A Competition-Based Proposal for Chinese Non-Local Reflexivization 3.1 Introduction

Chapter Three

A Competition-Based Proposal for Chinese Non-Local Reflexivization 3.1 Introduction

In this chapter, I will lay out an approach to the Chinese LD reflexive ziji based on Hu and Pan’s (2002) NP prominence theory, which in turn draws its inspiration from Xu’s (1993, 1994). As Hu and Pan’s approach has been introduced in detail in Chapter 2, the present chapter focuses on how the forthcoming theory differs from theirs. It will be demonstrated that five points of divergence are necessary as certain data suggest in section 3.2. One difference is the retention of an independent local binding component in the current proposal, whereas Hu and Pan’s approach is an ambitious attempt at unifying local and long-distance binding, which runs into problems with the most typical instance (See Chapter 2 for an example and 3.2.).

A second difference concerns the eliminability of [+prominent]. Hu and Pan assume that once an NP is marked as [+prominent], the feature will not be removed even when an NP next to it ranks higher in terms of prominence. By contrast, the current approach assumes that an NP marked as [+prominent] can have this feature removed when it is less prominent than an NP next to it. It is also assumed here that [+dominating, +animate] outranks [+subject].

A third difference, concomitant with the second one, relates to the necessity of certain prominence factors. [+first/second person] is eliminated from consideration because its sole purpose is to derive the asymmetrical blocking effect and nothing more. Further, while [+agent] and [+local] have a place in Hu and Pan’s theory, in the current theory, their roles are called into question; it is found that if we allow

[+dominating, +animate] to outrank [+subject], we can do away with [+agent] and [+local], thereby considerably downsizing the inventory of prominence factors.

Fourthly, the candidate as defined in Hu and Pan’s Antecedent-Seeking

Mechanism has to be refined in such a way that when the XP c-commanding ziji is a non-subject, the XP itself (when it is a nominal phrase) and the XP-internal NP

c-commanding the head of the XP count as candidates. When the XP is a subject, both it and any subject NP within it are candidates.

A fifth difference lies in the order in which candidates compete for prominence.

Hu and Pan adopt a strictly linear order, whereas I propose that after prominence competition has taken place among the various candidates within an NP which is itself a candidate, only the NP, not the leftmost candidate within it, goes on to compete with a structurally higher candidate on the left.

Section 3.3 examines the proposed mechanism against further data. I will argue that subject orientation and binding by sub-commanding NPs follow from the current account.

In section 3.4 I will examine some sentence types apparently imcompatible with the Antecedent-Seeking Mechanism, such as ziji in adverbial clauses adjoined to the matrix IP, psych-sentences, and ba/bei constructions. It will be shown that these cases can be handled if we adopt some assumptions, such as some functional phrases.

Section 3.5 is devoted to issues regarding the nature of the Antecedent-Seeking Mechanism. I will demonstrate that it is essentially a syntactic operation because it displays two parallels characteristic of syntactic operations, locality and upwardness.

Section 3.6 applies the Antecedent-Seeking Mechanism to Japanese and Korean.

Some modifications will be shown to be necessary in the presence of some data.

Below these points will be fleshed out and accompanied by examples that justify the necessity of these differences.

43 Section 3.2 might read somewhat like a review of Hu and Pan’s approach and thus might be

incorporated in Chapter 2. However, new ideas have been discussed here, and the discussion would be impossible if the inadequacies of their proposal were not highlighted to illustrate the necessity of the modifications to be proposed below. For a complete review of Hu and Pan’s approach, please refer back to the relevant sections in Chapter 2.

3.2 Modifications to Hu and Pan (2002)

As it is established in Hu and Pan (2002) that prominence plays a role in

licensing certain NPs as binders, the question now is how to modify their proposal in such a way as to accommodate the facts reported in the literature.

Firstly, the current proposal assumes that sub-commanding binding should be treated separately from local binding, but on a par with long-distance binding. The assumption runs counter to the claim implicit in the previous Chomskyan analyses such as Huang and Tang (1991), Huang and Liu (2001), Sung and Cole (1994) et al.

that sub-commanding binding by an animate NP is assimilated into local binding,

but gains support from the Singaporean Teochew example (1a) below in Cole et al.

(2001) and Icelandic examples (1b-c) in Maling (1984):

(1) a. Ah Meng

gai chia

hai-liao kaki

. Ah Meng’s car harm-Perf self

Ah Meng

’s car

harmed *him/*itself.

b. Skoðun Siggu

er að sig

vanti hæfileika Opinion Sigga’s is that self lacks-subj. talent Sigga’s opinion is that she lacks talent.

c. Jón

segir að Ólafur

hafi ekki enn fundið vinnu, sem sér

líki.

Jon says that Olaf has-subj. not yet found a-job, that self likes Jon says that Olaf has not yet found a job that self likes.

According to Cole et al., (1a) indicates that the subject Ah Meng is not a legitimate antecedent because certain logophoric requirements are lacking.

This suggests that

44 They either adopted sub-commanding in the definition of c-command (cf. Huang and Tang (1991)) or resorted to feature percolation (cf. Sung and Cole (1994)) so that the sub-commanding animate antecedent in the same clause as ziji is treated like a local antecedent in the grammar.

45 Although Cole et al. cited the example to illustrate the presence of logophoric requirements on sub-commanding antecedents in Teochew, they did not seem aware that this example (their (12a)) demonstrates that their feature percolation principle (FPP), whose function is to convert a

sub-commanding NP into a local c-commanding antecedent here, stands at odds with the fact that local antecedents need no logophoric requirements, unlike LD and sub-commanding antecedents.

Confusingly enough, they refer to the sub-commanding antecedent of (1) as an LD antecedent, although FPP seems to make it otherwise. Also recall that in Chapter Two I raised doubts about Cole et

both sub-commanding and LD antecedents need non-syntactic requirements (although these may differ for both cases) in the language and can be treated alike, whereas local binders require a separate treatment because they do not need to fulfill such requirements. (1b) and (1c) show that sub-commanding binding and long-distance binding are alike in that for most Icelandic speakers, both require the subjunctive tense as one licensing factor, in contrast to local binding (Maling, 1984).

Our assumption here also differs from that of Hu and Pan (2002), who treat binding, whether local, sub-commanding, or long-distance, in a uniform manner. One example against this claim is (81b) mentioned in Chapter 2, repeated below for ease of reference:

(2) Zhangsan

zhidao Lisi

xihuan ziji

. Zhangsan know Lisi like self.

Zhangsan

knows Lisi

likes him

/himself

.

As the reader can verify (see chapter 2), Hu and Pan’s theory would incorrectly predict long-distance binding to be impossible in this sentence.

