Perspective is syntactic : evidence from anaphora

This paper argues that grammatical perspective, expressed along the spatio-temporal and mental dimensions, has a syntactic component. Evidence for this is provided from non-local anaphora in the Dravidian language Tamil which is perspective-driven: i.e. the antecedent of a successfully bound anaphor in Tamil must denote a mental or spatio-temporal perspective-holder toward some predication containing this anaphor. I will argue that, in Tamil, the agreement marking that obtains on the clausemate verb of the anaphor, when this anaphor occurs in nominative case, seems to be anomalously triggered, not by the anaphor or by its antecedent, but by a silent perspectival pronoun local to the verb. Assuming that agreement is a morphosyntactic process, such a thesis, if correct, then entails that perspective must be syntactically (i.e. structurally and featurally) instantiated. Based on such evidence, I propose that perspectival anaphora is a composite consisting of variable-binding + discourse-pronominal reference at two distinct stages of grammar. Empirical evidence for such a model comes from the (seemingly) schizophrenic pronominal and bound-variable nature of such dependencies, diagnoseable by the usual syntactic and semantic tests.


Introduction
The goal of this paper is to argue that perspective, expressed along the mental and spatio-temporal dimensions, is syntactically represented and can, as such, drive syntactic dependencies.To this end, I present evidence from a linguistic phenomenon where grammatical perspective has long been observed to play a central role -namely, non-local anaphora (a cover-term not only for long-distance anaphora and backward anaphora but also for logophora).I refer to this class of items as "perspectival anaphora": (1) Definition of perspectival anaphora: In every instance of perspectival anaphora, the anaphor is properly contained within a predication which is evaluated relative to the perspective, mental or spatial, of some sentient individual.This individual must be aware of the eventuality described by this predication, at the time it happens.The antecedent of the anaphor must denote this individual.
The evidence that I provide comes primarily from Tamil, a language of the Dravidian family, spoken predominantly in South India.In this paper, I will argue that the agreement marking that obtains on the clausemate verb of the anaphor in Tamil, when this anaphor occurs in nominative case (the case that generally feeds agreement), seems to be anomalously triggered, not by the anaphor or by its antecedent, but by a silent pronoun (pro), local to the verb, introduced in the specifier of a perspectival head, Persp.Assuming that agreement is a morphosyntactic process, such a thesis, if correct, then leads to the conclusion that perspectival information must be visible early enough to drive this morphosyntactic process: i.e. this perspectival information must itself be syntactically (i.e.structurally and featurally) represented.
Building on this idea, I will additionally propose that this pronoun plays a central role in deriving the perspectival nature of anaphoric dependencies in Tamil, proposing that it mediates the relationship between the anaphor and its antecedent, coreferring with them in different ways.The pronoun's relationship with the anaphor is distinguished by its being local: as such, it Agrees with the anaphor in syntax, which triggers binding at LF.However, the antecedent of the anaphor is not local to the anaphor or to the pronoun: thus, the pronoun-antecedent relationship is a discourse-pragmatic one (essentially just pronominal reference) that is not structurally constrained in any way.The anaphor and its antecedent thus corefer only by transitivity -purely by virtue of their independent referential relationships with pro.This is illustrated below: (2) Two stage model of perspectival anaphora: Discourse-pragmatic coreference

Syntactic Agree + LF Binding
Non-local anaphora is infamous for its hybrid syntactic-pragmatic behavior which resists a unified analysis: certain properties, like the crosslinguistically robust antilocality constraint on anaphoric antecedence, suggest that the dependency is structurally regulated; but yet others, like the fact that the antecedent of the anaphor need not c-command the anaphor, or that minimality restrictions on antecedence are not obeyed, or the non-locality itself, or the fact that discourse-pragmatic factors such as perspective or empathy govern choice of antecedent, suggest that structure does not play a role after all.A two-stage model of non-local anaphora such as the one I propose here, with one stage being purely formal/structural, and the other being discourse-pragmatic, derives this dual nature in a unified manner.This model can potentially also be extended to other languages with perspectival anaphora like Icelandic (Hellan 1988;Sigurðsson 1991;Reuland 2001), Italian (Bianchi 2003;Giorgi 2006;2010), Japanese (Kuno 1987;Oshima 2004;Nishigauchi 2014) Norwegian (Hellan 1988;Lødrup 2007), Abe (Koopman & Sportiche 1989), French (Charnavel 2017), and Ewe (Pearson 2013), among others.Toward the end of the paper, I provide independent evidence for the existence of a perspectival pro with a mediating role such as that described above.If this proposal is correct, the anaphor-antecedent relationship should display the empirical fingerprint of pronominal reference, rather than anaphora.At the same time, the anaphor should itself not behave like a regular pronoun, but like a bound variable.I show that both these predictions are fulfilled using empirical diagnostics like split antecedence tests and bound variable vs. strict readings under definite DPs on the one hand, and antilocality effects under reflexivity and structural constraints on binding domains, on the other.
2 Background: Perspectival anaphora I use the moniker "perspectival anaphora" as a cover-term for all (nominal) anaphoric dependencies that are regulated by their sensitivity to grammatical perspective, defined along the mental and spatial dimensions.Below, I present some of the background on per-spectival anaphora and also discuss why this phenomenon has long posed such a unique challenge for generative linguists seeking to provide a unified analysis for its curious medley of discourse-pragmatic and syntactic properties.

Perspectival anaphora: Core properties
In the realm of anaphora, the notion of perspective is perhaps typically invoked in the context of logophoric dependencies, the term "logophor" denoting a designated pro-form referring to an entity "whose speech, thoughts, feelings, or general state of consciousness are reported" (Clements 1975: 141).
(3) illustrates this for Tuburi, a Chadic language (Sells 1987: 447): the plural logophor sā:rā represents the mental perspective of the sayer denoted by the matrix subject "they": (3) À (rínɡ) wò gā tí sā:rā t∫Í sā:rā.pro (say) pl comp head log hurt log 'They i said [ CP that they i had headaches].' The term has since been appropriated to refer to dependencies where the anaphor corefers with an extra-sentential nominal that denotes a discourse-salient individual from whose "inner mind" the narrative is reported, as in the free indirect discourse scenario (Banfield 1982;Schlenker 2004) from Jane Austen's Emma (Austen 1816 Chapter XVIII, 321) in (4): (4) "With Tuesday came the agreeable prospect of seeing him again, and for a longer time than hitherto; of judging of his general manners, and by inference, of the meaning of his manners towards herself; of guessing how soon it might be necessary for her to throw coldness into her air …" "Long-distance anaphora" -i.e.dependencies involving bound variable pro-forms that are anteceded by another nominal in the same sentence (albeit, crucially, not in the same local clause) -can also be similarly perspectival.This is illustrated by the striking contrast in the Icelandic sentences below (taken from Reuland 2001: 345): (5) Barnið i lét ekki í ljós [að það hef-ði verið hugsað vel um sig {i,*j} ]. child.defput not in light that there had-sbjv been thought well about anaph '[The child] i didn't reveal [ CP that she {i,*j} had been taken good care of].' (6) *Barnið i bar þess ekki merki [að það hef-ði verið hugsað vel um child.defbore of it not signs that there had-sbjv been thought well about sig i ]. anaph '[The child] i didn't look [ CP as if she i had been taken good care of].' Reuland (2001: 345), describing the sentences in ( 5)-( 6), reports that: "The difference in acceptability between [(5)] and [(6)] can be attributed to the fact that in [(5)] the report is made from the child's point of view, i.e., it is the child, and not the speaker, who didn't reveal that he/she had been taken good care of, whereas in [( 6)], it is the speaker who reports that the child didn't look as if he/she had been taken good care of." The role of mental perspective in long-distance anaphora has been observed for a range of other languages (see e.g.Koopman & Sportiche 1989;Pearson 2013;Kuno 1987;Oshima 2007;Bianchi 2003;Giorgi 2010;Jayaseelan 1997 for data and discussion on Abe, Ewe, Japanese, Italian, and Malayalam, respectively).Anaphoric dependencies may be governed by their sensitivity to spatial perspective, as well, as in Norwegian where "the simple reflexive [seg] is used when the physical aspect of the referent of the binder is in focus" (Lødrup 2007: 183; see also Rooryck & vanden Wyngaerd 2011 for data and discussion on the role of spatial perspective in Dutch anaphora).This is nicely illustrated in the pairing below, involving the preposition mot which is homophonous between a spatial vs. a more abstract, non-spatial meaning: (7) mot (toward, against): a. Han i dra-r den mot seg {i,*j} .He [nom] pull-prs it towards anaph 'He i pulls it towards himself {i,*j} .' b.Forbrukerråd-et i argumentere-r mot [seg selv] {i,*j} consumer.council-defargue-prs against anaph self.'[The consumer council] i argues against itself {i,*j} .' The simplex seg form is used only when the preposition is interpreted as spatial, and its antecedent is obligatorily interpreted as the spatial perspective-holder with respect to the spatial PP containing the anaphor.In all other cases, seg selv is used, making this form the elsewhere case.In the rest of the paper, I label all pro-form dependencies where the antecedence of the pro-form is regulated by perspective, in the manner illustrated above, as instances of "perspectival anaphora" and the pro-form in question in each case as a "perspectival anaphor".
Finally, it has been pointed out that perspectival anaphors in languages like Japanese (Kuno 1973;Nishigauchi 2014) are subject to an "awareness condition".Kuno (1973: 322) describes that the Japanese anaphor zibun in a subordinate clause may corefer with another nominal in the matrix only if the former "represents an action or state that the referent of [the nominal] is aware of at the time it takes place or has come to be aware of at some later time."Thus, Japanese zibun is licit in (8) where Takasi is aware of the election happening, but not in (9), where he is asleep, and therefore not (Nishigauchi 2014: 167, Exx. 24-25): (8) Iinkai-ga zibun i -o erab-I soo ni nat-ta toki, Takasi i -wa committee[nom] anaph-acc elect likely become-pst when Takasi-top huanni nat-ta.worried become-pst 'When it came to be likely that the committee might elect self i , Takashi i became anxious.' (9) *Iinkai-ga zibun i -o erab-i soo ni nat-ta toki, Takasi i -wa committee[nom] anaph-acc elect likely become-pst when Takasi-top gussuri nemut-te-i-ta.fast asleep be-pst 'When it came to be likely that the committee might elect self i , Takashi i was fast asleep.' It is also important to note that the individual denoted by the antecedent of a perspectival anaphor doesn't need to actually hold a (mental or spatial) perspective with respect to the predication containing the anaphor.Rather, the predication containing this anaphor must be evaluated (or determined) relative to the perspective of this individual.To appreciate the distinction, observe that it is possible in English to bind an anaphor under a negated attitude verb, as in (10), below:1 (10) (Due to her advanced Alzheimer's) Susan i doesn't realize yet that there's a letter from herself {i,*j} , from 40 years ago, that will be opened on her i 90th birthday.
In ( 10), Susan has no realization, thus no perspective, on the fact that there's a letter from herself.Yet, her perspective is still involved in evaluating the predication containing herself.Similar arguments can be made for anaphora that is regulated by spatial perspective (for discussion, see Levinson 2003;Kracht 2008;Barlew 2016).
Given these considerations and the rest of the discussion above, I define perspectival anaphora as in (11) below: (11) Definition of perspectival anaphora: In every instance of perspectival anaphora, the anaphor is properly contained within a predication which is evaluated relative to the perspective, mental or spatial, of some sentient individual.This individual must be aware of the eventuality described by this predication, at the time it happens.The antecedent of the anaphor must denote this individual.

