Investigating locality constraints on wh-in situ in French: an experiment

1.1 Wh-in situ questions across languages and in French

Colloquial French can be described as a ‘mixed’ language, as it shows both in situ and ex situ wh-phrases in normal information seeking questions (e.g., Dryer 2013) in contrast to other languages that show either only wh-in situ items (e.g., Mandarin Chinese or Japanese: [1]) or only wh-ex situ items (e.g., English: [2]).

In many languages the absence of wh-movement correlates with a peculiarity concerning the nature of the wh-elements. In Japanese and in Mandarin Chinese, wh-words such as dare ‘who’ and nani ‘what’ in (3a) and shenme ‘what’ in (3b) do not carry an interrogative force per se and can be used as indefinite pronouns when produced in a non-interrogative context.

This is not the case in French, where seemingly identical non-ambiguous wh-elements can be either left in situ or moved at the left periphery in a variety of different questions strategies, illustrated in (4).

A number of differences, variously described in the literature, suggest that the two strategies, in situ and ex situ, are not simply two options of the same grammar in French. First of all, they partially belong to different registers or varieties. While ex situ strategies (illustrated above in [4a, c, d]) belong to registers of varying degrees of formality, French wh-in situ in (4b) only belongs to a register we might call Colloquial French, which is mostly oral. Even when focusing on this register, though, wh-in situ and wh-ex situ questions display a number of semantic and syntactic differences. Wh-in situ questions have been widely described as more constrained than fronted wh-questions (Boeckx 1999; Bošković 2000; Chang 1997; Cheng and Rooryck 2000; Mathieu 1999). Many researchers have argued that wh-in situ in French is only a root phenomenon and cannot be found (i) in embedded clauses introduced by the complementizer que, (ii) in embedded infinitival clauses introduced by de or à, (iii) with negation, (iv) with quantifiers, and (v) with adverbs. Other scholars disagree and describe embedded in situ wh-questions with negation, quantifiers, and adverbs as attested in the relevant register (Adli 2006; Baunaz 2011; Starke 2001). This sharp disagreement may be due to the fact that the linguists who have worked on the topic do not describe the same dialect of French. We shall leave aside this important controversy as not immediately relevant to the research described here, which only focuses on direct positive interrogative clauses.

Researchers importantly agree that in general processing wh-questions involves constructing a non-local dependency between the scope position of the question operator and the position where the wh-phrase receives its thematic role. In wh-ex situ questions, an overt dependency is built between the moved wh-phrase and its original position (i.e., the gap: [5a]), while in wh-in situ questions, a covert dependency is formed: (5b).

There are three main approaches to covert wh-dependencies: LF-movement, unselective binding, and heavy pied-piping.

In a nutshell, the LF-movement hypothesis (stemming from Huang 1982) proposes that both in situ and ex situ questions involve wh-movement, but that it happens abstractly at Logical Form (LF) in in situ wh-questions, and concretely in overt syntax in ex situ wh-questions. In Minimalist terms, this can be rephrased saying that in ex situ questions it is the higher copy of the wh-element that is pronounced, while the lower copy is pronounced in in situ questions (see Chomsky 1995). As a consequence, wh-questions, whether in situ or ex situ, are globally expected to exhibit a shared sensitivity to the locality constraints associated with movement under this view.

The unselective binding hypothesis (Pesetsky 1987; Watanabe 1992) claims that the difference between ex situ and in situ wh-questions is more profound since the dependency associated with wh-in situ is an anaphoric dependency, not a movement one. In situ wh-phrases do not move (or internally merge), but instead receive their quantificational force by being bound by an operator in the clausal periphery. Crucially, under this view, in situ wh-questions are not expected to obey the same locality constraints as ex situ wh-questions, which involve movement.

The heavy pied-piping hypothesis (Richards 2000) claims that what moves covertly is not the single wh-element, but the entire clause containing it. Wh-phrases pied-pipe their clause up to their scope position at LF. As a consequence, they are expected to escape many locality constraints (and notably a number of islands) that restrict simple wh-movement.

Bayer and Cheng (2017) combine the first two hypotheses (LF-movement and unselective binding) as parametric options depending on how scope marking works in any given language. For languages with obligatory overt scope markers, like Japanese, they propose that wh-in situ is derived without movement through unselective binding. No syntactic constraint associated with movement (e.g., islands) is observed. For languages with no overt scope marker or no obligatory scope marker, like Mandarin Chinese, wh-in situ is derived through LF-movement, and as a result, the dependency is constrained as a movement dependency.

