Case-mismatching versus D-linking of ATB wh-questions in Korean

2.1 Asymmetric ATB wh-questions

ATB constructions are characterized by an asymmetric dependency involving one landing site and two (or more) extraction sites (Ross 1967; Williams 1978). A central issue in ATB dependencies is to understand how the movement of two (or more) elements can result in a single filler. One approach to analyzing ATB wh-movement is illustrated below (with elided material shown in angle brackets):

(9)

​

In (9), the two distinct wh-phrases are extracted from each conjunct. To create the one-to-many relationship between the wh-filler and the corresponding gaps, one of these wh-phrases is deleted at the PF level. This deletion must be mandatory, which is unusual since PF deletion is typically optional.

Several linguists have proposed that the conjunction head has its own projection, such as Boolean Phrase, Conjunction Phrase, or &P (Johannessen 1998; Kayne 1994; Munn 1993; Zoerner 1995, among others). For example, Munn (1993, 2001] suggests that ATB wh-questions can be analyzed as a type of parasitic gap construction, where one conjunct undergoes wh-movement, while the other conjunct undergoes null operator (Op) movement. In this view, conjunctions are functional heads that take the second conjunct as their complement and project to a Boolean Phrase (BP), with the entire BP then being adjoined to the first conjunct:^[2]

(10)

According to Munn, the identification of the null operator by the overt operator is enabled by the overt operator’s licensing of the null operator, which necessitates case agreement (cf. Franks 1993, 1995]). It is thus likely that the different case morphology between the overt wh-operator and the null operator lowers the acceptability of ATB wh-questions. However, this case agreement is not mandatory in all instances, as it is possible to have case mismatches between the licenser and the null operator in relative clauses or tough constructions, unlike in ATB dependency:

As is widely known, certain types of ATB wh-questions require symmetric reconstruction in coordinate structures. For instance, the wh-filler of ATB questions cannot cross over a coindexed pronoun that appears in both conjuncts:

The ungrammatical status of (12) suggests that the principle of strong crossover necessitates the ATB-moved wh-filler to be reconstructed symmetrically into both conjuncts.^[3]

On the other hand, asymmetric reconstruction occurs when there is reconstruction into the first conjunct but not into the second conjunct with respect to the following examples for Principle C (13), for weak crossover (14), or for anaphor binding (15). This has been called a proximity effect (Williams 1990):

For instance, the LFs of (13) would be represented as in (16).

As shown in (16a), (13a) is expected to be grammatical because the fronted wh-filler is reconstructed into the first conjunct. However, (13b) is not grammatical because it violates Principle C when the fronted wh-filler is reconstructed into the first conjunct as in (16b). If the fronted wh-filler were to be reconstructed into the second conjunct, the contrast between (13a) and (13b) could not be captured. In short, reconstruction may occur only into the first conjunct but cannot only into the second conjunct.

2.2 Previous analyses

Several theories of ATB movement have been proposed in the literature, which can broadly be categorized into two main approaches: symmetric and asymmetric. Symmetric approaches assume that extraction occurs from both conjuncts in parallel, while asymmetric approaches posit that genuine extraction happens only from one conjunct, with the other gap not directly related to movement. In this section, we will explore three different analyses of ATB wh-questions: sideward movement, multidominance, and parasitic gap.

First, Nunes (2004) presents an analysis that permits a constituent to move into an “unconnected workspace” (i.e., a separate syntactic domain of the local tree) via a copy operation. This allows for ATB dependency to be formed in the following way:

(18a)

sideward movement from a workspace to another workspace

(18b)

two vPs form a conjunction in the same workspace

(18c)

wh-movement from the first conjunct

As illustrated in (18a), the wh-phrase initially merges within the second conjunct, which is constructed in its own workspace. It then moves sideward (rather than upward) to the unconnected first conjunct, which is created in a separate workspace. Sideward movement is a unique operation that involves copying a constituent from one phrase marker to a different, unconnected phrase marker. This type of copying is possible when the initial numeration lacks sufficient elements to fulfill the lexical requirements of other predicates or to maintain parallelism. In a subsequent step, the two vPs are conjoined within the same workspace, as shown in (18b). Finally, the wh-phrase in the first conjunct undergoes operator movement to Spec of CP. For the purpose of linearization, the two copies (i.e., one within the first conjunct and the other within the second conjunct) of the wh-operator should be deleted via chain reduction.

This approach is partly symmetric as extraction occurs from both conjuncts, but also partly asymmetric as movement to Spec of CP proceeds only from the first conjunct. Due to this dual nature, it is debated whether this approach can adequately explain the asymmetric reconstruction effects observed in (13), (14), and (15). As a reviewer pointed out, if sideward movement is considered a form of successive cyclic movement and reconstruction to an intermediate position is possible, this approach might predict that asymmetric reconstruction to the first conjunct could occur.

