Acceptability and preference in the interpretation of anaphors(*).

KIM, SOO-YEON

Abstract

The purpose of this paper is to consider how the preference in the interpretation of anaphors and their distribution properties can interact and be best explained. To reach the goal, this paper presents the prominence hierarchy for computing a different degree of preference when there is more than one option for anaphor interpretations. This paper also argues that the coreferential possibility between the Korean anaphor caki and its antecedent is determined by the prominence principle, which is stated in terms of the prominence hierarchy: caki must be coreferential with a more prominent antecedent only if there exists such an antecedent. Finally, this paper extends its proposal to anaphors in other languages such as English, Icelandic, Japanese, and Chinese.

1. Introduction

A major theme in the recent literature on anaphors has been how to account for various uses of anaphors that are not predicted by the standard binding theory in Chomsky (1981). One of the most widespread phenomena regarding this issue is the existence of a so-called long-distance anaphor that includes caki in Korean. This is illustrated in the following Korean sentence, where caki refers to a long-distance antecedent:

(1) [TEXT NOT REPRODUCIBLE IN ASCII]

The most conspicuous attempt to resolve the long-distance anaphor in recent literature can be characterized as an LF-movement approach. According to the LF-movement analysis, anaphors undergo covert movement to an appropriate position near near antecedents (see Lebeaux 1983; Pica 1987; Cole et al. 1990; Cole and Sung 1994; Cole and Wang 1996; Hestvik 1992, among others). For example, caki in (1) is supposed to undergo covert movement to a position close to its antecedent Cheli. This attempt has an advantage in that independently needed syntactic processes such as head movement and spec--head agreement can account for the locality problem in long-distance anaphora. A substantial amount of redundancy of constraints could be eliminated if the chaotic uses of anaphors were uniformly accounted for by the LF-movement approach. However, the variety in the distribution of the Korean anaphor caki is not limited to long-distance anaphora. Caki can have a non-c-commanding antecedent, as in (2a), and can also have its antecedent in the preceding discourse, as in (2b):

(2) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Concerning the coreferential relations illustrated above, there have been a series of attempts to distinguish discourse-dependent anaphors from structure-dependent anaphors. The former are claimed to be constrained by nonstructural constraints such as logophoricity, point of view, or empathy (see Clements 1975; Sells 1987; Kuno 1987; Zribi-Herts 1989; Reinhart and Reuland 1993, among others) while the latter are constrained by a c-command requirement or requirements on head movement. If this is the case, the next question is, what are the criteria distinguishing these two types of anaphor? We cannot simply conclude that it is a certain discourse context such as logophoricity that allows the various uses of caki that seem to be exempt from violations of condition A.(1) The interpretation of caki illustrated in (2) does not require any specific discourse context: the sentences in (2) do not have to be stated) in Suni's point of view.

Given these, this paper will propose a unified approach that covers both instances in (1) and (2): the coreferential possibility between an anaphor and its antecedent can be captured by the prominence principle, which is stated in terms of the prominence hierarchy. The prominence principle on caki can be obviated if there is no potential antecedent.(2) This provides a straightforward account for an otherwise puzzling aspect of the following minimal pair. Subject caki in (3) can refer to genitive Yengi only when there is another potential antecedent in the same sentence:

(3) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

The interpretation of caki is sensitive to the existence of a potential antecedent in the given sentence. The genitive Yengi in (3a) cannot be an antecedent of caki because of the existence of a third person topic Cheli. However, caki in (3b), which is exactly the same structure as that in (3a), can be coreferential with Yengi.

The existence of the prominence hierarchy in grammar is motivated by the need to account for a degree of preference in the interpretation of caki when caki is ambiguous. For example, caki in (4) can refer to either Cheli or Yengi. When there is no special discourse context given, caki strongly prefers Cheli to Yengi as its antecedent; still, Yengi can be an antecedent of caki if a proper context is given.

(4) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

This paper is organized as follows: section 2 illustrates various uses of the Korean anaphor caki and suggests the prominence principle for anaphors (hereafter, PP-A) in Korean, which is identified in terms of the prominence hierarchy. Extending the PP-A to English anaphors, section 3 proposes a parameterized definition of a potential antecedent (PA) and parameterized pragmatic conditions to filter out overgenerated coreferential relations. In section 4, the PP-A in tandem with pragmatic filters will be expanded to anaphors in Chinese, Japanese, and Icelandic. Section 5 provides concluding remarks.

2. The prominence hierarchy

2.1. Preference in the interpretation of caki

This section discusses one of the most important properties of caki interpretation: there is a clear-cut difference in the degree of preference when a given sentence is ambiguous. In (5), topic Cheli is strongly preferred to subject Yengi as an antecedent of possessive caki:

(5) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Still, subject Yengi can be an antecedent of caki if an appropriate discourse context is given, as illustrated in (6):

(6) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

We propose that the grammatical role of a given expression plays a role in determining the preference of the interpretation. Caki prefers a grammatically more prominent expression to a less prominent one as its antecedent. The relative grammatical prominence comes from the prominence hierarchy:(3)

(7) Prominence hierarchy:

 topic > subject > object of verb > object of postposition > genitive NP >
 object of comparative

(8) a. Caki can be interpreted as coreferential with any third person NP that it does not precede in the prominence hierarchy.

b. When there is more than one competing interpretation of caki, the larger the gap between an antecedent and caki in the prominence hierarchy, the more preferred the interpretation.

The gap is the distance between the position of an antecedent and that of caki in the prominence hierarchy. If the former is lower than the latter, the gap assumes a negative value. The gap is computed by counting the number of symbols ">" between the antecedent and caki.

Now, let us see how the degree of preference in antecedenthood can be computed. Caki in (5a) (repeated here for ease of reference) is an object and has two more potential antecedents: topic Cheli and subject Yengi. According to the statement in (8a), caki in (5a) can be coreferential with either of these two and prefers topic as its antecedent since topic is higher in the hierarchy than subject.

(5) a. [TEXT NOT REPRODUCIBLE IN ASCII]

a'. b. [TEXT NOT REPRODUCIBLE IN ASCII]

Likewise, caki in (5b) is a genitive that can be coreferential with either topic Cheli or subject Yengi. Topic is preferred to subject as an antecedent of genitive caki:(4)

(5) b. [TEXT NOT REPRODUCIBLE IN ASCII]

b'. [TEXT NOT REPRODUCIBLE IN ASCII]

It is of importance to note that the domain for applying the prominence hierarchy in interpretation of a given anaphor is not a clause but a sentence. Sentence differs from clause in that the former denotes a freestanding clause that may contain more than one clause.(5) Since prominence is compared regardless of clause boundary, caki in the embedded clause in (5) can refer to the antecedent in the matrix clause.

We have discussed how we can compute the relative degree of preference when the interpretation of caki is ambiguous. Caki prefers a relatively more prominent antecedent to a less prominent one. Caki can also be coreferential with a relatively less prominent antecedent if a proper discourse context that favors the specific interpretation is given. In what follows, we examine cases where an interpretation of caki is unacceptable. The prominence hierarchy employed to account for the preference can also be used to explain the acceptability of caki interpretations. Unacceptable interpretations differ from unpreferred interpretations in that the former cannot be remedied by giving a discourse context that favors the interpretation, while the latter can.

2.2. Acceptability in the interpretation of caki

This paper defines an anaphor as an expression whose interpretation requires reference to its linguistic antecedent.(6) Among anaphoric expressions, demonstratives whose reference is from an entity in the world are excluded from the domain of research. Given this definition, the following shows that the Korean anaphoric expression, caki, clearly belongs to the category of anaphor that cannot be used deictically, as illustrated in (9):

(9) [TEXT NOT REPRODUCIBLE IN ASCII]

It is well known that syntactic constraints on the distribution of the Korean caki are loose, in the sense that caki can refer to almost any third person NP depending on the discourse context. These descriptions of caki come from the following uses of caki, where caki refers to a long-distance antecedent as in (10) or a non-c-commanding antecedent as in (11):

(10) [TEXT NOT REPRODUCIBLE IN ASCII]

(11) [TEXT NOT REPRODUCIBLE IN ASCII]

Apparently, caki in (10) seems to show the general properties of a simplex anaphor that is standardly classified as a long-distance anaphor.(7) First, caki in the embedded clause refers to a matrix subject, which is not local. Second, caki does not refer to the object Mini, implying that a long-distance anaphor caki is subject-oriented, like other long-distance anaphors illustrated in (12):

(12) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

According to the LF-movement approach that is based on the proposals in Lebeaux (1983) and Chomsky (1986), long-distance anaphors undergo coven movement to a position close to their antecedents. Since anaphors receive their interpretation through spec-head agreement between subject and an anaphor in Infl position, long-distance anaphors exhibit subject orientation (Cole et al. 1990; Hestvik 1992). If this is the case, we can be free from the burden of answering many questions about long-distance anaphors by adopting the LF-movement approach. For instance, apparently, the LF-movement approach can answer why monomorphemic anaphors can refer to a long-distance antecedent while multimorphemic anaphors in general cannot. It also seems to answer why they are subject-oriented when they refer to a long-distance antecedent. However, this relief comes with another burden in anaphor resolution: how are we going to account for the behavior of anaphors illustrated in (11)? Caki in (11) is coreferential with a non-c-commanding antecedent, John. One way to solve this problem is to admit the existence of a different type of category, such as discourse anaphors (Reinhart and Reuland 1993; Kuno 1987) or logophoric anaphors (Clements 1975; Sells 1987). It has been proposed that anaphors seem to be allowed to obviate condition A in many languages under certain discourse contexts, such as point-of-view, logophoricity, or focus (Thrainsson 1991: Hellan 1991; Reinhart and Reuland 1993). Interestingly however, caki in (11) does not require any specific discourse context. Still, it requires a linguistic antecedent, as we can see in the following:

(13) [TEXT NOT REPRODUCIBLE IN ASCII]

Therefore, we cannot simply rely on the structural approach with a strategy to allow exceptions for context-dependent (or discourse) anaphors.

