2.1 Materials

All experimental items were wh-questions in which the wh-dependency crossed a temporal adjunct island boundary. Adjunct type was held constant across all materials to avoid confounds associated with variation across adjunct classes (e.g., temporal vs. manner or reason adjuncts), which have been shown to differ in their susceptibility to extraction and amelioration effects (Chaves & Putnam 2020).

The materials were further constructed to ensure discourse relevance between the matrix predicate and the event described in the adjunct clause. Across all conditions, the adjunct event plausibly motivated the emotional or evaluative predicate in the matrix clause, yielding coherent event interpretations and minimizing the likelihood that pragmatic infelicity or event incoherence independently influenced acceptability judgments or processing measures (Truswell 2011; Chaves & Putnam 2020).

The experiment was not designed to investigate satiation effects. Lexicalization, repetition, and spacing between critical items were not controlled in a manner that would permit assessment of exposure-based changes in acceptability (Chaves & Putnam 2020); accordingly, no claims regarding satiation are made on the basis of the present data.

The experiment employed a 2 × 2 within-subjects design crossing the factors Animacy (animate vs. inanimate) and Tail (gap vs. resumptive pronoun), yielding four experimental conditions. The materials consisted of eight items (hereafter referred to as lexicalizations in the statistical analyses), each corresponding to a single base sentence frame with identical lexical content across conditions. Each item was instantiated in all four experimental conditions, resulting in 32 experimental sentences in total. Each participant was presented with 16 experimental sentences (four per condition), together with 32 filler sentences, for a total of 48 trials per participant.

All experimental sentences were bi-clausal to control for structural complexity and to keep dependency length constant across conditions. The same lexicalizations were used across conditions to isolate the effects of the manipulated factors. A three-word adjunct phrase (preposition + DP) was included following the dependency tail as a controlled spillover region, allowing potential reading-time effects extending beyond the integration site to be captured.

A summary of the experimental conditions is illustrated in Example (5), and the full list of experimental sentences is presented in Appendix A. Sentences predicted to be unacceptable under the Arabic theoretical literature are marked with an asterisk (*).

The filler items were also bi-clausal and matched the experimental sentences in length and structure. Half were fully acceptable (e.g., 6), and half were ungrammatical, containing a range of violations such as subject–verb agreement errors (7) and subcategorization violations (8). Including both grammatical and ungrammatical fillers ensured that participants were exposed to the full range of acceptability, encouraging them to use the entire rating scale and reducing the likelihood of response bias.

In total, 32 filler items were included. The fillers were evenly divided between grammatical (n = 16) and ungrammatical (n = 16), yielding a 1:1 ratio of grammatical-to-ungrammatical fillers. This ensured participants were regularly exposed to both acceptable and unacceptable sentences.⁶

2.2 Participants

Eighty-three native speakers of Saudi Arabic (aged 18–25) participated in the study. All participants were undergraduate students enrolled in an English-language program and received course credit for their participation. None had received formal training in linguistics at the time of participation, as linguistics courses are taken in later stages of the degree.

Participants were fluent speakers who reported using Saudi Arabic regularly in their daily lives. They were informed about the purpose of the study and their rights as participants, and all gave informed consent before beginning the task.

2.3 Methodology

The study employed a repeated-measures design, with each participant exposed to all four experimental conditions. To control for lexical repetition, carryover, and order effects, we implemented Latin Square counterbalancing. The 32 critical sentences (8 per condition) were distributed across eight pseudo-randomized lists, ensuring that no item ‘lexicalization’ appeared more than once within any list.

Each participant completed one of two online experiments, each consisting of four lists. Across these lists, each item ‘lexicalization’ appeared in two different forms, but no participant saw more than one version of a given item ‘lexicalisation’ in one list. Trial order was randomized individually within each list. All sentences (critical and fillers) were presented word-by-word in the SPR task and were followed by an acceptability judgment.