Structural vs. pragmatic approaches to perspectival anaphora
There is a fundamental analytic tension in the literature between conceptual and structural approaches to perspectival anaphora.The perhaps more traditional conceptual view is motivated by considerations like: (i) a tacit assumption that discourse-pragmatic notions like "perspective" do not belong in the domain of syntax proper but are, in some sense, peripheral to it; and (ii) the observation that perspectival anaphora seems to violate cornerstones of structural wellformedness (in generative frameworks like GB and Minimalism), making a syntactic analysis seem in turn rather far-fetched.To elaborate on the latter, in sentences like Icelandic (5), the antecedent is not local to the anaphor; since syntactic relationships are held to be fed by locality, such structures pose a nontrivial challenge.In multiply embedded sentences, the anaphor may be anteceded by a nominal across another one that is closer to it, in apparent blatant violation of Relativized Minimality (Rizzi 1990), another structural wellformedness condition.In such sentences, there is often also more than one individual that satisfies the perspectival conditions laid out in (11), thus more than one potential antecedent; the choice of antecedent in such cases is thus also indeterminate, which violates the idea that syntactic derivations yield a deterministic output.In so called "backward binding" constructions (Minkoff 2003), which occur in psych predications, the antecedent, which takes on the role of experiencer in the psych predication, doesn't even c-command the anaphor on the surface -as shown for Italian (12) (Giorgi 2006) and English (13) (Minkoff 2003): (12) La-propria {i,*j} moglie preoccupa molto Gianni i .self's wife worries a lot Gianni 'Gianni i is worried by self's {i,*j} wife.' (13) That slanderous article about herself i tipped Sue i over the edge.
Finally, the problem with logophoric relationships like that in English (4) above is, if anything, even more challenging.Here, the antecedent of the perspectival anaphor is extra-sentential which poses a non-trivial challenge to a model of syntax that can only deal with sentence-bounded dependencies.
The obvious solution, given these challenges, would seem to be to derive perspectival anaphora through purely non-structural, discourse-pragmatic means.The problem is that the role of structure cannot be dismissed entirely: its relevance already makes itself known in one interesting way.There is a robust and systematic antilocality effect observed with most perspectival anaphors crosslinguistically.That is, long-distance anaphors (so called "se" anaphors in the Reinhart & Reuland 1993 parlance), many of which show perspectival properties of the kind discussed here, resist being bound reflexively (i.e. by a co-argument of the verb).Given that locality is a structural concept, the sensitivity to locality entails sensitivity to structure, by transitivity. 2 Motivated in part by such observations, 3 the structural view within the generative framework (e.g. the movement approach in Chomsky 1986a; Pica 1987;Huang & Tang 1991 and the Relativized Subject hypothess in Progovac 1993) argues that the involvement of perspective in anaphora is syntactically implemented.The other kind of argumentation for a structural treatment is a weaker one -namely that, in many cases, perspectival anaphora cannot be understood discourse-pragmatically. Koopman & Sportiche (1989) argue that perspectival anaphora in the Kwa language of Abe must be syntactically implemented because the types of verbs that select logophoric complements cannot be straightforwardly distinguished in terms of their lexical meaning.Rather, they all have the property that they select a clause with a particular kind of overt complementizer.Sells (1987) and Baker (2008) conclude the same, based on similar types of data from languages like Tuburi and Slave, respectively.Of course, underlying this type of reasoning is again the premise that discourse-pragmatic sensitivity and structural sensitivity are mutually incompatible.
In contrast to these theories, I will propose to make sense of this dual nature of perspectival anaphora by developing a model that exploits both structural and discoursepragmatic aspects of grammar, interacting in a sequential derivation.In particular, I will argue that every instance of perspectival anaphora involves two types of dependency: a structural (i.e.syntactic and LF-semantic) one involving anaphoric binding by a perspectival null pronoun (pro) and a discourse-pragmatic one, building on this, involving coreference between the anaphor's antecedent and pro (see Nishigauchi 2014;Charnavel 2017 for similar proposals).

Perspectival anaphora in Tamil: A (very!) quick primer
Here, I show that non-local anaphora in Tamil is indeed perspectival, as defined in (11) and that it displays the hybrid syntactic-pragmatic properties described above for perspectival anaphora crosslinguistically.
2 An anonymous reviewer notes that long-distance anaphors that display such properties are, first, not all perspectival (see e.g.discussion of long-distance anaphora into infinitives in Reuland 2011) and, second, that purely structural analyses for such phenomena already exist, which do not appeal to their perspectival properties (e.g.Reuland 2001; 2011 would derive this as a function of the monomorphemicity of the anaphor).With respect to the first point, the current paper has nothing to say: the focus of this paper is on the specific class of anaphora I am calling "perspectival anaphora" which behave distinctly from standard anaphora in a number of (other) respects.With respect to the second, I will propose, at the end of this paper (in Section 6.2.1) an analysis for the antilocal behavior of perspectival anaphora and argue that the data presented here cannot be easily accommodated by purely structural approaches like Reuland (2011). 3Another consideration is so-called "subject orientation": the idea that perspectival anaphors must be anteceded by syntactic subjects and not objects.However, robust crosslinguistic empirical evidence from expletive and non-sentient subjects (which cannot antecede such anaphors) and experiencer objects in psych predications (which can), among others, has shown that subject-orientation is neither a necessary nor sufficient condition for antecedence of such anaphors (see Jayaseelan 1997;Giorgi 2006; Sundaresan 2012 a.o.for discussion).

Anaphora in Tamil is perspectival
Here, I focus on the properties of the Tamil anaphor taan, a morphologically simplex form whose basic case and number paradigms are given in Table 1. 4 Ta(a)n can only take 3rdperson antecedents (gender irrelevant),5 as shown in (14 The anaphor ta(a)n in Tamil co-exists with other pro-forms (I classify these as pronouns) which differ from it in being able to refer deictically.Consider (15) below: (15) Raman i tann-ooɖae i eɖaedŭ-pakkattŭ-lae irŭ-nd-ae paamb-ae ko-nn-aan.
Raman anaph-dat left-side-loc be-pst-rel snake-acc kill-pst-3msg 'Raman i killed the snake that was to his {i,*j} left.' The obligatorily non-deictic nature of ta(a)n can be illustrated by comparing its behavior across the minimally contrasting discourse-scenarios below:  Native speakers judge (15) ungrammatical in the scenario in ( 16) where there is both pointing and the spatial relations are reversed with respect to Raman's spatial coordinates (the snake is to Raman's right, not to his left).But ( 17) is judged perfectly acceptable under the discourse scenario in (1), where there is no pointing and the leftness of the snake is evaluated relative to Raman.8A systematic difference arises when one contrasts (15) with a minimally varying sentence containing a deictic pronoun instead of the anaphor, as in ( 18): (18) Raman i avan-ŭkkŭ {i,j} eɖaedŭ-pakkattŭ-lae irŭ-nd-ae paamb-ae ko-nn-aan.Raman he-dat left-side-loc be-pst-rel snake-acc kill-pst-3msg 'Raman i killed the snake (that was) to his {i,j} left.' In (15), the "left-ness" of the snake is evaluated from Raman's spatial perspective; in (18), however, this leftness is evaluated from the spatial perspective of the (utterance-context) speaker or is underspecified (with respect to the perspective of the speaker vs. Raman).But the anaphor ta(a)n may only be licitly used in (15), where the spatial perspectiveholder is the individual denoted by the antecedent and the antecedent alone.These examples show that perspective-holding plays a central role in regulating anaphoric dependencies in such languages.
The perspective-sensitivity of anaphora along the mental dimension, e.g. in attitude contexts, in Tamil, can be illustrated by its interaction with other perspective-sensitive elements, like epithets.An epithet occurring in the scope of an attitude holder cannot denote that attitude-holder (Dubinsky & Hamilton 1998 a.o.): it is thus anti-attitudinal.Thus, if mental perspective-holding regulates anaphora in Tamil, an anaphor in the scope of an attitude verb should not only be able to denote the attitude-holder, it should also be unable to corefer with an epithet in the scope of that attitude verb (see Charnavel 2017 for parallel tests in French).The sentence in ( 21) is unacceptable under the discourse scenario in (19), where it is understood that the epithet andae muʈʈaaɭ ('that idiot') denotes the attitude-holder Sri.But it is acceptable under the discourse scenario in (20), where andae muʈʈaaɭ ('that idiot') doesn't denote the attitude-holder Sri, but his son.This shows that the epithet is anti-attitudinal: Andae muʈʈaaɭ neʤamaa-vee school-ae viʈʈ-aan-aa?that idiot[nom] really-emph school-acc leave-pst-3msg-q 'Had that idiot really dropped out of school?' Taan i andae muʈʈaaɭ-ae {*i,j} patti kanavŭka-ɳɖ-aan-aa?anaph.nomthat idiot-sg.accabout dream-pst-3msg-q 'Had he i dreamed about [that idiot] {*i,j} ?' Now, consider a sentence like (22) which contains both an anaphor and an epithet in the scope of an attitude verb (used in a free indirect discourse scenario).The logophoric ta(a)n must denote the attitude-holder (Sri) and is also obligatorily disjoint from andae muʈʈaaɭ ('that idiot').As such, it is licit with the discourse scenario in (20) but incompatible with that in (19).
Perspectival anaphora in Tamil obtains "long-distance" (across multiple clauses -modulo processing, the actual distance doesn't matter), logophorically, and in psych predications (yielding backward-binding structures involving a non-c-commanding experiencer antecedent).In all these structures, it can be shown with respect to diagnostics like those above that anaphora is perspectivally regulated along the mental or spatial dimensions.Additional supporting evidence comes from the fact that there is an animacy constraint on anaphoric antecedence: this follows naturally if antecedence is perspectivally regulated (see also Sundaresan & Pearson 2014 for further discussion and formalizations of this constraint for perspectival anaphora).I will thus take it to be uncontroversial that anaphora in Tamil is perspectival in the sense defined in (11) above.
Perspectival anaphora in Tamil is also subject to the awareness condition described for Japanese above.Thus, (24), analogous to Japanese (8), is licit under the discourse scenario in (23).However, ( 26) is illicit under this discourse scenario.However, the sentence becomes felicitous again, when ta(a)n is replaced with a coreferent (honorific) pronoun, as in the minimally varying sentence in ( 25 anaph.nomelect-do-inf come-when Raman[nom] tuuŋg-i-kkoɳɖ-iru-nd-aar.sleep-pst-asp-cop-pst-3hon.msg'When the committee was about to elect him i , Raman i slept.'Finally, ta(a)n may be licitly bound under a negated attitude verb.Thus, (28) may be felicitously uttered by the reporting journalist in the discourse scenario in (27): 10 9 This modification was actually suggested by a native speaker I consulted on the acceptability of (26).This speaker said that the sentence sounded bad to his ear because the person uttering it was the journalist (and not Raman): since Raman himself was asleep, he could have had no knowledge of this eventuality.He suggested that I replace the anaphor with the regular pronoun avan ('he') instead, which would then licitly allow coreference with Raman in such a scenario. 10An anonymous reviewer notes that the awareness condition may be too strong given the licitness of binding under negation in such sentences.When I asked the same native speaker from Fn. 9 why a sentence like (28) was licit given the explicitly stated lack of awareness, he responded that ta(a)n is licit in this sentence (corroborated by another native speaker) because Raman is aware that "he has two possibilities, to win or not to win, even if he doesn't know which one turned out to be true.The discussion here shows that anaphora in ta(a)n is perspective-sensitive, that it is regulated by sensitivity to spatial as well as mental perspective and, more specifically, that the role of perspective in anaphora in this language is as defined in (11).

