Licensing Case-mismatches and dependent plural markers in Korean left-node-raising

2.1 Locus

Nakao (2009, 2010 and Chung (2010) argue that the pivot in LNR syntactically belongs to both conjuncts because the pivot is a complement of each predicate simultaneously. More precisely, Nakao obtains this by assuming that the pivot originates as two separate XPs but has undergone ATB movement into one XP position as in (7), while Chung claims that the pivot is multidominated by each predicate and has moved out of the shared position as in (8).

(8)

multidominance analysis (Chung 2010)

In the symmetric analyses of Japanese/Korean LNR, any Case-mismatch is predicted to be ill-formed. Since the pivot is Case-assigned by each predicate independently, its Case-marker is supposed to be licensed by each predicate. If the predicates assign different Case – dative Case and accusative Case as in (7) and (8) – there would be a derivational crash due to the Case-mismatch, regardless of where the Case-mismatch occurs.

In contrast, Kim et al. (2020) argue that the pivot in Korean LNR belongs to the first conjunct only. As shown in (9), the pivot is the complement of the first conjunct predicate, and the missing complement in the second conjunct is a null pronominal pro:

Under the asymmetric analysis depicted in (9), the pivot is Case-assigned by the first conjunct predicate.

2.2 Case types

Chomsky (1986) proposes that unlike structural Case, inherent Case is related to θ-role assignment. More precisely, Chomsky (1995) argues that structural Case is assigned solely in terms of S-structure configuration whereas inherent Case is associated with θ-assignment. This argument leads us to an interesting hypothesis regarding the Case type of a pivot in Case-mismatched LNR. A dative predicate can assign a goal θ-role along with dative Case (i.e., inherent Case); therefore, the Case-mismatched accusative pivot would pay a penalty if it should function as an argument of a dative predicate because of Case-mismatches. In contrast, an accusative predicate would not give a penalty to the Case-mismatched dative pivot since structural Case is assigned irrespective of θ-assignment. The derivations in (10) schematize this hypothesis:

According to this hypothesis, when the pivot is marked with dative Case as in (10a), the accusative predicate macihayssta ‘welcomed_Acc’ would interpret the pivot as its (theme) argument. However, when the pivot carries an accusative Case-marker as in (10b), the dative predicate cwuessta ‘gave_Dat’ would pay a penalty for the interpretation of the pivot as its (goal) argument. This penalty would attenuate acceptability. In this sense, we hypothesize that Case-mismatches with a dative pivot would be preferred than those with an accusative pivot in terms of acceptability.

2.3 Dependent plural markers

As mentioned in Section 1, the acceptability (or grammaticality) of (6) in which the DPM is attached to the pivot may be an interesting test ground for each of the previous analyses. If Chung’s (2010) observation of (6) being grammatical is attested, then (6) might be nicely accounted for by his multidominance proposal, as schematized in (11).

(11)

Since the multidominated pivot was c-commanded by the split plural subject John + Mary prior to fronting, the dependent-plural-marked (DPMed) pivot was successfully licensed under the multidominance proposal.

On the other hand, it is not clear how the DPMed pivot might be licensed under either the ATB movement or the scrambling-plus-pro proposal. That is, it is hard to imagine a derivational step in which the DPMed pivot meets a c-commanding (split) plural subject. Notice, however, that Chung’s (2010) multidominance argument only holds as far as the DPMed pivot in LNR such as (7) is acceptable.

Based on these theoretical predictions of LNR, we conduct an experimental and theoretical investigation of Korean LNR.

3.1 Experiment 1

In Experiment 1, we replicated Kim et al.’s (2020) experiment.

3.1.4 Results and discussion

Figure 1 presents the mean z-scores for the acceptability judgments for the four experimental conditions of Experiment 1. The zero represents the overall mean acceptability rating; positive z-scores indicate that conditions are rated towards being acceptable, while negative z-scores indicate that conditions are rated towards being unacceptable.

Figure 1:

Mean acceptability of experimental conditions (error bars indicate SE).