Of course, (2) does not definitively point to the separate treatments for local and non-local binding, as the problem might be [+local]; as will be shown below shortly, if we do away without it, Zhangsan and Lisi would be correctly predicted to antecede ziji. However, consider the following example, which demonstrates that Hu and Pan’s uniform treatment for local and non-local binding is problematic, and this has nothing to do with [+local]:

(3) Zhangsan

renwei meige gongyuan

duo you ziji

-de tese.

al.’s claim that examples such as (1) are ungrammatical because of the absence of the SELF requirement on the intended antecedent. If this doubt is confirmed, then (1a) would not support my treating sub-commanding and local binding separately.

46 Here I do not exclude the possibility that a uniform approach to binding could be achieved if Hu and Pan’s proposal is modified in some way. The criticism discussed here applies only to their current formulation.

Zhangsan think every park all have self’s characteristics.

Zhangsan

thinks every park

has his

/its

own characteristics.

Note that both Zhangsan and meige gongyuan can be the antecedent.

If we apply Hu and Pan’s prominence computation, we cannot predict this possibility. According to their definition of the agent, which includes the experiencer role, Zhangsan counts as an agent. It outranks meige gongyuan because the former is [+subject, +agent] and [-dominating, +animate] whereas the latter is [+subject, -agent] and [-dominating, -animate]. [+local] has no role to play here. This predicts Zhangsan to be the only antecedent, contrary to fact.

If we follow Tang (1989) and treat meige gongyuan as an instance of

metaphorical extension (personification), as well as assuming that meige gongyuan is [+agent]

, then only meige gongyuan could be the antecedent because it is more prominent than Zhangsan after we invoke [+local]. If we assume meige gongyuan is [-agent] and keep rest of the above personification account, only Zhangsan could be the antecedent; Meige gongyuan would be [+subject, -agent] and [-dominating,

+animate] and Zhangsan would be [+subject, +agent] and [-dominating, +animate]. In other words, there is no way Hu and Pan’s theory could allow both Zhangsan and meige gongyuan to be legitimate antecedents. It seems that however Hu and Pan’s account is applied, a separate treatment for local binding is useful in order to handle such data, contra their position.

47 Tang (1989) first discussed the example and treated it as an instance of personification. Pan (2001) opposed such an account. Please refer to section 2.4, Chapter Two.

48 Although we are not sure whether the subject of the predicate you ziji-de tese can be considered an experiencer and [+agent], we are showing Hu and Pan’s formulation cannot deal with (3), no matter whether meige gongyuan is [+agent] or not.

49 Concomitant with the first difference is our elimination of Hu and Pan’s requirement that a reflexive searches for overt NPs as antecedents before it goes on to search for covert NPs as antecedents. This requirement is set up to deal with antecedent relations as in (i) and (ii):

(i) Laowangi bei nij suo zai zijii/j-de wuzi-li.

Laowang BEI you lock Pro at self’s home-Loc Laowang was locked by you at self’s home.

(ii) Zhangsani bi Lisij gei zijii/j gua huzi.

Let us turn to our second assumption, viz. that an NP that is assigned [+prominent] during computation can have this feature removed at the end of computation, provided a more prominent NP occurs next to it. First consider the following:

(4) Lisi

dui Zhangsan

biaoming zheben shu

dui ziji

mei yong.

Lisi to Zhangsan indicate this-CL book to self no use

Lisi indicated to Zhangsan that this book was useless to him.

Recall that experiencers are treated as [+agent] in Hu and Pan’s approach and

Zhangsan, an experiencer, is thus [+agent]. Let zheben shu and Zhangsan in (4) enter into prominence competition. As the former is [+subject, -agent] and [-dominating, -animate] and the latter is [-subject, +agent] and [-dominating, +animate], Zhangsan outranks zheben shu in terms of prominence. Since in Hu and Pan’s formulation an NP that is marked [+prominent] must retain this feature throughout the computation, Zhangsan is wrongly predicted to be able to antecede ziji, despite the presence of Lisi, a candidate that is supposed to outrank Zhangsan.

But so far we cannot definitively decide that the problem is the assumption that an NP marked as [+prominent] will retain this feature throughout the competition because Hu and Pan’s approach relies heavily on whether a candidate has a particular value for a prominence factor. That is,

Zhangsan force Lisi for self shave beard Zhangsan forced Lisi to shave for him/himself.

In (i), on the assumption that [+subject] and [+agent] are equally prominent in their approach, as ni outranks Laowang because the former is [-subject, +agent], [+animate], and [+second person] and the latter is [+subject, -agent], [+animate], and [-second person], Hu and Pan account for the coreference between Laowang and ziji by postulating an empty local subject bearing the index of Laowang and then the letting the requirement do the work. I believe their postulation of an empty local subject here is an ad hoc decision to make (i) parallel to (ii), and I will therefore refrain from positing any empty subject and treat (i) as an instance of local binding. As for (ii), Zhangsan outranks Lisi, and the presence of Pro and that requirement save the coindexation between Lisi and ziji. But the relation between Lisi and ziji is an instance of local binding, and the requirement therefore has no place in my approach, which is devoted to non-local binding.

50 It is still possible that Zhangsan could refer to ziji, but it would be an emphatic/contrastive use and fall outside the scope of the current inquiry.

perhaps by tinkering with the values of a certain prominence factor, we could avoid the problem noted above. Let us test whether this could be achieved by assuming Zhangsan to be [-agent]. If so, it would lose the competition to zheben shu. Zhangsan, being [-subject, -agent] and [-dominating, +animate], would tie with zheben shu, being [+subject, -agent] and [-dominating, -animate] if we did not take [+local] into consideration. Once we consider [+local], zheben shu would outrank Zhangsan, which would compete with Lisi, which would win the competition over Zhangsan as the antecedent. Everything seems right here, except that zheben shu is predicted to be an antecedent (a problem that has been discussed along with our first difference from Hu and Pan and is not our concern now). It seems that as long as we assumed Zhangsan to be [-agent], it would not be wrongly predicted to be [+prominent] and a legitimate antecedent. However, consider the following, which is almost identical to (4) except that the subject is an inanimate NP:

(5) Zhe

xiang Zhangsan

zhengming-le zheben shu

dui ziji

mei yong.

This to Zhangsan prove-Perf this-CL book to self no use

This proved to Zhangsan that this book was useless to him.

If we assume that Zhangsan is [-agent] and additionally that zhe is [+agent]

, it would be predicted that both could antecede zjji. This is because Zhangsan would be [-subject, -agent] and [-dominating, +animate], zhe would be [+subject, +agent] and [-dominating, -animate] and neither is more local than the other, as there is no subset

51 Some people might feel that ziji strongly refers to a salient entity in the discourse, or the speaker. A variant of the NP accessibility hierarchy (subject> object> oblique> genitive NPs; note that I collapse primary and secondary objects into the category object for the present purpose) to be discussed in Chapter Four seems at work. I propose that the lower an NP marked as [+prominent] is placed in the hierarchy, the more likely it is to refer to a discoursally salient entity.