Now, consider a case where the wh-phrase merges in the first conjunct and then undergoes sideward movement to the second conjunct. After the two conjuncts form a conjunction phrase, the wh-phrase in the second conjunct may undergo wh-movement to Spec of CP. If reconstruction to the intermediate position occurs in this scenario, it should target the second conjunct. To avoid this outcome, the approach might rely on a specific conjunction process, such as Boolean conjunction. However, this approach faces several technical challenges, particularly concerning cyclicity and activity in relation to the Agree mechanism (Fernández-Salgueiro 2008; Salzmann 2012a).

Second, Citko (2003, 2005] attempts to explain ATB dependency through a multidominance structure (Goodall 1987; Moltmann 1992; Riemsdijk 2006, etc.). Specifically, she proposes a parallel merge account which creates symmetric and multidominant structures, as long as the effects of multidominance are invisible at Spell-Out (cf. Chomsky 2000). For the purpose of linearization, the multidominated element must move overtly to a higher position outside the coordination site, obeying Kayne’s (1994) Linear Correspondence Axiom, 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. This is demonstrated in (19).

(19)

Which book did John like and Mary hate?

As Citko’s (2005]: 493) multidominance analysis only posits one copy inside the conjuncts, which should apparently result in symmetric reconstruction, it cannot account for the asymmetric reconstruction effect in (13), (14), and (15).

Third, Munn’s (1993], 2001]) parasitic gap analysis may explain the asymmetric reconstruction effect in ATB wh-questions. According to Munn, under coordination, the second conjunct is adjoined to the first, forming a Boolean structure. It is further assumed that the two coordinated conjuncts may have separate gaps as in (20).

Under this analysis, there is an overt operator (i.e., which book) movement, leaving a gap, within the first conjunct, and there is a covert operator (i.e., Op) movement, leaving a distinct gap, in the second conjunct. More precisely, the wh-filler in ATB questions satisfies the case and θ-requirements of the first-conjunct predicate like, while the null operator satisfies those of the second-conjunct predicate hate. As a result, only the first conjunct is affected by the reconstruction of the wh-filler because the wh-filler originated within the first conjunct and never existed in the second conjunct.

To summarize, previous analyses of English ATB wh-dependencies make distinct predictions regarding the first-conjunct sensitivity of asymmetric reconstruction. In this context, we will explore asymmetric case-mismatching in Korean left-node-raising, which may be connected to case-mismatching observed in Korean ATB wh-questions.

2.3 Locus of a pivot in left-node-raising

In English right-node-raising (RNR) constructions, a right-peripheral string may be shared by the two conjuncts, as illustrated in (21).

The shared element (i.e., pivot) a book in (21) appears only in the second conjunct, but can be interpreted in both conjuncts.

Chung (2010) reports that Korean has a mirror image of RNR, which is called left-node-raising (LNR) (Nakao 2010; Yatabe 2001 for Japanese LNR), as in (22).

The left-peripheral pivot chayk-ul ‘book-Acc’ in (22) can be interpreted as the missing element in both conjuncts. In order to explain the nature of LNR, Yatabe (2001) and others (Chung 2010; Kim et al. 2023; Nakao 2010, etc.) have proposed various analyses.

According to the symmetric approach, the pivot, which originated in both conjuncts, was subcategorized by both conjunct predicates. In this regard, Chung (2010) explores a multidominance account of LNR, adopting Citko’s (2003, 2005] theory of multidominance as follows:

(23)

The multidominated pivot in (23) was base-generated as an argument of each predicate. This naturally captures the case-matching requirement. As mentioned earlier, the pivot under Citko’s multidominance theory must move to a higher position in order to be linearized.

What is notable is that LNR is more acceptable when the case of the pivot is licensed in the first conjunct than in the second conjunct. Kim et al.’s (2020, 2023] experimental findings show that the case-matching requirement may be obviated as long as the case of the LNRed pivot is licensed in the first conjunct, as shown below:

In (24a), the dative case of the pivot is licensed by the first-conjunct predicate cwu(ta) ‘to give,’ but not by the second-conjunct predicate maciha(ta) ‘to welcome.’ Conversely, in (24b), the dative case of the pivot is not licensed by the first-conjunct predicate maciha(ta) but is licensed by the second-conjunct predicate cwu(ta). According to the multidominance approach, the pivot in LNR is simultaneously dominated by both conjuncts and then moved out of the shared position. As a result, case-mismatching of the pivot is predicted to be ill-formed regardless of where it occurs. However, the experimental results from Kim et al. (2020, 2023] show that case-matching in the first conjunct, as in (24a), is significantly more acceptable than case-matching in the second conjunct, as in (24b).

