2.1 The difference in the distribution of AE between Japanese and Chinese

As stated in the introduction, Oku (Reference Oku1998) observes that Japanese permits AE in both null subject and null object positions. This symmetrical AE pattern is illustrated by the availability of sloppy interpretations in these positions, as shown in (1) and (2), respectively.

In contrast to Japanese, Chinese exhibits a subject–object asymmetry with respect to AE. Example (3) illustrates that the null object allows AE.

As noted by Huang (Reference Huang and Freidin1991) and Takahashi (Reference Takahashi2007), the null subject does not allow AE, as evidenced by example (4). The null subject in (4) only accepts the strict interpretation that Zhangsan’s child liked Xiaoli.

The contrasting distribution between Japanese and Chinese with respect to subject AE is most convincingly demonstrated by (5)–(6). Here, the Japanese example in (5) is completely parallel to the Chinese example in (6), thereby excluding the possibility that lexical choices, tenses, or other possible non-syntactic factors facilitate or hinder sloppy interpretations and hence AE.

Tomioka (Reference Tomioka2014) observes that the impossibility of subject AE in Chinese cannot be attributed to pragmatic considerations, an observation which further supports the existence of some structural restriction imposed on this position. In English, singular personal pronouns typically do not permit sloppy interpretations, but it is well known that such an interpretation is rendered somewhat easier to obtain when it is preceded by an expression such as the everyone but X construction. This point is illustrated in (7).

Tomioka then shows that the empty subject position in Chinese resists a sloppy interpretation even in this context. In (8), the sloppy interpretation is impossible even though the everyone but X construction used there facilitates it. The sloppy interpretation is achieved only when the relevant position is overtly filled by a lexical subject, ziji-de laoshi ‘self’s teacher’, as is shown in (8).

2.2 A brief excursus on the absence of subject AE in Chinese: Simpson et al. (Reference Simpson, Choudhury and Menon2013)

The present paper attempts to build a new analysis of AE for Chinese as an alternative to the anti-agreement theory, crucially basing it on the core observation that AE is possible in object positions but not in subject positions. It is worthwhile to take some time here to confirm the robustness of this observation in light of Simpson et al.’s (Reference Simpson, Choudhury and Menon2013) report that ‘in Mandarin Chinese, the use of the anaphoric possessor ziji-de results in the availability of sloppy readings in embedded subject positions’. The examples which Simpson et al. mentioned in this connection are shown in (9)–(12), together with the authors’ reported acceptability judgements.

I consulted 17 native speakers of Mandarin Chinese about their judgements of the four examples above, and the result of my consultation is as follows (see also footnote 7 below): First, thirteen of them reported that they only can accept the strict interpretation for the null subject in all these examples. Second, one of the remaining four speakers told me that he could ‘feel’ the sloppy reading only in (10) and (12), but even so, the reading was still very marginal whereas another reported that she could get the sloppy reading in (9) and (11), but not in (10) and (12). Finally, the two remaining speakers were quite unsure what interpretation(s) the null-subject sentences permitted in those four examples. Despite the rather minor individual variation, my survey above makes it clear that the impossibility of subject AE seems to be indeed the dominant judgement pattern shared by Mandarin Chinese speakers. I will therefore continue to assume that the subject–object asymmetry in Chinese is a core empirical generalization for the majority of its speakers, which thereby calls for a principled explanation, leaving a large-scale investigation of the important question, why some idiolectal/dialectal variation occurs, for another occasion.