52 Some linguists might require agents to be sentient and thus consider zhe to be [-agent]. Although the definition of agenthood is notoriously controversial and difficult to pin down (see Saeed 2003), our current criticism against Hu and Pan’s approach does not crucially rest on it. As the reader can verify for himself, even if we assume zhe to be [-agent], the problem discussed in the text still exists. Thanks go to Hsieh Laoshi for bringing the issue on agenthood to my attention.

relationship between them. Consider (6) below: the path from ziji to the PP dominating Zhangsan is not a proper subset of the path from ziji to the matrix IP dominating zhe. This is a tied competition. In other words, Hu and Pan’s formulation cannot rule out zhe as an antecedent in (5), if we assume that Zhangsan is [-agent]; if we assume Zhangsan is [+agent] instead, we will incorrectly predict it to be able to antecede ziji in (4).

The above discussion suggests that the culprit for the problems is not whether a certain NP should be marked [+agent]. Consider again (5) and its structure (6) with the tentative assumption that [+agent] is excluded from the prominence factors.

(6)

Zhe NP

I

xiang P

Zhangsan NP P' PP

zhengming-le V

zheben shu dui ziji mei yong CP

V' V'

VP I' IP

Zheben shu, being [+subject], [-dominating, -animate], and [+local], outranks

Zhangsan, being [-subject], [-dominating, +animate], and [-local]. Zhangsan ties with

zhe, being [+subject] and [-dominating, -animate]. As neither is more local than the

other, [+local] does not apply here. This means that Zhangsan and zhe would be

predicted to be legitimate antecedents. Obviously, the problems noted in the previous

paragraphs cannot be solved by marking one NP as [+agent] and the other as [-agent],

or by rejecting [+agent] altogether from consideration.

As it will be argued below, what is wrong is the assumption that once an NP is marked as [+prominent], it will be an available antecedent even when an NP next to it ranks higher in terms of prominence. Let us propose the following preliminary

modifications to Hu and Pan’s formulation:

(7) Prominence Elimination (Preliminary Version):

An NP marked as [+prominent] during the computation can have this feature eliminated, provided that there is a more prominent NP next to it.

(8) Prominence Hierarchy:

A. i. [+subject, +agent] > [+subject, -agent]

ii. [-subject, +agent] > [-subject, -agent]

iii. [+subject, +agent] > [-subject, +agent]

iv. [+subject, -agent] > [-subject, -agent]

B i. [-dominating, +animate] > [-dominating, -animate]

ii. [+dominating, +animate] > [-dominating, +animate]

iii. [+dominating, -animate] > [-dominating, -animate]

iv. [-dominating, +animate] > [+dominating, -animate]

B > A

Given two candidates α and β
^ only β, not α, is marked as [+prominent] if α outranks β in terms of A but is outranked by β in terms of B.

Reconsider (4) and (5):

(4) Lisi

dui Zhangsan

biaoming zheben shu

dui ziji

mei yong.

Lisi to Zhangsan indicate this-CL book to self no use

Lisi indicated to Zhangsan that this book was useless to him.

(5) Zhe

xiang Zhangsan

zhengming-le zheben shu

dui ziji

mei yong.

This to Zhangsan prove-Perf this-CL book to self no use This proved to Zhangsan that this book was useless to him.

53 Such an outranking relation is already implicit in Pan (2001). He proposed that the antecedent of long-distance ziji must be a self-ascriber. When a matrix-clause subject is inanimate, then the matrix-clause object, if it happens to be an animate object of verbs such as gaosu, is the antecedent because it is the only self-ascriber. (inanimate NPs cannot be self-ascribers.) However, as long-distance antecedents need not be self-ascribers, his outranking relation, which relies on self-ascribers, needs to be rethought. (8) and the modification to it in the text represent such a revision.

In (4), assume for the sake of argument that Zhangsan is [-subject, +agent] and [-dominating, +animate]. According to (7) and (8), it outranks zheben shu, which is [+subject, -agent] and [-dominating, -animate], and is marked as [+prominent]. Now Lisi, being [+subject, +agent] and [-dominating, +animate], outranks Zhangsan, thereby depriving the latter of [+prominent]. Only Lisi can antecede ziji in a neutral context, as predicted. If we assumed that Zhangsan is [-subject, -agent], the result would still be the same, with Lisi being the only antecedent. Let us now turn to (5). As in (4), Zhangsan outranks zheben shu; but it also outranks the matrix subject zhe, which is [-dominating, -animate]. With (7) and (8) in place, we can successfully tackle cases which Hu and Pan’s formulation cannot.

The third assumption of our current approach is facilitated by (7) and the dissociation of local binding from non-local binding in our approach. I propose that [+agent] and [+local] be excluded from the inventory of prominence factors.

Therefore (8) can be simplified as follows:

(9) B>A

[+dominating, +animate]> [+subject]

If A outranks B in terms of [+subject] but is outranked by B in terms of [+dominating, +animate], only B, not A, is marked as [+prominent].

Consider the following:

(10) Zhejian shi

gaosu Lisi

zheben shu

hai-le ziji

. This-CL thing tell Lisi this-CL book harm-Perf self This incident told Lisi that this book harmed him.

With (9), zheben shu would lose the competition to Lisi (the former would be

[+subject] and [-dominating, -animate], whereas the latter would be [-subject] and

[-dominating, +animate]). Zhejian shi, being [+subject] and [-dominating, -animate]

would also be outranked by Lisi. Contrast (10) with (11):

(11) Zhangsan

gaosu Lisi

zheben shu

hai-le ziji

. Zhangsan tell Lisi this-CL book harm-Perf self Zhangsan

told Lisi that this book harmed him

.

(7) and (9) predict Zhangsan to be the only antecedent because Lisi, although outranking zheben shu, is outranked by Zhangsan. Zhangsan and Lisi are both [-dominating, +animate], so the crucial factor for determining which is [+prominent]

is [+subject]. Zhangsan outranks Lisi because it is a subject.

As for [+local], we have seen in Chapter 2 that it renders typical long-distance binding impossible. To illustrate why we can do without [+local] if we are equipped with (7) and (8), consider the following:

(12) Zhangsan

renwei Lisi

taoyan ziji

. Zhangsan think Lisi hate self

Zhangsan thinks Lisi hates him/himself.

As both Lisi and Zhangsan are [+subject] and [-dominating, +animate], they are qualified as antecedents of ziji. To bring in [+local], as in Hu and Pan’s formulation, would make it impossible for ziji to refer to Zhangsan because Lisi would be [+local]

and outrank Zhangsan.

The role of [+first/second person] is also suspect. In Chapter 2, we have seen that its purpose during prominence computation is only to derive the presence and absence of asymmetrical blocking in some dialects of Mandarin. Consider (13):

(13) Wo zhidao Lisi taoyan ziji.

I know Lisi hate self

I know Lisi hates me/himself.

