There is a long-standing generalization that certain structural domains restrict syntactic movement operations, a phenomenon termed island effects (Ross, 1967). Attempting to extract from islands results in degraded sentence acceptability, as in the examples in (1), all involving illicit wh-movement.1

The nature of these island effects has been a long-standing source of debate in the linguistic literature. The degraded acceptability of island-violating sentences like (1a-g) has been variably attributed to constraints in grammar (Bresnan, 1976; Chomsky, 1964, 1973, 1977, 1986; Huang, 1982; Rizzi, 1990; Ross, 1967; Sag, 1976) or processing (Culicover et al., 2022; Hofmeister et al., 2013; Hofmeister & Sag, 2010; Kluender, 1991; Kluender & Kutas, 1993).2 Grammar-based approaches to island effects claim that the sentences in (1) are ungrammatical because they violate certain grammatical constraints (e.g., the Subjacency Condition, the Phase Impenetrability Condition, etc.). Processing-based approaches to island effects claim that the sentences in (1) are grammatical, but are unacceptable due to the high processing burdens they incur (analogous to the difficulty in processing center-embedding sentences).

In addition to the debate over whether island effects are best explained as the result of grammatical or processing factors, there is a lack of consensus regarding whether certain islands form natural classes. For example, some have grouped subject and adjunct islands together as a natural class in opposition to the other island types, and have attributed the two distinct classes of islands to two different constraints in the grammar (Chomsky, 1986; Huang, 1982; Nunes & Uriagereka, 2000). Others, however, reject this grouping, either proposing that subject and adjunct island effects involve different grammatical constraints (Haegeman et al., 2014; Hiramatsu, 2001; Stepanov, 2007), or arguing for a unifying (syntactic, information structural, or processing-based) account of a larger set of island effects, including but not limited to subject and adjunct island effects (Abeillé et al., 2020; Bošković, 2016; Erteschik-Shir, 1973; Goldberg, 2013). In sum, the island literature lacks a consensus on the source and nature of island effects.

The effect of repeated exposure on the perceived acceptability of island-violating sentences has been brought to bear on the debate over grammatical vs. processing accounts of island effects. A linking assumption that has been (implicitly or explicitly) adopted by some. is that degraded acceptability due to grammatical violations should not be affected by repeated exposure. In contrast, if the source of degraded acceptability is processing difficulty, exposing participants to similar sentences of the same type should increase familiarity with this sentence type and ease the associated processing burden. In turn, acceptability should increase with repeated exposure (i.e., show the satiation effect).3 If we accept this linking hypothesis, whether or not the acceptability of island-violating sentences increases with exposure can been used to diagnose whether certain island effects are grammatical or the result of processing constraints.

Acceptability increase after exposure, or the satiation effect (Stromswold, 1986), was first demonstrated experimentally for island-violating sentences in Snyder (2000).4 The observation of satiation effects in island-violating sentences has subsequently been interpreted as evidence for the extra-grammatical nature of islands, including the Complex NP Constraint (Hofmeister & Sag, 2010), the superiority effect (Hofmeister et al., 2011), and subject islands (Chaves, 2022; Chaves & Dery, 2014, 2019).

Other studies assume a different linking hypothesis for the satiation effect, whereby certain grammatical constraints may also be “unlearned” or weakened throughout repeated exposure to sentences violating those constraints, and differences in satiation profiles reflect differences in the types of grammatical constraints involved (Braze, 2002; Goodall, 2011; Hiramatsu, 2001; Snyder, 2000). Assuming this hypothesis, satiation results cannot be used to inform the grammar vs. processing debate regarding the nature of islands. Instead, satiation can be used to probe for natural classes formed by different islands. If two island types show different patterns of satiation (e.g., one satiates, while the other does not), they are assumed to have different underlying sources of unacceptability and should not be grouped together as a natural class. This linking hypothesis underlies the argument against Huang’s (1982) proposal that subject and adjunct islands form a natural class (Hiramatsu, 2001; Stepanov, 2007).

In sum, while assuming slightly different linking hypotheses, the satiation effect has been used as evidence in multiple debates surrounding the nature of island effects.5 In the current study, our goal is not to further complicate the picture by taking sides in any of these debates. We also remain agnostic about which linking hypothesis for satiation is appropriate. Instead, we aim to clarify the empirical landscape on satiation so that satiation can be better leveraged as a source of evidence.

