In English and many other languages of Indo-European origin, a wh-phrase generally moves overtly to a clause-initial position in wh-interrogative sentences. As is well known, such a movement cannot cross islands, namely, the structures out of which a constituent cannot be extracted. One of the famous islands is the complex NP island, where the complex NP refers to the NP modified by a clause (Bošković, 2016). The complex NP island/constraint requires that extraction from a complex NP is not allowed (Ross, 1967). For example, the clauses ‘that wrote’ and ‘that John wrote’ modify the NP ‘the book’ in (1a) and (1b), respectively. They are islands. After who in (1a) moves out of the relative clause ‘that wrote,’ the complex NP island/constraint will be violated, and the derived sentence will be ungrammatical. The same is true of (1b).

However, in Mandarin Chinese (hereafter ‘Chinese’) and other East Asian languages, wh-elements stay in situ. Furthermore, different from adjunct wh’s-in-situ, argument wh’s-in-situ do not seem to be subject to island constraints, as (2) shows.

This difference in island sensitivity between argument and adjunct wh’s-in-situ is known as argument–adjunct asymmetry in the literature.¹ This phenomenon has drawn the interest of many scholars, and many influential hypotheses have been advanced to account for it (see Huang, 1982a,b; Lasnik and Saito, 1992; Aoun and Li, 1993a; Tsai, 1999; Cheng, 2009, among many others).

Inspired by the studies of Sprouse (2007), Sprouse et al. (2012), Sprouse and Hornstein (2013), Lu et al. (2020) used the acceptability judgment paradigm to investigate wh’s-in-situ in Chinese. They report that both argument wh’s-in-situ and adjunct ones are sensitive to the Complex NP Island. This study is interesting because if it is proved to be true, we would need to reconsider the existing theory of wh’s-in-situ that has been established based on the argument–adjunct asymmetry for wh’s-in-situ. Nevertheless, in this study we will point out that Lu et al.’s (2020) experimental design has some drawbacks, which make their findings unreliable. We critically note that when object wh’s-in-situ like shénme ‘what’ are located inside a relative clause, they are subject to a pragmatic constraint, suggesting that the VP (formed by a verb and its wh-object) in the relative clause tends to describe the prominent/salient feature of the relativized nominal head. Improving on the design in experimental materials, we conducted two experiments on Chinese wh’s-in-situ. The results of the experiments show that unlike adjunct wh’s-in-situ, argument wh’s-in-situ are not sensitive to island constraints, which is in line with the long-established findings on the issue at stake.

Let us first expound the logic of the factorial design for isolating island effects, on which the previous formal judgment study into wh-in-situ, namely Lu et al. (2020) is based. Sprouse and his colleagues argue that the lower acceptability of island violating sentences results not only from the violation of a grammatical constraint, but also from such (non-grammatical) processing factors as dependency length and structure. In other words, the processing of dependency length and complex structure also contributes to the low acceptability of island violating sentences. With the dependency length in a sentence becoming longer, the processing load of the sentence will increase and its acceptability will decrease. Likewise, the complex structure (i.e., the structure containing island) is more difficult to process than the simple one (i.e., the structure without an island). Given this, Sprouse and his colleagues developed a factorial experiment design to isolate island effects, which has been fruitfully adopted by many researchers to study island effects in various languages (Sprouse, 2007; Sprouse et al., 2011, 2012, 2016; Almeida, 2014; Michel, 2014; Atkinson et al., 2016; Kush et al., 2018, 2019; Stepanov et al., 2018; Keshev and Meltzer-Asscher, 2019; Pañeda et al., 2020; Kush and Dahl, 2022). The first factor in such an experiment paradigm is Dependency Length, which has two levels: long and short. In the short condition, a wh-phrase moves from a position in the matrix clause to the sentence-initial position, and in the long condition a wh-phrase moves from a position within the embedded clause to the sentence-initial position. The second factor is Structure, which also has two levels: island and non-island. The island condition contains an island, and the non-island condition does not. After the two levels of each factor are crossed, four conditions are created, as is demonstrated with a complex NP island below.

