Comparing syntactic and discourse accounts of islands and parasitic gaps: experimental evidence from acceptability judgments

1.1 Filler-gap dependencies

A major debate in both theoretical and experimental linguistics revolves around the question to what extent linguistic phenomena can be explained in terms of syntactic principles or non-syntactic factors, such as effects of discourse or general properties of human cognition. For over three decades, filler-gap dependencies have been at the center of this debate (Abeillé et al. 2020; Chaves 2013; Chaves and Putnam 2020; Culicover and Winkler 2022; Culicover et al. 2022; Cuneo and Goldberg 2023; Deane 1991; Hawkins 1999; Hofmeister and Sag 2010; Kluender 1991, 1998, 2004; Kush et al. 2019; Liu et al. 2022; Phillips 2006, 2013; Pritchett 1992; Sprouse et al. 2012, 2016). Filler-gap dependencies are unbounded dependencies in which a displaced constituent – the filler – appears in a position different from its canonical position. The latter is called the gap site, which is an empty syntactic position licensed by the filler. The sentence in (1) illustrates an acceptable filler-gap dependency. Although the paper is interpreted as the complement of read, it appears in a position different from the postverbal position it would normally occur in (this gap site is indicated by the underscore). An important observation of generative linguistics is that not all constructions allow dependency formation of this type. Certain syntactic environments, including subjects, adjuncts, and relative clauses, are known to block the formation of filler-gap dependencies. An example of such a construction is shown in (2), whose acceptability is strongly reduced because a dependency needs to be formed between a filler and a gap in an adjunct clause.^[1]

Syntactic theories have shown that constraints on filler-gap dependencies can be formulated in terms of syntactic principles (e.g. Chomsky 1973, 1981, 1986; Huang 1982; Kayne 1983; Ross 1967). According to this view, (2) is unacceptable because it violates a grammatical constraint that blocks extraction from adjuncts (or non-complements, see Chomsky 1986; Huang 1982). Since Ross (1967), the domains from which extraction cannot occur are referred to as islands. Unacceptable filler-gap dependencies like the one in (2) are therefore called island violations. Although the term islands arguably suggests a syntactic source for the unacceptability, we will use it in a theory-neutral way, only to describe structures that resist filler-gap dependency formation.

Various alternative theories maintain that the unacceptability of (2) should be explained in terms of non-syntactic properties, like conflicting semantic or discourse-packaging conditions (Abeillé et al. 2020; Chaves and Putnam 2020; Cuneo and Goldberg 2023; Erteschik-Shir 1973; Goldberg 2006, 2013; Kuno 1987; Truswell 2007, 2011) or non-linguistic demands on cognitive processing (Deane 1991; Hofmeister and Sag 2010; Kluender 1991, 1998, 2004; Pritchett 1992). The specific claims and predictions of these non-syntactic approaches differ, but they share the underlying idea that the unacceptability of islands is due to factors other than the violation of a syntactic constraint.

A finding extensively discussed in the syntactic literature is that certain island violations can be ameliorated by a parasitic gap configuration (Culicover 2001; Engdahl 1983; Taraldsen 1981). A parasitic gap construction is a type of sentence in which one filler phrase is associated with two different gaps, one of which is located inside an island, and one of which is not. This is exemplified by sentence (3), in which the paper is the object of both read and filing.

The gap located inside the adjunct clause before filing _ is said to be parasitic on the gap in the complement position of read, as the acceptability of the former depends on the presence of the latter (cf. the unacceptability of (2)). The latter, by contrast, is a regular, licit gap, because it appears in a position that normally permits extraction (cf. the acceptability of (1)). Given that (2) is unacceptable, the amelioration effect in (3) is somewhat surprising: in both constructions, a dependency is established with a gap inside an island.

1.2 Syntactic constraints on filler-gap dependencies

The generative literature explains observations such as (1)–(3) in terms of the structural properties of these sentences. The unacceptability of extraction from subject- and adjunct islands, as in (2), is claimed to be due to a constraint that blocks extraction from non-complements, known as the Condition on Extraction Domain (Huang 1982). Likewise, in the Barriers framework (Chomsky 1986), non-complements are identified as Blocking Categories, because they do not receive a theta role from a lexical category under government and can therefore turn into barriers to extraction. A notion of structural government also underlies Kayne’s (1983) Connectedness Condition, which bans movement out of left branches. As such, this condition can account for the illegitimacy of extraction out of subjects, but Longobardi (1985) later extended Kayne’s machinery to apply to adjuncts as well. That different operations are at play in the two types of non-complements is also at the heart of Stepanov’s (2007) ‘eclectic’ account. Stepanov advocates distinguishing between subject- and adjunct clauses when describing island effects, because these clauses behave differently cross-linguistically: adjuncts are more robust islands than subjects. He therefore proposes that different mechanisms are at work in the derivation of subject- and adjunct islands. What these different formal approaches have in common is that extraction from adjuncts and subjects is ungrammatical and, consequently, unacceptable.

