Past research suggests that all three A’-dependency types, wh-movement, relativization, and topicalization, can cross into EQs in Norwegian. Maling & Zaenen (1982) report (10) as an acceptable example of wh-movement. Moreover, in a series of acceptability judgment studies, Kush et al. (2018) found that Norwegian participants frequently accept (argument) wh-movement dependencies from embedded polar questions like (11).

Although Kush et al. found that participants occasionally rejected wh-movement from an embedded question slightly more often for bare wh-arguments (hva) than complex wh-phrases (hvilke kaker), the authors concluded that the overall high probability of acceptance entails that wh-movement from EQs is syntactically possible in Norwegian.

In two large-scale judgment studies, Kush & Dahl (2020) found that relativization from EQs was judged as acceptable on average as relativization from declarative complement clauses. Kush & Dahl specifically investigated relativization of the subject from object EQs like (12) or adjunct EQs. Other gap sites and EQ types (e.g. polar, subject, etc.) were not investigated, but the presumption is that the acceptability of (12) implies the general possibility of argument movement from other EQ-internal positions.

Kush et al. (2019) find that topicalization from embedded polar questions like (13) is equally acceptable as long-distance topicalization from declarative complements (see Bondevik, Kush, & Lohndal 2020 for replication).

The studies mentioned above only investigated the acceptability of argument dependencies into EQs. We know of no studies that have tested the acceptability of adjunct dependencies into Norwegian EQs. Norwegian seems to follow the general cross-linguistic pattern that adjuncts cannot be extracted from EQs5 (Cinque 1990; Rizzi 1990; Szabolcsi & Zwarts 1993, a.o.). Informal judgments support this assumption:

Describing the distribution of acceptable A’-movement from Norwegian RCs is slightly more complex than with EQs because the extant literature provides conflicting reports on which dependencies are acceptable. Researchers agree that at least some RCs permit some A’-movement dependencies and that some movement from some RCs results in unacceptability. We discuss the uncontroversial cases first, before proceeding to more controversial cases.

RC-island violating dependencies most often feature topicalization from two types of RCs (e.g., Engdahl 1997; Lindahl 2017 for Swedish). The first type are presentational/existential subject RCs like (15), which primarily function to introduce new referents to a discourse. The second type are (ii) it-clefts like (16), which place focus on the head of the RC (Hedberg 2000; Prince 1978; see Gundel 2002 and Johansson 2001 for cross-linguistic comparison of English and Scandinavian it-clefts).6

RC-island violations are, however, not limited to presentational or cleft constructions. Engdahl (1997) observes that dependencies into predicate nominals like (17) are acceptable.

Multiple authors have argued that other embedding environments are possible. Naturalistic examples of topicalization from RCs attached to nominals in object position are attested (Erteschik-Shir 1973; Allwood 1982; Taraldsen 1982; Engdahl 1982, 1997; Lindahl 2014, 2017; Löwenadler 2015). Examples like (18a) with embedding predicates such as kjenne (‘to know/be acquainted with’) are relatively common and documented in the literature. Examples with other predicates such as (18b) and (18c) are rarer, though they are judged acceptable.

It is often observed, however, that filler-gap dependencies into structurally identical RCs, like (19), are likely to be judged unacceptable – particularly if presented in isolation.

The differences between (18) and (19) indicate that the acceptability of topicalization from RCs is at least partially sensitive to the embedding predicate. Importantly, however, many examples like (19) that informants initially reject become more acceptable given a sufficiently rich context (Engdahl 1997; Lindahl 2017). Given that examples like (19) can improve in context, and we assume that there is no principled structural distinction between the RCs under predicates in (18) and (19; contra Kush, Omaki, & Hornstein 2011), it would seem that the embedded predicate effects should be given an extra-syntactic explanation.

All the RCs above are subject RCs. Some researchers (e.g. Platzack 1999) have maintained that only subject RCs allow filler-gap dependencies, but others (Engdahl 1997; Lindahl 2017) have challenged this claim with naturalistic examples of filler-gap dependencies into non-subject RCs. The Norwegian (20), based on one of Engdahl’s examples, involves topicalization from an object RC. Our informants judge (20) acceptable, thereby arguing against a rigid subject-only restriction.