The results from the mixed-effects models with LOCUS and CASE as fixed effects, and random intercepts and slopes for participants and items in Table 1 confirmed that the main effect of LOCUS was extremely significant, indicating that the [1st] condition is significantly more acceptable than the [2nd] condition (mean: 0.197 vs. −0.522). The main effect of CASE also proved significant, confirming that the dative condition is significantly more acceptable than the accusative condition (mean: −0.072 vs. −0.252). However, the interaction between the two was not significant (χ²(1) = 0.071, p = 0.789).

Table 1:

Fixed effects summary for Experiment 1.

	Estimate	SE	t	p
(Intercept)	0.107	0.065	1.640	0.106
LOCUS	−0.719	0.082	−8.782	***
CASE	0.180	0.071	2.549	*

1. *p < 0.05, **p < 0.01, ***p < 0.001.

The results of Experiment 1 provided two findings. First, Case-mismatches in Korean LNR were conjunct-sensitive. Specifically, the Case-licensing in the first conjunct was more acceptable than that in the second conjunct. Second, the preference for the first conjunct as the Case-licensing locus of pivots was maintained regardless of Case types although Case-mismatched dative pivots were more acceptable than Case-mismatched accusative pivots.

Next, we investigate the effect of the DPM tul on the acceptability ratings of LNR, compared with those of RNR.

3.2 Experiment 2

3.2.2 Results and discussion

Figure 2 presents the mean z-scores for the acceptability judgments for the four experimental conditions of Experiment 2.

Figure 2:

Mean acceptability of experimental conditions (error bars indicate SE).

Table 2 presents the results of the mixed-effects model with SYNTAX and PLURALITY as fixed effects, and random intercepts and slopes for participants and items.

Table 2:

Fixed effects summary for Experiment 2.

	Estimate	SE	t	p
(Intercept)	0.636	0.079	8.079	***
SYNTAX	−0.655	0.095	−6.903	***
PLURALITY	−0.252	0.080	−3.161	**

The main effect of SYNTAX proved significant, confirming that RNR conditions were significantly more acceptable than LNR conditions (mean: 0.510 vs. −0.145). The main effect of PLURALITY was also significant, which confirms that the presence of DPM lowers the rating significantly (mean: 0.308 vs. 0.056). No interaction between the two factors was found (χ²(1) = 1.663, p = 0.197).

However, the results of pairwise comparisons reveal that there was no statistic difference between the [RNR | −DPM] condition and the [RNR | +DPM] condition (β = 0.152, SE = 0.112, t = 1.354, p = 1.000), but the [LNR | −DPM] condition was significantly more acceptable than the [LNR | +DPM] condition (β = 0.352, SE = 0.112, t = 3.140, p < 0.05).^[6] More precisely, as seen in Figure 2, the main effect of PLURALITY was produced by a very large difference between the two LNR conditions.

There were two main findings. First, the syntactic difference between RNR and LNR conditions modulated acceptability irrespective of the DPM. That is, RNR conditions were more acceptable than LNR conditions: (13a) was more acceptable than (13c), and (13b) was more acceptable than (13d). Second, the DPM did not modulate the acceptability of RNR conditions but it modulated the acceptability of LNR conditions: the acceptability of (13a) and (13b) was similar, but (13c) was more acceptable than (13d).

4.1 Case-licensing locus of a pivot

In Korean Case-mismatched LNR, Case-licensing in the first conjunct was more acceptable than Case-licensing in the second conjunct. This conjunct-sensitivity of Case-licensing presents a challenge to the symmetric account of Japanese/Korean LNR (Chung 2010; Nakao 2009, 2010), while supporting the asymmetric account (Kim et al. 2020).

The symmetric account argues that a pivot syntactically belongs to both conjuncts in Japanese/Korean LNR, being simultaneously subcategorized by each conjunct predicate. There are two popular analyses in this approach: the ATB scrambling analysis (Nakao 2009, 2010) and the multidominance analysis (Chung 2010; Nakao 2010).

The ATB scrambling analysis argues that Japanese/Korean LNR is derived via ATB scrambling as in (14).