The question we aim to address empirically here is: what about mixed languages like Colloquial French, where wh-in situ is only one strategy for questions alongside wh-ex situ?

Baunaz (2011: 75) describes in situ and ex situ structures in non-standard Colloquial French as syntactically similar, both involving movement and presupposing identical processing loads (following Adli 2006). Faure and Palasis (2021) also describe wh-in situ questions in Colloquial French as derived through covert wh-movement. They show, however, that movement is less restricted in wh-ex situ questions, as evidenced by lower sensitivity to weak islands and intervention. They attribute these differences to the presence of a different, lower, trigger in ex situ questions, called Exclusivity.

Both analyses remain purely theoretical in their method and scope.

1.2 Processing wh-questions

Studies about processing covert dependencies confirm the prediction of theoretical analyses, that posit the existence of a long-distance dependency in wh-in situ, which has a processing cost. Aoshima et al. (2004), Ueno and Kluender (2009) found in Japanese that longer wh-in situ dependencies implied more processing costs compared to shorter ones. Xiang et al. (2014, 2015) found that Chinese wh-in situ questions incurred more processing costs than their declarative counterparts. When involving one embedding, they were processed slower than with no embedding. Similar results were obtained with both simplex (‘who’) and complex (‘which’) in situ wh-questions in French (Pablos et al. 2018).

As for ex situ wh-questions, subject questions are in general easier to comprehend than object questions (see van Gompel 2013 for adults’ typical populations, Guasti 2004 for a discussion about children; Deevy and Leonard 2004; Levy and Friedmann 2009; van der Lely and Battell 2003 for children with language impairments; Sheppard et al. 2015 among others for agrammatic adults). There are various accounts for this subject advantage.

For one, the subject question in (6a) is easier to parse over the object question in (6b) because of the shorter linear distance between the wh-word and its gap.

A more structural account consists of appealing to hierarchical rather than to linear distance and suggesting that more deeply embedded constituents require more computational effort to parse than less deeply embedded ones: the subject question in (7a) is easier to parse than the object question in (7b) because of the shorter structural distance between the wh-word and its gap. These two accounts make different predictions when associated with word order variation. In particular, the latter predicts a universal subject gap that has so far proved to be correct (see Lau and Tanaka 2021 for a systematic review).

Among structural accounts, Featural Relativized Minimality (Friedmann and Rizzi 2009; Rizzi 1990) explains the subject advantage in yet a different, empirically accurate, way. According to Featural Relativized Minimality, the local relation between an extracted element and its trace is disrupted when it crosses an intervening element whose morphosyntactic featural specification matches the specification of the elements it separates. This approach naturally leads to a system that can capture degrees of processing difficulty: the relative acceptability of an intervention configuration will vary as a function of the total, partial, or zero featural overlap between the intervener and the target. In a nutshell, configurations involving less featural overlap are predicted to be easier than sentences involving a higher degree of overlap (Villata et al. 2016). Subject questions like (8a) are easy because no intervener with any overlap with the target holds in the relevant configuration.

Bare wh-object questions like (8b) are more difficult because of the partial featural identity between the object and the intervening subject: both the subject and the wh-element are DPs and as such they share the same categorial [+D] feature. Since the intervention is only partial (concerning only one feature), the sentence can still be processed, although with some additional difficulty. But the processing difficulty associated with object questions is predicted to be stronger in which-object questions as (8c), where an additional [+NP] feature is shared with the non-wh lexical subject. The pattern of acquisition of wh-questions in children appears to conform to this prediction, as among object questions, wh-questions with lexical restrictions (e.g., ‘which dog’) emerge significantly later in L1 acquisition than bare wh-questions (e.g., ‘who’) and are associated with difficulty until about age seven in monolingual development across languages (see, e.g., Avrutin 2000; Stromswold 1995; Tracy 1994).^[1] Finally, (8d), displaying a total featural overlap configuration, corresponds to ungrammaticality (a wh-island).

Going back to the subject advantage, it has also been suggested that object questions are more difficult, in SVO languages at least, because after movement the arguments in the sentence are no longer in a canonical word order (Diessel and Tomasello 2005). This makes it more difficult to infer a meaning based on surface form. Subject questions are easier presumably because there are no overt word-order differences between the interrogative and declarative forms.