2.3 Issues with the agreement analysis of the subject–object asymmetry in Chinese AE

Miyagawa (Reference Miyagawa2013) and Takahashi (Reference Takahashi and Saito2014) argue that the impossibility of subject AE in Chinese follows from Saito’s (Reference Saito2007) anti-agreement theory. Saito’s theory is framed within Chomsky’s (Reference Chomsky, Martin, Michaels and Uriagereka2000) Probe–Goal framework, according to which a probe with uninterpretable $\unicode[STIX]{x03C6}$ -features, either T or v, searches for a goal with interpretable $\unicode[STIX]{x03C6}$ -features. The matching of the $\unicode[STIX]{x03C6}$ -feature sets induces the deletion of the uninterpretable $\unicode[STIX]{x03C6}$ -features of the probe by Agree, which must, in turn, be activated by the presence of an uninterpretable Case feature of the goal. This condition is known as the Activation Condition. The Case feature of the goal is deleted as a reflex of the $\unicode[STIX]{x03C6}$ -feature agreement established between the probe and the goal. Saito proposes that this Probe–Goal system blocks AE, which he implements as an LF Copy process (see Oku Reference Oku1998), from targeting any syntactic position associated with $\unicode[STIX]{x03C6}$ -feature agreement. LF Copy is an operation that copies an argument from a full-fledged linguistic antecedent onto an empty argument spot in an elliptical clause to save an otherwise uninterpretable structure at LF. To understand how Saito’s theory works, consider three steps of the syntactic derivation required for LF Copy, shown in (13a–c); $e$ stands for the position of an elided argument.

In (13a), the probe F $_{1}$ , endowed with uninterpretable $\unicode[STIX]{x03C6}$ -features, searches for the goal DP $_{1}$ bearing the matching interpretable $\unicode[STIX]{x03C6}$ -features. Triggered by the presence of the uninterpretable Case feature on the goal, Agree takes place to delete the uninterpretable $\unicode[STIX]{x03C6}$ -features of the probe and of the uninterpretable Case feature of the goal. Suppose now that the DP in (13a) is copied at LF onto the null argument position designated by $e$ in (13b). Note that the Activation Condition prevents this DP $_{1}$ from participating in another agreement relationship with the new probe F $_{2}$ since the uninterpretable Case feature of the DP $_{1}$ has already been deleted before LF Copy takes place. As a result, the uninterpretable $\unicode[STIX]{x03C6}$ -features of the F $_{2}$ remain unchecked, causing the derivation to crash, as shown in (13c).

This anti-agreement theory of Saito (Reference Saito2007) correctly predicts the difference between English and Japanese with respect to the availability of AE. English blocks AE in both object and subject positions, as shown in (14a, b), respectively, whereas Japanese allows this process in both grammatical positions, as shown above, in (1a, b) and (2a, b).

Recall Chomsky’s (Reference Chomsky, Martin, Michaels and Uriagereka2000) central assumption that (structural) Case is inexorably tied to $\unicode[STIX]{x03C6}$ -feature agreement; Case checking is simply a side effect of a goal DP entering into Agree with an appropriate functional head such as T or v. The presence of (structural) nominative and accusative Cases in English indicates, then, that both T and v are associated with abstract uninterpretable $\unicode[STIX]{x03C6}$ -features. The presence of these features thus causes the derivation involving AE/LF Copy to crash, as schematically illustrated in (13c) above. In Japanese, on the other hand, AE/LF Copy is allowed to apply to both null subject and null object positions because T and v in this language need not have $\unicode[STIX]{x03C6}$ -feature agreement. Indeed, this is more or less a standard assumption since Kuroda (Reference Kuroda1988), who first proposed that agreement is not forced in Japanese.

Returning now to Chinese, Miyagawa (Reference Miyagawa2013) and Takahashi (Reference Takahashi and Saito2014) argue that the impossibility of subject AE in the language falls into place if it possesses abstract $\unicode[STIX]{x03C6}$ -feature agreement associated with the subject position. The hypothesis that Chinese possesses subject agreement was first argued for by Miyagawa (Reference Miyagawa2010) based on the so-called blocking effect on ziji ‘self’ (Tang Reference Tang1985, Reference Tang1989). Consider examples (15)–(16).

