2.1 Rationale

There were two goals for Experiment 1. The first goal was to demonstrate that gapped predicates are strongly preferred to gapless ones if a sentence contains a filler dependency. The second goal was to determine whether the availability of resumption would improve ratings for gapless predicates. In Experiment 1, I purposefully constructed materials that contained short filler dependencies, which should be unlikely to decay in memory. Thus, I predicted that participants would overwhelmingly reject fillers with gapless predicates, i.e., there should be no detectable effect of resumption in Experiment 1. In Experiments 2–4, I changed aspects of this design, which resulted in increased acceptance of resumption.

Previous work on resumption in English (McDaniel & Cowart 1999; Mc-Kee & McDaniel 2001; Alexopoulou & Keller 2007; Omaki & Nakao 2010; Heestand et al. 2011; Han et al. 2012; Beltrama & Xiang 2016; Ackerman et al. 2018) primarily compared resumptive pronouns in syntactic islands with gaps in the same position in order to assess the ability of resumption to ameliorate syntactic island violations. The results from previous studies have demonstrated a complex pattern of judgments. Overall, however, resumptive pronouns in syntactic islands appear to be rejected at similar rates to gaps in syntactic islands, if not more often.

For instance, in an oral judgment task, McKee & McDaniel (2001) found that resumptive pronouns were largely rejected in the subject, direct object, and prepositional object positions. By contrast, resumptive pronouns were accepted at rates greater than 60% in genitive object positions, and near ceiling in subject positions in wh-islands and tensed adjuncts. McDaniel & Cowart (1999) found that gaps and resumptive pronouns were rated equally low in subject islands, but found a preference for resumptive pronouns in the subject position of a wh-island, which they attributed to amelioration of the Empty Category Principle (ECP)/that-trace constraint (Chomsky 1981) (That’s the girl that I wonder when {she/___} met you). This finding was also replicated by Omaki & Nakao (2010).

In contrast to these findings, Alexopoulou & Keller (2007) found that resumptive pronouns were often rated worse than gaps outside of island contexts. With further embeddings, the ratings between gaps and resumptive pronouns approached convergence. In syntactic island contexts, ratings for both gaps and resumptive pronouns were equally low, regardless of the number of embeddings. A similar profile was found by Heestand et al. (2011). Their results showed that resumption was dispreferred to gaps outside of syntactic islands, and both were equally low inside syntactic islands.

In contrast to these findings, Han et al. (2012) and Keffala (2013) found a relatively consistent profile of low judgments for resumptive pronouns in a variety of syntactic configurations. However, they found that the unacceptability of gaps varied more dramatically in the same structures. Keffala (2013) suggested that this may underlie the apparent selective acceptability of resumption across constructions. On her account, moderately unacceptable resumptive dependencies may be preferable to strongly unacceptable island violations, but dispreferred compared to mildly unacceptable island violations.

Finally, Beltrama & Xiang (2016) found that resumption improved ratings for sentences when participants were asked to judge the comprehensibility of the sentence as opposed to its acceptability. These findings may suggest that resumption facilitates comprehension without being grammatical. However, in a forced-choice task, Ackerman et al. (2018) found that participants were more likely to prefer resumption over gaps. They argued that previous failures to find an amelioration effect of resumption was a methodological artifact of comparison across trials in experiments with randomized stimulus presentation. When the two alternatives were significantly salient, a preference for resumption emerged.

Thus, although the pattern of acceptance rates for resumption is mixed, resumption appears to generally be disfavored in English. In all these cases, the crucial comparisons were between a gap and a resumptive pronoun in the same syntactic position. From my perspective, the more crucial question is whether the comprehender prefers resolving a filler dependency with a pronoun over the possibility of a later, well-formed gap, and how acceptability of resumption compares to the baseline of a filler dependency being unresolved entirely. For this reason, the crucial manipulation in Experiments 1–4 was the presence of a gap after a (potentially) resumptive pronoun, and the possible interpretations of this pronoun. If resumption in English is ungrammatical, then it was predicted that gapped sentences would be overwhelmingly preferred to gapless sentences, regardless of whether a resumptive pronoun was present. However, if resumption is grammatical, then I predicted that comprehenders may be more likely to accept the pronoun as a suitable resolution site for the dependency if the pronoun and filler NP corefer.

2.2 Participants

There were 53 participants recruited for Experiment 1 from Amazon’s Mechanical Turk platform (http://www.mturk.com). They must have been from the United States, and have completed over 500 HITs (Amazon Mechanical Turk tasks) with 95% approval rating or higher. All participants self-identified as native English-speakers, and the mean age of participants was 38.

2.3 Materials

For Experiment 1, I designed 36 sets of target stimuli, and 32 filler items. The fillers were 50% grammatical. Most filler sentences contained two clauses, and most contained some kind of long-distance dependency, such as a cleft, pseudo-cleft, or relativization. Ungrammatical sentences had a variety of errors, but mostly used subcategorization errors, or local morphosyntactic errors such as agreement or case mismatches. Some filler items had multiple errors.

For the target stimuli, I manipulated ±Gap and Pronoun Reference (Ambiguous, Filler, Subject), yielding a 2 × 3 design. All target sentences contained a cleft (focused relativization) dependency and a pronoun. The gender of the pronoun was counterbalanced across all conditions across items. The pronoun was always a possessor in the subject of the embedded clause. This position was chosen because extraction from genitives is not allowed in English, and English-speakers readily produce resumptive pronouns in genitive NP positions (Zukowski & Larsen 2004). Furthermore, I assumed that processing an NP with a pronominal possessor is unlikely to introduce any significant processing costs. The findings from Chacón (2015) and previous pilot studies found that other sentence structures produced substantially noisy results, which I attribute to the additional complexity associated with the the structures of these sentences. Thus, I chose to use a simpler design to minimize the effect of syntactic complexity.

For the ±Gap factor, I manipulated whether the end of the sentence contained a gap for the open filler dependency (+Gap) or not (–Gap). In the +Gap conditions, I used verbs that I judged to have a strong bias for taking an NP object, and that could plausibly take the filler as an argument. For the –Gap conditions, I used verbs that had a strong bias against taking an NP object, or transitive constructions that had NP objects. This was to ensure that comprehenders could quickly detect that the filler was not an argument of the critical verb upon entering the VP, at least if they still maintained a prediction for a gapped structure. All target materials were matched in length.

