2.1 Islands at the intersection between syntactic, discourse-based, and processing accounts

Long-distance dependencies in general, and islands in particular, have been at the core of many debates in the literature for decades. The contrast between sentences like (1a) and (1b) has been accounted for in different ways. Proponents of grammatical accounts have put forth a series of syntactic constraints that somehow constrain extraction from island structures. One of the most influential syntactic analyses revolves around the Subjacency Condition (Chomsky 1973; 1977) that blocks movement over two or more ‘cyclic’ (or ‘bounding’) nodes, defined at the time as S (now IP) and NP (now DP). Central to this theory is the idea of locality: movement transformations never directly cross clause boundaries. A fronted wh-element instead moves in successive steps, using the specifiers of CP as ‘escape hatches’, rather than moving in one fell swoop. In the formation of (1a), for example, the wh-dependency is local because it is confined to the embedded CP what that Mary ate ___ in a first derivational step. Then, the distance between the overt filler and the gap formed in the specifier of the intermediate CP is structurally short as well: only one cyclic node is crossed in each movement step. Unlike other complementizers, however, wh-elements that introduce an interrogative clause are taken to occupy the specifier position of CP. This position is occupied by the embedded wh-phrase in (1b), and successive movement is hence blocked (because the position cannot be used as an escape hatch). The wh-filler is consequently forced to move to sentence-initial position in one fell swoop, and in so doing, it crosses two cyclic nodes: the two IPs corresponding to the matrix and embedded predicates. The sentence in (1b) thereby violates the Subjacency Condition and is ungrammatical, which in turn leads to unacceptability.

Rizzi (1982), however, shares examples from Italian in which a relative pronoun is extracted from an embedded interrogative, thereby crossing more than one cyclic node, see (3). Because che ‘what’ introducing the embedded interrogative clause prevents the specifier position from serving as an escape hatch, fell-swoop movement is forced in this case. The dependency consequently crosses two cyclic nodes (the IPs of raccontare ‘tell’ and domandarsi ‘wonder’, respectively), which violates the Subjacency condition. The sentence should therefore be ungrammatical, contrary to fact.

Thus, Rizzi (1982) proposes that Subjacency might be subject to parametric variation: what qualifies as a cyclic node may differ from one language to the next, and in Italian (or the Romance languages more generally) it may not be IP (then S), but CP (then S̄) that is a hurdle to movement. Not crossing the CP boundary of the matrix clause, the filler in (3) only moves over a single cyclic node on Rizzi’s account.

Notice that the sentence in (3) demonstrates a relative clause (RC) dependency. Wh-dependencies across island structures, as in (4), are by contrast unacceptable in Italian, which Rizzi (1982) attributes to independent factors related to question formation in Italian: sentences with two or more wh-phrases are (presumably) rejected in Italian. Sprouse et al. (2016: 312) note, however, that this restriction is no longer problematic in many varieties of Italian spoken today, and moreover that island types other than interrogative islands do not have a second wh-element. Rizzi’s (1982) theory (which was transposed to French by Sportiche 1981 and to Spanish by Torrego 1984) thus predicts no difference ceteris paribus between wh-dependencies and RC-dependencies (in fact, Torrego 1984 claims that wh-dependencies in Spanish essentially show the same pattern as RC-dependencies in Italian).

In and of itself, the Subjacency Condition does not account for the adjunct and sentential subject island effects identified by Ross (1967).³ Huang (1982) consequently postulated the Condition on Extraction Domains (CED) which blocks movement out of non-complements. In the Barriers framework (Chomsky 1986), adjuncts and subjects are similarly considered barriers that impede movement – a notion that in contrast to the formerly postulated cyclic nodes is structurally defined. Thus, if Romanian behaves like other Romance languages, the syntax-based Subjacency and Barriers accounts predict island effects to emerge when wh-dependencies are formed across adjuncts and subjects (as they are inherent barriers to movement), and complex NPs (i.e. NPs modified by a relative clause; as per the Subjacency Condition), but not across interrogative clause boundaries (see the discussion on (3) above).

According to Newmeyer (2016: 191), the particular behavior of islands has only played “a fairly minor role in theoretical discussions among formal syntacticians” since the late 1980s, although we can assume that the hypothesis that islandhood may be related to formal constraints persists. Indeed, notable locality principles such as (featural) Relativized Minimality (Rizzi 1990; Starke 2001) and the Minimal Link Condition (Chomsky 1995) have since been proposed, which constrain movement in terms of the structural properties of the intervening material. However, space limitations prevent us from a detailed description of these accounts. In what follows, we instead discuss non-syntactic accounts that have been developed over the years, which for the most part appeal to certain conditions on discourse packaging and/or processing.