In (15), ziji can be bound to the matrix subject across the embedded third-person subject. This long-distance binding becomes impossible, however, when the local potential antecedent is replaced by a first- or second-person subject, as in (16). Adopting the LF head-movement analysis of subject-oriented anaphors in Chinese (Battistella Reference Battistella1989, Cole, Hermon & Sung Reference Cole, Hermon and Sung1990), Miyagawa (Reference Miyagawa2010) assumes that ziji raises to its local T to be assigned a person-feature value from the T, which, in turn, receives this value from its specifier via Spec–Head Agreement; the anaphor subsequently moves to the higher T for the long-distance construal. This latter movement converges when the person-feature value of the higher T matches that of the lower T, as shown in (15). The same movement crashes in (16), on the other hand, because the two person-feature values do not match. Note that this analysis of the blocking effect crucially hinges on the assumption that Chinese possesses abstract person feature agreement associated with the subject position. Thus, to the extent that the subject agreement hypothesis holds true, the anti-agreement theory is sufficient to account for the subject–object asymmetry in Chinese.

However, there are three issues with the agreement analysis. The first two issues arise when the LF head-movement analysis of anaphors is adopted as an analytical premise for the blocking effect; the third issue concerns a wrong prediction made by the agreement hypothesis with respect to the distinction between syntactic and logophoric uses of ziji. First, the analysis in question leads us to predict that the availability of the long-distance reading should be governed by constraints, such as island effects, which are independently known to block other instances of LF head-movement in Chinese. However, Huang & Tang (Reference Huang, Tang, Koster and Reuland1991) show that this prediction is not borne out. In (17), ziji may take the matrix subject as its antecedent, in theory, by moving to the matrix T from within the Adjunct Island.

However, the relevant movement should be impossible because such a movement obeys this constraint elsewhere, as shown in (18). The sentence in (18) is out because the interrogative operator within the A-not-A question, which is under T (Huang Reference Huang1982), undergoes LF movement from within the Adjunct Island.Footnote ^[2]

The second issue with the LF movement analysis of the blocking effect, and with the T-agreement analysis, by extension, concerns the observation made by Xue, Pollard & Sag (Reference Xue, Pollard, Sag, Aranovich, Byrne, Preuss and Senturia1994) that the effect is created by non-subject arguments as well. Recall that the movement analysis capitalizes on the Spec–Head Agreement between T and its local subject to allow/block the long-distance construal of the subject-oriented anaphor ziji. Consequently, the analysis predicts that non-subject arguments should not play a role in blocking the long-distance reading. Xue et al. show that this prediction is false. Examples (19)–(20) show that the relevant reading is blocked by the intervening first- and second-person direct objects selected by the intervening verbs.

The third and final issue is concerned with Huang & Liu’s (Reference Huang, Liu, Cole, Hermon and Huang2001) observation, which was recently discussed by Li (Reference Li2014) as problematic for the subject agreement theory of Chinese. Huang & Liu’s observation is two-fold. First, ziji has two distinct dferent uses, one as a syntactic anaphor and the other as a logophoric anaphor. Second, the blocking effect obtains only when ziji is used as a logophoric anaphor due to the conflict of perspective-taking. The distinction between the two uses of ziji lies in the standardly postulated locality condition imposed on anaphor binding: a syntactic anaphor is bound within its governing category whereas a logophoric anaphor is not. The governing category is defined as the minimal TP or NP that contains an anaphor, its governor and a SUBJECT (the subject of TP/NP or Agr). With this definition in place, consider now example (21).

Li (Reference Li2014) shows that the contrast in (21a, b) is problematic for the view that Chinese has abstract subject agreement. However, this contrast is correctly accounted for if there is no subject agreement in Chinese, whereby the governing category for the anaphor is the entire sentence in (21a), but the embedded clause in (21b). This means that ziji in (21a) is a syntactic anaphor whereas ziji in (21b) is a logophoric anaphor. It follows correctly then, that only ziji in (21b) triggers the blocking effect. If there were subject agreement in Chinese, however, the relevant contrast would become mysterious, for the governing category for the anaphor in (21a) would now become the embedded TP, with the wrong result that ziji there would trigger the blocking effect on a par with ziji in (21b).