What do these different accounts predict as far as wh-in situ in a mixed language like French is concerned, and how does this interact with the various approaches to wh-in situ we have discussed briefly in the preceding section?

Starting from the canonical word order approach, this predicts no subject advantage in wh-in situ questions no matter the underlying derivation, since by definition in situ questions display canonical word order across the board.

On the other hand, if wh-in situ involves covert movement, all accounts predict wh-in situ questions to display a subject advantage as it obeys the same constraint as overt movement by hypothesis. If Featural Relativized Minimality is on the right track, moreover, we further expect object ‘which’ questions to be more difficult to comprehend than object ‘who’ questions. These effects are not expected if wh-in situ involves an unselective binding dependency, since a quantifier binds all unbound variables in its scope and a full definite NP in subject position is not expected to intervene in such dependency. Faure and Palasis (2021)’s description, which associates wh-ex situ and wh-in situ to different features, would be compatible with different featural Minimality effects for the two covert dependencies.

On the empirical side, little has been done to investigate whether wh-in situ questions display a subject advantage and whether this is modulated by the type of wh-element involved. A recent study on French (Bentea 2016) found a subject advantage only in ex situ questions. Object which-ex situ questions yielded lower response accuracy scores than subject which-ex situ questions, while who-questions displayed no significant difference with respect to object ones, due to a ceiling effect. In contrast, higher accuracy scores were obtained with object in situ questions, irrespective of whether the wh-element was ‘who’ (qui) or ‘which’ (quel). French Sign Language (LSF) on the other hand have recently been shown to exhibit a subject advantage in in situ wh-questions (Hauser et al. 2023): authors found a significantly higher response accuracy in subject questions than in object questions and this effect was stronger in wh-questions with lexical restrictions (equivalent to which-questions) than in bare wh-questions (equivalent to who-questions).

The goal of our study is to verify if there is a subject-object asymmetry in the processing of French in situ wh-questions, as well as in ex situ wh-questions. If wh-in situ in French is derived through the same covert movement operation as the one deriving overt wh-questions, we expect to observe a subject advantage in the comprehension of wh-in situ questions as well as in that of wh-ex situ questions. We expect moreover to find a stronger effect in which-questions than in bare wh-questions, due to Featural Relativized Minimality effects. If Faure and Palasis (2021) are on the right track, though, and the movement path for wh-ex situ and for wh-in situ is not the same, we might also expect to see some difference between the two strategies.

We constructed two comprehension and acceptability experiments including in situ and ex situ questions, with the aim of verifying whether they display the same pattern.

2.1 The quest for the ideal paradigm

While preparing our experiment, we soon realized that constructing unambiguous wh-questions in French is not easy, as confirmed by several pilot studies involving about five native French speakers each. Our aim was to construct items that were unambiguous both for their status (in situ vs. ex situ) and for their interpretation (subject vs. object), and finally both with qui ‘who’ and with quel ‘which’. We faced two major problems in trying to build such a complex paradigm: (i) the lack of quel-questions in subject ex situ conditions and (ii) the ambiguity in the comprehension and in status of in situ subject questions.

Starting with ex situ questions, in fact, particle questions, i.e., questions with est-ce que/qui, which have the advantage of clearly marking the ex situ strategy, are defective: they are not available with quel ‘which’ subject questions:

Since French does not allow quel-subject to be used with est-ce que, we tried to provide a complete paradigm by resorting to another strategy, namely interrogatives with inversion. However, as shown by (10a), which contrasts minimally with the corresponding object condition in (10b), subject questions with inversion are not grammatical either.^[2]

The participants to our pilots clearly preferred the non-inverted version, as in Quelle dame pousse le monsieur? ‘Which lady pushes the gentleman?’ This strategy of subject questions is however ambiguous both for its derivational status and for its meaning. As for its status, it could correspond both to an in situ and an ex situ derivation.^[3] As for its interpretation, it can be interpreted both as a subject question or an object question with inverted subject, as represented in (11).

(11)

In (11b), the verb and wh-object move to the front of the subject or in other words, the subject is posited post verbally in sentences where the initial constituent (quelle dame) is extracted. This inversion, generally known as stylistic inversion (Kayne and Pollock 1978), is optional and occurs in interrogative (12a) or exclamative sentences (12b).