As both wo and Lisi are [+subject, +agent] and [-dominating, +animate], we need to consider [+local] and/or [+first/second person] to determine which outranks which, according to Hu and Pan. Those native speakers who consider [+first/second person]

would allow ziji to refer to wo, whereas those who consider [+local] would allow only Lisi as the antecedent. Lisi, the subject of the local clause, is correctly predicted to be the blocker which forbids ziji to refer to wo for the latter group of speakers. However, one immediate problem is that the former type of speaker would be incorrectly

predicted to disallow Lisi, which loses the competition to wo, as an antecedent.

This inadequacy and the fact that [+first/second person] plays no role in determining other coreference relations during prominence competition (see below) cast doubts on the necessity of this prominence factor.

Also consider how wo competes against zhejian shi for prominence in (14a-b).

Although we have eliminated [+agent] from the stock of prominence factors, let us tentatively keep it and the rest of Hu and Pan’s account to see why [+first/second person] is useless. Wo, being [-subject, +agent] and [-dominating, +animate],

necessarily outranks zhejian shi, being [+subject, -agent] and [-dominating, -animate]

in (14a).

Therefore the fact that wo can antecede ziji in (14) without invoking

54 Incidentally, note that this result, while unfavorable to Hu and Pan’s approach, supports our assertion that local and non-local binding should be handled differently. Because the former type of speaker is wrongly predicted to allow only wo, the winner of the competition, to antecede ziji, a mechanism devoted to local binding is needed to account for the binding relation between the local subject, Lisi, and ziji.

55 Reference of ziji to wo in (14a-b) is presumably correct in the dialect that Pan (2001) investigated.

56 Some people might consider zhejian shi, the subject of the predicate gaosu, an agent in (14a). Even if this is correct, all we need to make wo an antecedent is [+local]. Zhejian shi, being [+subject, +agent]

and [-dominating, -animate] would tie with wo, [-subject, +agent] and [-dominating, +animate] if we did not consider [+local] or [+first/second person] under Hu and Pan’s account. Consider [+local] first.

The path from ziji to the matrix VP, the minimal maximal projection dominating wo, is a proper subset of the path from ziji to the minimal maximal projection dominating zhejian shi. This demonstrates that [+first/second person] is unnecessary. On the other hand, however, [+local] is also problematic as the foregoing text has shown. (14b) might suggest that [+first/second person] must be taken into account, because there is no proper subset relation between zhejian shi and wo here. Hence [+local] appears irrelevant, and the coreference between wo and ziji seems to hinge on [+first person]. However,

[+first/second person] adds to the claim that [+first/second person] is entirely unnecessary.

(14) a. Zhejian shi gaosu wo Lisi taoyan ziji.

This-CL event tell I Lisi hate self

This event told me that Lisi hated himself/me.

b. Zhejian shi dui wo zhengming-le Lisi taoyan ziji.

This-CL event to me prove-Perf Lisi hate self This event proved to me that Lisi hated himself/me.

Let us now turn to how to select candidate NPs for prominence competition.

According to Hu and Pan (2002), a candidate is an NP contained within an XP c-commanding ziji. This means that candidacy is not limited to NPs c-commanding ziji; genitive NPs within nominal phrases are also candidates, for example. As long as an NP falls within an XP c-commanding ziji, it is qualified for prominence

competition. Candidates in their formulation would therefore include NPs in object position within an NP c-commanding ziji. However, such NPs do not seem to be candidates, even when they outrank their containing NPs. Consider the following:

(15) *Zheben hai-le Zhangsan

-de shu benlai dui ziji

you yi.

This-CL harm-Perf Zhangsan’s book originally to self have benefit.

This book that harmed Zhangsan could have benefited self.

According to the revisions we have made to Hu and Pan’s account so far, Zhangsan is [+animate] and outranks Zheben hai-le Zhangsan

-de shu, which is [-animate].

consider the following:

(i) Zhejian shi dui Zhangsan zhengming-le Lisi taoyan ziji.

This-CL event to Zhangsan prove-Perf Lisi hate self This event proved to Zhangsan that Lisi hated himself/him.

(i) differs from (14b) only in that the matrix object is a third person NP. As there is no proper subset relation between zhejian shi and Zhangsan either and [+first person] is irrelevant here, the reference of ziji to Zhangsan in (i) and to wo in (14b) has nothing to do with [+first person].

Nevertheless, the former is still not a legitimate antecedent. Indeed, (15) might be appropriate in intensification or contrastive contexts, but such uses of reflexives generally fall outside the realm of sentence grammar and instead require

discourse-pragmatic conditions.

I assume, pace Hu and Pan, the following procedure for selecting candidate NPs:

(16) Candidate Selection Procedure A:

α is a candidate if it is any subject NP contained within an XP functioning as a subject and c-commanding ziji, and any argument containing α is a subject.

(16) means that when the XP c-commanding ziji is a subject NP, both it and the subject NPs it dominates, e.g. genitive NPs or the subject NPs of the relative clauses of the subject NP c-commanding ziji, are candidates, and the candidates cannot occur in object position. Consider the relevant structure represented in (17).

The fact that Zhangsan cannot be the antecedent in (15) follows from our new assumption (16) and (17), where Zhangsan, a non-subject, within the subject Zheben hai-le Zhangsan

-de shu, is not a candidate and hence cannot enter into prominence competition with Zheben hai-le Zhangsan

-de shu.

(17)

57 Another instance is provided by Xu (1994, p.118):

(i) Wo wen-guo tai ji bian zijii-de mingzi.

I ask-Exp he several times self’s name I asked him his name several times.

ziji can refer to the matrix object ta here. However, in his endnote 5 Xu shows its special status by noting that some people read ziji as an emphatic reflexive. This suggests that such reflexives behave differently from non-emphatic, non-contrastive reflexives.

58 Here I omit irrelevant projections such as Classifier Phrase and ignore the issues on the internal structure of nominal phrases. This decision is arbitrary, but nothing crucial hinges on it.

59 The assumption presented here is a mere stipulation. It is unknown why subject-orientation seems necessary only within subject NPs; NPs in matrix-clause object position can be an antecedent to long-distance bound ziji, provided that it is prominent, as the current theory has shown. We leave this peculiar property of subject NPs for future research.

Note that (16) is similar to Tang’s (1989) notion of sub-command, but their implications are different. Sub-command, or Feature Percolation is intended to

maintain a c-command configuration between an antecedent and ziji. As Chapter 2 has shown, however, both devices rest on the assumption that inanimate NPs cannot antecede ziji. Pan (2001) has demonstrated that this claim is falsified in the presence of examples like meige gongyuan dou you ziji-de tese, alluded to earlier this section.

Speakers of Mandarin Chinese typically do not regard the subject as personified in any way. On the empirical front, sub-command or feature percolation cannot explain the possibility of a PP-internal NP to antecede ziji, as it is not a subject and cannot fulfill the definition of sub-command or feature percolation. See Chapter 2 for fuller discussion.