We have discussed in the previous section that the degree of preference in the interpretation of cam can be calculated by counting the gaps between positions in the prominence hierarchy. In all cases, caki refers to a more prominent expression in terms of grammatical function. To put it in terms based on the prominence hierarchy, caki requires interpretation as coreferential with a third person NP that it does not precede in the prominence hierarchy. In (10), direct object caki refers to topic Cheli-uy hyeng, while it can be coreferential neither with genitive Cheli nor with indirect object Mini. So-called subject orientation in long-distance anaphors is due to the prominence of the subject. Likewise, the object of comparative caki in (11) refers to the genitive antecedent John regardless of c-commanding relations between them:

(10') the gap between Cheli-uy heng (topic) and caki (object) = 2
 the gap between Cheli (genitive) and cam (object) = -2
 the gap between Mini (object) and caki (object) = 0

(11') the gap between John (genitive) and cam (object
 of comparative) = 1

We can see that caki cannot refer to an expression if the gap value between them is negative.

However, it is not always the case that caki cannot refer to an expression when the gap between caki and the expression has a negative value, as the contrast between the (a) and (b) sentences in (14)-(15) shows. The interpretations given in (14a) and (15a) fall under the generalization in (8a): the subject caki cannot refer to the less prominent NP Yengi. Why are the interpretations in (14b) and (15b) acceptable, where caki refers to a less prominent NP Yengi?

(14) [TEXT NOT REPRODUCIBLE IN ASCII]

(14) [TEXT NOT REPRODUCIBLE IN ASCII]

The only difference between the (a) and the (b) cases in the examples (14)-(15) lies in the person feature of the topic. The former has a third person NP as a topic, whereas the latter has a first person NP that cannot be an antecedent of caki as a topic. Caki can refer to a less prominent third person NP when there does not exist a more prominent NP than itself (caki). In (14b) and (15b), caki does not have a more prominent third person NP than itself in the given sentences. Hence, it is allowed to refer to a less prominent NP. In other words, caki refers to an NP yielding a negative gap value only if there are no competing antecedents yielding a positive gap value.

Let us suggest that there is a potential antecedent (hereafter PA) for anaphors in a language. PA is a term that is technically defined: in Korean, PA is a third person NP that is more prominent than caki. It is significant to note that a real (actual) antecedent does not necessarily have to be a PA. A PA does not have to be a real antecedent either. In other words, when there does not exist a PA in a given discourse, caki can refer to an NP that is not a PA as long as the NP meets other requirements such as those of morphology or pragmatics: caki can refer to an NP with a negative gap value only if there is no PA. I propose the following generalization for the interpretation of caki:

(16) Prominence principle for anaphors in Korean (PP-A):

 Caki must be coreferential with a potential antecedent (PA) only if there
 exists a PA.

 N.B. 1. A PA for caki is a third person NP that is more prominent than
 caki.

 2. [Alpha] is more prominent that [Beta] iff [Alpha] proceeds [Beta] in the
 prominence hierarchy.

(17) Interpretation rule for anaphors in Korean (IR):

When there is more than one competing antecedent for caki, the larger the gap between an antecedent and caki in the prominence hierarchy, the more preferred the interpretation.

 N.B. 1. The gap is the distance between the position of antecedent and that
 of caki in the prominence hierarchy. If the former is lower than the
 latter, the gap assumes a negative value.

 2. A higher pair (e.g. topic antecedent and subject caki) is more natural
 than a lower pair (genitive NP antecedent and object of comparative caki).

(18) Prominence hierarchy:

 topic > subject > object of verb > object of post position > genitive NP

Now let us see how (16) works. As schematically represented in (19), caki can be interpreted as coreferential with an NP that is not a PA if there does not exist a PA. If there is more than one third person NP and if only one NP is a PA, caki has to refer to the PA. If there is more than one PA, caki has an ambiguous interpretation with a degree of preference according to the interpretation rule in (17):

(19) a. ... [NP.sub.i] ... [caki.sub.i] ...
 [-PA]

 if there does not exist a PA for caki, then caki can refer to
 [NP.sub.i].

 b. ... [NP.sub.i] ... [NP.sub.j] ... [caki.sub.i/*j] ...
 [+PA] [-PA]

 if [NP.sub.i] is the only PA for caki, then caki must
 refer to the PA,
 [NP.sub.i].

 c. ... [NP.sub.i] ... [NP.sub.j] ... [caki.sub.i > j] ...
 [+PA] [+PA]

 if both [NP.sub.i] and [NP.sub.i] are PAs for caki and if
 [NP.sub.i] is more prominent than [NP.sub.j], then caki has an
 ambiguous interpretation and [caki.sub.i] is prefered to
 [caki.sub.j].

In other words, the PP-A in (16) implies that if there does not exist a PA for caki in a given sentence, caki can refer to a less prominent NP that is not a PA. Again, there is no c-command requirement on a PA for caki: any third person NP that is more prominent than caki can be a PA for caki. It is now clear why caki in (14b) and (15b) can refer to Yengi, whereas caki in (14a) and (15a) cannot. In (14b) and (15b), there is no PA for caki. Hence caki in (14b) and (15b) is exempt and is allowed to take a less prominent antecedent. In contrast, there is a PA, Cheli, in (14a) and (15a); and caki has to refer to the PA, according to the prominence principle in (16).

So far, we have seen that the prominence principle for anaphors (PP-A) can account for problematic sentences for previous approaches as well as the degree of preference in the caki interpretation. The PP-A does not employ the structural c-command condition, allowing for caki to refer to a non-c-commanding antecedent. In addition, the PP-A applies only when there is a more prominent antecedent than caki. As a consequence, the constraints on caki interpretation in syntax are so weakened that there is overgeneration of some unacceptable coreferential relations. In what follows, we propose pragmatic conditions that filter out overgenerated coreferential relations.

2.3. A pragmatic condition

According to the PP-A in (16) in the previous section, caki can be coreferential with any NP as long as this NP is more prominent than caki. If there is more than one PA, caki has an optimal interpretation when it refers to the most prominent NP among the PAs (cf. the interpretation rule in [17]). Here, we have to note that there are cases where caki cannot refer to a more prominent NP than itself because of the presence of an intervening NP. A case in point occurs in scrambled structures of the type in (20). In (20), there are two potential antecedents of cam that are more prominent than caki. The PP-A predicts that caki can be coreferential with either of them. However, if we scramble the indirect object NP Suni in (20a) over the topic NP Cheli-nun, the acceptability changes, as in (20b):

(20) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

(21) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Interestingly, this contrast does not appear when the subject is first person na `I', as in (21). In short, a more prominent PA, topic Cheli, seems to block a coreferential relation between the indirect object Suni and caki in (20b), whereas na `I', which is not a PA, does not. To capture this, I suggest the following condition:(8)

(22) Prominence link condition (tentative):

 Given that [A.sub.i] and [B.sub.j] are PAs, when [A.sub.i] is not more
 prominent than [B.sub.j], the acceptability of the interpretation of the
 anaphor is degraded.

 [... [A.sub.i] ... [B.sub.j] ... [anaphor.sub.i] ...]

The prominence link condition defined in (22) implies that distance weakens antecedenthood: the farther away the PA is, the weaker the antecedenthood becomes. It should be noted that a violation of pragmatic filters differs from a violation of grammatical conditions: the former only degrades the acceptability, whereas the latter makes the sentence unacceptable. If we set up a discourse context where the given anaphor may well refer to a certain antecedent under an interpretation of caki that violates the prominence link condition, that interpretation can be acceptable. For example, caki in (23) violates the prominence link condition when it refers to Suni: object Suni is not more prominent than subject Cheli. However, if we give a discourse context, as in (24), the given interpretation becomes acceptable:

(23) [TEXT NOT REPRODUCIBLE IN ASCII]

(24) [TEXT NOT REPRODUCIBLE IN ASCII]

In short, the interpretation of the Korean anaphor caki is clearly accounted for by the PP-A in (16), which is defined in terms of the prominence hierarchy. A pragmatic condition such as the one in (22) interacts with grammatical conditions (i.e. PP-A) in such a way as to degrade acceptability for a set of sentences that are otherwise grammatical.(9)

2.4. The domain for the prominence principle

There remains a question regarding the domain for determining a possible PA for caki. In section 2.2, we claimed that the domain of applying the prominence hierarchy is a sentence. For instance, caki in the embedded clause in (25) prefers a topic in the matrix clause to the subject in the embedded clause as its antecedent since the domain of checking its prominence is not restricted to the clause where caki appears:

(25) [TEXT NOT REPRODUCIBLE IN ASCII]

At this point, we can raise a question whether the domain for applying the prominence principle in Korean can be extended to the previous sentences that share the same discourse topic. In this paper, we restrict the topic in the prominence hierarchy to a topic-marked NP in a given sentence. If a discourse topic were included in the domain of applying the prominence principle, Cheli in (26) would be the most preferred antecedent for caki:

(26) [TEXT NOT REPRODUCIBLE IN ASCII]

However, as shown in (26), caki cannot refer to the discourse topic but refers to the subject of the root sentence. Since the prominence principle applies to the sentence level, the interpretation given in (26) is accounted for. Likewise, if the domain of the prominence principle were extended to the previous sentence, which shares the same topic, the following case where caki can refer to John would not be explained: there is a more prominent PA, Bill, and caki has to be coreferential with it.