Each target stimulus contained two NPs before the critical pronoun: the subject NP of the main clause, and the filler NP. Both NPs could be potential antecedents for the critical pronoun. I manipulated the stereotypical genders of these NPs, using the strongly masculine and feminine stereotyped nouns from the norming study in Kennison & Trofe (2003). In the Ambiguous conditions, both the subject NP of the sentence and the filler NP matched in gender with the pronoun. In the Filler conditions, only the filler NP matched in stereotypical gender with pronoun, and in the Subject conditions, only the subject matched. The items were designed to keep both interpretations relatively plausible and salient. A sample set of stimuli are given in (10), and the full set of stimuli for Experiments 1–4, including fillers, are given in the Appendix. In the Ambiguous and the Filler conditions, the pronoun was potentially a resumptive pronoun, since reference between the filler and the pronoun was licensed by the stereotypical genders of the NPs. However, in the Subject condition, the pronoun was unlikely to refer to the filler, thus preventing a resumptive analysis, and strongly favoring an analysis in which the pronoun was anaphoric to the non-filler subject NP.

I predicted that there should be an effect of ±Gap, with +Gap conditions strongly preferred to –Gap conditions. This is because I understood resumption to be ungrammatical in English, and thus the addition of a pronoun would likely not reduce the “penalty” for gapless sentences. However, if comprehenders could resolve the filler dependency with the pronoun, then acceptance rates should increased for the –Gap, Ambiguous and –Gap, Filler sentences compared to the –Gap, Subject sentences. If resumptive dependencies are only constructed in contexts of high memory demand, then I did not expect the pronoun’s interpretation to have much effect in Experiment 1. This is because the sentences were specifically designed to not include structures that introduced exceptional processing difficulty, even though they were contained in a syntactic island. In other words, I predicted no interaction effects between ±Gap and Pronoun, given that processing demands were designed to be minimal in Experiment 1.

2.5 Results

For the filler items, I constructed a logit mixed-effects model using Grammaticality as the factor, and random effects for participant and item.¹ Grammaticality was significant (β = 3.96, SE = 0.34, z = 11.61, p < 0.01). The mean proportion of acceptance for Grammatical sentences was 88.1 ± 1.1%, and the mean proportion of acceptance for Ungrammatical sentences was 19.1 ± 1.3%. Thus, participants overwhelmingly accepted grammatical sentences, and rejected ungrammatical sentences.

For the target items, I constructed a logit mixed-effects models using the lme4 package in R (R Development Core Team 2008; Bates et al. 2015). I used a maximal random effects structure (Barr et al. 2013). The model included ±Gap, Pronoun, and their interaction term as factors. The model also included random slopes for ±Gap, Pronoun, and their interaction for participants and for items.² Because the factor Pronoun had three levels without an obvious choice for the reference level, I sum-coded these factors around 0. This resulted in two separate coefficients for the factor Pronoun. The ±Gap factor was fit with +Gap coded as 1 and –Gap coded as –1. The Pronoun factor was fit with two coefficients, with Ambiguous fit as 1 and 0 for the two coefficients respectively, Filler as 0 and 1, and Subject as –1 and –1. This complicates the interpretation of the mixed-effects model, but planned pairwise comparisons between each of the three Pronoun levels were compared within each of the ±Gap levels. The key comparison was between the –Gap, Filler resumptive condition compared to –Gap, Subject anaphoric condition. Pairwise comparisons used the Tukey honestly significant difference correction for multiple comparisons. The mean acceptance rates by condition are given in Figure 2. The results of the logit mixed effects model fit to the acceptability data are given in Table 1, and the results of the pairwise comparisons are given in Table 2.

Figure 2

Mean acceptance rates by condition from Experiment 1, with error bars representing one standard error from the mean.

Most strikingly, the +Gap conditions were overwhelmingly accepted, and the –Gap conditions were overwhelmingly rejected. This was predicted, because I assumed that resumption is ungrammatical in English, i.e., participants would only accept sentences in which the filler dependency resolves with a gap. Participants did not appear to treat the pronoun in the subject NP as a suitable resolution site for the cleft dependency.

Performance on Experiment 1 could also be understood as a signal detection task, i.e., correctly distinguishing grammatical sentences from ungrammatical sentences. For this reason, I computed the sensitivity index, or d′, to assess accuracy. D′-scores are ‘a measure of a participant’s ability to discriminate between the signal, while taking into account participant response biases (Macmillan & Creelman 2005). To calculate d′-scores, I treated acceptances of +Gap target items and Grammatical filler items as ‘hits’, and rejections of –Gap target items and Ungrammatical filler items as ‘correct rejections’. The average d′-scores across subjects for all items in Experiment 1 was 2.47 ± 0.14, indicating moderately high ability to discriminate between grammatical and ungrammatical sentences. In the filler items specifically, the average d′-score was 2.41 ± 0.14. Across all target items, the average d′-score was 3.25 ± 0.24, suggesting a robust ability to distinguish grammatical +Gap target items from ungrammatical –Gap target items.

2.6 Discussion

The goal of Experiment 1 was to determine whether the availability of resumption improved the acceptability of gapless sentences. The results from Experiment 1 found no such improvement. Previous work demonstrated that sentences with resumptive dependencies often were assigned low ratings in off-line judgments (Alexopoulou & Keller 2007; Omaki & Nakao 2010; Heestand et al. 2011; Han et al. 2012; Keffala 2013; Beltrama & Xiang 2016), which is consistent with the proposal that resumption is not grammatically available in English. However, these findings compared gaps with resumptive pronouns in the same syntactic positions. In this study, I compared resumptive dependencies with filler-gap dependencies, and anaphoric interpretation of pronouns with resumptive interpretations of pronouns, which is a novel comparison. The results from Experiment 1 further demonstrated that comprehenders were reluctant to interpret a pronoun as a resolution site for short filler dependencies. To my knowledge, this is also the first demonstration of the bias against resumptive dependencies using speeded acceptability judgment tasks. This is a novel finding, because it demonstrates that placing a participant under time pressure is not sufficient for increasing the acceptability of resumption.

One potentially surprising finding was that there was no detectable weak cross-over effect in Experiment 1 (Postal 1971; Wasow 1972). Weak cross-over effects are the penalty in acceptability associated with a filler NP that binds a pronoun and a gap, as in (11). If weak crossover were reliably detectable in judgments, then I would have expected lower ratings for the +Gap, Filler sentences compared to the +Gap, Subject sentences. This is because +Gap, Filler sentences contained a filler NP that was bound both the pronoun and the later gap.

