Glossa Psycholinguistics

5.1.1 A quick glance at frequency

It is theoretically possible that the difference in acceptability between the complementizers could simply be the result of collocation, such that participants found extraction from hvis (‘if’)-clauses better than that from fordi (‘because’)-clauses due to e.g. overrasket (‘surprised’) occurring more frequently with hvis than because. We looked up the collocations between the predicative adjectives in the study and the four different complementizers that they co-occur with in the stimuli in the Danish written language corpus KorpusDK; see Table 13. The total number of occurrences of the set of predicative adjectives was 5,547, which means that, overall, they occurred with one of the four complementizers 18.3% of the times.

The numbers in Table 13, though significantly different (χ² = 350.96, df = 3, p < 0.0001), do not explain the attested acceptability differences across complementizers. Overall, at ‘that’ is most frequent and most acceptable, and fordi ‘because’ is least frequent as well as least acceptable. However, da ‘when’ is more frequent but less acceptable than hvis ‘if’. (For a detailed discussion of frequency, acceptability, and complexity, see Christensen & Nyvad, 2024.)

We also searched the corpus to see if there were examples with any matrix predicative adjective (not just those in our study) followed by any preposition followed by hvis/da/fordi, i.e. examples where the adjunct clause was realized as a complement clause under a preposition, parallel to clauses with at. This search, however, resulted in only a single example with hvis (zero hits for da and fordi), which is in line the standard analysis of hvis, da, and fordi as adjunct complementizers.⁸

5.1.2 No effect of context

One might speculate that the variation in acceptability of extraction across the different complementizers may be dependent on context. Keller (2000) argues that grammatical constraints can be divided into those that are affected by the presence of a context (and which should be in effect crosslinguistically) and those that are not. As pointed out by Sprouse (2007, p. 48), many frameworks have analyzed the context-dependent constraints as being essentially semantic or pragmatic in nature, whereas context-independent constraints can be interpreted as relating to syntax (or grammar) proper. This line of reasoning leads Erteschik-Shir (2006), among many others, to argue that context-independent grammatical constraints should yield a binary output (i.e. either grammatical or ungrammatical). While Sprouse (2007) found context not to interact with island structures involving wh-movement, several other studies have indeed found a positive effect of supplying a facilitating context on extractability (Kristensen et al., 2013; Kush et al., 2019).

In the present study, however, there was no significant effect of context (cf. Table 10). Though it is possible that we might have seen an effect with a pronominal filler instead of a full DP (e.g. den ‘that’ instead of den prøve ‘that test’), and though it could be slightly more unnatural to repeat a full DP in the target sentences when it has already been mentioned in the context, the lack of significant context effects is remarkable and suggests that the variation found between complementizers is fairly independent of this pragmatic factor.

5.1.3 No effect of dependency type

Another unexpected result was the lack of effect on the overall acceptability patterns when changing the dependency type from relativization to topicalization. Importantly, the dependency lengths in the relativizations, (3), are longer than in the topicalization structures, (6), and hence there is a difference in complexity between the two, and the consequent difference in processing load is not reflected in the acceptability levels. (This may, in part, be explained by the fact that the corresponding baselines for the topicalization study are structurally simpler.) Sprouse et al. (2016) found an effect of dependency type for if-clauses in English, such that relativization was more acceptable than wh-movement, and they argue that dependency type is key to the understanding of the variation in the acceptability of extraction from islands. However, we found no difference between topicalization and relativization. As shown in Figure 6, if, when, and because are equally unaffected by the absence of context.