In this analysis, as shown in (14), the pivot was base-generated in both conjuncts and has moved simultaneously to the front. Therefore, it should satisfy the requirements for the missing complement position of both predicates, mandating the requirements to be symmetric between the predicates. Aforementioned, one of Nakao’s major arguments for this analysis was the Case-matching effect – the pivot should match with the Case-requirement of both predicates simultaneously, as exemplified again in (15):^[7]

According to Nakao’s intuition, (15) is much degraded because the Case of the pivot Mary-ni ‘Mary-Dat’ is not licensed in the second conjunct.

This (apparent but not genuine) Case-matching requirement would also be well captured under the multidominance analysis. As admitted by Nakao herself (2009, 2010), the ATB scrambling analysis of Japanese/Korean LNR may be reformulated from the perspective of multidominance. Nakao (2010) explores this perspective, adopting Citko’s (2005) multidominance analysis of LNR. According to Citko, the pivot is parallelly merged with the predicate in each conjunct, as shown in (16).

(16)

multidominance analysis of LNR

Similar to the ATB scrambling analysis, the pivot in (16) was base-generated as the complement of the predicate in each conjunct, capturing the Case-matching requirement. In this analysis, the pivot must move to a higher position for linearization, obeying Kayne’s (1994) Linear Correspondence Axiom (LCA), which dictates that the precedence in linearization be directly mapped from the c-command relation: XP precedes YP if and only if XP asymmetrically c-commands YP. Accordingly, if a multidominated element stays in-situ, it cannot be linearized.

Above all, our experimental finding suggests that the Case-matching requirement may be violated in certain environments. The symmetric accounts regard the pivot to syntactically originate from both conjuncts, requiring both predicates to assign the same Case to the pivot; thus, any Case-mismatches in LNR would be ruled out. However, this prediction is not compatible with our experimental results:

In (17), which repeats (12a) and (12c), the accusative predicate macihayssta ‘welcomed_Acc’ places a Case requirement that mismatches with the Case morphology of the pivot. The symmetric accounts would invariably rule out both (17a) and (17b), regardless of the location of the mismatching predicate. However, our results showed that the Case-mismatch in the first conjunct as in (17b) was significantly worse than that in the second conjunct as in (17a), demonstrating that Case-licensing in Case-mismatched LNR is conjunct-sensitive.

This conjunct-sensitive finding would be properly explained by an asymmetric account of LNR. In fact, Nakao (2010) explores this possibility. However, she claims that the acceptability of the LNR construction in (18) is different from that of the Null Object Construction (NOC) in (19), and gives up the option of reducing LNR to NOC.

Nakao reports that Japanese/Korean NOC as in (19), which resembles LNR as in (18) except that NOC consists of two sentences without coordination, allows the null pronoun pro and its antecedent to have different Case.

Nevertheless, Nakao (2010) adopts a scrambling-plus-resumptive pro analysis, essentially similar to Kim et al.’s (2020) proposal, in accounting for the LNR involving an island, as contrasted in (20) and (21):

In (20) the pivot was based-generated within an island in both conjuncts and has scrambled ATB across a relative island, which is ruled out according to her ATB scrambling analysis. Although admitting interspeaker variation, Nakao judges the Japanese counterpart of (21) as being acceptable, where only the second conjunct has a relative island. She claims that (21) does not display the properties of typical LNR, while proposing that this non-typical apparent LNR resorts to a resumptive pro strategy (Ishii 1991) to avoid an island violation. Consider next the example in (22) where the Case-mismatched pivot was base-generated within an island only in the second conjunct:

Nakao (2010) first reports that certain speakers including herself accept the Japanese counterpart of (22) and then claims that the second conjunct gap/trace is a null resumptive pronoun, evading an island violation. However, if resumptive pro is not much different from the so-called small pro, the emerging option is that all instances of Japanese/Korean LNR can be accommodated under the umbrella of the NOC proposal: scrambling-plus-pro analysis.^[8]