1.3 The Uninvited Guest Hypothesis

Culicover and Winkler (2022) maintain that the difference between islands and parasitic gap constructions is a matter of processing complexity, not a matter of syntactic well-formedness. Focusing on extraction from subject islands, as in (9), they propose the Uninvited Guest Hypothesis. For consistency with examples (1)–(3), we also added an example of a regular gap construction (8) and a parasitic gap construction (10) here.

According to Culicover and Winkler (2022), the island violation in (9) is unacceptable not because it violates a grammatical constraint that blocks extraction from subjects, but because it contains multiple sources of processing difficulty, which together lead to a strong reduction in acceptability. Specifically, (9) contains a referring argument in the predicate, her colleague, which they call the Guest. When (9) is presented in isolation, the Guest is Uninvited; it refers to a novel discourse entity and therefore adds complexity to the interpretation of the sentence (Gibson 2000; Warren and Gibson 2002). Extraction from a subject island also induces a processing cost, because it leads to a discourse clash between the foregrounded status of the extracted element and the backgrounded status of the domain of extraction (Abeillé et al. 2020; Erteschik-Shir 1973; Goldberg 2006, 2013; see Section 5.2.1.1 for further discussion). The combined cost associated with introducing a novel discourse referent and (wh-)extraction from backgrounded material then leads to full unacceptability.

Culicover and Winkler (2022) provide two types of evidence to support their hypothesis. First, they show that the acceptability of the sentence is improved when there is no overt expression in the predicate. This is the case in the parasitic gap construction in (10), which contains no Guest, because the object in the predicate is replaced by a gap. Besides extraction from a subject island, (10) contains no other source of processing complexity, and it is consequently considered more acceptable than (9).

Second, they show with a corpus analysis using Google searches that, when subject island violations are produced and the predicate contains a transitive verb, the argument in that predicate does not denote a novel discourse entity. These accessible or discourse-given arguments are referred to as Invited Guests. Extraction from subject islands occurs with two types of transitive predicates. One type attributes a property to the subject, as in …those who spending time with always feels like a positive experience. The other type contains a noun phrase (NP) that refers to an antecedent in the preceding discourse. Consider two examples from Culicover and Winkler’s (2022) Google search. Example (11) (their (13a)) contains a subject island violation (in bold) in which the predicate (underlined) contains the pronoun him, whose antecedent is immediately available in the discourse. It does not introduce a new discourse entity and is therefore an Invited Guest.

Example (12) (their (15c)) shows a subject island violation in which the predicate contains a NP the postulated meaning which bears on the topic of the discourse. The NP is again an Invited Guest; it does not introduce a discourse entity and has the discourse status ‘given’. In cases such as these, no additional processing cost is incurred by the predicate, because the referent of the argument is accessible in the discourse context.

Thus, Culicover and Winkler (2022) argue, island- and parasitic gap phenomena should not be explained in grammatical terms. Extraction from islands is grammatical and therefore acceptable in principle, but it is associated with increased demands on processing. In cases where this leads to a strong reduction in acceptability, the additional cost associated with processing an Uninvited Guest can push the sentence below the threshold of acceptability, giving rise to the appearance of ungrammaticality. Furthermore, they claim that there is no parasitic gap mechanism that can make island extraction grammatical; islands are simply more difficult to process than parasitic gap constructions due to the presence of an Uninvited Guest. Replacing the Guest with a true gap keeps the sentence above the threshold of acceptability (see also Culicover 2013). While this may seem like some form of grammatical amelioration, according to Culicover and Winkler (2022) the amelioration associated with the parasitic gap configuration is only a consequence of reduced processing complexity.

1.4 Current study

We asked participants to rate regular gap constructions, parasitic gap constructions, and island gap constructions, which were preceded by a short context that manipulated the discourse status of a referring argument in the sentence, the Guest. An example of a stimulus item is given in Table 1. In the target sentences, the NP the church is the Guest. Depending on whether the Guest has been introduced in the context or not, it is Invited or Uninvited. When the context introduces a church, the church in the subsequent target sentences is discourse-given and therefore Invited. In the alternative context, church is replaced by football stadium, making the church in the target sentences discourse-novel and therefore Uninvited.^[4]

In terms of form, the target sentences only differed in the presence and position of the Guest. In the regular gap construction, the Guest is embedded in a clausal adjunct (without visiting the church in Table 1). The regular gap in the complement position of the main verb is bound by the filler the city, which is the topicalized object of the sentence. In the parasitic gap construction, the Guest is entirely absent. The city is linked to two gaps, one in the adjunct island (the parasitic gap) and one in the complement position of the main verb (the regular gap). In the island gap construction, the Guest is the complement of the main verb. The only gap there is in the adjunct, making this sentence an adjunct island violation. Using short discourse contexts in two experiments, we thus evaluate whether the discourse status and accessibility of the Guest in the target sentence affects the acceptability of different kinds of filler-gap constructions.

