It has long been observed that a number of constraints apply to the formation of long-distance filler-gap dependencies (Ross, 1967; Chomsky, 1973, 1986; Rizzi, 1990; Boeckx, 2012). One such constraint is relevant to complex noun phrases. Consider the examples in (1). In sentences like (1a), the NP object, the so-called “filler” (those kinds of flowers), can be extracted from its thematic position, the “gap” ([–]), in the argument clause and “moved” long-distance to a left-peripheral position in the matrix clause. However, a similar extraction is prevented when the resulting gap would be in a relative clause (RC) that is part of a complex noun phrase (1b).

We will henceforth refer to the latter extraction as RC extraction (RCE). Complex noun phrases with RCs are standardly considered to be strong islands for extraction: syntactic domains from which extraction is not possible (Ross, 1967). The unacceptability that arises from an illicit extraction (e.g., 1b) is commonly referred to as an island effect.

Island effects have been observed in a great number of languages, spanning numerous language families. Consequently, island constraints have generally been considered to be universal. And yet, a number of recent studies have presented evidence that island effects may show some form of cross-linguistic variation (e.g., Phillips, 2013; Sprouse et al., 2016), raising the question as to whether the constraints themselves are variable. Phillips (2013) presents two general approaches to account for variation in island effects. The first, which we will call the surface variation approach, involves appealing to independently motivated syntactic properties to demonstrate that the violation is only apparent, thus maintaining the universality of island constraints. For example, apparent cases of RC extraction in the East Asian languages have been explained by positing control of a null subject inside the RC by a topic NP outside the RC (Han and Kim, 2004). The second approach entails allowing for deep variation in island constraints to exist between languages. That is, languages may show true underlying variance with regard to which structural configurations constitute syntactic islands (see e.g., Rizzi, 1982).

Certain languages have proved resistant to any surface approach for explaining the absence of expected island effects. Swedish provides just such a case. It has been widely attested that RCEs are acceptable in Swedish (2), as well as in the other Mainland Scandinavian languages (Erteschik-Shir, 1973; Allwood, 1976; Taraldsen, 1981; Engdahl, 1982, 1997; Christensen and Nyvad, 2014), with examples being found in both spoken and written language (Lindahl, 2017)¹.

To account for this apparent acceptability, a number of surface variation analyses have been proposed, including a discourse organization account (Erteschik-Shir, 1973), a covert resumption analysis (Cinque, 1990), and a small clause structure analysis (Kush et al., 2013). However, it has been demonstrated that none holds up under closer scrutiny (Engdahl, 1997; Boeckx, 2012; Christensen and Nyvad, 2014; Heinat and Wiklund, 2015; Lindahl, 2015; Müller, 2015). Thus, from a syntactic perspective, it remains unclear whether the apparent acceptability of Swedish RCEs represents a case of deep variation in island constraints or whether the issue continues to be a matter of finding a tenable surface structure account for the data.

As a means of guiding future syntactic research, the current study uses eyetracking while reading to investigate whether the apparent acceptability of Swedish RCEs has any processing correlate at the point of filler integration (the gap) compared to uncontroversial strong island violations. One possibility is that although RCEs are intuitively acceptable in Swedish, they may actually be processed more similarly to island-like structures. This would suggest that their apparent acceptability is attributable to some type of off-line repair mechanism, perhaps dependent on language specific discourse organization or contextual factors, and that the exceptionality of these structures therefore does not comprise a case of deep variation in island constraints. A second possibility is that the processing of Swedish RCEs may match their apparent acceptability, and that they will be processed more similarly to non-islands. Such an outcome would suggest that either a tenable surface account is still needed or that deep variation, on some level, does indeed exist.

No previous online processing studies of Swedish RC islands exist; however numerous studies have used online methods to investigate filler-gap dependencies that extend into island domains in other languages (primarily English). One point of general consensus from this body of research is that the parser appears to respect island constraints at the earliest stages of processing, effectively blocking the formation of filler-gap dependencies that extend to island domains while allowing them in licit structures (Stowe, 1986; Bourdages, 1992; Traxler and Pickering, 1996; McElree and Griffith, 1998; Omaki and Schulz, 2011; Omaki et al., 2015; cf. Freedman and Forster, 1985; Pickering et al., 1994; Phillips, 2006; Wagers and Phillips, 2009).

