Too True to Be Good? The Non-Uniformity of Extraction from Adjunct Clauses in English

Abstract

Adjunct clauses are traditionally assumed to be strong islands for extraction across languages. However, the universal island status of adjunct clauses has been challenged by studies showing that extraction is possible from finite adjunct clauses in the Mainland Scandinavian languages. The possibility of extraction in these languages appears to be affected by various factors, including the type of adjunct clause, the type of extraction dependency, and the presence of contextual facilitation. These findings call for a re-evaluation of the islandhood of adjunct clauses in English. We conducted an acceptability judgment study on relativization from three types of finite adjunct clauses in English (if-, when-, and because-clauses) in the presence of supporting context. We found that the three clause types showed rather non-uniform acceptability patterns: extraction from when- and because-clauses both yielded significantly lower ratings than extraction from if-clauses, which patterned with non-island that-clauses. Our results suggest that at least for relativization, if- and when-adjuncts are not invariably strong islands in English, and that extra-grammatical factors may be key in understanding island structures traditionally assumed to be purely syntactic in nature.

1. Introduction

One of the defining characteristics of human language is that an element in a sentence can appear in a different position than the one in which it is interpreted, e.g., in long-distance extraction:

1.	Who do you think [that I saw __]?

In this example, the question word who has been displaced to the front of the matrix clause, but it is interpreted as the direct object of the main verb in the embedded clause. The link between these two positions is called a filler-gap dependency and such dependencies are in principle unbounded. However, Ross (1967) identified a range of structures that (to varying degrees) block extraction which he termed “islands”. It has since become common to distinguish between weak (or selective) and strong (or absolute) islands (e.g., Kluender 1998; Szabolcsi 2006; Szabolcsi and Lohndal 2017), depending on the degree to which it is possible to extract from such structures.

Island effects are widespread across the languages of the world and they have been used as an argument for the existence of innate linguistic principles. The language-acquiring child is allegedly not exposed to these types of extractions and the lack of negative evidence has been taken to suggest that certain constraints on filler-gap dependencies are part of universal grammar, either as principles or parameters.

Effects of repeated exposure (satiation, priming, and/or training effects) have been widely attested in the literature on island structures (Chaves and Putnam 2020; Christensen et al. 2013; Snyder 2000, 2022). Such effects may suggest that the constructions in question are in fact not illicit, but rather that they are unusual and/or highly complex, giving rise to initial processing difficulty and resulting in decreased acceptability, which is then ultimately ameliorated as speakers adapt to them.

In the syntax literature, finite adjunct clauses have traditionally been assumed to be strong islands cross-linguistically, as exemplified in (2) from Szabolcsi (2006, p. 481):

2.	*Which topic did you leave [because Mary talked about __]?

The ungrammaticality of (2) has been accounted for under the Condition on Extraction Domain (CED, Huang 1982, p. 505). This constraint states that a phrase can only be extracted out of a domain that is properly governed, and hence both subjects and adjuncts should be island environments (cf. the Empty Category Principle, Chomsky 1981, p. 205ff; Haegeman 1994, p. 442). The CED has been thought to be an innate, universal principle (along with a number of other island constraints, subsequently subsumed under e.g., subjacency, Chomsky 1973), given the learnability problem imposed by the purported lack of negative evidence in the input during language acquisition. However, counterexamples abound in the Mainland Scandinavian (MSc.) languages. Recent studies on Norwegian (Kush et al. 2019; Bondevik et al. 2020) have found that topicalization (but not wh-movement) out of conditional adjunct clauses headed by om (‘if’) appears to be possible. In addition, Müller (2017) found that topicalization from adjunct clauses headed by om (‘if’) and efter (‘after’) in Swedish yielded acceptability ratings on the upper end of the scale, provided that the matrix and the adjunct clause event can be interpreted to be in a causal, coherent relation e.g., (3a), rather than in a purely temporal one e.g., (3b).

3.

a.

Det där röda vinet

mådde

jag

lite

illa

this there red wine.the

felt

I

a.little

sick

[efter

att

jeg

hade

druckit

___

sist].

after

that

I

had

drunk

last

‘I felt a little sick after I had drunk that red wine last time.’

b.

Det röda

kan

vi

ju

gå

ut

på

stan

the red

can

we

prt1

go

out

on

town.the

[efter

att

vi

har

druckit

___].

after

that

we

have

drunk

‘We can go out on the town after we have drunk the red wine.’

However, other types of adjunct clauses—specifically, clauses introduced by fordi ‘because’ in Norwegian and eftersom ‘because’ in Swedish—were found to be much less permissive in terms of extraction in these studies in the sense that most participants rated topicalization from them on the low end of the scale (Bondevik et al. 2020; Müller 2017). This has given rise to conjectures that different types of adjunct clauses, or adjunct clauses headed by different complementizers, may vary with regard to extraction possibilities (Bondevik 2018, p. 201; Bondevik et al. 2020; Müller 2017, 2019). Moreover, the studies by Kush et al. (2018, 2019) and Bondevik et al. (2020), among others, suggest that topicalization from adjunct clauses is easier to accept than question formation by wh-extraction, indicating that the acceptability of extraction might also depend on the type of extraction (see also Lindahl 2017; Müller 2019; and Abeillé et al. 2020 for suggestions along these lines).

Another factor that has been suggested to be important for the felicity of at least some island extractions in MSc. in formal acceptability studies is whether the stimulus sentences are preceded by one or more sentences providing a facilitating context. More specifically, some of the formal studies of adjunct or relative clause extraction in MSc. have yielded acceptability ratings that are unexpectedly low, given the examples provided in the literature and the informal observations reporting such structures to be acceptable in MSc. (e.g., Christensen and Nyvad 2014; Kush et al. 2018). It has been suggested that these unexpectedly low ratings can be partly explained by the lack of contextual cues in formal settings (e.g., Tutunjian et al. 2017; Wiklund et al. 2017, p. 207; Kush et al. 2018; Müller 2019, p. 209; Bondevik et al. 2020). Contextual cues may be required for the felicity of at least some of these extractions if for instance a context is necessary to license certain interpretational properties of the extracted phrase, such as D(iscourse) linking or specificity (Pesetsky 1987; Szabolcsi 2006). The licensing of such interpretational properties has been claimed to be relevant for the felicity of weak island extractions (see e.g., Starke 2001; Szabolcsi 2006; Abrusán 2014). In a recent study on Norwegian, Kush et al. (2019) found support for the beneficial effect of context on ratings for at least some adjunct clause extractions: A significant island effect (in terms of the factorial definition of island effects developed by Sprouse 2007) was observed for decontextualized extraction from conditional adjuncts, but no significant island effect occurred once a contextual preamble was added to the stimuli sentences. In fact, the participants in this study rejected sentences involving contrastive topicalization from complement clauses, uncontroversially considered to be grammatical, presumably at least in part because they were presented in vacuo. More general evidence in favor of the facilitative role of supporting context for the comprehension of fronted elements is presented by Engelkamp and Zimmer (1983), Kristensen et al. (2014), and Lau et al. (2020), but see Bernardy et al. (2018). The question remains whether English sentences such as (2) can be ameliorated by being presented in a facilitating context.

Kush et al. (2018) found significant inter-individual variation relating to the acceptability judgements of sentences involving wh-movement out of embedded polar questions in Norwegian (headed by om ‘if’), including a large number of participants not exhibiting any island sensitivity to this type of extraction. Kush et al. (2018) conclude that this pattern of inter-participant variation is incompatible with embedded questions being syntactic islands in Norwegian and they argue that the variability in acceptability is actually indicative of extra-syntactic factors being at play. Similarly, Bondevik et al. (2020) and Kush et al. (2019) found both inter- and intra-participant variation as well as variation between and within adjunct types in Norwegian.