Taken together, the data suggest that Norwegian children must learn that topicalization from RCs is, in principle, acceptable in different embedding contexts (existential and presentational constructions, clefts, and RCs attached to NPs in object position) and with different RC types (subject, object, etc.). They must also learn the (as yet) poorly understood additional factors that lead some RC-island violating dependencies to be judged unacceptable, such as the fact that the embedding predicate appears to affect the acceptability of dependencies into RC. We return to this issue in the general discussion.

The last issue to address is the acceptability of different filler-gap dependencies into RCs. In the prior literature, most, if not all, attested examples involve topicalization. Kush et al. (2018) found that Norwegians generally reject wh-movement from non-presentational RCs presented out of context. We do not know of any work that has formally investigated the acceptability of relativization from RCs. In light of the relative abundance of examples with topicalization, the dearth of wh-movement and relativization examples is noteworthy. As we see it, there are at least three ways to interpret the asymmetry: First, we could assume that movement dependencies are uniformly blocked from all RCs and that the apparent cases of topicalization from RCs are not true cases of movement. Second, we could assume that specifically wh-movement, relativization, or both are blocked from RCs. Third, we could assume that wh-movement and relativization from RCs is possible in principle, but individual examples are ruled out by supplemental restrictions, similar to the issue of embedding verb choice.

We reject the first possibility: Topicalization from RCs exhibits characteristics of a standard movement dependency (see Engdahl 1997; Lindahl 2015, 2017): for example, topicalization from RCs exhibits case-connectivity effects (21a), PPs (21b) and certain adjuncts (21c) can be topicalized, and gaps cannot be replaced with resumptive pronouns (21d). Gaps of topicalization also license parasitic gaps (21e; see also Lindahl 2017).

We also dismiss the possibility that either wh-movement and relativization from RCs is banned outright. It is relatively easy to show that relativization from RCs is possible: Naturally-occurring examples can, for instance, be found online:7

Naturally-occurring examples of wh-movement from RCs is acceptable are scant. However, it appears possible to construct acceptable examples. For example, we constructed (23) – and similar examples – and consulted 10 different native speaker informants. All accepted the examples.

The existence of acceptable cases of relativization and wh-movement from RCs argues against a wholesale ban on these movement dependencies from RC. Nevertheless, it seems that the frequency of examples differs starkly by dependency type: Topicalization from RC is well attested, relativization significantly less frequent, and wh-movement is extremely rarely observed in the wild: For example, Lindahl (2017) found that a collection of 270 naturally-occurring Swedish examples included 93% topicalization, 7% relativization, and no instances of wh-movement from RC (p. 150). The fact that some examples of relativization or wh-movement are acceptable, but other structurally similar sentences are unacceptable, suggests that the distribution of these dependencies is likely governed by supplemental contextual factors above and beyond simple syntactic well-formedness (Allwood 1982; Engdahl 1997; Lindahl 2017; Kush et al. 2018, 2019). Such extra-syntactic factors might also be able to explain the cross-dependency differences in frequency. We return to this idea in the general discussion.

Based on formal and informal judgments, we conclude that all six combinations of filler-gap dependency type and island-type (EQ, RC) are, in principle, allowed in Norwegian, though certain combinations appear less frequently and their acceptability seems to be subject to fine-grained contextual factors. Given that the target grammar appears to allow all three types of dependencies from both island types, we would expect the input to children to contain positive examples of the six different dependency-island combinations if the target distribution must be learned via direct positive evidence.

The NorGramBank children’s fiction corpus contains text from 155 children’s books taken from bokhylla.no, managed by the National Library of Norway. The corpus was created as part of the Norwegian Infrastructure for the Exploration of Syntax and Semantics (INESS) project (Rosén, De Smedt, Meurer, & Dyvik 2012; accessible at http://clarino.uib.no/iness). The corpus comprises 4111212 words and 389556 sentences automatically parsed in the LFG formalism with XLE (Xerox Linguistic Environment; Maxwell & Kaplan 1993; Kaplan et al. 2002) and the LFG Parsebanker tool (Rosén, Meurer, & De Smedt 2009). Sentences are annotated with constituent-structure and functional-structure (henceforth f-structure) features which can be queried using the INESS-search interface.