3.1 The definite subject restriction and operator–variable topic chains in Chinese

It is widely acknowledged in the literature on Chinese grammar (Chao Reference Chao1968; Li & Thompson Reference Li, Thompson and Li1976, Reference Li and Thompson1981; Tsao Reference Tsao1977; Huang Reference Huang, Reuland and ter Meulen1987; Shi Reference Shi2000; Chou Reference Chou2004; Cheng Reference Cheng2013) that Chinese manifests the definite subject restriction. This restriction is illustrated in (22a) below. To express the indefinite reading, the existential marker you ‘to exist’ has to be used instead to introduce the subject in a post-verbal position, as shown in (22b). Examples (23) show that the same semantic restriction is imposed on the embedded subject position.

The restriction in question follows if, in the absence of any other topic within the sentence, the subject, by default, undergoes string-vacuous topicalization to form an operator–variable relationship between [Spec, TP] and [Spec, Top], as schematically indicated in (24).

The restriction is now derived as a natural consequence of the discourse-oriented requirement that topics, by definition, are definite, a requirement which is independently illustrated in (25) by the inability of indefinite non-subject arguments to undergo topicalization (see footnote 3 below).

3.2 Operator–variable chains, LF Copy, and the subject–object asymmetry in Chinese AE

I will now show how the establishment of an operator–variable topic chain blocks subject AE in Chinese. My analysis builds on Saito’s (Reference Saito, Shibatani, Miyagawa and Noda2017) analysis whereby AE, technically analyzed as LF Copy, cannot target an operator–variable structure as its input, to which I now turn.

Saito (Reference Saito, Shibatani, Miyagawa and Noda2017) examines a number of cases reported by Hoji (Reference Hoji1998) and Funakoshi (Reference Funakoshi2012, Reference Funakoshi and Huber2013) as potential problems for the AE analysis of null objects in Japanese. Saito notes that AE cannot apply to a phrase which requires an operator–variable configuration, showing how this generalization follows from the LF Copy theory of AE (Oku Reference Oku1998). For reasons of space, I will only illustrate this generalization and Saito’s analysis thereof using wh-questions. Consider (26). Example (26b) shows that an interrogative wh-phrase cannot undergo AE.

Saito assumes that a wh-phrase contains an interrogative–operator pair, as in (27a) below, and that the operator and variable parts are interpreted in two positions linked by wh-movement, as in (27b). More specifically, the operator is interpreted in [Spec, CP] whereas the variable is interpreted in the base position. This point is illustrated in (27c), where the words in bold indicate the position at which the two materials are interpreted at LF.

Now, if we apply this chain-based interpretive mechanism to the embedded question in (26a), the representation in (28a) below results. Given this, if the operator in (28a) is copied to the ellipsis site in (26b), (28b) will be obtained. If the variable in (28a) is copied there instead, (28c) will be obtained. Crucially, neither representation is interpretable: the former involves the operator in an argument position whereas the latter involves a free variable not bound to any operator. It follows that AE, analyzed as LF Copy, cannot target an operator–variable structure.

Now it should be clear why subject AE is blocked in Chinese: the subject position involves an operator–variable configuration, which in turn blocks AE as an LF Copy process. Consider how this analysis works, using (6) above as an example. Given my analysis, the antecedent clause in (6) has the LF representation in (29a) below. The embedded clause in (6), then, involves the schematic operator–variable chain in (29b). Now, depending on whether we copy the operator or the variable portions of the object in (29b) to the null argument position in (6), we would get the two representations in (29c, d). Neither representation is interpretable at LF.

My analysis maintains that AE/LF Copy is blocked from applying to an argument position when its input forms an operator–variable structure. We diagnosed the presence of such a chain involving the null subject position through the definite subject restriction. With this in mind, the example in (30) shows that the object position is free from this restriction.