The experiment was administered online using PsychoPy (version 2023.2.1) (Peirce et al. 2019) and hosted on Pavlovia (Open Science Tools Ltd. 2023). Participants completed the task in approximately 25 minutes. The self-paced reading task used a moving-window design, where sentences were read one word at a time. After each sentence, participants rated its acceptability on a 7-point Likert scale (1 = completely unacceptable, 7 = fully acceptable). The instructions emphasized that participants should base their judgments on their own intuitions.

2.4 Results

2.4.1 Acceptability judgment analysis

Before turning to the experimental results, we note that participants performed the task as expected: grammatical fillers received high acceptability ratings (mean = 6.43), whereas ungrammatical fillers were rated substantially lower (mean = 2.77). A full analysis of filler responses is reported at the end of this section.

We now turn to the descriptive results of the acceptability judgment task. The two factors manipulated in this study were Animacy (Animate vs. Inanimate) and Tail (Gap vs. RP). The primary research question was whether animacy improves the acceptability of island-violating dependencies, and whether this improvement is modulated by the type of dependency tail. Table 1 presents the mean acceptability ratings across the four experimental conditions.

Table 1: Mean acceptability ratings by Animacy and Tail.

Animacy	Tail	Mean	SD	N
Animate	Gap	2.64	1.93	332
Animate	RP	3.20	2.17	332
Inanimate	Gap	2.50	1.89	332
Inanimate	RP	2.42	1.80	332

As shown, Animate–RP conditions yielded the highest mean acceptability, while Inanimate–RP conditions received the lowest scores. Gap conditions were rated lower overall, but ratings for Animate–Gap were slightly higher than for Inanimate–Gap.

To visualize these trends, raw ratings were plotted for each Animacy × Tail condition in Figure 1.

Figure 1: Acceptability ratings by Animacy × Tail condition.

To test whether the observed patterns were statistically reliable, we fitted a series of ordinal cumulative link mixed models (CLMMs) using the ordinal package in RStudio. Likelihood ratio tests showed that adding Tail significantly improved model fit (p = .002), adding Animacy provided a further improvement (p < .001), and including the Animacy × Tail interaction yielded an additional significant improvement (p = .0004). Random intercepts were included for Participant and Lexicalisation. Post-hoc comparisons confirmed that adding random slopes did not improve model fit (all p > .55), so the simpler random-intercepts-only model was retained. The final model thus included both main effects and their interaction, with random intercepts:

Acceptability ~ Animacy * Tail + (1|Participant) + (1|Lexicalisation)

A summary of the fixed effects is given in Table 2.

Table 2: Fixed effects from the best-fitting CLMM (Inanimate and Gap were reference levels).

Predictor	Estimate	Std. Error	z-value	p-value
Tail: RP	–0.053	0.156	–0.342	0.732
Animacy: Animate	0.210	0.155	1.355	0.175
RP × Animate Interaction	0.770	0.219	3.522	0.0004

To aid interpretation, Figure 2 plots the predicted probabilities of each acceptability rating (1–7) across conditions. Each vertical line corresponds to a specific Animacy × Tail condition. Points where vertical lines intersect the rating curves represent the model-estimated probability of each rating value.

Figure 2: Ordinal cumulative link model showing predicted probabilities of each acceptability rating (1–7) for every Animacy × Tail condition. The horizontal axis indexes discrete experimental conditions; the curves represent model-estimated probabilities rather than trends over a continuous dimension. Conditions that appear close in the figure have similar predicted probability distributions across rating levels.

The plot revealed a clear pattern:

• Inanimate–RP and Inanimate–Gap were the least acceptable conditions, with a high probability of receiving the lowest rating.

• Animate–Gap was rated slightly higher; the probability of receiving the lowest rating decreased compared to Inanimate–RP and Inanimate–Gap.

• Animate–RP produced the lowest probability of low ratings.

Post-hoc Tukey-adjusted pairwise comparisons confirmed this pattern (Table 3). Animate–RP was rated significantly higher than Animate–Gap (p < .0001) and Inanimate–RP (p < .0001), whereas no reliable differences emerged between Inanimate–Gap and Inanimate–RP, or between Inanimate–Gap and Animate–Gap.

Table 3: Post-hoc contrasts.