As mentioned earlier, the satiation literature abounds with mixed findings. Statistically significant satiation of island effects has been observed with some experimental procedures and items, but these effects have been inconsistent (see Snyder, 2022b, for a comprehensive overview). For example, the adjunct island is found to satiate in Chaves & Putnam (2020), but not in Crawford (2012); Francom (2009); Hiramatsu (2001); Snyder (2000); Sprouse (2009, inter alia). For a comprehensive list of past satiation studies and whether they observed satiation in each island type, the reader is referred to Tables 4 and 5 from Snyder (2022a). The current study reports a quantitative meta-analysis of past satiation studies, with the aim of summarizing and aggregating past findings in the service of assessing which island types reliably satiate.

Below, we report the meta-analysis results for satiation studies examining the seven different island effects listed above: the Left Branch Condition (LBC), adjunct islands, the Complex NP Constraint (CNPC), subject islands, the that-trace construction, the want-for construction, and whether-islands.

In this study, we conducted a meta-analysis of past experimental studies on the satiation of island-violating sentences in English. The results of this study provide answers to the following three questions. First, which types of islands reliably satiate? Second, which island types display heterogeneity in satiation, and which factors contribute to this heterogeneity? Third, is there evidence for a publication bias in the satiation literature?

To answer the first question, random effects meta-analytic models revealed significant acceptability increases for repetition in sentences violating constraints on adjunct islands, complex noun phrase islands, subject islands, the that-trace effect, the want-for construction, and whether-islands. In contrast, the models did not reveal evidence of acceptability increases for sentences violating the Left Branch Condition (see Figure 3).

To answer the second question, significant cross-study heterogeneity was detected among the CNPC studies, the subject island studies, and the whether-island studies. Following speculation by Snyder (2022a), we tested the contribution of three moderators (presence of context sentences in the experiment, total number of repetitions, and the scale type) to the heterogeneity. For CNPC studies, there was no evidence for any of the moderators modulating satiation; for subject island studies and whether-island studies, the presence of a context sentence increased the rate of satiation, while the other two moderators did not reach significance. For all three groups of studies, residual heterogeneity persisted even when the three moderators were included in the mixed-effects model, suggesting that additional moderators contribute to cross-study heterogeneity.

One such moderator may be the number of fillers. The increase in acceptability of the critical items due to exposure might gradually decay when participants see many unrelated filler items, similar to the observed effect of number of intervening items between prime and target on the strength of priming effects. Thus, the rate of satiation might be smaller when more filler items are included in the experiment. However, since not all selected studies reported filler information in the original papers or in the data files shared by the authors, we could not include the number of fillers as a moderator. Another factor that could affect satiation is the inclusion of multiple island types in the same experiment. It has been demonstrated that exposure to one island-violating sentence type leads to a change in the acceptability of another island-violating sentence type (Lu et al., 2022).17 Therefore, it is possible that when multiple island sentence types are tested together in a single experiment, they might influence each other’s rate of satiation. Cross-study heterogeneity could then arise as a result of different studies testing different sets of island sentence types.

To answer the third question, there is possible publication bias favoring the studies reporting significant satiation for sentences violating the CNPC and the whether-island. However, the evidence is not strong and is based on marginally significant results from Egger’s regression test of funnel plot asymmetry. We did not find a funnel plot asymmetry in any other group of studies, suggesting that there is no evidence for a publication bias in those groups of studies. However, we should note that these groups tended to also be the ones with fewer studies. It is possible that for island types with fewer studies, we might not have had the statistical power to detect publication bias even if the bias exists, creating a false picture that there is no publication bias in those groups of studies. An exception is subject islands, which have the largest number of studies devoted to them and yet didn’t show evidence of publication bias.

The results of this meta-analysis are valuable in at least four different ways: in the debate over grammatical vs. processing accounts of islands, in the debate between different linking hypotheses for satiation, in the debate over whether subject and adjunct islands form a natural class in the taxonomy of islands, and in revealing the varying rates of satiation of different islands as a future direction for research.