(27) [TEXT NOT REPRODUCIBLE IN ASCII]

Since the search for a PA stops at the sentence level according to the current theory, (27) will be the. case where the prominence principle will be vacuously satisfied.(10)

2.5. Other hierarchies in grammar

A thematic hierarchy has been claimed to play a role in anaphora theories such as in Jackendoff (1972), Giorigi (1984), and Kiss (1991). For instance, Kiss (1991: 249) argues that one of the factors that are relevant to coreferential possibilities between anaphors and their antecedent is lexical thematic prominence:

(28) An anaphor must have an antecedent that precedes in the lexical argument hierarchy.

The lexical thematic hierarchy of Kiss (1991) is claimed to be the hierarchy of [Theta]-roles associated with a given head in the lexicon. For instance, Kiss (1991:251) illustrates the [Theta]-grid of the verb megmutat `show', which explains the acceptability of the two sentences in (30), as follows:

(29) megmutat, + V - N

 (i) <1,2,3,E> agent (1), theme (2), beneficiary (3)

 (ii) <1,2,3,E> agent (1), beneficiary (2), theme (3)

(30) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Despite the differences in the details in the theory, scholars who argue for the role of argument hierarchy agree on the fact that agent is higher than theme in the argument hierarchy. However, in case of anaphors in Korean, [Theta]-role does not seem to be a crucial factor in determining the acceptability or preference of the coreferential relations. First of all, consider the following pairs of sentences:

(31) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

(32) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

In the (a) sentences above, caki, which is an object and a theme, can be coreferential with the subject NP, which happens to be an agent. In the (b) sentences, however, caki, which is a subject and a theme, cannot be coreferential with an agent that is lower than subject caki in the prominence hierarchy. The acceptability of the sentences given in (31) and (32) can be a piece of evidence to support the importance of the role of the prominence hierarchy in the domain of Korean anaphora.

Moreover, there is another piece of evidence that favors the prominence hierarchy over the argument hierarchy. As discussed in the previous section, there is a difference in the degree of preference when caki has an ambiguous interpretation. For instance, in (33), the prominence principle can predict that Cheli is strongly preferred to Yengi:

(33) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

In both cases, Cheli is more prominent than Yengi. If we compare the O-roles of Cheli and Yengi, Cheli in (33a) and Yengi in (33b) are positioned high in the argument hierarchy since they are the agent. The data set in (33) at least shows that the argument hierarchy is not relevant to determining preference in the interpretation of ambiguous caki.

It is of importance to note that there are many proposals that make reference to hierarchies based on the grammatical function of an expression in grammar. For instance, Keenan and Comrie (1977: 66) proposed the accessibility hierarchy to express relative accessibility to relativization of NP positions:

(34) SU [is greater than] DO [is greater than] IO [is greater than] OBL [is greater than] GEN [is greater than] OCOMP

Their proposal of the accessibility hierarchy is based on data from 50 languages. The accessibility hierarchy in (34) is similar to the prominence hierarchy of this paper in that both rely on the grammatical function of an NP. Keenan and Comrie (1977) claim that the accessibility hierarchy directly reflects psychological ease of comprehension. If this is the case, the accessibility hierarchy can be further supporting evidence to confirm the necessity of a hierarchy based on grammatical functions in the theory of grammar.

3. The parameterization of the prominence principle for anaphors

3.1. The parametric definition of a potential antecedent: a PA in English

Pollard and Sag (1992) (hereafter, P&S) propose an account of English anaphor binding in terms of an obliqueness hierarchy that is based on relational obliqueness:(11)

(35) subject [is less than] primary object [is less than] second object [is less than] other complements

(36) Principle A:

A locally o-commanded anaphor must be locally o-bound.

Principle A of P&S can be loosely interpreted as follows: when an anaphor has a coindexed NP in a subcategorization list (i.e. SUBCAT lists), it has to be more oblique (in our terms, less prominent) than the coindexed NP (i.e. its antecedent). P&S distinguished an anaphor that obeys principle A in (36) from an exempt anaphor that does not have to be bound by a less oblique NP. Their main idea is that an anaphor must be coindexed with a less oblique coargument, if there is one. In other words, their locally o-bound requirement on anaphors is restricted to locally o-commanded anaphors. Hence, an anaphor that does not have a locally o-commanding antecedent (i.e. less oblique antecedent in its SUBCAT list) is free from principle A of P&S in (36).(12) P&S use the term coargument to represent NPs in the same SUBCAT list. In this section, after defining coargument as in (37), I show how P&S's analysis works:(13)

(37) [Alpha] and [Beta] are coarguments iff [Alpha] and [Beta] are arguments of the same predicate.

(38) a. [John.sub.i] hates [himself.sub.i].

b. *[Bill.sub.i] thinks that [John.sub.j] hates [himself.sub.i].

c. *[John.sub.i] found [Bill.sub.j]'s picture of [himself.sub.i].

d. *[John.sub.i] said that Bill's pictures of [himself.sub.i] were on sale.

e. [John.sub.i] found a picture of [himself.sub.i].

f. [John.sub.i] said that pictures of [himself.sub.i] were on sale.

g. [John.sub.i] considers [himself.sub.i] intelligent.

Whenever an anaphor is more oblique than one or more referential elements on a SUBCAT list, then it must be coindexed with one of them (P&S 1992: 287). According to P&S, himself in (38a)-(38d) has to obey principle A in (36) because it has a less oblique coargument: in (38a) and in (38b), John is less oblique than himself. In (38c) and in (38d), the subject of NP, Bill, is less oblique than himself On the other hand, they regard himself in (38e) and in (38f), which does not have a less oblique coargument, as an exempt anaphor. In (38g), John and himself are the coarguments of a predicate consider intelligent.(14)

Likewise, each other in (39a)-(39c) is an exempt anaphor that does not have any less oblique (more prominent) coarguments. Following P&S and many others, we regard possessors in English as subjects. Possessor NPs are more prominent than object NPs in English.(15,16) Hence, each instance is exempt from principle A: it does not have to refer to less oblique coarguments because it does not have any less oblique coargument. Each other in (39d) has to refer to a less oblique argument because the subject the agreement and the primary object trading rights are present.

(39) a. The agreement that [[Iran and Iraq].sub.i] reached guaranteed each [other.sub.i]'s trading rights in the disputed waters until the year 2010.

b. [[John and Mary].sub.i]'s houses appealed to each [other.sub.i]'s taste (Reinhart and Reuland 1993).

c. [[John and Mary].sub.i]'s aggressive tactics weakened each [other.sub.i]'s positions in their arguments.

d. *The agreement that [[Iran and Iraq].sub.i] reached gave trading rights to each [other.sub.i] (P&S 1992).

Here, we can see a striking similarity between the interpretation of English and Korean anaphors: each other in (39a) is coreferential with a non-coargument, Iran and Iraq, because there does not exist a more prominent coargument of each other. On the other hand, each other in (39d) cannot refer to Iran and Iraq because of its coarguments, the agreement and trading rights. In Korean, caki in (40b) can refer to a less prominent NP, Suni, because there does not exist a more prominent third person NP, whereas caki in (40a) cannot refer to less prominent Suni because there is a more prominent NP, topic Cheli:

(40) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

The difference between English and Korean lies in the definition of potential antecedents (PAs). The former has a more prominent coargument as its PA, whereas the latter has a more prominent third person NP as its PA. Hence, we can generalize the prominence principle for anaphors (PP-A) as follows:(17)

(41) The prominence principle for anaphors:

If there exists an anaphor [Alpha] and a set of potential antecedent (PA) S, S = {[PA.sub.1], [PA.sub.2], [PA.sub.3], ...}, [Alpha] must be coreferential with a member of the set S.

(42) A parametric definition of PA:

Korean: more prominent third person NP English: more prominent coargument

It is crucial that the term PA be used in a technical sense. For example, according to the definition of PA in (42), even I in (43) can be a potential antecedent for himself because I is a coargument of himself.

(43) *[I.sub.i] like [himself.sub.i].