To avoid both types of ambiguities, we first tried to embed the interrogative under a believe-verb, as in (13):

(13)

Note that French does not allow the complementizer que ‘that’ to be followed by a gap, but it is possible to use the form qui instead (Rizzi and Shlonsky 2007; Rooryck 2000; Taraldsen 2001). This allowed us to construct a complete set of items filling the 4 conditions.

Sentences in (14c-d) display no ambiguities in interpretation nor in status. The problem is that our participants disliked this strategy when applied to qui-questions. They rated badly most of the items in the qui-paradigm:

Our participants would rather use Qui pousse le monsieur? ‘Who is pushing the gentleman?’ and Par qui le monsieur est-il poussé? ‘By whom is the gentleman pushed?’ instead of (15a) and (15c). (15b) is slightly better than the others but less acceptable than Qui est-ce que le monsieur pousse? and Qui le monsieur pousse- t-il ? ‘Who is the gentleman pushing?’.

The final decision was then to use clefted questions as in (16).

These questions have the double advantage of being clearly ex situ and displaying no ambiguity in interpretation. The status and interpretation of clefted questions is far from clear, however. Some scholars have claimed that wh-elements in clefted questions do not move as far as to their scope position (Chang 1997), and they are thus not genuine instances of wh-ex situ questions. We acknowledge this complication but decided to stick to these items in lack of better alternatives. Indeed, it is uncontroversial that the wh-element is not in its in situ position either, and that the clefted question contains a gap.^[4] We shall consider this to be the minimal defining property of any ex situ strategy. We are of course aware that clefted questions are more complex than questions with no clefting. But recall that the aim of our experiment here is to verify if there is a subject-object asymmetry to be observed within each of the strategies of interrogatives, and not to compare them.

There is some normative stigma associated with the use of double qui in c’est qui qui since qui donc, qui est-ce qui, qui c’est qui or a single qui are deemed more elegant. This is one of the reasons why we decided to develop an oral version of the experiment, assuming that oral French is closer to Colloquial French and thus less prone to normative pressures (see below).

Turning to in situ conditions, we found that simple intonation questions are still potentially ambiguous in various ways. Subject questions with simple intonation can be either in situ or ex situ, and several interpretation ambiguities also arise. A sentence like Qui lave le chat? can be interpreted not only as a subject question ‘Who is washing the cat?’ but also an object question with subject inversion ‘Who is the cat washing?’ (as in [11]). We avoided these potential ambiguities by inserting a clitic left dislocation structure (CLLD), as in (17):

In situ questions in (17a-b) are preceded by a left dislocated (i.e., topicalized) masculine NP, associated with a resumptive proclitic, and this addition makes them non-ambiguous and natural.

For the sake of uniformity, we decided to add this topicalization pattern across the board in all items. Ex situ questions (18a-b) also included topicalized masculine NPs besides being clefted to avoid the ambiguities discussed above.

The result we obtained is a homogenous set of questions, all containing a clitic left dislocation structure (CLLD), displaying no ambiguity in interpretation, and a distinction between ex situ and in situ which is again non ambiguous and capitalizes on clefting versus non-clefting.

We also eliminated verbs starting with a vowel (e.g., appeler ‘call’, aimer ‘love’, …) and only used verbs starting with a consonant (e.g., prévenir ‘inform’, désirer ‘desire’…) to avoid perceptual ambiguities in sequences such as qui l’appelle ‘who calls him’ and qu’il appelle ‘who he calls’:/i/in the latter is a little longer than in the former but this difference is very subtle and difficult to distinguish. We then avoided female referents and kept only male referents to avoid any gender bias that might play a role wherever social gender is involved: men are more likely to be in a subject position and to be referred to via pronouns (Richy and Burnett 2020).

A potential issue with this data setting is that it is particularly natural in spoken Colloquial French but may be somewhat stigmatized in written French. This is why we decided to replicate the experiment in two modalities: one being a reading task, and the other a listening task. We expect any effect observed in the results to be relatively larger in the listening task compared to the reading task.