The above-mentioned studies (Bondevik et al. 2020; Kush et al. 2018, 2019; Müller 2017) indicate that adjunct clauses in the MSc. languages display a non-uniform behavior when it comes to their island sensitivity, and that the possibility of extraction is affected by various factors, including the type of adjunct clause, the type of extraction dependency, contextual facilitation, and semantic coherence2. These findings seem to call for a re-evaluation of the situation in English regarding adjunct islandhood. To date, only few formal acceptability studies of extraction from finite adjunct clauses in English exist, and almost all of them focus on wh-extraction from only one type of adjunct clause and do not provide any contextual facilitation (e.g., Sprouse 2009; Sprouse et al. 2012; Michel and Goodall 2013). Though Sprouse et al. (2016) also investigated extraction by relativization from if-adjuncts, they did not compare it to extraction from other types of adjunct clauses. In order to adequately evaluate whether English finite adjunct clauses really are uniformly strong islands, as reported in the traditional literature (and to allow for a fair comparison between English and MSc. adjunct islands), an investigation that tests adjunct clause extraction in English controlling for the above-mentioned factors (type of adjunct, dependency type, presence/absence of supporting context, and coherence) is necessary. In this paper, we present a study that aims to fill (parts of) this gap, by testing extraction in the form of relativization from three different types of finite adjunct clauses in English (if-, when-, and because-clauses) in the presence of supporting context.

We chose to test extraction in the form of relativization rather than topicalization since topicalization is a fairly marked structure in English compared to the MSc. languages (e.g., Engdahl 1997; Poole 2017, p. 15). There are some indications that relativization behaves on a par with topicalization in that both seem to facilitate extraction from certain islands, compared to wh-movement. Sprouse et al. (2016) compared wh-extraction and relativization out of adjunct clauses in English and found that relativization from if-clauses does not result in island effects in terms of the factorial definition of islands (although ratings for both relativization and question formation remained relatively low). Recently, Abeillé et al. (2020) showed that relativization (but not question formation) from subject islands in English as well as in French received ratings on par with grammatical controls. Moreover, there are indications that spontaneously produced cases of adjunct clause extractions primarily feature relativization from the adjunct: to see whether any authentic examples of adjunct island extraction can be found in English, Müller and Eggers (2022) conducted an exploratory corpus study on adjunct extraction using the Corpus of Contemporary American English (COCA; Davies 2008). All cases of adjunct clause extraction found in naturalistic English in their study involved relativization from the island, such as the examples in (4), both from COCA.

4.	a.	Many of the exercises are ones that I would be surprised if even 1 percent of
			healthy women can do.
	b.	Now, those are things that I feel very warm when I look at, and I wouldn’t
			want to live in a house that they—a house that didn’t have room for those.

On a purely syntactic account, the level of acceptability of extraction out of finite adjunct clauses should be uniform across adjunct types, given that they have the same syntactic structure and that they are assumed to be adjoined to the matrix clause in the same way. Furthermore, gradient acceptability of extraction from adjunct clauses is not predicted by the CED either.

As an alternative to purely syntactic accounts of island constraints, Chaves and Putnam (2020) propose that a host of islands traditionally assumed to be of syntactic nature are instead “Relevance Islands”: the filler must be “relevant for the main action that the proposition conveys” and if it is not, unacceptability ensues as a natural consequence, without the need for postulating constraints in the syntax (Chaves and Putnam 2020, p. 120; in line with other semantic or discourse-pragmatic accounts such as Erteschik-Shir 1973; Deane 1991). Prima facie, even a semantico-pragmatic account of island effects such as the one proposed by Chaves and Putnam would predict uniformly low acceptability of adjunct clause extraction since adjunct clauses are typically used to express backgrounded or presupposed information (see also e.g., Van Valin 1994; Erteschik-Shir 2006; Goldberg 2013, p. 203 for accounts in a similar vein).

2. Predictions

In light of the counterexamples to supposed universal constraints found in the MSc. languages, Christensen and Nyvad (2019, 2022) re-examined wh-islands and relative clause extraction in English in order to investigate whether the same island-insensitivity and graded acceptability could also be demonstrated for English. However, the results support the hypothesis that the syntactic configurations in question (embedded wh-questions and relative clauses) are strong islands in this language, as is standardly assumed in the syntactic literature (but see Vincent et al. 2022). Given this compatibility of experimental findings in English with the standard assumptions in the syntax literature (the Wh-Island Constraint and the Complex NP Constraint, respectively), our predictions regarding extraction from adjunct clauses are based on the CED, which treats all adjunct clauses as a coherent class in terms of islandhood, potentially ruling out any scope for variation:

Prediction 1.

Adjunct clauses are strong islands, i.e., their acceptability level is consistently low.

Prediction 2.

There is no variation in acceptability as a function of adjunct clause type.

Prediction 3.

There is no between-participant variation.

Prediction 4.

There is no trial effect. The acceptability of extraction is not positively correlated with repeated exposure over time in the experiment, regardless of the type of adjunct clause.

We included subject islands and coordinate structure islands as fillers and as points of comparison, given that these structures are also assumed to be strong islands in English, but with different strengths. The Subject Condition (Ross 1967; Chomsky 1973) has been proposed not to be active across all types of constructions: Kush et al. (2018, 2019) found substantial island effects in the acceptability rating of constructions involving movement out of subject phrases. However, extraction from subjects has been demonstrated to be grammatical in certain instances, depending on referential processing (Culicover and Winkler 2022; Culicover et al. 2022), discourse function (Abeillé et al. 2020), and repeated exposure (Chaves and Putnam 2020, p. 221). The Coordinate Structure Constraint, on the other hand, appears to hold cross-linguistically, and no robust counterexamples seem to have been identified (Chaves and Putnam 2020, p. 72).3 This way, we had two different points of comparison in terms of patterns of ungrammaticality.

3. Materials and Methods

Our design was based on a 2 × 2 stimulus design with the structure outlined in

Table 1. Experimental design.

	Complementizer	[−Extraction]	[+Extraction]
Non-island	That	Type 1	Type 5
Islands	If	Type 2	Type 6
	When	Type 3	Type 7
	Because	Type 4	Type 8

, with ±Island and ±Extraction as the two factors; this is similar to the factorial design introduced by Sprouse (2007) and subsequently employed by e.g., Sprouse et al. (2011, 2016) and Bondevik et al. (2020). In addition, our design had a Complementizer factor with four different levels: that, if, when, and because. That-clauses are not assumed to be islands, since they are complement clauses, whereas the other three types introduce adjunct clauses. This design gave us eight target types (see Table 1), which we used as the levels for a Type factor for our statistical model in order to test for interactions between extraction and the four clause types.

Each target stimulus sentence was preceded by a short facilitating context. We constructed the materials in such a way that all eight types could be preceded by the same context, as shown in (5):

5.		Context: In the latest workout routine I designed for Emma, I really wanted to make it impossible for her and included another set of particularly brutal pull-ups.
		Non-island structure, [-Extraction]:
	a.	It’s obvious that I was surprised that she actually completed this exercise.
		Island structure, [-Extraction]:
	b.	It’s obvious that I would be surprised if she actually completed this exercise.
	c.	It’s obvious that I was surprised when she actually completed this exercise.
	d.	It’s obvious that I was surprised because she actually completed this exercise.
		Non-island structure, [+Extraction]:
	e.	This is the exercise that I was surprised that she actually completed __.
		Island structure, [+Extraction]:
	f.	This is the exercise that I would be surprised if she actually completed __
	g.	This is the exercise that I was surprised when she actually completed __.
	h.	This is the exercise that I was surprised because she actually completed __.