Contrast	Estimate	SE	z-ratio	p-value
Inanimate Gap – Animate Gap	–0.210	0.155	–1.355	0.528
Inanimate Gap – Inanimate RP	0.053	0.156	0.342	0.986
Animate Gap – Animate RP	–0.717	0.152	–4.716	<.0001
Inanimate RP – Animate RP	–0.980	0.154	–6.346	<.0001

For completeness, estimated marginal means illustrating this interaction are provided in Appendix B.

Random effects revealed that most variance was attributable to participants (τ₀₀ = 2.79), while lexical items contributed relatively little (τ₀₀ = 0.11). The intra-class correlation (ICC = 0.47) indicates that nearly half the variance stemmed from participant-level clustering. Thus, while the animacy effect was broadly consistent across items, ratings varied considerably across individuals.

Figure 3 illustrates this variability. It plots raw ratings across conditions, highlighting broad dispersion but also a tendency toward higher ratings in the Animate–RP condition. This underscores that the interaction effect is not driven by outliers but reflects a robust group-level trend.

Figure 3: Raw acceptability ratings by participant across the four conditions.

To further illustrate the distribution of acceptability judgments across participants, histograms of participant-level mean ratings by condition are provided in Figure C1 in Appendix C.

The full statistical output of the CLMM is reported in Appendix D. We further tested a model with the maximally converging random effects structure (Animacy * Tail | Participant; Animacy * Tail | Lexicalisation), since models using only random intercepts are known to elicit increased Type 1 error rates (Barr et al. 2013). This model converged and yielded a fixed-effect pattern that was qualitatively consistent with the simpler random-intercepts model. In particular, the Animacy × Tail interaction remained significant and in the same direction in both models (β = 0.77, p < .001 in the intercept-only model; β = 0.98, p < .001 in the maximal model). Full results of the maximal model are reported in Appendix E.

For the filler sentences, participants rated the grammatical items consistently at the top of the scale, with most responses clustering at 7. By contrast, the ungrammatical fillers, regardless of the type of violation illustrated in (3)–(5), were rated toward the lower end, with the majority of responses at 1–2 and progressively fewer at higher values. The distributional plots (Figure 4) show that participants used the full 1–7 scale and that fillers produced the expected contrast between grammatical and ungrammatical sentences.

Figure 4: Boxplot contrasting the acceptability patterns of grammatical vs. ungrammatical fillers.

2.4.2 Self-paced reading (SPR) analysis

We evaluated reading times at the spillover region immediately following the RP or gap, which consisted of a three-word adjunct phrase. We targeted the spillover region rather than the integration site itself because verbs hosting resumptive pronouns involve cliticization, which independently inflates reading times relative to gap conditions and could therefore obscure effects of interest. An example stimulus is shown below, with the spillover region in bold:

1. (9)

1. meen / ʔayš
2. who / what
1. tafaʔʒaʔt-i
2. got.surprised-you
1. lamma
2. when
1. al-mudīra
2. the-principal
1. intaqadat __/-uh
2. criticized __/(him/it)
1. fߓī
2. in
1. ijtimāʕ
2. meeting
1. al-idāra?
2. the-administration
1. (‘Who/what did you get surprised about when the principal criticized __ / him in the administration meeting?’)

Table 4 illustrates how region labels (N1–N4, RT0–RT3) correspond to individual words in the sentence.

Table 4: Illustration of region labels (N1–N4, RT0–RT3) used in the self-paced reading analysis, mapped onto the words of an example experimental sentence.

Region	Word (example sentence)	Word (English translation)
N1	meen / ʔayš	Who / What
N2	tafaʔʒaʔt-i	got.surprised-you
N3	lamma	when
N4	al-mudīra	the-principal
RT0 (integration site)	Intaqadat__/-uh	criticized-/(him/it)
RT1 (spillover region)	fī	in
RT2 (spillover region)	ijtimāʕ	meeting
RT3 (spillover region)	al-idāra	the-administration

Reaction times (RTs) were cleaned at the region level by removing values below 100 ms and above 3000 ms for all regions. Extremely fast responses were taken to reflect accidental key presses, while very slow ones likely signalled distraction. After trimming, reaction time at the spillover region was computed as the sum of the RTs of the three words following the integration site, so that this measure was based on the same cleaned trials. This procedure kept about 78% of the data, which is typical for self-paced reading studies, and left us with a dataset that better reflects natural sentence processing.