On the one hand, island effects have traditionally been thought to be due to universal constraints on A-bar-movement (Chomsky, 1973, 1977, 2000; Cinque, 1990; Huang, 1982), and as such, we should, in principle, expect topicalization to be as unacceptable as wh-movement from islands. On the other hand, in the MSc. languages, it is more common to front topical DPs than wh-phrases (Lindahl, 2022). In addition, topicalization generally receives a higher acceptability rating. Furthermore, both relativization and topicalization often pattern together, compared to question formation (see e.g. Williams, 2011), a connection that has also been highlighted by Kuno (1976) and Abels (2012). Douglas (2016, p. 147) even argues that relativization and topicalization share a discourse-linking feature absent in wh-questions. Indeed, Müller and Eggers (2022) showed that relative clause extraction is more prevalent than adjunct clause extraction in both Danish and English. While they found more examples of topicalization out of relative clauses than relativization, the prevalence of relativization and topicalization out of adjunct clauses was at the same level in Danish. This picture fits our experimental results, which do not suggest a distinction between the two dependency types when it comes to extractability from adjunct clauses.⁹

5.1.4 The lexico-semantics of the complementizers

Bondevik et al. (2020) replicated the results of Kush et al. (2019) by finding island effects in connection with topicalization from temporal and causal adjunct clauses, but not conditional clauses. However, in our study, even though we appear to find the same overall pattern, the temporal adjunct clauses introduced by når in the Norwegian study differed in their island effect from the causal adjuncts (when > because), whereas our temporal and causal adjunct clauses were on a par in the raw scores. The DD scores, however, show that the island effects for da and fordi are actually different (see Figure 3 (D)). This may be related to the fact that our temporal adjuncts were introduced by da, not når (both translatable to ‘when’). Danish, like Norwegian, is a “two-‘when’ language” (Vikner, 2004, p. 133). Whereas da covers episodic clauses describing a unique identifiable eventuality in the past, når can be used in both past and present habitual clauses, as well as for future clauses. Simply put, the temporal da-clause presupposes that there is only one situation which would correspond to the description it expresses, which is not the case with når in the past (see the detailed discussion in Vikner, 2004). However, like when in English, da also has an explicit causal reading, and in the stimuli of the present study, it is not possible to completely distinguish between the temporal (informal) and the causal (formal) readings.¹⁰ But then again, the density plots in Figure 4 do not indicate ambiguity, because the distributions are not bimodal or spread-out. It may indeed be possible that all participants got a causal interpretation on top of the temporal one.

According to Truswell (2011, p. 44), causation between matrix and embedded clause may be a crucial factor in extractability from adjunct clauses. Similarly, Chaves and Putnam (2020) argue that the degree of extractability from adjunct clauses is correlated with Van Valin’s (2005, pp. 208–209) “inter-clausal relations hierarchy”, which states that causal relations express more cohesion or a tighter semantic connection than purely temporal ones. However, this causative connection was a potential in all our adjunct clause types, both in the English study (Nyvad et al., 2022) and the present Danish one, given that they were embedded under emotive factive predicates which elicit a causal reading, and, as such, extraction from all three adjunct clause types should be equally licit, contrary to our results. It could be argued, however, that for the da-clauses, a causative reading was not forced, and that the between-participant and between-item variation may be partially explained by this. However, it does not appear as though this factor is determinant in the variation that we have found overall, although a causative link between the adjunct clause and the matrix may be one of the preconditions for extraction (Bondevik et al., 2020).

Our data support recent proposals that the notion of gradience should be properly operationalized and integrated into accounts of island phenomena in a falsifiable way. Our data suggest that variation attested across adjunct clause types may be due to the lexical semantics of the complementizers interacting with factors relating to information structure and discourse function.

5.1.5 The effect of information structure/discourse function

Information-structural and discourse-functional accounts of island phenomena have long held that the gap in filler-gap dependencies generally must occur within the potential focus domain of the sentence or in the part of the utterance that is “asserted”, “focused” or “dominant” (Deane, 1991; Erteschik-Shir, 1973; Van Valin, 1996). As a result, adjunct clauses are not necessarily expected to exhibit uniform island effects. The degree to which an adjunct clause can be argued to have this special status may interact in intricate ways with a range of lexical and semantic factors. In addition, we would argue, it is possible that this “privileged” status that facilitates extraction is a product of an interplay between the semantics of the complementizer and the dependency type.

Fillers in long-distance dependencies typically occur in “discourse-prominent slots” (Goldberg, 2005, p. 135), as extraction, in and of itself, places a discourse element in a prominent position at the left edge of a clause. Adjunct clauses typically provide circumstantial information about the main event which may cause them to be considered non-essential in relation to the situational knowledge described by the matrix predication, which may, in turn, affect extractability.