It follows that the object position does not incur an operator–variable chain and hence provides a legitimate object for LF Copy to merge onto the null object position.

One might ask in this connection why a certain hypothetical derivation is blocked in my analysis. In this derivation, the operator and variable portions in (29b) above are copied at LF onto the topic and null subject positions of the elliptical clause in (6), respectively. The resulting representation, shown in (31), looks like a legitimate operator–variable configuration.

This derivation, I maintain, is blocked by the independently motivated assumption, dating back to Oku (Reference Oku1998), that AE can apply only to argument positions, which is supported by the well-known observation that, across languages, adjuncts themselves cannot participate in this phenomenon. Consequently, copying any (part of) LF-object onto a non-argument position such as the topic position is an impossible derivational choice, to begin with.

3.3 Cross-linguistic implications: Subject ellipsis in Japanese, Korean and Mongolian

Recall that my analysis maintains that the LF Copy underlying AE is blocked from applying to the null subject position in Chinese by an operator–variable chain, the presence of which is diagnosed by the definite subject restriction. This analysis has an important implication for the cross-linguistic distribution of subject AE in other languages, beyond Chinese. Specifically, if an AE language is not subject to the definite subject restriction, the language should not require the topic chain, and hence should permit subject AE. I suggest that subject ellipsis is allowed in Japanese precisely for this reason. Example (32) shows that the nominative subject does not have to be interpreted as definite, unlike in Chinese.

The same analysis can be extended to subject ellipsis in Korean and Mongolian, two other languages which have been shown by Takahashi (Reference Takahashi2007, Reference Takahashi2013a) to permit AE in both null subject and null object positions, as shown in (33)–(34) and (35)–(36), respectively.

It is significant that neither Korean nor Mongolian requires the subject to be definite, as the next two examples show. (The example in (38) was provided by Lina Bao through Yuta Sakamoto (personal communication, May 2015).)

3.4 Subject ellipsis in Chinese: Hanging topics, relative clauses and conditional clauses

Returning now to Chinese, my analysis makes an important prediction that, contrary to conventional wisdom in the literature, the null subject position in this language should, in principle, permit AE, as in Japanese, Korean and Mongolian, as long as there is no operator–variable relationship between the topic and the null subject. I will show that this prediction is indeed borne out in hanging topic constructions, relative clauses and conditional clauses.

In the hanging topic construction, the sentence-initial topic is followed by the comment clause consisting of the logical subject and the predicate. Example (39) is a case in point:

It is hard to imagine the derivation for this construction where the topic moves from within the comment clause to the sentence-initial position because there is no suitable argument position from which the topic could have moved, to begin with. Xu & Langendoen (Reference Xu and Langendoen1985) present other empirical arguments that this construction violates well-known syntactic constraints on movement such as subjacency and the bijection principle, which suggest that the topic is base-generated in the topic position without any movement transformation.

The most significant property of this construction for our present purposes is that the subject does not have to be definite, as illustrated in (40)–(41).

In (40), for instance, the DP na-zhong douzi ‘that kind of beans’ is presented as the topic of the sentence, and the rest of the sentence is predicated of this DP as the comment clause. The topichood of the sentence-initial DP is confirmed by the fact that its indefinite counterpart in the same position leads to ungrammaticality, as shown in (41). The logical subject in (40), on the other hand, is clearly not definite. This means that in my analysis, it does not participate in an operator–variable relationship with the topic.

Given this property, then, my analysis predicts that the null subject position should allow AE in this context. Examples (42)–(43) show that this prediction is indeed borne out.