Although a portion of the corpus was disambiguated by annotators at INESS after automatic parsing, most sentences in the corpus are associated with multiple candidate parses. INESS-search queries all parses associated with each sentence, which can lead to a large number of false positives for complicated queries. To counteract the effect of false positives on our frequency estimates, we manually checked all results for queries attempting to identify island violations. For very broad searches (e.g. estimating the total number of wh-, relative clause, or topicalization dependencies), we did not manually check if all examples were correctly identified.

Previous corpus work on the distribution of long-distance dependencies in children’s input has focused on wh-question dependencies (e.g. Pearl & Sprouse 2013), but we collected frequencies for three different dependency types: wh-questions, relative clauses, and topicalization. Queries were conducted using both constituent-structure features (e.g. phrasal categories, dominance relations) and f-structure features, such as clause type, dependency type, and functional role annotation (e.g. subject, object).

The corpus contains a diverse array of reading material intended for different age groups, from picture books to young adult novels. Since text characteristics may differ by target age group and aggregate frequencies for the whole corpus may not accurately reflect the relative probabilities across age groups, we separated books into four rough age groups. The age group for each book was taken from its designated reading level (målgruppe) at the Oslo Public Library website (deichman.no). If a reading level was not available from the library, we consulted a variety of online booksellers. Finally, in the rare event that these searches did not provide a classification, acquaintances with a background in child education were consulted. The four age groups used were: ages 3–5, 6–8, 9–11, 12–18.

The term ‘child-directed’ is typically reserved for input to language learners under the age of 9 or younger. This reflects the common assumption that children complete the acquisition of major syntactic constructions in their native language at a relatively young age. The 9–11 and the 12–18 age groups fall outside the standard range for ‘child-directed’ language. We nevertheless opted to include frequencies from these age groups as a point of comparison against the ‘child-directed’ text in the two lower age groups to get a better sense of whether and how relevant frequencies varied by age. We base our conclusions, however, on the results from the two lower age groups.

Table 1 gives the total number of sentences in each age-separated sub-corpus, as well as the total number of embedded declarative and interrogative clauses. The subcorpus of texts for 3–5 year olds has the smallest number of sentences (partly reflecting the fact that books in this subcorpus were significantly shorter than the rest). Embedded declarative clauses and embedded questions are relatively infrequent, as has been observed in previous corpus work in Norwegian (Westergaard & Bentzen 2007; Ringstad 2019).9

We began by calculating the base frequencies of the three movement dependencies, so that we could subsequently report the relative frequencies of island violations by dependency type. We restricted our search for wh-questions to sentences where a wh-word was followed by a verb (see Pearl & Sprouse 2013). This was done to exclude fragment questions (e.g. ‘Who/What/Where/For what?’), which do not constitute overt evidence of a wh-dependency. Our search terms for topicalization specified sentences in which the constituent marked as the topic was not the matrix subject, as, again, such sentences do not provide overt evidence for movement. The counts for each dependency type split up by age group are found in Table 2.

The overall frequency of main wh-questions varies between 1.8–2.7% across age groups in the corpus. Westergaard (2005) reports that main questions make up roughly 8% of Norwegian CDS, similar to reports that main questions comprise between 8%–16% of English CDS (Cameron-Faulkner & Noble 2013). We can thus estimate that main wh-questions are 4–8 times less likely in Norwegian book text than in CDS. This estimate aligns with findings from Cameron-Faulkner & Noble (2013), who found that main questions were roughly six times more common in English CDS than in a sample of picture books for 2-year-old children.

Relative clauses were observed more frequently than main questions. Following previous findings, we assume that relativization structures are more common in text than they are in CDS, perhaps by as much as an order of magnitude (see Montag 2019).