Figure 5 shows the region-by-region reading times across conditions. Reading times in the pre-critical regions (N1–N4) were comparable across conditions, indicating that participants were engaged with the task and that differences did not emerge prior to the integration site. At the integration region (RT0), which corresponds to the subcategorizing verb, RTs were elevated in RP-animate condition, but it could be due to the cliticization process. In the following spillover regions (RT1–RT3), reading times were systematically longer in the animate conditions compared to the inanimate conditions, while no consistent differences were observed between gap and RP conditions.

Figure 5: Region-by-region reading times across conditions (Left panel: gap conditions; right panel: RP conditions). Mean RTs (in seconds) are plotted for pre-critical regions (N1–N4), the integration region (RT0), and spillover regions (RT1–RT3). Lines indicate condition means (Animate–Gap, Animate–RP, Inanimate–Gap, Inanimate–RP), with 95% confidence intervals shown as vertical bars. Reading times were comparable across conditions in the pre-critical regions, elevated at the integration region (RT0) in Animate-RP condition, and consistently longer for animate fillers in the spillover regions.⁷

Figure 6 presents the mean raw reading times for the three spillover words (RT_Sum) across conditions (the sum of RT1, RT2, RT3). The descriptive pattern suggests longer reading times in the Animate conditions relative to the Inanimate ones.

Figure 6: Raw Spillover RT across all conditions.

To correct for skewness in the RT_Sum distribution, a Box–Cox transformation was applied. The estimated λ was –0.18, close to zero, indicating that a log transformation provides a suitable approximation (Box & Cox 1964). Distributional diagnostics of the transformed RTs are provided in Appendix F.

Transformed reaction times at the spillover region (i.e. RT_Sum) were analyzed using Generalized Additive Models (GAMs) (Wood 2011) implemented in the mgcv package in RStudio (Version 1.1.419). While Box–Cox transformation reduces skewness in reaction times, it does not address other non-linear properties of reading data, such as practice and fatigue effects, spillover between regions, and individual variability in temporal dynamics. GAMs allow flexible smooths for these predictors and random smooths for participants, making them well suited for psycholinguistic time-course data (Baayen et al. 2016).

Likelihood ratio comparisons between nested GAMs showed that the best-fitting model included Animacy as a fixed effect, with random smooths for Participant and Lexicalisation. The smooth for Participant was highly significant (edf ≈ 70, F = 8.86, p < .001), indicating substantial individual variability, whereas the smooth for Lexicalisation was not (edf ≈ 2, F = 0.40, p = .20). Adding Tail or the Animacy × Tail interaction did not improve model fit. Thus, Animacy emerged as the sole significant predictor of reading times.

Specifically, sentences with inanimate fillers were read significantly faster than those with animate fillers (β = –0.020, SE = 0.005, t = –4.05, p < .001). The model explained 45% of the deviance, with an adjusted R² of 0.409 (REML = –1066.4). This effect is visualized in Figure 7, which shows that median RTs were consistently lower for inanimate fillers.

Figure 7: Distribution of Box–Cox–transformed reading times (RTs) by animacy condition. Each violin plot shows the full distribution of RTs, with an embedded boxplot marking the interquartile range and median. Median RTs are consistently lower for inanimate fillers, reflecting the significant animacy effect identified in the GAM analysis.

Model diagnostics supported the adequacy of the fit. Fitted versus observed values showed a strong linear trend with no signs of heteroscedasticity, and residuals were symmetrically distributed, further confirming the reliability of the model (see Appendix G).

To ensure the robustness of these results, we also ran supplementary linear mixed-effects models using the lme4 package (Bates et al. 2015). The results of these models, reported in Appendix H, converge with the GAM analysis in showing a reliable animacy effect and no evidence for a tail effect or interaction.