First, the results may inform the debate over grammatical and processing accounts of islands. As discussed in Section 1, it is often implicitly assumed that when a degraded sentence type satiates, the source of the degradation should come from extra-grammatical factors like high working memory burden (Hofmeister et al., 2013; Hofmeister & Sag, 2010) or low frequency (Chaves & Dery, 2014, 2019). Assuming that this linking hypothesis is accurate (and we shall return to the possibility that it is not), the results reported here can be used to inform theories of island effects: the reliable satiation effects for adjunct islands, CNPC, subject islands, the that-trace construction, the want-for construction, and whether-islands suggests they should be considered grammatical but degraded due to processing factors. In contrast, the LBC does not satiate, and therefore should be considered ungrammatical. These grammaticality statuses pose a challenge to syntactic theories that predict the ungrammaticality of the satiating island types. These include the Government and Binding (GB) theory, which attributes island effects to grammatical constraints like the Subjacency Condition and the Empty Category Principle (Chomsky, 1986; Huang, 1982), and syntactic theories under the minimalist program that attribute island effects to the cyclic nature of spell-out and linearization (Fox & Pesetsky, 2005; Nunes & Uriagereka, 2000). In contrast, our results are, in general, compatible with syntactic theories without non-local syntactic constraints. For example, in certain versions of the Head-driven Phrase Structure Grammar (HPSG) (Boas & Sag, 2012; Michaelis, 2013), only island effects that can be framed in terms of local constraints are predicted to exist without arbitrary stipulations of filtering constraints. Most island-violating sentences, including the ones that reliably satiate according to this meta-analysis, are predicted to be grammatical in this framework (see Chaves & Putnam (2020) for a detailed discussion of the predicted island effects under this framework). In contrast, the LBC, where we found no evidence for satiation, can be captured in HPSG by an independently motivated local constraint requiring that only elements in the arg(ument)-st(ructure) feature list of a head can appear in the gap feature list of the same head. Possessors and modifiers are not part of the arg-st list of an N head, and thus cannot appear in the gap list of N (i.e., cannot be extracted from an NP, Chaves & Putnam, 2020; Runner et al., 2006; Sag, 2012). Nevertheless, one should note that the that-trace construction and the want-for construction are largely left out of the debate between processing-level and grammar-level accounts of island effects, and processing-based accounts of their pre-satiation degradedness are yet to be worked out.18 This may be particularly challenging for the want-for construction, where the declarative form (example (2)) without the long-distance dependency is also considered by many speakers to be degraded, suggesting that the unacceptability of the want-for construction cannot be attributed to a hard-to-process dependency.

As for the that-trace construction, past accounts mostly attributed its degradedness to syntactic constraints (e.g., the anti-locality constraint) or phonological constraints (e.g., a phonological requirement that a complementizer cannot be linearly adjacent to a gap).19 It remains unclear how the that-trace effect can be captured by a processing-level explanation. If one believes that processing-level accounts for the want-for and the that-trace constructions are implausible, the satiation of these two sentence types may also be taken as evidence against the linking hypothesis that satiation reflects grammaticality.

Second, the discussion up to this point assumes the linking hypothesis that grammaticality decides the satiability of degraded sentences. However, this hypothesis is not unchallenged. Other factors have been claimed to affect the satiability of sentences, including whether the grammatical constraint violated is part of Universal Grammar (UG) (Braze, 2002; Hiramatsu, 2001; Snyder, 2000), the surface similarity with a grammatical alternative (Sprouse, 2007), and sentence interpretability (Francom, 2009). Instead of assuming a particular linking hypothesis under which to test theories of island effects, one can also use the meta-analysis results to inform the linking hypothesis itself. For example, the hypothesis that grammatical principles in UG determine satiability is rejected by our results. Under this hypothesis, the LBC, which is the only non-satiating island type among the ones investigated, should be a principle in the UG. However, the LBC is in fact subject to variation cross-linguistically. Left branch extraction (extraction of modifiers or possessors out of an NP) is permitted in many Slavic languages (Bošković, 2005). For example, Serbo-Croatian allows the equivalent of “Whose_i does Petko like t_i car?” as shown in (3). This shows that the LBC cannot be a grammatical principle encoded in UG, thus rejecting the hypothesis that satiation diagnoses UG principles.

Consider also the linking hypothesis that satiation is underlyingly linguistic adaptation, whereby participants update their beliefs about the frequency distribution of degraded sentence types throughout acceptability judgment experiments (Lu et al., 2021, 2022). Under this proposal, participants should be able to update their expectations for any given sentence type whose underlying linguistic representations can be recovered. This predicts that any sentence type, with the exception of word-salad sentences from which no abstract linguistic representation can be recovered, should demonstrate satiation. However, our results show that the LBC sentences resist satiation altogether, which is not predicted by the vanilla version of the adaptation-based linking hypothesis. It is likely that there are other gate-keeping factors at play, restricting the satiation of an otherwise perfectly representable structure.

One possible candidate for such gate-keeping factors is the availability of grammatical alternatives: when the intended meaning of a sentence can be expressed by a more acceptable alternative sentence, satiation of the sentence is suppressed, perhaps because the participants do not adapt to sentences produced by hypothetically non-cooperative speakers. Among all seven sentence types investigated, the LBC sentences are the only type with a clear meaning-equivalent acceptable alternative. For example, the unacceptable LBC-violating sentence (4a) has the meaning-equivalent acceptable alternative (4b).

Whether the availability of acceptable alternatives is indeed the reason why the LBC sentences resist satiation is beyond the scope of the current study, but further studies should test this possibility.