Dual syntactico-pragmatic behavior
It was noted in Section 2.2 that perspectival anaphora crosslinguistically is characterized by a hybrid mixture of structural and discourse-pragmatic properties.Here, I present evidence to show that perspectival anaphora in Tamil exhibits the same behavior in this respect, as well.
Below, I show that long-distance anaphors in Tamil violate locality and, frequently, minimality and antecedence determinacy.In (29b), Krishnan antecedes ta(a)n across several other DPs that are structurally closer to the anaphor, at least one of which (namely Raman) also readily qualifies as a potential antecedent to it.Being two clauses higher, Krishnan is also clearly non-local to the clause containing ta(a)n.Thus, (29b) attests to apparent violations of non-locality and non-minimality and also to antecedence optionality.The latter is more clearly illustrated in (30c): here, either Krishnan or Raman may antecede ta(a)n as the referential indices indicate.Backward binding structures involving psych predications show us apparent violations of c-command: in (31b), Raman can antecede ta(a)n despite being embedded as a possessor DP inside the experiencer -thus clearly not c-commanding the anaphor.Finally, logophoric dependencies such as that illustrated in (33) show that the antecedent doesn't need to be in the same sentence as the anaphor -but can be elsewhere in the salient discourse: b.Antecedent: extra-sentential (logophoric) Seetha has had a string of bad luck lately.On an especially cold winter evening, she is feeling particularly sorry for herself.Her thoughts run along the lines of (33).
Seetha-vŭkkŭ i oɳɳum purija-lae.Taan {i,*j} maʈʈum een ivvaɭavŭ Seetha-dat anything understand-neg.anaph.nomalone why this.much kašʈappaɖa-ɳum?suffer-must 'Seetha i didn't understand at all.Why must she {i,*j} alone suffer this much?' The larger take-home message from these empirical patterns is the same as before: such structures pose a genuine challenge to analyses that seek to derive these anaphora through purely structural means.But here again, as before, the role of structure cannot be dismissed out of hand.As has been noted elsewhere (see e.g.Schiffman 2006;Annamalai 2000), Tamil ta(a)n cannot be locally bound as is, without something extra, specifically a verbal suffix kol on ta(a)n's clausemate verb, which is often classified as a kind of middle marker (see Sundaresan 2016), being added:11 (34) *Raman i tann-ae i paar-tt-aan.Raman[nom] anaph-acc see-pst-3msg Intended: 'Ramani saw himself i .' A similar situation seems to hold in the related Dravidian language Kannada, as discussed in detail in Lidz (2001;2004: et seq).I discuss perspectival reflexives at the end of this paper in detail and try to derive the antilocality in terms of the model of perspectival anaphora developed here.At this point, it suffices to note that the mere existence of this pattern suggests that a structural restriction (some form of antilocality) is at work.
To sum up, then, we are left with the same mixed bag of properties in the case of perspectival anaphors in Tamil, as we were with the others: i.e. dependencies involving seemingly unruly syntactic behavior that nevertheless show sensitivity to structure (specifically to syntactic locality) and are simultaneously regulated by their sensitivity to discourse-pragmatic perspective.

Insights from Tamil verbal agreement
This section presents and discusses the core data of the paper.The main goal is to argue, on the strength of evidence from verbal agreement triggered in the scope of the nominative anaphor taan in Tamil, that grammatical perspective is represented in the syntax, in the form of a silent pronoun (or pro), in the same local perspectival predication as the anaphor.This pro is visible to, thus can participate in syntactic processes, including but not limited to anaphora.On the strength of this conclusion, I will propose a two-step model of perspectival anaphora whereby only one stage of the perspectival anaphoric dependency is instantiated in the syntax proper, involving an Agree relation between pro and the anaphor in syntax which in turn triggers binding at LF.The perspectival pro, not the antecedent, is thus the true binder of the anaphor.The second stage of the process is not implemented in the syntax at all, but at the broader interpretive and discoursepragmatic levels, and involves discourse-pronominal coreference between pro and the antecedent.Such a model allows us to elegantly capture the hybrid syntactico-pragmatic nature of perspectival anaphora in Tamil and other languages, described above.

Verbal agreement under ta(a)n in Tamil
In Tamil, verbal agreement for person, number, and gender (i.e.φ-agreement) is typically triggered by a local nominal in the nominative.Thus, in (35), the matrix verb reflects 3msg agreement and is triggered by the nominative pronoun avan ('he') whereas the embedded verb, marked 2sg, matches the 2sg features of the embedded nominative subject nii ('you').In (36), the embedded subject has been changed to avaɭ ('she') and reflects 3fsg features on its clausemate verb: [Avaɭ j paris-ae tookkapoo-r-aaɭ-ŭnnŭ] avan i namb-in-aan.she[nom] prize-acc lose.go-prs-3fsg-comphe[nom] believe-pst-3msg 'He i believed [ CP that she j would lose the prize].'But when the nominative nominal is the anaphor taan, the agreement on its clausemate verb varies in an interesting way.The sentence in ( 38) is licit only under the discourse scenario in (37a); ( 39) is licit only under the discourse scenario in (37b): (37) Maya and Raman are the two final contestants at a music competition.Maya ends up winning the contest, and the prize.Maya later shows her two sons that Raman believed all along that: a. she (Maya) would actually lose the prize.I can report this as in: (38), but (39).b. he (Raman) would actually lose the prize.I can report this as in: (39), but (38).
( When the intended antecedent is avaɭ ('she') (as in ( 38)), the agreement under ta(a)n is 3fsg.(39) varies minimally from (38) with the only difference lying in the choice of antecedent for ta(a)n -the medial subject avan ('he') instead of the matrix subject avaɭ ('she').Here, the verbal agreement under the anaphor tracks this choice, with the agreement changing to 3msg in (39).In (40), ta(a)n refers logophorically to the extra-sentential attitude-holder koɻaendae ('child') which triggers neuter agreement on its clausemate verb; although ta(a)n is in a different sentence, the agreement triggered under it must still reflect the φ-features of this antecedent: if koɻaendae were replaced by avan ('he'), the agreement-marking in the following ta(a)n-sentence would be 3msg -aan instead.The agreement patterns above thus suggest the following: (41) Antecedence tracking generalization: Nominatives trigger agreement in Tamil.
When the anaphor ta(a)n occurs in the nominative, the agreement on its clausemate verb tracks the antecedent of ta(a)n.

Agreement is not triggered by the antecedent
An obvious candidate for the source of agreement, given the antecedent-tracking effect of the agreement, given in (41), is the antecedent of the anaphor.Following e.g.Kratzer (2009) and others, we might propose that this is a case of φ-feature transmission from the antecedent to the embedded verb in the ta(a)n-clause (perhaps cyclically, via intermediate functional heads).
An immediate, potentially fatal problem for this view is that, in Tamil perspectival anaphora, the antecedent may be several clauses away, need not c-command the ta(a)n-clause and, in structures involving the logophoric use of ta(a)n, is extra-sentential (cf.( 29b)-(30c)).Further evidence against such an account comes from seemingly mismatched agreement in sentences like ( 42 (42) obtains under tightly constrained structural conditions, specifically only in the clausal complement of a speech predicate.The anaphor ta(a)n is its nominative subject and it takes an antecedent, the matrix subject Raman, which has 3msg features, and triggers 3msg agreement on the matrix verb.But the φ-agreement on the clausemate verb of ta(a)n is 1sg.A feature-transmission account cannot explain (much less derive) the mismatch between the features of the antecedent and those on the embedded verb.
( What the contrast above shows is that the thematic properties and structural position of the antecedent directly affect the nature of agreement on the embedded verb.The antecedent must be the agent of a speech predicate and, furthermore, must be an argument of the clause that directly selects the ta(a)n clause as its complement.
Additional supporting evidence for this same point comes from number marking on the verb.When the agent of the selecting speech predicate (which also serves as the antecedent of the anaphor) is marked plural, the agreement on the verb under ta(a)n is 1pl not 1sg: Pasaŋ-gaɭ i [taaŋ-gaɭ {i,*j} viiʈʈŭkkŭ tanijaa poo-r-oom-ŭnnŭ] so-nn-aaŋ-gaɭ.boy-pl.nom[anaph-pl.nomi home-dat alone go-prs-1pl-comp] say-pst-3m-pl Literal: 'The boys i said [ CP that themselves {i,*j} are going home alone].'Reading: 'The boys i said [ CP that they {i,*j} are going home alone].'Sundaresan (2012) argues that sentences like (42), (44), and (46) involve indexical shift (von Stechow 2002;Schlenker 2003a;Anand 2006;Shklovsky & Sudo 2014) for 1st-person in the embedded clause: i.e. the 1st-person forms are evaluated against the speech index introduced by a selecting speech predicate, rather than against the utterance context.Such sentences show that the antecedent-tracking generalization in (41) does indeed hold: but the nature of agreement triggered in each case is different.In the standard (or elsewhere) case, antecedent-tracking yields φ-matching.In the more tightly constrained clausal complement of a speech verb (where indexical shift obtains), and where the antecedent denotes the agent of the speech verb, agreement is 1st-person.

