Spanish embedded question island effects revisited: an experimental study

1.1 Bounding and EQ-islands in Spanish

Bounding approaches to island effects (Chomsky 1973, 1977, 1986, 2000) are based on the idea that movement must be local. For instance, the principle of Subjacency (Chomsky 1973) bars movement in one fell swoop across more than one bounding node, i.e., S (currently TP) or NP/DP in English. Movement across several bounding nodes is only possible if it proceeds successive-cyclically, making intermediate stops at unfilled spec,CP positions. From this perspective, island effects arise because extractees are forced to cross more than one bounding node without making intermediate stops. This occurs in English wh-questions with extraction from EQs because the wh-word crosses two TPs and the intermediate spec, CP position is already filled by the wh-phrase or operator introducing the EQ (see Figure 1).

Figure 1:

Structure of an English wh-question with extraction from EQs. Bounding nodes are circled.

Sentences like (3) motivated the idea that EQ-island sensitivity could vary across languages (Suñer 1991; Torrego 1984). This variation was squared with Subjacency in languages like Spanish via the proposal that allowed bounding nodes to vary parametrically (Rizzi 1982; Torrego 1984). If S′ (currently CP) is a bounding node in Spanish, as opposed to S (TP) in English, A′-movement from EQs would only cross a single bounding node. Such movement would comply with the Subjacency condition.

As discussed above Torrego (1984) and Suñer (1991) report that some Spanish EQs nevertheless disallow extraction: FGDs into EQs under rogative verbs like ask or wonder are supposedly unacceptable (5), (6). Suñer (1991) explains the supposed islandhood of rogative-EQs as a Bounding/Subjacency effect: EQs embedded under rogative verbs contain additional functional structure in their left periphery, which blocks successive-cyclic movement out (Figure 2). Specifically, Suñer argues that EQs under rogative verbs have an additional CP layer (labeled CP in Figure 2).

Figure 2:

Structure of Spanish embedded questions under responsive and rogative verbs, according to Suñer (1991). The bounding nodes crossed by an extracted wh-word are circled.

Evidence for this additional CP layer comes from the fact that Spanish questions embedded under rogative verbs can be preceded by the complementizer que ‘that’ in (7), whereas questions embedded under responsive verbs cannot as in (8).

In EQs under rogative verbs, both spec,CP positions are assumed to be filled: the lower spec,CP contains the wh-phrase/operator introducing the EQ, and the higher spec,CP contains a null question operator, which Suñer assumes is required to mark the double CP structure as an indirect question. Moving out of a rogative EQ should therefore cross two CPs. Since CP is considered a bounding node in Spanish, extraction from these EQs violates Bounding constraints.

Thus, according to Suñer (1991), EQs under rogative verbs are structural/syntactic islands, while EQs under responsive verbs are not. This predicts that extraction from EQs under rogative verbs should yield structural island effects, typically associated with stark unacceptability, whereas extraction from EQs under responsive verbs should yield no structural island effects and should obtain acceptable ratings. Her judgments suggest that this is the case in (4), (5), (6), but this contrast has never been tested experimentally.

1.2 Recent experimental work

Two recent experimental studies tested extraction from Spanish si ‘whether’ EQs embedded under the verb preguntar(se) ‘to ask/wonder’ (López-Sancio 2015; Pañeda et al. 2020). Both studies compared the acceptability of sentences with FGDs in a 2 × 2 factorial design (discussed in greater detail below). An example set of test sentences (taken from López-Sancio 2015) is in (9).

Both studies found whether EQ-island effects: Participants rejected (9d) at significantly higher rates than the other sentences in the paradigm – and at rates analogous to other less controversial syntactic island violations such as extractions from adjuncts.^[2]

The results of López-Sancio (2015) and Pañeda et al. (2020) are compatible with the claim that EQs under rogative verbs are syntactic islands, though they do not establish that island effects are verb-dependent, since they did not compare responsive and rogative verbs.

Further, these studies do not address the source of the island effects observed. The results are compatible with a syntactic source, e.g., they could result from a syntactic Bounding constraint violation. They are alternatively compatible with extra-syntactic sources of unacceptability, such as semantic or pragmatic constraint violations or processing difficulty.