Distinct from the factorial design introduced by Sprouse (2007) and subsequently employed by e.g., Sprouse et al. (2011, p. 201; 2016), Bondevik et al. (2020), the main matrix clauses of the stimulus sentences are not minimally different, but they are as minimally different as they can be. The sentences without extraction (the baseline versions) all begin with It’s obvious that… or It’s clear that…, whereas the sentences with extraction all begin with This is the X that… (where the X is the extracted filler). The stimulus set was created this way because we prioritized having relativization out of the adjunct clauses (see Section 1 above). There is no direct unextracted counterpart to relativization that our [+Extraction] sentences could be compared to, but the It’s obvious that… or It’s clear that… constructions used for the matrix in our baseline sentences are as similar as possible in terms of length and complexity to the cleft structure used in the relativization condition. Note that both matrix structures add very little to the overall structure and meaning (we used pronouns and copula be). We acknowledge that this makes the ±Extraction contrast not minimally different, but there are no a priori reasons to suspect that these small differences would affect the acceptability ratings.

As illustrated in (5) and in Table 1, we wanted to compare three different types of adjunct islands: if-clauses, when-clauses and because-clauses. The adjunct clauses and non-island that-clauses were embedded under adjectival psych-predicates such as be surprised or be happy, which can be expected to trigger a causal and thus semantically coherent reading of the matrix and embedded event. The presence of a causal, coherent relation between the matrix and the adjunct clause has previously been shown to facilitate extraction from adjuncts (Truswell 2011; Tanaka 2015; Müller 2019). In order to allow for a felicitous use of the if-clauses, the verbal structure in the matrix clause was slightly altered in the conditions involving if-adjuncts (5b,f) compared to the other conditions: instead of past tense was, the structure would be + psych-adjective was used, signaling the hypothetical.

Twenty-four sets of items of the type in (5) were constructed and distributed across eight lists in a Latin square design. In this way, each list contained only one version of the same scenario; since each participant saw only one list, each participant was exposed to three tokens of each type. The order of sentences on each list was randomized. As mentioned in the introduction, we also added two types of fillers: eight sets involving extraction from NP subjects, filler 1, as in (6), and eight sets involving extraction from coordinate structures, filler 2, as in (7). The NP subjects and the coordinate structures were embedded in the same four clause types that we used in target items, i.e., non-island that-clauses, and adjunct clauses introduced by if, when, and because (see Table 1). Each of these conditions with extraction was compared to the corresponding version without extraction, resulting in eight control conditions that could easily be compared to our target items. (The entire stimulus set is available online, see Data Availability Statement).

6.		Filler 1.
		Context: My team has developed a COVID-19 vaccine in record time, and I think we deserve some recognition.
		Subject island + non-island that-clause, [-Extraction]:
	a.	It’s clear that we were pleased that our vaccine against this virus finally got the Nobel Prize.
		Subject island + adjunct island, [-Extraction]:
	b.	It’s clear that we would be pleased if our vaccine against this virus finally got the Nobel Prize.
	c.	It’s clear that we were pleased when our vaccine against this virus finally got the Nobel Prize.
	d.	It’s clear that we were pleased because our vaccine against this virus finally got the Nobel Prize.
		Subject island + non-island that-clause, [+Extraction]:
	e.	This is the virus that we were pleased that our vaccine against __ finally got the Nobel Prize.
		Subject island + adjunct island, [+Extraction]:
	f.	This is the virus that we would be pleased if our vaccine against __ finally got the Nobel Prize.
	g.	This is the virus that we were pleased when our vaccine against __ finally got the Nobel Prize.
	h.	This is the virus that we were pleased because our vaccine against __ finally got the Nobel Prize.
.		Filler 2.
		Context: I had promised a friend to watch his pets for a week. Unfortunately, I accidentally left the front door open just a bit too long and both his cat and his prize-winning show dog ran out.
		Coordinate structure island + non-island that-clause, [-Extraction]:
	a.	It’s obvious that I was ashamed that I actually lost both the cat and this dog in one day.
		Coordinate structure island + adjunct island, [-Extraction]:
	b.	It’s obvious that I would be ashamed if I actually lost both the cat and this dog in one day.
	c.	It’s obvious that I was ashamed when I actually lost both the cat and this dog in one day.
	d.	It’s obvious that I was ashamed because I actually lost both the cat and this dog in one day.
		Coordinate structure island + non-island that-clause, [+Extraction]:
	e.	This is the dog that I was ashamed that I actually lost both the cat and __ in one day.
		Coordinate structure island + adjunct island, [+Extraction]:
	f.	This is the dog that I would be ashamed if I actually lost both the cat and __ in one day.
	g.	This is the dog that I was ashamed when I actually lost both the cat and __ in one day.
	h.	This is the dog that I was ashamed because I actually lost both the cat and __ in one day.

Participants were instructed to base their rating only on the sentence following the context. The context in our stimuli was constructed so that it triggered a cohesive interpretation of the extracted DP in the discourse in the [+Extraction] conditions.

Participants were recruited via professional connections and through various social media platforms and were randomly assigned to one of the eight presentation lists. They rated the sentences using an online questionnaire created with Google Forms. Judgments were given on a seven-point Likert scale ranging from 1 = ‘completely unacceptable’ to 7 = ‘completely acceptable’. The participants had the opportunity to leave comments on every test sentence.

4. Results

A total of 235 native speakers of English (188 female, 41 male, 6 other) participated in the experiment on a voluntary basis. Participant age ranged from 18 to 74 years (mean = 33.19, SD = 10.79). The level of education in years (including e.g., primary and secondary school, college, university, PhD, etc.) ranged from 9 to 30 years (mean = 18.00, SD = 3.19). The number of participants per list ranged between 24 and 39 (list 1 = 25, list 2 = 28, list 3 = 30, list 4 = 29, list 5 = 39, list 6 = 32, list 7 = 24, list 8 = 28).

Prior to analysis, we removed one of the target sets (a set of 8 sentences with the same context) because it turned out the that [−Extraction] sentence (as in (5a) above) was pragmatically illicit. Since we wanted to use that [−Extraction] as baseline for the island effect size measure (i.e. the DD-score in Figure 4 below), including this in the analysis would skew the result.

Table 2. Mean acceptability scores on a 7-point scale. [±Ex] = ±Extraction.

		Target Sentences		Filler 1		Filler 2
		Acceptability	SD	Acceptability	SD	Acceptability	SD
That	[−Ex]	5.90	1.41	5.81	1.45	5.58	1.42
If	[−Ex]	6.27	1.11	5.83	1.42	5.77	1.27
When	[−Ex]	6.14	1.24	5.99	1.35	5.88	1.19
Because	[−Ex]	5.82	1.39	5.77	1.51	5.51	1.55
That	[+Ex]	4.97	1.72	2.78	1.69	1.91	1.23
If	[+Ex]	4.97	1.76	2.80	1.63	1.83	1.29
When	[+Ex]	3.92	1.84	2.37	1.55	1.86	1.32
Because	[+Ex]	3.39	1.74	2.11	1.42	1.70	1.19

contains the mean acceptability scores for all the conditions.

The data was analyzed using R version 4.1.2 (R Core Team 2021) with the lmerTest Package (Kuznetsova et al. 2017) for mixed effects models and the MASS Package (Venables and Ripley 2002) for sliding, pair-wise contrasts of the type factor.

Extraction from adjunct clauses (target sentences as in (5) above), extraction from embedded subjects (filler type 1, as in (6)), and extraction from the second conjunct in a coordinate structure (filler 2, as in (7)) were analyzed using separate mixed-effects models. (We had no prior hypotheses about statistical differences between the three, and since the three sentence types were not minimally different, direct comparisons would also be inappropriate).

All plots were made in R using ggplot2 (Wickham 2016) and gridExtra (Augie 2016).

4.1. Target Type: Adjunct Islands

The model contained acceptability as an outcome variable and type as a predictor (with sliding contrasts) and random intercepts for participant and item, and random slopes for type and trial by participant, and random slopes for trial by item. The summary of the results is provided in

Table 3. Summary of the mixed effects model for the raw ratings of the target sentences (adjunct islands). Significance indicators: *** p < 0.001, ** p < 0.01, * p < 0.05.