One influential account that proposes that island status is predicted by discourse factors comes from Goldberg (2006; 2013). She claims that the extent to which a construction backgrounds its content is indicative of island status,⁴ formalized in the BCI Condition in (5). Long-distance dependency formation foregrounds the filler (inside its domain), and so if the gap is located inside a backgrounded construction (i.e. an island), this element is infelicitously both foregrounded and backgrounded at the same time. The ‘discourse clash’, Goldberg claims, is what leads to unacceptability. The BCI approach finds its roots in earlier works by e.g. Erteschik-Shir (1973) and Erteschik-Shir & Lapin (1979).

Another account that attributes island effects to a clash of discourse functions comes from Abeillé et al. (2020), in the form of the Focus-Background Conflict in (6). The main difference between these proposals is that the FBC predicts island effects in dependencies with focused fillers only (i.e. wh-dependencies). Long-distance dependencies with a non-focused filler (such as in RC-dependencies) by contrast comply with the criterion and are hence not ruled out. That the two dependency types may lead to distinct island effects is corroborated in experimental work on various languages. Abeillé et al. themselves tested PP extractions from subjects and objects in French and English. They found, for both languages, that wh-dependencies which originate in subjects yielded lower acceptability ratings than those which originate in objects, but they crucially did not find this same pattern for RC-dependencies. López Sancio (2015) and Sprouse et al. (2016) similarly observed subject island effects in Spanish and Italian, but only in items with wh-extraction, not in relative clauses. Sprouse et al. moreover report adjunct island effects for wh-dependencies, but not for RC-dependencies, in English.⁵,⁶ Kobzeva et al. (2022), finally, report small island effects for wh-extraction but not relativization out of Norwegian conditional adjuncts and embedded interrogatives. If these clauses are more backgrounded than their declarative counterparts, wh-extraction from them should induce a penalty under the condition in (6).

While the results of Kobzeva et al.’s (2022) experiment are in line with this prediction, the authors argue that their findings do not provide unequivocal support for the FBC condition: the interaction effect in the items with wh-extraction from embedded interrogatives was only small and possibly due to a ceiling effect in two of the conditions (the short-distance movement cases), while the acceptability patterns in the other two conditions (the long-distance movement cases) were nearly identical. Moreover, the predicted penalty for wh-extraction from (backgrounded) adjuncts was inconsistent and did not emerge in the majority of trials. Furthermore, Cuneo & Goldberg (2023) find strong correlations between the acceptability ratings of wh- and RC-dependencies in various (English) constructions, including different types of adjuncts.⁷ The empirical facts so far thus do not unequivocally support the FBC condition in (6).

According to Chaves & Putnam (2020), the acceptability of sentences with a dependency crossing an island boundary is influenced by the structural frequency and plausibility of the sentence. Coherent and prototypical complex structures are easier to process, which translates into increased acceptability. That cognitive constraints on processing may explain the (un)acceptability of long-distance dependencies has also been claimed by Deane (1991), Kluender (1991; 1992; 1998; 2004), Pritchett (1991), Kluender & Kutas (1993a,b), and Hofmeister & Sag (2010) – studies that pioneered the ‘experimental turn’ in island research. These accounts are based on the observation that the acceptability of island-violating sentences is much higher when factors known to induce processing costs are controlled for. ‘Island effects’ are claimed to be due to an abundance of performance factors which all independently constrain the processing system: the observed superadditivity can be regarded as the cost of exceeding processing limitations. Thus, processing demands for island-violating sentences are too high, as numerous processing cost-inducing tasks must be tended to (see Hofmeister & Sag 2010 for a review of some factors). The success of sentence processing is then contingent on the sum and difficulty of demands imposed by (parts of) the material, given that cognitive resources spent on a given processing task cannot simultaneously be spent on other processing tasks as well. Note that these accounts naturally predict the often-observed gradience in acceptability ratings, and moreover place the focus on features of the whole sentence – not just the filler or gap site, but also the intervening material that may constrain working memory resources (such as “the frequency, specificity or ‘semantic richness’, verb class, and argument selection properties of verbs on the extraction path”, Hofmeister & Sag 2010: 379). Crucially, some processing accounts maintain that no grammatical rules are necessary at all to explain island phenomena (i.e. reductionist accounts, see Sprouse et al. 2012a and Phillips 2013 for discussion). However, processing accounts like Hofmeister and Sag’s (2010) are in principle compatible with the existence of syntactic rules (as well as discourse conditions). Grounded processing approaches, by contrast, consider island effects to be caused by formal syntactic constraints, but appeal to properties of the parser as well. In particular, these approaches consider the grammatical rules to have developed as a consequence of processing difficulty, in that the processing burdens have grammaticized into island constraints.

In summary, island effects and island insensitivity have been explained from different theoretical perspectives. One theoretical camp claims that specific syntactic constraints block particular types of (wh-)extraction. Other researchers propose that the unacceptability of (certain) long-distance dependencies is due to conflicts in the discourse representation, either because extraction happens from a backgrounded domain (Goldberg 2006; 2013), or because an extracted wh-element is both focused and non-focused at the same time (Abeillé et al. 2020). A third group of researchers submit that, when parsing long-distance dependencies, various interacting performance factors may cause the parser to ‘break down’, so to say. Regarding the wh-dependencies we tested in the experiment to be presented in Section 3, most accounts predict island sensitivity in Romanian, with the exception of the Subjacency account (under the assumption that the cyclic nodes of Romanian are CP and DP, like in Italian, Spanish, and French); this account does not predict island effects to emerge when the gap site is located inside an interrogative clause. We discuss our linking hypotheses in more detail in Section 3.1.