Agreement is not triggered by the anaphor
Given that agreement is uniformly triggered by the nominative elsewhere in Tamil (see again ( 35)), a more reasonable claim might be that agreement under nominative ta(a)n is simply triggered by ta(a)n itself.Here, I argue against this conclusion on two grounds: (i) This would require us to claim that there is an anaphor and a 1st-person shifted indexical, both of which syncretize as ta(a)n -to deal with agreement contrasts like that between ( 38)-( 40), ( 45), on the one hand, and ( 42), (44), and ( 46), on the other.While not impossible, this would be a difficult syncretism to capture formally because these categories do not seem to form a natural class.(ii) 1st-person agreement can also be triggered under a 2nd-person indexical nii in Tamil.It is not possible to extend a syncretism account to deal with this pattern because this would involve claiming that nii is simultaneously an unshifted 2nd-person indexical and a shifted 1st-person indexical.
Turning to (i), we have just seen that the agreement on the clausemate verb under ta(a)n always tracks the antecedent of the anaphor (as described in ( 41)) but does so in different ways.In the clausal complement of a speech predicate, it matches the number of the antecedent but not its person; rather it shows up as 1st-person.In all other scenarios (the elsewhere case), it fully matches the φ-features of the antecedent.As mentioned, Sundaresan (2012) treats the 1st-person agreement cases as involving indexical shift in the complement of the speech predicate.Assuming this is correct, proposing that ta(a)n is the agreement trigger in such sentences would entail that ta(a)n is a shifted 1st-person indexical.Furthermore, recall that ta(a)n itself cannot take a 1st-person or 2nd-person antecedent (cf.( 42)-( 46)).This means that ta(a)n wouldn't just be a 1st-person form that can be shifted: rather, it would have to be the spell-out of an obligatorily shifted 1st-person indexical.Of course, this wouldn't yield the φ-matching agreement pattern in the elsewhere case.So here, we would have to propose that ta(a)n spells out a 3rdperson pro-form (with additional gender and number features).In other words, if ta(a)n is the controller of both patterns of verbal agreement, it should be able to bear either a 3rd-person feature ( + gender and number features) or an obligatorily shifted 1st-person feature.While we could set up post-syntactic spell-out rules that yield this kind of syncretism, it would be quite difficult to do so in a principled way, as 3rd-person and obligatorily shifted 1st-person do not seem to form a natural class (which could e.g.be defined in terms of a common set of features, with underspecification for all others).12Moving now to (ii), even more compelling evidence that the shifted indexical triggering 1st-person agreement in sentences like ( 42)-( 46 In (47), we have the 2nd-person indexical nominative pronoun nii ('you') in the clausal complement of a speech verb which is coreferent with the agent of this speech predicate.
In ( 46)), we again have nii as the embedded subject, but the agreement triggered under nii in the embedded clause is 1sg.Such seemingly mismatched 1sg agreement only obtains under ta(a)n and under nii, under highly constrained structural conditions (namely, in the clausal complement of a speech predicate), as illustrated here; it doesn't obtain under other nominative DPs (e.g. a 3rd-person pronoun like avan ('he') or avaɭ ('she')) in the same structural configuration: [ CP avaɭ i viiʈʈ-ŭkkŭ tanijaa poo-r-een-nnŭ] so-nn-aaɭ.she.sg.nom she.sg.nom house-dat alone go-prs-1sg-comp say-pst-3fsg Literal: 'She i said [ CP that she i am going home alone].'Intended reading: 'She i said [ CP that she i is going home alone].' Given the many parallels, the structure in ( 47) again looks like it involves indexical shift in the embedded clause: note that the 1st-person agreement does not denote the utterance-context speaker but, rather, the speaker of the intensional index associated with the matrix speech predicate (see Messick 2016 for parallel examples from related Dravidian language Telugu, which are also analyzed in terms of indexical shift).But claiming that the shifted 1st-person indexical is embedded nii in ( 47) is even harder to maintain than it was with ta(a)n, since nii has an explicitly 2nd-person form and denotes the Addressee of the utterance-context.I.e. it looks like a well-behaved unshifted 2nd-person indexical in Tamil.This means that, in sentences like (47), the embedded nii cannot be the source of 1sg agreement on its clausemate verb: rather, this suggests that there is some other element in the local domain, specifically a shifted 1st-person indexical, that triggers this agreement.
This should immediately make apparent that any syncretism account adopted to handle agreement under ta(a)n cannot be readily extended to handle agreement under nii.For instance, an analysis that postulated that ta(a)n syncretically spells out an obligatorily shifted 1st-person indexical and a 3rd-person anaphor, while nii spells out 2sg would fail to yield the 1st-person agreement under nii in (47).On the other hand, we could extend the analysis that we are led to posit for (47), namely that 1st-person agreement is triggered by some other shifted 1st-person indexical (presumably silent) in the embedded clause, to the ta(a)n sentences in (( 42)-( 46)) as well as those in ( 38)-( 40) and (45).I.e.rather than having different flavors of ta(a)n triggering the agreement (3rd vs. 1st) (which is independently non-trivial, as discussed earlier), we could propose that some other nominal in the local domain, which is sometimes a shifted 1st-person indexical, does so.This would yield a unified analysis of verbal agreement in embedded clauses, including in sentences like (47).

A silent, perspectival pronoun
The previous section has presented evidence suggesting that, in sentences involving perspectival anaphora with ta(a)n, agreement on the clausemate verb of ta(a)n is triggered, not by the nominative anaphor or its antecedent, but by some other nominal.Let us see what this would entail.This nominal must, of course, itself have valued φ-features so that it can trigger them on the verb; we don't see it overtly on the surface, so it must be silent.Finally, given that agreement is local (formalized via Agree in Minimalism, see Chomsky 2001 et seq.), this nominal must be syntactically local to the T head on which it triggers agreement.Putting these properties together, we arrive at the conclusion that the nominal must be a silent pronoun or pro (i.e. a silent form of a pronoun like 'he', 'she', 'it' etc.) in the local clause of the verb.This section will explore a theory that takes the existence of such a pronoun at its heart.I will show that this pronoun is perspectival and, furthermore, mediates the relationship between the anaphor and its antecedent, coreferring with the latter in discourse and binding the former in syntax-semantics.This will allow a straightforward account of perspectival anaphora that captures their hybrid structural and discourse-pragmatic properties as well as their interesting relationship with verbal agreement in Tamil.
In Section 6, I will present independent evidence (i.e.independent of agreement) for the existence of this silent pronoun.This evidence will show that the model of perspectival anaphora developed here makes correct empirical predictions with respect to the bound variable nature of the anaphor, on the one hand, and the pronominal nature of the pro, on the other.An account that assumes only the presence of an anaphor and its antecedent with no perspectival pronoun such as I describe in this section, will be unable to describe those properties.

A mediating pronoun
Recall that: (i) φ-agreement triggered under nominative ta(a)n always tracks the antecedent in different ways: in the clausal complement of a speech predicate, it is 1st-person, triggered by a shifted 1st-person pronoun, but still reflects the features of the agent of the speech predicate; everywhere else, it matches the φ-features of the antecedent.(ii) antecedent-tracking agreement only obtains when the clausemate subject is ta(a)n or, in the clausal complement of speech verbs, nii.In all other instances, agreement reflects the features of its clausemate nominative argument.
The most straightforward way to derive the antecedent-tracking effect in (i) would be to have the pro that (putatively) triggers verbal agreement corefer with the antecedent of the anaphor.In the default scenario, the φ-feature sets of the two coreferring nominals are evaluated against the same context (default = utterance-context); thus, coreference entails φ-matching.This plays out as follows.In a sentence like (38), the pro corefers with the antecedent avaɭ ('she'); since both pro and the antecedent are evaluated against the same evaluation context, coreference entails φ-matching, thus pro also has 3fsg features.
Pro thus triggers 3fsg verbal agreement under nominative ta(a)n, and the agreement matches the φ-features of the antecedent, by transitivity.In a sentence like (42), we are assuming that the 1sg agreement on the embeded verb is triggered by a clausemate shifted 1st-person indexical.We are proposing that verbal agreement is triggered by pro, so this entails that the shifted 1st-person indexical is pro.But the φ-features of the antecedent -namely the agent of the selecting speech predicate -are evaluated against the unshifted utterance-context.I.e. in (42), pro is 1sg and denotes Raman in the shifted context, while Raman has 3msg features in the unshifted one, and also denotes Raman: thus both nominals corefer.More generally, context-shifting allows us to get coreference between pro and the antecedent (and thus the antecedence-tracking effect with verbal agreement) without the added entailment of φ-matching between them.
The observation in (ii) above -i.e. the fact that the perspectival pronoun triggers verbal agreement only when the anaphor (as opposed to some other nominal, e.g. a coreferent pronoun) occurs in the nominative, shows that it must also be sensitive to the presence of the anaphor in some way.This in turn demonstrates that it is not enough to have the pro interact with the antecedent alone; it must interact with the anaphor, as well.I propose, specifically, that the perspectival pronoun Agrees with the anaphor in syntax and binds it at LF -a position I elaborate on more explicitly in Section 5.3.Since pro corefers with the antecedent, we get coreference between the anaphor and its antecedent by transitivity (see Nishigauchi 2014; Charnavel 2017 for similar proposals for exempt anaphora in French and Japanese, respectively).In other words: (49) The silent pronoun in the local clause of the anaphor enters into two dependencies: one with the antecedent and the other with the anaphor, yielding identical reference with both in difference ways.It thus mediates the relationship between the anaphor and the antecedent, which thus corefer only indirectly, via this silent pronoun.

Enter grammatical perspective
Where does perspective fit into all this?Recall that the central property of anaphora in Tamil is that it is perspectival, defined in the sense of (11), repeated below: (50) Definition of perspectival anaphora: In every instance of perspectival anaphora, the anaphor is properly contained within a predication which is evaluated relative to the perspective, mental or spatial, of some sentient individual.This individual must be aware of the eventuality described by this predication, at the time it happens.The antecedent of the anaphor must denote this individual.
The most elegant way to combine the insights in ( 49) and ( 50) would be to propose that the silent pronoun that mediates between the anaphor and its antecedent is itself a perspectival pronoun.
Let us now try to be precise about what a perspectival pronoun is.Like any other pronoun, the perspectival pro will bear inherent φ-features: this is what allows it to trigger verbal agreement in sentences where ta(a)n occurs as the nominative subject.However, unlike standard pronouns, it bears an added restriction that the individual it denotes must be perspectival in the sense defined in (50).I propose that this restriction comes about purely as a function of where this pronoun is merged in the structure.Specifically, I argue that pro is introduced in the specifier of a perspectival head (Persp) which assigns it a perspective-holding "discourse role" with respect to the proposition in its complement, via Event Identification (this analysis takes much of its intuition from the structural implementation of point-of-view (POV) proposed in Speas 2004; see also Nishigauchi 2014; Charnavel 2017 for recent proposals along very similar lines).This is, incidentally, much like the Voice head assigning an Agent θ-role to the external argument in Kratzer (1996).The Persp head selects the perspectival predication in its complement and has the following denotation: (51) ⟦Persp⟧ c,g = λxλe.PerspHolder (e,x) Following Heim & Kratzer (1998), I assume that φ-features on pronouns are encoded as presuppositions, formalized as partial functions on their lexical entries.Thus, pro, if it were to be born with 3fsg features would have the lexical entry given in (52) in Tamil: (52) ⟦pro 3fsg ⟧ c,g = λx: ¬Participant(x) ∧ Female(x).x Once the pro in [Spec, Persp] composes with Persp -its reference gets restricted both by the presuppositions imposed by its own φ-features and the θ-role-like discourse information on perspective-holding contributed by the Persp head.The set of possible referents for pro is thus twice filtered -yielding a set of individuals who satisfy the φ-features on pro as well as the perspectival condition given in (50).
We observed at the outset that anaphoric perspective could be defined along the mental or spatial dimensions.Anaphora in Tamil can be regulated by either.Building on prior work concerning the semantics of self-ascription (Lewis 1979 a.o.), Sundaresan & Pearson (2014) propose that all perspectival predicates quantify over elements of a set that are designated by a sentient entity (the judge or perspective center) as candidates for her actual time, location or world.The difference between spatial vs. mental perspectiveholding lies merely in the choice (location vs. world, respectively) of this coordinate.Building on this analysis, I further tentatively propose that the choice of these coordinates is made in the Persp head.To be concrete, the Persp head inside the complement of a spatial predicate (e.g.inside a locative PP or DP) will contain only the spatial coordinate, yielding spatial alternatives; the Persp head in the complement of an attitude verb will contain the World coordinate (yielding Doxastic alternatives) and so on (see also work on discourse centers in Roberts 2014 for related ideas).