In a real-time processing study, Kush et al. (2017) found that the head of a an unresolved filler-gap dependency was considered as a suitable antecedent for a pronoun, even though this interpretation resulted in a weak cross-over violation. However, this interpretation was rejected in off-line measures. These findings suggested that the weak cross-over violations are not immediately detected in real-time processing. Because speeded acceptability judgment tasks are intended to give a measure of judgments earlier in processing, the failure to find a weak cross-over effect may be interpreted as replicating this result. However, in Experiment 3 and in the overall meta-analysis in Section 6, there was evidence of the weak cross-over effect.

In summary, the results of Experiment 1 failed to demonstrate any facilitatory effect for resumption, even under time pressure. However, this is unsurprising. I purposefully designed these materials to have shorter filler dependencies, and I purposefully selected a syntactic island that I supposed would not induce significant strain on the participants’ memory. In Experiment 1, the filler NP and the pronoun were only separated by one word, the complementizer that. Additionally, the non-filler NP is the main clause subject, a syntactically prominent position. This makes it a highly accessible antecedent for the pronoun (Gundel et al. 1993; Grosz et al. 1995; Arnold 2010; Arnold & Lao 2015), which is a determining factor for resumption in English (Erteschik-Shir 1992; Ariel 1999; Chacón 2015). In other words, it may have been difficult for the pronoun to access the filler NP as its antecedent, in favor of the subject NP.

In Experiment 2, the position of the cleft filler NP and the non-filler subject NP were changed, such that the cleft was in the main clause and the non-filler subject NP was in the embedded clause. This increased the length of the cleft dependency, and also placed the filler NP as a prominent argument in the main clause, making it a more accessible antecedent for the pronoun. These changes were predicted to make the resumptive analysis more likely.

3.1 Participants

There were 60 participants recruited for Experiment 2 from Amazon’s Mechanical Turk platform, using the same inclusion criteria as in Experiment 1. All participants self-identified as native English-speakers, and the mean age of participants was 36.

3.2 Materials

The materials in Experiment 2 were similar to the those in Experiment 1, except that the filler NP and the non-filler subject NP were re-ordered, such that the filler NP appeared first in the main clause, and the non-filler subject NP appeared second in the embedded clause. This change made the filler NP more syntactically prominent, and increased the length of the cleft dependency. The materials for Experiment 2 are exemplified in (12).

3.4 Results

For the filler items, I constructed a logit mixed-effects model using Grammaticality as the factor, and random effects for subject and item, as in Experiment 1. Grammaticality was significant (β = 3.50, SE = 0.34, z = 10.38, p < 0.01). The mean proportion of acceptance for Grammatical filler items was 88.3 ± 1.0%, and the mean proportion of acceptance for Ungrammatical filler items was 27.9 ± 1.4%. As before, participants overwhelmingly accepted Grammatical filler items and rejected Ungrammatical ones.

Due to an experimenter error, 40 observations out of 2160 target items had to be discarded. After excluding these observations, I constructed a logit mixed-effects model using the same procedure as in Experiment 1. The mean acceptance rates by condition are given in Figure 3. The results of the logit mixed effects model are given in Table 3, and the results of the pairwise comparisons are given in Table 4.

Figure 3

Mean acceptance rates by condition from Experiment 2, with error bars representing one standard error from the mean. Reported p-value corresponds to results from the pairwise comparisons.

As in Experiment 1, the most striking result is that the +Gap sentences were accepted significantly more often than the –Gap sentences. However, unlike in Experiment 1, there was a significant interaction between ±Gap and Pronoun, which was reflected in the increased acceptance rates for the –Gap, Filler condition compared to the –Gap, Subject condition. Thus, in Experiment 2, there appeared be a facilitation effect of resumption, which was not observed in Experiment 1.

The average d′-score across participants and all items for Experiment 2 overall was 1.61 ± 0.11. For filler items, the average d′-score was 2.27 ± 0.14. For target items, the mean score was 1.40 ± 0.15. This suggests that participants were moderately capable in discriminating between grammatical and ungrammatical items.

3.5 Discussion

The goal of Experiment 2 was to determine whether resumption would improve sentences with longer cleft dependencies than the sentences used in Experiment 1. In Experiment 1, there was no evidence that reference between the filler NP and pronoun improved acceptability, suggesting that comprehenders did not entertain a resumptive analysis. In Experiment 2, coreference between the filler NP and pronoun was facilitated, because the filler NP was in a syntactically more prominent position, which increased the likelihood of coreference between the pronoun and the filler (Gundel et al. 1993; Grosz et al. 1995; Arnold 2010). Additionally, the dependency was longer, which is independently observed to increase the acceptability of resumption (Alexopoulou & Keller 2007; Hofmeister & Norcliffe 2013). Thus, the design of Experiment 2 favored the acceptability of resumption.

In Experiments 3 and 4, I added the word list recall task to the speeded acceptability judgment task. This was designed to strain working memory resources while participants processed and judged a sentence. On the analysis that I propose, increased strain on memory sources was predicted to increase the acceptability of –Gap target items overall. This is because the additional strain on memory should result in quicker decay of the prediction of a gapped structure in memory. This then should result in diminished ability to determine whether a filler dependency was resolved in a syntactically licensed way, i.e., with a gap.

4.1 Rationale

The goal of Experiment 3 was to determine whether the acceptability of resumption increased with an increased strain on memory. As described earlier, many previous accounts of resumption have proposed that processing difficulty and syntactic complexity partially determine the distribution of resumption (Keenan & Comrie 1977; Ariel 1999; Hawkins 1999; 2004). Resumptive pronouns typically are used in syntactic island contexts (Ross 1967; Boeckx 2003; Ferreira & Swets 2005), which are often characterized as particularly syntactically complex (Kluender & Kutas 1993; Kluender 1998; 2004; Hofmeister & Sag 2010). Additionally, length affects the acceptability of resumption in English (Alexopoulou & Keller 2007; Hofmeister & Norcliffe 2013), which itself may be due to the increased processing difficulty of longer dependencies (Kluender & Kutas 1993; Hawkins 1999; 2004; Gibson 2000; Hofmeister & Sag 2010). On my account, syntactic processing difficulty favors resumption because it leads to difficulty in maintaining a predicted structure in memory. Syntactic island configurations and length are possible sources of this difficulty, but other kinds of demands on memory may also affect the representation.