	Estimate	SE	t	df	p
that [–Ex] vs. if [–Ex]	0.333	0.120	2.774	143.403	0.006	**
if [–Ex] vs. when [–Ex]	−0.108	0.119	−0.907	136.577	0.366
when [–Ex] vs. because [–Ex]	−0.293	0.118	−2.472	143.549	0.015	*
because [–Ex] vs. that [+Ex]	−0.826	0.138	−5.992	211.140	0.000	***
that [+Ex] vs. if [+Ex]	−0.043	0.130	−0.331	179.725	0.741
if [+Ex] vs. when [+Ex]	−1.041	0.129	−8.083	179.660	0.000	***
when [+Ex] vs. because [+Ex]	−0.487	0.121	−4.042	136.090	0.000	***

.

The results and sliding pairwise comparisons for the eight target sentence types are also shown in Figure 1A. The acceptability ratings for the four baseline conditions [−Ex] all had a rating above 5 on the 7-point scale. There was a significant drop in acceptability between the [−Ex] and [+Ex] types, indicating a significant negative main effect of extraction, [−Ex] > [+Ex], as shown by the contrast between because [−Ex] and that [+Ex] (p < 0.001, see also Figure 1). (Running the same model without sliding contrasts showed significant differences between that [−Ex] and that [+Ex] (p < 0.001) as well as between that [−Ex] and if/when [−Ex], while the difference between that [−Ex] and because [−Ex] was not significant). Within the [+Ex] types, while the ratings for that and if did not differ from each other, they were both rated significantly higher than when, which was rated significantly higher than because.

4.2. Filler Type 1: Subject Islands

As in the target condition, the model had acceptability as outcome variable and type as predictor (with sliding contrasts). However, the only convergent model had no random slopes, only random intercepts for participant and item. The model is summarized in

Table 4. Summary of the mixed effects model for filler type 1 (subject islands). Significance indicators: *** p < 0.001, marginal: · p < 0.1.

	Estimate	SE	t	df	p
that [–Ex] vs. if [–Ex]	0.033	0.237	0.137	52.854	0.891
if [–Ex] vs. when [–Ex]	0.134	0.237	0.565	52.854	0.575
when [–Ex] vs. because [–Ex]	−0.225	0.237	−0.947	52.854	0.348
because [–Ex] vs. that [+Ex]	−2.979	0.237	−12.549	52.854	0.000	***
that [+Ex] vs. if [+Ex]	0.046	0.237	0.192	52.854	0.849
if [+Ex] vs. when [+Ex]	−0.421	0.237	−1.773	52.854	0.082	·
when [+Ex] vs. because [+Ex]	−0.269	0.237	−1.134	52.854	0.262

.

The results and sliding pairwise comparisons for the eight sentence types are also shown in Figure 1B. There were no significant differences between the four baseline [−Ex] types, which were all rated in the top range. In contrast, the four [+Ex] sentence types were all in the lower range, and they were all significantly different from the [−Ex] types. In other words, there was a clear and significant negative main effect of extraction. There was also a marginally significant difference between if [+Ex] and when [+Ex] (p = 0.082).

4.3. Filler Type 2: Coordinate Structure

The converging model contained acceptability as outcome variable and type as predictor (with sliding contrasts) and random intercepts for participant and item, and random slopes for trial by participant and item. The model is summarized in

Table 5. Summary of the mixed effects model for filler type 2 (coordinate structure). Significance indicators: *** p<0.001, ** p < 0.01, marginal: · p < 0.1.

	Estimate	SE	t	df	p
that [–Ex] vs. if [–Ex]	0.268	0.150	1.790	54.999	0.079	·
if [–Ex] vs. when [–Ex]	0.052	0.150	0.347	52.842	0.730
when [–Ex] vs. because [–Ex]	−0.430	0.143	−3.003	47.354	0.004	**
because [–Ex] vs. that [+Ex]	−3.518	0.142	−24.805	43.364	0.000	***
that [+Ex] vs. if [+Ex]	−0.104	0.136	−0.762	37.712	0.451
if [+Ex] vs. when [+Ex]	0.073	0.130	0.560	34.270	0.579
when [+Ex] vs. because [+Ex]	−0.162	0.128	−1.263	30.727	0.216

.

See also Figure 1C for the results and sliding pairwise comparisons for the eight sentence types. Similar to the findings for filler 1, the acceptability ratings for the four baseline conditions [−Ex] were all in the top range, whereas the four [+Ex] sentence types were all in the lower range. There was a marginal difference between that [−Ex] and if [−Ex] and a significant difference between when [−Ex] and because [−Ex]. As with filler 1, the most dramatic effect is the negative main effect of extraction, as indicated by the difference between because [−Ex] and that [+Ex] (p < 0.001).

Figure 1. Mean (raw) acceptability ratings. *** p < 0.001, ** p < 0.01, marginal: · p < 0.1.

Figure 1. Mean (raw) acceptability ratings. *** p < 0.001, ** p < 0.01, marginal: · p < 0.1.

To see if the general acceptability pattern in the target condition (Figure 1A) was consistent across the 23 sets (or contexts, see Section 2), we plotted acceptability by type and participant for each set. As can be seen in Figure 2, the overall patterns look quite similar.

Next, in order to obtain an impression of the between-participant variation, we plotted the mean responses from each participant for the target sentences, see Figure 3. Although most of them show the same pattern with a downward slope from left to right, there is also a lot of variation. To control for this inter-participant variation (which is actually also taken into consideration as a random effect in the mixed-effects model) and to inspect it further, we applied z-transformation (next section).

Figure 2. Mean (raw) acceptability ratings by type (see Figure 2) for each of the 24 target sets.

Figure 2. Mean (raw) acceptability ratings by type (see Figure 2) for each of the 24 target sets.

Figure 3. Mean (raw) acceptability ratings by type for each of the 235 participants.

Figure 3. Mean (raw) acceptability ratings by type for each of the 235 participants.

4.4. Standardized (z-Transformed) Acceptability Ratings

For direct comparison with Sprouse et al. (2012, 2016) and a number of other studies, we z-score transformed the ratings by participant to control for potential individual scale bias, such as using only one end of the scale or a larger or smaller range (see also Kush et al. 2018, 2019; Bondevik et al. 2020). A z-transformed rating represents the number of standard deviations the raw (non-transformed) rating is from that participant’s mean rating. As with the raw scores, the z-transformed results were analyzed using a linear mixed effects model with sliding contrasts. The model contained type as a predictor and random intercepts for participant and item, and random slopes for type and trial by participant, and random slopes for trial by item. The summary of the results is provided in

Table 6. Summary of the mixed effects model with sliding contrasts for the z-transformed ratings of the target sentences (adjunct islands). Significance indicators: *** p < 0.001, ** p < 0.01, * p < 0.05.

	Estimate	SE	t	df	p
that [–Ex] vs. if [–Ex]	0.172	0.060	2.858	148.388	0.005	**
if [–Ex] vs. when [–Ex]	−0.061	0.060	−1.015	141.736	0.312
when [–Ex] vs. because [–Ex]	−0.155	0.060	−2.581	150.968	0.011	*
because [–Ex] vs. that [+Ex]	−0.377	0.067	−5.620	205.117	0.000	***
that [+Ex] vs. if [+Ex]	0.011	0.064	0.167	178.378	0.868
if [+Ex] vs. when [+Ex]	−0.536	0.064	−8.400	180.962	0.000	***
when [+Ex] vs. because [+Ex]	−0.248	0.061	−4.084	151.319	0.000	***

and plotted in Figure 4A. Note that the general acceptability pattern as well as the significant contrasts are the same as for the non-transformed ratings, compare Figure 1A and Figure 4A, and Table 3 and Table 6.