A two-stage model of anaphora
The state of affairs described in (49) sets the stage for a two stage model of anaphora, with a mediating perspectival pro at its heart.The dependency between pro and the anaphor is distinguished by its being local and structurally constrained while that holding between pro and the antecedent of the anaphor is a case of (non-structural) discourse-pragmatic pronominal reference.Below, I describe the nature of this two-stage model of anaphora in detail.

Stage I: The structural stage
Let us now zoom in on the nature of the structural relationship between pro and the anaphor, and when the anaphor is the nominative subject, between pro and T. I am working within a Minimalist framework (Chomsky 2001 et seq.) which assumes a Y-modular architecture of grammar (with a "narrow" syntactic module feeding the LF and PF interfaces).The pro triggers verbal agreement on T when the anaphor is in the nominative (yielding the antecedent-tracking effect); it also Agrees with the anaphor for a different formal feature in syntax -a dependency that triggers binding at LF.
I formalize this state-of-affairs as follows.Agree proceeds upward (see Zeijlstra 2012; Bjorkman & Zeijlstra 2014 for motivations for Upward Agree) with the perspectival pro (Goal) c-commanding the anaphor (Probe) and T (Probe).There are two relevant Agree relations: one between T and pro for φ-features, and another between pro and the anaphor, which feeds binding at LF.The main feature inventory consists of φ-features (valued/unvalued) on nominals and T; there is also a dep-feature, defined as follows: (53) The Dep feature: 13,14  i.A Dep feature marks two DPs that are in a syntactic binding dependency with one another.ii.Dep takes integers or letters as value.iii.Two elements with matching Dep values are construed to be in a binder-bindee relationship with one another at LF, and will thus denote the same entity in the evaluation context.
13 This is a marked deviation from the more traditional view that anaphoricity follows from φ-defectiveness (Reuland 2001;2011;Kratzer 2009;Rooryck & vanden Wyngaerd 2011 a.o.), but it is both deliberate and warranted.As Table 2 shows, only two types of nominals may bear dep: pro in [Spec, PerspP] which bears a valued dep and an anaphor which bears unvalued dep.Thus, Agree for dep will hold only between these elements.If a perspectival anaphor were defined in terms of unvalued φ-features, we would expect it, however, to Agree with the minimally closest c-commanding nominal with valued φ-features -but this would overgenerate greatly.We are, ultimately, dealing with a fundamentally different kind of anaphor.As such, there is no reason to assume that the features that define non-perspectival anaphors will necessarily define perspectival ones, as well: quite to the contrary. 14An anonymous reviewer wonders whether dep violates the Inclusiveness Condition in Chomsky (1995: 225), given that the value of dep on perspectival pro is not already specified in the lexicon but is known only after it is merged in the syntactic structure.Hicks (2009) proposes that Hicks ' [Var] which is formally essentially like dep doesn't violate Inclusiveness because: "The feature is present in the lexicon, just that the feature value in the listed entry is an instruction to be converted into an integer upon lexical selection.The feature value that the pronoun receives is not strictly present in the lexicon, but it is determined by its lexical properties."(Hicks 2009: 115-116).This reasoning can be unproblematically extended to dep.
iv.An anaphor is a nominal with an unvalued Dep feature -this is the syntactic correlate of anaphoricity; the pro in [Spec, PerspP] has a valued Dep feature (potentially a kind of selectional feature), by virtue of where it is merged in the structure.v.The anaphor may have one or more φ-features in addition to the Dep feature, some of which may themselves (but need not) be unvalued. 15  Table 2 illustrates all possible featural values on T, pro and the anaphor.In addition to the features listed below, I assume that both ta(a)n and the perspectival pronoun are endowed with a categorial D feature and case feature -not included here for reasons of space.
Since the anaphor has an unvalued dep-feature, it probes upward in its local search domain to get this valued: pro in [Spec, PerspP] values dep, so the anaphor and pro end up having matching dep-values.This triggers binding at LF, with pro binding the anaphor, since it asymmetrically c-commands it.The anaphor and pro will thus denote the same individual in the evaluation context.When the anaphor is ta(a)n, pro must be 3rd-person (elsewhere case) or a shifted 1st-or 2nd-person pronoun (1st-person agreement case).
If pro is an unshifted 1st-or 2nd-person pronoun, it will denote the Speaker and Addressee of the utterance-context, respectively.While pro itself is free to denote any individual (as long as it fulfills the perspectival condition), we have seen that ta(a)n cannot be taken by 1st-and 2nd-person (unshifted) antecedents.This is shown below for 1st-person naan: anaph[nom] school-dat go-prs-1sg-comp say-pst-1sg 'I i said [ CP that I i am going to school].' (Intended) This restriction cannot come from the dep feature on ta(a)n (given that nii is anaphoric and can patently denote the Addressee of the utterance-context).Rather, I propose that it is related to the notion that ta(a)n has an unvalued person feature. 16Specifically, I propose that ta(a)n has a presuppositional restriction in its lexical entry preventing it from denoting a Participant of the utterance-context, as in ( 55): (55) ⟦taan⟧ c,g = λx: ¬Participant*(x).x,for Participant* = Participant(c*) 15 Thus, φ-features, if any, presuppositionally restrict the domain of mapping possibilities for the reference index at LF (Heim & Kratzer 1998), but don't (directly) have anything to do with flagging a nominal as an anaphor in the syntax. 16This is a welcome result.Given that other perspective sensitive elements (spatial verbs, taste predicates and the like) can refer to utterance-context participants, it would be a priori unexpected if perspectival anaphors alone could not do so.Furthermore, it is intuitively appealing to derive a referential ban on certain types of person combinations as a function of other properties of person innate to the anaphor.[[g(17)]] c,g = Raman Verbal agreement under nominative ta(a)n is due to φ-Agree between T and pro, as mentioned.Nevertheless, such φ-agreement must be sensitive to the presence or absence of the anaphor, since pro triggers φ-agreement on T only when the nominative subject is an anaphor (i.e. in all other cases, the nominative subject triggers φ-agreement, cf. ( 35)-( 36)).I propose, in line with Koopman & Sportiche (1989); Speas (2004: but pace Nishigauchi 2014; Charnavel 2017 who propose it is merged lower in the clausal spine) and others, that the perspectival phrase is merged in the left periphery of the clausal spine, crucially above the subject.The T head probes to get its φ-features valued by the nominative DP, typically the subject in [Spec, TP].Typically (cf.( 35)-( 36)), this DP has inherently valued φ-features and can itself value the features on T. But in sentences like ( 38)-( 40), the nominative subject is the anaphor ta(a)n which has an unvalued dep feature and unvalued person and gender features.17As such, it cannot value these φ-features on T which thus keeps probing upward in its local domain until it finds the next closest nominal with valued φ-features, namely pro in Spec, PerspP. 18Pro values the unvalued φ-features on T with its own inherent features.In sentences, like ( 38)-( 40), pro is a(n) (unshifted) 3rd-person pronoun and thus triggers 3rd-person agreement on T. Since pro refers to the individual denoted by the anaphoric antecedent, antecedence φ-matching is the result.In ( 42) and ( 46), pro is a shifted 1st-person indexical and thus triggers 1st-person agreement on the verb.Nevertheless, pro still denotes the individual denoted by the anaphoric antecedent in these cases: thus, we still get the antecedent-tracking effect of verbal agreement observed in sentences like ( 44), ( 45) and ( 46).The agreement mechanism in the φ-matching scenario (Elsewhere case) is illustrated below: (59) Agree + binding between pro and ta(a)n in ( 57 When the nominative subject is anaphoric nii (as in a sentence like (47), in the clausal complement of a speech predicate, with 1st-person agreement on the clausemate verb), the derivation proceeds essentially analogously.The perspectival pro is a shifted 1st-person indexical (inherently valued as 1sg) and also has a valued dep-feature.It values the depfeature on its clausemate subject nii (which probes upward to get this feature valued) in the embedded clause: this leads pro to bind nii at LF. T probes upward to get its φ-features valued.As with ta(a)n, it first encounters the nominative DP nii in syntactic subject position.Unlike with ta(a)n, nii does have valued φ-features.However, it has an unvalued dep-feature.I propose that this prevents it from serving as a Goal for φ-valuation on the T Probe. 19The T head thus continues probing upward in its local search domain, just like in the ta(a)n case, until it reaches pro (the shifted indexical) in [Spec, PerspP], which values it as 1sg. 20