To test this, I introduced a word list recall task in Experiments 3 and 4. Participants were asked to memorize a list of words before judging a sentence as acceptable or unacceptable, and then respond to a probe word after judging the sentence. Crucially, the materials in Experiment 3 were identical to those in Experiment 1. On my proposal, the addition of this task should lead to a degraded representation of the gapped structure in memory, resulting in less sensitivity to the ungrammaticality of an unresolved filler-gap dependency.

4.2 Participants

There were 60 participants recruited for Experiment 3 from Amazon’s Mechanical Turk platform (http://www.mturk.com), using the same inclusion criteria as in Experiments 1 and 2. All participants self-identified as native English-speakers, and the mean age of participants was 35.

4.3 Materials

The sentences that participants were asked to judge were the same as in Experiment 1. For the word list recall task, there were three short nouns not mentioned in the stimulus sentence that preceded it. These nouns were randomly selected from a dictionary, and then hand-selected to avoid repetition between the word-list and the target sentence. I also avoiding using nouns that had clear semantic relations to the sentence or to each other. Most nouns were monomorphemic.

4.5 Results

In this section, I first describe performance on the word list recall task, and then performance on the acceptability judgment task. For the acceptability results, I first conducted analysis on the raw results. Afterwards, I reanalyzed the dataset, excluding trials with incorrect responses on the word list recall task. I decided to do this, because participants may have primarily attended to one of the two tasks. Primarily attending to the judgment task may produce results more similar to those in Experiment 1, but with low accuracy rates on the word list recall task. Conversely, attending to the word list recall task may produce noisier performance on the acceptability judgment task, or it may amplify the effect of memory strain on the acceptability judgments. Because I am interested in the effect of memory strain on judgments in Experiment 3, I chose to err on the side of excluding trials in which participants were not attending to the word list recall task. For transparency, I report on both results, and highlight any crucial differences.

For probe word recall accuracy, I report on both percentage of correct trials and average d′-scores across participants. Overall, probe word recall accuracy was 80.4 ± 0.4%. The average d′-score across all items and participants was 1.92 ± 0.10, suggesting moderate accuracy in the recall task in Experiment 3. For the filler items, probe word recall accuracy was 80.0 ± 0.06%. For Grammatical filler items, recall accuracy was 81.1 ± 0.9%, and for Ungrammatical filler items, recall accuracy was 78.9 ± 0.9%. In d′-scores, recall accuracy for filler items was 2.2 ± 0.17. For target items, the mean accuracy was 80.8 ± 0.06%. Mean accuracy by condition for target items is broken down in Table 5. A logit mixed effects model with ±Gap, Pronoun, and their interaction term as factors, and with participant and item as random effects,³ revealed higher recall accuracy for the Subject conditions (β = 0.47, SE = 0.14, z = 3.30, p < 0.01), but no other effects were found (all ps > 0.05). The average d′-score on the word list recall task for target items was 2.46 ± 0.26. This suggests that participants attended to the word list recall task with moderately high accuracy.

Next, I describe the acceptance rates in the acceptability judgment task. First, as before, I constructed a logit mixed effects model to analyze the filler items, with the same structure as Experiments 1 and 2. Grammatical filler items were accepted more often than Ungrammatical filler items, in both the dataset including incorrect trials on the word list recall task (β = 3.11, SE = 0.27, z = 11.41, p < 0.01), and the dataset excluding these trials (β = 3.26, SE = 0.31, z = 10.57, p < 0.01). Before excluding these trials, the Grammatical filler items were accepted 84.3 ± 0.8% of the time, and the Ungrammatical filler items were accepted 31.3 ± 1.1% of the time. After exclusion of incorrect trials, the Grammatical filler items were accepted 85.0 ± 0.9% of the time, and the Ungrammatical filler items were accepted 31.2 ± 0.01% of the time.

For the target items, I constructed a logit mixed-effects model for both datasets with the same structure as in Experiments 1 and 2, and conducted the same pairwise comparisons. The mean acceptance rates by condition after exclusion are plotted in Figure 5. The results of the logit mixed effects model after exclusion are given in Table 6, and the results of the pairwise comparisons are given in Table 7.

Figure 5

Mean acceptance rates by condition in Experiment 3, with error bars representing one standard error from the mean. Reported p-value corresponds to results from the pairwise comparisons.

As in Experiments 1 and 2, the most striking fact is the main effect of ±Gap, with the +Gap conditions accepted significantly more often than the –Gap conditions. Additionally, there was a main effect of Pronoun:2. In the dataset with incorrect memory trials excluded, pairwise comparisons revealed a dispreference for coreference between a filler NP and the pronoun within the +Gap level. This is consistent with the weak cross-over effect described in Section 2.6, since it reflects a bias against a filler NP binding both a pronoun and a gap in the same sentence. However, no comparisons were significant within the level –Gap.

The model fit to the dataset that included incorrect trials on the word list recall task showed a similar profile, with a main effect of ±Gap (β = 2.51, SE = 0.28, z = 8.91, p < 0.01), a marginal effect of Pronoun:1 (β = 0.45, SE = 0.25, z = 1.81, p = 0.07), a significant effect of Pronoun:2 (β = –0.59, SE = 0.18, z = 3.32, p < 0.01), and an interaction effect between ±Gap and Pronoun:2 only observed in this dataset (β = –0.54, SE = 0.19, z = 2.86, p < 0.01). Pairwise comparisons in this dataset again revealed a dispreference for interpreting the filler NP and the pronoun as coreferential within the level +Gap (β = 1.62, SE = 0.61, z-ratio = 2.67, p = 0.02). Additionally, Subject was accepted more often than Filler within the level +Gap, suggesting that Filler reference is specifically dispreferred (β = 1.81, SE = 0.61, z-ratio = 2.98, p = 0.01). Including incorrect memory trials appears to more clearly bring out the weak cross-over effect in Experiment 3.

Before exclusion, the average d′ score across all participants and items was 1.81 ± 0.14, and after exclusion, 1.84 ± 0.13. Thus, participants were moderately successful at discriminating between grammatical and ungrammatical sentences. For the filler items, the average d′-scores across participants was 1.86 ± 0.14 before exclusion, and 1.85 ± 0.13 after exclusion. For the target items, the average d′-score was 2.46 ± 0.26 before exclusion, and 2.62 ± 0.25 after exclusion.

