2.1 Previous findings

When islands were first characterized and described in detail in Ross (1967) and later by Chomsky (1973; 1977; 1986), islands were explained in terms of syntactic principles. The claim was that islands arose from innate, universal, syntactic constraints on general movement operations. The traditional syntactic accounts such as the Subjacency Condition (Chomsky 1973; 1977) and Barriers (Chomsky 1986) alongside Phases (e.g., Chomsky 2000) predict that there will be minimal variation between island domains and between languages, and that any variation observed must be due to independent syntactic differences (see e.g. Rizzi 1982). Much research has, however, questioned this clear set of predictions, both within and across languages.

According to many researchers, particularly within traditional syntactic approaches to islands, adjunct clauses have maintained their status as strong and universal islands (see e.g., Stepanov 2007, and the overview in Bode 2020). Thus, the empirical predictions that follow are (i) adjunct islands should have universal validity, unless there is (preferably independently observable) evidence of relevant structural differences between languages; (ii) the acceptability of adjunct island violations should be categorically low (though see Chomsky 1986: 28). Some formal investigations find exactly this. Both Sprouse et al. (2012; 2016) and Kush et al. (2018) find large island effects of forming a wh-dependency into finite adjunct clauses in English and Norwegian, respectively.

Despite the claimed universal validity of the Adjunct Island Condition, much variation has also been uncovered for this island type. Sprouse et al. (2016) find no island effect for finite adjunct clauses in an rc-dependency in English, and Goldberg (2006) and Chaves (2021), among others, provide examples of acceptable extraction from finite adjunct clauses in English.

Norwegian and Swedish have figured prominently in the literature as languages with exceptions to the universal validity of island constraints. The papers collected in Engdahl & Ejerhed (1982) demonstrate a range of variation in MSc languages, among them examples of licit extractions from finite adjunct islands in Norwegian and Swedish (see also e.g., Teleman et al. 1999; Faarlund 1992; Bermingrud 1979 etc.).³

Engdahl & Ejerhed claim that such data challenges the “proposed universal principles of rule application” (1982: 9). Nevertheless, Engdahl (1982) maintains that long-distance dependencies are not unbounded in Norwegian and Swedish as there are several examples of illicit extraction provided alongside licit examples (see e.g. Bermingrud 1979; Faarlund 1992; Teleman et al. 1999).

More recent formal investigations corroborate that there are both licit and illicit extractions from adjunct clauses in Norwegian. Kush et al. (2018) find island effects for finite adjunct clauses in Norwegian in a wh-dependency. In a second series of experiments, Kush et al. (2019) find island effects for topicalization out of finite adjunct clauses, but no island effects for finite adjunct clauses when a context sentence facilitating contrastive topicalization is presented alongside the test sentence. An example of their test material is provided in (5).

Kush et al. (2019: 406) report that contrastive topicalization from a finite adjunct clause with context, on average, was rated to be almost as acceptable as topicalization from embedded declarative clauses. In addition, they find that judgments varied between and within participants. Kush et al. (2019) conclude that conditional adjuncts are not islands for topicalization in Norwegian.

Bondevik et al. (2021) further investigate Kush et al.’s (2019) findings for finite adjunct clauses in a contrastive topicalization dependency with context. Bondevik et al. (2021) test three different adjunct clauses – conditional om ‘if’-clauses, habitual når ‘when’-clauses and causal fordi ‘because’-clauses. Overall, they replicate Kush et al.’s (2019) findings for om ‘if’ showing that om ‘if’ is not treated as an island in Norwegian. However, they find large island effects for fordi ‘because’-clauses, and variable effects for når ‘when’-clauses. They conclude that with regards to islandhood, “adjunct” does not behave as a uniform class in the manner predicted by traditional syntactic approaches.⁴ Additionally, Bondevik et al. (2021) find a wide distribution underlying the average judgments for om ‘if’, much like Kush et al. (2019). They also see this for når ‘when’. They find no predictor which reliably explains differences between participants, nor are they able to identify any syntactic, semantic, or pragmatic factors that reliably predict differences between items which could explain the wide distribution of ratings.

Two recent studies have investigated demonstrative rc-dependencies (dem rcs)⁵ into finite adjunct clauses. Nyvad et al. (2022) investigated English dem rcs into the same three finite adjunct clauses that were tested in Bondevik et al. (2021) – if, because and when. Despite the widely held assumption that all finite adjunct clauses are strong islands in English, the authors find non-uniformity between the different adjunct clause types. As Bondevik et al. (2021) found for Norwegian, they find that forming an A’-dependency into finite if-adjuncts in English is rated much higher than A’-dependencies formed into finite because- and when-clauses. It is worth noting that the same proportional relationship between adjunct clause types replicates across languages (Norwegian vs. English) and across a different dependency type as well (top vs. dem rc). Unlike Bondevik et al. (2021), Nyvad et al. (2022) find that when- and because-adjuncts yield intermediate⁶ island effects. Thus, they argue that their results indicate that all finite adjunct clause types require a gradient theory of adjunct islands.

Kobzeva et al. (2022) do not find a strong island effect for dem rcs in Norwegian conditional om ‘if’-adjuncts. They find a null effect, similar to Sprouse et al.’s (2016) findings for rc-dependencies in English, and average judgments on the “long, island” condition to be just below the acceptable range. In comparing conditional om ‘if’-adjuncts on dem rcs and wh-dependencies, Kobzeva et al. (2022) find that dem rcs yield lower acceptability ratings compared to wh-dependencies, contrary to previous findings in Kush et al. (2018) that both simple and complex wh-dependencies yield large island effects in Norwegian om ‘if’-adjuncts. Kobzeva et al. (2022) suggest that differences between studies might be related to the predicate types used in the different experiments.