The stimuli in the two experiments are highly similar, the only difference being that the adjunct in the target sentence is an untensed clause in Experiment 1 (as in Table 1) but a tensed clause with a pronominal subject in Experiment 2. From a discourse-processing perspective, the two experiments are minimally different, because the cognitive load introduced by the manipulation of tense is not only minimal (Gibson 2000; Kluender 1998; Warren and Gibson 2002) but also the same across the three types of constructions. Thus, to the extent that finite tense has an effect on acceptability judgments, it should concern a main effect that does not interact with the construction manipulation; i.e. the effect should be similar across constructions, according to the Uninvited Guest Hypothesis. According to Syntactic Theory, however, tense has a strong effect on the acceptability of certain filler-gap dependencies, in particular islands and parasitic gap constructions (see Section 1.2.2). Our two experiments combined therefore allow for a comprehensive comparison of the Uninvited Guest Hypothesis and Syntactic Theory. On the one hand, these theories make different predictions about the acceptability of different kinds of filler-gap constructions, especially when embedded in context. And on the other hand, they make different predictions about the effect of finite tense.

2.1 Methods

2.1.3 Procedure

The experiment was conducted online using the Qualtrics platform (Version 2024). Participants were instructed to locate themselves in a quiet environment for the duration of the experiment, and to carefully read the brief stories (discourse preambles) presented to them. They were stimulated to imagine that these stories were being told by a first language speaker of Dutch. After 7,500 ms, the corresponding target sentence and rating scale would appear. Presentation of the target sentence was delayed so as to stimulate participants to carefully read the story, and not to proceed with the experiment without paying much attention to this discourse context (Schoenmakers 2023). Participants were asked to judge the naturalness of the target sentence, which was presented on a new line in bold font (see Figure 2 for a sample trial), by relying on their first intuition; they were told that there were no right or wrong answers. Judgments were made on a semi-continuous sliding scale ranging from 0 (very unnatural) to 100 (very natural). The initial score value was set to 50, but the slider bar had to be moved in order to proceed with the experiment. The numerical value was hidden to the participants, so that participants had to rely on spatial rather than numerical reasoning when providing their response. We note that 100-point slider scales are no less sensitive than the Likert scales more commonly used in experiments investigating island effects. Systematic investigations of different scale types conclude that continuous and unnumbered graphic slider scales with few labels, such as the one used in our experiment, offer good sensitivity and yield reliable data (Cook et al. 2001; Marty et al. 2020). Finally, we checked whether participants carried out the experiment seriously by analyzing their responses following a number of exclusion criteria (Section 2.1.4).

Figure 2:

Sample trial. The context sentences are presented at the top. De target sentence (in bold) is presented in the middle, just above the judgment labels erg natuurlijk ‘very natural’ and erg onnatuurlijk ‘very unnatural’, which are presented on opposite ends of the sliding scale. In this trial, which is taken from the item presented in Table 1, the target sentence is an island gap construction with an Uninvited Guest.

2.2 Results

Figure 3 shows the acceptability judgment scores in each condition. The three-level factor construction was coded with two contrasts. The first contrast compared the (traditionally considered) ungrammatical condition (island gap) to the average of the grammatical conditions (regular and parasitic gap). As expected, this comparison showed that the constructions with an island gap (M = 14.2, SE = 0.83) were judged as less acceptable than the constructions with a parasitic gap (M = 58.3, SE = 1.08) or a regular gap (M = 59.5, SE = 1.11), β = −1.33, SE = 0.03, p < 0.001. The second contrast compared parasitic gaps to regular gaps. This comparison revealed no statistically significant difference between the two constructions, β = −0.04, SE = 0.03, p = 0.257, which received highly similar judgment scores (BF₀₁ = 9.35). Furthermore, the acceptability of the target sentences was modulated by the discourse context: sentences with an Invited Guest (M = 45.4, SE = 1.33) were judged as slightly more acceptable than sentences with an Uninvited Guest (M = 42.7, SE = 1.31), β = −0.08, SE = 0.03, p = 0.003. A Bayes factor analysis showed that there was moderate evidence for this difference, BF₁₀ = 3.60. Moreover, the effect of context was similar across constructions (see Table 2), showing that a supporting context did not increase the acceptability of sentences with an island gap more than it increased the acceptability of sentences with a regular or a parasitic gap. Importantly, the absence of a significant interaction (in the direction predicted by the Uninvited Guest Hypothesis) is not a matter of statistical power, as the effect of context is numerically larger for both regular and parasitic gaps than it is for island gaps. Indeed, the Bayes Factor analysis showed strong evidence for the null hypothesis, BF₀₁ = 28.87. The fixed and random effects of the full model are shown in Table 2.