More generally, we expect to observe a subject advantage in cleft structures (as in [18]), because they are associated with overt wh-extraction (although with the proviso discussed above). If the experiment is well-designed and sufficiently sensitive, the effect should be stronger with quel ‘which’ questions than with qui ‘who’ questions. If the data pattern is as predicted in the ex situ condition, we will be able to consider the results in the in situ condition as evidence for their processing and derivation. If in situ questions involve covert movement to the same position as wh-ex situ, the hypothesis is that they will display the same subject advantage, and with the same modulation based on wh-material. If the dependency involved is of a different type (or targeting a different position; see above) the hypothesis is that no subject advantage should hold.

In what follows, we present the two experiments only once, because the items were just the same with the only difference concerning the modality of representation. The results will be discussed separately, due to some minor differences.

2.2 Materials

In total, our experiment included 48 groups of wh-questions: 24 sets of who-questions (qui-questions) and 24 sets of which-questions (quel-questions). Each set of wh-questions followed a 2 × 2 design: (A) in situ (19a, b)/ex situ (19c, d), and (B) subject (19a, c)/object (19b, d).

The experimental task was twofold: participants had to rate each item, and then to reply to each question by selecting a character on a picture with three characters. Two of them were similar, either performing an action toward the third, distinct character standing between them, or experiencing the same action performed by the third character in the middle. An illustration is given in Figure 1. Details on the procedure are given in the next section.

Figure 1:

Picture illustrating the pushing event of (19): the woman on the right matches with the subject question in (19a, c) and the woman on the left matches with the object question in (19b, d).

Each participant saw only one item from each set and encountered each condition an equal number of times. In total, 48 items were shown to each participant. In addition to the test items, there were 14 where-questions (où-questions) included as fillers, as in (20). These filler questions always targeted the side characters, similar to the test items (see Figure 2). Four of the fillers were ungrammatical distractors (e.g., Mouton le, il où est? ‘sheep the, he where is?’) to ensure that participants were paying attention to the sentence items. These sentences were expected to be rated as 1 or 2. The data of participants who gave a high rating to these very bad sentences were excluded by the analysis.

Figure 2:

Picture corresponding to the filler where-question in (20): the princess on the right answers the question.

All items were randomized into four counterbalanced lists with the same fillers (see Appendix: https://osf.io/y9x2h). All items and fillers were recorded by a French native speaker in a soundproof room. Each sentence lasted approximately 3–4 s and was played at a volume of 20 dB.

2.3 Tasks and procedure

The test was hosted on PCibex (https://farm.pcibex.net/). It began with a consent page clarifying the purpose and procedure of the experiment,^[5] followed by a page gathering some basic information such as age, gender and native language and then proceeded to an instruction page. A short training phase, including four practice items, with no time limit, was provided before each trial, allowing participants to familiarize themselves with the experiment. Answers from the training phase were not considered in the data analysis. The duration of each testing session was approximately 15 min.

Since the experiment was conducted online, we were unable to provide face-to-face instructions to participants, who were informed that they needed to complete the task in a well-connected and quiet environment. Participants were also free to withdraw at any time due to unstable internet connections, interruptions, or fatigue.

For each trial, participants either read (if performing the reading task) or listened (if performing the listening task) to the question stimulus and had then to perform two tasks:

1. an acceptability rating task: they were asked to rate its acceptability on a 1–7 Likert scale. The time limit was 10 s.

2. a forced choice task: each question was followed by a picture from which participants had to choose a character to answer, with a response time limit of 5 s.

Acceptability rating tasks are known to have relatively higher statistical power with smaller sample sizes compared to other informal offline methods (e.g., forced-choice, magnitude estimation, yes-no tasks) (Sprouse and Almeida 2017: 28). The task was meant to allow us to establish a scale for the comprehension levels of different sentences, enabling comparisons across conditions and supporting our factorial experimental design.

The forced-choice task is known to be highly sensitive in detecting differences between two contrasting conditions (Sprouse and Almeida 2017: 25), so the task was also meant to allow us to identify differences across conditions. But response accuracy was also meant as a sanity check, to verify whether participants were properly reading and understanding the experimental items. This is why we used the combination of these two tasks.

We expected the acceptability rating task to yield more significant results than the forced-choice task, as the former is more sensitive when testing the same syntactic phenomena (Sprouse and Almeida 2017: 28). We also expected the results of the forced-choice task to align with those of the acceptability judgment, with only minor differences. Specifically, higher response accuracy should correspond to higher ratings. Given our research hypothesis, if all wh-questions in French displayed a subject advantage, participants would be expected to answer subject questions more accurately than object questions and to rate subject questions higher than object ones.