(16) determines the set of candidates when the XP c-commanding ziji is a subject, but candidates occur elsewhere too. Below the other procedure for candidate selection is given:

(18) Candidate Selection Procedure B:

Of all the XPs c-commanding ziji, candidates can be (i) any NPs c-commanding OP

t

NP

Agr

hai-le V

Zhangsan NP VP Agr' AgrP

de C C' CP

shu

N

N'

NP

ziji; or (ii) any NPs c-commanding the head of the c-commanding XP when the XP is a non-NP.

The following examples show (18) is necessary:

(19) a. Zhangsan

zai Lisi

-de jia-li tingshuo Wangwu taoyan ziji

. Zhangsan at Lisi’s house hear Wangwu hate self

Zhangsan heard in Lisi’s house that Wangwu hated him.

b. Zhejian

shi xiang Zhangsan

zhengming-le Lisi

taoyan ziji

. This-CL thing to Zhangsan prove-Perf Lisi hate self

This event proved to Zhangsan that Lisi hated him.

c. Wangwu

dui Zhangsan

biaoming Lisi

taoyan ziji

. Wangwu to Zhangsan indicate Lisi hate self

Wangwu indicated to Zhangsan that Lisi hated him.

According to Hu and Pan’s formulation, in (19a), Lisi, being [+subject, -agent]

and [-dominating, +animate] outranks Lisi-de jiali, being [-subject, -agent] and

[-dominating, -animate]. Lisi is therefore [+prominent]. Recall that Hu and Pan assume that once an NP is [+prominent], this feature will be retained throughout the computation. This means that Lisi can antecede ziji, contrary to fact. On the other hand, according to (7) and (9), Lisi would tie with Zhangsan, as both are [+subject]

and [-dominating, +animate]. Once we consider (18), however, Lisi would not be a candidate, as it does not c-command the head of the PP, which is a non-NP and c-commands ziji. Therefore Zhangsan competes only with Lisi-de jia, not Lisi. As predicted, only Zhangsan can be the long-distance antecedent of ziji, because it is [+subject] and [-dominating, +animate] and outranks Lisi-de jia, which is [-subject]

and [-dominating, -animate].

(18) is confirmed if we consider the contrast between (19b) and (19c), where Zhangsan c-commands the head of the PP c-commanding ziji and is therefore a

60 Note that genitive NPs are considered subjects.

candidate that enters into prominence competition with zhejian shi. Being [-subject]

and [-dominating, +animate], Zhangsan outranks zhejian shi, which is [+subject] and [-dominating, -animate] in (19b). It follows that only Zhangsan can long-distance antecede ziji. In (19c), Zhangsan loses the competition to Wangwu, which ends up as the only long-distance antecedent.

Let us now turn to the order in which candidates compete for prominence. Recall that Hu and Pan adopted a strictly linear procedure for prominence competition:

(20) A Linear Procedure for the Antecedent-Seeking Mechanism:

If there is an NP within an XP so that the latter c-commands ziji and the former is a member along with other such NPs in the linear NP sequence, i.e.

NP=( α

,…, α

, α), the reflexive ziji begins the search for the antecedent by comparing the prominence of α and α

and coindexes the NP marked as [+prominent] before running the computation on α

and α

and coindexes the NP marked as [+prominent] and so on.

However, with (20) Hu and Pan’s approach cannot work well if we try to deal with the following:

(21) Zhejian shi zhengming Zhangsan

-de baba

zhidao Lisi

bu xihuan ziji

. This-CL event prove Zhangsan’s father know Lisi not like self

This event proved that Zhangsan

’s father

knew that Lisi

didn’t like him

/himself

.

If we follow Hu and Pan’s proposal as it is tentatively, Lisi, being [+subject, +agent]

and [-dominating, +animate], ties with Zhangsan-de baba, being [+subject, +agent]

and [-dominating, +animate]. Lisi and Zhangsan-de baba are marked as [+prominent]

61 The label is my own invention, for ease of reference. (20) is one component of Hu and Pan’s Antecedent-Seeking Mechanism, although they have not named it separately.

62 Recall that agents include experiencers in their approach.

and correctly predicted to be antecedents of ziji. Zhangsan-de baba, being [+subject, +agent] and [+dominating, +animate], outranks Zhangsan, being [+subject, -agent]

and [-dominating, +animate]. Now comes the problem for (20): nothing bars

Zhangsan from entering into competition with zhejian shi. Zhangsan, being [+subject, -agent] and [-dominating, +animate] outranks zhejian shi, being [+subject, -agent] and [-dominating, -animate].

This means that Zhangsan is also predicted to be an

antecedent, contrary to fact.

On the other hand, we can make correct predictions for sentences such as (21) if we reject the strictly linear view of competition order and instead adopt the following procedure:

(22) A Not Totally Linear Procedure for the Antecedent-Seeking Mechanism:

(Let (16) and (18) determine the set of candidates first.) After prominence competition has taken place among the various candidates within an NP which is itself a candidate, only the NP, not the leftmost candidate within it, goes on to compete with a candidate on its left. Computation proceeds linearly for candidates not embedded in an NP.

According to (22), after Zhangsan has lost the competition to Zhangsan-de baba, the whole NP, not Zhangsan, goes on to compete with the matrix subject zhejian shi. The fact that Zhangsan cannot antecede ziji falls out from (22).

In this section, I have shown that some changes are necessary for a

competition-based approach to non-local binding à la Hu and Pan (2002). All these

63 According to Hu and Pan’s account, even if zhejian shi is treated as [+agent], Zhangsan would still incorrectly be predicted to be an antecedent. Zhejian shi, being [+subject, +agent] and [-dominating, -animate] would ultimately tie with Zhangsan, being [+subject, -agent] and [-dominating, +animate].

[+local] does not apply here, because there is no proper subset relation between their paths; the path from ziji to the minimal maximal projection dominating Zhangsan includes Zhangsan-de baba, but the path from ziji to the matrix IP, which dominates zhejian shi, does not pass through Zhangsan-de baba.

Rather, it passes through the embedded IP. This means that both are predicted to be able to antecede ziji—still an incorrect result. This discussion is intended to prove beyond reasonable doubt that (20) cannot be correct. The same conclusion can be reached even if we assume Zhangsan to be [-subject], a possibility that one might entertain because Hu and Pan do not explicitly state whether genitive NPs are considered subjects in their approach. I shall leave it to the reader himself to verify this point.

have been justified by empirical data. In the next section, I shall piece together all the modifications I have made to their approach and test them against more data.

3.3 Evaluating the Current Algorithm for Non-Local Reference of Ziji

In section 3.2, I proposed some major changes to Hu and Pan’s (2002) proposal, including (i) the separation of local binding from non-local binding; (ii) the

eliminability of prominence; (iii) the rejection of [+agent], [+local] and [+first/second person] from the stock of prominence factors; (iv) the refined candidate selection procedures; and (v) a not totally linear order for prominence computation. Below I offer an integrated picture of the results of the preceding section. The reader is recommended to compare the following with Hu and Pan’s (2002) original formulation.