The island effect can be isolated with the logic of subtraction. First, the total effect is quantified by calculating the difference between (3a) and (3d) (i.e., [3a − 3d]), and then the effect of dependency length can be captured by calculating the difference between (3a) and (3b) (i.e., [3a − 3b]), and the effect of structure can be isolated by calculating the difference between (3a) and (3c) (i.e., [3a − 3c]). The island effect can now be obtained by subtracting the effect of dependency length and that of structure from the total effect. Put differently, the island effect can be quantified with the following formula: island effect = [3a − 3d] − [3a − 3b] − [3a − 3c]. If there is no island effect, the score for the island effect will be zero (in mathematic terms, [3a − 3d] = [3a − 3b] + [3a − 3c]). By contrast, if there is an island effect, the score for the island effect will be larger than zero.² In other words, if the island effect is present, the total effect will be greater than the sum of the dependency length effect and the structure effect (in mathematic terms, [3a − 3d] > [3a − 3b] + [3a − 3c]), and the island effect is reflected by the super-additive effect. The island/super-additive effect can be identified statistically. If the island effect is present, there will be a significant interaction effect between the two factors: Dependency Length and Structure.

The formula for the island effect also has an equivalence like ‘island effect = [3d − 3c] − [3b − 3a].’ The formula can be interpreted as follows. [3b − 3a] can capture the effect of dependency length. If there is an island/super-additive effect, the difference between (3c) and (3d) should be greater than the difference between (3a) and (3b) though (3c) and (3d) appear to be different only by dependency length (Keshev and Meltzer-Asscher, 2019).

Inspired by the experimental paradigm in studying island effects, Lu et al. (2020) adopted a 2 × 2 × 2 factorial experiment design, involving the three factors such as Dependency Length (short vs. long), Structure (non-island vs. island) and Wh-Category (argument vs. adjunct). This gave the following eight conditions: (i) short + non-island + argument, (ii) long + non-island + argument, (iii) short + island + argument, (iv) long + island + argument, (v) short + non-island + adjunct, (vi) long + non-island + adjunct, (vii) short + island + adjunct, (viii) long + island + adjunct.

Consider the following sentences cited from their experiment, representing one set of their stimuli.

In these sentences, the verb xiǎngzhīdào ‘wonder’ takes an interrogative clause as its complement. For the sake of exposition, we will name the complement of xiǎngzhīdào ‘wonder’ as CP₁. In the short condition, SpecCP₁/C₁ and the wh-phrase form mono-clausal dependency. If the wh-phrase is to move covertly to SpecCP₁, such a movement in (4), (6), (8), and (10) will be a short-distance one. By contrast, in the long condition, SpecCP₁/C₁ and the wh-phrase form bi-clausal dependency. If the wh-phrase is to move covertly to SpecCP₁, such a movement in (5), (7), (9), and (11) will be a long-distance one. The factor Structure controls for whether the complement of xiǎngzhīdào ‘wonder’ contains an island (the island condition) or not (the non-island condition). The factor Wh-Category controls for whether the wh-phrase at hand serves the role of an argument or an adjunct.