4.6 Discussion

Participants performed accurately on the word list recall task, while maintaining sensitivity to the distinction between grammatical and ungrammatical sentences, both in the filler items and in the target items. This demonstrated that participants were capable of maintaining the word list in memory while simultaneously processing the complex sentences. Interestingly, participants were more accurate at the word list recall task in the Subject conditions than other target items, which was unpredicted.⁴ One possible explanation for this finding is that the Subject sentences were easier to process, thereby mitigating their impact on maintaining the word list in memory. In the Ambiguous and Filler conditions, coreference was possible between the filler NP and the pronoun. This may have engendered some processing difficulty, perhaps because participants momentarily considered the resumptive interpretation. This increased processing difficulty may have in turn resulted in degradation of the word list in memory, resulting in lower performance on the Ambiguous and Filler conditions. However, this is speculation, and this effect was not predicted.

In Experiment 1, participants overwhelmingly rejected the –Gap sentences, regardless of the pronoun’s interpretation. This was likely because the materials were designed to not induce significant demand on memory resources. By contrast, it was predicted that the addition of the word list recall task would result in increased acceptability for the –Gap sentences and for resumption generally. However, there was no specific increase in judgments for the –Gap, Subject sentences in the analysis of either dataset, as I had predicted. One possibility is that dependency length is a precondition for resumption in English. Thus, although participants were placed under memory strain in Experiment 3, they did not analyze the pronoun as resumptive, due to the dependency being too short. Relatedly, length may be a precondition for resumption because decay is a significant factor to degradation of the representation in memory. In other words, the addition of the word list recall task was not sufficiently difficult enough to engender resumption. Alternatively, participants may not have considered a resumptive analysis due to relative salience of the filler NP as an antecedent. Like Experiment 1, the filler NP was not located in the main clause. Thus, the non-filler subject NP may been significantly more accessible, and therefore participants were less able to access the filler NP as an antecedent.

The account of resumption sketched in this paper also predicted that –Gap sentences may be accepted more frequently overall with the addition of the word list recall task. Qualitatively, the –Gap target item acceptance rates increased compared to Experiment 1. However, as in Experiment 2, the acceptance rates for Ungrammatical filler sentences also increased, suggesting that participants may have simply guessed more frequently in Experiment 3. I turn to this issue in Section 6.

Next, the goal of Experiment 4 was to further explore the interaction between memory strain and length. In Experiment 4, I used the same items as in Experiment 2. These items were designed to favor resumption. If the ambiguity advantage in Experiment 3 resulted from failure to maintain both possible referents, then it is predicted that a similar ambiguity advantage should surface in Experiment 4.

5.1 Participants

There were 60 participants recruited for Experiment 4 from Amazon’s Mechanical Turk platform (http://www.mturk.com), using the same inclusion criteria as in Experiments 1–3. As before, all participants self-identified as native English-speakers, and the mean age of participants was 35. One participant was excluded for reliably low results on the word list recall task (<50%).

5.2 Materials

The sentences that participants were asked to judge were the same as in Experiment 2. For the word list recall task, the same three short inanimate nouns that were generated in Experiment 3 were used.

5.4 Results

As in Section 4.5, I first report on the word list recall results, and then the acceptability judgment results. I first conducted analysis on the raw acceptability results, and then excluding trials with incorrect memory recall.

Overall, the mean word list recall accuracy was 80.6 ± 0.4%. The average d′-score across all items and participants was 2.06 ± 0.14. This suggests that there was moderately high accuracy in the recall task. For the filler items, recall accuracy was 81.0 ± 0.7% overall. For Grammatical filler items, the recall accuracy was 82.0 ± 0.9%, and for Ungrammatical filler items, the recall accuracy was 80.0 ± 0.9%. In d′-scores, memory recall accuracy for filler items was 2.28 ± 0.17. For target items, the mean accuracy was 80.3 ± 0.6%. The mean accuracy by condition for target items is broken down in Table 8. As in Experiment 3, recall accuracy was higher in Subject target items (β = 0.60, SE = 0.16, z = 3.73, p < 0.01). The mean d′-score for the recall task on the target items was 2.24 ± 0.18.

The acceptability judgment data was analyzed with the same methods as in Experiments 1–3. For the filler items, Grammatical filler items were accepted more often than Ungrammatical filler items, in both the dataset including incorrect trials on the memory recall task (β = 3.24, SE = 0.29, z = 11.02, p < 0.01), and the dataset excluding these trials (β = 3.13, SE = 0.28, z = 11.15, p < 0.01). Before excluding incorrect recall trials, Grammatical filler items were accepted 88.2 ± 0.01% of the time, and the Ungrammatical filler items were accepted 32.6 ± 0.01% of the time. After exclusion of incorrect trials, the Grammatical fillers were accepted 88.6 ± 0.8% of the time, and Ungrammatical fillers were accepted 32.2 ± 1.2% of the time.

For the target items, I constructed a logit mixed-effects models for both datasets with the same structure as in Experiments 1–3. The mean acceptance rates by condition after exclusion are plotted in Figure 6. The results of the logit mixed effects model after exclusion are given in Table 9, and the results of the pairwise comparisons is given in Table 10.

Figure 6

Mean acceptance rates by condition in Experiment 4, with error bars representing one standard error from the mean. Reported p-value corresponds to results from the pairwise comparisons.

As in the previous three experiments, there was a strong preference for +Gap target items over –Gap target items. Unlike in Experiment 3 however, there was no main effect of Pronoun. There was a significant interaction between ±Gap and Pronoun, which was reflected in the increased acceptance rates for the –Gap, Filler sentences compared to the –Gap, Subject sentences. Thus, in Experiment 4, resumptive pronouns detectably facilitated the acceptability of a sentence, as in Experiment 2. Unlike Experiment 3, there was no evidence of the weak cross-over effect.

The model fit to the data that did not exclude incorrect memory recall trials had a similar pattern. There was a main effect of ±Gap (β = 1.32, SE = 0.20, z = 6.64, p < 0.01), and an interaction effect between ±Gap and Pronoun:2 (β = 0.41, SE = 0.14, z = 2.94, p < 0.01). Pairwise comparisons for each level of Pronoun nested within each level of ±Gap did not yield any significant differences, but there was a marginal difference between the resumptive –Gap, Filler condition and the anaphoric –Gap, Subject condition (β = 0.66, SE = 0.31, z-ratio = 2.13, p = 0.08). Thus, the crucial comparison was significant in the dataset that excluded incorrect memory recall trials, but only marginally significant in the full dataset.