The overall frequency of topicalization structures is high due to Norwegian’s propensity to front non-subjects to sentence-initial position. The rate of topicalization observed in our corpus is roughly the same as rates of topicalization reported to occur in Norwegian CDS (23.4% according to Westergaard 2005). This rate includes topicalization of all phrase types (e.g. both arguments and adjuncts).

The numbers in Table 2 aggregate over local movement dependencies and long-distance movement dependencies (i.e. where the filler occupies a position outside the clause containing its gap). Island violations are long-distance dependencies, so establishing that long-distance movement is possible is a precondition for entertaining the possibility of island violations. Evidence of basic long-distance movement has been shown to be relatively infrequent in the input to children in languages like English (Pearl & Sprouse 2013; Bates & Pearl submitted). We wished to confirm whether the relative rarity of long-distance movement was also characteristic of Norwegian child-directed text.

To obtain the approximate frequency of standard long-distance movement dependencies, we searched for sentences with one or more complement clauses that also contained a filler-gap dependency. We then manually identified sentences in the results that contained examples such that a filler was associated with a gap across at least one embedded declarative complement clause. Results are in Table 3. The amount of text differs across the sub-corpora, but long-distance examples of each dependency type are found in each subcorpus.

We give examples of long-distance filler-gap dependencies into embedded declaratives from the corpus. We provide the book title and sentence number for all examples.

To summarize the basic results: When collapsing across local and long-distance dependencies, topicalization is most common, followed by relativization. Wh-movement dependencies are less frequent than in regular CDS, in line with previous findings regarding the differences between written and spoken corpora, but wh-movement dependencies still occur at non-trivial rates in the corpus. Long-distance dependencies from declarative complements are, on the whole, infrequent, but the absolute frequency of long wh-, relativization, and topicalization dependencies is comparable.

Having established the distribution of simple filler-gap dependencies, we turned to island violations. We conducted distinct searches for wh-, relativization, and topicalization dependencies into RCs and EQs. We used relatively broad search criteria to avoid missing potential hits and then manually sifted the results (see Appendix A for search queries). In the entire corpus we found 63 examples of dependencies into RCs and 42 examples into EQs. Table 4 presents the counts of movement dependencies from RCs and EQs split by age group, dependency and embedded clause type. It also includes the counts of dependencies from simple declarative clauses and the overall count of each clause type without a long-distance dependency to adjust for base rate differences across construction type. Finally, in each cell corresponding to a different movement-island pairing, we provide the expected count10 E[…] of tokens that should have been observed if movement from that island was comparable in frequency to movement from an embedded declarative.

As seen in Table 4, direct evidence for long-distance dependencies into RCs and EQs occurs in the input to children across age groups. The frequency of evidence differs both by embedded clause-type and dependency type.

For each age group we conducted two 2x4 Fisher Exact tests in R (R Core Team 2018), comparing the counts in the embedded declarative column to the corresponding RC and EQ column, respectively. For the youngest age-group, neither comparison was significant (p = .229, 1), presumably reflecting low power. For the 6–8 age group, the counts in the RC column were significantly lower than in the declarative column (p < .000), but no significant difference was found between the declarative v. EQ columns (p = .664). For the 9–11 age group, RC and EQ counts were lower than declarative counts (p’s < .000). The same held for the 12–18 year old group (p’s < .000).

Qualitatively, evidence for island-violating movement appears unevenly distributed across the dependency-island cells. For some dependency-clause-type combinations, direct evidence seems roughly as frequent as would be expected if movement out of an island was as probable as movement out of a simple declarative complement clause. For example, the number of attested examples of topicalization and relativization from EQs is rather close to the expected counts for most age groups. For other combinations, examples are entirely absent: Wh-movement from either RCs or EQs is unattested in the sample. In the younger age groups, long-distance wh-movement and EQs are infrequent enough that the absence of their conjunction is not surprising. However, in older age groups, the absence is conspicuous. Similarly, the fact that relativization from RCs is never observed in the older groups, paired with the disparity from the expected counts suggests that the gaps in the distribution are real.