There were 24 target items and 72 filler sentences for the experiment. The participants were instructed to rate the naturalness of sentences on a seven-point Likert scale, with 1 being completely unnatural and 7 being completely natural. Their findings show that for both argument and adjunct wh’s-in-situ, there was a significant super-additive interaction effect of Dependency Length × Structure. This is an intriguing result because if proved to be true, their findings pose a challenge to the long-standing generalization on argument–adjunct asymmetry in wh’s-in-situ. That is, argument wh’s in Chinese are not subject to island constraints, whereas adjunct wh’s are (see, for example, Tsai, 1999). However, a careful inspection of Lu et al.’s (2020) experiment design indicates that we cannot be highly positive about the significance of their experiment because there are some drawbacks with their study. First, argument wh’s-in-situ can occur in both subject and object positions. In their design, they used two argument wh’s-in-situ, namely shéi ‘who’ and shénme ‘what.’ Shéi ‘who’ occurred in the short condition, serving as a subject, while shénme ‘what’ occurred in the long condition, serving as an object. But only shénme ‘what’ was manipulated to occur inside an island (i.e., only the island effects on the object shénme ‘what’ was tested). With the Complex NP Island as a test case, they argued that both argument and adjunct wh’s are subject to island constraints. This conclusion was too hastily drawn because the subject shéi ‘who’ was placed outside an island in their design (i.e., the island effects on the subject shéi ‘who’ were not tested),³ and there is no compelling evidence that object wh’s-in-situ are a typical case for testing island sensitivity. Without a detailed study of island effects on subject wh’s-in-situ, it would be particularly inappropriate to jump to the conclusion that both wh-argument and wh-adjunct elements are sensitive to island constraints. Second, when the object wh-in-situ in the relative clause is shénme ‘what,’ the interpretation of the construction is susceptible to factors not bearing on island constraints. One of the factors is that when object wh’s-in-situ are in a relative clause, the construction is subject to a pragmatic constraint, suggesting that the VP (formed by a verb and its wh-object) in the relative clause tends to characterize the prominent feature of the relativized nominal head.⁴ To illustrate the point, consider (7). Its alleged low acceptability might be caused by pragmatic inappropriateness rather than by a violation of an island constraint. To be exact, (7) is taken to be unacceptable because—without a proper context—eating a certain thing is not the prominent feature of a girl. On the other hand, if we provide an appropriate context, (7) will become acceptable. Suppose that there are three girls in a street eating different kinds of things: One girl is eating a hamburger, another girl, an omelet, and the third girl, fried chicken. If both the speaker and the hearer share this common ground, (7) will be acceptable because, in this context, eating a certain thing is the characteristic feature of the girls at issue, and it is natural to ask what the girl that Bill met ate.

The argument to the effect of displaying the role of a pragmatic factor in the interpretation of an island-internal object wh-in-situ is that when we replace the lexical item in (7) with other appropriate ones, the resulting sentences such as (12)–(14) will become acceptable too.

For example, in the context of (12), the girl’s eating habit is important because after she and Bill got married, this would affect their relationship. Perhaps it will also affect John if he knows Bill. On this condition, it is natural to ask what the girl likes eating. The same kind of construal applies to (13). The content is a prominent feature of a book because in most cases whether the book can attract a person or not is dependent on its content. Consequently, it is valid to ask what the book that Bill bought is about. Likewise, teaching is the prominent feature of a teacher as the primary duty of a teacher is to teach. Given this, it is natural to ask what he taught, as in (14). However, Lu et al. (2020) fail to put the above pragmatic confounding factor under control.

To sum up, although Lu et al.’s (2020) results are intriguing, their experiment design still has at least two drawbacks: one is that they fail to study the island effects of subject wh’s-in-situ, and the other is that when studying the island effects of object wh’s-in-situ, they fail to put the pragmatic confounding factor under control. These drawbacks undermine their conclusions that argument wh’s-in-situ in Chinese are subject to an island constraint, and that there is no argument–adjunct asymmetry in wh’s-in-situ. In the next section, we will introduce our two experiments, in which the weaknesses of Lu et al. (2020) were resolved. In one experiment, the island sensitivity of subject wh’s-in-situ was tested, and in the other one, the island sensitivity of object wh’s-in-situ was tested with the pragmatic confounding factor under control.