To assess the reliability of the self-paced reading data under the acceptability-judgment task, we conducted an independent filler-based analysis. This analysis shows that sentence-level reading times reliably distinguish grammatical from ungrammatical filler items (see Appendix I).

Furthermore, we examined whether variation in offline acceptability judgments across participants was systematically related to variation in online processing difficulty. To this end, we conducted a participant-level correlation analysis between each participant’s mean acceptability ratings and their mean spillover-region reading times. This analysis revealed no reliable association between acceptability and reading times (Pearson’s r = .19, p = .097). The same conclusion was obtained using a rank-based correlation (Spearman’s ρ = .14, p = .218). These results suggest that acceptability judgments do not systematically track online processing difficulty at the participant level.

3.1 Syntactic accounts

The findings are not predicted by categorical syntactic accounts of island effects. Within theoretical syntax, island-violating wh-dependencies are assumed to be ungrammatical due to violations of movement constraints (Chomsky 1973; 1986; Huang 1982). In the context of Arabic syntax, Aoun et al. (2009) argue that while gap dependencies across islands violate movement constraints, RPs may rescue grammaticality through establishing binding relations, but only with wh-expressions that project a DP-layer (see Section 2), as with animate who but not inanimate what.

Our results diverge from the predictions of this account. The analysis in Aoun et al. (2009) predicts a sharp contrast between licit and illicit dependencies: animate–RP conditions should be judged acceptable, whereas all gap conditions and inanimate–RP conditions should be unacceptable. Instead, we observe a gradient pattern of acceptability. Animate–RP sentences receive higher ratings than the other conditions, but their acceptability remains substantially below full acceptability. The modest, graded improvement observed in the acceptability of animate–RP sentences challenges the syntactic claim that resumptive pronouns in Arabic license island-violating who-questions through binding. If such a binding relation were fully responsible for licensing these dependencies, we might expect a stronger or more categorical improvement in acceptability than what is observed, rather than the relatively subtle shift in mean ratings. This subtle, gradient nature of the animacy–resumption interaction suggests that their ameliorating effect operates beyond the domain of categorical syntactic licensing.

However, this does not imply that the results are incompatible with syntactic explanations per se, but rather that they call for syntactic models that allow graded well-formedness. As noted in the Background (see Note 1), although island constraints are traditionally treated as categorical in generative syntax, an increasing body of work has argued that grammatical well-formedness itself may be gradient, even within the competence grammar (Haegeman & Jiménez-Fernández & Radford 2014; Villata & Sprouse & Tabor 2019; Villata & Tabor 2022).

Within this perspective, the present findings can be naturally interpreted under the Self-Organized Sentence Processing (SOSP) model (Villata et al. 2019; Villata & Tabor 2022). SOSP assumes a flexible structure-formation system in which the parser attempts to assemble an interpretable representation even under suboptimal conditions by coercing elements into roles that are not ideally licensed. When no sufficiently compatible grammatical structure is available, coercion fails and acceptability remains low. When a partially compatible structure exists, however, the system may converge on a strained but interpretable parse, yielding intermediate acceptability rather than absolute ungrammaticality. From this perspective, the intermediate ratings observed in the Animate–RP condition reflect a reduction in structural strain: although adjunct island constraints continue to block full dependency formation, the presence of a resumptive pronoun—particularly with an animate filler—provides a configuration that is similar enough to independently licensed resumptive structures in Arabic, such as relative clauses (Aoun et al. 2009), to support partial convergence.

A remaining question is why coercion succeeds in the Animate–RP condition but fails in the Inanimate–RP condition. Under the SOSP account, this asymmetry can be explained by considering the constraints governing the grammatical distribution of resumptive constructions in Arabic. Independently licensed resumptive dependencies in Arabic, most prominently in relative clauses, are systematically associated with referential, highly individuated fillers (Aoun et al. 2009). As a result, animate fillers paired with resumptive pronouns inside adjunct islands provide a closer match to an existing grammatical template, reducing structural strain and allowing partial convergence. In contrast, in the Inanimate–RP condition, no sufficiently similar grammatical template is available. Although a resumptive pronoun is present, inanimate fillers are poorly matched to the types of resumptive dependencies independently licensed by the grammar, preventing successful coercion. Consequently, structural strain remains high and acceptability remains low.