On the other hand, as we can see in (39a)-(39c) and in (40b), the antecedent of an anaphor is not a PA. A PA is not necessarily an expression that is morphologically a possible antecedent of a specific anaphor; neither is a PA a real antecedent of an anaphor. A PA is simply a coargument that is more prominent than the relevant anaphor in English, and a third person NP that is more prominent than caki in Korean.

There is another important similarity in the interpretation of anaphors in Korean and English. There is a difference in the degree of preference when the interpretation of an anaphor is ambiguous. Consider (44):(18)

(44) [Mary.sub.i] talked to [Jane.sub.j] about [herself.sub.i [is greater than] j].

(44') the gap between Mary (subject) and herself (object of PP) = 2 the gap between Jane (object) and herself (object of PP) = 1

The sentence in (44) has an ambiguous interpretation since there is more than one PA that is more prominent than the reflexive. The reflexive herself in (44) can be and must be coreferential with either of its coarguments because it is the object of a preposition, which is less prominent than an (in)direct object or subject. As the interpretation rule in (17) predicts, Mary is clearly preferred to Jane as an antecedent of herself.

By defining PA as in (42), we can distinguish an anaphor that has to obey the PP-A from one that is not subject to the PP-A. In English, an anaphor that does not have a more prominent coargument will not violate the PP-A. Likewise, in Korean, when there is not a more prominent third person NP in a given discourse, an anaphor can be coreferential with any NP as long as the coreferential relation obeys the rules of other subcomponents of grammar such as pragmatics, semantics, and morphology.

3.2. Parametric definition of the prominence link condition

Compared to the standard binding theory of Chomsky (1981), the PP-A in (41) is loose enough to allow unacceptable coreference relations of the type in (45).

(45) a. *[Tom.sub.i] said that [Bill.sub.j] thought that pictures of [himself.sub.i] would be on sale.

b. *[They.sub.i] made sure that [we.sub.j] would prevent each [other.sub.i]'s pictures from being on sale.

In other words, once we allow an anaphor that does not have a more prominent coargument to be exempt from the PP-A, we permit the coreferential relations in (45). The coreferential relations represented in (45) seem to be blocked by the presence of another PA, Bill and we.(19)

The unacceptable relations in (45) seem to be similar to the Korean cases. For instance, when there is a more prominent PA Cheli(20) between an anaphor and its antecedent Suni, acceptability is degraded, as in the following:

(46) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

(46') the gap between Cheli (topic) and caki (genitive) = 3 the gap between Suni (indirect object) and caki (genitive) = 1

There are two PAs, Cheli and Suni, for caki in (46). Cheli is more prominent than Suni because Cheli is a topic, which is ranked in the highest position of the prominence hierarchy. In (46a), where Suni is closer to caki than Cheli is, the coreferential relation between a relatively less prominent PA Suni and caki is acceptable.(21) If we change the word order in (46b), the acceptability is degraded: Cheli, which is a more prominent PA, intervenes between indirect object Suni and caki. In contrast, the acceptability of the coreferential relation in (47a) is not affected by a change in word order: the topic-marked NP na `I' in (47), which is not a PA in Korean, does not intervene in the coreferential relation. We see that a more prominent PA acts as an intervener in Korean.

(47) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Let us consider the English sentences in (48). In English, more prominent NPs do not intervene in the coreferential relation. Instead, animacy is a factor in such an intervention:

(48) a. *[They.sub.i] made sure that [we.sub.j] would not prevent each [other.sub.i]'s pictures from being on sale.

b. [They.sub.i] made sure that nothing would prevent each [other.sub.i]'s pictures from being on sale.

Each other in (48) does not have a PA, a more prominent coargument: any coreferential relation is possible as far as the PP-A is concerned.(22) However, the coreferential relation in (48a) is unacceptable, whereas that in (48b) is acceptable. In English, the animate NP we in (48a) disrupts the coreferential relation between they and each other, whereas the inanimate NP nothing does not.

Kuno and Takami (1993) claimed that semantic transparency is a relevant factor for intervener status: the specific semantic import that an intervener carries affects the coreferential relation between NPs. Their claim is significant in that the unacceptability caused by the presence of an intervener is shown as a continuum rather than a clear-cut difference. However, given the following example, where a semantically nontransparent NP does not block coreferential relations, I retain the proposal that animacy is a factor in determining the intervener:(23)

(49) a. (?)[They.sub.i] made sure that these two events that would take place around the same time would not prevent each [other.sub.i]'s pictures from being on sale.

b. *[They.sub.i] made sure that Mary's pictures on sale would not prevent each [other.sub.i]'s pictures from being on sale.

In our terms, Mary in (49b), which is an animate subject, blocks the coreferential relation between they and each other, which was allowed by the PP-A. On the other hand, the subject of the intermediate clause two events in (49a) does not disrupt the coreferential relation.

Based on these observations, we can apply to English the prominence link condition, which I have proposed for Korean, by parameterizing the definition of intervener:(24)

(50) Prominence link condition:

a. [... [A.sub.i] ... [B.sub.j] ... [anaphor.sub.i] ...]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

When [B.sub.j] is an intervener and [A.sub.i] is not more prominent than [B.sub.j], then the acceptability of the interpretation of the anaphor is degraded.

(51) A parametric definition of intervener:

 Korean: third person NP English: animate NP

In (48a), both they and we are subjects. (48a) fits the template in (50a): they is not more prominent than we. Hence, the animate NP we blocks the coreferential relation between they and each other in (48a). In contrast, the inanimate subject NP nothing in (48b) does not disrupt the coreferential relation between the matrix subject and the anaphor in the embedded clause. The unacceptable coreferential relations in (45) can be explained in the same manner. In (45), Tom and they are not more prominent than Bill and we respectively. Hence the inanimate subject in the embedded clause blocks the coreferential relation in (45).

The template in (50b) is suggested for cases like the following:

(52) a. [Their.sub.i] [brothers.sub.j] like each [other.sub.*i/j]'s toys.

b. [Their.sub.i] unfriendliness angered each [other.sub.i]'s wives.

c. [Their.sub.i] [brothers.sub.j]' unfriendliness angered each [other.sub.*i/j]'s wives.

Recall that genitives in English are ranked as high as subjects in the prominence hierarchy. The reciprocals in (52) do not violate the PP-A, because they do not have a more prominent coargument. According to (50), their brothers in (52a) is an intervener in the coreferential relation between their and each other because their is not more prominent than their brothers. In (52b), on the other hand, their unfriendliness cannot be an intervener because it is inanimate. Finally, we can see why their brothers in (52c) is the only possible antecedent for each other: their brothers' unfriendliness is an inanimate NP that does not qualify as an intervener. Their brothers is an intervener in the coreferential relation between each other and their according to (50b), because their is not more prominent than their brothers. Hence, only their brothers can be an antecedent for each other in (52c). Note that their brothers in (52c) is not a PA but an antecedent. As we saw earlier, PA is a technical term and does not imply capacity for antecedenthood. The same phenomena are observed in Korean:

(53) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

c. [TEXT NOT REPRODUCIBLE IN ASCII]

The subject caki in (53) does not have a PA: it cannot have an antecedent yielding a positive gap value. Hence, it can refer to a less prominent NP. In (53a), caki must be coreferential with Cheli, which is morphologically the only possible antecedent of caki. In (53b), there are two third person NPs, Cheli and Cheli-uy-hyeng `Cheli's brother'. Caki can be coreferential only with Cheli-uy hyeng. Caki in (53b) cannot refer to Cheli in the presence of Cheli-uy hyeng `Cheli's brother' because Cheli is not more prominent than Cheli-uy hyeng.(25)

Finally, the PP-A and the prominence link condition can account for the following contrast:(26,27)

(54) a. [John.sub.i]'s campaign requires that pictures of [himself.sub.i] be placed all over the town.

b. *[John.sub.i]'s father requires that pictures of [himself.sub.i] be placed all over the town

Because himself in (54) does not have a PA, it does not violate the PP-A. The PP-A allows himself to refer to John in both (54a) and (54b). In (54b), however, an animate NP, John's father, disrupts the coreferential relation between John and himself: John's father is an animate NP and John is not more prominent than John's father. According to the prominence link condition in (50b), the interpretation is degraded. John's campaign in (54a) does not act as an intervener because it is inanimate.

In sum, a parametric definition of PA in tandem with the prominence link condition can account for the apparently diverging distribution of anaphors in English and in Korean.

4. Extension of the prominence principle

4.1. Icelandic

Now let us examine how the prominence principle in (41) accounts for other anaphors such as Icelandic sig. Sig is well known as a reflexive that can have a long-distance antecedent under certain conditions.(28) Interestingly, sig differs from other long-distance anaphors in that sig with a local antecedent behaves differently from sig with a long-distance antecedent.

First of all, sig must have a subject antecedent when it refers to a long-distance antecedent:

(55) [Jon.sub.i] sag??i [Haraldi.sub.j] a?? Maria elska??i [sig.sub.i/*j].

[John.sub.i] told [Harold.sub.j] that Mary likes(subj) [self.sub.i/*j]

`[John.sub.i] told [Harold.sub.j] that Mary likes (subj) [self.sub.i/*j].'

However, when sig is coreferential with an expression within a clause, as in (56), it can refer to an object.