This study brings to light several important findings. First, a significant three-way interaction effect of Structure × Wh-Category × Dependency Length was observed. In this regard, our study renders reinforcing support to the long-standing generalization concerning the argument–adjunct asymmetry in wh’s-in-situ in Chinese (Huang, 1982a,b; Tsai, 1999; Cheng, 2009), contrary to the findings of Lu et al. (2020). In accordance with Sprouse (2007), Sprouse and Hornstein (2013), and Sprouse et al. (2016), to name a few, syntactic island effects can be measured systematically using the factorial design, and they are represented by the super-additive effects that stem from combining both the effects of Structure and those of Dependency Length. Adopting this experimental paradigm, we have demonstrated that the wh-adjunct wèishénme in the current study displays island effects as there arises a significant Structure × Dependency Length interaction giving rise to the super-additive effect. This, in turn, supports the well-acknowledged claim that wh-adjuncts are sensitive to island constraints (Huang, 1982a,b), or that covert (operator) movement is involved in the derivation of wh-adjuncts (Aoun and Li, 1993a; Tsai, 1999). By contrast, in the current study no super-additive effect is observed for wh-arguments, which counters the findings of Lu et al. (2020). This means that the traditional generalization should be re-endorsed—that is, wh-arguments in Chinese are not sensitive to island constraints. The underlying reason behind this re-endorsement is that wh-arguments are derived by unselective binding (Tsai, 1994, 1999), or that an operator licensing them is base-generated in a position external to islands (Aoun and Li, 1993a). Put differently, the significant difference between the adjunct wh’s-in-situ and the argument wh’s-in-situ in the long + island condition reflects the different syntactic properties of the adjunct wh’s-in-situ and the argument wh’s-in-situ. The argument wh’s-in-situ do not undergo covert movement. Without violating the island constraint, the relevant sentences are rated to be natural. By contrast, the adjunct wh’s-in-situ undergo covert movement, violating the island constraint. Consequently, the relevant sentences are rated to be unnatural.⁹

Second, Lu et al.’s (2020) experiment result shows that although the mean acceptability rating score of the long + island + argument condition is low, it is still a little higher than that of the long + island + adjunct condition. Given this, they acknowledge that Huang (1982a,b) and Tsai (1994, 1999) are correct in noting that argument wh’s-in-situ in islands are higher in acceptability than adjunct wh’s-in-situ in islands. Nevertheless, they argue that it does not support the argument-adjunct asymmetry in that it may be a reflection of the main effect of Wh-Category. They further argue that there is no real argument–adjunct asymmetry in wh’s-in-situ in terms of island sensitivity. The argument-adjunct asymmetry in wh’s-in-situ reported in the literature is attributed to the methodology of acceptability judgment. In their opinion, while making judgments, linguists implicitly construct minimal pairs, i.e., they implicitly employ a minimal-pair experiment design. In the experiment employing such a design, the preference for wh-arguments in situ rather than wh-adjuncts in situ would be a reflection of the main effect of Wh-Category (argument vs. adjunct) rather than wh-adjuncts’ sensitivity to island constraints. Our two experiments have demonstrated that the validity of their claim needs to be re-evaluated. Recall the results of Experiment 1 and Experiment 2. They reveal that the wh-arguments were not always rated to be better in acceptability than the wh-adjunct, which is different from the findings made in Lu et al. (2020). Particularly, our two experiments have found that in the short + non-island condition, the wh-adjunct was judged to be more acceptable than the wh-arguments. Further, no significant difference between the wh-argument and the wh-adjunct was observed in the short + island condition of Experiment 1, and the wh-adjunct was slightly better than the wh-argument in the same condition of Experiment 2.¹⁰ Based on the findings in our study, we can conclude that even if Lu et al.’s (2020) assumption is reasonable that while making judgments, linguists implicitly employ a minimal-pair experiment design, the preference for argument wh’s-in-situ in the island condition cannot be attributed to the effects of Wh-Category.

In addition, it is worth noting that previous studies often focus only on island effects that arise on adjunct wh’s-in-situ, without explicit reference to minimal pairs involving both adjunct wh’s-in-situ and argument wh’s-in-situ. It is, then, implausible to claim that the participants in these studies always implicitly construct a minimal pair, which—as Lu et al. (2020) claim—leads to argument–adjunct asymmetry in island sensitivity.