This first-conjunct-sensitivity of case-licensing in LNR could pose a dilemma for the multidominance approach (Chung 2010; Nakao 2010). Following Kim et al. (2020, 2023], we maintain that Korean LNR constructions are a type of null object constructions, where the pivot is asymmetrically scrambled only from the first conjunct and there is a null pronoun pro in the second conjunct, which is anaphoric to the LNRed pivot. The following shows the derivation of (24a).

The pivot in (25) was base-generated only in the first conjunct and assigned case exclusively from the first-conjunct predicate. The missing argument in the second conjunct is pro.^[4]

To summarize, the scrambling-plus-pro account would be better than the multidominance account with respect to the case-mismatching property in Korean LNR. We will show later on that the first-conjunct case-licensing preference of the pivot in LNR can be carried over to ATB wh-dependency in Korean.

2.4 D-linking

As previously mentioned, there has been an issue regarding the difference between bare wh-phrases and D-linked wh-phrases in relation to wh-dependency, which relates to the gaps that are necessarily linked to wh-phrase fillers. D-linked wh-phrases have generally been considered as being connected to previous discourse in some way, but for our purposes here, we will use it primarily as a term for which-phrases that indicate a group of individuals. In the syntax literature, it has been noticed that extracting the unmarked wh-phrase what from the wh-island results in a sentence that is considered unacceptable as in (26a). However, the extraction of the D-linked wh-phrase which problem in (26b) is seen as acceptable. This suggests that the wh-island violation in (26b) is somehow repaired, although it is still less acceptable than the wh-extraction in (26c).

The phenomenon of D-linking thus poses an intriguing problem, and several explanations have been put forward to account for it.

One popular approach suggests that island violations like (26a) are unacceptable due to a syntactic issue. According to Rizzi (1990, 2004], the dependency between the filler what and its gap site in (26a) violates a fundamental property of syntax called relativized minimality, which does not allow for the dependency between a filler and a gap when there is another intervening filler, such as who in the case of (26a), that could also potentially form a similar dependency with the gap. However, D-linked wh-phrases share similarities with fronted topics, which are not affected by relativized minimality effects. This is because D-linked wh-phrases also contain additional lexical material beyond the wh-word itself and rely on previously mentioned elements in the discourse. Therefore, if D-linked wh-phrases are interpreted as topics, they should be able to bypass the relativized minimality requirement, which would increase their acceptability.

The second popular approach proposes that the reason for unacceptability of island violations like (26a) is due to limitations in working memory (Goodall 2015; Hofmeister and Sag 2010; Kluender 1998; Kluender and Kutas 1993, etc.). The filler what must be held in working memory until it can be reintegrated into the structure at the gap site in the embedded clause. This is difficult because maintaining the filler what in working memory, while processing a clause boundary and the intervening filler who, overwhelms the limited capacity of the processor. Thus, reintegration of the filler what is less likely to succeed and the sentence is perceived as unacceptable. However, when the filler is D-linked as in (26b), it requires more initial processing due to the presence of lexical material, which gives the filler which problem a higher level of initial activation in working memory. This enables the filler which problem in (26b) to survive more successfully until it can be reintegrated at the gap site. There is substantial evidence that such a processing advantage for D-linked fillers exists, likely leading to higher acceptability (Diaconescu and Goodluck 2004; Frazier and Clifton 2002; Goodall 2015; Hofmeister and Sag 2010; Hofmeister et al. 2013; Kluender 1998, etc.).

The working memory approach to D-linking and islands differs from the syntactic approach in a crucial way. Both approaches assign unique properties to D-linked fillers, but only the working memory approach would expect these properties to enhance acceptability even when there is no island violation present. Specifically, D-linked fillers are thought to facilitate individuation, allowing them to avoid island constraints, but the syntactic approach does not anticipate the effect on acceptability outside of island environments. In contrast, the working memory approach proposes that D-linked fillers have a heightened level of activation, promoting better retention in working memory and reintegration at the gap position, regardless of the structure. It thus predicts that making a filler D-linked should enhance acceptability in both well-formed and ill-formed contexts. Therefore, the working memory approach and the syntactic approach have a fundamental divergence in that the former predicts that D-linking will increase acceptability in both well-formed and ill-formed contexts, whereas the latter does not.^[5]

As far as we know, no one has investigated whether D-linking affects the acceptability of case-mismatch violations in ATB wh-dependency, regardless of where the case-mismatching gaps occur. Our study aims to address this question.

3.1 Participants, materials, and design

We recruited 43 self-reported native Korean speakers, who were undergraduate students at a university in South Korea, with a mean age of 21.93 years (SD: 2.04). They agreed to participate in the study and provided informed consent in exchange for course credit. The online experiment was generally completed within 10 min. We excluded responses from three participants who were found to be inattentive during the task, using a specific procedure described in Section 3.2. Therefore, only the responses from 40 participants (10 from each of the four lists) were used in the analysis.