2.2 Dealing with variation

Above, we have seen that the traditional claim that all adjuncts are islands cross-linguistically is disputed by more recent evidence of cross-linguistic variation (Sprouse et al. 2016), variation between dependency types (Kush et al. 2018; 2019; Kobzeva et al. 2022), and even variation between and within adjunct clause types (Müller 2019; Bondevik et al. 2021; Nyvad et al. 2022). Variation poses a problem for traditional syntactic accounts, and variation in adjunct islands particularly so. On these approaches, adjunct clauses are constrained by general principles that restrict all adjuncts categorically. For instance, within Huang’s (1982: 505) Condition on Extraction Domain (CED), all adjuncts are islands based on the claim that no adjuncts are (properly) governed.

The notion of proper government has been abandoned in recent theoretical frameworks, but the idea remains that adjuncts are islands namely because adjuncts provide a special type of constituent that is less closely integrated with the matrix clause (see e.g., Bode 2020 for an overview). This is implemented in different ways in Minimalism (see e.g., Chomsky 2000; Stepanov 2007; Hornstein & Nunes 2008). Consequently, all adjuncts are islands simply because they are adjuncts. Thus, traditional syntactic approaches generally do not allow fine-grained variation between and within adjuncts.

Sprouse et al. (2016) review several syntactic approaches to islands looking at how each of these can account for variation in dependency type between languages. For each of the syntax-based approaches that they review, they find that their results are difficult to accommodate. This indicates that none of the syntax-based approaches can easily handle variability. However, they discuss the possibility that Relativized Minimality might have the power to account for differences in dependency types, but they do not provide an explicit analysis of differences between rc-dependencies and wh-dependencies into if-adjuncts in English. Nyvad et al. (2022) come to a similar conclusion as Sprouse et al. (2016) regarding syntax-based approaches. Bondevik et al. (2021) and Nyvad et al. (2022) also review some extra-syntactic approaches, but find that these struggle to readily handle the differences between adjunct clause types.

2.3 Research questions, predictions, and hypotheses

It seems clear that adjuncts are not categorical islands for all A’-dependencies as predicted by traditional syntactic accounts, but that there are some factors that facilitate variation across constructions, languages, and adjunct types. Our main aim is to map the empirical landscape of finite adjunct clauses in Norwegian. Finite adjunct clauses have been tested in a wh-dependency, a top-dependency and a dem rc-dependency in Norwegian. There is evidence of cross-dependency variation for finite adjunct clauses in Norwegian, such that top-dependencies and dem rcs are less sensitive to finite adjunct island effects compared to wh-dependencies (though see Kobzeva et al.’s findings for wh-dependencies). Sprouse et al.’s (2016) findings for English and Italian point in different directions as to whether or not rc-dependencies are sensitive to adjunct islands constraints. We therefore want to test different finite adjunct clause types in an rc-dependency in Norwegian.

In addition, Norwegian finite adjunct clauses provide an interesting case study for investigating the island sensitivity of rc-dependencies. Previous research documents systematic differences between adjunct clauses in Norwegian (Bondevik et al. 2021). It is therefore possible to test (i) whether rc-dependencies are sensitive to adjunct island effects in general, and (ii) whether rc-dependencies are sensitive to all adjunct clause types equally. These are important for two reasons – firstly, it is an empirical problem that we do not know the descriptive adequacy of these constructions, and secondly, by studying these two phenomena in tandem we can begin to create better models for capturing variation displayed by adjunct clauses. Specifically, our research questions are:

1. Are adjuncts islands for relativization in Norwegian?

2. Do different types of adjunct clauses behave like a uniform group for relativization?

The rest of the paper is organized as follows. In Section 3, we give a detailed overview of the experimental design employed in Experiments 1 and 2 and provide a detailed overview of methodology and results for Experiment 1. The second experiment is presented in Section 4. Section 5 provides a discussion of our research questions in view of both experiments. Finally, Section 6 concludes the paper.⁷

3.1 Experimental design

To investigate our research questions, we ran an acceptability judgment study following the 2 × 2 factorial design popularized by Sprouse and colleagues (Sprouse 2007; Sprouse et al. 2016).⁸ This allows for a direct comparison with previous findings for adjunct clauses in rc-dependencies in English (Sprouse et al. 2016), in dem rcs in Norwegian (Kobzeva et al. 2022) and in top-dependencies in Norwegian (Kush et al. 2019; Bondevik et al. 2021). The goal is to isolate any effects of an island violation that goes beyond potential processing difficulties involved with complex sentences. The design controls for two confounds that potentially put a strain on processing, and subsequently lower acceptability: (i) the length of time that a filler must be maintained in working memory before the gap is encountered (short vs. long), and; (ii) the complexity of the domain present in the sentence (no-island (declaratives) vs. island (domains claimed to be islands)). The idea is that domains claimed to be islands (e.g. adjunct clauses), irrespective of extraction, might be more complex to process than declaratives.

The 2 × 2 design crosses the two factors, each with two layers: Structure: no island vs. island × Distance: short vs. long. This yields four test conditions which together make up one test item. An example is provided in (7).

If the main effects illustrate linear additivity (i.e., no interaction effect), we will see that the decrease in acceptability is constant between the short and long conditions, and equally, that it is constant between the no-island and island conditions. This is illustrated in the interaction plot in Figure 1 under “No island effect”. Here, the lowered acceptability on the “long, island” condition can be explained by the linear sum of the processing costs.

Figure 1

Examples of interaction patterns.

If, however, the main effects illustrate a super-additive interaction, the effect of forming a filler-gap dependency into an island domain is larger than the sum of processing costs. This is termed an island effect and is illustrated in Figure 1 under “Island effect”. Here, the additional decrease in acceptability on the island violating sentence indicates that there is something outside of processing costs that causes an “unexpected” decrease in acceptability. Importantly, the effect is predicted to be directional such that the “long, island”-condition is rated as least acceptable.