(23) Antecedent-Seeking Mechanism:

In a candidate set ( α

,…, α

, α) defined by (24), ziji finds as its antecedent any NP marked as [+prominent] according to (26) and (27).

(24) Candidate Selection Procedure:

Given an XP c-commanding ziji, a candidate can be a. any NP c-commanding ziji;

b. any NP c-commanding the head of the XP when the latter is a non-NP;

c. any subject NP α contained within the XP β which is a subject when any argument containing α is a subject.

(25) Prominence Hierarchy:

A. [+subject] > [-subject]

B i. [-dominating, +animate] > [-dominating, -animate]

ii. [+dominating, +animate] > [-dominating, +animate]

iii. [+dominating, -animate] > [-dominating, -animate]

iv. [-dominating, +animate] > [+dominating, -animate]

64 Actually, we will demonstrate that Biii. is unnecessary in upcoming discussion.

B > A

Given two candidates α and β
^ only β, not α, is marked as [+prominent] if α outranks β in terms of A but is outranked by β in terms of B.

(26) Procedure for Prominence Competition

a. In a candidate set ( α

,…, α

, α), α and α

compete for prominence as per (25), before α

competes with α

in a linear fashion, until α

has competed with a

, except when (b) applies:

b. Given a candidate which itself dominates other candidates, only it goes on to compete with a candidate on its left, after prominence competition has taken place among the candidates it dominates.

c. When the competition is tied, both candidates are marked as [+prominent].

(27) Prominence Elimination:

A candidate marked as [+prominent] during the competition can have this feature eliminated, provided it is outranked by another candidate.

Having presented the algorithm in full, let us test the approach against some typical sentences. The most typical ones are those with animate subjects in both the local and the matrix clauses. Consider (28) and its structure in (29):

(28) Zhangsan

renwei Lisi

zhidao Wangwu

taoyan ziji

. Zhangsan think Lisi know Wangwu hate self

Zhangsan

thinks Lisi

knows Wangwu

hates him

/himself

.

(29)

65 The notion of animacy will be elaborated by an animacy hierarchy later in this chapter.

Zhangsan NP

I

renwei V

C

Lisi NP

I

zhidao V

Wangwu taoyan ziji CP

V' VP I' IP C' CP V' VP I' IP

Following (24), we have three candidates—Wangwu, Lisi, and Zhangsan. Each is the subject of their respective clause, and each c-commands ziji. (23) requires us to look for ziji’s antecedents as per (26) and (27). Let Wangwu compete with Lisi first, and this is a tied competition: both are [+subject] and [-dominating, +animate] and therefore both are marked as [+prominent] and are legitimate antecedents. Now let Lisi compete with Zhangsan, and again they tie for the match, as both are marked as [+prominent] in the way Wangwu and Lisi are. The possibility of long-distance binding is therefore predicted. At this point we see no advantages of our approach over previous ones, as they can handle (28) as well.

However, consider (30), a sentence with a sub-commanding NP as the antecedent of ziji:

(30) Zhangsan

-de jiaoao

hai-le ziji

(Tang 1989) Zhangsan’s pride hurt-Perf self

66 Of course, except for Hu and Pan (2002). I have demonstrated that their approach cannot predict the highest subject NP to be an antecedent in a long-distance context in Chapter 2 and section 3.1.

Zhangsan’s pride hurt him.

(31)

According to (24c), there are two candidates, Zhangsan and Zhangsan-de jiao, as both are subjects; the former is a genitive NP and the subject of the containing NP, and the latter is the subject of the sentence. Zhangsan outranks Zhangsan-de jiaoao, because the former is [+subject] and [-dominating, +animate] and the latter is [+subject] and [+dominating, -animate]. Zhangsan is therefore predicted to be the antecedent. Next consider (32), with the containing NP being an animate NP. The only way in which (32) differs from (30) is that the subject of (32) is animate. Zhangsan-de baba, being [+subject] and [+dominating, +animate], outranks Zhangsan, being [+subject] and [-dominating, +animate], and is marked as [+prominent] and able to antecede ziji, a result consistent with fact.

(32) Zhangsan

-de baba

dui ziji

mei xinxin. (Cole et al. 1993) Zhangsan’s father to self no confidence

Zhangsan’s father has no confidence in himself.

Although Tang’s (1989) notion of sub-command and Cole et al.’s (1993) feature percolation can account for the anaphoric relation, it is important to recall that they are based on the incorrect assumption that inanimate NPs cannot antecede ziji. Our

Zhangsan-de NP

jiaoao N N' NP

I

hai-le V

ziji NP V' VP I' IP

approach attributes the binding relation to the result of prominence competition and accounts for the following coreference relations, with our assumption that local binding and non-local binding are given separate treatments:

(33) Zhangsan

sheji-de zhezuo gongyuan

you ziji

-de tese.

Zhangsan design-DE this-CL park have self’s feature

The park that Zhangsan designed displays his/its own features.

(34)

Zhangsan sheji-de CP

zhezuo gongyuan NP

NP

I

you V

ziji-de tese NP V' VP I' IP

According to the Candidate Selection Procedure, Zhangsan sheji-de zhezuo gongyuan, a subject c-commanding ziji, and Zhangsan, the subject of the relative clause

contained by a subject, are candidates. Zhangsan, being [+subject] and [-dominating, +animate] competes with and outranks Zhangsan sheji-de zhezuo gongyuan, [+subject]

and [+dominating, -animate]. It is predicted that only Zhangsan can antecede ziji, as in the analyses of Cole et al. (1993), Huang and Tang (1991), and Huang and Liu (2001). But as they assume that inanimate NPs cannot antecede ziji, they cannot explain why ziji can refer to Zhangsan sheji-de zhezuo gongyuan, an inanimate NP.

On the other hand, since it is assumed in the current approach that inanimate NPs can

antecede ziji and local and non-local binding should receive separate treatments, the

coindexation in question is not ruled out as impossible; it is the task of local binding

to establish the coreference between ziji and Zhangsan sheji-de zhezuo gongyuan.

Let us now turn to a less typical case of long-distance binding, with an inanimate NP as the subject of an intermediate clause:

(35) a. Zhangsan

renwei zhejian shi

zhengming Lisi

taoyan ziji

. Zhangsan think this-CL event prove Lisi hate self

Zhangsan thought that this event proved that Lisi hated him/himself.

b. Ni

shuo-guo naben shu

hai-le ziji

ma? (Pan (2001)) You say-Exp that-Cl book hurt-Perf self Q

Did you say that that book hurt you?

In (35a), there are three candidates, Lisi, zhejian shi, and Zhangsan, each

c-commanding ziji. They compete for prominence as per (26a). Lisi, being [+subject]

and [-dominating, +animate] outranks zhejian shi, being [+subject] and [-dominating, -animate], which is outranked by Zhangsan, being [+subject] and [-dominating, +animate]. Both Lisi and Zhangsan, but not zhejian shi, are marked as [+prominent]

and legitimate antecedents. Previous analyses, especially the head-movement account along the lines of Peter Cole et al., neglect such sentences and instead focus only on typical cases like (28), resulting in a binding theory that fails to account for (35b).