2.2 Experimental work on island phenomena

In most experimental work on island phenomena, ratings of acceptability are collected for a larger set of experimental items from a larger sample of (non-linguist) participants. Many experiments cross two particular factors to isolate island effects: distance (i.e. dependency length) and the type of construction, as the usual definition of an ‘island effect’ involves a decrease in acceptability of sentences with a long-distance dependency with its gap located inside an alleged island construction (Sprouse & Villata 2021). Thus, a 2×2 design that manipulates i) the presence of a long-distance dependency, and ii) the presence of an alleged island, can be used to investigate island effects. It is worth noting here that the two factors may influence acceptability independently. The additional processing cost of keeping a long-distance dependency active in memory, compared to a short-distance dependency, may negatively affect the acceptability of a sentence, and the presence of an alleged island structure, typically a complex structure, may similarly impact sentence ratings. Therefore, we can only speak of an ‘island effect’ when the decrease in acceptability occurs on top of the individual effects of the factors. That is, statistically, we are interested in super-additivity, i.e. the interaction effect between the two factors.

Consider the sentences in (7) for illustration, in which the two factors construction (non-island vs. island) and distance (short vs. long distance) are crossed. Each factor might be associated with its own cost, but in the case of an island effect the interaction between the two factors is expected to give rise to a super-additive effect, which according to linguistic theory takes the shape of (7d) receiving lower acceptability ratings than predicted by the combined effects of construction and distance.

The size of island effects is usually reported as the difference-in-differences (DD) score on standardized acceptability ratings. The DD-score is computed as follows (see Maxwell & Delaney 2003: 684–687 for more detailed description). A first difference score (D1) is calculated by subtracting the acceptability score of the island/long condition (7d) from the non-island/long condition (7b). D1 quantifies the effect of having an island structure in sentences with a long-distance dependency. A second difference score (D2) is calculated by subtracting the mean acceptability rating of the island/short condition (7c) from the non-island/short condition (7a). D2 quantifies the effect of having an island structure in sentences with a short-distance dependency. Finally, the DD score is calculated by subtracting D2 from D1, and represents the difference between differences related to the two factors.

Since Sprouse (2007), the factorial design discussed above served as the basis for acceptability judgment experiments in investigations of island phenomena in various languages. Here, we focus on the Romance languages, and on Romanian in particular. We first discuss experimental studies on island effects across Romance languages in Section 2.3 and continue by providing novel data from Romanian.

2.3 Experimental studies on island effects in Romance languages

We summarize the main findings from several experiments on islands in the Romance languages in this section, but this is by no means an exhaustive review. Recall that theoretical linguists from the early 1980s claimed that Romance languages do not obey the same island constraints as English does (Rizzi 1982 for Italian, Sportiche 1981 for French, Torrego 1984 for Spanish). Thus, extraction from interrogative embedded clauses in particular is expected to be more acceptable in Romance languages, at least when a RC-dependency is involved, but the theory adopted in these studies can in principle be extended to other types of islands and to wh-dependencies as well (see also Sprouse et al. 2016). This section discusses experimental work investigating different types of island effects in the Romance languages.

Sprouse et al. (2016) conducted a series of judgment experiments to gauge the acceptability of extraction from four different island types (interrogative, complex NP, subject, and adjunct islands) in Italian and English. They tested both RC- and wh-dependencies, but in this paper we focus on the items with a wh-dependency. For each island type, the authors tested four conditions manipulated across the factors distance (short vs. long) and construction (island vs. non-island), with eight lexicalizations per experiment. Importantly, the design of the items with a subject island was different from the other island types, because Italian does not allow preposition stranding (see also Abeillé et al. 2020), and without a preposition following the object, there might be local ambiguity. Thus, Sprouse et al. (2016) tested the acceptability of extraction of simple subjects and objects, and from complex subjects and objects, see (8).

Sprouse et al. (2016) tested 193 native speakers of Italian (spread across four experiments) and found super-additive effects for all four island types with a wh-dependency in Italian, with varying effect sizes (complex NP islands DD 0.89; adjunct islands DD 1.31; subject islands DD 1.37; interrogative islands DD 1.69).