For example, Traxler and Pickering (1996) used eyetracking while reading to test whether the parser immediately employs structural island information to preclude filler-verb integration in sentences involving extraction from English RCs, which are known to be strong islands.

In their design, the optional transitivity of the embedded verb (wrote) creates a temporary ambiguity as to the actual location of the extraction site (gap) of the filler, which is ultimately disambiguated at a second gap position. By manipulating the plausibility of the NP filler (book vs. city) and monitoring for plausibility mismatch effects at the embedded verb (first gap), they were able to test whether the parser immediately acts to inhibit dependency formation at the embedded verb in strong islands (3a), relative to a non-island control (3b). They found that plausibility effects, manifest as longer fixation durations for implausible filler-verb relations (city-wrote) than plausible ones (book-wrote), were present for the non-island condition in both first fixation durations and total fixation durations at the embedded RC verb, but not for the island condition. From this, Traxler and Pickering drew two primary conclusions. First, that the parser actively attempts to complete unbounded dependencies and fill gaps at the earliest possible moment, regardless of the presence of an ambiguity. Second, that dependency formation is blocked in cases where the parser is unable to construct a dependency that will result in a licit constituent, such as those involving RC island extractions. The differential nature of this general finding in regards to islands and non-islands thus provides a heuristic for determining the island status of a given construction from a processing perspective.

In the current study, we enlist use of a structural manipulation (Experiment 1), as well as a variant of the plausibility heuristic from Traxler and Pickering (1996) (Experiment 2) to address the island status of Swedish RCEs and thereby direct future work on variation in island effects.

Experiment 1 uses an eyetracking while reading paradigm to investigate whether Swedish RCEs (4b) elicit processing costs similar to licit extractions from that-clauses (TCE) (4a) or instead pattern closer to extractions from non-restrictive RCs (NRCE) (4c), which are reported to be strong islands in Swedish (Engdahl, 1997; Teleman et al., 1999; Platzack, 2000).² For our comparison, we assume a link between differential results and successful or unsuccessful integration of an extracted filler (Såna där gamla skottkärror/“such old wheelbarrows”) at the first embedded verb in the dependent clause (tvättade/“washed”) and the following two-word PP region (på bensinmacken/“at the gas station”).

We also include an intransitive control condition, PCRCE (4d) (also allowing extraction), which is identical to RCE, but for which the embedded transitive verb is pseudo-coordinated with a preceding intransitive light verb (stod/“stood”) (Teleman et al., 1999; Wiklund, 2007). Thus, for PCRCE the first embedded verb region differs from the verb in the other three conditions (stod/“stood” vs. tvättade/“washed”). Being intransitive, the first embedded verb in PCRCE is expected to provide a baseline for what processing would look like in cases where no integration is attempted (cf. Omaki et al., 2015). In this way, PCRCE is expected to help assure that any noted differences between RCE and NRCE are related to integration and islandhood and do not solely represent any independent cost arising from structural complexity.

Although the words in our analysis regions are held maximally constant across structures (excepting the first embedded verb in the PCRCE control), there are differences in earlier sentential regions, which have the potential to affect fixation durations at the embedded verb. For example a different adverb is used for NRCE than in the other conditions (to induce the island), and relativization is present in all conditions except TCE. To best circumvent any potential confounds related to these differences, our approach takes as a starting point the widely accepted generalization that different eyetracking measures reflect different costs that are likely to emerge in the course of processing (see discussion in Clifton et al., 2007). By enlisting different measures together with an appropriate control condition, we expect to be able to tease apart structural effects related solely to differences in structural complexity (e.g., additional syntactic heads or underlying structure) and effects related to complexity and integration (i.e., actual dependency formation over intervening material), the latter of which we will assume more accurately reflects islandhood costs. We detail our assumptions about these costs below, and then follow with description of the eyetracking measures used in our study and their relation to different costs, as well as our specific predictions for each condition and measure.