2.4 Participants

A total of 68 participants completed our reading task. Most were recruited online through RISC (Le relais d’information sur les sciences de la cognition), while two participants joined after seeing the recruitment notice on campus. Among them, two were non-native speakers of French, and four had previously participated in the experiment as pilot testers (three of these four had low response accuracy: 54.76 %, 59.57 % and 62.50 %); their results were removed from the analyses. Three participants gave high ratings (6 or 7) to ungrammatical distractors. Although we removed their ratings, we retained their answers to the other questions since their response accuracy fell within the acceptable margin of error (i.e., greater than the mean minus two standard deviations, which is 67.49 %). The remaining valid participants were all native French speakers,^[6] consisting of 20 men and 42 women. Each participant received a €3 gift card as a reward. The average age was 28.27 years (range: 19 to 45).

A total of 62 participants consisting of 18 men and 44 women completed our listening task. Most were recruited online through RISC, Facebook, and SurveyCircle. Two participants were excluded due to low response accuracy (56.10 % and 60.42 %). Additionally, four participants gave high ratings (6 or 7) to ungrammatical distractors. Although we removed their ratings, we retained their answers to the wh-questions since their response accuracy fell within the acceptable margin of error (i.e., greater than the mean minus two standard deviations, which is 72.19 %). Each participant received a €3 gift card as a reward. The average age was 30.58 years (range: 18–66).

2.5 Results

2.5.1 The reading task

Results for who-questions (qui-questions) and which-questions (quel-questions) were analyzed separately. Ten sentence items (within a given list) with mean response accuracy smaller than the overall mean by more than two standard deviations (SD) (i.e., smaller than 64.56 %) were removed. The results were analyzed using R Studio. Response accuracy was analyzed using generalized linear mixed-effects models with glmer() from the lme4 package (Bates et al. 2015), while acceptability judgments were analyzed using cumulative link models with clm() from the ordinal package (Christensen 2019). Log-transformed response times log(RT) were analyzed using linear mixed-effects models with lmer() from the lmerTest package (Kuznetsova et al. 2017).^[7]

Random effects included intercepts for subjects and items. The fixed factors were question type (subject question vs. object question) and the position of the wh-element (in situ vs. ex situ). All effects with a p-value less than 0.05 were considered significant.

Starting with mean response accuracy, Figure 3 shows a significant interaction between the two main factors (subject/object and in situ/ex situ) for qui-questions (β = 0.6029, z = −2.498, p < 0.05), reflecting a significant subject-object asymmetry in the ex situ condition only (β = 0.2990, z = 3.352, p < 0.001). This indicates that there were more correct answers for subject questions than for object questions (i.e., subject advantage). For quel-questions, there is also a significant interaction between the two main factors (β = 0.5210, z = −2.175, p < 0.05). In this case, a subject-object asymmetry appears only in the in situ condition (β = 0.4017, z = −1.997, p < 0.05), with object questions answered more accurately than subject questions (i.e., object advantage).

Figure 3:

Mean accuracy rates (%) for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (blue) and object (pink) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

The mean ratings of qui/quel-questions are presented in Figure 4.

Figure 4:

Mean ratings for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (blue) and object (orange) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

Figure 4 shows a slight but not significant interaction between subject/object and in situ/ex situ conditions for qui-questions (β = 0.19392, z = −1.647, p < 0.1). A subject-object asymmetry is present, but only in the in situ condition (β = 0.1380, z = −2.867, p < 0.01). This interaction is significant for quel-questions (β = 0.195204, z = −4.060, p < 0.001), reflecting a subject-object asymmetry also in the in situ condition (β = 0.1405, z = −5.608, p < 0.001). The mean ratings for object questions are significantly higher than those for subject questions (i.e., object advantage).

The mean response time to qui/quel-questions are presented in Figure 5.

Figure 5:

Mean response time (log milliseconds) for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (green) and object (yellow) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

In Figure 5, there is no interaction between subject/object and in situ/ex situ conditions in either case (qui: β = 2.548e-02, t = 0.316, ns; quel: β = 0.023468, t = 0.257, ns). We observe no subject-object asymmetry in qui-questions (β = 1.808e-02, t = −0.028, ns) or in quel-questions (β = 0.016499, t = 1.355, ns). For qui-questions, there is a slight but not significant difference between in situ and ex situ conditions (β = 1.790e-02, t = −1.734, p < 0.1). This difference is not present in quel-questions (β = 0.016526, t = −1.012, ns).