Experimental items were wh-questions and were prepared using a 2 × 2 design, crossing LOCUS and FILLER. The four conditions are exemplified in (27).

In the [1st | D-linked] condition, the dative case of the D-linked wh-filler is licensed in the first conjunct, while in the [2nd | D-linked] condition, it is licensed in the second conjunct. Similarly, in the [1st | Bare] condition, the dative case of the bare wh-filler is licensed in the first conjunct, while in the [2nd | Bare] condition, it is licensed in the second conjunct. The full list of experimental items is available online.^[6]

Sixteen sets of experimental conditions were created, with each set containing four conditions that were matched for lexical content. These sets were then assigned to four lists using a Latin square design to ensure that each list had one item from each set. Each list contained 16 experimental items and 48 filler items, with a ratio of 1:3 for experimental items to fillers. The filler items were of comparable length to the experimental items but varied in acceptability. Overall, each list contained 64 sentences.

3.2 Procedure

The experiment was conducted using a web-based platform called PCIbex (Zehr and Schwarz 2018). Participants viewed sentences one at a time on a computer screen and made acceptability judgments on a 1–7 Likert scale. The experiment consisted of 16 sets of test items and 16 “gold standard” filler items, including eight good and eight bad filler items which had previously been assessed on about 200 participants to determine their expected value (1 for bad, 7 for good). For each gold standard item, the difference between each participant’s response and its expected value was calculated, squared, and summed to obtain a sum-of-the-squared-differences value. Participants with values greater than two standard deviations from the mean were excluded from the analysis as they were deemed not paying attention during the task (cf. Sprouse et al. 2022).

3.3 Data analysis

Before analyzing the data, the raw judgment ratings, including both experimental and filler items, were converted to z-scores to eliminate potential scale biases between participants (Schütze and Sprouse 2013). This was done to standardize all participants’ ratings to the same scale and correct for the possibility that individual participants treat the scale differently. The data were analyzed using linear mixed-effects (regression) models in the R software environment (R Core Team 2020), estimated with the lme4 package (Bates et al. 2015). Linear mixed-effects models allow for the simultaneous inclusion of random participant and random item variables (Baayen et al. 2008). The maximally convergent random effect structure with participant and item was used (Barr et al. 2013), and p-value estimates for the fixed and random effects were calculated by Satterthwaite’s approximation (Kuznetsova et al. 2017).

3.4 Results

The data from 40 participants, which equates to a total of 640 tokens for the four conditions, were analyzed. Figure 1 displays the average z-scores for the acceptability ratings for each of the four experimental conditions. A score of zero reflects the overall mean acceptability rating. A positive z-score indicates that the conditions are more acceptable, while a negative z-score indicates that the conditions are less acceptable. The LOCUS effect is shown by the downward slope of the lines, while the FILLER effect is shown by a vertical gap between the lines. The interaction can be seen in the non-parallel lines that emerge when the four conditions are plotted together.

Figure 1:

Mean acceptability scores for the experiment (error bars = SE).

Table 1 provides an overview of the linear mixed-effects model used in the analysis. The model includes LOCUS and FILLER as fixed effects, as well as their interaction. Random intercepts for both participants and items were included, along with the maximum number of random slopes justified by the data.

The analysis revealed a significant effect of LOCUS, indicating that the [1st] condition was rated as more acceptable than the [2nd] condition. There was no significant effect of FILLER, but there was a significant interaction between LOCUS and FILLER. The results of the interaction showed that the D-linking effect was significant in the [2nd | D-linked] condition where the first conjunct case-licensing requirement was violated, while it was not in the [1st | D-linked] condition where the first conjunct case-licensing requirement was satisfied. Planned pairwise comparisons indicated that the two 1st conditions did not significantly differ from each other (mean: 0.209 vs. 0.099; β = 0.110, SE = 0.084, t = 1.313, p = 1.000), while the [2nd | D-linked] condition was rated as significantly more acceptable than the [2nd | Bare] condition (mean: 0.032 vs. −0.364; β = 0.396, SE = 0.084, t = 4.744, p < 0.001).^[7]

To summarize, there were two main findings. First, the case-licensing of bare wh-fillers in the first conjunct is far more acceptable than that in the second conjunct, supporting the claim that the first-conjunct case-licensing of bare wh-fillers is crucial in Korean ATB wh-questions. Second, the [2nd] conditions, which violate the first-conjunct case-licensing requirement in ATB dependency, are not always unacceptable; they showed super-additive degradation when the bare wh-filler, but not the D-linked wh-filler, was the pivot.