It is perhaps worthwhile to reiterate here that the availability of AE in the null subject position in Chinese is mysterious under the agreement theory of AE reviewed in Section 2.3 above. Since the analysis assumes that T in this language is invariably endowed with agreement, it predicts that the logical subject in the hanging topic construction should block AE on a par with the logical subject in archetypical subject ellipsis cases (examples (4), (6) and (8)–(12)).Footnote ^[3]

The prediction stated above is further supported by the recent finding reported by Abe & Park (Reference Abe and Park2017). Given Kuno’s (Reference Kuno1973) observation that subjects in relative and conditional clauses do not render topic readings in Japanese, Abe & Park assume that the subject is not linked to the topic position through movement in such cases in Chinese as well.Footnote ^[4] I take this to mean that the subject position does not enter into an operator–variable chain with the topic. It is significant then that the null subject allows sloppy interpretations precisely in these two contexts, as shown in (44) and (45), just as predicted by my analysis.

One may counter that the sloppy interpretation in (42)–(43) can be accounted for without invoking AE if the null subject there is not elliptic but represents an indefinite pronoun in the sense of Hoji (Reference Hoji1998). Hoji (Reference Hoji1998) argues that the sloppy interpretation of the null object in Japanese examples such as (1b) above is merely a ‘sloppy-like’ reading, which is derived though the indefinite use of pro ( $pro_{\text{NP}}$ ) on a par with indefinite bare nominals. At first brush, this analysis appears to be well-suited for the null subject in (42b) and (43b), given my observation that the subject position of the hanging topic construction may host an indefinite expression.Footnote ^[5] However, this analysis makes wrong predictions when tested against sentences with quantificational null arguments, another well-known diagnostic for AE together with sloppy interpretations (Takahashi Reference Takahashi, Miyagawa and Saito2008a, Reference Takahashib). Examples (46) are a case in point:

The sentence in (46b) permits the quantificational interpretation, where Rizzi published more than three of papers in NLLT. The indefinite pro analysis incorrectly predicts that this quantificational interpretation should not be available in (46b) because the overt, bare indefinite nominal – wenzhang ‘paper’ – does not allow this interpretation, as witnessed in (46c). Furthermore, (46b) does not allow the interpretation that Rizzi published (any number of) his papers in NLLT, an interpretation which should be fine if the empty argument there were filled by the indefinite nominal wenzhang ‘paper’, given the independently attested possibility of this interpretation in (46c). The impossibility of this indefinite interpretation in (46b) falls neatly into place under my analysis, however, since AE/LF Copy takes the overt quantificational expression in the antecedent clause in (46a) as its input argument.

It is important to stress in this connection that my native Chinese consultants all agreed that sloppy readings in the null subject positions in (42)–(43) are readily available without much contextual priming. This point calls for emphasis. Simpson et al. (Reference Simpson, Choudhury and Menon2013) report that sloppy readings in the null subject position can be rendered available in Hindi and Bangla – two languages which otherwise prohibit subject AE – with the help of heavily enriched contexts which facilitate such readings. Example (47) from Hindi illustrates their observation.Footnote ^[6]

Based on their observation that sloppy interpretations require heavily enriched contexts in null subject positions, as opposed to non-subject positions, Simpson et al. (Reference Simpson, Choudhury and Menon2013) discount the null subject case in Hindi and Bangla as genuine AE and suggest that they involve pros. The sloppy reading in such cases is merely a ‘pseudo-sloppy’ reading made available to hearers simply by contextual priming, which allows a pro-subject to refer back to one of the referents made prominent in the discourse. In (47), both Ram’s daughter and Ram’s brother’s daughter are explicitly introduced into the background context so that the null subject may simply pick up the latter discourse reference as its antecedent, yielding what we call a ‘sloppy’ reading.

However, subject ellipsis cases in (42)–(43) clearly involve genuine AE instead of enriched pros because they do not require the similar amount of detailed background contexts that make sloppy interpretations available as do the Hindi subject ellipsis case. The 13 Mandarin native speakers I corresponded with (see Section 2.2 above) reported that they could easily obtain the sloppy interpretation in (42)–(43) where the null subject is anchored to the subject in the matrix clause without the help of such contextual cues. Recall also that the very availability of quantificational interpretations as in (46b) already independently mitigates against the (uniform) pro-analysis of the null subject in Chinese.Footnote ^[7]