While the acceptability pattern in the Animate–RP condition is naturally captured by SOSP, the online processing data are not fully accounted for by this analysis. Under SOSP, coercion is expected to arise when an alternative, independently licensed structure is sufficiently compatible with the input. In such cases, increased reading times are interpreted as reflecting the additional processing effort involved in coercing the input into a partially compatible representation. However, increased reading times are also observed for animate fillers in gap conditions, where no such compatible structure is available and coercion is therefore not expected to occur.

3.2 Memory-based processing accounts

Memory-based accounts do not straightforwardly predict the observed pattern. These approaches attribute the unacceptability of island-violating sentences to working-memory limitations involved in maintaining and retrieving fillers during dependency formation (Kluender 1991; Kluender & Kutas 1993a; 1993b; Hofmeister & Sag 2010); on this view, animacy should increase the accessibility of the filler and thereby ease processing, since accessible fillers are expected to be easier to maintain and retrieve during dependency formation. Resumptive pronouns should further reduce processing demands by providing an overt integration site for the filler (McDaniel & Cowart 1999; Chacón 2019; Hammerly 2021).

Our results diverge from these predictions. If accessibility of filler phrases in memory were the driving factor, animate wh-fillers should speed processing at the integration site, regardless of tail type, yielding shorter RTs. Contrary to the predictions of memory-based accounts, our results reveal the opposite trend: animacy consistently slowed processing in both gap and RP conditions. This reversal is not just about the size of the effect—it’s about its direction. Rather than facilitating processing, animacy imposed a cost. Such a pattern is not predicted by the core assumption that animacy enhances retrieval by increasing accessibility. In fact, our findings suggest a misalignment between offline and online measures: animacy improves acceptability in the RP condition but slows processing across the board. This asymmetry is not straightforwardly predicted by memory-based accounts, which assume that higher accessibility should make retrieval easier, not harder.

Taken together, the self-paced reading results do not provide evidence for the facilitation predicted by memory-based accessibility accounts, as animacy was associated with longer reading times across conditions. Nevertheless, this pattern should be interpreted with caution, since task demands (the combination of self-paced reading and acceptability judgment) and the relatively high proportion of unacceptable sentences may have influenced readers’ processing strategies and made accessibility-based facilitation effects harder to detect in the self-paced reading data (see the Limitations section below for discussion).

An alternative interpretation of the self-paced reading results comes from similarity-based interference accounts of dependency formation, which predict increased processing cost when a dependency must be maintained or retrieved in the presence of featurally similar interveners (Jäger & Engelmann & Vasishth 2017). In the present materials, animate wh-fillers (e.g., meen ‘who’) introduce a discourse-prominent animate representation that must be maintained in the presence of other animate referents (e.g., second-person pronouns and lexical NPs in the adjunct clause such as the principal in 10), increasing similarity along features such as [+animate] and [+human].

From this perspective, the observed slowdown for animate fillers across both gap and resumptive-pronoun conditions follows naturally: increased reading times reflect interference arising from maintaining a salient animate representation amid competing animate discourse entities, rather than difficulty associated with dependency resolution per se. Importantly, this account also predicts the absence of a processing advantage for resumptive pronouns. In Arabic, bare wh-fillers are not morphologically specified for gender or number, whereas resumptive pronouns are marked for these features but do not encode animacy. Consequently, resumptive pronouns do not provide retrieval cues that distinguish animate from inanimate fillers under similarity-based accounts and are therefore not expected to facilitate online processing selectively for either type. If resumptive pronouns exert any effect under similarity-based interference, it is more likely to increase interference when their morphosyntactic features overlap with those of intervening constituents, rather than to reduce it.⁸

At the same time, this explanation leaves the acceptability pattern unaccounted for. While similarity-based interference predicts increased processing cost for animate fillers regardless of tail type, it does not explain why acceptability improves selectively in the Animate–RP condition. The resulting dissociation between online processing difficulty and offline acceptability suggests that additional interpretive or discourse-level factors contribute to RP-specific amelioration.

