophors which includes URCs, the logophoric properties laid out in this section will be explained (or, at least, be made less mysterious).

This section is divided in two parts. In the first, I review some of the rationale that supports a reconceptualization of logophoric reflexives along the lines of Menuzzi (1999; 2004). The resulting alternative falls within the reflexivity-and-chains approach (Reinhart & Reuland 1993), but differs from its previous implementations in crucial respects. In the second part, I proceed to explain how this new view of logophoric reflexives extends to unexpected reflexivization contrasts (URCs), which were not hitherto addressed within the reflexivity-and-chains framework.

This article begun by exploring a set of unexpected reflexivization contrasts (URCs) which have been problematic throughout the history of Generative Grammar. A popular proposal based on a (semantically primitive or syntactically derived) thematic hierarchy was reviewed and dismissed in virtue of some empirical drawbacks. Given these results, the alleged effects of the thematic hierarchy must be treated as epiphenomena of a more fine-grained interpretive factor.

I followed the thread initiated by Grimshaw (1990), Jackendoff (1992) and Safir (2004) in claiming that URCs have a peculiar semantic property: they all involve predicates which, for some reason or other, have proxy functions applied to one of their semantic arguments. I argued that this constitutes a logophoric context (i.e. a context in which CA cannot be fulfilled). Reflexives are nonetheless mandatory, because pronouns in these configurations violate the CC. But, since CA fails, the only way to interpret reflexives in URCs is to resort to a non-literal interpretive strategy – i.e. to what I called the Logophoric Strategy (LS). This, however, is only possible for cases in which the antecedent is a perspective bearer – that is, when it is the reflexive (and not the antecedent) which is linked to a proxy. When the antecedent is mapped into an inanimate proxy, the LS fails to provide an interpretation for the reflexive and the result is invariably bad.

There is still some vagueness in my formulation of LS. I did not provide, for example, a fully explicit characterization of anti-reflexive predicates or of how logophoric roles are attributed to discourse referents. Most of these loose ends can only be overcome by embedding the analysis developed here into a more formally rigorous framework. I have tried, as much as possible, to make my proposal neutral with respect to many details of formal implementation. For instance, I have not addressed here whether logophoric roles should be represented in narrow syntax (Ross 1970; Speas 2004; Sundaresan 2018; Charnavel 2019) or in separate Discourse Representation Structures (Sells 1987; Stirling 1993; Huang 2000).

Regardless of these issues, if my analysis ultimately works, it derives the conflicting judgements in URCs from more principled and independently motivated notions than previous approaches. If reflexives in URCs are logophoric – as many reflexives in many languages are –, URCs are simply explained on the basis of what counts as a possible logophoric antecedent. Moreover, this account embeds this explanation within a robust overall theory of anaphora: the reflexivity-and-chains approach of Reinhart & Reuland (1993).28