Our cost assumptions are largely cast within Dependency Locality Theory (DLT) (Gibson, 1998, 2000), which posits that sentence complexity related to dependency formation and integration is affected by increases in memory load and can be coarsely determined via the number of new discourse referents intervening between a head/antecedent and its integration site. To the extent that DLT specifies, the intervening material in our sentences is equivalent between conditions. However, it has been proposed that the DLT may not sufficiently predict processing costs on its own. For example, Demberg and Keller (2007) argue that an extra source of processing complexity arises when the verb-filler dependency extends over a clause boundary; and Alexopoulou and Keller (2007) provide evidence that other types of heads, such as auxiliary verbs, contribute to complexity. Our materials bear features that are similarly not accounted for in the DLT, but which could also have relevance to complexity and integration. First, RCE, PCRCE, and NRCE all contain relativization, whereas our TCE condition does not. Relativization has been demonstrated to increase memory load, which may become manifest as locality effects if working memory resources are sufficiently taxed (Levy and Keller, 2013). Second, our NRCE condition is likely to have additional underlying structure (Haegeman, 2002, 2010; de Vries, 2012), which may or may not contribute to its overall complexity (depending on what counts as “complex”), but which is arguably instrumental to its structural island status. Both of these differences could accrue additional cost. In addition, it has been argued that processing is best explained not only by locality principles but also by expectation (e.g., Demberg and Keller, 2007; Levy et al., 2012; Levy and Keller, 2013). Our PCRCE control utilizes an intransitive verb in the critical region, whereas RCE, TCE, and NRCE all use the same transitive verb, a difference which could conceivably be reflected in expectation-based costs. We return to these differences below, when stating our predictions, but before doing so, it is necessary to first present the eyetracking measures used in the study.

It is generally held that different eyetracking measures reflect different processing costs. We analyze four such measures in our experiment. First fixation duration (the duration of the first fixation on a region during first pass reading) has been shown to be sensitive to lower level processes such as word recognition and general increases in complexity (Rayner et al., 1989; Hyönä and Vainio, 2001; Vainio et al., 2003; Staub et al., 2006; see review in Clifton et al., 2007). Gaze duration (the sum of all fixations in a region during first pass reading) is also often treated as a measure of complexity, but it has additionally been shown to reflect costs associated with actual integration during dependency formation (Rayner et al., 2004; Staub et al., 2007; Warren and McConnell, 2007; Demberg and Keller, 2008). Likewise, measures targeting regressive eye movements such as regression path duration (the sum of all fixations on the target region plus regressive fixations to prior regions until the target region is exited to the right during first pass reading) are also thought to reflect integrative stages (Rayner et al., 2004; Traxler et al., 2005; Gordon et al., 2006; see discussion in Staub, 2010). Finally, total duration (the sum of all fixations on a region) is held to reflect textual integration processes, lexical and syntactic/semantic processing, and general comprehension difficulty, and thus the overall holistic, cohesive success of the sentence (Boston et al., 2008; Demberg and Keller, 2008). The first two measures are often referred to as “early” measures and the last two as “late” measures, as a generalization about the stages of processing they roughly represent.

We turn now to our predictions for Experiment 1, beginning with the clearest contrast, that between TCE and NRCE. Though not tested directly in their studies, both Omaki and Schulz (2011) and Traxler and Pickering (1996), using materials roughly similar to our TCE non-island and NRCE strong island conditions, showed a numerical advantage for their plausible non-islands relative to their plausible islands. This pattern was seen in first fixation durations and total durations at the verb region for Traxler and Pickering (1996), and in self-paced reading RTs (roughly analogous to eyetracking measures marking integration) in post-verb regions in Omaki and Schulz (2011). Given these numerical distinctions and the differences noted above for NRCE relative to TCE (relativization, additional structure/complexity), which acting together we assume contribute to its islandhood status, we expect that TCE will show greater facilitation than NRCE at the verb region (and a following PP region), across the four eyetracking measures³. However, we also expect that this difference will not always be for the same reason. In order to make direct comparisons between islands and potential non-islands, it is critical to first tease apart costs that are purely related to structural complexity and those that involve complexity insofar as it relates to integration and dependency formation, the latter of which we view as a construal of islandhood. For this reason, we include our PCRCE control.