Crucially, Nakao (2010) points out that the Case-mismatched pivot in (22), unlike that in (18), is acceptable. In (22) the first conjunct predicate wiloha ‘comfort_Acc’ assigns accusative Case, while the second conjunct predicate khisuha ‘kiss_Dat’ inside the island assigns dative Case. Nakao acknowledges that (22) is as acceptable as (21) for the speakers who accept (21) in spite of the Case-mismatch. This could be well captured by the scrambling-plus-pro analysis. If the apparent pivots ku cicap-ul ‘the wallet-Acc’ in (21) and ku yepaywul-lul ‘the actress-Acc’ in (22) were to move in an ATB fashion, both sentences should have been ruled out. Under the current proposal, the apparent pivots have moved only in the first conjunct via scrambling, and the gap within the island in the second conjunct is pro. Since there is no movement out of islands, (21) and (22) are acceptable.

In this respect, in line with Kim et al. (2020), we propose that Korean LNR constructions can be reduced to NOC constructions, where the pivot is asymmetrically scrambled from the first conjunct and there is a null pronoun pro in the second conjunct, which is anaphoric to the LNRed pivot as in (23):

In (23) the pivot NP₁ is base-generated only in the first conjunct TP₁ and assigned Case from the first conjunct predicate. In the second conjunct TP₂, the superficially missing argument is pro, which does not care about the morphological Case of its antecedent, i.e., the pivot.^[9] In short, this asymmetric scrambling-plus-pro analysis would be more favorable than the symmetric analyses to explaining the acceptable instances of Case-mismatches in Korean LNR.

4.2 Case type of a pivot

Our second finding was that while the preference for the first conjunct as the Case-licensing locus of pivots was maintained regardless of Case types, Case-mismatched dative pivots were more acceptable than Case-mismatched accusative pivots: (12a) > (12b). This finding may be addressed from the perspective of processing since a perception of acceptability was maintained in (12a) and (12b). In order to account for the difference of acceptability ratings between dative pivots and accusative pivots in LNR, it is reasonable to speculate, therefore, that the difference between (12a) and (12b) is not categorical but gradable in nature. Below, to evaluate the impact of the scrambled pivot on the processing and acceptability of LNRed sentences, we resort to the syntactic and semantic complexity of the filler-phrase.

Syntactically, it has been generally assumed that (24b) and (24c) involve a movement (i.e., scrambling) of an object to the clause-initial position (Miyagawa 2001; Saito 1985). We stress that scrambling does not influence the propositional meaning in (24) (Saito 1985):

In sentence processing, however, several studies on various languages have proved that the processing cost of scrambled word order is higher than that of canonical word order: German (Weyerts et al. 2002), Russian (Sekerina 2003), etc.

There are several explanations for this difference. One possible explanation is that complex structures require heavier processing costs than simple structures do. For example, according to Gibson (1998), processing non-canonical structures requires greater memory resources than processing canonical structures because the former is associated with higher reading times and is therefore more difficult to process. Notice that scrambled phrases may form a sort of filler-gap dependency with gaps. We thus assume that the pivot, fronted via scrambling, constitutes a filler-gap dependency with each gap in LNRed constructions. Under this explanation, the cause of difficulty for scrambling is related to working memory resources.

The difficulty of processing filler-gap dependencies depends not only on the length of the dependency (cf. Alexopoulou and Keller 2007), but also on the property of the filler itself. As an illustration, Hofmeister (2011) proves that complex filler-phrases turn out to facilitate processing at retrieval points. For example, the clefted indefinite in (25b), which is syntactically and semantically more complex than that in (25a), were found to produce faster reading times around the subcategorizing verb banned:

To account for this pattern, Hofmeister (2011) argues that linguistic expressions with more syntactic and semantic features contribute to retrieval process in language comprehension, owing to increased activation and resistance to interference. This contribution ultimately leads to increasing acceptability.