Because the present study did not independently manipulate the animacy, gender, or number features of intervening constituents, feature-based interference effects cannot be isolated directly. Accordingly, while interference-based accounts can plausibly explain the animacy-related processing cost observed here, they do not by themselves account for the selective offline acceptability improvement observed only in the RP condition. A more direct test of the interference-based explanation would require future experiments that independently manipulate intervener similarity with the wh-filler (e.g., animacy) and with the resumptive pronoun (e.g., gender or number).

3.3 A multivariate account

To account for these findings, we propose a multivariate account.⁹ Animate fillers (like who) are naturally more prominent because they refer to specific, individuated entities that are easy to anchor and track in the discourse, unlike inanimate fillers like what, which tend to denote kinds or categories (Alexopoulou & Keller 2013). When a resumptive pronoun appears, it can link to this salient filler and form an interpretable anaphoric dependency (Erteschik-Shir 1992; Frazier & Clifton 2002). That is, the resumptive pronoun functions as an anaphoric pronoun that is linked back to a highly salient referent, resulting in a more interpretable sentence. This explains the higher acceptability ratings observed in the animate–RP condition.

At the same time, the process is not cost-free. Constructing a discourse-based dependency requires additional effort, as the parser shifts away from syntactic dependency resolution toward an anaphoric, discourse-based interpretation (Alexopoulou 2010). This additional interpretive work is reflected in the slower reading times observed for animate fillers across both gap and RP conditions. In other words, animacy improves interpretability but increases processing demands, a trade-off that captures the mismatch between acceptability judgments and reaction times in our data.

Specifically, the parser treats who as a strongly individuated discourse entity that can naturally serve as the antecedent of an anaphoric pronoun. Even though grammar blocks gap-search inside islands (Stowe 1986; Frazier & d’Arcais 1989; Traxler & Pickering 1996), parsers may still anticipate a pronoun as a possible non-grammatical resolution strategy (Keshev & Meltzer-Asscher 2017). When no resumptive pronoun appears (gap condition), the parser is left with an unfulfilled prediction, forcing it to sustain unresolved search pressure, which slows processing and contributes to degraded acceptability. In what-questions, by contrast, no such pronoun search is triggered, since inanimates are less natural antecedents for anaphoric pronouns in anaphoric dependencies. Thus, while both animate and inanimate gaps remain unresolved, only who-gap involves an unmet anaphoric expectation, which explains the longer reading times.

The lack of processing differences between the animate–gap and animate–RP conditions can be understood in this light. In the animate–RP condition, the parser appears to consume memory resources in search of an antecedent, which increases processing load—just as it does with gaps. But unlike gaps, resumptive pronouns offer a partial discourse-level repair: they enable the parser to construct an anaphoric link back to the salient animate filler. This seems to support a modest rise in acceptability, even if the extra processing demands remain. Put differently, both tails in the animate condition increase processing costs—gaps because they leave anaphoric predictions unmet, and resumptives because they require additional linking operations. Yet only the latter offers a path to a more coherent interpretation.

This explanation aligns well with the eclectic model of island effects proposed by Chaves & Putnam (2020), which treats syntax, processing, and discourse as jointly shaping both acceptability and real-time comprehension. A purely syntactic account struggles to explain why animacy modestly improves RP sentences, and a memory-based account falls short of explaining why that improvement comes with a processing penalty.

Our findings suggest that animacy here acts as a discourse cue rather than a processing facilitator: it enhances interpretability without easing integration effort. The lack of a processing benefit for RPs further underscores their role as interpretive, not syntactic or processing devices in Saudi Arabic. They function like discourse pronouns that improve the interpretability of island-violating dependencies, but only when the filler is semantically rich enough to support such an interpretation. This perspective aligns with an anaphoric view of resumption, in which RPs behave more like referential pronouns than structural devices (Erteschik-Shir 1992; Frazier & Clifton 2002).