1.3 The present study

In this paper we test for effects of verb choice, in order to determine whether extraction from Spanish EQs is influenced by the type of matrix verb (ask vs. know). We carried out two acceptability judgment experiments. Experiment 1 tests extraction from si (‘whether’) EQs, in order to follow up on previous experimental findings that extraction from whether EQs under rogative verbs is unacceptable (López-Sancio 2015; Pañeda et al. 2020). Experiment 2 examines whether the results extend to a different interrogative, cuándo (‘when’).

To assess whether effects of verb choice are likely to reflect a structural difference, we minimize the influence of non-structural factors on acceptability by presenting test sentences in a context designed to make them semantically and pragmatically motivated, and by reducing processing difficulty as much as possible. It has been proposed that sentences that are rejected due to non-structural constraints become more acceptable in context, whereas sentences which are rejected due to structural constraints do not (Erteschik-Shir 2006; Keller 2000; Sorace and Keller 2005).

With the same aim, we also examine judgment distributions. Kush et al. (2018, 2019 have recently suggested that judgment distributions can provide useful information about whether a domain is necessarily a structural island. The authors observed that small island effects can arise from variable/inconsistent judgments across trials. Kush and colleagues argue that if participants accept, on a significant proportion of trials, sentences with dependencies out of a supposed island domain, X, then one can minimally conclude that there exists a non-island structural analysis of domain X. Residual variability in judgments on such trials could either reflect that participants occasionally fail to adopt the non-island analysis of X or that extra-syntactic factors conspire to reduce acceptability. Applying this logic to EQs, we make the following predictions: If EQs under rogative verbs are necessarily analyzed as structural islands, extraction from EQs under rogative verbs should be consistently judged unacceptable and we expect to see large island effects. If, on the other hand, EQs under rogative verbs are not structural islands we either expect to see (i) no island effects at all, or (ii) smaller island effects characterized by inconsistent judgment distributions. In the last case we could conclude that there is a non-island analysis of EQs under rogative verbs (i.e., there is not a categorical syntactic ban on extraction).

As a control, we test extraction from adjuncts and relative clauses under the same conditions. Adjunct and relative clause island effects are generally attributed to a structural constraint violation (e.g., Chomsky 1973, 1977, 1986, 2001; Huang 1982; Nunes and Uriagereka 2002), so we expect them to yield unacceptability and large island effects. EQs should show similar effects if they are necessarily analyzed as structural islands, but they should behave differently if they are not.

1.4 General experimental design: the factorial definition of islands

Our experiments adopt a version of the experimental paradigm introduced by Sprouse et al. (2011) and which has subsequently been used to collect formal acceptability judgments of island sentences in various languages (Sprouse et al. 2012, 2016; Aldosari 2015; Almeida 2014; Keshev and Meltzer-Asscher 2019; Kush et al. 2019, 2018; López-Sancio 2015; Michel 2014; Ortega-Santos et al. 2018; Pañeda et al. 2020; Stepanov et al. 2018; Tucker et al. 2019). This design isolates island effects from two factors that can affect acceptability, namely the distance between the filler and the gap (short/long) and the type of structure where the gap is located (non-island/island). An example from Sprouse et al. (2016) is shown in (10).

In the short conditions (10a), (10b), the gap is in the matrix sentence, whereas in the long conditions (10c), (10d), the gap is in the embedded clause. In the non-island conditions (10a), (10c), the embedded structure is a declarative clause, which does not give rise to island effects, whereas in the island conditions (10b), (10d), it is an EQ, expected to yield island effects. Sentences with extraction from an island are represented by condition island/long (10d), in which a long-distance dependency is established inside an EQ.