A similar picture emerges in experimental work on Spanish. López Sancio (2015) replicated Sprouse et al.’s (2016) experiment for Spanish in his Master’s thesis, and asked 217 native speakers of Spanish (distributed over four experiments) to rate stimulus items, testing two dependency types (wh- and RC-dependencies) and four island types (interrogative, complex NP, subject, and adjunct islands). The stimulus items in this experiment were constructed following Sprouse et al.’s (2016) items, with a slight adaptation to the subject island items: the subject and object DPs were made non-specific (e.g. ¿De quién crees que varios escándalos han suscitado rumores sobre el candidato? ‘Of whom do you think various scandals have sparked rumors about the candidate?’, cf. (8)). Regarding the items with a wh-dependency, López Sancio reports significant island effects for all four island types (complex NP islands DD 1.41; adjunct islands DD 1.71; subject islands DD 0.80; interrogative islands DD 1.65).

Island effects in Spanish are also reported by Pañeda et al. (2020), who tested wh-dependencies across the same four constructions, but who used a speeded acceptability judgment experiment instead. Participants in Pañeda et al.’s experiment were asked to evaluate items as fast as possible after fast-paced word-by-word presentation on a computer screen, on a binary scale, instead of a 7-point scale. Regarding the subject island items, however, the experimental design used in Pañeda et al. was crucially different from that used in Sprouse et al. (2016) and López Sancio (2015). It instead crossed the familiar factors distance and construction, see (9): the filler either originated in the matrix clause (9a,b) or in the embedded clause (9c,d), and was only extracted from a subject in (9d). Notice that the conditions have different modifiers; this was done so that each condition has a comparable number of words in it.

Pañeda et al. (2020) tested 80 native speakers of Spanish, who each saw all island types, and found super-additive effects for the subject, adjunct, and interrogative islands,⁸ with subject island effects larger than the other two island types, but no clear evidence for island effects in the items with a complex NP island.⁹ Notice that Pañeda et al.’s findings do not fully corroborate López Sancio’s (2015) findings: the subject island effect that López Sancio reports is the weakest of all four investigated island types, whereas Pañeda et al. found the strongest interaction effect in the items with a subject island. Pañeda et al. argue that these differences might be due to the alternative design used to test subject islands, as the design in (9) might overestimate the island effect size because the interaction term may also contain independent effects of subextraction from complex phrases (as pointed out by Kush et al. 2018). The design in (8), by contrast, might underestimate the effect size, since the complexity of the added DPs in (8c) and (8d) may reduce the acceptability ratings to the extent that floor effects reduce the interaction effect (Kush et al. 2018).

Abeillé et al. (2020) designed a series of experiments to test RC- and wh-extraction from subject islands in English and French. Abeillé et al. crossed the factors grammatical function (subject vs. object) and extraction (PP vs. no extraction),¹⁰ resulting in yet another design to test subject islands, see (10). Still, Abeillé et al.’s design is rather close to that in (8), the main difference being that the non-extracted subjects and objects remain in situ in (10a,b) but are fronted in (8a,b), and the subjects and objects are swapped between conditions in (10). The situation described in (10) is the same across conditions.

Abeillé et al. (2020) tested 47 native speakers of French, and found significant subject island effects in the wh-dependency experiment. As Abeillé et al. posted their research data online, we could calculate the corresponding DD-score for this experiment, which was 0.68, but note that this DD score is not fully comparable to those reported for the other experiments because of the alternative experimental design. Crucial for the authors’ purposes was that they did not find significant island effects in the items with a RC-dependency, and that the patterns for French and English are highly similar. We, however, focus on the fact that their results corroborate the findings for Italian (Sprouse et al. 2016) and Spanish (López Sancio 2015; Pañeda et al. 2020), in that wh-extraction from subjects received worse ratings than would be expected, at least to the extent that the experimental designs can be compared.

The experimental studies discussed above show that island effects are consistently reported for wh-dependencies in various island constructions in the Romance languages that have been investigated so far. The reported island effect sizes are different, however, which in many cases is (presumably) due to differences in the design of the stimulus materials. Our experiments (reported in Section 3) followed the design used in Sprouse et al. (2016) and López Sancio (2015) to a large extent; the main difference is that we tested two island types with a wh-dependency per experiment, while Sprouse et al. and López Sancio tested one island type with a wh-dependency and another with a RC-dependency per experiment. Table 1 reproduces the island effect sizes that Sprouse et al. and López Sancio found in their items with a wh-dependency, per construction type; the findings from Pañeda et al. (2020) and Abeillé et al. (2020) are not represented here because of their alternative research designs. Our investigation contributes data from wh-extraction from four different island types in Romanian to the archipelago (in contrast with Stoica & Schoenmakers 2024, who only tested interrogative islands). We report on our experiments in Section 3, but before that, we turn to the literature on islands in Romanian.

2.4 Islands in Romanian

As noted, island phenomena are relatively understudied in Romanian. The theoretical literature largely acknowledges the presence of island effects with respect to adjuncts, interrogative clauses, and complex NPs (Pană Dindelegan 2013), illustrated with wh-dependencies in (11).¹¹ The example in (12), provided by the native-speaking author of this article, is added to illustrate that extraction from subject islands is presumably also unacceptable in Romanian.