Although it is beyond the scope of this paper to make precise the explanation of the processing cost of scrambling in Korean LNR, we speculate that the processing cost of scrambling in LNR can be alleviated by the degree of filler complexity. Assuming that the dative marker eykey (‘to’ or ‘toward’) is semantically and syntactically more complex (in that it behaves like a postposition) than the structural accusative marker (l)ul, we suggest that the processing cost of LNRed accusative pivot is higher than that of LNRed dative pivot due to the filler complexity of the latter.^[10] To put it differently, we suggest that the complexity of the pivot under the filler-gap dependency in LNR may lower processing cost because the complex filler is easier to retrieve at the site of the gap and this leads to the amelioration in acceptability in (12a) in relation to (12b). To the extent that acceptability is partly derived in terms of processing difficulty, the observed contrast caused by the complexity of the LNRed pivot may play a key role in the perception of the acceptability of the entire sentence. In addition, the increase in acceptability is uniform between dative pivot conditions and accusative pivot conditions: (12a) > (12b) and (12c) > (12d).

Summing up, at the point of the gap site, complex fillers are easier to access and then integrate into the existing structure. We suggest that processing factors have the potential to account for this otherwise unexplained variation in acceptability judgments.

4.3 Dependent plural marking of a pivot

As mentioned before, Nakao (2010) briefly explores to reformulate her ATB scrambling analysis of Japanese LNR from the perspective from Citko’s (2005) multidominance proposal. According to Nakao’s multidominance reanalysis of LNR, the pivot is parallelly merged with the predicate in each conjunct, capturing the Case-matching requirement. In this analysis, the pivot must move to a higher position for linearization, obeying Kayne’s (1994) LCA. This perspective is echoed in Chung (2010), except that the movement of LNRed pivots is forced by Wilder’s (1999, 2008 version of the LCA. Departing from Citko (2005), Chung argues that multidominated elements do not necessarily leave their base position, establishing a derivational connection between RNR as in (26) and LNR as in (27).

The derivations are represented in (28), respectively:

(28a)

RNR

(28b)

LNR

In (28a) the multidominated pivot VP ‘diligently-DPM studied’ remains in-situ, which would provoke a contradiction in linearization under Kayne’s (1994) LCA; since the pivot is dominated by both conjuncts, it has to precede itself, violating reflexivity. Chung, however, argues that there arises no linearization problem under Wilder’s (1999, 2008) LCA, according to which an in-situ pivot gets linearized at the right-edge of the conjuncts, deriving an RNR construction as in (26). According to Chung, moving out of the conjunct involving multidominance is an alternative way to deal with the linearization problem, which is employed in LNR constructions, as shown in (28b); the multidominated pivot yelsimhi-tul ‘diligently-DPM’ moves leftward out of the coordination structure.

Chung (2010) argues that the multidominance analysis has an advantage over the ATB scrambling analysis; it can accommodate the distribution of DPM as well as the Case-matching requirement. Based on Choe’s (1988) observation, he points out that the DPM tul, which is licensed by a c-commanding plural antecedent in the local domain, is licensed in both RNR (Chung 2004) and LNR even when the subject of each conjunct is singular as in (26) and (27), respectively. According to Chung, the well-formedness of such constructions can be explained via multidominance; the two singular subjects in the two conjuncts simultaneously c-command and license the pivot, collectively functioning as a plural antecedent.

However, as attested by Experiment 2, there is an acceptability contrast between RNR and LNR involving the DPM. The LNR construction in (27) is much worse than the RNR construction in (26), which raises an empirical question of whether the DPMed pivot in RNR and LNR is licensed in the same way. Recall that (28a) is acceptable according to the results of Experiment 2. Thus, the RNRed DPM can be licensed under the multidominance proposal in that its in-situ position is co-c-commanded by the split antecedents John and Mary. However, since (28b) is predicted to be acceptable under the multidominance proposal, the experimental result of the DPMed pivot in LNR is not compatible with Chung’s (2010) multidominance analysis.

Instead, the unacceptability of (28b) is predicted by the current scrambling-plus-pro analysis, as shown in (29).

Under the scrambling-plus-pro analysis, there is no chance of the DPMed pivot to be c-commanded by a (split) plural subject. In short, the unacceptable status of the DPMed pivot could be evidence of the scrambling-plus-pro analysis, but not of the multidominance analysis.^[11]