The design in (10) quantifies the independent effects of a long-distance dependency and an EQ, and it can thus be used to assess to what extent these two factors contribute to the (un)-acceptability of the island/long condition. If the island/long condition is as (un)-acceptable as predicted by the sum of the effects of the two factors, there is no evidence that EQs yield island effects, understood as specific effects caused by establishing a long-distance dependency inside an EQ. However, if the island/long condition is less acceptable than predicted by the addition of the effects of distance and the EQ structure (i.e., if there is “super-additivity”), this is taken as an indication that there are island effects, thus supporting the need to assume additional constraints (syntactic or otherwise) to explain them. An island effect is represented statistically by an interaction between distance and structure. Island effect sizes can be computed with differences-in-differences (DD) scores as shown in (10).^[3]

In our experiments, we deployed a 3 × 2 version of the design in (10) to address the impact of the type of matrix verb. Following most previous work on Spanish and English, our materials involved object extraction and a wh-question configuration.

2.1 Method

2.2 Results

The grammatical fillers were rated in the acceptable range (z = 0.80, SD: 0.36) and the ungrammatical fillers were rated in the unacceptable range (z = −1.08, SD: 0.56), showing that participants understood the task and carried it out as expected.

Acceptability by condition and DD scores in the adjunct, relative clause and EQ-islands are shown in Figure 3. Statistical results are reported in Table 1.

Figure 3:

Acceptability by condition in the adjunct, relative clause and embedded question (EQ) islands of Experiment 1. Error bars show 95% confidence intervals. Non-parallel lines are indicative of an island effect.

We found significant super-additive Distance × Structure interactions both in adjunct (β = −1.521, SE = 0.104, p < 0.001) and relative clause islands (β = −1.827, SE = 0.080, p < 0.001), consistent with island effects. In both cases, island effects were large (adjunct islands: DD = 1.46; relative clause islands: DD = 1.82) and characterized by unacceptability in the island/long conditions (adjunct islands: −0.95; relative clause islands: −1.34).

EQs also yielded a super-additive Distance × Structure interaction, indicating that extraction from them was less acceptable on average than extraction from embedded declarative clauses (β = −0.599, SE = 0.103, p < 0.001). However, these island effects were smaller than adjunct and relative clause island effects (know-islands: DD = 0.22, ask-islands: DD = 0.38). The smaller interaction effects reflect the fact that ratings of the island/long conditions were well above z = 0, the midpoint of the scale, arguably in the acceptable range (know-island/long: 0.55, ask-island/long: 0.40). These ratings were lower than ratings to the other (uncontroversially grammatical) EQ-island conditions, but similar to ratings to the (also grammatical) adjunct and relative clause non-island/long conditions (adjunct: 0.50, relative clause: 0.48). Regarding the effects of verb choice, the interaction was significantly greater with ask than with know (β = −0.159, SE = 0.059, p = 0.008), indicating that island effects were greater in the former case. However, the verb did not cause a qualitative difference in acceptability.

The DD scores and condition means thus indicate that extraction from EQs is less acceptable on average than extraction from matched declarative clauses, but more acceptable than extraction from adjuncts and relative clauses. This could indicate that EQ-island effects are “subliminal” in Spanish, i.e., consistently mild and in the acceptable range (Almeida 2014). However, aggregate means alone are not enough to determine whether an effect is consistently mild, as they can either reflect a central tendency in judgments or obscure a pattern of variability that is incompatible with a “subliminal” interpretation (Kush et al. 2018, 2019). To see whether the acceptability means in the two EQ-island/long conditions reflect a central tendency in judgments, we plotted the distribution of z-scores in these conditions and compared it to the distribution of ratings in (i) the uncontroversially grammatical non-island/long conditions for all three islands and (ii) the adjunct and relative clause island/long conditions. Comparison plots are in Figure 4.

Figure 4:

Distribution of z-scores in the non-island/long and island/long conditions of Experiment 1.

Figure 4 shows that filler-gap dependencies into EQs were rated acceptable on most trials. The distribution of ratings in both EQ island/long conditions is similar to the distribution of ratings in the EQ non-island/long condition: most judgments cluster around the highest scores (close to z = 1). The two EQ-island/long conditions differ from the non-island/long condition in that there was a higher probability of a lower rating or rejection. Thus, their slightly lower mean acceptability seems to reflect variability in judgments rather than a consistent preference for lower/intermediate ratings. Note also that the EQ island/long conditions patterned very similarly to the grammatical adjunct and relative clause non-island/long conditions, which obtained comparable acceptability means, and they differed strongly from the corresponding ungrammatical island/long conditions, in which judgments clustered around the lowest scores.