3.2 Test material

We tested three different adjunct clause types in an rc-dependency, each introduced by a different complementizer om – conditional ‘if’, fordi – causal ‘because’ and når – habitual ‘when’. In addition, we included two control clause types for baseline comparisons: complex subjects, which have been shown to yield large and robust island effects in Norwegian, and complement om ‘whether’ clauses, which have been shown to yield small or no island effects in Norwegian in a top-dependency (Kush et al. 2019; Bondevik et al. 2021).

A relative clause is a clause in which the nominal phrase is associated with a position both in the matrix and the subordinate clause. Unlike Kobzeva et al. (2022) and Nyvad et al. (2022), we tested restrictive relative clauses in which the head noun is the object of the matrix verb. The most common type of restrictive relative clauses in Norwegian are som-relatives, introduced by the complementizer som (Åfarli 1994: 82).

(8) shows that the relative complementizer is obligatory in subject relative clauses, but not in object relative clauses. All target items were created with rc-dependencies forming restrictive relative clauses. For the object relative clauses, the complementizer som ‘who/which/that’ was included to maintain as much of the structure as identical as possible across subject and object relative clauses.

The test items were modelled on previous experiments with this design (Sprouse et al. 2016; Kush et al. 2018; 2019; Bondevik et al. 2021). Specifically, the items followed the structure in Sprouse et al. (2016) for testing island violations in an rc-dependency, where there are three clauses – a matrix clause, a relative clause modifying the object in the matrix clause and finally a finite adjunct clause embedded under the relative clause. The finite verb in each clause will henceforth be referred to as Vmatrix, Vrel and Vadjunct, respectively. An example item for om ‘if’ is provided in (9).⁹

The items are matched on several syntactic and semantic parameters that might influence acceptability. Every verb phrase is in the present tense, none of the embedded clauses are negated (Szabolcsi & Lohndal 2017), and every relative clause head is a definite DP. Finally, all adjunct clauses can be classified as Central Adverbial Clauses in the sense of Haegeman (2012) (see also Müller 2019). There are minor differences between items such as type of subject (e.g., indefinite determiners noen ‘someone’, full NPs studentene ‘the students’, general 3^rd person pronouns de ‘they’) in Vmatrix, Vrel or Vadjunct. This means that items are not minimally distinct, but items are matched on the features that have been suggested in the literature to be relevant for judgments of islandhood.

As pointed out to us by an anonymous reviewer, the conditions are not minimally different on two important aspects which potentially confound the results: (i) there are different lexicalizations across the different conditions, and (ii) on the short conditions, the gap is in the subject position of the relative clause. Regarding the first point, we believe that this fact is not detrimental since the different lexicalizations are the same for two and two conditions within each item. Thus, any effects of word choice will subtract (see Sprouse & Villata 2021). Turning to (ii), this means that there is a subject gap in the relative clause in the short conditions and an object gap in the clause embedded within the relative clause in the long conditions. Thus, the DISTANCE factor controls both whether there is a subject gap or an object gap and whether the filler-gap dependency is short or long. Given the subtractive logic of the 2 × 2 factorial design (see e.g., Sprouse 2016: 314), the main effect of DISTANCE can be attributed to the difference in length or the difference in argument structure properties. This design will not be able to distinguish between these two possibilities.

3.3 Participants

100 Participants were recruited through Prolific and offered 7 GBP for participation. The study was made available to all participants who registered “Norway” as their nationality on Prolific. A background survey collected data on language history and demographics. Participants were asked to briefly describe how to get to their closest bus stop. Here three participants were excluded for providing a written reply that did not comply with Norwegian written standards. Next, 14 participants who self-reported being ‘bilingual’ were excluded.¹⁰ In addition, among the 14 participants who reported living outside of Norway, we excluded five participants who reported having lived abroad for a long period of time and/or who reported rarely speaking Norwegian. Participants were rewarded regardless of their responses. Finally, we excluded three participants for having >5 responses with <1000 ms. reaction times. We consider <1000 ms. insufficient time to read and judge any of our test sentences. After the exclusion criteria were applied, 76 participants were included in our data set.

Out of 76 participants, 30 reported being in the 18–24 age group, 30 between 25–34, 12 between 35–44, 3 between 45–54 and one older than 65. Participants were also asked to report dialectal background. Dialects were grouped into 10 larger dialectal areas based on Mæhlum & Røyneland’s (2012: 179) map of dialectal areas in Norwegian. In addition, bergensk ‘Bergen-dialect’ and ingen av disse ‘none of these’ were added as possible responses. All dialectal areas were represented in the study, the most frequent response being østlandsk ‘Eastern Norwegian’ (40 responses).

3.4 Procedure

16 items were tested for each adjunct clause types (16 items × 3 clause types = 48 adjunct items), while 8 items were tested for each of the control clause types (8 items × 2 clause types = 16 control items). Items were distributed across 4 lists in a Latin Square procedure, such that participants only saw one condition per item. This left 64 test sentences in each list. Under the assumption that every island violating sentence is unacceptable, the ratio between acceptable and unacceptable sentences was 3:1 for the target test sentences within each list.

The experiment was designed to be balanced both with regards to the ratio of target to filler sentences, and acceptable and unacceptable sentences. The experiment included 64 fillers, of which 48 were created to be unacceptable fillers. The bad fillers included syntactic, semantic, and orthographic violations. The good fillers included relative clauses and finite adjunct clauses that differed from target sentences, e.g., non-restrictive relative clauses, other adjunct clause types. All fillers were used across all four lists. Test sentences and fillers were pseudo-randomized by list for every individual participant by condition.