According to Goldberg (2005, p. 135, (14)), it is pragmatically anomalous for an element to be both discourse-prominent and backgrounded at the same time, and she formulates the following discourse generalization: Backgrounded constituents may not serve as gaps in filler-gap constructions (Backgrounded Constructions are Islands, henceforth BCI) (see also Cuneo & Goldberg, 2023). She argues that such backgrounded constituents are characterized by being neither “primary topic” nor part of the “focus domain” of the sentence, but rather by being presupposed. Presuppositions, she asserts, can be revealed through negation tests, as they should be implied by both the positive and negative form of the sentence (Goldberg, 2005, p. 130) (for a critical review of the negation test, see Pérez-Leroux & Kahnemuyipour, 2014, pp. e121–e122). The BCI, however, does not predict extractability to vary as a function of the dependency type involved, unlike Abeillé et al.’s (2020) Focus-Background Conflict Constraint (FBCC), according to which a backgrounded constituent cannot function as a gap for a filler-gap dependency when the filler is focalized, because that leads to a discourse-functional “clash” between filler and gap. This means that variation between dependency types is expected under the FBCC: Wh-extraction from a backgrounded constituent should be unacceptable, while relativization should be acceptable, since a relative clause typically just adds information about a “given” NP and does not focalize the filler (i.e. the correlate), and any type of fronting not involving focalizing should be equally unaffected by the FBCC. As also pointed out by Lindahl (2017, pp. 160–161), the information-structural/discourse-functional approaches appear to face a problem when it comes to cleft structures, given that the clefted constituent is focused and the negation test identifies the embedded clause as presupposed (hence, as an island), even though this type of construction generally seems to allow extraction in the MSc. languages.

The three adjunct clause types in our stimuli have distinct semantico-pragmatic profiles, and they are not equivalent in terms of discourse function: The temporal and causal clauses are presupposed, while the conditional clause is not. Hence, what our study appears to demonstrate is that the BCI (or FBCC) would have to be gradient in order to account for the variability in the acceptability patterns attested, such that there is a spectrum of backgroundedness (as indeed suggested by Abeillé et al., 2020; Goldberg, 2013; Winckel et al., 2025) and this continuum interacts with the possibility of extraction from e.g. an adjunct clause. As pointed out by Schütze et al. (2015, p. 35), however, the degree of backgroundedness of a particular constituent is “not fixed for a given sentence” and may vary significantly as a function of the containing discourse. Hence, we need a way to operationalize the concept of backgroundedness, which to our knowledge does not yet exist. For example, though Abeillé et al. (2020, p. 21) argue that the FBCC is gradient, they say “a gradient discourse model is outside the scope of this paper”. (But see Cuneo & Goldberg, 2023.)

The adjunct clauses in our stimuli were constructed such that they express a cause or condition for the psychological state expressed by the matrix clause, and, as such, the adjunct clauses express at-issue information for the main action (Chaves & Putnam, 2020, p. 91). Following Simons (2007), Kush et al. (2021) suggest that extractability may instead depend on the clause containing the gap contributing “wholly or partially new information” (Kush et al., 2021, p. 39), or being the main point of utterance (MPU) of the sentence. An embedded clause, according to Simons (2007), can be MPU, especially if the predicate in the matrix clause expresses how the speaker feels about the information in the embedded clause. This is reminiscent of Kluender (2004, pp. 121–122), who argues that the filler’s relationship with the predicate in the main clause is as important as the relationship that it has to the gap position.

5.2.1 Processing differences?

Island phenomena have been explained by means of processing-based accounts, dating back (a least) to Pritchett (1991) and Kluender (1991) (see also Fodor, 1983, p. 310; Ross, 1987, p. 190). Kluender and Kutas (1993) argued that island effects may be the result of memory load exceeding capacity in connection with the processing of complex structures in e.g. adjunct island configurations. More recent processing-based accounts have been proposed, which draw on the importance of factors such as “accessibility” (Arnold, 2010), “predictability” (Hale, 2001; Levy, 2008) and “surprisal”’ (Culicover et al., 2022). Overall, from a processing perspective, the relatively low frequency of island structures (Lindahl, 2017; Müller & Eggers, 2022) would predict low levels of acceptability for these configurations, and given that islands involve higher syntactic complexity, they should incur a higher processing cost (Christensen & Nyvad, 2024).