Perhaps, at this moment one may wonder why the wh-adjunct in the short + non-island condition was judged to be more acceptable than the wh-argument in our experiments and why there was a difference in this regard from Lu et al. (2020). As to the first question, we think the answer might lie in different sensitivity to dependency length between argument wh’s-in-situ and adjunct wh’s-in-situ. As can be seen from the two interaction plots of our experiments, the dotted lines for argument wh’s-in-situ in the left panels are almost flat, and the dotted ones for adjunct wh’s-in-situ in the right panels slope downward visibly. The identical result is also found in Lu et al. (2020). This suggests that adjunct wh’s-in-situ are more sensitive to dependency length than argument wh’s-in-situ. Therefore, it is highly likely that the mean acceptability rating score for wh-adjuncts is higher than or approximately equal to that of wh-arguments (when the processing difference between adjunct wh’s-in-situ and argument wh’s-in-situ is small) in short dependency condition, and the rating score for wh-adjuncts is lower than that of wh-arguments in long dependency condition. Given this, it will be unsurprising that in the short + non-island condition, wh-adjuncts were judged to be more acceptable than wh-arguments.

If we continue to investigate why there is a difference in sensitivity to dependency length, we speculate that the reason might be that wh-adjuncts and wh-arguments establish dependency in different ways: wh-adjuncts establish dependency through LF movement and wh-arguments through unselective binding (Tsai, 1999). It will be a very interesting topic to study whether varying degrees of sensitivity to dependency length can be used as a criterion for determining how dependency is established. We leave this topic for future study.

As to the question why argument wh’s-in-situ are preferred over adjunct wh’s-in-situ in the short + non-island condition in Lu et al.’s experiment, a result different from ours, we speculate that one reason lies in the choice of an embedded subject. The embedded subject in the adjunct + short + non-island condition of their experiment is a proper noun, as in (8). It has been reported in the literature that a proper noun/definite phrase provokes a higher processing cost (Warrena and Gibson, 2002; Hofmeister and Sag, 2010), which will reduce the acceptability of this type of sentence. Therefore, it is reasonable that the mean rating score for argument wh’s-in-situ is greater than that of adjunct wh’s-in-situ in Lu et al.’s experiment. In contrast to the stimuli in Lu et al.’s experiment, the embedded subject in the adjunct + short + non-island condition of our experiment is a pronoun, as in (19). The embedded subject may thus not have induced any additional processing cost. This may have resulted in the explicit short dependency effect of adjunct wh’s-in-situ. Consequently, the mean rating score for adjunct wh’s-in-situ is greater than that of argument wh’s-in-situ.

Third, we suspect, as pointed above, that the low acceptability score reported by Lu et al. (2020) concerning wh-arguments in the long + island condition results from pragmatic inappropriateness of their experimental items rather than the island constraint at stake. In other words, we propose that when the object wh-in-situ is in a relative clause/an island, the processing of such a sentence is susceptible to an additional pragmatic constraint. This is confirmed by Experiment 2, in which the experiment items were carefully designed so that the pragmatic confounding factor could be removed. The results of this experiment have shown that once the pragmatic confounding factor is under control, the acceptability score for the object wh-in-situ in the long + island condition is very high.