The experiment was distributed via Prolific and run on JATOS with JsPsych (de Leeuw 2015). Following previous experiments using this design, the experiment was designed as an acceptability judgment task where each test sentence was presented alone. Judgments were given on a labelled 1–7 Likert Scale with end points given as 7 god ‘good’ and 1 dårlig ‘bad’ (i.e., a full Likert Scale as defined in Marty et al. 2020).¹¹

Inside the experiment, the background survey was presented first. Next, task instructions were given. Specifically, participants were instructed to imagine a context in which the sentence was uttered by someone in their own dialect. Moreover, the instructions specified that long sentences are not necessarily unacceptable and short sentences are not necessarily acceptable. An example of a grammatical, but long sentence was shown and rated 7, and an example of a short, but ungrammatical sentence, rated 1.

Two unmarked practice items initiated the experimentation phase: one was clearly grammatical, the second ungrammatical.

3.5 Data analysis

The data was analyzed using similar procedures as previous experiments following this design (e.g. Sprouse et al. 2016). The raw responses were z-score transformed by participant prior to analysis. Following Sprouse et al. (2016), there are three procedures for identifying island effects within this design: (i) a visual inspection of the relationship between conditions: a superadditive pattern vs. a linear additive pattern; (ii) a numerical identification process of calculating differences-in-differences scores (DD-scores) (see e.g., Sprouse & Villata 2021: 230 for a detailed explanation of the DD-score): a score above 0 is indicative of an island effect, while a score below 0 is characterized by Sprouse et al. (2011) as a reverse island effect, and; (iii) a statistical procedure fitting linear mixed effects models.

Data visualizations for visual inspection were created with ggplot2 (Wickham 2016). The size of the island effect for each island type was calculated with a DD-score.¹² Linear mixed effects models were fitted with lmer() from the lme4 package (Bates et al. 2015) in R (R Core Team 2021). An omnibus model was fit with a three-way interaction term crossing the main effects island type, distance, and structure. We included the three-way interaction term as we predict that the interaction of the main effects will differ by island type. By-item and by-participant varying slopes and intercepts were estimated as random effects. The model was simplified in a stepwise fashion to arrive at a model that converged without warning messages (though see Winter 2020: 266–267 for problems with such an approach). The categorical predictors were contrast coded –1 and 1. The omnibus model returns the results for the reference level (which is alphabethically set to fordi ‘because’) and the rest of the model must be interpreted in relation to the reference level. To measure the island effect for each specific island clause type, we also fit separate models for each island type with a two-way interaction term crossing the main effects distance and structure.

We also checked to see if there was satiation of judgments. Satiation is a term used to describe the “perception of acceptability after repeated exposures to the same sentence or the same structure” (Sprouse & Villata 2021: 242). Several studies on English have found that there are no satiation effects for adjunct islands (see overview in Sprouse & Villata 2021). Chaves & Putnam (2020), however, found satiation effects with 24 exposures to the same adjunct island structure. Moreover, they found that conditional adjunct clauses satiated at a higher rate than causal and temporal adjunct clauses. Given that participants were only exposed to 4 test sentences of the same structure in Experiment 1, we predict that we will not see any satiation effects for either adjunct clause type. Nevertheless, we want to exclude this as a potential source of variation. We looked for this in two ways: (i) we checked if the results in Experiment 1 replicated when only the first two responses to each condition were included in a partial data set.¹³ As participants were only presented with two test sentences per control clause type in the full data set, the control clause types are the same for partial and full data sets. (ii) Following Chaves & Putnam (2020), we fit linear mixed effects models for each of the adjunct island’s “long, island” condition crossing z-scores and trial index as main effects and fitting by-subject and by-item varying intercepts.

3.6 Results

The bad fillers received an average rating of z = –0.834, while the good fillers received an average rating of z = 0.859, both yielding narrow distributions of scores. Table 1 provides an overview of the main results of the omnibus model.

The omnibus model returned a significant interaction effect between the three main effects – island type, distance, and structure. In addition, there was a main effect of distance and structure. On the interaction term, the model did not distinguish between fordi ‘because’-adjunct clauses (= the alphabetically determined reference level), the når ‘when’-adjunct clauses and the subject-islands. There were, however, significant differences between the om ‘if’-adjunct clauses and the fordi ‘because’-clauses, and similarly between the ‘whether’-clauses and the fordi ‘because’-clauses. This indicates that the interaction of distance and structure is statistically significantly different between fordi ‘because’- and når ‘when’-adjunct clauses on the one hand, and om ‘if’-adjunct clauses on the other.

Looking at each island type separately, we ran separate linear mixed effects models for each island type and calculated DD-scores. We found significant island effects for all island types except for the control ‘whether’-clauses. For the ‘whether’-clauses only the main effect of distance was significant. The subject-island, the other control condition, yielded significant island effects and the largest effect size of all clause types. See Table 2 reports the results for the control clause types.

All three adjunct clause types yielded significant interaction effects. However, as the omnibus model indicated, there are differences between adjunct clause types: Fordi ‘because’ and når ‘when’ on the one hand show large DD-scores, while om ‘if’ shows a much smaller score. Table 3 provides an overview of the model output for each target clause type, while the interaction plot in Figure 2 visualizes the island effect and the effect size for each island type.

Figure 2

Interaction plot for all island types, Experiment 1 – average ratings on every condition for each clause.

We see that the average z-score for the “long, island” condition varies between island type, while the average ratings for the three non-island violating conditions are relatively stable across clause types. The average z-score on the “long, island” condition for the subject-island is low and for the ‘whether’ island it is high. Again, fordi ‘because’ and når ‘when’ pattern together with average ratings well below 0, while the island condition in the om ‘if’-items received average ratings just above 0.