To test for a superadditive ‘island effect’, we calculated the differences-in-differences (DD) score for each [+Ex] condition by subtracting the D score for that, measured as the difference in rating between that [–Ex] and that [+Ex] (i.e., the extraction effect for the non-island type), from the D scores for the three adjunct island types: if, when, and because, see Figure 4C,D (Sprouse et al. 2012; Kush et al. 2018, 2019; Bondevik et al. 2020). DD score proves a standardized measure of the island effect size which can be used for comparisons between different island types across experiments and languages. According to Kush et al. (2019, p. 401), “DD scores for island effects typically fall within the range of 0.75–1.25”, and “any intermediate-sized island effect bears closer scrutiny”. To test for significance, we analyzed the z-transformed ratings again using a linear mixed-effects model, this time with fixed effects of Complementizer and Extraction and their interaction, and random intercepts for participant and item, and random slopes for type and trial by participant, and random slopes for trial by item, first with that as intercept (summarized in

Table 7. Summary of the mixed effects model of Complementizer and Extraction (z-transformed ratings, target sentences (adjunct islands)), using that as intercept. *** p < 0.001, ** p < 0.01, marginal: · p < 0.1.

	Estimate	SE	t	df	p
Comp = that (Intercept)	0.598	0.047	12.787	181.579	0.000	***
Comp = if	0.190	0.066	2.895	176.200	0.004	**
Comp = when	0.118	0.066	1.787	182.055	0.076	·
Comp = because	−0.050	0.066	−0.746	185.248	0.457
Extraction	−0.435	0.065	−6.702	171.523	0.000	***
Comp = if x Extraction	−0.162	0.092	−1.760	171.650	0.080	·
Comp = when x Extraction	−0.626	0.092	−6.812	171.573	0.000	***
Comp = because x Extraction	−0.724	0.092	−7.886	171.364	0.000	***

), then with when as intercept (summarized in

Table 8. Summary of the mixed effects model of Complementizer and Extraction (z-transformed ratings, target sentences (adjunct islands)), using when as intercept. *** p < 0.001, * p < 0.05, marginal: · p < 0.1.

	Estimate	SE	t	df	p
Comp = when (Intercept)	0.716	0.046	15.456	175.595	0.000	***
Comp = if	0.071	0.066	1.087	175.994	0.279
Comp = because	−0.168	0.065	−2.565	175.022	0.011	*
Comp = that	−0.119	0.065	−1.813	175.326	0.072	·
Extraction	−1.061	0.065	−16.212	175.326	0.000	***
Comp = if x Extraction	0.464	0.093	5.016	175.413	0.000	***
Comp = because x Extraction	−0.099	0.093	−1.068	175.109	0.287
Comp = that x Extraction	0.626	0.093	6.759	175.227	0.000	***

). (For when, the model had to be simplified such that it included random slopes only for extraction by participant).

Figure 4A shows the mean of the z-transformed acceptability ratings for the target condition, with and without extraction (compare Figure 1), 4B the extraction effects for the four complementizers, 4C the extraction effect sizes (all significant, p < 001, and that, if, and when significantly different form each other, p < 0.001), and 4D the DD scores indicating island effects. The DD score for if is only marginally significant (cf. Comp = if x Extraction, p = 0.080, in Table 7), whereas the DD scores for when and because were significant (Comp = when x Extraction and Comp = because x Extraction in Table 7, both p < 0.001). As indicated by Comp = if x Extraction, p < 0.001, in Table 8, the DD scores for if and when were significantly different from each other (and when and because were not, cf. Comp = because x Extraction, p = 0.287). Note that the DD scores are all smaller than 0.75, which, according to Kush et al. (2019, p. 401), is the lower threshold value for typical island effects.

In line with Kush et al. (2019) and Bondevik et al. (2020), in order to examine the underlying judgment pattern for each of the mean scores shown in Figure 1, we investigated the distribution of acceptability ratings by condition, as shown in the density plots in Figure 5. Uniform syntactic islands should show a unimodal distribution narrowly centered around z = −1, indicating that the participants consistently rated them as unacceptable. This is indeed the case for the two filler types with extraction (Figure 5B,C, light blue). The distribution of ratings for extractions from subject islands and coordinate structures is clearly skewed toward the left. Conversely, unambiguously well-formed types should show a unimodal distribution centered narrowly around z = +1, indicating that the participants consistently rated them as completely acceptable. Indeed, this is the case for all the baseline [−Ex] sentence types (Figure 5, dark blue): the ratings are clearly skewed towards the right. The same goes for extraction from complement clauses headed by that, though the distribution is more even and flat than the corresponding baselines (it is less pointy, suggesting more inter-participant variation). That leaves the three adjunct island extractions: if, when, and because. Extraction from if shows the same pattern as extraction from that, suggesting that extraction from if is equally acceptable (as shown in Figure 1 and Table 3). Extraction from because shows the inverse pattern, with a somewhat less pointy but clearly left-skewed unimodal distribution suggestive of islandhood. Finally, when stands out with a left-skewed but somewhat ‘flatter’ or even distribution, tending towards being bimodal with peaks at z = −1 and z = 0.

Figure 4. (A) Mean z-transformed acceptability ratings. (B) Extraction effect. (C) Extraction effect size (difference between [+Ex] and [–Ex]). (D) Island effect size measured in DD scores (Complementizer x Extraction interaction) between that and the other three complementizers). *** p < 0.001, ** p < 0.01, * p < 0.05, marginal: · p < 0.1.

Figure 4. (A) Mean z-transformed acceptability ratings. (B) Extraction effect. (C) Extraction effect size (difference between [+Ex] and [–Ex]). (D) Island effect size measured in DD scores (Complementizer x Extraction interaction) between that and the other three complementizers). *** p < 0.001, ** p < 0.01, * p < 0.05, marginal: · p < 0.1.

4.5. Trial Effects

In order to test for trial effects, in particular positive correlations between repeated exposure of type and acceptability ratings, we plotted the mean acceptability (across participants on each list) of each occurrence of the four extraction types in the target category, see Figure 6A. That, when, and because showed a positive effect of trial, but the effect was only significant for when: in the course of the experiment, extraction across when increased 0.6 points on the 7-point scale (R² = 0.17, p < 0.05).

Under the (standard) assumption that adjunct clauses headed by if, when, and because have the same syntactic structure, we combined the three sets to see if their combined number of trials i.e., (9) would show a significant effect. This was not the case (p > 0.1, see Figure 6B).4

Figure 6. (A) Acceptability as a function of repetition (trial) of extractions from complement clauses headed by that and from adjunct clauses headed by if, when and because in the target category. Grey shading: 95% confidence interval. (B) Trial effect of extraction from target adjunct clauses headed by either if, when or because. Grey shading: 95% confidence interval.

Figure 6. (A) Acceptability as a function of repetition (trial) of extractions from complement clauses headed by that and from adjunct clauses headed by if, when and because in the target category. Grey shading: 95% confidence interval. (B) Trial effect of extraction from target adjunct clauses headed by either if, when or because. Grey shading: 95% confidence interval.

5. Discussion

We investigated the acceptability of extraction from three different types of finite adjunct clauses (if-, when- and because-clauses) in English. Extraction was tested in the form of relativization (rather than wh-extraction) and in the presence of supporting context, since previous studies on island constructions in MSc. indicate that type of extraction dependency and contextual facilitation appear to have an impact on the possibility of at least some island extractions. Our study thus allowed us to examine whether finite adjunct clauses in English indeed constitute uniformly strong islands, as the traditional literature suggests, even when the potentially facilitating role of dependency type and context is taken into consideration. We predicted consistently low ratings for relativization from adjunct clauses, with no variation in acceptability across different adjunct clause types, and no trial effects. These predictions were not confirmed: the three different types of adjunct clauses tested in this study displayed a rather non-uniform behavior with regard to their acceptability. Specifically, extraction from when- and because-clauses both yielded significantly lower ratings than extraction from if-clauses. If-clauses appeared to pattern with non-island that-clauses instead, in that extraction from both if- and that-clauses yielded ratings above the middle range, and a similar, right-skewed distribution of acceptability ratings (as can be seen in the density plots in Figure 5). These findings suggest that, at least for relativization, if-adjuncts are not absolute islands in English. However, when- and because-clauses in turn yielded not only significantly different ratings with extraction, but these structures also resulted in somewhat different acceptability distributions (with because showing a left-skewed distribution, whereas extraction from when resulted in a flatter distribution, cf. Figure 5). This picture is corroborated by a difference in the effects of repeated exposure: only when-clause extraction yielded a significant trial effect in our experiment, whereas the other clause types did not. When-clause extraction thus seems to stand out, both in terms of showing a trial effect and in showing a rather flat, somewhat bimodal distribution of acceptability ratings. This suggests that the reduced acceptability levels of the three clause types have different underlying causes (see also Snyder 2022).