(56) [TEXT NOT REPRODUCIBLE IN ASCII]

Before defining PAs in Icelandic, we need to note that a structural requirement such as c-command plays a role only when sig has a local antecedent:(29)

(57) [TEXT NOT REPRODUCIBLE IN ASCII]

Sina cannot refer to the object Mariu in (57b), whereas sig in (57a) can refer to the object Haraldi. The same phenomena are observed when sig refers to a subject NP:(30.31)

(58) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

The unacceptable interpretation of ser in (58b) shows that it cannot be coreferential with a non-c-commanding NP Jon when the antecedent is within the same clause.(32) Note that sig can refer to a non-c-commanding NP when the antecedent is not within the same clause, as shown in (58a). However, it is not the case that any long-distance subject can be an antecedent of an anaphor in the embedded clause.(33)

(59) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Sig can refer to a long-distance antecedent such as Jon in (58a). However, sig in (59a) cannot be coreferential with the long-distance subject Jon; nor can sig in (59b) be coreferential with Mariu. To explain these data, it has been claimed that Icelandic anaphors are constrained by conditions involving logophoricity, point of view, or perspective holding (see Thrainsson 1976; Maling 1984; Sells 1987; Sigur??sson 1990, among others).

Now let us see how the prominence principle works in accounting for Icelandic reflexives. Recall that PA is a technical term designed for the prominence principle: a PA is not a real antecedent of an anaphor. If we defined PA as a subject in Icelandic, we could not account for the fact that an anaphor can have a local object as its antecedent even when its subject is present (e.g. [55]). If we defined PA in Icelandic as a more prominent coargument, we would face a problem regarding sig with a long-distance antecedent: sig can have a long-distance antecedent despite the presence of a local coargument:

(60) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Despite the fact that sig in (60) has a more prominent coargument, sig refers to a long-distance antecedent. Therefore I propose that there are two types of PA in Icelandic:(34)

(61) PA in Icelandic: subject, coargument.

If there is a subject in the given discourse or a more prominent coargument, anaphors in Icelandic must obey the prominence principle.(35) As expected, the prominence principle with a PA defined as in (61) will allow unacceptable coreferential relations. For example, the PP-A allows the coreferential relations in (58b) and (59b) (repeated here as [62a] and [62b], respectively):

(62) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Ser in (62a) satisfies the PP-A since it refers to a PA, the subject Jons.(36) Likewise, the interpretation under which sig in (62b) refers to the subject Mariu satisfies the PP-A. Based on these, conditions like the following are suggested for Icelandic sig:(37)

(63) A local antecedent must (PP-visible) c-command its anaphor in Icelandic.

(64) A long-distance antecedent of an anaphor must be a logophoric subject in Icelandic.

N.B. [Alpha] is a logophoric subject iff the embedded clause of the matrix verb that [Alpha] is the subject of represents the thought, speech, and perception of [Alpha] other than the speaker-narrator (Clements 1975: 169).

According to (63), ser in (62a) cannot refer to Jons because Jons does not c-command ser. (64) predicts that Mariu cannot be an antecedent of sig, because it is not a logophoric subject.(38,39)

Before closing this section, I briefly discuss the apparent subjunctive mood requirement for long-distance anaphors in Icelandic. It has been assumed that sig can refer to remote subjects only when the embedded clause is in the subjunctive mood:

 (65) a. [Jon.sub.i] segir a?? Maria elski [sig.sub.i].
 Jon says(ind) that Mary loves(subj) self
 `[John.sub.i] says that Mary loves [self.sub.i]'
 (Sigur??sson 1990: 310).

 b. ?[Jon.sub.i] veit a?? Maria elskar [sig.sub.i].
 Jon knows(ind) that Mary loves(ind) self
 `[John.sub.i] knows that Mary loves [self.sub.i]'
 (Sigur??sson 1990: 311).

Structure-based approaches such as the parametric approach of Wexler and Manzini (1987) and the LF-movement approach of Pica (1987, 1991) identify the subjunctive mood (more precisely, nonidicative mood) as a necessary condition for sig to refer to remote antecedents. However, Sigur??sson (1990: 313) mentioned that there are some speakers, including himself, who find indicative clauses containing a long-distance anaphor sig acceptable or marginally acceptable as in (65b). For those speakers, subjunctive mood is not a necessary condition for the long-distance interpretation of sig.

We have also seen that the subjunctive mood is not a sufficient condition for the long-distance interpretation of sig (e.g. [59]). Thrainsson (1976) was the first to mention that there are cases where sig in a subjunctive clause cannot refer to a remote subject when the clause containing sig is adverbial (see also Hellan 1991; Maling 1984; Sells 1987; Thrainsson 1991, among others).

 (66) a. *[Jon.sub.i] er her enn bo a?? eg skammi [sig.sub.i].
 John is here still although I scold(subj) self
 `[John.sub.i] is still here, although I scold
 (subjunctive) [self.sub.i].'

 b. [Maria.sub.i] segir a?? Jon se Jon her enn bo a?? eg
 Mary says(ind) that John is(subj) here still although I
 skammi [sig.sub.i].
 scold(subj) self

 `[Mary.sub.i] says that John is (subjunctive) still
 here, although I scold (subjunctive) [self.sub.i].'

Sig in (66a) cannot refer to Jon despite the subjunctive mood. Note that when the whole clause (the matrix and the adverbial clause in [66a]) is embedded under a verb of saying, as illustrated in (66b), sig in the embedded clause can refer to the remote subject Maria. Now, it becomes clear that the subjunctive mood is neither a sufficient nor a necessary condition for long-distance interpretations of sig. The factor that controls long-distance coreferential relations between sig and its antecedent seems to derive from the logophoricity condition mentioned in (64).

4.2. Chinese

Now let us examine how the prominence principle for anaphors accounts for Chinese ziji. Ziji, which can refer to any NP regardless of person morphologically, seems to have at least two conditions on its interpretation. First of all, the antecedent of ziji has to be an animate NP:

(67) a. [TEXT NOT REPRODUCIBLE IN ASCII]

Second, ziji shows a strong subject orientation.(40) Unlike Icelandic sig, which can refer to a nonsubject NP if the antecedent is in the same clause, ziji cannot refer to a local object:

(68) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

In (68a), ziji can refer only to the subject wo `I'. Likewise, ziji in (68b) cannot refer to the indirect object Lisi.

Now let us consider how the PP-A can apply to Chinese ziji. Based on the properties mentioned here, we suggest that a PA in Chinese is an animate subject.(41) The PP-A predicts that if animate subjects are present, ziji has to refer to one of them. Since a possessive NP such as Lisi in (67b) is regarded as a subject of NP in Chinese just as in English, ziji in (67b) refers to the only PA, Lisi. When there is more than one PA, ziji will have an ambiguous interpretation (e.g. [68b]).

As in other languages, the PP-A in Chinese will allow unacceptable coreferential relations, illustrated in (69), that is, [ziji.sub.i]. The examples in (69) show the so-called blocking effect in Chinese:

(69) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

In an LF-movement approach to long-distance anaphors, it is claimed that the blocking effect derives from a feature conflict.(42) However, the blocking effect shown in (69b) is a problem in an LF-movement approach: it can neither explain why ziji cannot refer to Zhangsan nor account for why ziji can refer to wo because there cannot be any person agreement between when ziji adjoins to the embedded Infl in (69b). Even if there were, it could not be the first person feature because wo `I' is not the subject of the embedded clause. Recall that both Korean and English have the prominence link condition as a filter that applies to overgenerated coreferential relations. If we define the intervener in Chinese as any animate NP with different [Phi]-features from the antecedent of ziji, the prominence link condition in Chinese, (70), will explain the unacceptable interpretations in (69):

(70) Prominence link condition in Chinese:

 a. [... [A.sub.i] ... [B.sub.j] ... [anaphor.sub.i] ...]

 b. [[[... [A.sub.i] ...].sup.[B.sub.j]] ... [anaphor.sub.i] ...] When
 [B.sub.j] is an animate subject with a different [Phi]-feature from
 [A.sub.i], and when [A.sub.i] is not more prominent than [B.sub.j], then
 the acceptability of the interpretation of the anaphor is degraded.

In (69a), the subject wo `I' intervenes in the coreferential relation between Zhangsan and ziji. Even when an intervener is embedded in the subject NP, as in (69b), a coreferential relation between Zhangsan and ziji is banned: because wo in (69b) is an animate subject, ziji refers to wo.(43) The fact that (70) is a pragmatic condition explains certain issues raised by an LF-movement approach: first of all, regardless of a c-command relation, an animate subject NP in (69b) intervenes in the coreferential relations, according to the prominence link condition in (70). Second, because (70) is a pragmatic condition, a violation of the condition in (70) does not lead to as severe a deviance as does a violation of the structural condition, especially when an appropriate pragmatic context is given. Therefore, in a context where Lisi's use of herself is not plausible, the following coreferential relation is marginally allowed, despite the violation of (70):

(71) [TEXT NOT REPRODUCIBLE IN ASCII]

In sum, Chinese also supports the proposal in this paper: structural constraints need to be weakened in explaining the interpretations of anaphors. Pragmatic filters rule out syntactically permitted coreferential relations.