Following findings for topicalization, we expect to see inter-trial variation, especially for om ‘if’ (Kush et al. 2019; Bondevik et al. 2021) and partly for når ‘when’ (Bondevik et al. 2021). We therefore investigated the distribution of z-scored ratings for each condition for each island type. In Figure 3, the distribution of z-scored ratings for each condition for each clause type is plotted.

Figure 3

Comparing the distribution of z-scores on the no-island and the island conditions for the long and short conditions separately, Experiment 1.

We see a unimodal and quite narrow distribution for the ‘whether’ island condition. The distribution of scores for the “long, island” condition largely overlaps with the distribution for the “long, no-island” condition, where scores predominantly fall well above 0.¹⁴ On the “long, no-island” condition, there is a mostly unimodal distribution around –1 for the subject island.

Again, the distribution of scores is similar between fordi ‘because’ and når ‘when’ on the “long, island” condition, such that the majority of scores fall below 0. However, the leftward tail for når ‘when’ is wider than for fordi ‘because’, indicating that there is some variation between trials for når ‘when’ that is not observed for fordi ‘because’.¹⁵

The ratings for om ‘if’ have a wide, bimodal distribution: the biggest cluster of scores falls above 0, and a smaller cluster of scores below 0. The distribution of scores on the “long, island” condition resembles the distribution of scores on the “long, no-island” condition, but there is more variation for the “long, island” condition.¹⁶

Investigating the raw scores, we see the same pattern that we do for the z-scored ratings. In Figure 4 we see that om ‘if’ is different from the two other adjunct clause types – while fordi ‘because’ and når ‘when’ resemble the subject clause type, om ‘if’ resembles the ‘whether’-clauses.

Figure 4

Barplots displaying the count of raw responses per condition for Experiment 2.

Checking for satiation effects, we find the exact same pattern for the partial data set that we find for the full data set (see Supplementary file). The omnibus model returns a significant interaction effect, main effects of distance and structure. The model finds om ‘if’- and ‘whether’-clauses to be significantly different from the reference level (fordi ‘because’). Running a linear mixed effects model modelling z-score on the island violating condition by trial index for each adjunct clause type reveals that there is a significant effect of z-score by trial index, but that as Sprouse & Villata (2021) point out, it is very small across adjunct clause type, see model output in Table 4.

This means that for each repetition, the z-score is predicted to rise by > 0.005 for each of the island conditions. As we presentented participants with 4 repetitions of the same structure, we exclude satiation as having any effect on ratings.

In the plots in Figure 5 (based on Chaves & Putnam 2020), we see judgments for items by block for each of the adjunct clause types. Block 1 contains the first two responses given to a certain condition, and block 2 the last two. We do see differences between blocks, such that some items show an increase in acceptability from block 1 to block 2. However, we also see instances of a decrease in acceptability between blocks. We understand this to mean that overall there is a slight increase in acceptability as the experiment proceeds, but as the model demonstrates, the increase is very small.

Figure 5

Boxplot illustrating average judgments on the “long, island” condition by item for each island type. The dashed line highlights the border between adjunct clause types. The plot legend provides the explanation of the colors.

3.7 Intermediate summary

Experiment 1 reveals that rc-dependencies are sensitive to island constraints in Norwegian. Collapsing across island types, we find island effects of forming a relative clause dependency into these domains. Fitting separate models for each island type, we find statistically significant island effects for all adjunct clauses and for the subject island, while the ‘whether’-island did not yield any significant interaction effects. As such, findings for the control island types replicate previous findings for top-dependencies in Norwegian (Kush et al. 2019; Bondevik et al. 2021).

Though we find island effects across the three adjunct clause types, we see clear indications that fordi ‘because’, når ‘when’ and om ‘if’ do not behave like a group in rc-dependencies. We find statistically significant differences between fordi ‘because’ and når ‘when’ on the one hand, and om ‘if’ on the other. While om ‘if’ shows a small island effect size, z-scored ratings clustering above 0 and a distribution of scores indicating variation between trials, fordi ‘because’ and når ‘when’ show large island effect sizes and z-scored ratings clustering well below 0. Thus, our findings substantiate Bondevik et al.’s (2021) and Nyvad et al.’s (2022) findings: adjuncts do not behave like a uniform group with regard to islandhood.

As previously discussed, many theories of islands predict that there will be a categorical split between islands and non-islands, such that islands should be clearly unacceptable, while non-islands should be clearly acceptable. To that end, the intermediate island effect that we see for om ‘if’ is problematic for these theories. Om ‘if’ seems to fall in an intermediate position between acceptable (null effects) and unacceptable (large island effects). Thus, we need some way of accounting for om ‘if’.

One possible interpretation of the intermediate effect size is that intermediacy is caused by averaging over variable results. The other studies testing om ‘if’ in Norwegian report substantial variation between trials. We see indications of this too in the distribution of scores for om ‘if’ on the “long, island” condition. Kush et al. (2019) suggest that the variation might be caused by inconsistent raters, i.e., either between- or within-speaker variation. Another option implied by Bondevik et al. (2021) is that there is variation between items. However, Bondevik et al. (2021) fail to find any factor across items that can explain said variation. If om ‘if’ sporadically induces island effects depending on certain factors (that we have yet to identify), which yield intermediate effects when averaged over, om ‘if’ is an adjunct type that variably causes large or small-to-nonexistent island effects. Such an interpretation predicts groupings of judgments on either side of the scale.

Another possibility is that the intermediate result we uncovered for om ‘if’ is a true representation of the acceptability of extraction from om ‘if’. This means that extraction from om ‘if’ is systematically judged to be less acceptable than extraction from embedded complements (‘whether’ and declarative-clauses) and systematically more acceptable than extraction from fordi ‘because’ and når ‘when’ clauses. If this is true, we predict that there will be normal distribution around an intermediate score, i.e., variation between trials will be within the expected range.