Our finding that extraction from non-island that-clauses and from if-clauses both yielded positive and non-distinct ratings suggests that at least for relativization, if-clauses behave similarly to that-clauses in English in that they do not seem to universally block extraction. This finding challenges the strong island status of if-adjuncts. On the other hand, there was a statistically significant difference between that [−Ex] and if [−Ex], indicating that the baseline conditions for that- and if-clauses may not have been as minimally different as we intended. Some of the comments left by the participants in Google Forms reveal that a few participants were bothered by that occurring twice in relatively close proximity in our that-condition, as in (8). Both instances are actually optional, and some participants would prefer the sentences in which (at least) the first that was left out.

8.	It’s obvious that I was surprised that she actually completed this exercise.

The comments furthermore indicate that in a few items, the context provided by us may have been more felicitous with the conditional structure used in the if-clauses than with the indicative that-clause. Both of these circumstances could explain why our baseline that-clauses were rated significantly worse than the baseline if-clauses. Nevertheless, our conclusion that if-clauses do not appear to be absolute islands for relativization in English is supported by our finding that extraction from if-clauses yielded acceptability ratings that were clearly above the middle range and showed a right-skewed distribution of ratings, characteristic of acceptable sentences. These findings echo those of Kush et al. (2019), who found that the acceptability level of structures involving topicalization out of a conditional adjunct in Norwegian was on a par with the acceptability of topicalizing from a non-island embedded clause.

In the following sections, we discuss our findings in relation to dependency type (Section 5.1), the role of semantics and pragmatics in extractability (Section 5.2), processing factors (Section 5.3), and variation between complementizers and the internal structure of adjunct clauses (Section 5.4).

5.1. Extraction and Dependency Type

Our finding that adjuncts are not uniformly strong islands might similarly call for more fine-grained classifications or for a differentiated approach not only within the group of constructions traditionally referred to as adjunct islands, but also concerning the effect of dependency type on extractability. The fact that extraction from an if-clause appears to be rather acceptable in English may tell us something quite significant not only about the importance of applying a suitable context in this type of experiment, but also about the potential role of the type of dependency that links the filler and the gap. Previous experiments investigating extraction from finite if-clauses in English have primarily tested wh-extraction, and consistently yielded very low ratings for this type of extraction (e.g., Sprouse et al. 2012; Sprouse et al. 2016). A possible reason for the relatively high acceptability of extraction from if-adjuncts in our study is that extraction was in the form of relativization rather than question formation, and that extraction out of adjuncts may be possible for some dependencies such as relativization, but not for wh-movement. If this assumption is on the right track, the results from our study contribute to a growing body of evidence that extraction possibilities seem to differ across dependency types (for English, see Sprouse et al. 2016; Abeillé et al. 2020; for Norwegian, see Kush et al. 2019).

This difference in extraction dependencies is unexpected under traditional syntactic accounts of adjunct islands such as the CED: since wh-movement, topicalization and relativization are all instances of A’-dependencies, they should adhere to the same syntactic locality conditions. However, a syntactic analysis of this phenomenon may nevertheless be possible under the framework of Relativized Minimality (RM; Rizzi 1990), specifically under newer incarnations of RM called featural RM (e.g., Starke 2001; Rizzi 2013; Villata et al. 2016). Featural RM accounts for island effects in terms of feature-based intervention effects: a constituent cannot be extracted if the movement path crosses an intervening element that has the same featural specifications as the moved element. However, intervention effects can be overcome if the extracted phrase is more richly specified in morphosyntactic features than the potential intervener. An intervention account along these lines could potentially be applied to adjunct islands, in particular in the light of previous suggestions that some adjunct clauses (including conditional [if] and temporal clauses [when]) involve movement of an operator to the left periphery of the adjunct clause (e.g., Demirdache and Uribe-Etxebarria 2004; Bhatt and Pancheva 2006; Haegeman 2010, 2012). A potential asymmetry between different extraction dependencies into adjunct clauses could thus be accounted for under the assumption that the operator involved in the adjunct clause in question creates an intervention effect for wh-movement, but not for relativization across it. Such an account presupposes that relativization differs from wh-extraction in involving movement of an element that carries a richer feature specification than the operator present in the relevant adjunct clause (see also Sprouse et al. 2016 for a proposal along those lines).

A problem that this type of analysis faces is that it is not obvious how the featural setup of the relative operator (in relativization) or of the wh-phrase (in wh-movement) interacts with the type of operators that have been suggested to move in adjunct clauses in order to create (or overcome) intervention effects. Assuming that the proposed operators in if- and when-adjunct clauses differ in their feature specifications, it might provide a partial explanation of their different acceptability ratings. Crucially, however, in order to account for the gradience across all three types of adjuncts in our experiment, we would be forced to stipulate some kind of operator in because-clauses.

5.2. The Role of Semantics and Pragmatics in Extractability

Gradient acceptability may, however, not be surprising under a discourse-based account, such as the ones proposed recently by e.g., Chaves and Putnam (2020) (for a range of island phenomena) and Abeillé et al. (2020) (for subject islands specifically), echoing the longstanding idea that backgroundedness (Goldberg 2005) or pragmatic dominance (Erteschik-Shir 1973) are determining factors when it comes to extractability.

More specifically, Goldberg (2005, p. 135) argues that “Backgrounded Constructions are Islands” (BCI), which according to her definition means that it should be difficult to extract from a constituent that is neither primary topic nor part of the potential focus domain (Goldberg 2005, p. 130). Similarly, Abeillé et al. (2020) suggest a pragmatic account termed the Focus-Background Conflict (FBC) constraint, arguing that the island effects attested for certain constructions are due to the discourse clash that occurs when a focused element is part of a backgrounded constituent. They specifically apply this constraint to subject islands, where it can account for the relative acceptability of relativization from DP subjects in English and French, compared to wh-extraction from the same constituents: Wh-movement, but not relativization, puts the fronted element into focus, which clashes with the backgrounded nature of subjects. (Note, however, the very low scores for relativization from subjects in our experiment, cf. Figure 1B). An approach of this nature could be extended to cover the apparent possibility of relativizing from some adjunct clauses that seem to resist wh-extraction, given that adjunct clauses also typically express backgrounded information.

The three adjunct clause types in our stimuli are not discourse-functionally equivalent but have different semantico-pragmatic profiles. If, when, and because all introduce backgrounded adjunct clauses in the sense that the filler refers to a single member of a set in the context, but they differ in terms of veridicality and presupposition: when- and because-clauses are veridical (they have positive polarity and a positive truth value), whereas if-clauses are not (they cannot be assigned a truth value). Furthermore, when-clauses and because-clauses (at least central ones) are presupposed, if-clauses are not (as they are not veridical).