4.3. Japanese

It has been assumed that only a subject can be an antecedent of zibun in Japanese:(44)

(72) [TEXT NOT REPRODUCIBLE IN ASCII]

However, as pointed out in McCawley (1976) (see also Iida 1992), subjecthood is neither a sufficient nor a necessary condition of zibun binding:

(73) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Zibun referring to a nonsubject in (73) demonstrates that subjecthood is not a necessary condition for an antecedent of zibun.

Sells (1987) compared Icelandic sig and Japanese zibun, regarding the logophoricity requirement. Sells cited the following examples from Kameyama (1984, 1985),(45) who claimed that the antecedent of zibun must be a subject or a logophoric NP:

(74) a. [TEXT NOT REPRODUCIBLE IN ASCII]

b. [TEXT NOT REPRODUCIBLE IN ASCII]

Unlike Icelandic sig in (74b), which does not allow a remote subject to be an antecedent of an anaphor unless it is a logophor, the nonlogophoric subject Taroo in (74a) can be an antecedent of zibun. Zibun can have a remote nonsubject antecedent if the antecedent is a logophoric NP, as shown in the following example:

(75) [TEXT NOT REPRODUCIBLE IN ASCII]

Nonsubject Takasi in (75) can be an antecedent of zibun when it is logophoric, that is, when the whole clause that contains zibun represents speech of Takasi. On the basis of these data, I propose that a PA for Japanese zibun is any subject or logophoric NP. If there is neither a subject nor a logophoric NP, zibun can refer to an expression that is neither of these (cf. [73]).

In Japanese, too, logophoricity (Sells 1987), deitic perspective (Iida 1992), or empathy (Kuno 1987) plays a role, but its role is different from that in Icelandic. In Icelandic, logophoricity is a condition that renders some coreferential relations unacceptable. In Japanese, logophoricity is a property that enables an NP to be a PA. Again, Japanese zibun supports the claim that syntactic conditions have to be weakened enough to allow all acceptable sentences as well as some unacceptable ones in the interpretation of an anaphor. Overgenerated sentences are filtered out by pragmatic conditions.

5. Conclusion

In sum, seemingly chaotic uses of anaphors are captured by the prominence principle that makes use of the prominence hierarchy based on grammatical function. The prominence hierarchy that is necessitated for capturing the different degree of preference in caki interpretation is employed to determine the acceptability of a given interpretation. Anaphors must refer to a potential antecedent (PA) if a PA exists. When there is more than one PA for a given anaphor, the anaphor has an ambiguous interpretation. The larger the gap between an antecedent and an anaphor in the prominence hierarchy, the more preferred the interpretation. The prominence principle applies to anaphors in other languages such as English, Icelandic, Chinese, and Japanese, employing the parameterized definition of a potential anaphor in a given language.

The current proposal has some advantages compared to earlier attempts to account for anaphor resolution. Empirically, it covers various uses of anaphors in a unified manner without any distinction between types of anaphors. As is well known, the goal that has been commonly pursued in treatments of anaphors is that all anaphors are subject to a single constraint such as condition A. However, if we pursue an approach based on structural constraints only, we end up having sets of acceptable sentences despite the violation of the constraints. For instance, there are anaphors with a non-c-commanding antecedent or even with an antecedent in the previous discourse. The current proposal, which does not rely on a structural requirement such as c-command, can capture the acceptability of the sentences without any problem. Moreover, it also captures the degree of preference, which the structural approach cannot. The numerical gap value clearly distinguishes the preferred interpretation from the unpreferred one. Theoretically, the current proposal distinguishes an unacceptable interpretation from an unpreferred one. Unacceptability, which is due to the violation of rules of syntax, remains constant regardless of the discourse context. In contrast, an unpreferred interpretation that comes from violations of semantic and pragmatic principles can be improved by changing the discourse context. Put another way, sentences that violate rules of semantics or pragmatics have unpreferred interpretations when they are produced out of the blue. However, they can become perfectly acceptable if the discourse context that strongly favors the given (unpreferred) interpretation is previously established. On the other hand, the degree of deviance that comes from a violation of the syntactic rules stays constant regardless of the change of discourse context: an unacceptable interpretation cannot become acceptable no matter how we set up the discourse context for it. In other words, we need to check where the deviance comes from when we discuss the data.

Since it is clear that we have the prominence hierarchy in grammar for the interpretation of anaphors, employing the same hierarchy in controlling the distribution of anaphors does not add any extra burden in the theory of grammar. Rather, it eliminates an unnecessary distinction between anaphors that seem to be exempt from violation of condition A and anaphors that do not. Insofar as there is a clear-cut preference in the interpretation of caki when caki is ambiguous, and insofar as native speakers of Korean uniformly show that preference, we need a hierarchy to explain those interpretations, implying that the prominence hierarchy is required in another component of grammar. If this is the case, employing the prominence hierarchy to explain the distribution of anaphors is not necessarily more costly in the grammar.

Sejong University

Received 4 May 1999 Revised version received 18 January 2000

Notes

(*) I am greatly indebted to Susumu Kuno and Hoskuldur Thrainsson, whose challenges and insightful discussion have improved almost every analysis presented here. I am also grateful to anonymous reviewers for their invaluable comments on the content of this article and its presentation. For helpful discussion, I would like to thank participants at the meetings of the Korean Generative Grammar Circle and the Korean Society for Language and Information, both in 1998. A preliminary report of the material developed in this article is available as Kim (1996 and i.p.). Correspondence address: Department of English Language and Literature, Sejong University, Kunja-Dong 98, KwangJin-Ku, Seoul, Korea 143-747. E-mail: kimsy@kunja.sejong.ac.kr.

(1.) Reinhart and Reuland (1993) distinguish two types of logophor that seem to be exempt from condition A. One is a "perspective logophor" that appears in the point-of-view contexts in the sense of Zribi-Hertz (1989) as illustrated in (i); the other is a "focus logophor" in (ii):

 (i) a. Max boasted that the queen invited Lucie and himself for a drink.

 b. It angered him that she -- tried to attract a man like himself.

 (ii) Bismarck's impulsiveness has, as so often, rebounded against himself.

As illustrated in (2), however, Korean caki does not require any specific discourse context such as perspective or focus. See also note 2.

(2.) A view similar to ours has been independently proposed for English anaphors (see Pollard and Sag 1992; Reinhart and Reuland 1993). For instance, condition A of Reinhart and Reuland (1993) concerns only a reflexive-marked predicate in English. If an anaphor does not occupy an argument position, condition A never applies (i.e. vacuously applies). The anaphors in examples (i) and (ii) of note 1 do not reflexive-mark a predicate (i.e. not in an argument position) and appear in a specific discourse context such as focus or point-of-view. However, the minimal pair in (3) seems to be different from English examples: first, both can be in the same discourse context. Second, both have exactly the same structure.

Pollard and Sag (1992) also proposed that condition A for English anaphors be restricted to those in argument position. The proposal of Pollard and Sag (1992) will be discussed in section 3.

(3.) On independent grounds, hierarchies based on grammatical function of a given NP have been proposed. For instance, Keenan and Comrie (1977) proposed the accessibility hierarchy, in (i), that determines the degree of accessibility to relative clause formation. The topic interpretation hierarchy of Kuno (1976) that shows which NP is most likely to become a topic, as in (ii), is also proposed. Despite minor differences among these, these hierarchies in tandem with the prominence hierarchy seem to give a strong emphasis to the importance of the grammatical role in the grammar.

(i) subject > direct object > indirect object > major oblique case NP > genitive > object of comparison

(ii) subject > direct object > indirect object > object of preposition > genitive > object of comparison

We will discuss these further in section 2.5.

(4.) The higher pair (e.g. topic antecedent > subject anaphor) is more natural than lower pairs (e.g. genitive NP antecedent > object of comparative anaphor) even when the gap values are the same.

 (i) [TEXT NOT REPRODUCIBLE IN ASCII]
 Cheli-Top self-Nom the best-be think
 lit. `[Cheli.sub.i] thinks that [self.sub.i] is the best.'

 (ii) [TEXT NOT REPRODUCIBLE IN ASCII].
 Yengi-Gen shoes-Top self-than more big-seemed
 lit. `[Yengi.sub.i]'s shoes seemed bigger than [self.sub.i]'

(5.) Other hierarchies in the grammar that employ the grammatical relation of the expression do not seem to be the same concerning the domain. For instance, the accessibility hierarchy in Keenan and Comrie (1977) limits its application to a simplex main clause. It is not clear why the domain of applying the prominence hierarchy in determining the preferred reading is extended to sentence level in Korean anaphora. This needs future research.

(6.) See Chomsky (1986) and Reinhart and Reuland (1993) for similar definitions of anaphors.

(7.) Pica (1985, 1987) claims that anaphors are classified into two groups: one is a simplex anaphor (or [X.sup.0]-anaphor, monomorphemic anaphor) that can have a long-distance antecedent and the other is a complex anaphor (or XP-anaphor, multimorphemic anaphor) that requires a local antecedent. According to this subcategorization, caki in Korean is included in the category of [X.sup.0]-anaphors with other anaphors such as ziji in Chinese, zibun in Japanese, and sig in Icelandic.