The average d′ across all items and participants was 1.55 ± 0.14 before exclusion incorrect recall trials, and 1.96 ± 0.16 after exclusion. The average d′-score for filler items was 2.25 ± 0.20 before exclusion, and 2.36 ± 0.22 after exclusion. The d′-scores for all target items was 1.34 ± 0.20 before exclusion, and 1.31 ± 0.22 after exclusion.

5.5 Discussion

The goal of Experiment 4 was to determine whether acceptance rates increased for sentences with a possible resumptive analysis. The general pattern of results from Experiment 2 were replicated in Experiment 4, i.e., –Gap, Filler target items were accepted more often than –Gap, Subject items. This is unsurprising, since the same materials were used across the two experiments.

Qualitatively, the results of Experiment 4 suggested a greater acceptance rates for the –Gap target items compared to Experiments 1–3. On the proposal in this paper, the length of the filler dependency and the word list recall task both strain memory resources, which results in a diminished ability to detect the ungrammatical resolution of the filler dependency. I turn to the cross-experiment comparisons in the next section to demonstrate that this is a reliable pattern across Experiments 1–4.

6.1 Acceptability ratings

First, I examined the acceptance rates of the target items across Experiments 1–4. I collated the data from the four experiments, including the trials with incorrect word list recall performance in Experiment 3 and Experiment 4. I then constructed a logit mixed effects model with the acceptance rates of the target items as the dependent variable, and with ±Gap and Pronoun as within-participant factors. I also included two new between-participant factors, Memory Load and Length. Memory Load had levels –Memory Load (Experiments 1 and 2) and +Memory Load (Experiments 3 and 4). Length had levels Short (Experiments 1 and 3) and Long (Experiments 2 and 4). I also included random slopes for ±Gap × Pronoun by participant, and random effects by item.⁵ More complex models with random slopes by item did not converge. The results of mixed effects model are given in Table 11. Due to the complexity of this model, I will first discuss main effects, then significant interaction effects, then pairwise comparisons.

6.1.3 Pairwise comparisons

First, I conducted pairwise comparisons of the three levels in Pronoun nested within the two levels of ±Gap. This revealed the expected pattern. Within the +Gap conditions, Ambiguous items were accepted more than Filler items (β = 0.68, SE = 0.17, z-ratio = 3.90, p < 0.01), and Subject items were also accepted more than Filler items (β = 0.48, SE = 0.16, z-ratio = 2.97, p = 0.01). This reflects sensitivity to the weak cross-over constraint, because it demonstrates a bias against a filler NP binding a pronoun and a gap in the same sentence. Within –Gap conditions, Filler items were accepted more than Subject items (β = 0.47, SE = 0.13, z-ratio = 3.49, p < 0.01), and the other comparisons were not significant (p > 0.10). This demonstrates that forced coreference between a filler NP and a pronoun improves the penalty of an otherwise unresolved filler dependency, consistent with a resumptive analysis. In other words, the effect of resumption on repairing an unresolved filler dependency was observed in the meta-analysis.

Next, I discuss the interaction effect between Length and ±Gap. Pairwise comparisons revealed that Long sentences were accepted more often than Short sentences within the level –Gap (β = 1.98, SE = 0.20, z-ratio = 10.01, p < 0.01). Conversely, Long sentences were accepted less often than Short sentences within the level +Gap (β = –0.92, SE = 0.17, z-ratio = –5.50, p < 0.01). In other words, longer filler dependencies increased the acceptance of gapless resolutions, and lowered the acceptance of gapped resolutions. This implies that sensitivity to whether a filler dependency resolved in a grammatically-licensed way was diminished for longer dependencies, consistent with this proposal.

Next, I examined the three-way interaction between ±Gap, Length, and Memory Load in several pairwise comparisons. First, I examine the effect of Length within Memory Load and ±Gap. Within the level –Gap. Within this level, Short, +Memory Load items were assigned higher ratings than Short, –Memory Load (Experiment 1 vs. Experiment 3; β = 1.89, SE = 0.33, z-ratio = 5.71, p < 0.01). This suggests that the increased processing of the working memory task improved the –Gap sentences, even in shorter dependencies. Similarly, within the level +Gap, the addition of the memory recall task increased acceptance rates for longer dependencies, as revealed in the pairwise comparison between Long, +Memory Load and Long, -Short Memory Load (Experiment 2 vs. Experiment 4; β = 1.18, SE = 0.25, z-ratio = 4.77, p < 0.01).

Next, I compared the effect of Memory Load nested within the two levels of Length and ±Gap. Within the level +Gap, Memory Load facilitated acceptance rates for Long sentences (β = 0.77, SE = 0.23, z-ratio = 3.28, p < 0.01). This shows that the additional strain induced by the word list recall task had an effect on improving acceptance rates for the gapless resolutions for longer dependencies only. However, there was no effect of Memory Load on the short dependencies within the level of +Gap (p > 0.10). This implies that memory load is likely insufficient for entirely inducing enough working memory demand to facilitate resumption. Instead, additional working memory demand, as induced by the word list recall task, may magnify the processing demand on longer filler dependencies. The effect of the word list recall task is plotted in Figure 7.

Figure 7

Effect of word list recall task on acceptance rates by condition. The y-axis corresponds to the difference between mean acceptance rates between Experiments 3 and 1 (top) and Experiments 4 and 2 (bottom).

6.2 Sensitivity scores

One possible concern about the comparison between Experiments 1–4 is that the increased acceptance rates of –Gap target items may not reflect anything specific about the processing of filler dependencies. Rather, these results may reflect noisier performance induced by increased task difficulty. Increased guessing on acceptability trials should raise the acceptance rates of –Gap items, due to higher proportion of at-chance performance. The finding that +Gap, Long items were accepted less often than +Gap, Short items may also be consistent with this explanation, i.e., this may show that additional processing cost associated with longer filler dependencies may result in both –Gap and +Gap acceptance rates drawing closer to chance.

To distinguish these hypotheses, I next examine the distribution of d′-scores across participants in the four experiments. As I described in Section 2.6, the d′-score is a way of quantifying individual participants’ sensitivity that takes into account individual response bias. If the cross-experiment results were due to increased task difficulty resulting in noisier performance, then there should be a decline in d′-scores with the addition of the word list recall task and the longer filler dependencies. First, I discuss the d′-scores for target items. Then, I turn my attention to the d′-scores for filler items.