Moreover, a comparison of the results of Experiment 1 and Experiment 2 also suggests that we should be on the right track in assuming that when the object wh-in-situ is in a relative clause/an island, it is subject to an additional pragmatic constraint, and the low acceptability score reported by Lu et al. (2020) concerning wh-arguments in the long + island condition results from pragmatic inappropriateness. The results of Experiment 1 and Experiment 2 have revealed that the subject wh-in-situ and the object wh-in-situ behave almost in the same way. For example, in the long + island condition, both the subject wh-in-situ and the object wh-in-situ are judged to be better than the adjunct wh-in-situ, and in the short + non-island condition, both the subject wh-in-situ and the object wh-in-situ are considered to be worse than the adjunct wh-in-situ. Their parallel behaviors are well expected because both subject and object wh’s-in-situ serve as arguments. Unlike Experiment 2, Lu et al. (2020) left the pragmatic confounding factor untouched. As a result, they reported that both the object wh-in-situ and the adjunct wh-in-situ in the long + island condition are rated to be very low in acceptability. If their findings are reasonable, then object wh’s-in-situ and adjunct wh’s-in-situ are distinguished in acceptability from the subject wh-in-situ in the long + island condition. Such a difference is surprising given that both subject wh-in-situ and object wh-in-situ are arguments. After fixing the experimental materials so that the pragmatic confounding effect is controlled for or eliminated, Experiment 2 has shown that the object wh-in-situ exhibits the same characteristics as the subject wh-in-situ does. This argues strongly in favor of our claim that it is the pragmatic confounding factor that prevents Lu et al. (2020) from uncovering the real nature of object wh’s-in-situ. Put differently, under the impact of the pragmatic confounding factor, their result shows that just like adjunct wh’s-in-situ, there is also a significant Structure × Dependency Length interaction for argument wh’s-in-situ. Consequently, they fail to find a significant three-way interaction of Structure × Wh-Category × Length. Actually, the significant Structure × Dependency Length interaction for argument wh’s-in-situ in Lu et al. (2020) is a reflection of pragmatic constraint rather than syntactic island constraint.

Our claim can also explain the practice that theoretical linguists in Chinese generally rely on: when intending to show that argument wh’s-in-situ in this language are not sensitive to island constraints, they usually use subject wh’s-in-situ as in (37) rather than object wh’s-in-situ as a test case.¹¹

According to the findings of our experiments, the reason lies in that unlike subject wh’s-in-situ, when object wh’s-in-situ occur in a relative clause, namely an island, they are more likely subject to an additional pragmatic constraint. Therefore, it is much easier to create sentences with appropriate subject wh’s-in-situ than sentences with object wh’s-in-situ to examine the island sensitivity of argument wh’s-in-situ.

Different from their counterparts in English, wh-elements in Chinese remain in situ in question sentences. Argument wh’s-in-situ in Chinese are insensitive to island constraints, unlike adjunct wh’s-in-situ. This has led to the celebrated generalization regarding the argument and adjunct asymmetry in wh’s-in-situ in this language. With the acceptability judgment experiment, Lu et al. (2020) challenge this long-established generalization. They argue that this asymmetry is an illusion, and that both the adjunct wh-in-situ and the argument wh’s-in-situ are subject to island constraints. In this study, we point out that their results are not convincing because their experimental design has some drawbacks. We redesigned the experimental materials in question to examine island effects on wh-elements in situ in Chinese. The results of the two experiments in this study show that the argument versus adjunct asymmetry in wh’s-in-situ is present. Furthermore, the argument–adjunct asymmetry at issue cannot be attributed to the main effect of Wh-Category.

On top of supporting the traditional theoretical generalization on Chinese wh’s-in-situ, this study also discovers that when object wh’s-in-situ are located inside a relative clause, they are subject to a pragmatic constraint, suggesting that the verb phrase in the relative clause tends to describe the prominent/salient feature of the relativized nominal head. This contributes to the understanding of the processing of relative clauses and wh-in-situ sentences.

Finally, it will be much fair to point out that although Lu et al.’s (2020) findings are not supported by our experiments, their research is very enlightening. Inspired by their study, in the future we will examine whether the argument-adjunct asymmetry is also observed in other island environments and whether our findings in this study may shed some light on other theories related to asymmetry in unselecting binding, covert LF movement, ECP, etc.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

QT conceived and designed the study, implemented the experiments, analyzed the data, drafted, and revised the manuscript. M-KP provided theoretical guidance when necessary and made revisions to the manuscript. XY participated in the conception of the study, statistical analysis, as well as the revision of part of the manuscript. All authors contributed to the article and approved the submitted version.