To determine whether the variability in the ratings of the EQ island/long conditions was associated to differences between or within participants, we plotted each participant’s highest z-score against her/his lowest z-score in the island/long and non-island/long conditions in Figure 5 (cf. Bondevik et al. 2020; Kush and Dahl 2020; Kush et al. 2018, 2019).

Figure 5:

Highest against lowest z-score by participant in the island/long and non-island/long conditions of Experiment 1.

Note that there were three ratings by participant and condition, but for ease of visualization we only show the highest and the lowest rating, as these already comprise the maximum range of variability within individuals. Assuming that z-scores above 0 are acceptable and z-scores below 0 are unacceptable, a participant with both z-scores above 0 (upper right quadrant) is a consistent accepter, a participant with both z-scores below 0 (bottom left quadrant) is a consistent rejecter, and a participant with one z-score above and the other below 0 (upper left quadrant) is an inconsistent rater. In the EQ island/long conditions, most individuals were consistent accepters, and almost all others were inconsistent. In this, EQ island/long differs from its non-island/long counterpart, consistently accepted by all participants, but it resembles the non-island/long conditions of the adjunct and relative clause subdesigns, which were also rated inconsistently by some individuals. Note also that in the EQ island/long conditions there were almost no consistent rejecters, in contrast to their adjunct and relative clause counterparts, in which consistent rejection was the norm.

Finally, to explore the significant difference in island effects with ask and know, in Figure 6 we plotted the relationship between each participant’s mean DD score with know and their mean DD score with ask.

Figure 6:

Individual differences-in-differences (DD) scores in know whether embedded question islands against individual DD scores in ask whether embedded question islands in Experiment 1. Each circle represents a participant. A DD score of 0 or below indicates no super-additive island effect. DD scores greater than 0 are indicative of super-additive island effects.

Participants on the diagonal line had the same DD score with both types of EQs. Participants on the left side of the diagonal line had a greater DD score with ask, and participants on its right side had a greater DD score with know. The plot shows that many individuals had greater island effects with ask, although some displayed a different pattern, with greater know island effects or the same effect in both constructions.

2.3 Discussion

In Experiment 1, extraction from si ‘whether’ EQs yielded small island effects. Sentences obtained mean ratings in the acceptable range and mostly high scores, like grammatical control conditions. In contrast, extraction from adjunct and relative clause islands – which we take to be ungrammatical – yielded larger island effects, general unacceptability and mostly low ratings. Because EQs patterned more like grammatical conditions than like ungrammatical controls, we suggest that whether-EQs, on the whole, are not syntactic islands in Spanish.

The small whether EQ-island effect seems inconsistent with the notion of “subliminal” island effects (Almeida 2014), since the effect was not consistently mild across trials and speakers. Rather, it was caused by a combination of high and a few low ratings. Most participants accepted extraction from whether EQs consistently, and thus it appears that for many speakers, extraction from whether EQs is (almost) as acceptable as extraction from declarative clauses. Low ratings came by and large from inconsistent participants. We think this suggests that some individuals’ judgments were particularly affected by factors unrelated to a syntactic ban on extraction, as there was a similar rate of inconsistent participants in the adjunct non-island/long condition, in which extraction is syntactically allowed.

Island effects were numerically larger with the matrix verb ask than with know, indicating that know sentences had a higher acceptance rate. This effect was observable in most participants. Even so, the matrix verb did not cause any qualitative differences, as extraction was acceptable in both cases and ratings were distributed in a very similar way with both verbs.

The preponderance of acceptable ratings in both EQ-island conditions suggests that neither EQ is a syntactic island in Spanish, contrary to the analysis in Suñer (1991), where it was proposed that ask-EQs were syntactic islands and know-EQs were not. We note, however, that the effect goes in the direction predicted by Suñer, suggesting that ask may introduce some additional factor that increases the probability of mild degradation.