Certain factors have been argued to alleviate island effects in Romanian (and the Romance languages more generally). Based on the fact that Romanian allows multiple fronted wh-elements (or, it allows a multiply filled COMP), Comorovski (1986: 177) originally predicted that the language “will not obey any form of the Wh-island Constraint”; Rudin (1988) similarly claimed that Romanian allows extraction from interrogative clauses, illustrated in (13). Later, however, Comorovski (1989) and Alboiu (2002) argued that this claim only holds for D(iscourse)-linked elements (Pesetsky 1987), i.e. elements whose reference can be reconstructed from the existing discourse context. Fillers that contain the phrase which, for instance, are contextually more salient than fillers that contain the phrase what, and their extraction from island constructions is consequently more acceptable, as in (14). Sevcenco (2006) suggests that D-linking ameliorates extraction from complex NPs as well, see (15). Notice that the extracted object in (15a) is clitic-doubled because of the form of the filler (Dobrovie-Sorin 1990); the sentence in (15b) is ungrammatical regardless of the presence of this clitic.

Stoica & Schoenmakers (2024) experimentally investigated the claim that D-linking alleviates island effects that are due to extraction from Romanian interrogative clauses headed by dacă ‘whether’ and când ‘when’. In their experiment, they manipulated the filler type between bare wh, what N, and which N fillers, predicting a gradual increase in acceptability, the more D-linked the filler is. Although Stoica and Schoenmakers found significant island effects in each sub-experiment, their predictions regarding the discourse embedding of the fillers were not borne out entirely. That is, in line with their predictions, they found that which N fillers led to weaker island effects than bare wh fillers; however, what N fillers led to unexpectedly strong island effects – stronger, in fact, than which N fillers. The authors conclude that D-linking accounts cannot straightforwardly explain their results. However, they also note that long-distance wh-extraction from non-island constructions (16) yielded unexpectedly high ratings compared to the other conditions, especially in the items with a what N filler, which seems to have been the driving factor behind the differences in the island effect sizes, even though no theory would a priori predict this pattern. We return to this observation in Section 3.6.2. The present paper only investigated bare ce ‘what’ fillers, in line with what López Sancio (2015) and Sprouse et al. (2016) did for Spanish and Italian,¹² but unlike Stoica & Schoenmakers (2024) we tested four different island types.

Finally, Sevcenco (2006) claims that Romanian island effects are also alleviated when a subjunctive is used in a complex NP, see (17), and this appears to hold for other island types as well: the sentences in (18) with an interrogative and a subject island become unacceptable when the subjunctive is changed to an indicative (see also Baunaz & Puskás 2014). As far as we know, this difference has not yet been investigated experimentally. The alleged islands in our experiment were in the indicative.

In sum, it has been noted that Romanian wh-questions may be sensitive to various types of islands, with factors such as D-linking and verbal mood playing a role. With the exception of Stoica & Schoenmakers (2024), to our knowledge there have been no experimental investigations of island effects in Romanian. The next section reports on two experiments that followed the same basic design as in López Sancio (2015) and Sprouse et al. (2016), measuring the acceptability of wh-dependencies across four different types of island structures.

3.1 Linking hypotheses

Under the assumption that the cyclic nodes in Romance languages are CP and DP (extrapolating from Sportiche 1981, Rizzi 1982, and Torrego 1984), Chomsky’s (1973; 1977) Subjacency Condition predicts that interrogative islands are unproblematic in Romanian, whereas complex NP islands resist extraction. Specifically, while initial movement to the edge of the embedded CP (within the complex noun phrase) is allowed, subsequent movement into sentence-initial position would cross two cyclic nodes (viz. the CP and DP boundaries). This would consequently be disallowed under Subjacency. With interrogative clauses, by contrast, only one cyclic node is crossed at a time (see the discussion on (3)). Experimental data from Spanish and Italian (López Sancio 2015; Sprouse et al. 2016), however, suggest that these languages do in fact resist wh-extraction over interrogative clause boundaries. Further, extraction from adjuncts and subjects is expected to be disallowed according to the CED (Huang 1982) and the Barriers framework (Chomsky 1986), as adjuncts and subjects are both non-complement constituents.

According to Goldberg’s (2006; 2013) BCI condition (see (5)), all island constructions that we investigate are backgrounded constituents and resist extraction, as long-distance dependency formation foregrounds the filler while the gap is located inside a backgrounded domain. Similarly, Abeillé et al.’s (2020) Focus-Background Conflict predicts that extraction from the island constructions investigated here yields unacceptability, given that wh-dependency formation involves focus movement. The filler-gap dependency would therefore simultaneously be focused (because the filler is a wh-phrase) and non-focused (because the gap site is located inside an island).