Figure 3:

Acceptability judgment scores of Experiment 1, grouped by construction and context. The black dot encodes the mean, the error bars indicate 95 % confidence intervals around the mean. The average judgment scores of the grammatical and ungrammatical filler items are added as rough points of reference. These should not be interpreted as reflecting thresholds of acceptability or unacceptability.

2.3 Discussion

The results of Experiment 1 reveal that topicalization out of adjunct islands is rated as unacceptable. By contrast, regular and parasitic gap constructions received (equally) high acceptability scores, even though they were minimally different from the island gap constructions. Moreover, the effect of context was not different across the three types of constructions. These results are in line with Syntactic Theory in the sense that what is deemed grammatically licensed is judged as acceptable, and what violates a grammatical principle is judged as unacceptable. In contrast, this categorical pattern does not corroborate the gradient pattern predicted by the Uninvited Guest Hypothesis (see Figure 1). According to this hypothesis, parasitic gap constructions should be degraded (compared to regular gap constructions), because they contain extraction from an island. This is not what we observe. Moreover, island gap constructions should be judged as more acceptable when the Guest in the predicate is Invited compared to when it is Uninvited. This prediction is not borne out either; while island gap constructions were judged as slightly more acceptable when they contained an Invited Guest, a similar effect of context was observed in the other two constructions. This suggests that the introduction of a novel discourse referent is not what makes island gap constructions unacceptable or difficult to process.

Relatedly, according to the Uninvited Guest Hypothesis, the increase in acceptability that is commonly seen when the Uninvited Guest is replaced by a true gap, as in a parasitic gap configuration, is simply a matter of reduced processing complexity. A prediction following from this idea is that the island violation cannot be ‘rescued’ by a parasitic gap configuration when the processing demands incurred by island extraction for a given sentence are already very high. In that case, the unacceptability is too strong to show sensitivity to the presence or absence of the Uninvited Guest, and the parasitic gap construction should stay below the threshold of acceptability (see also Culicover 2013). Our results do not corroborate this prediction. For each individual item, it is the case that its acceptability substantially improves with a parasitic gap configuration (see Figure A1 in the Appendix): no matter how unacceptable (and/or difficult to process) the island gap construction, replacing the Guest by a true gap strongly improves the acceptability. In all, the results of Experiment 1 do not suggest that the difference between islands and parasitic gap constructions is a matter of processing complexity involving a new discourse entity.

3.1 Methods

3.2 Results

Figure 5 shows the acceptability judgment scores in each condition. The factor construction was again coded with two contrasts, the first of which compared the ungrammatical condition (island gap) to the average of the grammatical conditions (regular and parasitic gap). Like in Experiment 1, this comparison showed that the constructions with an island gap (M = 15.8, SE = 0.84) were judged as less acceptable than the constructions with a parasitic gap (M = 38.9, SE = 1.17) or a regular gap (M = 60.7, SE = 1.27), β = −1.00, SE = 0.03, p < 0.001. In contrast to Experiment 1, however, the second contrast showed a significant effect as well: parasitic gap constructions were judged as less acceptable than regular gap constructions, β = −0.64, SE = 0.04, p < 0.001. The effect of discourse context in Experiment 2 was numerically similar to that of Experiment 1, with sentences with an Invited Guest (M = 39.4, SE = 1.31) receiving slightly higher acceptability scores than sentences with an Uninvited Guest (M = 37.5, SE = 1.30). However, this effect was not significant, β = −0.06, SE = 0.03, p = 0.054, and the Bayes Factor analysis showed moderate evidence for the null hypothesis, BF₀₁ = 3.23. The interactions between construction and context were also not significant, and this null effect was very strongly supported by the Bayes Factor analysis, BF₀₁ = 74.48. The fixed and random effects of the full model are shown in Table 3.

Figure 5:

Acceptability judgment scores of Experiment 2, grouped by construction and context. The black dot encodes the mean, the error bars indicate 95 % confidence intervals around the mean. The average judgment scores of the grammatical and ungrammatical filler items are added as rough points of reference. These should not be interpreted as reflecting thresholds of acceptability or unacceptability.