6.2.2 Filler sensitivity scores

To test this, I fit a linear mixed effects model to the d′-scores for the filler items with the same structure. Memory Load had no effect on the d′-scores (β = –0.01, SE = 0.22, t(216) = –0.06, p = 0.95), nor did Length (β = 0.15, SE = 0.22, t(222) = 0.68, p = 0.51). However, there was a marginal interaction effect between Memory Load and Length (β = –0.54, SE = 0.31, t(222) = –1.75, p = 0.09). Pairwise comparisons revealed that +Memory Load, Short had reduced d′-scores compared to –Memory Load, Short (β = 0.79, SE = 0.31, t(223) = 2.60, p = 0.01), but there was no difference between +Memory Load and –Memory Load within the level Long (p > 0.10). The mean d′-scores by experiment and by item type (Filler vs. Target) is shown in Figure 8.

Figure 8

Comparisons of d′-scores across Experiments 1–4. Error bars correspond to one standard error from the mean. The results from Experiments 1 and 2 were coded as -Memory Load, and Experiments 3 and 4 were coded as +Memory Load. The results from Experiments 1 and 3 were coded as Short, and the results from Experiments 2 and 4 were coded as Long.

6.3 Discussion

Overall, this meta-analysis suggested that the cross-experiment manipulations of memory load and dependency length affected the acceptability of –Gap sentences. Moreover, it confirmed the general pattern for preferring resumptive dependencies over anaphoric dependencies, when no syntactically licensed resolution site is available. Secondly, I argued that comparing the sensitivity scores between target items and filler items across the four experiments revealed a more extreme profile on the target items. Although performance may have been noisier with more complex experimental paradigms, overall, the effect was magnified for the –Gap sentences in the target items. This suggests that the cross-experiment manipulation of processing difficulty specifically impacted comprehenders’ ability to detect an unresolved filler-gap dependency.

Finally, it is worth pointing out that there was no interaction between Pronoun and the critical factors that manipulated processing difficulty, Memory Load and Length. My account would likely be more strongly supported with such an interaction. On my account, it is expected that the facilitatory effect of co-reference between a pronoun and a filler phrase should be increased when the comprehender is placed under processing difficulty. However, this meta-analysis did not demonstrate this. This may partially be due in part to the lack of resumptive effect in Experiments 1 and 3, and the small magnitude of the effect overall. Thus, I leave investigating this more systematically for future research.

7.1 What’s stored in memory, and how?

The account I provide here makes some assumptions that are not shared by other work on the processing of filler dependencies. In this paper, I argued that filler-gap dependency processing depends on active maintenance of a predicted structure. Typically, this process is characterized as active maintenance of a filler NP while searching for a resolution site (Wanner & Maratsos 1978; Crain & Fodor 1985; Frazier 1987; Wagers & Phillips 2014). On this view, active gap formation is motivated by a need to discharge the dependency, so that a thematic role may be assigned to the filler NP (Aoshima et al. 2004), or to alleviate the burden of maintaining it in memory. On this generally accepted view, the eagerness to discharge the dependency results in the filled-gap effects, as discussed in Section 1, and other evidence of early interpretation, such as the plausibility mismatch effect (Traxler & Pickering 1996). Similarly, EEG results show a sustained left anterior negativity, which has been interpreted as reflecting active maintenance of the filler NP in working memory (King & Kutas 1995; Fiebach et al. 2001; 2002; Phillips et al. 2005). Finally, as discussed in section 1, previous accounts of resumption specifically attribute the availability of resumption to low accessibility of the filler NP (Ariel 1999).

There are two reasons why it’s important that the predicted resolution site is maintained in working memory for my account. First, I argued that acceptability of resumption in part depends on a diminished sensitivity to the grammatical requirement that a filler binds a gap, as reflected in the variable acceptance rates for –Gap target items across Experiment 1–4. On my view, detecting the ungrammaticality of an unresolved filler dependency relies on comparing the predicted structure in memory against the bottom-up input. On the traditional view, it presumably relies on a failed prospective search for a gap. Moreover, in order for resumption to facilitate acceptability, the filler NP must be a sufficiently accessible antecedent for the pronoun (Erteschik-Shir 1992; Ariel 1999). If resumption is only licensed when the filler NP has sufficiently degraded in memory, then it is also less likely that the filler NP will be retrieved as an antecedent.

To clarify the difference, consider the contrast in (13). On my proposal, (13a) is recognized as acceptable when the comprehender is able to detect that the predicted structure matches the input, and (13b) is recognized as unacceptable when the comprehender detects that the expected gap is not available. If this prediction has sufficiently degraded, then (13a) may be perceived as a selectional violation.⁷ In (13b), if the expectation for a gap has sufficiently degraded, then the sentence will be perceived as more acceptable. Then, coreference between the pronoun and the filler NP may further increase acceptance, by means of recovering a coherent interpretation. On theories that posit that the filler NP is a fragile representation maintained in memory, the explanation for identifying the grammaticality of (13a) and the ungrammaticality of (13b) is largely the same. By contrast, in (13b), if the filler NP has been forgotten, then the pronoun would be most likely interpreted as a pronoun lacking an antecedent. On the assumption that antecedent retrieval relies on a cue-based retrieval strategy that is sensitive to the accessibility of target representations in memory (cf. Lewis & Vasishth 2005), a degraded filler NP should be difficult to retrieve as the antecedent of it, leaving it without an antecedent. Thus, in order for resumption to be perceived as acceptable while producing a coherent interpretation, it is important to distinguish the status of the filler NP in memory from the ability to detect whether a filler dependency has grammatically resolved.

The existing evidence on active gap formation processes is largely equivocal on what information is stored in working memory. For instance, filled gap effects are sometimes characterized as supporting an account in which the comprehender maintains a representation of the filler NP in memory while prospectively searching for a gap. However, filled gap effects could follow from incrementally checking the predicted structure maintained in memory against the input sentence. Similarly, sustained anterior negativity results may reflect the active maintenance of the structured prediction, and not the filler NP as such.