What our data ultimately suggest is that the concept of gradience needs to be properly integrated into an account of island phenomena. Given that the acceptability of extracting from if-, when-, and because-clauses exhibits surprising variation, we still need to identify what exactly determines the differences in the acceptability of extraction from the adjunct clauses under investigation. According to Goldberg (2013, pp. 224–25), “backgrounded constituents of a sentence are not part of what is asserted by the sentence”, and backgroundedness is gradient and independent of truth value, as being true is “not a requirement”. In other words, to some extent, ‘backgrounded’ means ‘not asserted’. One possible way of explaining the variation in the data while maintaining the basic assumptions of the BCI or the FBC could be to appeal to the idea of gradience in backgroundedness: the because-clauses could be argued to be somehow more ‘asserted’ than the when-clauses which in turn could be said to be more ‘asserted’ than the if-clauses (where asserted means ascribed a truth value). In that case, the clash between the adjunct clause and the filler-gap dependency relation by way of relativization should be greater for the because-clauses than e.g., the if-clauses, and this could give us the differences in acceptability found in the present study. It might be argued, then, that the more asserted a clause is, the less acceptable the extraction is predicted to be. If-clauses are not asserted, and they have no truth-value, and therefore, they would be predicted to be permeable and allow for extraction. However, a problem is, then, that the extractions from when-adjunct clauses in our experiment are much more acceptable than expected under this assumption; in a sense, our when-clauses are too ‘good’ (acceptable) to be ‘true’ (asserted), while our because-clauses are too ‘true’ (asserted) to be ‘good’ (acceptable). However, it is not clear that the relevant factor distinguishing between the three adjunct clause types is necessarily degree of assertion and/or backgroundedness (and, in any case, both terms still need to be properly defined and operationalized in terms of scalability). Other contenders might be degree of factivity (Kiparsky and Kiparsky 1970) or projectivity (Tonhauser et al. 2018)5, realis/irrealis (Elliott 2000; Palmer 2001), veridicality (Giannakidou 1998), or indeed relevance, which we turn to now.6

As mentioned above, Chaves and Putnam (2020) argue that many island constraints traditionally assumed to be of a syntactic nature (e.g., the CED) can be reduced to “Relevance Islands”: since adjuncts typically convey backgrounded information, extraction from them tends to violate a pragmatic principle requiring the referent that is singled out by the extraction to be sufficiently relevant for the main action, event, or state of affairs (as originally observed for extractions from complex subjects by (Kluender 2004). However, this pragmatic constraint can, according to Chaves and Putnam (2020, p. 91), be circumvented if extraction occurs in the form of relativization, since relatives ”express assertions rather than backgrounded information”. Chaves and King (2019) found a “positive correlation between the plausibility of the proposition (as expressed by a declarative clause) and the extractability of a deeply embedded referent therein” (Chaves and Putnam 2020, p. 222). The difference in acceptability between the three types of adjunct island complementizers found in the present study might thus be related to the proposition that they create. It seems reasonable to assume that the semantic plausibility of a proposition plays a role in connection with extractability. It is, of course, important that the corresponding declarative or non-extracted version of a sentence be plausible and felicitous (i.e., coherent) in order for extraction to be acceptable (see also Engdahl 1997; Kehl 2022). However, more specifically, the more prototypical and coherent the relations between the “semantic components” are, the easier it is to conceive of a referent as relevant to the main event and hence to extract it (Chaves and Putnam 2020, pp. 222–23). This approach may account for some of the variability attested at the participant level, given that the likelihood of an extracted referent being considered relevant to the main event by the comprehender may be a matter of degree and thus subject to between-participant variation. However, our results suggest rather uniform interpretations across participants, except for extraction from adjunct clauses headed by when (see Figure 3 and Figure 5).

Furthermore, the account presented in Chaves and Putnam (2020) suffers from the same problems that other semantico-pragmatic accounts have, namely the difficulty in operationalizing the approach in a falsifiable way. In fact, one of the few tools that they do suggest makes only partially correct predictions. Specifically, Chaves and Putnam (2020, p. 235) suggest that the acceptability of extracting from an adjunct clause is correlated with the inter-clausal semantic relations hierarchy in Van Valin (2005, pp. 208–9), according to which causal relations express a tighter semantic link or a higher degree of cohesion than purely temporal relations. However, as mentioned above, all three adjunct clause types in our study were embedded under psych-predicates that trigger a causal reading of the events in the matrix and embedded clause. If a causal reading is equally available for all three types of adjunct clauses in our study, they should rank equally high on Van Valin’s hierarchy of semantic relations and also allow extraction equally easily, which is not what we found. Possibly, when-clauses can be argued to express a smaller degree of cohesion than if- and because-clauses since the causal relation is not explicitly encoded in when-clauses but must be inferred pragmatically. If- and because-clauses, on the other hand, linguistically encode causation by means of the introducing element (if and because). This would explain why extraction from when-clauses scored significantly lower ratings than extraction from if-clauses in our study. However, it leaves unexplained why extraction from because-clauses was rated significantly worse than both if- and when-clause extraction.7

5.3. Processing Factors

When- and if-clauses may, however, differ from because-clauses in terms of processing, given that they can occur as a complement or argument. In examples such as (9a) and (9b), the matrix verb appreciate appears to select if- and when-clauses which hence function as arguments in the superordinate clause. Crucially, however, this is not possible with because-clauses, (9b) and (10b) (see Teleman et al. 1999, pp. 568–93 for a similar observation in relation to Swedish). (Examples (9a) and (10a) are from COCA).

9.	a.	I would appreciate if you can post a link to my article to everyone in your
		network
	b.	*I appreciate because you can post a link to my article to everyone in your
		network.

10.	a.	Sir, we always appreciate when you come on.
	b.	*Sir, we always appreciate because you come on.

Based on these data, it seems possible that speakers may be able to assign a structural parse to some sentences involving if- and when-clauses where the if- or when-clause has the status of an argument or complement (similar to the that-clauses used in our stimuli), whereas this is not an option for because-clauses. While the selectional properties of predicative adjectives are different from those of transitive verbs such as appreciate, it may be the case that the parser, holding a filler, postulates an intermediate attachment site at the left edge of the adjunct clause, for example via the principle of Attach Anyway (Fodor and Inoue 1998). This phenomenon is reminiscent of the effects of matrix verb compatibility found for long-distance dependencies: when it is plausible that the filler originates in the matrix clause and intermediate attachment is possible, the resulting acceptability level is higher than if the matrix verb is not compatible with the filler (see Christensen et al. 2013 and references cited there). This in turn lends credence to the CED: while it seems nearly impossible to salvage a purely syntactic version of CED, the latter may be couched in processing or computational terms instead. It is simply easier to parse extractions from complements than adjuncts, and the data in our study may at least partially be explained as a misparse of the adjunct clauses as complement clauses, thus a local ambiguity leading to a global (positive) effect on acceptability. The variation found between if, when and because might be due to the frequency with which they occur as complement clauses (for instance, COCA returns 2.130 hits for “appreciate that” with that specified as a complementizer, 200 for “appreciate if” and 103 for “appreciate when”, while there are no hits of the relevant type for “appreciate because”). The possibility of extracting from adjunct clauses may thus be modulated by frequency, as the parser may use probabilistic knowledge in the online processing of the structures involving extraction from adjunct clauses in the sense that it may temporarily postulate intermediate attachment of if- and when-clauses as complements rather than adjuncts, based on previous exposure, and perhaps this may ameliorate the acceptability level overall. Future research, e.g., using G-Maze (Witzel et al. 2012), could investigate the online processing of these structures and whether the comprehender has a slower reaction time at the complementizer because than for if and when, which would lend support to the idea that the syntactic integration is more difficult, potentially resulting in a decrease in acceptability.

On the other hand, the potential amelioration of temporarily interpreting adjuncts as complements would be the opposite of what has been observed for garden-path resolution. Studies on parsing and ambiguity resolution has shown that garden-path sentences are particularly difficult because the position where the ambiguous constituent is initially integrated does not contain the (correct) target position (Pritchett 1992). In a (non-garden path) sentence such as She knew Tom liked Cats, Tom is temporarily attached as direct object of knew, but has to be reanalyzed as the subject of the embedded clause in the object position. In the garden-path sentence While she dressed Tom made a sandwich, Tom is again first analyzed as object, but has to be reanalyzed as the subject of the matrix clause in which while she dressed is embedded. This reanalysis is effortful and costly, leading to a severe reduction in acceptability.