Processing accounts argue that the combination of multiple factors that induce a processing cost impede parsing in structures with a wh-dependency across an island boundary. The super-additive effect observed in experimental research is then viewed as the cost of exceeding cognitive resources. All else being equal, processing accounts also predict island effects in our experiments, but these island effects are expected to correlate with individual participants’ working memory capacity (Sprouse et al. 2012a). Since we did not include a working memory task in our experiment, which would help separate the predictions of processing accounts from those from the syntax-based accounts, we instead refer the reader to the discussion between Hofmeister et al. (2012a,b; 2014) and Sprouse et al. (2012a,b) (see also Michel 2014 and Pham et al. 2020).

To recap, the Subjacency account that adopts parametrization predicts sensitivity in the items with adjunct, subject, or complex NP islands, but not in the items with interrogative islands. Accounts that are based on discourse clashes or resource overload predict island sensitivity in all construction types. Earlier experimental work on Romance languages found island effects in each of the item sets that we tested (López Sancio 2015; Sprouse et al. 2016). Given that our experiments employ the same design and also test a Romance language, we expect to find similar results. That is, we expect to find island effects as defined in Sprouse & Villata (2021) as an interaction effect between the factors distance (long vs. short) and construction (island vs. non-island) in the adjunct, embedded interrogative, and complex NP item sets, and as an interaction effect between gap site (subject vs. object) and complexity (simple vs. complex) for subject islands (Sprouse et al. 2016).

3.2 Participants

211 native speakers of Romanian without language deficits completed one of two versions of the questionnaire, either with adjunct and complex NP islands or with subject and interrogative islands, and received course credit upon completion. The participants were unfamiliar with syntactic theorizing or island phenomena. Data from participants were removed if their average score for the grammatical filler items was lower than 4 (on a 7-point scale), if their average score for the ungrammatical filler items was higher than 4, or their standard deviation was lower than 1. This was the case for seven participants. We entered data from 103 participants in the adjunct/complex NP-island task (M_age: 20.6, SD: 3.1, range 18–46) and 101 participants in the subject/whether-island task (M_age: 21.7, SD: 5.1, range 18–47) into statistical analysis.

3.3 Materials

We conducted two acceptability judgment tasks, one testing adjunct islands and complex NP islands, the other testing interrogative islands and subject islands. Both tasks followed the factorial design used in Sprouse et al. (2016), which isolates the factors distance (short vs. long) and construction (island vs. non-island). Sample items for each island type are given in (19) – (22); here we added the boldface font, brackets, and underscores as a visual aid. The stimuli followed the same pattern as in Sprouse et al.’s experiment, but were adapted in order to control for definiteness and tense (the full stimulus list can be found in our OSF repository). All islands were in the indicative. All adjunct island items contained conditional dacă ‘if’ clauses, and all interrogative island items contained dacă ‘whether’ clauses. As discussed in Section 2.1, we also follow Sprouse et al.’s alternative design for the subject island items, which instead crossed the factors gap site (subject vs. object) and complexity (simple vs. complex). The reason to use this design is that the Romance languages do not allow preposition stranding, and the alternative design minimizes the possibility that participants interpret the sentences on the wrong reading (Sprouse et al. 2016).¹³ Notice that the subjects and objects in the complex conditions in (22) are both modified by a PP. The alternative design moreover controls for subextraction (unlike the one adopted in e.g. Kush et al. 2018; Kobzeva et al. 2022): by manipulating the extraction domain in the complex conditions, it permits a comparison of subextraction from subjects vs. subextraction from constituents in different positions. The embedded clause in the condition with subextraction from a complex object (20b) was VOS for processing reasons. If the sentence were Despre ce crezi că a iscat raportul despre dictator o discuție ___ ? ‘What do you believe the report about the dictator sparked a conversation about?’, the second despre-PP before the gap site might induce higher processing costs, because it introduces another DP before the filler-gap dependency is resolved. Moreover, this maximizes similarity with the condition subextracting from a complex subject (22d) in terms of dependency length. The embedded clauses in all other conditions were VSO sentences. An anonymous reviewer points out that the subjects in the subject island item set were always definite, while the objects were always indefinite. Since definite DPs are less transparent than indefinite DPs (‘definite islands’ in Ross 1967; see Neal & Dillon 2021 and Shen & Lim 2024 for experimental investigations), a putative island effect may include a definiteness effect. We return to this confound in Section 3.6.1, noting here that the effect sizes in the cited experimental investigations were only small.

For both sub-experiments, eight items per island type (i.e. sixteen experimental items in total) were distributed over four experimental lists according to a Latin Square design (i.e. eight experimental lists in total). Following Sprouse et al. (2016), twelve grammatical and twenty ungrammatical fillers were added to each experimental list, resulting in a 2:1 ratio of experimental and filler items in each list. The number of grammatical and ungrammatical fillers was chosen on the assumption that three of the four experimental conditions are acceptable, and one unacceptable. Each experimental list thus contained 24 (presumed) acceptable and 24 (presumed) unacceptable items. Sample grammatical and ungrammatical filler items are given in (23); (23b) is ungrammatical because the adjunct până în ziua de azi ‘up until today’ intervenes between the fronted wh-phrase and the verb (Alboiu 2002).