3.3 Discussion

Like in Experiment 1, island gap constructions were rated as unacceptable and regular gap constructions were rated as acceptable. Compared to Experiment 1, however, the acceptability of parasitic gap constructions was substantially reduced, such that the average score is exactly in between the scores of island gap and regular gap constructions. Moreover, by-item acceptability scores show that for some items, a parasitic gap in a tensed adjunct clause is even completely unacceptable (see Figure A2 in the Appendix). This suggests that the tense of the adjunct clause, which was the only difference between Experiments 1 and 2, has an asymmetric effect on the acceptability of filler-gap constructions, in line with previous experimental observations in English (Boxell and Felser 2017; Kurtzman and Crawford 1991; Phillips 2006). Experiment 2 showed no effect of context, although the small difference in acceptability of sentences with an Invited Guest and sentences with an Uninvited Guest was similar to that in Experiment 1.

4.1 Methods

Comparing the results of Experiments 1 and 2, it appears that the tense of the adjunct clause only affects the acceptability of parasitic gap constructions. To test this impression, we pooled the data from both experiments together and analyzed the three-way interaction between tense (a between-subjects variable), construction and context (both within-subjects variables) in a linear mixed-effects model. The factors construction (island gap, parasitic gap, regular gap), tense (untensed, tensed) and context (invited, uninvited (guest)) were entered as fixed effects, together with their interactions. As in Experiments 1 and 2, we used Helmert coding for construction and deviation coding for context. The two-level factor tense was also deviation coded (−0.5, 0.5). We initially included random intercepts for both items and participants. However, because these models led to singularity issues, the by-participant random intercept was removed and only the by-item random intercept was retained.

4.2 Results

The fixed and random effects of the model with the three-way interaction are presented in Table 4. Leaving aside the effects already present in the data from Experiments 1 and 2, two effects are worth discussing. First, there was a main effect of tense: constructions with untensed adjunct clauses (M = 44.0, SE = 0.70) were perceived as more acceptable than constructions with tensed adjuncts (M = 38.3, SE = 0.70), β = 0.11, SE = 0.02, p < 0.001. However, as is clear from Figure 6, the effect of tense is not consistently the same across the different constructions. Parasitic gap constructions are rated as substantially less acceptable when the parasitic gap is in a tensed clause than when it is in an untensed clause, but the effect of tense seems to go in the other direction for both island gap and regular gap constructions. Indeed, both interactions between tense and construction were significant (Construction1 × Tense: β = −0.33, SE = 0.04, p < 0.001; Construction2 × Tense: β = 0.61, SE = 0.05, p < 0.001). Post-hoc pairwise comparisons (Tukey-corrected) confirm the pattern observed in Figure 6: parasitic gap constructions are rated as less acceptable when they are tensed (M = 41.5, SE = 1.18) than when they are untensed (M = 55.5, SE = 1.08), β = −0.52, SE = 0.03, z = −14.9, p < 0.001. In contrast, island gap constructions are rated as slightly more acceptable when they are tensed (M = 18.5, SE = 0.85) than when they are untensed (M = 11.3, SE = 0.83), β = 0.11, SE = 0.03, z = 3.29, p = 0.013 (BF₁₀ = 260.07). The effect of tense on regular gap constructions was not significant, β = 0.09, SE = 0.03, z = 2.50, p = 0.125. The latter two results likely reflect the fact that the two experiments tested different participants. The participants in Experiment 2 gave higher judgments overall, both for the island violations (on average, 7.2 points higher than in Experiment 1) as well as for the regular gap constructions (on average, 6.8 points higher than in Experiment 1).^[6] The latter difference was probably not significant because of the larger within-subjects variance in the responses to regular gap constructions (indeed, there was only anecdotal evidence for the null, BF₀₁ = 1.21). Given these methodological considerations, we do not think that the significant effect of tense on the acceptability of island violations is theoretically informative. However, because the effect of tense on the acceptability of parasitic gap constructions goes in the other direction, this difference cannot be explained by the fact that the participants in Experiment 2 gave higher judgments in general. Instead, it suggests that parasitic gaps are genuinely less acceptable when they are located in a tensed clause.

Figure 6:

Acceptability judgment scores of Experiments 1 and 2 combined, collapsed over context and grouped by construction and tense. The error bars indicate 95 % confidence intervals around the mean. *p < 0.05, ***p < 0.001.

4.3 Discussion

A number of observations in the data of both experiments deserve discussion. To begin with, none of the experimental items were rated as highly acceptable. In Experiments 1 and 2, regular gap constructions were rated with an average acceptability of around 60, whereas the grammatical fillers received an acceptability score of around 80. Even though this difference is quite substantial, we do not think that it indicates that the regular filler-gap constructions were perceived as ill-formed. The order of the words in filler-gap dependencies deviates from the canonical word order, due to which these constructions are commonly perceived as complex and difficult to process (Gibson 1998; Hawkins 1999), and receive low acceptability judgments even if they are unquestionably grammatical (Kurtzman and Crawford 1991; Phillips 2006). By contrast, the grammatical fillers were short, simple, and transparently related to the discourse context, so these differences are likely due to the length and complexity of the experimental items.