One reason to suspect that the comprehender actively maintains features of the filler NP is the profile of “D-linked” wh-phrases. D-linked wh-phrases presume a set of alternatives available in the discourse (Pesetsky 1987). Additionally, they are argued to be more robustly represented in memory, which in turn mitigates the impact that distance and syntactic islands have on dependencies headed by these structures (Goodall 2015; see also Frazier & Clifton Jr. 2002; Diaconescu & Goodluck 2004; Hofmeister & Sag 2010). Compellingly, Alexopoulou & Keller (2007; 2013) demonstrated that D-linking and resumption interact, such that D-linking may increase the acceptability of resumption in syntactic islands. Similarly, it has been demonstrated that feature overlap between the the filler NP and subsequent linguistic material results in interference effects (Rizzi 2013; Atkinson et al. 2015; Villata et al. 2016), suggesting that these features are maintained in working memory. Thus, I take these findings to indicate that some working memory resources are devoted to maintenance of the filler NP. However, it is still possible that the comprehender also maintains an active prediction of the resolution site with some representation of the filler NP.

Another tension between my proposal and previous literature is the interpretation of length effects. On my proposal, length adversely affects maintenance of predicted structure in memory, and this predicted structure is required for determining whether a filler dependency has a syntactically licensed resolution. This predicts that length may diminish filled-gap effects. Wagers & Phillips (2014) and Chow & Zhou (2018) examined filler-gap dependencies that cross variable distances. In both sets of studies, they found evidence of active dependency formation at all dependency lengths. Wagers & Phillips (2014) found that filled-gap effects were observed over variable distances, but detection of implausible filler-verb relations was delayed until after the reactivating the filler NP for longer dependencies. By contrast, Chow & Zhou (2018) found evidence of sensitivity to implausible filler-verb relations at all distances.

On my account, it is unlikely that the representation of the gap completely degrades. For instance, if the syntactic prediction had completely degraded, then the –Gap target items should have been rated near ceiling. Conversely, the +Gap target items should have been accepted at lower rates. This is because all +Gap items contained verbs or prepositions that required objects. Thus, if the prediction for a gap had completely disappeared, then these items should have been perceived as selectional violations, e.g., an obligatorily transitive verb missing a required argument NP. However, even in Experiment 4, there was still a preference for +Gap over –Gap, implying that comprehenders still distinguished between grammatical and ungrammatical resolutions. This could reflect residual representation of the predicted structure in memory, at least on some trials. Even so, my account predicts that active dependency formation should diminish over time, if not vanish. However, this does not seem to match the findings by Wagers & Phillips (2014) and Chow & Zhou (2018). The plausibility mismatch effect in Chow & Zhou (2018) was similar in magnitude across filler dependency lengths, implying no substantial degradation of the semantic information of the filler NP. It’s crucial to point out that this tension extends beyond my proposal. If comprehenders are capable of executing active dependency formation across long distances, then the findings by Alexopoulou & Keller (2007) and Hofmeister & Norcliffe (2013) similarly need explanation. If grammatically-sensitive active dependency formation processes extend across long distances without losing sensitivity, then it is unclear why these processes entertain ungrammatical dependencies for long dependencies.

Relatedly, an anonymous reviewer points out that the findings by Wagers & Phillips (2009) may be problematic for my account. As I discussed in Section 1, they found that active dependency formation persists into the second conjunct in ATB constructions. This may be surprising on my account, since conjoined VPs are arguably complex configurations. Wagers & Phillips (2009) may have detected continued active dependency formation in the second conjunct of ATB configurations for two possible reasons. First, it may be that conjoined VPs are not particularly complex, for some relevant metric of complexity. Second, it may be that after detecting the conjunct and, the comprehender generated a new prediction for upcoming structure, conditioned on how the first conjunct was produced. Previous work has argued for a processing-based parallelism preference on ATB configurations (Frazier et al. 2000; Frazier & Clifton 2001; Apel et al. 2007; Sturt et al. 2010; Knoeferle 2014), that is, a preference that a gap in the subsequent conjuncts is located in the same relative position as it was in the first conjunct. This parallelism preference may be attributed to a predictive mechanism (Frazier et al. 2000). Upon detecting the conjunct and, the comprehender may generate a new prediction before entering the second VP. If so, then this second prediction is likely to be very active. Importantly, Parker (2017) found that active dependency formation in ATB configurations obeys parallelism, i.e., comprehenders expect gaps in the the second conjunct to be located in the same position as in the first conjunct. If active dependency formation partially depends on structured prediction, then this may be expected.

Finally, I have remained agnostic about the nature of the degradation of representations in memory and the structure of short-term memory. Classically, theories of working memory postulate privileged memory buffers for maintaining information accessible to computations (Baddeley & Hitch 1974; Baddeley 2000). Limitations on working memory follow from limited working memory resources, and storing representations in these buffers over time is costly. For instance, Wanner & Maratsos (1978) proposed that processing of filler-gap dependencies requires maintaining the filler in a specialized buffer for later retrieval at the gap. However, more recent cue-based retrieval models reject privileged memory buffers (Anderson 1983; Lewis & Vasishth 2005; Jonides et al. 2005; McElree 2006). On these accounts, recently built representations are stored in memory alongside all others. If recently built structures are more accessible, this is due to residual activation from having been recently processed in the focus of attention. However, as memory representations go unused or unactivated, they may degrade over time (decay). Additionally, difficulty can arise form trying to access a representation that shares properties with other active representations in memory (interference; see, e.g., Van Dyke & Lewis 2003).

The results of this study are likely uninformative for arbitrating between these approaches to short term memory. For instance, the effects of length could either be cast as decay of the predicted structure in memory or as interference between the predicted structure, the linguistic material processed after the filler NP, and the word list. Similarly, the effect of the word list could be cast as interference between the word list and the material stored in working memory, or as competition for limited working memory resources.

However, if the impact of the word list recall task is ultimately due to interference, then the form of the stimuli in the list may have differential effects on acceptability. For instance, asking participants to memorize sequences of non-linguistic stimuli, such as numbers or shapes, may be less likely to improve ratings for gapless filler dependencies. If so, this may suggest that linguistic interference underlies the improved acceptability of Memory Load in these studies. Similarly, one way to directly isolate decay as a factor would be to keep the linguistic form of the sentences the same, but modulate the stimulus onset synchrony (SOA), such that there is increased time between the filler NP and the gapless predicate. If increasing the length of time improves the ratings of –Gap predicates, then this demonstrates that decay of the representation over time uniquely contributes to the loss of the predicted structure in memory. Finally, it may be possible to directly modulate the effect of interference on the prediction maintained in memory while keeping sentence length the same by selecting different island types, given an independent measure of syntactic complexity. Alternatively, interference may be increased without affecting sentence length by manipulating the degree of feature overlap between the filler NP and the subsequent linguistic material, to induce similarity-based interference.