Stage II: The discourse-pragmatic stage
The second stage in the perspectival anaphoric dependency involves the relationship between the perspectival pro in [Spec, PerspP] and the individual denoted by the antecedent of the anaphor.As we have seen, there are no (obvious) structural constraints placed on the distribution of the antecedent in Tamil (or Icelandic, or the other languages with perspectival anaphoric systems discussed here): i.e. the antecedent may be extrasentential (logophoric), non-c-commanding, non-local, non-minimal, and indeterminate.
19 This is, admittedly, a stipulation.But it is not altogether devoid of independent merit.The Anaphor Agreement Effect (AAE) (Rizzi 1990) observes that anaphors cannot occur in positions construed with φ-agreement.Subsequent research (see e.g.Woolford 1999;Legate 2002;Haegeman 2004;Baker 2008;Deal 2010;Tucker 2011;Shiraki 2005) on a wide range of languages has since revealed that the descriptive generalization is more that anaphors cannot trigger covarying φ-agreement.An easy way to derive the AAE would have been to propose that anaphors are themselves φ-featurally minimal, thus cannot value φ-features on probing heads (T/v), as Kratzer (2009) suggests.But while such an analysis might work for ta(a)n, it will not work for anaphoric nii in sentences like (47), where its clausemate verb surfaces with 1sg agreement, since this would require us to posit that nii must simultaneously be a shifted 1st-person indexical and an unshifted 2nd-person indexical, as discussed in Section 4.3.Thus nii must have inherent 2sg φ-features, as indicated in Table 2.But if having any unvalued feature on a nominal prevents it from valuing φ-features, then the AAE with nii would follow, since anaphoric nii has an unvalued dep-feature. 20Unlike the Agree relation for φ-features between pro and T head which could take place in the post-syntactic PF module (Bobaljik 2008), the Agree relation for dep between pro and the anaphor must take place in narrow syntax, since its output must feed operations at both LF and PF.
The relationship between the antecedent and the perspectival pro in the local phase of the anaphor must thus necessarily be non-structural.We capture this by proposing that the relationship between the perspectival pro in [Spec, PerspP] and the individual denoted by the antecedent, is just discourse-pronominal reference.The perspectival pro can refer, just like any standard pronoun can, to such an individual, as long as it has been made discourse-salient by another nominal (R-expression or pronoun).Such a nominal could have invoked this individual in the c-commanding syntactic structure (as in standard cases of long-distance anaphora), in non-c-commanding structure (as in psych predications), or in the preceding discourse-context (as in cases of logophora).The φ-features inherent to pro will restrict the domain of individuals it may refer to in the context of evaluation.In addition, given the perspectival discourse role pro is assigned in [Spec, PerspP] from Persp, the set of entities it may denote is further perspectivally restricted as described in (50).The nominal (R-expression or pronoun) that introduces the perspectival individual in the sentential structure or salient discourse will thus corefer with pro.In the structural stage of perspectival anaphora, discussed in Section 5.3.1, we observed that pro Agrees with the anaphor in syntax, which then leads to its variable-binding it at LF (see again ( 58)).The anaphor will thus necessarily take on the same reference as pro and also corefer with this nominal, which will come to be construed as the antecedent of the anaphor.
The sentence in (60) below, repeated from ( 57), illustrates this more concretely: In (60), pro happens to be born with 3msg features.There are two R-expressions and one pronoun in the sentence structure c-commanding the minimal PerspP containing the anaphor -namely the matrix subject Raman, the medial subject avaɭ ('she'), and the matrix object Krishnan.These denote three salient individuals, Raman, a (previously invoked) female individual, and Krishnan, in the evaluation context.The female individual is automatically ruled out as a possible referent because pro's own φ-features presuppositionally restrict its reference to atomic, male individuals.That leaves Raman and Krishnan.However, Krishnan doesn't satisfy the perspectival condition in (50): the Per-spP inside the innermost CP is not evaluated relative to Krishnan's perspective, but to Raman's.Thus, pro denotes Raman, and corefers with the matrix subject, the R-expression Raman.Since pro Agrees with ta(a)n in syntax and binds it at LF (in Stage I), ta(a)n also denotes Raman.Raman is thus construed as the antecedent of the anaphor.
But this is just in the pragmatically unmarked discourse scenario.Let us suppose that the propositional content of ( 60) is uttered directly after the free indirect discourse scenario in (60), which is reported from Krishnan's inner perspective: Krishnan k stayed upset that whole day.Getting that prize would have meant a lot of money for the family.But Raman seemed to have some pretty solid inside knowledge about how the results would turn out.
In this scenario, both Krishnan and Raman fulfill the perspectival condition in (50).
Krishnan could be upset because Raman himself lost the prize, as (60) indicates -e.g. if Raman is his son.But Krishnan could also be upset because he himself lost the prize.In the former, we have an instance of long-distance anaphora, in the latter, one of logophora. 21The advantage of the current model is that both types of dependency are derived in a precisely parallel fashion.The updated referential possibilities against this discourse scenario are thus as given in ( 62 The assignment function at LF can thus have the pro-ta(a)n pairing at LF denote either Raman or Krishnan: the choice of one over the other depends only on speaker-intent and related criteria (just like with standard reference assignment for pronouns).The set {Raman,Krishnan} constitutes the domain of (salient) potential antecedents for ta(a)n, and the one that is actually chosen, Raman in (60) and Krishnan in (62), its actual antecedent in a given context.
Finally, if pro were born with 3fsg features, it would be presuppositionally restricted to denote the female individual denoted by the medial pronominal subject avaɭ ('she'), and any other salient atomic female individuals.Raman and Krishnan would be ruled out on φ-featural grounds.Additional filtering would be imposed as before by the perspectival condition in (50).Since pro is a pronoun with inherent valued φ-features, there are no restrictions on the φ-features it may be born with.The syntax is assumed to overgenerate; any incompatible combinations (e.g. if pro were born with 3nsg features in (60)/(62) and there were no salient individual that could satisfy the presuppositional restrictions of those features) are assumed to be filtered out at the interfaces.
When the context of evaluation of the antecedent and that of pro are identical, coreference between the two entails φ-matching.We saw an instance of this in ( 60) and (62).Taking (60) as expository, the context of evaluation for pro in the innermost CP = the context of evaluation for the antecedent Raman in the matrix CP = the utterance-context.Thus, the φ-features of both pro and Raman are evaluated against the utterance-context.This means that coreference between pro-ta(a)n and the antecedent yields φ-matching: both must be specified 3msg, which is what we get.
We have seen (cf.42) and ( 46)) that, when ta(a)n is in the clausal complement of a speech predicate, the clausemate verb of ta(a)n shows 1sg agreement.The perspectival pro that triggers 1sg agreement on the embedded verb must thus be born with 1sg φ-features.But recall from Section 4.3 that the embedded clause in sentences like (63) involves indexical shift (Sundaresan 2012): I will assume that formally, indexical shift is due to the presence of a "monster" ( ) operator (Kaplan 1989;Schlenker 2003a;Anand 2006;Shklovsky & Sudo 2014) introduced by the selecting speech predicate soll ('say') in its complement.This operator replaces the context of utterance-context with the intensional index of this predicate: i.e. ⟦ α⟧ c,i,g = ⟦α⟧ i,i,g .Pro, a 1st-person indexical, must thus be merged in the scope of this monster, causing it to be shifted.( 63) presents the resulting underlying structure of such a sentence: As a result of indexical shift, the 1sg φ-feature on pro in (63) will denote, not the (unique) Author of the utterance-context, but the Author of the index associated with the speechpredicate, namely Sai.As always, in addition to the presuppositional φ-restriction, the perspectival condition must also be fulfilled: given that Sai is an attitude-holder with respect to the PerspP containing ta(a)n, this condition is also met.The referential assignment of ta(a)n (which pro has already bound at LF in the structural stage) to Sai is thus felicitous.Sai is introduced as a possible referent in the discourse by the matrix subject, the R-expression Sai, and pro thus corefers with it; ta(a)n takes Sai as its antecedent.Crucially, however, Sai, being in the root clause, is evaluated against the utterance-context.It denotes a non-participant in the utterance-clause, thus has 3msg φ-features (as also indicated by the 3msg φ-features it triggers on its clausemate matrix verb soll ('say')).Thus, here we have a scenario involving coreference between pro/ta(a)n and the anaphor's antecedent Sai without the added entailment of φ-matching between the two.Now consider ( 64): The structure in (64) varies from that in (63) only to the extent that the matrix and embedded subjects are now the unshifting 2nd-person indexical pronoun nii ('you') instead of Sai and ta(a)n, respectively.But the structure of the embedded CP remains unchanged.I.e. the embedded CP, which is selected by the matrix speech predicate is contextually shifted by a monster.pro, which again has 1sg features is merged in the scope of this monster as before and gets shifted, as before.Thus, instead of denoting the unique Author of the utterance-context, pro denotes the unique Author of the index associated with soll ('say') -namely, nii (which also happens to be the Addressee of the utterance-context).As before, nii also fulfills the perspectival condition with respect to the PerspP containing embedded nii: thus, pro and embedded nii, which it binds (due to nii's unvalued dep-feature), can felicitously denote matrix nii.Embedded nii thus takes matrix nii as its antecedent.

Summing up
We observed at the outset of this paper that perspectival anaphora displays hybrid syntaxpragmatic properties that seem to resist a unified analysis.At times, these properties were shown even to flagrantly violate what are considered cornerstones of structural wellformedness (c-command, (Relativized) Minimality, syntactic determinacy, and sentenceboundedness).
The two stage model of perspectival anaphora proposed here allows us to reconcile these hybrid properties with one another, by arguing that every instance of perspectival anaphora (logophoric, long-distance, backward etc) is serially restricted by both syntactic and discourse-pragmatic factors.Perspectival anaphora, in other words, represents a hybrid syntactico-pragmatic phenomenon that is comprised of two separate, sequential dependencies, as depicted below: (65) Two stage model of perspectival anaphora: Discourse-pragmatic coreference Syntactic Agree + LF Binding Stage I: A local, structural (i.e."narrow" syntactic and LF-semantic) relationship between the anaphor and the perspectival pro introduced in the specifier of the minimal Perspectival Phrase (PerspP) containing the anaphor.This minimal PerspP characterizes the local binding domain of the anaphor.The pro Agrees with the anaphor for a dep feature which triggers binding at LF. Stage II: A discourse-pragmatic relationship, holding between the perspectival pro and the individual denoted by the antecedent of the anaphor.The relationship between pro and this individual is restricted in two ways: (i) There is a presuppositional restriction on reference assignment, contributed by the inherent φ-features on pro.(ii) In addition, there is a perspectival restriction relative to the minimal PerspP containing pro and the anaphor, contributed by the perspectival discourse-role that the Persp head assigns to pro in its specifier.Thus, the denotation of pro is twice filtered: as such, it is both well-formed with respect to its φ-features and characterizes the perspective of the minimal PerspP containing itself and the anaphor.The nominal (call it XP) that introduces this individual in the sentence structure or salient discourse corefers with pro.If XP and pro are both evaluated against the same context, such coreference entails φ-matching.But if they are evaluated against different contexts (e.g. if one of them is in a domain that is shifted), then coreference can obtain in the absence of φ-matching.

Stage I feeds into Stage II:
The structural component (Stage I) feeds into the discoursepragmatic one (Stage II) in the derivation.As a result, the anaphor which is bound by pro in Stage I also ends up coreferring with XP from Stage II.XP is construed as the antecedent of the anaphor.There is thus no structural relationship between the antecedent on the one hand, and the anaphor/pro, on the other.This relationship is just discourse reference.C-command, Relativized Minimality, and lack of optionality for antecedence are thus not expected.The relationship between the anaphor and pro is structural.Thus, structural sensitivity (e.g. the agreement patterns in Tamil when ta(a)n is in the nominative) can be explained.

Independent predictions of a two-stage model
The two stage model of anaphora centrally revolves around the notion of a mediating, perspectival pro, as we have seen.Indeed, one of the main empirical goals of the paper thus far has been to motivate, through a careful investigation of the agreement patterns on the clausemate verb of the anaphor, the existence of such an element in syntax.Nevertheless, it is important to be explicit about what this model entails.In particular, it is potentially hazardous in two ways.First, it could violate Occam's Razor.All the mediating burden needed to make the current model work is carried on the shoulders of a silent nominal.All else being equal, it would be simpler to remove pro from the equation altogether and have the anaphor take over the agreement-triggering properties that are now being attributed to it.But as discussed in detail in Section 4.3, such a stance is independently problematic.
To briefly reiterate, this would require us to postulate an inelegant sycretism between an obligatorily shifted 1st-person indexical (for 1st-person agreement under ta(a)n as in (63) and a 3rd-person pro-form).Additional evidence comes from 1st-person agreement triggered under nii ('you'), as in ( 64), where even such an analysis, however inelegant, fails: i.e. nii itself is patently 2sg in this sentence, thus couldn't have triggered 1st-person agreement on the verb (short of us postulating that nii is simultaneously 2nd-person and 1st-person).
Section 4 thus addresses the Occam's Razor challenge by arguing that all else is not equal: there is independent motivation for why the elements that are visible -namely the anaphor's antecedent and the anaphor itself -are not sufficient to derive the agreement patterns, and why a third element, even if it is silent, has to be postulated to derive these patterns.But if we simply propose an element that precisely meets the needs of that motivation, we potentially run into the second problem, namely that of circularity: i.e. assuming an empty element with tailor-made properties to fit the observed phenomena (in this case, the embedded agreement patterns observed for Tamil under the nominative ta(a)n and under nii).Ideally, therefore, we should find confirmation or evidence that is independent of the agreement patterns for the existence of a mediating nominal with the properties ascribed to pro.
We take these challenges seriously.In this section, I thus present independent evidence from anaphora in Tamil, and crosslinguistically, for the presence of a mediating perspectival pronoun with properties precisely such as those proposed here. 22The two-stage model predicts that perspectival anaphora should display dual referential behavior.The relationship between the antecedent and pro should exhibit the properties of pronominal coreference, since the pro, being on the "outside" discourse-pragmatically corefers with the antecedent the way a regular pronoun does.At the same time, the anaphor should behave like a bound variable, since it is locally bound by pro at LF. Below I present empirical diagnostics to show that these predictions are met.Furthermore, I show that a model that does not presuppose the existence of a mediating pro would be incapable of making these same predictions.