We inspected filler-gap dependencies into embedded questions, breaking them down by the type of embedded question (e.g. subject, object, polar question11) and the base position of the long-distance moved element. Table 5 provides the counts for each combination.

Collapsing across relativization and topicalization, long-distance movement of the highest subject out of an embedded question is most common (25/28 cases of relativization, 11/14 cases of topicalization). Long-distance subject movement is most common from copular clauses where the copula’s complement is questioned. (27) provides examples:

This specific configuration is most frequent, but there is variation. Subjects phrases are moved out of a variety of different embedded question types: polar (28a), object (28b), and adjunct (28c).

Moreover, the corpus contains evidence that non-subjects may also be moved long-distance out of EQs, though the amount of evidence for this possibility is markedly less frequent.

Given the relative diversity of embedded question types that allow long-distance movement, and that phrases with different grammatical roles can be extracted, there may be sufficient evidence that all embedded questions are non-islands for the relativization and topicalization of arguments. The absence of attested examples of wh-movement from EQs, however, means that the corpus lacks examples of at least some sentence types that are acceptable in the target grammar.

Consistent with trends noted in Engdahl (1997) and Lindahl (2017), all 63 examples of RC-island violations occurred either in a presentational or it-cleft configuration. Because both constructions use the expletive pronoun det, followed by the copula, some examples were potentially superficially ambiguous between existential and cleft construction. We categorized each token as either a presentational or it-cleft based on properties of the RC head and whether the sentence could be paraphrased using an alternate existential construction (Søfteland 2014). Like English existential constructions (Milsark 1977), presentational constructions allow weak determiners, but not strong determiners or proper names post-verbally (30). Like English clefts, Norwegian clefts generally disallow weak determiners in post-verbal position.

The copula in existential constructions can be replaced with the existential verb å finnes (Søfteland 2014), but the same replacement is not possible with clefts (31):

After categorization, we found that presentational and it-clefts were both well represented among our examples (2838 and 3832 examples, respectively; see Table B1 in Appendix B). There are two possible explanations for why RC-island violations were only observed with presentational and cleft constructions. The over-representation could indicate that the constructions have one or more properties that make their RCs particularly amenable to extraction compared to other RCs. Such a finding would potentially provide insight into the conditions governing acceptable extraction. Alternatively, their over-representation could simply reflect a difference in base-rate: perhaps presentational and cleft constructions are simply the most frequent types of RC in the input to children. Under this line of reasoning, the frequency of extraction from RCs would be consistent across constructions containing RCs, but non-presentational and non-existential RC constructions are so infrequent that the probability of observing an extraction is just too low.

We compared the overall frequency of presentational and it-cleft constructions to other RCs to see if simple base frequency could account for the absence of island violations from other RCs. We conservatively excluded RCs attached to subject NPs because subjects may be islands for independent reasons (e.g. Huang 1982). Presentational and it-clefts made up roughly 20% of the remaining RCs (6732 of 29522 eligible RCs, see Table B1 in Appendix B). As cleft RCs do not constitute an overwhelming majority of all RCs, it is unlikely that base frequency alone can explain their overrepresentation in island violations.

Movement from RCs was only observed from subject RCs (a tendency described in earlier work, Engdahl 1997; Lindahl 2017; Christensen 1982; Nordgård 1991). Closer inspection showed that subject RCs were roughly twice as likely as non-subject RCs in both presentational and it-cleft constructions, but non-subject RCs were still relatively common (see Table B2 in Appendix B). All else equal, it seems unlikely, then, that baseline frequency differences are to blame for the over-representation of island violations among subject RCs.12

Finally, we also report the phrasal types and roles of the constituents that were topicalized from presentational and clefted RCs. Table 6 shows that the corpus contains examples of a range of different phrase types with different roles being extracted from RCs. Direct object NPs are the phrase most commonly extracted from RCs. Nominal arguments are the most commonly topicalized elements (56/62 examples), however topicalization of (declarative and interrogative) CP complements are also attested, as well as at least one instance of a locative PP.