5.4. Variation between Complementizers and the Internal Structure of Adjunct Clauses

As discussed in Section 5.1, the variation in acceptability as a function of the choice of complementizer (if, when, because) is difficult to explain under a purely syntactic account (or solely under a processing account, see Section 5.3) given that the underlying structure of the adjunct island constructions is generally assumed to be basically the same. However, some have argued that different types of adjunct clauses are adjoined at different positions. For instance, Haegeman (2012) distinguishes between central and peripheral adverbial clauses, which contrast in their relation to the clause that they modify, as the central adverbial clauses are event structuring (modifying the matrix clause at the event level), while the peripheral ones are discourse structuring. In other words, central adjunct clauses modify the event encoded in the host proposition, while peripheral adjunct clauses provide a contextually accessible background assumption contributing evidence for the relevance of the matrix domain (Badan and Haegeman 2022). They differ in attachment height, such that peripheral adjuncts are adjoined to CP, central ones to TP or vP (Haegeman 2012, p. 171), and correspondingly, in their degree of opacity for syntactic operations. However, all three types of adjunct clauses used in this study (if-, when- and because-adjuncts) are central adjunct clauses that are merged at the same height according to Haegeman’s (2012, p. 164) classification. Other proposals connect differences in attachment height to differences in whether or not the adjunct can be interpreted as being in a coherent (e.g., causal) single-event relation with the matrix clause (e.g., Ernst 2001; Narita 2011; Truswell 2011; Sheehan 2013; Brown 2017). However, since all of our target sentences were construed to trigger a semantically coherent interpretation between the matrix and adjunct clause events, these proposals cannot capture the differences across different adjunct clauses observed in our study.

There are some indications that causal adjuncts (because-clauses) have more elaborate syntactic structures than conditional (if-clauses) and temporal adjuncts (when-clauses). In terms of Haegeman’s (2012) classification, causal clauses are merged at the same height as other central adjunct clauses, but seem to display the internal syntax of peripheral adverbial clauses (Müller 2019, p. 98). For example, causal clauses can contain epistemic modal markers, which are usually only possible in peripheral clauses (Ros 2005, p. 98). Causal clauses are also compatible with V2 word order in the MSc. languages, see Teleman (1967) for Swedish, Julien (2015) for Norwegian, and Nyvad et al. (2017, p. 462) for Danish. Based on this evidence, it has been suggested that causal clauses are root-like in that they can encode independent illocutionary force (Müller 2019), which is also in line with the observation that causal clauses may be asserted independently (Hooper and Thompson 1973; Sawada and Larson 2004), whereas temporal and conditional clauses are considered non-assertive (see also Section 5.2).

An account that may be compatible with the presented facts is the cP/CP-distinction first proposed by Nyvad et al. (2017) for Danish: these authors propose an account which credits evident island extractability in Danish (and presumably the other Mainland Scandinavian languages) to the possibility of CP-recursion. Given the apparent acceptability of the structures in question, the grammar must be able to generate them (i.e., the syntax must provide a licit hierarchical representation), and given that cross-clausal filler-gap dependencies adhere to the principle of successive-cyclic movement (e.g., Chomsky 1973; 1995, 2001; Den Dikken 2009; van Urk 2020; Davis 2021), an ‘escape hatch’ must be present. The latter may be provided by cP (“little cP”), a projection found in embedded clauses, which hosts functional elements such as complementizers and which can also be recursive, thus accounting for both the widespread use of complementizer stacking and otherwise unexplained island extractability in e.g., Danish. “Big” CP, on the other hand, hosts lexical elements such as finite verbs in (embedded and main clause) V2. Under this analysis, CPs are always finite, they carry illocutionary force (or illocutionary potential) and they are strong islands. Hence, in this framework, CPs are root clauses or root-like clauses. A similar observation is made in connection with Haegeman’s (2002) proposal that (root-like) peripheral adjunct clauses may have a Force projection, allowing embedded topicalization and V2, while central adjunct clauses do not. See Haegeman (2012, p. 155) for examples of the impossibility of argument fronting in central if- and when-clauses in English.

It may hence be the case that significantly decreased acceptability found for extraction from because-clauses (as compared to if- and when-clauses) is due to the internal syntax of the because-clauses being more complex compared to if- and when-clauses. Whether that is the result of the because-clauses having a Force projection, a root-like C-projection or simply a feature encoding illocutionary force is still an open question, as is the question of how similar English really is to the Mainland Scandinavian languages when it comes to island extractability: while the Mainland Scandinavian languages appear to be able to extract fillers from embedded questions (e.g., Christensen et al. 2013), relative clauses (e.g., Lindahl 2017) and adjunct clauses (e.g., Müller 2019) more or less independently of dependency type (i.e., topicalization, wh-movement or relativization), English may be more limited in its range of possibilities.

Overall, however, relativization out of adjunct clauses appears to be possible in English when presented with a supporting context. If-, when- and because-clauses may thus not be strong islands after all, and they do not form a uniform class, as they seem to have different profiles in our data: the acceptability of extraction from if-clauses was not significantly different from that of that-clauses, even though the two types of structures are viewed as islands (adjuncts) and non-islands (complements), respectively. With regards to the when-clauses, they stand out in a different way: the density plot (acceptability distribution) for this clause type is flatter and slightly bimodal, indicating more variation between participants. In addition, there was a trial effect for the when-clauses such that the participants in the study found extraction significantly better over time, but not for because-clauses, which suggests different underlying causes for reductions in acceptability (Snyder 2022). Following Sprouse (2007, p. 124), this trial effect suggests that extraction from when-clauses is licit, since only grammatical structures should exhibit priming effects (see also Christensen et al. 2013; Christensen and Nyvad 2014). The acceptability of because-clauses did not improve significantly over time, but the DD-score for this structure type was actually below the threshold of 0.75 for islandhood proposed by Kush et al. (2019).

In short, none of the adjunct clause types showed clear signs of strong islandhood. The present study also illustrates the importance of applying different types of statistical analyses. Without this type of ‘profiling’, the multifactorial profiles of extraction from English adjunct clauses would go undetected. We suspect that the acceptability ratings result from interactions between subtle syntactic contrasts, pragmatics, and processing factors. It is very difficult, if not impossible, to tease apart the exact contribution of each of the individual factors to the observed pattern of acceptability.

6. Conclusions

Universal constraints should hold cross-linguistically and across constructions, and certainly across different syntactic constructions of the same underlying type. Apparent cross-linguistic variation in adjunct islands, as displayed by the MSc. languages, has posed a challenge for both formal syntactic accounts and processing accounts of island constraints. However, the reports of crosslinguistic variation are based on comparisons of rather different structures in English and Scandinavian (usually, non-contextualized extraction by wh-movement in English vs. contextually facilitated topicalization in MSc.). Once the impact of different extraction dependencies and of the presence of context on extractability is taken into consideration, the emerging picture is that the crosslinguistic variation between these languages regarding adjunct islandhood may have been exaggerated. Even though our experiment does not directly compare English to other languages, the results are compatible with the view that English is similar to MSc. in that different adjunct clauses display a non-uniform behavior when it comes to extraction. More specifically, the pattern observed in our study for different adjunct clauses matches the patterns found for topicalization from the corresponding adjuncts in Swedish (Müller 2017) and Norwegian (Bondevik 2018; Bondevik et al. 2020): extraction from Swedish and Norwegian if-clauses yielded smaller island effects and better average ratings than extraction from when-clauses, which in turn yielded smaller island effects and better ratings than extraction from because-clauses. This suggests that English and the MSc. languages might be more similar than previously assumed with regards to adjunct islandhood.

The heterogeneity in the acceptability patterns related to adjunct island violations suggests that syntax alone cannot account for the data. Syntactic theory may possibly need to be modified in the face of mounting evidence suggesting that configurational syntax alone cannot explain the variability in acceptability attested across a wide range of traditional island structures and that extra-grammatical factors are also essential in understanding these phenomena.