The embedded intransitive verb in PCRCE is arguably less structurally complex than the corresponding transitive verb in NRCE, TCE, and RCE. PCRCE should thus show the shortest durations (after adjusting for word length via a residualization procedure-described in section Data Analysis) for any measure solely reflecting complexity (e.g., first fixation duration). In contrast, for measures involving actual integration and dependency formation, encountering and accessing the argument structure of the intransitive verb in PCRCE should comprise a costly breach of expectation, provided such integration is attempted, as it lacks the necessary transitivity to accommodate the open dependency. The presence of a cost for PCRCE relative to the transitive conditions would thus provide verification that processes of integration and dependency formation were at least considered by the parser. In turn, any cost then shown by NRCE relative to TCE for that measure could be attributed to integrative processes and issues with dependency formation related to the islandhood of NRCE. And most importantly, the patterning of RCE could be assessed relative to TCE and NRCE. Based on the attested acceptability of RCEs, our prima facie expectation is that RCE will pattern together with TCE as faster than NRCE in integrative measures involving dependency formation, as confirmed by the patterning of PCRCE, and that this will be interpretable as support that RCE involves a licit extraction.

In Experiment 2 we again use an eyetracking while reading paradigm to investigate the processing of RCEs, but enact two key changes to our materials to address aspects of our experimental design that may have engendered forced integration in the NRCE condition and which prevented us from unequivocally drawing conclusions regarding RCE in Experiment 1. First, to reduce the risk of forced integration, we implement a dual-gap design with a filler-verb plausibility manipulation similar to that used in Traxler and Pickering (1996) and Omaki et al. (2015). This also allows for a greater degree of comparison to previous studies, in that we can now focus on comparisons within a particular structure and match those studies in terms of task demands. Second, we include a within-structure plausibility manipulation to remove the necessity of using maximally similar structures and to allow us to use another type of strong island, extractions from subjects (SRCE). SRCE has traditionally been considered to comprise a strong island violation, the Subject Condition (Huang, 1982; Chomsky, 1986), and unlike RCE, this violation has been observed to be respected in Swedish (Engdahl, 1982). In addition, SRCE provides an island structure that involves extraction from an RC, and is thus consistent in terms of relativization compared to RCEs, but is not dependent on the presence of a specific type of adverb to induce degradation, as was the case with NRCE.

Our design thus includes two Structure variants: RCE (6c, 6d) and SRCE (6a, 6b) and a plausibility manipulation (Plausibility). For this, each sentence began with a filler-NP headed by a concrete noun which functioned as either a semantically/pragmatically congruent filler (Plausible) to the embedded RC verb (möbler/“furniture”, as in 6a for SRCE; 6c for RCE) or as an incongruent filler (Implausible) (flyttlådor/“moving boxes”) to the verb (6b for SRCE; 6d for RCE).

One tradeoff to enlisting a dual-gap design is that the presence of a second gap potentially opens the door for parasitic gapping (PG), a phenomenon whereby an illicit gap is rescued in the presence of an additional licit gap position in the sentence, and both gaps are linked to the same filler phrase (Ross, 1967; Culicover, 2001; Phillips, 2006). Parasitic gapping inside complex subjects has been claimed to be possible in Swedish (Engdahl, 1983), as well as in English. The illicit gap in (7a) (the subject island violation) can be rescued by being parasitically linked to a licit gap, thus rendering the sentence acceptable, cf. (7b) (example sentences borrowed from Phillips, 2006 p. 802).