Examples from the corpus that illustrate the variety of topicalizations from RC are below:

Usage-based or item-based learning models treat filler-gap acquisition as learning a set of bespoke constructions or templates for well-formed dependencies (Tomasello 2003; Dąbrowska 2004, 2008, a.o.). Templates are generated as follows: learners first construct highly specific formulae/frames from single-items and subsequently collapse multiple formulae into more general abstract templates or constructions. The original formulae can be seen as the conjunction of lexical, morphological, syntactic, semantic, and pragmatic features of a single item at varying granularities (e.g., Dąbrowska & Lieven 2005). Generalization involves collapsing over equivalence classes of features across formulae, though the exact mechanism of generalization is not well understood. Generalization yields a complex well-formedness condition that abstracts over variable features, but maintains highly specific features along dimensions that were invariant in the input.

If examples of wh-movement from EQs are not attested in children’s input, then a constructivist learning strategy, which generalizes from direct evidence, would arguably have a difficult time learning that the filler-gap dependencies are possible. In principle, a constructivist strategy could be engineered to generalize that wh-movement out of EQs is acceptable having observed relativization and topicalization from the same domains. However, such generalization would require collapsing across dependency-types with different semantic features and discursive functions.13 Alternatively, if examples of wh-movement from EQs do infrequently occur in CDS, children could potentially learn their distribution. If there are stricter felicity conditions on wh-movement from EQs than other dependency types, which there seem to be, a usage-based learner might be in a position to learn the narrower restrictions from the smaller number of attested examples, as there would be significantly less variation in the fine-grained structural and semantic features shared across the observed dependencies.

Usage-based learning models would have more difficulty learning the appropriate distribution of acceptable filler-gap dependencies into RCs. We illustrate this point focusing on what superficial lexical and syntactic features a usage-based learning model could extract from the input as necessary conditions on filler-gap dependencies into RCs. Assuming that it was minimally sensitive to features such as (i) dependency type, (ii) embedded clause type, (iii) phrase type, and (iv) the lexical content of the embedding clauses (Dąbrowska 2004, 2008; Verhagen 2006; Löwenadler 2015), a usage-based learner might end up with a construction/template like (33) for RC-island violations. (Constructions are expected to be even more rarefied including additional semantic and pragmatic conditions on application, see e.g., Löwenadler 2015).

The construction/template in (33) closely tracks the fine-grained features of RC-island violations in the input. As a result, it overfits the input distribution, excluding dependencies that adults accept (e.g. dependencies into non-cleft/presentational and non-subject RCs).

Usage-based models maintain that knowledge of the distribution of acceptable filler-gap dependencies consists of bespoke, construction-specific conditions that incorporate all manner of features across grain-size and grammatical type. Other learning models restrict their attention to a subset of syntactic features in the input. We consider how two different approaches to syntactic learning would fare on the Norwegian data. Both models learn syntactic generalizations that, in isolation, predict a wider range of acceptable island violations than are observed in the corpus – and to some extent the target distribution of acceptable filler-gap dependencies.

We inspected how well the fine-grained characteristics of island violations in the input represented the target distribution of acceptable dependencies in the Norwegian adult grammar. The input distribution was compatible with very narrow generalizations that would not extend to the full range of forms that adults judge to be acceptable. We concluded that learners generalize beyond what is seen in their input.

Our conclusion that children’s input does not provide sufficient evidence for the full range of acceptable dependencies in the target grammar is only justified if the absence of particular forms in the written corpus is representative of their absence (or extreme infrequency) in CDS, children’s primary linguistic input. Island violations may be less frequent in written text than in speech (if, for example, they are associated with a more informal register). The occurrence of violations in our sample argues against categorical prohibitions against their use in writing and preliminary research suggests that island violations are also found in formal text like newspaper articles (Sant, Strætkvern, & Kush 2019). Nevertheless, their overall occurrence could still be less frequent, or particular types of examples could be under-represented in text relative to speech.

We discussed ways in which the statistics of written corpora deviate from CDS: studies in English show that main questions are less frequent in written texts than in CDS, while complex constructions such as RCs are more frequent (Cameron-Faulkner & Noble 2013; Noble, Cameron-Faulkner, & Lieven 2018; Montag 2019). The divergences between text and speech characteristics cut both ways.