More broadly, the multivariate account developed here fits naturally within recent approaches to locality phenomena that treat acceptability and processing as the outcome of interacting syntactic, discourse, and processing constraints. In a large-scale experimental investigation of bridge effects in English, (Huang et al. 2025) show that bridge effects emerge from the interaction of morphosyntactic licensing, information-structural constraints, and processing costs, with no single factor proving sufficient; their multivariate model substantially outperforms single-predictor accounts. The present findings extend this conclusion to adjunct island violations in Arabic. While animacy and resumption interact to improve interpretability at the level of judgment, they do not alleviate—and may even increase—online processing cost. This dissociation supports a view in which discourse-level factors modulate interpretability without eliminating structural constraints or processing difficulty, reinforcing the need for a multivariate account of island effects.

3.4 Implications

3.5 Limitations

A growing body of work shows that reading behavior in self-paced reading tasks is sensitive to task demands, particularly to whether readers anticipate providing acceptability judgments or answering comprehension questions. Such task manipulations have been shown to influence processing strategies and grammaticality bias, thereby shaping online reading-time profiles (Hammerly & Staub & Dillon 2019; Laurinavichyute et al. 2023; Laurinavichyute & von der Malsburg 2024). In particular, Laurinavichyute & von der Malsburg (2024) argue that acceptability-based tasks reduce default assumptions of well-formedness and are associated with more conservative, analytic processing, in which readers remain sensitive to structural uncertainty during incremental parsing rather than relying on shallow heuristics that often suffice in comprehension-based paradigms.

Evidence for the impact of task configuration also comes from prior work using closely comparable materials in Baha Arabic. In that work, self-paced reading paired with comprehension questions yielded no significant effects of resumption, D-linking, or islandhood, whereas pairing the same paradigm with acceptability judgments revealed systematic and theoretically interpretable reading-time effects (Hayyas 2020). Together, these findings suggest that task demands can systematically modulate online processing strategies in island-violating dependencies.

A related limitation concerns the distribution of sentence types in the experimental materials. Participants were exposed to a relatively high proportion of unacceptable configurations, due both to the inclusion of island-violating test sentences and to the number of ungrammatical fillers. This distribution differs from that typically adopted in self-paced reading studies aimed at approximating naturalistic comprehension, where ungrammatical sentences are often kept below approximately 15–20%. Prior work shows that increasing the proportion of unacceptable material can alter participants’ expectations about sentence well-formedness and modulate online reading behavior (Hammerly et al. 2019; Laurinavichyute et al. 2023; Laurinavichyute & von der Malsburg 2024). Such input statistics may lead readers to adopt more cautious parsing strategies, yielding reading-time profiles that differ from those observed under more naturalistic distributions.

In light of these task-related and distributional considerations, the dissociation observed between online reading times and offline acceptability judgments in the experimental items suggests that reading times in the present study should not be treated as direct proxies for perceived sentence quality. Rather, they appear to index the processing effort involved in maintaining and interpreting island-violating wh-dependencies during incremental parsing. Importantly, this dissociation is not observed for filler items: reading times reliably distinguish grammatical from ungrammatical fillers and closely track acceptability judgments. The divergence in the experimental conditions therefore appears to reflect properties specific to island-violating dependencies, rather than a general limitation of the self-paced reading method under evaluative task demands.

Crucially, these task-related factors do not undermine the interpretation of the animacy-related slowdown reported here but instead help clarify how the task may have shaped the processing strategies through which this effect became visible. Because acceptability-based reading encourages readers to remain attentive to unresolved dependencies during incremental parsing, the longer reading times observed for animate fillers remain consistent with the interpretation that animacy increased the effort associated with maintaining and resolving an anaphoric dependency rather than facilitating retrieval.

Given that task demands and input distributions shape online processing strategies, future work employing alternative methods—such as eye-tracking, maze tasks, or forced-choice paradigms—would be valuable for assessing the robustness of animacy and resumption effects under conditions that more closely approximate incremental comprehension.