The items were pseudorandomized using the software Mix (van Casteren & Davis 2006) and each list was presented in one of two distinct orders. Each list started with three practice items that were excluded from statistical analysis. The experiments were conducted in Qualtrics.

3.4 Procedure

The experiments were online questionnaires that started with a general information and consent form. Participants provided demographic data and were assigned to one of the experimental lists at random. They were instructed to judge on a 7-point scale how natural the target sentences would sound when uttered by a native speaker of Romanian, where 1 represented foarte rău ‘very bad’ and 7 foarte bine ‘very good’. The instructions mentioned that participants could rely on their first intuition and that there are no right or wrong answers. Participants could provide their answer using radio buttons, which were presented with the corresponding value (from 1 to 7). The verbal labels foarte rău and foarte bine were presented at the extremes only (1 and 7). Each trial was presented on a new page. The experiments started with three practice trials before continuing to the actual experimental task, and concluded with the opportunity for participants to provide qualitative comments about the experiment.

3.5 Analysis

We performed a series of linear mixed effect models on the z-transformed judgment scores using R (version 4.2.3, R Core Team 2022) and the lme4 package (Bates et al. 2015). For the items with adjunct, complex NP, and interrogative islands, we entered the two-level variables distance (short vs. long) and construction (island vs. non-island) into the models as fixed effects; for the items with subject islands we entered the variables gap site (subject vs. object) and complexity (simple vs. complex). All variables were encoded using deviation contrasts (–0.5; 0.5). We started out with base models that included the predictors of interest and their interaction. By-item and by-participant random intercepts were included in the model only if they improved the model fit (assessed by Likelihood Ratio Tests), followed by by-participant random slopes. The models did not include by-item random slopes, because the experiments tested only eight items per island type (i.e. two per condition). The final model for the adjunct island item set included a by-item random intercept. The final model for the complex NP island item set was a simple linear model, because inclusion of random effects did not improve the model fit. The final model for the subject island item set included by-participant and by-item random intercepts. The final model for the interrogative island item set included by-item and by-participant random intercepts as well as a by-participant random slope for the effect of construction. In case an interaction effect was significant, we ran post-hoc analyses with Bonferroni-corrected p values using the emmeans package (Lenth 2024).

3.6 Results

The mean (non-standardized) judgment scores and standard deviations for the four conditions are given for each island type in Tables 2 and 3. The mean rating for the grammatical filler items was 6.1 (SD = 1.5) and the mean rating for the ungrammatical filler items was 2.3 (SD = 1.7). The (z-transformed) data are visually represented in Figures 1 and 2. In what follows, we discuss the results for each island type in turn.

Figure 1: Interaction plots for adjunct, complex NP, and interrogative islands in Romanian (wh-dependencies).

Figure 2: Interaction plots for subject islands in Romanian (wh-dependencies).

3.6 Discussion

Our results show that super-additive effects arise in all islands constructions tested in our experiment, providing first experimental evidence for three types of island sensitivity in Romanian wh-dependencies (see also Stoica & Schoenmakers 2024 for interrogative islands). We find the strongest island effect in items with an interrogative island (DD 1.80), with gradually weaker effects in items with complex NP (DD 1.40) or adjunct (DD 1.10) islands, and the weakest effect in items with a subject island (DD 0.41). However, recall that the latter item set followed a different design, which has been claimed to underestimate the island effect size (Kush et al. 2018).

Interestingly, closer inspection of Figure 1 shows that the interrogative island and adjunct island effects are not only driven by the low ratings for wh-extraction from island constructions; they are also associated with unexpectedly high ratings for wh-extraction from non-island constructions (compared to the other conditions). Post-hoc analyses confirm that the non-island/long conditions received significantly higher ratings than the two short-distance conditions (adjunct islands: t_island/short = 4.32, t_{non-island/short} = 4.77; interrogative islands: t_island/short = 8.10, t_{non-island/short} = 11.90; all p < .001). Although this finding does not dispute the existence of the island effects in Romanian per se, it does call into question the reported DD-scores (cf. Stoica & Schoenmakers 2024). We address this issue in Section 3.6.2.

4.1 Results

Despite the unexpected patterns discussed in Section 3.6.2, based on the full data pattern in Figure 1 we conclude that Romanian displays island sensitivity: the theoretically marked island/long conditions received lower acceptability ratings in each item set, as confirmed by post-hoc pairwise comparisons. More specifically, Romanian exhibits at least interrogative, adjunct, and complex NP island effects. While the adjunct and complex NP island effects were expected based on the literature, the interrogative island effect contradicts the original claim that the language does not obey the Wh-island Constraint (Comorovski 1986, Rudin 1988). Our stimuli were presented free of context, however, and the fronted elements were bare wh-phrases. It has been claimed that extraction from interrogative clauses is possible on the condition that the extracted element is D-linked (Comorovski 1989; Alboiu 2002) and a recent experimental study (Stoica & Schoenmakers 2024) shows that D-linking (bare wh vs. which N) indeed ameliorates extraction from whether- and when-islands in Romanian (although the island effect in the D-linked conditions is still significant). The degree of discourse embedding may thus impact the island effect size, which may also play a role in other island types (as claimed in Sevcenco 2006 for Romanian complex NP islands).