In order to classify om ‘if’ with regard to islandhood, it is important to understand the source of the intermediate effects. Experiment 1 does not reveal much about the source of the intermediate effect. Thus, we carried out a follow-up experiment where we controlled for between- and within-speaker and -item variation.

4.1. Test material

In Experiment 2, only om ‘if’ was tested with the same exact 16 items as were tested in Experiment 1. We also re-used the fillers.

4.2 Participants

100 participants completed the study. The exclusion criteria applied in Experiment 1 were also applied in Experiment 2. Six participants were excluded for reporting being bilingual. One participant was excluded for failing to report being a native Norwegian speaker. 37 participants were excluded for having >5 responses below 1000 ms. We characterized these respondents as “false respondents” as they typically had >50 responses below 1000 ms.

In total, 56 participants were included in the data material. Out of 56, 49 participants reported being aged between 18–24. All dialect groups were represented, with the most frequent reply being østlandsk ‘Eastern Norwegian’ (14 responses).

4.3 Procedure

The study followed the same procedure as Experiment 1, with two exceptions. First, items were not distributed across different lists. The Latin Square distribution of test sentences in Experiment 1 makes it impossible to distinguish participant variation from item variation. To control for this, every participant was presented with all test-sentences in experiment 2 in the exact same randomized order. Such a design allows us to control for (i) participant effects, which will be the same across items, (ii) item effects, which will be the same across participants, and finally (iii) potential ordering effects, which will be the same across items and participants. Participants saw 64 (16 × 4) test sentences for om ‘if’, 64 fillers (48 bad, 16 good) and 2 unmarked practice sentences.

Second, participants were recruited through NTNUs internal student platforms and one external student’s social media platforms. We think it is highly unlikely for someone to have participated in both Experiments 1 and 2. Participants received monetary reward for completing the study (150 NOK).

4.4 Data analysis

Data analysis was conducted as for Experiment 1. A linear mixed-effects model was fit with a two-way interaction term crossing the main effects distance and structure. We also fit a linear mixed effects model that included item as a fixed effect in an interaction with distance and structure. Here the model makes item 1 the reference level, and the model outputs must be read in relation to this reference level. We calculated by-participant DD-scores aggregated over all items and by-item DD-scores aggregated over all participants. As we did for Experiment 1, we checked for satiation effects. As test sentences were given in the same order across participants, satiation effects are conflated with potential item effects. Thus, we will not rely too heavily on any results of these analyses here. We ran a model for target conditions, modelling z-scores by trial index, with by-subject and by-item varying intercepts. We also ran separate models for each target condition and for bad fillers, checking whether trial index co-varied with z-scores for each condition. Here, we also fit by-subject and by-item intercepts. Based on the evidence in Chaves & Putnam (2020) for conditional clauses, since participants were exposed to 16 island violating conditions we hypothesized that we would see some evidence of satiation for the “long, island”-condition.

4.5 Results

4.5.1 Overall results

The bad fillers received low ratings, and the good fillers received high ratings. Table 5 provides an overview of average ratings for each condition included in Experiment 2.

Table 5

Overview of average ratings (z-scored) and standard deviations for every condition, Experiment 2.

Condition	Mean z-score	SD
Bad fillers	–0.896	0.726
Good fillers	0.761	0.666
Short, no-island	0.736	0.569
Long, no-island	0.675	0.581
Short, island	0.484	0.673
Long, island	0.031	0.762

The linear mixed effects model with a two-way interaction between distance and structure returned a significant interaction effect, in addition to significant main effects of distance and structure (see Table 6). The model indicates, through the size of t, that the main effect of structure is greater than the effect of distance. We also see an intermediate effect size, and an average z-scored rating of the “long, island” condition just above 0. This implies that om ‘if’ yields intermediate island effects in Norwegian, as can be visually confirmed in Figure 6. As such, Experiment 2 replicates Experiment 1.

Figure 6

Interaction plot for om ‘if’, Experiment 2.

Table 6

Main results of the linear mixed effects model, Experiment 2.

	Estimate	SE	t	p
Intercept	0.030	0.095	0.308	0.791
distance: short	0.454	0.105	4.320	<0.001
structure: no-island	0.645	0.085	7.558	<0.0001
distance × structure	–0.392	0.106	–3.680	0.002

We also investigated the distribution of z-scores on the four conditions, which shows that there is more variation on the “long, island” condition compared to the three baseline conditions. For the three baseline conditions there is a narrow distribution around z = 1, with a thin right-ward tail indicating some variation. For the “long, island” condition, however, we see a wide distribution.

The density plot in Figure 7 shows that a portion of scores on the “long, island” condition overlaps with the “long, no-island” condition, indicating that for some portion of the trials, the “long, island” condition is indistinguishable from the “long, no-island” condition. An analysis with overlap() from the overlapping-package in R (Pastore 2018) shows that these distributions are 44% different (following the procedure detailed in Pastore & Calcagnì 2019). This means that the distributions of the scores for the “long, no-island” and the “long, island” conditions are more similar than they are different.

Figure 7

Comparing the distribution of z-scores for om ‘if’ on the no-island and the island conditions for the long and short conditions separately, Experiment 2.

Comparing Figure 7 to the distribution of scores on the bad and good fillers, we see the way in which scores are distributed for two conditions that are consistently distinguished by participants.¹⁷ Figure 8 shows that there is only marginal overlap between z-scores for the filler conditions, meaning that the fillers were consistently distinguished across trials. An overlap analysis finds that they are 85% different.

Figure 8

Distribution of z-scores for the fillers, Experiment 2. The bad fillers show a narrow distribution around –1.5. The good fillers show a narrow distribution around 1.