In Experiment 1, the acceptability of the target sentences was modulated by the discourse context: sentences with an Invited Guest received higher ratings than sentences with an Uninvited Guest. While this main effect of context was not explicitly predicted (see Figure 1), it is not incompatible with either the Uninvited Guest Hypothesis or with Syntactic Theory. A regular gap construction like the city, he rated without visiting the church is more likely to follow a context in which a church has been introduced (Invited, see Table 1) than a context in which a football stadium has been introduced (Uninvited), in particular given the definiteness of the church, which implies givenness (Schoenmakers 2023). The effect of context also shows that the discourse manipulation successfully manipulated the discourse accessibility of the guest. Importantly, it did not affect the acceptability of island gap constructions more strongly than it affected regular gap and parasitic gap constructions. As we noted in Section 2.2, the absence of the predicted interaction is not a matter of statistical power, as the effect of context is numerically larger for both regular and parasitic gap constructions than it is for island gap constructions, and a Bayes Factor analysis showed strong evidence for the null hypothesis of no interaction.

The pattern of results observed in Experiment 2 was slightly different. There was no effect of context, and the acceptability of parasitic gap constructions was substantially reduced compared to Experiment 1. Pooling the data from both experiments together, we found an interaction between tense and construction, showing that the manipulation of tense affected the constructions differently: parasitic gap constructions are rated as substantially less acceptable when the parasitic gap is in a tensed clause as compared to when it is in an untensed clause, whereas island gap constructions are rated as slightly more acceptable when the gap is in a tensed clause. The fact that tense has a strongly asymmetrical effect on the acceptability of these constructions is inconsistent with the discourse-processing nature of the Uninvited Guest Hypothesis. Discourse-processing theories argue that the reduced acceptability of sentences with tensed clauses is due to the discourse requirements of tense, which indexes a new discourse referent and therefore incurs additional processing cost (Chaves 2013; Culicover and Winkler 2022; Gibson 1998, 2000; Hofmeister et al. 2013). According to this view, and contrary to what we observe, the effect of finite tense should apply to all three constructions alike, because the overt subject of the adjunct clause intervenes between the filler and the gap in all three constructions.

The effect of tense on parasitic gap constructions is in line with Syntactic Theory, which holds that tense hampers filler-gap dependency formation and therefore predicts reduced acceptability for constructions with parasitic gaps in tensed clauses. However, this predicted effect of tense is categorical: parasitic gap constructions with tensed clauses are thought to be ungrammatical (Bennis and Hoekstra 1985; Huybregts and van Riemsdijk 1985). The gradient effect of tense on the acceptability of parasitic gap constructions (which has been observed previously; Boxell and Felser 2017; Kurtzman and Crawford 1991; Momma and Yoshida 2023; Phillips 2006) is therefore not entirely consistent with Syntactic Theory either. We discuss these results further in Section 5.1.

5.1 The Uninvited Guest Hypothesis versus Syntactic Theory

Experiment 1 showed that island gap constructions were judged as unacceptable, confirming the status of adjuncts as strong islands in Dutch. Regular gap and parasitic gap constructions instead received (equally) high acceptability scores. Moreover, the unacceptability of islands was not affected by supporting discourse; islands received low ratings, regardless of whether the referring argument was discourse-given or discourse-new. Neither the similar acceptability of regular gap and parasitic gap constructions, nor the absence of an effect of context on the acceptability of island gap constructions, is in line with the predictions of the Uninvited Guest Hypothesis (Figure 1). Experiment 2 showed a very similar pattern of results, except that the acceptability of parasitic gap constructions with tensed adjuncts was now in between the acceptability of (unacceptable) island gap constructions and (acceptable) regular gap constructions. This construction-specific effect of tense on the acceptability of filler-gap dependencies was also not predicted by the Uninvited Guest Hypothesis. Discourse-processing accounts tend to explain effects of tense on the acceptability of filler-gap dependencies by noting that tensed clauses are referential; they introduce referents that are tracked in the discourse and thereby add a measurable cognitive load to processing (Chaves 2013; Culicover and Winkler 2022; Hofmeister et al. 2013; Kluender 1998). Filler-gap dependencies targeting tensed clauses are therefore both more difficult to process and less acceptable than filler-gap dependencies targeting untensed clauses, as shown by the contrast between examples (19a) and (19b).

This account fails to explain our results, because the effect of finite tense on parasitic gap constructions was opposite to its effect on island gap constructions. If the discourse-processing demands imposed by tensed domains lead to a reduction in acceptability, we should have seen a similar effect in regular gap and island gap constructions.