We also find a weak but significant subject island effect, but we treat this effect with caution. The alternative experimental design we chose may have underestimated the size of the island effect, because of floor effects in the complex conditions (Kush et al. 2018), while a definiteness effect may have overestimated it.

Our results are in line with earlier experimental work on Italian (Sprouse et al. 2016) and Spanish (López Sancio 2015), although there is variation between the languages with regard to the island effect sizes (which are in part due to the unexpected pattern in the Romanian data, see Section 3.6.2). In what follows, we discuss potential avenues for future research.

4.2. Limitations and future directions

Our experiments provide initial evidence for island sensitivity in Romanian, based on experimental data from four construction types. This outcome was predicted by most theories of island phenomena. Future research can be geared towards more precise claims from linguistic theories. We suggest various ways forward.

First, all our stimuli contained island structures in the indicative. Indicatives clauses, however, have been argued to be less permeable to extraction than subjunctive clauses (see Section 2.4; Baunaz & Puskás 2014). This claim can be tested experimentally in future work.

Second, recall that Abeillé et al.’s (2020) Focus-Background Conflict criterion in (6) predicts that a focused element should not be part of a backgrounded constituent. As such, the condition rules out wh-dependencies, but not other types of dependencies which involve non-focused movement, such as relativization or topicalization (pace Goldberg 2006; 2013). Our experiment tested wh-dependencies only. All else being equal, the FBC criterion predicts that other dependency types are not resisted in the same way as wh-dependencies. Future work could therefore investigate other dependency types, as has been done for other languages as well (e.g. López Sancio 2015; Sprouse et al. 2016; Kush et al. 2019; Abeillé et al. 2020; Kobzeva et al. 2022; Cuneo & Goldberg 2023).

Third, our experiment did not include a working memory task, so we cannot investigate the potential correlation between the island effect size and working memory capacity, as would be predicted by processing accounts of island phenomena. Future experiments may include a working memory task, although the design of the experiment as well as the choice of working memory task are matter of debate (Hofmeister 2012a,b; 2014; Sprouse et al. 2012a,b). Earlier experimental investigations have found that individual cognitive resources explain only little variation in the observed island effects (Sprouse et al. 2012a; Michel 2014; Pham et al. 2020).

Processing accounts also argue that processing is easier with stronger discourse embedding or accessibility (see Hofmeister & Sag 2010). Complex DP fillers (e.g. which book instead of what) are therefore expected to yield higher acceptability ratings than bare wh fillers, because they are D-linked (Pesetsky 1987) (or because there are less demanding interference effects, Villata et al. 2016; Chesi et al. 2023; Stoica & Schoenmakers 2024). Experimental studies on various languages have included the filler’s complexity as a factor in their design and indeed found amelioration effects (Goodall 2015; Sprouse et al. 2016 for English; Kush et al. 2018; Stoica & Schoenmakers 2024), but the island effect persisted when the fillers were complex instead of bare wh-phrases. As noted, Stoica & Schoenmakers (2024) tested the claim for Romanian interrogative islands, but the question has not been addressed for other island types in the language.

Finally, another approach to testing claims from processing accounts, suggested by a reviewer, is to investigate the impact of verb frame frequencies in the experimental items (Liu et al. 2022a). The [V + that] frame may be more frequent than the [V + whether] frame, for example, which may in turn increase the acceptability of the conditions with an embedded declarative, on the assumption that these sentences resemble the stored template of a prototypical structure more closely (after Dąbrowska 2008). While post-hoc computations of this kind would be a viable option for our interrogative island item set, our experiment did not manipulate the verb forms (we only tested crede că ‘believe that’ and se întreaba dacă ‘wonder whether’). The results would consequently not be theoretically informative. Verb frame frequences may provide potentially insightful information, however, as they did for English in Liu et al. (2022a), and so this hypothesis can be tested for other languages in dedicated investigations in the future.

We conclude this section with a methodological note. Our experiments as well as several earlier experimental studies tested only eight items per island type. In a 2×2 design, this amounts to participants rating only two distinct items per condition. We urge future researchers to test larger numbers of items, so as to avoid that the acceptability ratings reflect lexical properties of the stimuli or other idiosyncratic factors. More generally, statistical models may struggle to provide accurate predictions when only a small number of items are tested (i.e. low statistical power). Although it is difficult to specify an ideal number of items required in experiments like the one presented in this paper, it is worth noting that larger numbers of items increase the reliability of the data-set. As Maxwell & Delaney (2003: 25) illustrate, the chances of passing a multiple-choice quiz by simply guessing the answers are slimmer when the test is longer. Increased numbers of items in experiments investigating island phenomena would thus allow for closer scrutiny of the variation patterns in them.