This is so, because they aim at maintaining a strict local relation between the

antecedent and the reflexive by movement to Agr or Infl and some feature checking

mechanism, and they encounter intractable difficulty in dealing with sentences such as

(35b), which has an inanimate intermediate subject that does not agree in phi-features

with the subject of the higher clause. Blocking is predicted to be observed, contrary to

fact. Also note that the strict locality thesis between the antecedent and the reflexive

via LF head-movement construes long-distance binding as a sequence of local

bindings, and this is fundamentally incompatible with (35b); ziji, once moved to the

Infl or Agr of the intermediate clause, finds no local antecedent, which is an NP that

agrees with it with respect to phi-features under Cole et al.’s assumption.

Let us now turn our attention to a sentence with a matrix object as well as an embedded clause in which ziji is an argument. Consider (36) and its structure (37):

(36) Zhangsan

gaosu Lisi

nei ge shagua Wangwu

shanghai ziji

(Cole et al. 2001) Zhangsan tell Lisi that CL fool Wangwu harm self

Zhangsan told Lisi that that fool Wangwu harmed him/himself.

As there are three NPs c-commanding ziji, there are three candidates in (36). Neige shagua Wangwu, being [+subject] and [-dominating, +animate], outranks Lisi, being [-subject] and [-dominating, +animate]. Therefore the former is marked as

[+prominent]. Lisi then goes on to compete with Zhangsan. As the latter is [+subject]

and [-dominating, +animate], it outranks the latter, being [-subject] and [-dominating, +animate]. Zhangsan, not Lisi, is then correctly predicted to be the long-distance antecedent of ziji. Such sentences as (36), however, do not sufficiently demonstrate the superiority of the current account over others, as the head-movement analysis can make the same prediction regarding (36). Once ziji adjoins to the matrix Infl or Agr as a head, the only NP c-commanding it is the matrix subject, Zhangsan. Therefore Zhangsan is correctly predicted as the long-distance antecedent, as in our account.

(37)

Zhangsan NP

I

gaosu V

Lisi NP

C

neige shagua Wangwu NP

I

shanghai V

ziji NP V' VP I' IP

C' CP V'

VP I' IP

However, consider (38), a sentence identical to (36) in structure:

(38) Zhejian shi

gaosu Lisi

Wangwu

taoyan ziji

. This-CL thing tell Lisi Wangwu hate self

This event told Lisi that Wangwu hated him/himself.

As in (36), Wangwu outranks Lisi, but Lisi outranks zhejian shi and is marked as [+prominent] because Lisi is [-subject] and [-dominating, +animate] and zhejian shi is [+subject] and [-dominating, -animate]. Recall that our Prominence Hierarchy

requires that [+dominating, +animate] > [+subject], i.e. a candidate A outranks a

candidate B if the former is outranked by the latter in terms of [+subject] but outranks

the latter in terms of [+dominating, +animate]. As a result, Lisi is correctly predicted

as the long-distance antecedent of ziji. On the other hand, (38) is not amenable to the

head-movement account, because the landing site of ziji, the matrix Agr, cannot be

c-commanded by the object, Lisi. Also note that the IP adjunction analysis can allow

Lisi to antecede ziji, as ziji is c-commanded by Lisi when it adjoins to the lower IP.

But the analysis cannot rule out Lisi in (36) as a binder, as ziji is c-commanded by Lisi as in (38). Thus, it seems that movement analyses are facing a dilemma: subject orientation of (36) has to be accounted for, but at the same time the non-subject antecedent of (38) should not be excluded. Only a competition-based approach seems able to handle both (36) and (38).

Next, consider the following sentence with a sub-commanding and a long-distance antecedent:

(39) Zhangsan

shuo Lisi

-de baogao

hai-le ziji

(Hu and Pan (2002)) Zhangsan say Lisi’s report harm-Perf self

Zhangsan said Lisi’s report harmed him.

There are three candidates, because both Zhangsan and Lisi-de baogao are NPs c-commanding ziji, and the latter is a subject which contains another candidate Lisi, the subject of the subject NP. Let Lisi-de baogao compete with Lisi. The latter, being [+subject] and [-dominating, +animate], outranks the former, being [+subject] and [+dominating, -animate]. Lisi is therefore predicted as an antecedent. Then Lisi-de baogao, not Lisi, competes with Zhangsan, according to (26b). As predicted, Zhangsan is a legitimate antecedent because it, being [+subject] and [-dominating, +animate] outranks Lisi-de baogao, being [+subject] and [-dominating, -animate].

In the above we have seen how the animate genitive NP contained inside a subject becomes [+prominent] and a legitimate antecedent by the mechanism I have proposed. Next we shall examine the case in which an animate subject is contained inside another animate subject. Consider the following:

(40) Zhangsan

shuo Lisi

-de baba

hai-le ziji

Zhangsan say Lisi’s father hate self

Zhangsan said Lisi’s father hated him/himself.

(40) contains three candidates—Zhangsan, Lisi, and Lisi-de baba. Lisi-de baba, being [+subject] and [+dominating, +animate], outranks Lisi, being [+subject] and

[-dominating, +animate]. According to our procedure for prominence competition, Lisi-de baba goes on to compete with Zhangsan. Both are marked as [+prominent]

because both are [+subject] and [-dominating, +animate]. It is therefore correctly predicted, as in all the analyses available, that Zhangsan, not Lisi, can be a non-local binder.

Next, let us consider a case in which a candidate contains three candidates (including itself). Our approach also handles it correctly.

(41) Zhang xiansheng

-de baba

-de yinmou

bei ziji

-de pengyou shipo-le. (Hu and Pan (2002))

Zhang Mr.’s father’s plot BEI self ’s friend discover-Perf Mr.Zhang’s father’s plot was discovered by self’s friend.

Let Zhang xiansheng compete with Zhang xiansheng-de baba first. The former, being [+subject] and [-dominating, +animate], is outranked by the latter, being [+subject]

and [+dominating, +animate]. Zhang xiansheng-de baba is marked as [+prominent].

Then it competes with Zhang xiansheng-de baba-de yinmou, being [+subject] and [+dominating, -animate]. Zhang xiansheng-de baba, being [+subject] and

[-dominating, +animate], outranks Zhang xiansheng-de baba-de yinmou, and is therefore marked as [+prominent] and a legitimate antecedent.

67 In this connection, compare (41) with the structurally identical example (66b) on p. 35. The latter allows the most embedded candidate to antecede ziji, whereas the former does not. Our approach cannot predict the coindexation in (66b). If we assume that syntactic binding reflects contextlessly preferable judgments, then we are not obliged to deal with (66b) because (66a), which is identical to (66b) except for lack of context, conforms to the preferable judgment predicted by our approach. Note that our treatment presupposes Pollard and Xue’s (1998) thesis that syntactic binding is optional.