Our whether-island results differ strongly from those of López-Sancio (2015) and Pañeda et al. (2020), as these studies found large island effects and unacceptability in sentences with extraction from these clauses. They also differ from findings recently reported by Rodríguez and Goodall (2020), who similarly obtained large whether-island effects in Spanish. We attribute this difference to the use of context and complex fillers in our materials, which we included to motivate the sentences pragmatically and reduce the burden of several factors on acceptability. Note that the control adjunct and relative clause island sentences were strongly unacceptable even though they were presented under the same conditions, indicating that context and complex fillers do not make sentences acceptable when they are – by assumption – ungrammatical (cf. Erteschik-Shir 2006; Keller 2000; Sorace and Keller 2005). In our view, the acceptability contrast between the experimental sentences and the control sentences indicates that there is no syntactic constraint on extraction from whether-EQs in Spanish. Thus, the large island effects in previous studies may have reflected semantic, pragmatic, and/or processing factors.

As an anonymous reviewer notes, our results are compatible with the classic observation that embedded questions are weak/selective islands, i.e., islands that allow extraction of D(iscourse)-linked arguments, but not of non-D-linked arguments or adjuncts (Cinque 1990; Rizzi 1990). This is particularly so considering the large island effects and unacceptability that previous studies obtained with bare fillers, which were presumably non-D-linked. We can only agree on this observation. However, we would like to add that, in our view, the fact that these sentences are more acceptable when the wh-filler is D-linked is in itself an indication that they are not syntactic violations, as D-linking can be considered a sort of contextualization and should thus have similar effects as context. We are aware that the acceptability contrast between island sentences with and without D-linking has been given syntactic explanations: for instance, Cinque (1990) and Rizzi (1990) have argued that filler-gap dependencies inside weak islands are syntactically licit with D-linked/referential fillers, but syntactically ill-formed with non-D-linked/non-referential fillers (Cinque 1990; Rizzi 1990). However, such explanations seem ad hoc, as there are no obvious syntactic differences between island sentences with and without D-linking. Thus, we believe that, unless evidence for a syntactic difference is provided, it is more parsimonious to assume that all weak island sentences have the same syntactic status and that acceptability differences related to D-linking have extra-syntactic causes.

3.1 Method

3.2 Results

The grammatical fillers were rated in the acceptable range (0.95, SD: 0.29) and the ungrammatical fillers were rated in the unacceptable range (−0.94, SD: 0.54), showing that participants understood the task and carried it out as expected.

Acceptability by condition in the adjunct, relative clause and when-islands is shown in Figure 7. Statistical results are reported in Table 2.

Figure 7:

Acceptability by condition in the adjunct, relative clause and embedded question (EQ) islands of Experiment 2. Error bars show 95% confidence intervals. Non-parallel lines are indicative of an island effect.

Similar to Experiment 1, we found large, significant super-additive Distance × Structure interactions in the control conditions: the adjunct islands (β = −1.775, SE = 0.076, p < 0.001, DD: 1.78, z-score: −1.03) and relative clause islands (β = −1.795, SE = 0.082, p < 0.001, DD: 1.79, z-score: −1.21).

In the when-EQs, we also obtained a super-additive Distance × Structure interaction (β = −2.477, SE = 0.137, p < 0.001), but island effects were smaller than in adjunct and relative clause constructions (DDs: 1.09 with know and 1.39 with ask). The when-EQ island/long conditions obtained higher ratings than their adjunct and relative clause counterparts. However, in contrast to Experiment 1, these were below 0 and thus arguably unacceptable (know-island/long: −0.23, ask-island/long: −0.61). As in Experiment 1, the interaction was significantly greater with ask than with know (β = −0.306, SE = 0.079, p < 0.001), indicating that island effects were greater under ask.

In order to see the pattern of judgments that underlies the aggregate means, in Figure 8 we plotted the distribution of z-scores in the non-island/long and island/long conditions.

Figure 8:

Distribution of z-scores in the non-island/long and island/long conditions of Experiment 2.

The two when EQ-island/long conditions pattern differently from all the non-island/long conditions: judgments do not cluster at the high end of the scale. Judgment distributions for both EQ-island conditions exhibit bimodality, with judgments falling on both ends of the scale, though in differing proportions. With the verb ask, most judgments cluster around the low end of the scale, although there are a small number of judgments above 0. With the verb know, ratings of when EQ-island/long sentences in the two modes are comparable: reflecting an equal proportion of low and high ratings.