Unlike English, Swedish is claimed to allow parasitic gapping in finite clauses to some degree (Engdahl, 1983). It is consequently more difficult to preclude the occurrence of parasitic gapping in our sentences than it would be, for example, in an experiment using English language materials that was also based on the general Traxler and Pickering (1996) design.

However, there are two reasons why we are optimistic that the threat of parasitic gapping will be minimal. First, relativization in combination with tense (as in our RCs) appears at the lowest end of the hierarchy of accessible parasitic gap domains, as proposed in Engdahl (1983). Second, our use of a PP (på landet/“at the countryside”) between the two gap positions could help to strengthen the parser's commitment to an intransitive analysis of the embedded RC verb and thus diminish the potential for parasitic gapping (see Chaves, 2013). Despite this, participants may become aware of the dual-gaps over the course of the experiment, thus potentially engendering parasitic gapping. To address this, we model the effect of trial in each of our analyses. This allows us to monitor parasitic gapping to the extent that it may arise from a learning effect over the course of the experiment.

A second tradeoff is that the number of dependents crossed is no longer equal; RCEs cross an extra dependency relative to SRCEs (jag/“I”) at the first gap. However, this imbalance is critically biased against RCEs integrating more successfully than SRCEs.

If RCEs are licit, non-island structures, as Experiment 1 suggests, and if the dual-gap design reduces the potential of forced integration at the embedded RC verb region (R01) and the following PP region (R02) without introducing parasitic gapping, then interactions should be present in early and late measures whereby RCEs, but not SRCEs (on account of their island status) show facilitation for Plausible relative to Implausible at R01/R02.

The current study is the first of its kind to investigate the processing of relative clause extractions in Swedish. In Experiment 1, we directly compared the processing pattern for RCEs to that of a strong island violation (NRCE), as well to licit extractions from that-clauses (TCE) and an intransitive control (PCRCE) condition. RCEs were seen to pattern differently from NRCE strong islands and were often processed in a similar way to licit that-clause extractions (TCE), despite RCEs being more similar to the island conditions in terms of relativization. In Experiment 2, we manipulated the plausibility of filler-verb pairings for RCEs and another type of strong island violation (SRCE), similar to the approach taken by Traxler and Pickering (1996), allowing us to monitor for signs of semantic integration independently within each structure. Integration-based processing differences were observed between RCEs and SRCE strong islands, present both as simple effects of Plausibility and Structure across eyetracking measures, and as differences in potential learning effects over increases in trial as participants proceeded through the experiment. Structural facilitation was also observed for RCE relative to SRCE despite any cost for crossing an additional discourse referent in the RCE structures. Taken together, the results from Experiments 1 and 2 suggest that Swedish RCEs are processed differently than at least two types of strong island structures, and therefore are not likely to comprise a case of strong island violation. Our findings thus provide a processing correlate to the intuitive acceptability of RCEs, as reported in the literature.

However, it was not clear from our findings that Swedish RCEs should be treated as non-islands, on par with TCEs. In total durations (Experiment 1), a measure of the overall cohesive success of the sentence, RCEs were seen to pattern in-between non-islands and islands, which echoes their below-midpoint ratings presented herein and in Müller (2015). This difference may simply reflect a late cost of relativization, but it could also signal that RCEs are more island-like than TCEs in the sense that integration is associated with more cost in these structures. We first explore the latter possibility, and then discuss the implications of our findings in terms of deep vs. surface variation.

In a recent proposal, Lindahl (2015) treats RCEs as a type of licit weak island extraction (WIE). Descriptively speaking, weak islands differ from strong islands in allowing certain extractions to occur from the island, while banning others (Rizzi, 1990). A paradigmatic case is wh-islands, (8) (from den Dikken and Szabolcsi, 2002 p. 219), which marginally allow argument extraction but not adjunct extraction.