We will now consider a case in which a candidate is in a prepositional phrase.

Consider the following contrast.

(42) a. Wangwu

xiang Zhangsan

zhengming-le Lisi

taoyan ziji

. Wangwu to Zhangsan prove-Perf Lisi hate self

Wangwu proved to Zhangsan that Lisi hated him/himself.

b. Zhe

xiang Zhangsan

zhengming-le Lisi

taoyan ziji

. This to Zhangsan prove-Perf Lisi hate self

This proved to Zhangsan that Lisi hated him/himself.

(43)

Wangwu NP

I

xiang P

Zhangsan NP PP

zhengming-le V

C

Lisi NP

I

taoyan V

ziji NP VP I' IP CP V'

V' VP I' IP

(42a) and (42b) share the same structure (43). Let us see how prominence competition

results in different coindexation possibilities. Consider (42a) first. There are three

candidates, Lisi, Zhangsan, and Wangwu. Zhangsan is a candidate because it fulfills

(24b), i.e. it c-commands the head of a non-NP XP c-commanding ziji. Let Lisi

compete with Zhangsan first. The former, being [+subject] and [-dominating,

+animate], outranks the latter, being [-subject] and [-dominating, +animate]. Lisi is

therefore [+prominent]. Zhangsan goes on to compete with Wangwu. The latter, being

[+subject] and [-dominating, +animate], outranks the latter, being [-subject] and [-dominating, +animate]. Therefore, Wangwu, not Zhangsan, is marked as

[+prominent]. The binding possibilities are correctly predicted. Now consider (42b).

Again, Lisi outranks Zhangsan. Zhangsan then competes with zhe. The former, being [-subject] and [-dominating, +animate], outranks the latter, being [+subject] and [-dominating, -animate]. Therefore, Zhangsan is marked as [+prominent] and

correctly predicted to be an antecedent. As with (36) and (38), (42) demonstrates how subject orientation is derived; it is a consequence of prominence competition.

Next, consider the following case, which contains a prepositional phrase, but the

“object” is not a candidate:

(44) Zhangsan

cong Lisi

nar tingshuo Mali

hen taoyan ziji

. (Cole et al. (2001)) Zhangsan from Lisi there hear Mali very hate self

Zhangsan heard from Lisi that Mary hated him/herself.

In contrast to (42), there are only two candidates—Mali and Zhangsan, both

c-commanding ziji. Lisi is not a candidate, because it does not c-command cong, the head of the PP c-commanding ziji. Here I treat nar as a monosyllabic localizer like shang, xia, etc., heading its own projection LP (cf. Huang et al. (2004)), which itself is a complement to a preposition. If so, then Lisi is contained inside an LP and cannot c-command cong. Now let Mali compete with Zhangsan. Both are [+subject] and [-dominating, +animate], so both are [+prominent] and are legitimate antecedents. A similar situation holds in (45) below. Here Lisi is further contained than in (44); now it is dominated under the complement to the localizer li. Therefore it is not a candidate and does not participate in prominence competition.

(45) Zhangsan

zai Lisi

-de jia-li tingshuo Mali

hen taoyan ziji

.

Zhangsan at Lisi’s home-inside hear Mali very hate self Zhangsan heard at Lisi’s home that Mary hated him/herself.

Now consider an apparent problem for our approach. In (46) below, Lisi does not c-command the head of the PP c-commanding ziji, but somehow it is a legitimate antecedent.

(46) Zhejian shi

dui Lisi

laishuo zhengming-le Wangwu

taoyan ziji

. This-CL event to Lisi about prove-Perf Wangwu hate self

This event proved, as far as Lisi was concerned, that Wangwu hated him/himself.

It appears that Lisi competes with zhejian shi and outranks the latter by way of being [-subject] and [-dominating, +animate]. However, pending an exact analysis, it is very likely that laishuo and Lisi form a phrasal constituent that is a complement to the preposition dui. If so, Lisi cannot c-command dui, and thus is not a candidate. How could it compete with zhejian shi and be marked as [+prominent]? I will argue below that it is actually not a candidate, and it is a discourse antecedent whose occurrence is licensed by the pragmatic requirement of being salient in the discourse (cf. Baker (1995)). Indeed, dui…laishuo is an expression whereby the speaker identifies with a third person NP and views things in his perspective.

But similar pragmatic, or rather logophoric, requirements are also argued to exist for Mandarin syntactically licensed long-distance antecedents (Cole et al. (2001); Pollard and Xue (2001)), so this is not a sufficient reason for treating Lisi as a discourse antecedent. It seems that we can only argue in a theory-internal manner. Compare (47) and (48). The only difference is the placement of dui Lisi laishuo.

68 Note, incidentally, that Huang and Liu’s (2001) functional phrases, e.g. a SELF phrase, cannot apply here and predict the grammaticality of (46). If we posit a SELF phrase above the lowest IP, the Spec of which hosts the reflexive, Lisi still would not be able to c-command ziji.

(47) Zhangsan

renwei zhejian shi

dui Lisi

laishuo zhengming-le Wangwu

taoyan ziji

.

Zhangsan think this-CL event to Lisi about prove-Perf Wangwu hate self Zhangsan thought that this event, as far as Lisi was concerned, proved that

Wangwu hated him/himself.

(48) Zhangsan

renwei dui Lisi

laishuo zhejian shi

zhengming-le Wangwu

taoyan ziji

.

Zhangsan think to Lisi about this-CL event prove-Perf Wangwu hate self Zhangsan thought, as far as Lisi was concerned, that this event proved that

Wangwu hated him/himself.

In both sentences, ziji can refer to Lisi. If Lisi were a candidate, we would expect it to be outranked by Zhangsan in (48), because the latter is [+subject] and [-dominating, +animate]. It therefore seems that Lisi does not really participate in prominence competition. Perhaps some speakers might feel such a coindexation is of reduced acceptability in (48), but the fact that it is still possible suggests the function of dui…laishuo is to empathize with a third person NP.

Next, let us consider a sentence with a locative phrase in subject position:

(49) Zhangsan

renwei

[Lisi

-de zhuo shang] you yizhang ziji

-de huaxiang.

Zhangsan think Lisi’s table on have one-CL self’s picture Zhangsan thought there was a picture of self on Lisi’s table.

Recall that a subject contained inside another subject is a candidate if any argument containing the former is a subject. The question is whether Lisi is a candidate. It has turned out that the answer hinges on how we analyze the localizer phrase. If Lisi-de zhuo is treated as the complement and the object argument to the localizer shang,

then Lisi would not be a subject NP and thus not a candidate.

If Lisi-de zhuo is

69 Huang et al. (2004) consider localizers to be lexical heads.

70 There are indeed speakers who question the acceptability of ziji referring to Lisi. For them, it is possible that Lisi-de zhuo is an argument to shang.