(8.) There have been proposals in the same spirit as the prominence link condition of this paper. They include the intervention constraint of Grinder (1970) and the semantico-syntactic chain-of-command principle of Kuno (1987). Basically, those two conditions are designed to filter out sentences where a weak potential binder has been chosen over a much stronger potential binder.

(9.) Pragmatic conditions include the humanness hierarchy condition and the empathy constraint proposed in Kuno (1987) (see also Kuno and Kaburaki 1977; Kuno and Takami 1993; Kim 1996 for additional conditions).

(10.) There still remains a question: what blocks the coreferential relation between caki and Mary in (27)? According to the prominence principle, caki in (27) freely refers to Mary also. When we change John in (27) to na `I' like the following, caki also refers to Bill only. This will raise the same question as (27) does regarding the unacceptable coreferential relation between caki and Mary.

 (i) [TEXT NOT REPRODUCIBLE IN ASCII].
 Bill-Top yesterday strange story-Acc said.

 [TEXT NOT REPRODUCIBLE IN ASCII]
 I-nom Mary-to self-Nom

 [TEXT NOT REPRODUCIBLE IN ASCII].
 had won-Comp said-Rel thing-be
 lit. `[Bill.sub.i] told a strange story yesterday.
 (The thing is) that I told [Mary.sub.j] that [self.sub.i/*j]
 had won.'

The concept of "logophoricity" seems to play a role here because topic Bill is the person who said the sentence in (i) and the person who heard it in (27). Still, it is not clear how Bill, which is both a logophoric NP and a topic, blocks the coreferential relation between Mary and caki. This requires further research. I am indebted to S. Kuno and to an anonymous reviewer for Linguistics for addressing this question.

(11.) Definitions of local o-command and local o-binding (Pollard and Sag 1992: 287) are

(i) A locally o-commands B just in case the content of A is a referential parameter and there is a SUBCAT list on which A precedes (i.e. is less oblique than) B.

(ii) A locally o-binds B just in case A and B are coindexed and A locally o-commands B. If B is not locally o-bound, then it is said to be locally o-free.

That is, a phrase A o-commands everything that is a more oblique complement of the same head. They assume that verbs and other lexical items that head phrases bear a lexical specification for a feature SUBCAT, which takes as its value a list of specifications corresponding to the various complements that the word in question combines with in order to form a grammatically complete phrasal projection. The order of elements on the SUBCAT list does not necessarily correspond to surface order but instead corresponds to the order of relative obliqueness, with more oblique elements appearing later than (i.e. to the right of) less oblique elements (P&S 1992: 280). Each constituent has, in addition to its syntactic category, another component called CONTENT, which contains linguistic information that is relevant to the determination of the phrase's semantic interpretation. Because their account is based on head-driven phrase-structure grammar (HPSG), I do not discuss specifically how their definition works.

(12.) P&S followed a suggestion in Brame (1977) that reflexive anaphors in English have only accusative forms. Therefore they can occur neither as finite clause subjects nor as possessors:

(i) *Himself left.

(ii) *John likes himself's book.

(13.) For example, John and Bill are coarguments in (i) and (ii):

(i) I think John likes Bill.

(ii) I like John's picture of Bill.

The local o-binding domain of himself in (i) is the embedded clause and that in (ii) is the nominal phrase John's picture of himself.

(14.) I am indebted to an anonymous reviewer for the discussion of this small-clause construction.

(15.) In Korean, however, possessors that are realized as genitive NPs are less prominent than objects.

(16.) This means that genitives are ranked as high as subjects regardless of their semantic function. For instance, their in (i) is not more prominent than an objective possessor their friends since both of them are genitives that are ranked as high as subjects.

(i) [Their.sub.i] [friends.sub.j]' defeat angered each [other.sub.*i/j]'s mothers.

I am indebted to an anonymous reviewer of Linguistics for this example.

(17.) The parametric definition of the potential antecedent (PA) in (42) is heterogeneous: the PA of Korean involves agreement and (under-)specification in terms of [Phi]-features restricting its candidate to third person NPs. On the other hand, the PA of English involves domains, as shown in the term coargument.

(18.) A reviewer for Linguistics does not see the clear-cut difference in the preference regarding the interpretation of herself. It can be due to the fact that the gap value between the two interpretations in (44') is trivial (i.e. 1).

(19.) P&S suggested that the intervention constraint of Grinder (1970) might account for the overgeneration in (45). In the same vein, Kuno (1987: 96) modified Langacker's (1969) chain-of-command condition and argued that it applies with varying degree of strength, depending upon the relative strength as a controller for reflexives.

(20.) In this case, we are comparing the prominence of two PAs: Cheli is more prominent than Suni.

(21.) It is not an optimal interpretation, due to the interpretation rule in (17).

(22.) In the Korean case that we discussed, the prominence link condition applies to a sentence where there is more than one PA for the anaphor. In the English data in (48), the prominence link condition applies to sentences where a PA for the anaphor does not exist.

(23.) I am indebted to S. Kuno (personal communication) for these examples. If there is a speaker who judges (49a) as unacceptable as (49b) is, semantic transparency seems to play a role in intervention for the speaker.

(24.) Interestingly, when there is an inanimate anaphor, inanimate NPs can be an intervener:

(i) *[[The [book.sub.i]'s] reviews] provided a useful critique of [itself.sub.i].

Since itself is an inanimate anaphor, inanimate NP reviews intervenes in the coreferential relations between the book and itself. If it is the case, we can suggest [[Alpha] animate NP] is an intervener for [[Alpha] animate anaphor].

(25.) Tang (1989) observed a similar phenomenon in Chinese:

(i) [TEXT NOT REPRODUCIBLE IN ASCII].

Ziji refers only to Zhangsan's father. It blocks the coreferential relations between ziji and Zhangsan.

(26.) The minimal pair in (54) can imply that the agenthood be the relevant factor here. However, the following sentence favors the animacy as a relevant factor to agenthood.

(i) [John.sub.i]'s [father.sub.j] is guilty by the fact that pictures of [himself.sub.*i/j] were found on the victim.

I am indebted for this example to an anonymous reviewer for Linguistics.

(27.) On independent grounds, Huang and Tang (1991: 270) suggest the following generalization to capture the blocking effect in Chinese, which will be discussed in section 4.2:

(i) The set of potential blockers of long-distance ziji is exactly the set of its potential local, or less remote, binders.

(ii) [TEXT NOT REPRODUCIBLE IN ASCII]

By suggesting (i), they tried to solve problems of an LF-movement analysis. Of course, their motivation differs from ours; they needed (i) to account for the fact that an intervening person NP (ni in [ii]), which does not agree in its [Phi]-features with a more remote potential binder (Zhangsan in [ii]), blocks a long-distance binding relation. Even though we cannot explain at this point why there is no blocking effect when an intervening NP agrees in its [Phi]-features with a remote PA, it must be noted that the general idea in (i) is similar to ours.

(28.) There are various suggestions regarding this condition, one of which is that the mood of the embedded clause has to be subjunctive. I discuss this later in this section.

(29.) I am indebted to H. Thrainsson (personal communication) for these examples and for invaluable discussions about Icelandic anaphora.

(30.) Just as in English, a possessive NP in Icelandic is ranked as high as a subject in the prominence hierarchy. Note that in Korean, a possessive NP is treated simply as a genitive NP, which is less prominent than a subject.

(31.) R. Kayne (personal communication) suggested the possibility that (58a) is a case of extraposition and that, at some level of representation, sig is c-commanded by its antecedent.

(32.) For this reason, Thrainsson (1991) proposed that we have to distinguish a locally bound sig from a LDA sig.

(33.) H. Thrainsson (personal communication) judges (59b) with [sig.sub.j] as marginally acceptable.

(34.) The definition of Icelandic PA in (61) does not mean that a PA must be a subject and a coargument of a reflexive at the same time. It simply means that any NP that functions as subject and any coargument of a reflexive can be a PA in Icelandic.

(35.) The emphatic reflexive in Icelandic can be in the subject position:

(i) [TEXT NOT REPRODUCIBLE IN ASCII].

It is not clear to me whether we have to treat this emphatic reflexive differently from other reflexives in Icelandic.

(36.) As was mentioned in note 30, a possessive NP in Icelandic is regarded as a subject. Because Icelandic sig does not have any animacy restriction, as shown in (i), the reflexive in (62) should be allowed to refer to sko??un Jons:

(i) [TEXT NOT REPRODUCIBLE IN ASCII].

(37.) The c-command condition in (63) on a local antecedent for reflexive sig runs contrary to the goals of this paper: we have tried to show that a structural requirement such as c-command is not required to account for the coreferential possibilities of the range of anaphors. The analysis in this section shows that we cannot completely abandon the c-command requirement on anaphora. I leave this question open, focusing in this subsection on the manner in which the PP-A, which is motivated by Korean anaphora and supported by English anaphora, applies similarly to anaphora in various languages.

(38.) H. Thrainsson observes that when an antecedent of sig is embedded, as in (i), it become worse:

(i) [TEXT NOT REPRODUCIBLE IN ASCII].