The absence of island-violating wh-movements in the corpus, could reflect, in part, the reduced frequency of wh-movement in text. It is possible that examples of island-violating wh-movement are to be found in CDS, given the higher base rate of wh-movement. Thus, we cannot definitively conclude that children’s total input lacks evidence of island-violating wh-movement. If examples are present in CDS, the question remains whether they are frequent enough to drive reliable acquisition (e.g., distinguishable from errors or noise; Legate & Yang 2002).

On the other hand, the fact that RCs are more common in text than in CDS could be taken as evidence that our results overestimate the frequency of island violations that feature RCs. Our corpus findings might therefore overestimate the frequency of direct evidence of relativization from EQs. Our findings may also overestimate the frequency of RC-island violations given the higher base rate of RCs in text.

When it comes to RC-island violations, we feel relatively confident in extrapolating from the corpus to the conclusion that children’s input does not provide sufficient direct evidence of the full range of acceptable RC-island violations. In order to have direct evidence of the full range of acceptable RC-island violations, children would need to observe filler-gap dependencies into non-presentational/cleft RCs. But such RCs are even less likely in CDS than in text: Diessel & Tomasello (2000) note that presentational RCs make up close to 70% of the RCs in (German) CDS, compared to the 20% in our sample. Based on these numbers, we tentatively conclude that direct evidence of non-presentational RC-island violations is unlikely to be sufficiently frequent in CDS to drive direct acquisition of the knowledge that non-presentational RCs also allow extraction.

We considered how learning models with different inductive biases could recover the target generalizations based on our input. We reasoned that a simple usage-based learning model would overfit the input distribution because it would be too reliant on features of attested examples and would therefore preclude generalization to a number of the construction types that adults accept. Syntactic learners, such as the computational learner of Pearl and Sprouse (2013) or parameter-setting models (Yang 2002; Sakas & Fodor 2001, 2012; Gould 2015), learn generalizations that go beyond the empirical distribution in the input.

The fact that the syntactic learning accounts generalize broadly allows them to account for dependencies that are acceptable in the target grammar but unattested in the corpus. We consider this a welcome result: the models arguably learn the correct basic syntactic generalization that movement is allowed out of both RCs and EQs. However, the basic syntactic generalizations alone are insufficient to explain the finer-grained distribution of acceptable and unacceptable dependencies into EQs and RCs. In subsection 6.4 we sketch how the overall distribution of acceptable sentences might be modeled by incorporating semantic (e.g. Szabolcsi & Zwarts 1993; Abrusán 2014) and discursive/functional constraints (Erteschik-Shir 1973, 1982; Erteschik-Shir & Lappin 1979, Kuno 1987; Van Valin 1995, 1998; Ambridge & Goldberg 2008; Goldberg 2006, 2013) as supplemental filters on the output of the syntax (as suggested by Lindahl 2017 and Kush et al. 2018).17

Although we have expressed support for Generative learning accounts, we acknowledge the possibility that a usage-based model that attends to a different, select subset of features might not be prone to the same overfitting problem. It is also possible that filler-gap acquisition involves a stage of feature-reduction or selection wherein previously-posited highly specific features are abandoned in favor of more general features. At present, we know of no usage-based models that have a fleshed-out account of such a feature-selection process in filler-gap acquisition, but we encourage proposals in this direction. Insofar as we do not have longitudinal data on the abstractness of Norwegian children’s generalizations, we cannot rule out that they initially pursue fine-grained generalizations that overfit the data, only to adopt broader generalizations at a later age (Boyd & Goldberg 2011; Tomasello 2003).

In the introduction we asked what role the frequency of island violations might play in the acquisition of such structures. We found that children encounter island violations in written text. Do our results indicate that children see such violations frequently enough to learn island insensitivity from them alone? We cannot offer a definitive answer to this question, but provide discussion of the relevant challenges.