Figure 9 shows each participant’s highest z-score against her/his lowest z-score in the same two conditions.

Figure 9:

Highest against lowest z-score by participant in the non-island/long and island/long conditions of Experiment 2.

Here, the ask and know when island/long conditions were similar: most participants were either consistent rejecters or inconsistent raters, though there were also a few consistent accepters. The when-island/long conditions differ from their non-island counterpart, which was consistently accepted by all participants. They also contrast with the other non-island/long conditions, which were mostly consistently rated. The when-island/long conditions further differ from the adjunct and relative clause island/long conditions, which were consistently rejected by almost all speakers.

Finally, Figure 10 shows the relationship between individual DD scores with know and ask islands. As in Experiment 1, most participants are on the left of the diagonal line, indicating that they exhibited greater island effects with ask, but the effect does not hold consistently for all individuals.

Figure 10:

Individual differences-in-differences (DD) scores in know when embedded question islands against individual DD scores in ask when embedded question islands in Experiment 2. Each circle represents a participant. A DD score of 0 indicates no island effect. DD scores greater than 0 are indicative of island effects.

3.3 Discussion

Replicating the results of Experiment 1, in Experiment 2 island effects were greater and caused more degradation with ask than with know, but the verb did not yield a qualitative difference in acceptability. The contrast between ask and know was caused by a higher number of rejections with ask, and could be seen in most participants, suggesting again that ask introduces an additional constraint that increases degradation.

Extraction from cuándo (interrogative ‘when’) EQs yielded island effects and mean unacceptable ratings. When-EQs differed from adjunct and relative clause islands in that they yielded smaller island effects. Underlying the smaller EQ-island effects was a rather variable rating distribution. Judgments of EQ-islands exhibited bimodality, suggesting globally inconsistent judgments of either acceptance or rejection. There were low scores and consistent rejecters, as in ungrammatical conditions, some high scores and consistent accepters, as in grammatical conditions, and many inconsistent participants. The fact that there was a significant number of high scores suggests that there is at least one structural analysis under which when EQs are not islands. The cases of rejection may reflect (i) the impact of non-structural factors on acceptability, like semantic-pragmatic infelicity or processing difficulty, or (ii) an alternative analysis under which when-EQs are structural islands.

For instance, Torrego (1984) claims that EQs with subject-verb inversion are islands even if the extraction observes bounding constraints. If Torrego’s observation is correct, participants would be expected to have rejected extraction from EQs they analyzed as having undergone inversion. Our EQs did not contain any overt subjects, so our stimuli were, in principle, compatible with both an inversion and a non-inversion analysis. It is possible that participants rejected extraction from when-EQs that they analyzed as having undergone inversion. Torrego suggests that when forces inversion when the predicate is not “heavy”, but in that case, we would have expected participants to uniformly reject extraction from when-EQs. The fact that we saw variation in acceptance could mean that individual participants applied inversion inconsistently across trials, or that there is inter-speaker variation in whether when triggers inversion. This is an empirical question yet to be investigated.

4.1 Matrix verb effects

Our findings from Experiment 1 (on whether islands) argue against Suñer’s (1991) syntactic analysis of selective verb-driven island effects. If extraction from an EQ under a rogative verb resulted in a syntactic constraint violation, as claimed by Suñer, participants should have rejected wh-extraction from ask-EQs and we should have observed larger island effects. Instead, extraction from ask-EQs obtained average ratings in the acceptable range and yielded small effects. Seventy seven percent of trials obtained ratings of 5 or above on a 7 pt scale (compared to 14% in adjuncts and 1% in relative clauses), and 64% obtained z-scores of 0.5 or above (compared to 10% in adjuncts and 1% in relative clauses). Furthermore, most participants accepted these sentences consistently, and there were very few consistent rejecters, at rates comparable to those of consistent rejecters in uncontroversially grammatical sentences. This suggests that extraction from ask-EQs is syntactically allowed, like extraction from know-EQs, which was rated in a similar way.