These processing accounts seem to struggle, however, in the face of the overall crosslinguistic variation found in English and Danish, assuming that the island configurations, which are presumably relatively uniform across the two languages, also incur more or less the same processing load. Nevertheless, one could imagine that the three different types of adjunct clause may vary in processing difficulty due to other (semantic or pragmatic) factors, not just within but also across the languages. However, to the best of our knowledge, no processing account has been proposed which is able to capture and operationalize variation across constructions and across languages.

In a broader sense, it could be argued that island extraction might be relatively common overall in the MSc. Languages, because of the frequency of topicalization structures in these languages, cf. 4.2 above. (See also Engdahl, 1997; Erteschik-Shir, 1973, pp. 126–127.) Though frequency in itself is not a predictor (it is what needs to be explained), the fact that topicalization is so widespread and common in Danish but marked in English might suggest that it is easier to extract from at least some of the configurations in Danish that are more island-like in English. In this light, it is even more surprising that the acceptability ratings for island extractions are lower in Danish than in English.

From the perspective of a processing account, the lack of a difference between topicalization and relativization out of adjunct clauses in our Danish study is surprising. For example, Gibson (1998) and Lewis & Vasishth (2005) would not make a distinction between relativization and topicalization as such, since it is the length of the dependency (distance and/or number of discourse referents between filler and gap) that is important in their frameworks. However, in our study, the topicalization structures are shorter and the dependency lengths are longer in the relativizations. In other words, the lack of a difference goes against processing accounts based on dependency length.

5.2.2 Effect of translation

The stimuli employed in the Danish study were translated as closely as possible from the English stimulus set in Nyvad et al. (2022). However, in trying to keep the stimulus sets in the two languages as minimally different as possible, structurally as well as lexically, the formality level may have been inadvertently affected, potentially leading to differences in acceptability, such that the Danish version of the original English stimulus set is more stilted or formal, at least to some participants.

We attempted to address this question, at least in part, by running the experiment again without the context to see if the crosslinguistic difference might be an effect of the context somehow being worse in Danish than in English. The context was intended to simulate a turn in a conversation where someone is talking about a certain event and we thus aimed for a low formality level, which is why we modelled the actual target sentences on naturally occurring examples found in corpora.

If it is indeed the case that the translation introduced an unintended stiltedness in the Danish stimuli, then it is not surprising that the unextracted versions of the stimulus sentences in Danish appear to be on a par with the corresponding results in English, while the extracted version is worse in Danish: Extraction from islands is typically considered a spoken language phenomenon and hence a higher level of formality in the Danish stimulus items would clash more with the phenomenon under investigation, which is most often found in a colloquial context.¹¹

It may also be the case that unintended variation in formality level between the two languages played a role which differed across the conditions: For instance, the complementizer da (‘when’) can introduce both temporal and causal adjunct clauses, but its use as a causal complementizer is associated with an elevated level of formality, and in our stimuli, it is actually not possible to tell the temporal and causal interpretations apart. In other words, the lower-than-expected raw scores for extraction from clauses introduced by da (‘when’), may be related to a clash in formality level.

However, though we cannot completely rule out that it may have contributed to the attested difference between Danish and English, the fact that there was no significant effect of context (4.1) and no significant effect of dependency type (4.2) speaks against the translation as such being problematic. In addition, the sentences were constructed and carefully evaluated by native speakers, who did not find them problematic or stilted.

5.2.3 Syntactic difference between English and Danish

The lack of effect of removing the supporting context and of changing dependency type from relativization to topicalization could speak to a syntactic account of the results. However, adjunct islands are typically treated as a uniform class by syntactic theories (e.g. Chomsky, 1986; Huang, 1982; Rizzi, 1990). The variation attested between English and Danish in the two studies is thus difficult to explain with reference to syntactic structure alone or to standard generative analysis: The sentences containing extraction from adjunct clauses are generally thought to be structurally the same, both crosslinguistically in English and Danish, as well as across the three types of adjunct clause, if/hvis, when/da and because/fordi.