"On the outside": Pronominal
Under the two-stage model, the relationship between pro in [Spec, PerspP] and the individual denoted by the anaphor's antecedent is just one of discourse-pronominal reference (restricted by a perspectival presupposition introduced by the Persp head).The antecedent of the anaphor and pro corefer as a result; similarly, the anaphor and pro corefer (since pro binds it).This predicts that the relationship between the antecedent and the anaphor shouldn't fulfill any of the standard tests associated with standard bound-variable anaphora since the anaphor, in fact, has no direct relationship with the antecedent.Rather, it should display the characteristics of pronominal coreference.
Here, I show that this prediction is fulfilled.Bound-variable anaphors have been observed to be incapable of taking split antecedents.They must also obligatorily yield obligatory bound-variable readings when c-commanded by definite DPs such as R-expressions (Reinhart 1983).In contrast, regular pronouns may take split antecedents and may yield bound-variable as well as "strict" (due to their ability to refer discourse-pragmatically) readings under definite DPs.Such tests are thus commonly used to distinguish between pronominal and anaphoric uses of a term, when this is difficult to diagnose on the surface.Below, I show that the antecedence-ta(a)n relationship displays the characteristics of pronominal reference with respect to these diagnostics, rather than those of bound-variable anaphora (see Charnavel 2017 for similar evidence from exempt anaphora in French).
In Tamil, sentences involving ta(a)n can take split antecedents (see also Annamalai 2000) and can also yield bound-variable or strict readings:23 (66) Kumar bought a house that just came on the market as a surprise for his wife.Last week, Kumar showed his wife the house he bought for them.I can report this to you as in: (67).

"On the inside": Anaphoric
Strictly speaking, the patterns given above are perfectly consistent with the notion that there is no mediating pro.They could also be explained under the assumption that the perspectival pronoun is ta(a)n itself.Under such a model, ta(a)n, being a free pronoun, would have inherent φ-features and no unvalued dep (or other unvalued) feature, since it wouldn't enter into an Agree relation that feeds binding in syntax-semantics.It would thus also be able to trigger agreement on the verb when it occurs in the nominative.The only hiccup in the analysis would be the independent difficulties with a syncretism analysis that would go hand in hand with having ta(a)n be the source of agreement, as discussed in Section 4.3.Nevertheless, the fact that we get strict and sloppy reference under ellipsis, and that ta(a)n can take split antecedents would both be predicted.
Here, I present empirical arguments against this alternative.To this end, I show that, despite the pronominal nature of the relationship between ta(a)n and its antecedent, the pronoun is not ta(a)n itself.Rather, ta(a)n is a locally bound anaphor (a bound variable), which is bound by a pro in its local PerspP, just as argued in this proposal.There are two kinds of evidence I present to this end: (i) Multiple occurrences of ta(a)n within a single PerspP cannot take distinct antecedents: they are forced to take the same antecedent.(ii) The notion that ta(a)n is bound by a pro in [Spec, PerspP] coupled with the idea that the structural position where pro is merged is higher than [Spec, TP], the syntactic position of the clausal subject -predicts that object ta(a)n should not be capable of being locally anteceded by its clausemate subject.I.e. it predicts that surface reflexivity should be banned with ta(a)n.
In the sections below, I show that both predictions are met.

Antilocality restriction
If ta(a)n is bound by a pro in its local domain which is, furthermore, merged in the clausal left periphery, higher than the subject as argued here -then it is predicted that object ta(a)n should not be capable of being anteceded by its clausemate subject.This is because, in such a configuration, pro would asymmetrically c-command this antecedent in addition to coreferring with it.If the antecedent is an R-expression, this would yield a Condition C violation; if it is also a pronoun, this would yield a Condition B violation.Thus, we predict that ta(a)n should not be capable of being locally anteceded.
This prediction is fulfilled (see also Annamalai 2000), as we have already seen in ( 34), repeated here: 24 (74) Raman is watching TV coverage of a cricket match he had attended when he suddenly spots himself on TV.I cannot report this state-of-affairs as in: (75a) or (76a).
( The only way to salvage sentences like (75)-( 76) within the perspectival model proposed here would be if the perspectival binding domain (the PerspP) were smaller than a CP, and could, specifically, intervene between the subject and the anaphor.Assuming that the subject is merged as the external argument in [Spec, VoiceP] (or [Spec, vP]) and the internal argument (which is the anaphor) is merged as the complement of V, as is standard, this would thus be a position between v/Voice and V.
24 As mentioned earlier, (75a)-(76a) can become grammatical with the addition of a verbal suffix "kol" on the verb.Following Sundaresan (2016), I will assume that kol is a thematic raising predicate (in the sense of Ramchand 2008).In transitive constructions, it raises the external argument in Spec, VoiceP into its own specifier and assigns it a new θ-role.Sundaresan proposes that, in reflexive structures, this raising operation allows the external argument to escape the minimal PerspP containing the anaphoric object and the external argument (in its base position).As such, the external argument can, from its new A-position in the Spec of kol serve as a potential antecedent for the anaphor without violating antilocality.Such an anaphor may, in other words, be reflexively bound under the addition of kol. 25 An anonymous reviewer asks whether Condition C violations may be expected elsewhere with perspectival anaphora.Under the current model, the perspectival pro may denote any individual in the salient discourse, as long as this individual is also a perspective-holder toward the ta(a)n-predication.Similarly, an R-expression, e.g.John in the sentence or salient discourse, may also denote such an individual (assuming the φ-features are compatible in the evaluation context).This yields coreference between pro and the R-expression.If the antecedent is logophoric (i.e.extra-sentential), then there is no problem for Condition C. In a case of long-distance anaphora, the antecedent is intra-sentential (but outside the binding domain).
Here, the only configuration we have to worry about is one where pro c-commands the R-expression antecedent.In so-called "backward binding" constructions (see again Exx. ( 12)-( 13), the anaphor superficially c-commands its antecedent.But there is no reason to think that this represents the underlying c-command relation given work arguing that experiencers may be merged higher or move to a higher position (Beletti & Rizzi 1988).As far as I know, there aren't other instances of such violation -which is telling in itself.I thank the reviewer for helping me think through this more clearly.
However, relevant crosslinguistic evidence has been recently brought to bear in Bylinina et al. (2014) and Bylinina & Sudo (2015), based on data involving perspective-shifting with respect to various structural domains, arguing precisely against this possibility.A central notion of these works is that certain structural domains involve the presence of a perspectival operator (which would instantiate the Persp head in this model) which shifts the perspective of a perspective sensitive item (PSI) in its scope from the default perspective (that of the utterance-context speaker) to that of the attitude-holder associated with this operator. 26The shiftability of a PSI in a given structural domain can thus be taken to diagnose the presence or absence of a perspectival center/Persp.Crucially, such diagnostics show that VP is not a shifting domain because, when a perspectival item appears as the main predicate, it cannot shift its perspectival center to the subject of that sentence.The authors provide examples like "John is handsome", where the (perspectival) taste-predicate handsome has to be evaluated from the utterance-context speaker's perspective and cannot be evaluated from that of John. 27Under the current proposal, this would translate to saying that there is no Persp between v/Voice and V, i.e. between the internal and external arguments.In reflexive structures, the anaphor and its co-argument are thus contained inside the same minimal PerspP (or binding domain).The structural configurations of sentences like (75a) and (76a), as given in (75b) and (76b) above, thus predict ungrammaticality.
Potential further evidence that the ungrammaticality of sentences like (75a) and (76a) has to do with the antilocality of the relationship between pro and the antecedent rather than that between the antecedent and the anaphor (as is more traditionally assumed), comes from the fact that, when the antecedent is another anaphor, the antilocality restriction is lifted and reflexive binding becomes possible (crucially without the addition of the verbal suffix kol).Compare (77b) with (75a)/(76a): 28 (77) No antilocality with anaphoric subject: a. Raman is watching TV coverage of a cricket match he had attended, when he thinks he sees himself on TV!I can report this as in: (77b 26 The idea takes its intuitions from context-overwriting approaches of indexical shift due to monstrous operators, as we have already seen (Anand 2006;Shklovsky & Sudo 2014). 27In contrast, in a sentence like "If a handsome man comes in, John will be startled", the PSI handsome is ambiguous and may be evaluated either from the speaker's perspective or from John's, showing that there is a perspectival center introduced at the level of the CP by the attitude verb. 28It is tempting to dismiss (77b) as just another instance of long-distance anaphora.But assuming that derivations are built bottom-up, the lower CP will be computed before the antecedent is merged, thus any antilocality effect between the structural positions of the antecedent and the anaphor in the embedded CP should kick in first.Furthermore, even if this is a case of long-distance anaphora, there is still a local reflexive relation between subject and object instances of ta(a)n which should trigger antilocality.As far as I can see, sentences like (77b) would thus be problematic under an analysis like Reuland (2011) which would treat the antilocality in ta(a)n-reflexives as resulting from the monomorphemic status of ta(a)n -specifically, as resulting from the notion that it would form an irrecoverable A-chain with its antecedent.Note that such an analysis would also need to account for the independent use of kol and the perspectival properties of ta(a)n throughout, including in reflexive structures.Under the current proposal, this is exactly as predicted.There is no Condition B or Condition C violation, since the antecedent, being itself an anaphor, can be locally bound by the perspectival pro in (77b).

Unique binder restriction
A tacit assumption of the current proposal is that there is a unique pro per PerspP (the binding domain of the anaphor).There is independent empirical evidence for this idea coming from a "Shift Together" constraint on perspective-sensitive items This restriction automatically follows if there is a unique perspectival center per binding domain (the PerspP).29 Given this, a prediction that the two-stage approach for anaphora proposed here makes, is that multiple occurrences of an anaphor within a single PerspP should be restricted to taking the same antecedent.This would be an instance of Shift Together for anaphora.Conversely, when the multiple occurrences of ta(a)n belong to distinct structural domains, the restriction is lifted: the different occurrences can now denote distinct antecedents, just as predicted: (83) Raman and Seetha are travelling on the train, each carrying a lot of cash.To avoid pickpockets, they decide that Seetha should hide both her cash and Raman's cash in a safe place.Raman thought that: a. Seetha hid her cash near herself.b.Seetha hid his cash near herself.c.Seetha hid his cash near himself.d.Seetha hid her cash near himself.
(84) The sentence above involves a mental Persp introduced by the matrix attitude-predicate nenae ('think') and a spatial Persp introduced by the locative preposition pakkattŭ ('near').Each instance of ta(a)n is crucially in the scope of a different Persp, as indicated.Under the current model, this means that each will be bound by a different perspectival pro and will thus be able to corefer with a different antecedent.This prediction is again fulfilled, as illustrated above.In contrast to (80), the sentence in ( 84) is four-ways, not two-ways, ambiguous.