The results of Experiment 1 are in line with Syntactic Theory in the sense that what is (deemed) grammatically licensed is judged as acceptable, and what violates a grammatical principle is judged as unacceptable. However, the gradient effect of tense on the acceptability of parasitic gap constructions in Experiment 2 was not predicted by Syntactic Theory. In Section 1.2.2 we discussed two syntactic accounts of parasitic gap constructions in Dutch. These accounts resort to different theoretical notions to explain differences in the distribution of parasitic gaps in Dutch and English. In particular, Huybregts and van Riemsdijk (1985) rely, among other things, on differences between the two languages in the government properties of prepositions to explain why Dutch does not have parasitic gaps of the type found in English. They analyze (what appear to be) parasitic gaps in Dutch as ATB extraction from coordinates. From this, it follows rather naturally that parasitic gaps in tensed clauses, which would involve subordination, are impossible. While Bennis and Hoekstra (1985) present a different analysis, they also rely on cross-linguistic differences in the government properties of prepositions in order to explain why parasitic gaps in tensed clauses are grammatical in English but not in Dutch. However, it is not so clear that Dutch and English actually differ in this respect. Experimental studies have shown that parasitic gaps in tensed clauses in English are judged as degraded but not fully unacceptable (Boxell and Felser 2017; Kurtzman and Crawford 1991; Momma and Yoshida 2023; Phillips 2006), just like what we find for Dutch. Our Experiment 2 showed that although tensed parasitic gap constructions are less acceptable than their untensed counterparts, they are not rejected to the same extent as island violations.^[7] This gradient effect of tense is not captured by formal accounts of either English or Dutch, and we suggest that it should not be explained in syntactic terms.

Indeed, a possible alternative can be found outside of syntax, by adopting Truswell’s (2011) proposal, which concerns a semantic condition. That is, it might be the case that our participants attempted to coerce the separate event readings triggered by the sentences with a tensed adjunct clause into grouped event readings, so as to achieve a meaningful interpretation that complies with semantic conditions (cf. De Swart 1998); the Single Event Grouping Condition in particular (see (7) in Section 1.2.2). Such coercion processes are computationally costly (Paczynski et al. 2014; Piñango et al. 1999) and might lead to degraded acceptability ratings. The latter happens with parasitic gap constructions in particular, because the filler in those constructions is the semantic complement of the verb in both of the matrix clause and the adjunct clause. The predicates in both clauses are therefore more strongly related in parasitic gap constructions than in both island gap and real gap constructions, in which the complements of the matrix and adjunct verb are different. People might therefore be tempted more strongly to coerce the separate events of the two clauses in parasitic gap constructions, explaining why tense has an asymmetrically degrading effect on acceptability. We do note that the details of this non-syntactic proposal are to be worked out in future research, because the relationship between processing costs and acceptability judgments is not straightforward (Sprouse et al. 2012).

5.2 Alternative non-syntactic accounts of filler-gap dependencies

In this section, we evaluate how alternative non-syntactic theories of filler-gap constructions fare with respect to our results. Following Abeillé et al. (2020) and Liu et al. (2022), these non-structural proposals are grouped into discourse-pragmatic accounts and processing accounts. The accounts are different in spirit, but they share the underlying idea that island violations are not ungrammatical and that their low acceptability is due to non-grammatical factors.

5.3 Conclusion and outlook: multiple gap constructions

In the preceding sections we noted that non-syntactic accounts cannot readily explain why parasitic gap constructions are acceptable, even if these accounts can explain variation in the acceptability of certain types of island constructions. This suggests that an analysis of parasitic gap constructions minimally requires a syntactic base. However, a final challenge to such syntactic accounts is the existence of acceptable constructions that contain multiple gaps, none of which are parasitic. These constructions have recently been used to argue in favor of a non-syntactic account of the dependencies that are classically labeled as parasitic gaps (Chaves 2013; Chaves and Putnam 2020). In this section, we explain why the existence of such non-parasitic multiple gap constructions does not provide a compelling argument against syntactic accounts of parasitic gaps.

To explain why parasitic gap constructions are relatively acceptable even though they contain a gap inside an island, it has been suggested that when the comprehender has reactivated the filler at the first gap, processing of the second gap is facilitated (Chaves 2013; Liu et al. 2022). Evidence in favor of this idea is that island constructions improve not only when a regular, licit gap is added, but also when an additional illicit gap inside an island is added. For instance, (30) and (31) are thought to be acceptable, even though they contain two illicit gaps. In both cases, the two gaps are linked to the same filler, but these are not parasitic gap constructions, because both gaps are located inside an island (Chaves and Putnam 2020; Culicover 2013).