One reason frequency is presumed to be important is that it can help learners distinguish signal (e.g., target sentences) from noise (e.g. errors) in the input. Target constructions should occur more reliably than speech errors that should be ignored (e.g. Legate & Yang 2002). In edited text, noise is assumed to be either non-existent or very low, so low frequency constructions can be taken as potentially ‘sufficient’ evidence evaluated relative to the corpus. However, we do not know how corpus frequencies compare to frequencies in CDS, so we are not licensed to conclude that children observe sufficient direct evidence on the whole to distinguish island violations from noise.

Even if children are able to distinguish target sentences from noise, the frequency of a construction can still affect its learnability. As we discuss below, however, what constitutes ‘sufficient’ evidence for acquisition varies by model (and model parameters).

Under usage-based models frequency determines the degree to which generalizations are abstract – and possibly whether generalization occurs at all (Boyd & Goldberg 2011). If frequency drives abstractness of generalization (perhaps because feature-selection processes require larger amounts of data), then the relative infrequency of island violations would lead to predictions of highly specific constructions. The relative infrequency of island violations should also entail relatively late onset of generalization/acquisition, though a later age of acquisition is not a unique prediction of usage-based models.

For Pearl & Sprouse’s learner, islands are structures with container node sequences with unattested trigrams of zero (or nearly zero) probability. Since the distinction between island and non-island essentially amounts to ‘seen v. unseen’ trigrams, the exact frequency of direct evidence is less important (as long as it is frequent enough to be reliably observed). Of course, the more frequent island violations are in the input, the higher probability the relevant trigrams will be, which will thereby increase the predicted acceptability of similar island violations. Thus, frequency is important for tuning the acceptability predictions of the model. Finding the required frequencies to best model human judgment patterns is a project we leave for future work.

Multiple factors influence estimation of sufficiency under parameter-setting models including: the ambiguity of the data (Fodor 1998), the number of hypotheses under consideration, the model’s learning rate (e.g. Yang 2002), and the prior probabilities or weights assigned to each hypothesis. Island violations are relatively unambiguous evidence against the large class of conceivable grammars that disallow filler-gap dependencies into islands. We speculate that the positive evidence of island violations would be ‘enough’ to learn the right syntactic generalizations in Norwegian if learners must only set a single parameter, such as whether the complementizer domain of the language provides an extra escape hatch for successive cyclic movement. We have not, however, conducted the simulations to argue for this conclusion strongly.

Different models’ learning rates would affect how quickly children learned island insensitivity. Triggering models (e.g., Gibson & Wexler 1994) could successfully set the parameter with little data, since parameter settings can, in principle, be changed in response to exposure to individual sentences (or to a small number of sentences).18 Models where the learning rate can vary in relation to the strength of the evidence (e.g. Bayesian models like Pearl & Lidz 2009 and Perfors, Tenenbaum, & Regier 2011) could also learn the generalization quickly. Models with smaller, fixed learning rates (e.g. Yang 2002; 2004) require more data to set the parameter, resulting in a protracted acquisition period. This is particularly true if the prior on a parameter setting (like having an extra specifier) is especially low. A general preference for simpler grammatical analyses might bias children against the parameter options that allow for island-insensitivity, unless the input forced the analysis. The stronger the bias for a simpler model, the more data would be required to overcome that bias. Future work should explore interactions of frequency and parameter settings in implemented models.

If we assume that generative models treat wh-dependencies, relativization, and topicalization as underlying instances of A’-movement, the models above learn the arguably correct syntactic generalization that phrases of any type can be wh-moved, relativized, or topicalized out of any EQ or RC. We saw, however, that some dependencies that the model could generate were consistently judged unacceptable and others were extremely rare, if not unattested. How can we explain the distribution of observed dependencies from Norwegian EQs and RCs if we adopt a generative model that, on its own, allows for unrestricted movement from EQs and RCs?

One possibility is to assume that the responsibility for explaining the residual restrictions on Norwegian island violations in Norwegian should not fall to the syntax. Instead, unacceptable forms that the syntax generates should be ‘filtered out’ by independent semantic and pragmatic conditions. We sketch below how different distributional restrictions can be seen as arising from semantic and pragmatic conditions.