Hence, we need other conditions to account for (i), because Icelandic sig is more restricted than we expect it to be. R. Kayne (personal communication) suggested that if (58a) is a case of extraposition, (i) would be marginal because the extra relative clause makes it marginal as extraposition.

(39.) Note that the interpretation of the Icelandic reflexive sig does not seem to be affected by the intervention condition. Consider the following:

(i) [TEXT NOT REPRODUCIBLE IN ASCII].

(ii) [TEXT NOT REPRODUCIBLE IN ASCII]

(40.) According to Cole and Sung (1994), the polymorphemic reflexive ([X.sup.max]) ta ziji in (i) can refer to Lisi:

(i) [TEXT NOT REPRODUCIBLE IN ASCII]

Any properties that differ between monomorphemic and polymorphemic reflexives will be a problem for the proposal in this paper.

(41.) In Chinese, the relative notions of more/less prominent do not seem to play a role. However, there is no undesirable result when we leave the notion more/less in the definition of PP-A.

(42.) The blocking effect is claimed to be accounted for in the following way (see Cole et al. 1990). If the [Phi]-features of [subject.sub.2] do not agree with those of [subject.sub.3], neither [subject.sub.2] nor [subject.sub.1] can be an antecedent of ziji, as represented in (i):

(i) [TEXT NOT REPRODUCIBLE IN ASCII]

(ii) [TEXT NOT REPRODUCIBLE IN ASCII]

(iii) [TEXT NOT REPRODUCIBLE IN ASCII]

Assuming that ziji is marked for [Phi]-features, Cole et al. (1990) claim that when ziji undergoes successive cyclic movement from Infl to Infl, each Infl, must agree with ziji in [Phi]-features. Infl, at the same time, agrees with the subject of its own clause in [Phi]-features. Ziji in (ii) cannot refer to Zhangsan because ziji agrees with Infl of Wangwu when it first moves to the t' position. Once ziji agrees with the third person singular [Phi]-feature in the Infl of the embedded clause, it cannot adjoin to another Infl whose feature is first person because there would be conflict in features.

(43.) If we regarded the possessive NP as a nonsubject that cannot be a PA in Chinese, we could not explain the coreferential relation between wo and ziji in (69b). The PP-A would predict that ziji has to refer to the animate subject Zhangsan in (69b), regardless of the existence of the nonsubject NP wo.

(44.) Like ziji in Chinese, morphologically zibun can refer to first person, second person, and third person NPs.

(45.) This claim is first made in Kuno (1972).

References

Brame, Michael (1977). Alternatives to the tensed-S and specified subject conditions. Linguistics and Philosophy 1, 381-411.

Chomsky, Noam (1981). Lectures on Government and Binding. Dordrecht: Foris.

--(1986). Barriers. Cambridge, MA: MIT Press.

Clements, N. (1975). The logophoric pronoun in Ewe: its role in discourse. Journal of West African Languages 10, 141-177.

Cole, Peter; Hermon, Gabriella; and Sung, Li-May (1990). Principles and parameters of long distance reflexives. Linguistic Inquiry 21, 1-22.

--; and Sung, Li-May (1994). Head movement and long distance reflexives. Linguistic inquiry 25, 355-406.

--; and Wang, Chengchi (1996). Antecedents and blockers of long-distance reflexives: the case of Ziji. Linguistic Inquiry 27, 357-390.

Giorgi, A. (1984). Towards a theory of long distance anaphors: a GB approach. Linguistic Review 3, 307-359.

Grinder, John (1970). Super-Equi-NP-deletion. In Papers from the 6th Regional Meeting of the Chicago Linguistic Society, M. Campbell et al. (eds.), 297-317. Chicago: CLS.

Hellan, Lars (1991). Containment and connectedness anaphors. In Long-Distance Anaphora, Jan Koster and Eric Reuland (eds.), 27-49. Cambridge: Cambridge University Press.

Hestvik, Arild (1992). LF movement of pronouns and antisubject orientation. Linguistic Inquiry 23, 557-594.

--(1995). Reflexives and ellipsis. Natural Language Semantics 3, 211-237.

Huang, C.-T. James; and Tang, C.-C. Jane (1991). The Local Nature of the Long-Distance Reflexive in Chinese. In Koster and Reuland 1991, 263-282.

Iida, Masayo (1992). A unified theory of zibun-binding. Unpublished manuscript, Hewlett-Packard Laboratories.

Jackendoff, Ray (1972). Semantic Interpretation in Generative Grammar. Cambridge, MA: MIT Press.

Kameyama, M. (1984). Subjective/logophoric bound anaphor zibun. In Papers from the 20th Regional Meeting, J. Drogo, V. Mishra, and D. Testen (eds.), 228-238. Chicago: Chicago Linguistic Society.

--(1985). Zero anaphora: the case of Japanese. Unpublished doctoral dissertation, Stanford University.

Keenan, Edward (1987). On semantics and the binding theory. Unpublished manuscript.

--; and Comrie, Bernard (1977). Noun phrase accessibility and universal grammar. Linguistic Inquiry 8, 63-99.

Kim, Soo-Yeon (1991). Two reflexives in Korean: caki and casin. Unpublished manuscript, Harvard University.

--(1996). Dependencies: A Study of Anaphoricity and Scrambling. Seoul: Hanshin.

--(i.p.). What should we do with condition A? Korean Journal of Linguistics 25 (1).

Kiss, Katalin E. (1991). The primacy condition of anaphor and pronominal variable binding. In Long-Distance Anaphora, Jan Koster and Eric Reuland (eds.), 245-262. Cambridge: Cambridge University Press.

Koster, Jan; and Reuland, Eric (1991). Long-Distance Anaphora. Cambridge: Cambridge University Press.

Kuno, Susumu (1972). Pronominalization, reflexivization, and direct discourse. Linguistic Inquiry 3, 161-195.

--(1976). Subject, theme and the speaker's empathy: a reexamination of relativization phenomena. In Subject and Topic, C. Li (ed.). New York: Academic Press.

--(1987). Functional Syntax: Anaphora, Discourse and Empathy. Chicago: University of Chicago Press.

--; and Kaburaki, Etsuko (1977). Empathy and syntax. Linguistic Inquiry 8, 627-672.

--; and Takami, Ken-Ichi (1993) Grammar and Discourse Principles: Functional Syntax and GB Theory. Chicago: University of Chicago Press.

Langacker, R. (1969). On pronominalization and the chain of command. In Modern Studies in English. D. A. Reidel and S. A. Schane (eds.). Englewood Cliffs, NJ: Prentice Hall.

Lebeaux, David (1983). A distributional difference between reciprocals and reflexives. Linguistic Inquiry 14, 723-730.

Maling, Joan (1984). Non-clause-bonded reflexives in Icelandic. Linguistics and Philosophy 7, 211-241.

McCawley, Noriko Akatsuka (1976). Reflexivization: a transformational approach. In Syntax and Semantics 5, Japanese Generative Grammar, Masayoshi Shibatani (ed.), 51-115. New York: Academic Press.

Pica, Pierre (1985). Subject, tense, and truth: towards an approach to binding. In Grammatical Representation, Jacqueline Gueron, Hans-Georg Obenauer, and Jean-Yves Pollock (eds.), 259-291. Dordrecht: Foris.

--(1987). On the nature of reflexivization cycle. In NELS 17, J. McDonough and B. Plunkett (eds.), 483-499. Amherst, MA: GLSA.

--(1991). On the interaction between antecedent-government and binding: the case of long-distance reflexivization. In Long-Distance Anaphora, Jan Koster and Eric Reuland (eds.), 119-136. Cambridge: Cambridge University Press.

Pollard Carl; and Sag, Ivan A. (1992). Anaphors in English and the scope of binding theory. Linguistic Inquiry 23, 261-303.

Reinhart, Tanya; and Reuland, Eric (1993). Reflexivity. Linguistic Inquiry 24, 657-720.

Sells, Peter (1987). Aspects of logophoricity. Linguistic Inquiry 18, 445-481.

Sigur??sson, Halldor, Armann (1990). Long-distance reflexives and moods in Icelandic. In Modern Icelandic Syntax, Joan Maling and Annie Zeanen (eds.), 309-346. San Diego: Academic Press.

Tang, C.-C. Jane (1989). Chinese reflexives. Unpublished doctoral dissertation, University of Leiden.

Thrainsson, Hoskuldur (1976). Reflexives and subjunctives in Icelandic. In NELS 6, 225-239. Amherst, MA: GLSA.

--(1991). Long-distance reflexives and the typology of NPs. In Long-Distance Anaphora, Jan Koster and Eric Reuland (eds.), 49-75. Cambridge: Cambridge University Press.

Wexler, Kenneth; and Manzini, Maria Rita (1987). Parameters and learnability in binding theory. In Parameter Setting, Thomas Roeper and Edwin Williams (eds.), 41-76. Dordrecht: Reidel.

Wu, Yuru (1992). Anaphora and quantifier scope in Mandarin Chinese. Unpublished doctoral dissertation, Harvard University.

Zribi-Herz, Anne (1989). Anaphor binding and narrative point of view: English reflexive pronouns in sentence and discourse. Language 65(4), 695-727.