Summing up
In this section, I have presented independent evidence to support the proposal that perspectival anaphora involves a two-stage process with a mediating pro at its core: a structural one involving a variable-binding relation between the anaphor and a perspectival pro in its local domain and a discourse pragmatic one involving regular pronominal (co)-reference between pro and the antecedent of the anaphor.
Although the initial motivation for this proposal was evidence involving verbal agreement triggered under anaphora in Tamil, I have argued in this section that this model makes the right empirical predictions with respect to perspectival anaphora in Tamil and languages like it.In particular, it predicts that the relationship between the antecedent and pro should display the empirical fingerprint of (discourse-)pronominal reference while that between pro and the anaphor should display that of bound-variable anaphora.I have attempted to show at length that these predictions are fulfilled.With respect to the former, ta(a)n can take split antecedents and yield strict readings in the domain of definite DPs.With respect to the latter, I show that multiple occurrences of ta(a)n within a single PerspP (binding domain) cannot take distinct antecedents: this follows from the independently supported notion that each PerspP has exactly one pro binder.I also argued that the two-stage model predicts that standard reflexivity should be ruled out within a perspectival system, as a function of antilocality (violations of Conditions B or C).This prediction is also borne out.

Conclusion
The goal of this paper has been to argue that grammatical perspective, instantiated either mentally or spatio-temporally, is structurally represented.Evidence for this came from the Dravidian language Tamil where it was argued that grammatical perspective could directly affect the shape of morphosyntactic agreement on the verb.On the strength of this, I have proposed a two-stage model of perspectival anaphora mediated by a perspectival pronoun that corefers discourse-pragmatically with the antecedent of the anaphor, and variable-binds the anaphor in its local domain at LF.The antecedence-anaphora relationship is thus actually an epiphenomenon of two independent referential relationships.In addition to explaining the agreement facts that motivated the analysis in the first place, this model also has the independent advantage of being able to explain hitherto problematic aspects of perspectival anaphora, to wit that it is structurally well-behaved in some respects (e.g. with respect to respecting locality domains for anaphors) and ill-behaved in all others (e.g. with respecting to violating locality, minimality, c-command, antecedent determinism and so on) -properties which make it hard to analyse in either purely structural or purely discourse-pragmatic terms.It also predicts that such anaphora should be pronominal "on the outside" (i.e. with respect to antecedence) and anaphoric (like a bound variable) "on the inside" (i.e. with respect to anaphora).
There is another sense in which, under the current system, perspective-taking is structural. 30The perspectival pro derives is perspectival properties as a function of being merged as the specifier of a functional head (Persp).This functional head assigns pro a perspectival role, much like Voice assigns the external argument in its specifier a θ-role.This state-of-affairs has two consequences.First, given that Persp is unique to the clausal (and, in languages like Tamil, also the nominal) extended projection, this automatically ensures that there is a unique pro per phase.Second, it entails that there is no need to distinguish between perspectival and non-perspectival pronouns in the lexicon of a given language.There are simply pronouns, null and overt: when a null pronoun is merged in the Spec of a head like Persp and gets assigned a perspectival role, it becomes perspectival. 31he perspectival pro derives its perspectival properties directly as a function of its structural position under the current analysis, as just discussed.At the same time, nothing forces us to say that this pro must be the Spec of a Perspectival Phrase (PerspP): it can be the specifier of any functional head that is capable of assigning a perspectival role to its specifier.Languages may, indeed, vary in their choice of what such a head might be.Although the analysis here has been based primarily on evidence from Tamil, it can be easily extended to model (mental or spatial) perspectival anaphora in other languages.For instance, it has been noted (see e.g.Hicks 2009) for Icelandic, that the identity of the perspective-holder also seems to condition the choice of subjunctive vs. indicative marking on the clausemate verb of the chosen antecedent.Interestingly, the role of the subjunctive in Icelandic seems to be "to signal that the perspective-holder of a given construction is distinct from the [utterance-context] speaker" (Hellan 1988: 89) or, as Sigurðsson (2010: 50): "In modern Icelandic, the most important factor that triggers subjunctive marking in these complements is that the speaker does not take responsibility for their truthfulness" (Sigurðsson 2010: 50).An elegant way to model these facts would be to propose that, in Icelandic, the "Persp" head that introduces the perspectival pro and associates it with a perspective-holder role is, in fact, nothing other than the Mood head that is responsible for yielding subjunctive marking.In the indicative, the perspectival pronoun is pre-set to denote the utterance-context speaker, but in the subjunctive this default is obviated or shifted, allowing it to corefer with the antecedent of the anaphor.In other languages, perspectival anaphora is intimately tied with properties of (aboutness) topichood: here, the Persp head might be Topic.
If perspective is syntactically represented, as this paper has aimed to show, we expect it to make its presence felt not only semantically but also morphologically.Indeed, clauses containing logophors are often introduced by special complementizers (Sells 1987;Koopman & Sportiche 1989), perspectival anaphors are often distinct from their nonperspectival counterparts and in certain dialects of Tamil, two types of anaphoric form seem to be attested, often occurring in the same environments: the only difference is that one of them is perspectival, while the other is not.Even more compelling evidence that perspective is represented inside a dedicated structural projection, as argued in this paper, comes from Spadine (2017).Spadine presents evidence from perspectival anaphora in Tigrinya to argue that both Persp and the perspectival pronoun in its specifier may be overtly represented.
Claiming that perspective is structurally represented, however, has the implication that it should be able to influence not only anaphoric dependencies but also other types of (morpho)syntactic phenomena.A striking parallel to this phenomenon is found in the realm of control -broadly speaking another kind of referential dependency between nominals.Landau (2015) indeed argues that instances of non-obligatory control crosslinguistically should be analysed as a kind of "logophoric control" (see also Frascarelli 2007) involving a perspectival pronoun which has a mediating function that is strikingly similar to that of the perspectival pro in the current model.The precise extent and nature of the differences between these phenomena will require careful empirical investigation which also take seriously the roles of clausal finiteness and predicate selection into consideration.In principle, every predication that is evaluated relative to a judge or perspectival center should include the representation of a Perspectival Phrase (PerspP) with a Persp head that introduces a pro in its specifier.This suggests that this model could, in theory, be extended to derive other perspectival phenomena in grammar such as "taste" predications (Stephenson 2007), modal auxiliaries (Speas & Tenny 2003), evidentials (though see Korotkova 2016 for a discussion of why a judge-based treatment of evidentials is problematic), and so on.
-local and non-minimal: a. Krishnan, Raman and Anand and I are drinking together at a bar after work.I watch as Krishnan eavesdrops on Raman who is telling Anand that our friend, Seetha, saved Krishnan from falling off a cliff last week.Later, I say: (29b).

Table 1 :
Case and number paradigms for Tamil taan.
Krishnan i [ CP Seetha tann-ae {i,j} kaadali-kkir-aaɭ-ŭnnŭ] Krishnan[nom] Seetha[nom] anaph-acc love-prs-3fsg-comp Raman-ae j nenekka-vej-tt-aan.Raman-acc think-caus-pst-3msg 'Krishnan i made Raman j believe [ CP that Seetha loved him {i,j} ].' CP [ DP Taan {i,j} avvaɭavŭ eeɻae-jaaga irŭnd-adŭ] [ DP Raman-ooɖae i anaph[nom] so poor-adj be-pst-3nsg.nomRaman-gen aɳɳaav-ae] j rombae-vee baadi-jirŭ-kkir-adŭ.] brother-acc very-emph affect-be-prs-3nsg '[ DP His {i,j} having been so poor] has really affected [ DP [ DP Raman i ]'s brother] j .' 30) Choice of antecedent: indeterminate: a. Krishnan and Raman are both in love with Seetha.Krishnan, who is quite manipulative, recently convinced Raman that Seetha actually loves him (Krishnan), hoping to get Raman off his back.I later describe this to you as in: (30c).b.Krishnan and Raman are both in love with Seetha.Krishnan, who is quite pessimistic, recently convinced Raman that Seetha actually loves him (Raman).Later, I describe this to you as in: (30c).c.Antecedent: non c-commanding: a. Raman and his brother both invested very foolishly in the stock-market and are now both broke, where they were once quite well-off.Their family doctor cautions Raman's wife regarding Raman's health, saying (31b).A little later, she meets with Raman's brother's boyfriend and cautions him the same way regarding the brother's health, uttering (31b).b. [ themselves k would lose the prize]].' (literal) prize-acc lose.go-prs-3msg-compnamb-in-aan-ŭnnŭ] [pasaŋ-gaɭ-kiʈʈae] k kaaʈʈ-in-aaɭ.believe-pst-3msg boy-3pl-all show-pst-3fsg 'She i showed [the boys] k [ CP that he j believed [ CP that himself j /*herself i /*themselves k would lose the prize]].' (literal) The child i reflected about what had happened.Why had itself {i,*j} suffered?' Raman i said [CP that self {i,*j} am going home alone].'Reading: 'Raman i said [CP that he {i,*j} is going home alone].' 42) superficially seems to violate the antecedence-tracking generalization, described in (41).But a closer look at the interpretation of such examples shows that this is not the case: Raman i said [ CP that Krishnan j said [ CP that he {j,*i} loves Seetha]].' i said [ CP that Krishnan j said [ CP that he i loves Seetha]].' Did you Addr* say [ CP that you Addr* am going home alone]?' Reading: 'Did you Addr* say [ CP that you Addr* are going home alone]?'
CP ...[ PerspP pro i ...[ TP Raman i tann-ae i paar-tt-aan]]] CP ...[ PerspP pro i ...[ TP avan i tann-ae i paar-tt-aan]]] [ [ Raman i thought [ CP that self {i,*j} saw self {i,*j} ].' Reading: 'Raman i thought [ CP that he {i,*j} saw himself {i,*j} ].' Persp Utt-Speaker ) b. John read a book by an author who John thinks is talented and who is from a different country than John.(talented: Persp John ; foreigner: Persp John ) c. John read a book by an author who I think is talented and who is from a different country than John.(talented: Persp Utt-Speaker ; foreigner: Persp John ) d. John read a book by an author who John thinks is talented and who is from a different country than me.(talented: Persp (PSIs): "i.e.PSIs in the same [local] domain must refer to the same PC [perspectival center]" (Bylinina et al. 2014: 10) -illustrated below (Bylinina et al. 2014: 12, Ex. 40): (78) John read a book by a talented EvidentialPSI foreigner PronominalPSI .a. John read a book by an author who I think is talented and who is from a different country than me.(talented: Persp Utt-Speaker ; foreigner: John ; foreigner: Persp Utt-Speaker ) Below, I show that this prediction is indeed fulfilled: Sri i [Mia j [ GerP [ PerspP ta(a)n {i,j} tann-ae {i,j} aɖi-čč-adaagae]] Sri i Mia j anaph[nom] anaph-acc hit-pst-nmlz kanavuka-ɳɖ-aaɭ-ŭnnŭ] so-nn-aan.dream-pst-3fsg-comp say.3msgLiteral: 'Sri i said [ CP that Mia j dreamed [ GerP of self {i,j} hitting self {i,j} ]].'Under a proposal where ta(a)n is a free (albeit perspectivally restricted) pronoun, the fact that Krishnan cannot report (80) to mean either (79c) or (79d), is unexpected.Indeed, if we replace the anaphor ta(a)n with the deictic pronouns avan ('he') and avaɭ ('she'), the other two readings become available, as shown below: Sri i said [ CP that Mia j dreamed [ CP that she j hit him i ]].' (81), as reported by Krishnan, is compatible with: (79c).Sri i said [ CP that Mia j dreamed [ CP that he i hit her j ]].' (82), as reported by Krishnan, is compatible with: (79d).