In Experiment 2, extraction from ask-EQs had a lower acceptance rate, which fell below the midpoint of the scale. Still, acceptability was higher than in the control islands (27% of trials obtained ratings of 5 or above, compared to 10% in adjuncts and 4% in relative clauses, and 16% had z-scores of 0.5 or above, compared to 3% in adjuncts and 2% in relative clauses). Furthermore, just like in Experiment 1, ask-EQ islands patterned more with know-EQ islands than with the control islands, both in terms of average acceptability and in the rating distribution, which was variable. Because of all this, we believe that our results provide strong evidence against the hypothesis that the structural islandhood of Spanish EQs depends on the type of matrix verb.^[5]

Even so, island effects were quantitatively greater with ask than with know in both experiments. Thus, it seems that the rogative/responsive distinction does have a (non-categorical) effect on acceptability. It remains to be determined what to attribute this distinction to. One possibility is that the difference in the acceptability of extraction from ask and know questions has a semantic cause. Indeed, questions embedded under each type of verb have different readings: they are interpreted as “real” indirect questions under ask, but as propositions under know. In relation to this, ask-EQs may contain a question feature that their know counterparts lack. This feature may cause mild intervention effects or interfere with the retrieval of the filler at the gap position (see de Cuba and MacDonald 2013 for a related proposal).

Another possibility within the domain of semantic approaches to island effects comes from Abrusán (2014). Abrusán notes that different embedding predicates induce stronger wh-island effects depending on how difficult the lexical semantics of the predicate makes satisfying a maximal informativity requirement on wh-questions. It is possible, under certain conditions, to meet the requirement with verbs such as know (Abrusán 2014: 139–143), but the conditions under which questions with ask satisfy the requirement may be more narrow or exacting. We leave working this out to future research.

Note also that, unlike know-EQs, ask-EQs may be interpreted as uttered questions or “question acts” (Krifka 2001; Lahiri 2002). This possibility was favored in our study by the context, which presented the questions as reported speech. Question acts have been argued to have a simplified Boolean algebra that only allows some types of quantifiers to scope out of them (Krifka 2001). This may interact with island effects, as these have also been considered a scope phenomenon related to a simplified Boolean algebra (Szabolcsi and Zwarts 1993). Future research should address whether extraction from embedded question acts is expected to be less acceptable than extraction from other types of EQs on account of their semantic differences.^[6]

Our findings are further relevant for the interpretation of the results from previous studies on Spanish islands. Both López-Sancio (2015) and Pañeda et al. (2020) found large island effects in whether questions embedded under the verb preguntar(se) ‘to ask/to wonder’. Since the effects of this verb on extraction had not been directly assessed, it was unclear to what extent it could have influenced the results. Our results suggest that even if it may have indeed increased the size of the effects, the fundamental cause for these large island effects must lie elsewhere.

4.2 Whether vs. when embedded questions

While the presence of EQ-island effects in Spanish does not seem to be strongly determined by the type of embedding verb, the type of interrogative word introducing the EQ seems to have a larger impact: si ‘whether’ island effects were small, obtained mean ratings in the acceptable range and yielded relatively little variation, whereas cuándo ‘when’ island effects were greater, yielded mean unacceptable ratings and much more variation. This difference suggests that the picture of extraction from Spanish EQs is more complex than commonly thought.

Note, though, that under classical bounding approaches (Chomsky 1973, 1977), this contrast does not reflect a bounding difference. From this perspective, EQ island effects arise because wh-movement is forced to cross more than one (TP) bounding node in a single movement. This occurs because the intervening specifier,CP “escape hatch” is occupied by the embedded wh-phrase. If Spanish chooses CP as its left-peripheral bounding node (Rizzi 1982; Torrego 1984) however, direct movement across the filled intermediate specifier,CP should not result in a Subjacency violation. If indeed CP is a bounding node in Spanish, then extraction from all EQs should be syntactically licit.