The mean reading time to qui/quel-questions is presented in Figure 6.

Figure 6:

Mean reading time (log milliseconds) for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (green) and object (yellow) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

Figure 6 shows an interaction between subject/object and in situ/ex situ conditions for qui-questions (β = 0.01708, t = 4.211, p < 0.001), while this interaction is not significant for quel-questions (β = 0.01730, t = 1.696, ns). For qui-questions, subject questions are read faster than object questions in ex situ conditions (β = 0.009898, t = −4.545, p < 0.001), indicating a subject advantage. There is no significant difference in in situ conditions (β = 0.01370, t = 1.853, ns). For both types of questions, the mean reading time for ex situ questions is significantly longer than for in situ questions (qui: β = 0.01209, t = −6.360, p < 0.001; quel: β = 0.01221, t = −4.707, p < 0.001). There is no significant subject-object asymmetry within any condition.

2.5.2 The listening task

Results for whom-questions (qui-questions) and which-questions (quel-questions) were analyzed separately. Seventeen items with mean response accuracy smaller than the overall mean by more than two standard deviations (SD) (i.e., smaller than 64.16 %) were removed. The models used are the same as those in Section 2.5.1 of the reading task.

Figure 7 represents the mean response accuracy. There is an interaction between the two main factors (subject/object and in situ/ex situ) in both cases (qui: β = 0.9074, z = −2.913, p < 0.001; quel: β = 0.5946, z = −2.615, p < 0.001). The difference between subject and object questions in the ex situ condition is significant. It shows a subject advantage in both types of wh-questions (qui: β = 0.5256, z = 3.366, p < 0.001; quel: β = 0.3268, z = 4.375, p < 0.001). This means that there are more correct answers in subject questions than in object questions (i.e., subject advantage). No significant asymmetry shows in in situ conditions.

Figure 7:

Mean accuracy rates (%) for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (blue) and object (pink) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

The mean ratings of qui/quel-questions are presented in Figure 8.

Figure 8:

Mean ratings for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (blue) and object (orange) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

Figure 8 shows almost no interaction between subject/object and in situ/ex situ conditions for qui-questions (β = 0.21599, z = −0.662, ns) and for quel-questions (β = 0.21790, z = −1.817, p < 0.1). However, there is an object advantage in quel-questions, but only in in situ conditions (β = 0.1564, z = −3.111, p < 0.01). The mean ratings for object questions are significantly higher than those for subject questions (object advantage).

The mean response time to qui/quel-questions are presented in Figure 9.

Figure 9:

Mean response time (log milliseconds) for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (green) and object (yellow) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

In Figure 9, there is no interaction between subject/object and in situ/ex situ conditions in either case (qui: β = 0.025017, t = 0.048, ns; quel: β = 0.024767, t = 1.067, ns). This indicates that there is no subject-object asymmetry in qui-questions (β = 0.018871, t = 1.593, ns), nor in quel-questions (β = 0.017719, t = −0.156, ns). For quel-questions, there is a significant difference between in situ and ex situ conditions (β = 0.017708, t = −2.596, p < 0.05). Ex situ questions have longer response times, indicating that they are relatively more difficult to process. This difference is not present in qui-questions (β = 0.018843, t = −0.537, ns).

The mean listening time to qui/quel-questions is presented in Figure 10.

Figure 10:

Mean listening time (log milliseconds) for qui-questions (left panel) and quel-questions (right panel) for in situ (right in each panel) and ex situ (left) questions across subject (green) and object (yellow) conditions. Error bars represent standard errors of the mean and stars are used to indicate statistically significant results.

Figure 10 shows no interaction between subject/object and in situ/ex situ conditions for qui-questions (β = 0.011159, t = 0.206, ns) and quel-questions (β = 0.010794, t = 0.395, ns), indicating no subject-object asymmetry in qui-questions (β = 0.007896, t = 0.851, ns) or quel-questions (β = 0.007612, t = −0.284, ns). Additionally, there is no difference between in situ and ex situ conditions (for qui: β = 0.007886, t = −0.525, ns; for quel: β = 0.007624, t = 0.633, ns).