A WIE approach has in fact also been made to explain the acceptability (for some speakers) of extractions out of indefinite NPs with a relative clause in English (Postal, 1998). A challenge to the application of a WIE analysis for Swedish RCEs has been that many of the suggested restrictions on RCEs (including those relating to the head noun, the matrix predicate, the matrix subject, and the position of the relative clause) either do not hold up to scrutiny or are derivable from independent factors (Heinat and Wiklund, 2015). Moreover, Swedish seems to differ from the English varieties analyzed in Postal (1998) in allowing extraction even from definite DPs and in allowing PP-/adjunct extraction from the RC (Heinat and Wiklund, 2015). Despite this exceptionally liberal behavior of Swedish RCEs, Lindahl (2017) identifies at least one restriction that applies to them. She shows that certain wh-adjuncts (varför/“why”) cannot be extracted from the relative clause (9), and in doing so provides an argument that Swedish RCEs, while not strong islands, are weak islands in the sense presented above.

A WIE approach to RCEs is compatible with the results in the current study, in that we would expect licit WIEs to be processed either on a par with our non-island condition, or between the non-island and the islands condition if for example the cost for relativization is high enough to influence the pattern of results, or the absence of context affects weak islands more than non-islands. The weak island approach, but less clearly the non-island approach, also appears compatible with the somewhat low ratings that RCEs receive in formal experiment situations, as WIEs are expected to require contextual conditioning which is often absent (for practical reasons) from acceptability experiments. Successful WIEs have often been claimed to be conditioned on specific readings in relation to the filler, for example, an existential presupposition reading (den Dikken and Szabolcsi, 2002; Boeckx, 2012). This suggests that contextual conditioning is very important and is a factor which should receive more attention when studying the processing of RCEs.

Turning to variation in island effects and the concomitant question of whether true (deep) variation in island constraints exists to explain the former variation, our findings do not preclude any conclusion in support of deep variation. However, if Swedish RCEs involve licit weak island extractions, as we have suggested above, then these structures instead appear to conform to the proposed universality of island constraints, proven to be quite robust despite a few exceptions. On this account, Complex NPs with relative clauses (indefinite and definite) are islands in Swedish, like in English, only much weaker islands than their English counterparts, the latter either strong islands (standardly) or weak islands (in the varieties reported on in Postal 1998).

Insofar as we understand Phillips' (2013) characterization of variation in relation to island effects, any variation cast in terms of island strength must derive from surface variation (structure) rather than deep variation in the island constraints themselves. Quite standardly, a non-island domain (to the extent that this extreme exists) is a structure simple enough to allow all extraction. A strong island domain (to the extent that this extreme exists) is a structure complex enough to ban all extraction. Different types of weak island domains fall in between these two extremes but they share the property of allowing some and banning other extractions. Which extractions are allowed, depends on the (intervening) features/structures that create the island (Starke, 2001; den Dikken and Szabolcsi, 2002; Boeckx, 2012).

If relative clauses are weak islands, and the set of admissible extractions varies between languages as a function of the features/structures present in the relative clause, then we need to identify what features of the relative clause comprise the locus of variation and the different structures that these features project to create the variation in evidence. The relevant features need not necessarily relate solely to relativization. One possibility is that relativization in combination with finiteness is what induces island effects in relative clauses (generally across languages), but that features relating to finiteness comprise a locus for variation in regard to RCE. The latter receives support from a recent study on extraction from adjunct islands by Müller (2017), where it is shown that extraction from finite (contingent) adjuncts is possible in Swedish, despite being impossible in English (Truswell, 2011). Observations like these, taken in consideration with the results from the current study, suggest that surface variation rather than deep variation in island constraints may be responsible for any variation; and that further investigation at the micro-level (features) is the right way to move forward toward a full understanding of variation in island effects across languages.

This study was carried out in accordance with the recommendations of The Swedish Research Council (see Swedish Research Council, 2002: “Forskningsetiska principer inom humanistisk-samhällsvetenskaplig forskning”) with written informed consent from all subjects. The study protocol did not require approval under Swedish national ethics board (Etikprövningsnämnden) rules, and was thus not submitted for review.

DT experimental design, experiment building, data collection, statistical analysis, interpretation, writing; FH experimental design, stimuli creation, data collection; EK stimuli creation, data collection; A-LW experimental design, interpretation, writing.