We see here that for participants om ‘if’-adjuncts are not unacceptable in the same way as the bad fillers, nor acceptable in same way as the good fillers.

Looking at satiation effects, we ran a similar linear mixed effects model investigating the effect of trial index on z-score as we did for the “long, island”-conditions in Experiment 1 (see Table 4). An overview of model outputs is provided in Table 7. We see an overall satiation effect across conditions, but the estimate is very low. With an estimate of 0.0025, each new test sentence will see a very small increase in rating (across all conditions), which means that after being exposed to 64 test sentences a z-score of e.g. z = 0.2 will increase to z = 0.34. Fitting models for each condition separately, we do not see a significant increase in rating as the experiment proceeded.

Table 7

Overview of results from linear mixed effects models testing for satiation, Experiment 2.

	Number of data points	Estimate	t	p
Overall	3570	0.0025	6.440	<0.0001
Short, no-island	892	0.0005	0.289	0.776
Long, no-island	891	0.0020	1.348	0.1989
Short, island	896	0.0007	0.340	0.7388
Long, island	891	0.0007	0.286	0.779
Bad fillers	2677	0.0024	1.352	0.183

Separating the responses on the “long, island”-condition into four blocks (the first four responses in block 1, etc.), we see the same pattern that we see in the model for this condition, i.e., no indication that late blocks are rated better than earlier blocks. This is illustrated in Figure 9 below.

Figure 9

Boxplot illustrating the average judgments on the “long, island” condition by item. The different shades of blue indicate the different block numbers.

4.5.2 Results – variation

The average results for om ‘if’ are in the intermediate range. However, in the distribution of scores we see variation between trials. The distribution of scores is wider than the distribution for any of the filler and other target conditions. Therefore, we want to investigate this variation more closely to see if there are any meaningful patterns either between participants or between items. If so, we expect to see grouping of participants and/or items.

First, we looked at variation between items. We fit a linear mixed effects model on our data in a three-way interaction between item, distance, and structure. The model did not return a significant interaction effect and found only a significant effect of structure. The model returned significant differences between item 1 (reference level) and several items, but there were also several items that were found not be distinguishable from item 1 (see Supplementary file).

Visually inspecting the items in an interaction plot in Figure 10, it is clear why the model did not return a significant interaction effect, nor significant main effects when item 1 was set as the reference level. For item 1, there is only minimal linear additivity between conditions, which is reflected in a DD-score of –0.09. Linear additivity and DD-scores close to 0 are the common denominators for items that the model did not distinguish from item 1. In comparison, the items that were found to be distinct from item 1 show super-additivity. These also have DD-scores well above 0. There is, however, variation in the size of the DD-score between the items that the model distinguished from item 1.

Figure 10

Interaction plot by item for om ‘if’, Experiment 2. The items that the model did not distinguish from item 1 are labelled “Not different”, while the items that the model did distinguish from item 1, “Sign. different”.

Looking at each item separately in this manner, we see that there are differences between items. As participant- and ordering-effects are kept constant across the experiment, the variation can in fact be attributed to item variation. Nevertheless, investigating the distribution of the DD-scores by items aggregrated over participants in Figure 11b, we see that there is in fact normal distribution (with a positive skew) around an intermediate score. In other words, we do not see indications of item grouping. This suggests that the variation we see in Figure 10 might be random variation that we can expect to see by chance.

Figure 11

DD-scores calculated by participant across 16 items (a) and DD-scores calculated by item across 56 participants (b), Experiment 2. Histograms are plotted with geom_histogram(), boundary = 0, binwidt = 0.25.¹⁸

This does not exclude the possibility that there is variation at the participant level. The design allows us to calculate DD-scores for each participant aggregated over the same 16 items. This means that we have a large sample of items that make up the average DD-scores per participant. As such, if we see differences between DD-scores we will assume that these reflect real differences between participants. Investigating the range of DD-scores in a histogram we see that there is a wide range of DD-scores ranging from an average score below 0 to an average above 1. However, the histogram in Figure 11a shows that participants’ DD-scores are widely, but normally distributed around the average DD-score (DD = 0.39). In other words, we do not see signs of participant grouping. Accordingly, we do not see indications in the variation between items or between participants that the intermediate effect is caused by aggregating over variable judgments.

Importantly, intermediate scores are also represented in the raw ratings.¹⁹ Looking at the raw ratings by condition in Figure 12, we can recognize the pattern of the z-scored ratings. They tell us that participants use the full range of the scale, but that the most frequent responses are in the intermediate range. We also see that there is a large portion of ratings on the “long, island” condition at 7, i.e., the highest score possible. In terms of absolute ratings of an island violating sentence, this tells us that for some items some participants did not find these island violations to be unacceptable. Comparing the raw scores of the test sentences to the fillers in Figure 12, we see that there is a larger proportion of intermediate ratings for the “long, island” condition than for bad fillers and fewer high ratings than for good fillers.

Figure 12

Barplots displaying the count of raw responses per condition, Experiment 2.

4.6 Intermediate summary

In Experiment 2 the on average intermediate island effects of forming an rc-dependency into om ‘if’-adjuncts in Norwegian were replicated. The results replicate in a design where participants see every test sentence, as opposed to distributing items in a Latin Square Design. Thus, it seems that the number of exposures to lexicalizations of the same test conditions does not influence acceptability. Overall, we find a significant interaction effect, a super-additive judgment pattern (see Figure 4), a DD-score of 0.39 and an average rating of the “long, island” condition just above 0. Though we see variation both at the item and participant level, there is normal distribution around an intermediate effect size. Such a distribution of DD-scores indicates that the average intermediate results for om ‘if’ do not conceal meaningful variation between items and/or participants or order of exposure. Thus, it seems that the intermediate effect is not caused by (the most obvious) extra-grammatical factors. Accordingly, the intermediate results for om ‘if’ seem to reflect the accurate underlying acceptability pattern for this adjunct clause type.