Any explanatory account of island phenomena needs to explain the variation attested across languages. It could be argued, as is done in Schütze et al. (2015, p. 35), that a parameterization of what counts as a bounding node may be able to capture crosslinguistic facts. Potential parametric variation in the structure of the left periphery in the MSc. languages has also been discussed in connection with island phenomena; Nyvad et al. (2017) suggest a cP/CP-distinction, which may help capture the extraction patterns found in Danish, as it affords a division between lexical, “root clause” CP, conceivably a bounding node for extraction, and the functional, potentially recursive “embedded clause” cP, which may not be (see also Vikner et al., 2017). This framework is reminiscent of Wiklund et al. (2007), who argue that the matrix predicate may select a clause with a more articulated left periphery (ForceP), thus carrying illocutionary force (corresponding to CP). Even so, this theoretical distinction cannot account for the low acceptability scores in Danish compared to English. In addition, in our data, the matrix predicate is invariably an emotive factive predicate, and the same predicate is kept constant across extractions. Moreover, all the adjunct clauses in our experiment were central (rather than peripheral) according to Haegeman’s (2012, pp. 161–164) definition, meaning that their structural position (and hence, semantic scope) should be the same, and due to this low adjunction site, extraction should be more easily acceptable (compared to peripheral clauses). (See also the discussion in Nyvad et al., 2022.) In other words, it would seem that the crucial factor determining the differences in the possibility of extracting must lie elsewhere.

It is, however, possible that a syntactic factor plays a role in the generally lower acceptability levels in Danish found in the present study, as compared to those found in English by Nyvad et al. (2022): In English, finite complement clauses headed by that normally cannot occur immediately after a preposition selected by an adjective (Quirk et al., 1985, pp. 658–661), a phenomenon known as the Case-Resistance Principle (Stowell, 1981), while the corresponding complement clauses in Danish require the presence of the preposition:

This basic syntactic asymmetry between English and Danish can potentially have led to the differences in the acceptability patterns attested, as this crosslinguistic difference might suggest a potentially stronger tendency towards a complement reading of the adverbial clause in the English stimuli than in the Danish, given that the complement reading in Danish would require the presence of a preposition in front of the adverbial clause; compare (7) and (8). In fact, a few of the comments left by the Danish-speaking participants revealed that some felt that a preposition was “missing” in the sentences containing adjunct island configurations. This could suggest that some of the participants might indeed have had a complement reading in mind at some point during the sentence processing. It is possible that a greater propensity towards an intermediate complement reading in the English stimulus sentences could have led to overall greater acceptability in the English data, but this is an empirical question for future research. Note again, however, that the density plots in Figure 4 for all three adjunct islands show narrow and unimodal distributions, which does not suggest speaker variation in parsing strategy. Furthermore, the corpus frequencies in 5.1.1 also show that the complement reading (Adj + Prep + Comp) is not likely for hvis, da, and fordi, and that examples with Adj + Prep make up less than half of the examples.

5.2.4 Demonstrative relative clause reading or cleft reading?

The crosslinguistic differences attested between English and Danish in the two studies could also be the result of a difference in the status of the filler (or correlate) that is being interpreted in the gap position of the adjunct clause. It may be the case that the English participants constructed a demonstrative relative clause, while the Danish participants constructed a cleft.

Proponents of the FBCC (e.g. Winckel et al., 2025) could argue that the discrepancy in acceptability levels in the English and Danish studies could be due to the English participants interpreting the fronting as a demonstrative, existential relative clause while a cleft reading is the only possibility in the Danish stimuli. Existential relative clauses are typically non-presuppositional and, hence, not backgrounded, which could have positive consequences for felicity. In clefts, on the other hand, the clefted constituent is focused and the negation test identifies the embedded clause as presupposed (see Lindahl, 2017, p. 160 ff for a discussion).