This processing account of parasitic gaps is also supported by examples like (32), in which the filler phrase corresponding to one gap itself contains a gap, which is linked to another filler (Chaves 2013; Pollard and Sag 1994). One of the gaps in (32) is illicit, the other is licit. Yet, (32) is not a parasitic gap construction, because the two gaps are not linked to the same filler.

The acceptability of these examples seems to support the idea that processing one gap makes it easier to process another gap, even when there is no obvious syntactic relation between the two gaps. Indeed, based on examples such as these, Chaves and Putnam (2020: 84) concluded that “there is no grammatical distinction between ‘parasitic’ and ‘real’ gaps.”

However, the existence of non-parasitic multiple gap constructions such as (30)–(32) does not provide an argument in favor of a processing account of parasitic gaps, because not all multiple gap constructions are processed in the same way. For instance, comprehension experiments by Wagers and Phillips (2009) show that when people read sentences with wh-extraction, they actively search for a second gap in ATB extraction from coordinated VPs, but they do not search for a gap in postverbal adjunct clauses. The two gaps in ATB constructions are governed by the Coordinate Structure Constraint (Ross 1967), which states that one gap must occur in each coordinated phrase, so the second gap is obligatory (e.g. which computer did you test _ and repair _/*it?). The second gap in postverbal adjuncts, which are a potential parasitic gap environment, is however optional (e.g. which computer did you test _ before repairing _/it?). It thus seems that processing a second gap is easier only when that second gap is mandatorily required by the grammar (here, the Coordinate Structure Constraint). On the production side, Momma and Yoshida (2023) showed that when speakers have to produce wh-questions, they plan the verb before sentence onset only if the verb and the wh-phrase engage in a direct syntactic relationship. More specifically, speakers plan the verb of a secondary clause before producing the sentence-initial wh-phrase when that wh-phrase is a syntactic complement of the verb, as is the case in ATB constructions (e.g. [which computer] _i did you test and repair _ _i ?), but not when it is a constituent that bears a theme/patient role but is not a syntactic complement, as in parasitic gap constructions (e.g. [which computer] _i did you test before repairing _ _i ?). These results provide experimental evidence against the idea that the syntactic relationship between filler and gap is shared between ATB and parasitic gap constructions (see also Postal 1993).

Psycholinguistic data such as these show that not all multiple gap constructions are alike, and that the acceptability of non-parasitic multiple gap constructions by itself is not an argument against a syntactic account of parasitic gaps. It must first be shown that constructions like (30)–(32) are processed in the same way as regular parasitic gap constructions, which need not be the case. For instance, the relative acceptability of (30) and (31) could be due to a preference for parallel extraction (Parker 2017).^[11] Likewise, the acceptability of (32) could be explained by the fact that it suggests a possible second reading, in which both gaps are linked to heroes, as if it is a parasitic gap configuration. The seeming availability of this co-referential interpretation might interfere with people’s judgment about the acceptability of the sentence under the correct reading indicated in (32), a kind of grammatical illusion (Phillips et al. 2011).^[12] Thus, before the acceptability of these examples can be used to support a non-syntactic account of parasitic gaps, alternative processing accounts of the corresponding judgments must first be ruled out (e.g. based on parallelism). Furthermore, it must be shown that non-parasitic multiple gap constructions exhibit the same properties as regular parasitic gap constructions, which are licensed under specific conditions (Section 1.2.1; see also Chomsky 1986; Culicover 2001; Engdahl 1983; Taraldsen 1981). For instance, parasitic gaps show typical island effects, as shown in (33a). This sentence is unacceptable, arguably because the parasitic gap is embedded in not one but two adjunct clauses. As such, (33a) violates a locality restriction on extraction, as does the adjunct island violation in (33b).

The unacceptability of (33a) suggests that reactivation of the filler at the first gap does not necessarily lead to facilitated processing of the second gap, at least not to the extent that the latter becomes acceptable (compared to the gap in (33b)). It again seems that it is only easier to process a second gap when it is licensed by the grammar. In all, the distribution of parasitic gaps is constrained, and not identical to the distribution of other multiple gap constructions, like those generated by ATB movement (Postal 1993). These empirical generalizations are the explanandum for any account of multiple gap constructions, be it a syntactic theory of parasitic gaps in particular or a non-syntactic theory of multiple gap constructions in general.

In conclusion, we empirically evaluated a recent account of constraints on filler-gap dependencies, which holds that the (un)acceptability of extraction from islands can be explained in terms of demands on discourse processing. The results of two acceptability judgment experiments are inconsistent with this proposal, suggesting that the difference between islands and parasitic gap constructions is not a matter of discourse-processing complexity. The data are instead most consistent with a syntactic account of parasitic gaps, though the empirical basis of such an account would be further strengthened if it is even more directly supported by the results of both linguistic and psycholinguistic experimentation.