5.1 Do adjunct clauses behave like a uniform group for relativization?

Following up on Bondevik et al.’s (2021) results where finite adjunct clauses introduced by fordi ‘because’, når ‘when’ and om ‘if’ did not behave as a uniform group for top-dependencies, the present study finds that these three adjunct clauses do not behave like a uniform group for rc-dependencies either. First, the linear mixed effects model did not distinguish between fordi ‘because’ and når ‘when’, but distinguished fordi ‘because’ and om ‘if’. This indicates that fordi ‘because’ and når ‘when’ received judgments that, on average, were similar enough to accept the null hypothesis that these behave alike in rc-dependencies. In addition, judgments on the “long, island” condition are similarly distributed around a negative z-score across the two adjunct clause types. There is slightly more variation in the scores for når ‘when’ than is seen in fordi ‘because’. Om ‘if’, on the other hand, yields smaller DD-scores across experiments 1 and 2, compared to fordi ‘because’ and når ‘when’. In addition, ratings of the “long, island” condition fall above 0. In other words, om ‘if’ yields intermediate results.

The data clearly shows that there are systematic differences between the adjunct clause types. accoringly, we need our theory of adjunct constraints to explain these differences. However, there are no traditional syntactic approaches to islands that can readily accommodate the distinction between adjunct clause types which our, Bondevik et al.’s (2021) and Nyvad et al.’s (2022) results necessitate. Comparing the recent findings, we find systematicity in the extraction patterns: Causal clauses (tested with because) and habitual clauses (tested with when) yield low ratings and island effects across three dependency types and across two languages (Norwegian: top, rc; English: dem rc, simple and complex wh). Conditional clauses (tested with if) yield high ratings, more closely resembling declarative clauses than the other adjunct clause types across dependency types and languages (Norwegian: top, rc, dem rc²⁰; English: dem rc, and no island effect in rc). We believe that the systematicity in the recent findings indicate that there is some identifiable and general constraint that governs extraction from adjunct clauses. However, none of the current theories provide a ready-made solution to this puzzle. Thus, we must explore more fine-grained versions of the current theoretical approaches.²¹

Investigating the semantics of each complementizer, we see clear differences between adjunct clause type. In (10) the same sentence is presented with the three different complementizers to allow for an easy comparison of the meaning of each complementizer.

Om ‘if’ introduces a conditional clause. The om ‘if’-clause specifies a condition, and the clause that it modifies conjectures an outcome of the fulfillment of the condition (Hornstein 1990: 74). A causal relationship between the conditional and relative clause is implied – if the condition is not satisfied, the event in Vrel (finite verb in relative clause)²³ might still occur, but not for the reason expressed in the conditional clause. Når ‘when’ introduces a habitual clause. The meaning of når ‘when’ in this use is “every time the event expressed in Vadjunct (finite verb in the adjunct clause) occurs, the event in Vrel also occurs”. It is presupposed that both the event in the relative clause and in the ‘when’-clause have minimally occurred once. A causal relationship between the adjunct and the relative clause is implied. Fordi ‘because’ introduces a causal clause that explicitly expresses the cause of the event in Vrel.

As each complementizer contributes different meanings to the sentence, it is possible that each complementizer conditions the adjunct clause’s opacity differently. As semantic conditions have been shown to govern extraction from non-finite adjunct clauses (Truswell 2011; Ernst 2022), it is not improbable that there are semantic conditions on finite adjunct clauses as well (see also Abrusán 2014 on semantic conditions in weak islands). Truswell (2007; 2011) proposes the semantic condition on adjunct islands that wh-extraction is only possible if two events can be construed as one event (2011: 157). This is captured in the Single Event Grouping Condition (SEGC). For events decribed in different clauses to be construed as a single event (i) the events described in the two clauses must have spatiotemporal overlap; and (ii) there can be maximally one agentive verb. Spatiotemporal overlap means that the event grouping must “[…] happen in a single place, as well as in a single time” (Truswell 2011: 48). Accordingly, Truswell’s SEGC can make distinctions between adjuncts that are structurally the same, but differ in their semantics. Truswell (2011) does not explicitly extend the SEGC to non wh-dependency types.

Truswell (2011) specifically shows that his condition does not apply to finite adjunct clauses, arguing that a finite operator blocks extraction from finite adjuncts regardless of the SEGC (2011: 118). Extending Truswell’s approach, Ernst (2022) relaxes the complete ban on extraction from finite adjunct clauses.²⁴ Ernst (2022) assumes that non-finite clauses also include a tense operator, and thus, he rejects that a tense operator in and of itself blocks movement from finite adjunct clauses.

Ernst’s (2022) extension of Truswell’s (2011) SEGC has the potential to explain the difference between adjunct clause types without additional machinery: The three different sentences, on their most natural readings, imply different temporal relations between Vrel and Vadjunct. While temporal location of the event time in either the relative or the conditional clause is possible with om ‘if’, temporal location of two separate event times is possible for both fordi ‘because’ and når ‘when’. Thus, although there is grammatical tense in the om ‘if’-clauses, the lack of temporal interpretation means that there are not two “independently determined” times that the clauses can be associated with, in fact the event times are undetermined. It is possible, following Ernst (2022), that extraction is facilitated as there is only one determined time (Vmatrix) in the om ‘if’-items. It would be interesting to test extraction from om ‘if’-adjuncts where both the event in Vrel and Vadjunct can be temporally located.²⁵

There is, however, nothing within this theory that can explain the robustness of the intermediate results for extraction from conditional clauses.

5.2 Are adjuncts islands for relativization in Norwegian?