In the English stimuli (Nyvad et al., 2022), the extraction sentences are most plausibly interpreted as demonstrative relative clauses (e.g. This is the exercise that I was surprised because she completed). According to Kobzeva et al. (2022, p. 7), demonstrative relative clauses are characterized by the head of the relative clause being definite and preceded by a demonstrative pronoun and a finite form of to be.¹² In the present Danish study, however, the target sentences were unambiguously cleft sentences, due to the use of the acute accent (dét program ‘THAT exercise’) and, as a result, the clefted constituent receives emphatic stress and the NP that is co-referent with the gap is focalized (cf. the examples in (3)).

Winckel et al. (2025) compared extraction from relative clauses and clefts (it-clefts in English and c’est-clefts in French) and found significant island effects in both languages, but only for the clefts. According to Abeillé et al. (2020), the extraction in the demonstrative relative clause reading leads to less of a discourse clash than the extraction in the cleft structure. In a similar vein, Abeillé et al.’s FBCC would predict that, all else being equal, the acceptability ratings of the unambiguous cleft structures in Danish should be worse than what was found for the parallel English sentences: The latter could potentially be read as a demonstrative relative clause, making associating the filler with the gap inside the adjunct clause more felicitous; conversely, in the former, associating a clefted, hence, focused filler with a gap inside a backgrounded adjunct clause is less felicitous and less acceptable. This means that the information-structural match between the discourse function of the dependency type and the status of the adjunct constituent containing the gap is better in English than in Danish. If so, the crosslinguistic pattern for relativization could potentially be accounted for with reference to Abeillé et al.’s (2020) FBCC.

This line of reasoning, however, begs the question of why we see similar acceptability levels when the dependency type is topicalization instead of relativization. In principle, the FBCC can be extended to account for the frequently attested topicalization out of island configurations in the MSc. languages, since the fronting typically involves non-focal elements, and can, thus, pattern with relativization in terms of the discourse-functional role of the dependency type (Müller & Eggers, 2022). In other words, in topicalization structures, the topical filler is typically backgrounded information, and, in general, this dependency type is less foregrounding than question formation, which is focus movement. However, in our Danish stimulus sentences, the topicalization involves a contrastive focus, in that the topic is also focalized (or contrastive): The context introduces a discourse set and the topicalization picks out a particular referent in that set, as shown in (9), repeated from (6f) above:

In this sense, the filler in our Danish stimuli could be argued to be focalized in both the relativization structures (interpreted as clefts) and in the topicalization structures (the filler is a referent picked from a discourse set in the context). Given that the adjunct clauses are indeed backgrounded, this would incur a discourse-functional clash not found in the English stimuli: Extraction by way of relativization is not predicted by the account in Abeillé et al. (2020) to incur a strong island effect, and the English stimuli involve a demonstrative, existential relative clause, which may not be at odds with the adjunct constituent containing the gap.

It thus appears that the discourse-functional role of the filler and the embedded clause containing the gap may indeed play a role in acceptability. The problem with this kind of explanatory model is, however, that the concept of focus is slippery, open to interpretation, and not easily operationalizable, at least not in a way that can capture the subtle differences in acceptability and make the FBCC falsifiable. The lack of reliable tests for focus and backgroundedness makes it tricky to evaluate the functional accounts in question in connection with the empirical data collected in our studies here (cf. the critique of the negation test, Lindahl, 2017, p. 160 ff; Pérez-Leroux & Kahnemuyipour, 2014, pp. e121–e122; but see Cuneo & Goldberg, 2023). In addition, discourse-functional alignment between filler and gap position may just be a precondition for felicitous extraction, interacting with a host of other factors. In other words, the way in which the grammatical system interacts with discourse-pragmatic factors, such as information structure and topic-focus articulation, could play a decisive role in island sensitivity. It is also possible that specific constraints or language-internal factors in Danish influence island sensitivity differently than in English. These constraints might further interact with structural and processing factors, contributing to the overall acceptability levels. In other words, a substantial amount of research is still needed to extricate the component parts of island effects, across constructions and languages.