An open question is whether a bounding explanation for the difference between whether and when-EQs is possible under more modern assumptions, for instance under the framework of Phase Theory (e.g., Chomsky 2000, 2001, 2008; see also Citko 2013). Phase Theory proposes that sentences are derived by phases and that once a phase is completed, part of it undergoes Spell-Out, becoming inaccessible for operations such as movement. Which categories count as phases is a matter of discussion. It is often argued that CPs are phases, but CPs are also assumed to have a rich internal structure, with multiple phrasal projections (Rizzi 1997, 2001). This opens up the possibility that only some CP projections are phases. In this context, the contrast between whether and when-EQs could be explained provided when occupies a phasal projection and whether does not. It has been suggested that si ‘whether’ occupies an Int(errogative) projection within the CP, whereas wh-words such as when occupy a Foc(us) projection (Rizzi 2001; see also Hernanz Carbó 2012). Given this, it could be argued that Foc is a phase head, while Int is not (see Yoshimoto 2012 for a related proposal). However, we are unaware of empirical and conceptual arguments that support this assumption. Assuming that the outermost specifier of a domain is the phase head (Bošković 2016), Int seems more likely to be a phase than Foc, as it is in a higher projection according to Rizzi (2001). This seems to predict that extraction from whether EQs should be more degraded than extraction from when EQs, against our findings.

We have argued that a syntactic explanation in terms of bounding is not possible, at least not under the common understanding of this notion. It remains to be determined whether the contrast between whether and when-EQs could be accounted for within syntax by other proposals. For instance, if, as argued by Torrego (1984), EQs with subject-verb inversion disallow extraction (see Section 3.3), the contrast between whether and when EQs might reflect differences in the extent to which each interrogative forces inversion.^[7] It remains to be established why inversion should interact with islandhood, as Torrego originally accounted for it by means of the Empty Category Principle (Chomsky 1981), which is not part of the theoretical apparatus anymore.

Another syntactic proposal that could account for the differences is featural Relativized Minimality (fRM; Belletti et al. 2012; Friedmann et al. 2009; Rizzi 1990, 2013; Villata et al. 2016). This theory explains EQ-island effects as intervention effects caused by the similarity between the wh-filler and the interrogative introducing the EQ. An explanation of the whether-when differences in terms of fRM would have to posit that Spanish cuándo ‘when’ is more similar to the wh-filler than si ‘whether’, and it is therefore a stronger intervener.^[8] However, we think that an fRM account will have trouble explaining inter-trial inconsistency in judgments.

It might also be argued that the differences between whether and when-EQs are not caused by any syntactic constraint and that they arise instead for grammatical, but extra-syntactic reasons, like a semantic factor. One potential avenue for explaining the differences has to do with presuppositions. When we had picked up __ in (14e), (14f) presupposes that something was picked up, whereas whether we had picked up __ in (11e), (11f) does not. Clauses that express presupposed content (such as complements of factive verbs) are known to yield weak island effects (Abrusán 2011, 2014; Cinque 1990; Szabolcsi 2006), so the larger island effects in when-EQs could be related to their presupposing the event in the embedded clause. We leave the exploration of semantic accounts for future research.

Finally, the contrast in the acceptability of extraction from whether and when-EQs could (at least partially) reflect that the former are easier to process than the latter. First, when-island sentences contain two filler-gap dependencies, while their whether counterparts contain only one, and an additional dependency can increase processing burden. Second, if when shares more features with the wh-filler than whether, as discussed above, this can cause similarity-based interference, hindering the retrieval of the wh-filler at the gap position (Atkinson et al. 2016; Keshev and Meltzer-Asscher 2019; Villata et al. 2016). Similarity-based interference may also explain why extraction from whether clauses was rejected more often than extraction from declarative clauses, because non-syntactic features also count for the computation of similarity under a processing account.^[9]

Finally, moving on to other EQ-islands: most of the discussed accounts for the difference between whether and when-island effects predict that other wh-interrogatives should pattern like when. Indeed, Ortega-Santos et al. (2018) provide evidence that por qué ‘why’ patterns like when, as they found that por qué yielded large island effects even though test sentences were presented in a context that arguably reduced semantic/pragmatic infelicity. Thus, it is possible that extraction from Spanish EQs is considerably degraded with all interrogatives except for whether, against previous theoretical claims, even if this degradation is not caused by the violation of a bounding constraint.^[10]