Third, our results also have implications for the debate regarding whether subject and adjunct islands are reducible to the same underlying constraint. In syntactic theories, among the various types of island effects, adjunct islands and subject islands are traditionally considered to form a natural class. For example, Huang (1982) attributes both adjunct and subject island effects to a single syntactic principle: the Condition on Extraction Domains (CED), which states that constituents that are not properly governed restrict extraction from within. Subject DPs and adjuncts are not properly governed, and therefore are both CED islands.20 Chomsky (1986), which aims to provide an analysis for all island effects using the concept of barriers, also grouped subject and adjunct islands together. Movements out of adjuncts and subjects need to cross two barriers, whereas movements out of other islands (e.g., whether-islands, complex NPs) cross only one barrier. Studies including Hiramatsu (2001) and Stepanov (2007) question Huang’s (1982) and Chomsky’s (1986) accounts on the grounds that subject island sentences satiate while adjunct island sentences do not. Assuming a linking hypothesis under which structures with a common source of unacceptability should show a similar satiation profile (Braze, 2002; Goodall, 2011; Snyder, 2000), we would expect subject and adjunct islands to either both satiate or both resist satiation, contrary to Hiramatsu’s (2001) observation. However, as is evident in the current meta-analysis, there is reliable evidence that both subject and adjunct island sentences satiate. Although this result cannot distinguish between the CED or the barriers account, at least it provides sufficient grounds for rejecting Hiramatsu’s (2001) counterargument.

Fourth, our results point to a future direction of research. As mentioned in Section 2.2.1, there appear to be varying rates of satiation among the satiating island types. The differences in the rates of satiation might signal differences in the linguistic properties underlying these sentence types, and could potentially become a useful diagnostic tool for experimental syntacticians. Differences in rates of satiation could reflect different sources of unacceptability (see Goodall, 2011, for a similar proposal). Further research is needed to determine which factors govern the rate of satiation.

Finally, we would like to acknowledge several limitations of the current study.

First, the results of the Egger’s test provide evidence for publication bias only among the CNPC and whether-island studies. This does not suggest that there is no publication bias for the other island types. It is possible that the Egger’s tests we conducted simply lack the statistical power to detect publication bias for the island types with fewer studies (e.g., the that-trace construction and the want-for construction both have fewer than 10 studies).

Second, the current meta-analysis reports results on the satiation of island-violating sentences (i.e., the increase in acceptability ratings for sentences that contain island violations), as opposed to the amelioration of island constraints (i.e., the amelioration of the acceptability degradation induced by violating an island constraint). This is because the current meta-analysis is constrained by the design of early satiation experiments that did not include the relevant grammatical control conditions that would enable us to probe for the amelioration of island constraints. However, we should note that the satiation of an island-violating sentence type does not necessarily reflect any change in the relevant island constraint, which is only one of various possible sources of degradation for the tested sentence type.

Below, we would like to point to a possible future direction for satiation researchers interested in teasing these two concepts apart. Recent experimental syntax studies on islands have largely adopted the “factorial design of measuring island effects” (Fukuda et al., 2022; Keshev & Meltzer-Asscher, 2019; Kim & Goodall, 2016; Ko et al., 2019; Kush et al., 2018, 2019; Lu et al., 2020; Sprouse et al., 2016, 2012; Stepanov et al., 2018, inter alia). Consider a CNPC sentence such as (5).

The degraded status of (5) could be due to any of the following factors: (a) a long-distance dependency crossing a clausal boundary, which may be cognitively taxing to process; (b) a Complex NP embedded clause structure, which may be structurally infrequent; (c) a special penalty for having a long-distance dependency crossing a Complex NP embedded clause structure (i.e., the violation of an island constraint). Under the factorial design of measuring island effects, these potential sources of degradation are teased apart using a 2 × 2 factorial design manipulating dependency distance (movement from embedded clause vs. movement from matrix) and embedded clause structure (island structure vs. non-island structure). An example set of stimuli is shown in (6), adapted from Sprouse et al. (2012).

If the island constraint violation contributes unacceptability, we should detect it as an interaction effect between dependency distance and clause structure, where the dependency distance penalty is larger in the island conditions than in the non-island conditions. This factorial design of measuring island effects can be valuable for future research on the satiation of island effects. If the island constraint itself is affected by repeated exposure, we should expect a three-way interaction of dependency distance, clause structure, and presentation order, in the direction where the super-additive interaction of dependency distance and clause structure decreases as presentation order increases. This would serve as a more rigorous test for whether the satiation of island-violating sentences truly reflects the amelioration of the underlying island constraints.