Glossa Psycholinguistics

2.1 Two different views on the timing of structural constraints

In Section 1, we adopted the widely held view that crossover is a structural constraint: regardless of its theoretical source, crossover effects depend on configurational relations among the operator, the trace, and the pronoun. Building on this assumption, this section examines how structural and non-structural constraints interact during sentence processing and how we probe the time-course of structural and agreement cues in crossover configurations.

During real-time pronoun resolution, the parser must establish a dependency between a pronoun and a potential antecedent. When multiple candidates are available, the parser relies on a range of constraints and cues to locate or retrieve an optimal candidate. For our purposes, the relevant constraints involved are the structural constraints (SCO and WCO constraints) and the gender agreement constraint. Although previous psycholinguistic work has shown that both structural and agreement information can guide antecedent retrieval (e.g., Badecker & Straub, 2002; Cunnings & Sturt, 2014; Felser & Cunnings, 2012; Sturt, 2003), there is no consensus on their relative timing during online processing. Below, we outline two different views. For ease of exposition, we illustrate their predictions using WCO configurations; both approaches make parallel predictions for SCO.

One prominent view holds that a wide variety of linguistic information, structural and non-structural alike, is available from the earliest stages of processing for antecedent retrieval (e.g., Badecker & Straub, 2002; Jäger et al., 2015; Kaiser et al., 2009; Lewis et al., 2006; Lewis & Vasishth, 2005; Patil et al., 2016; Runner et al., 2006). We refer to this class of approaches, irrespective of particular implementation details (e.g., cue-based memory retrieval), as the Uniform Multiple Constraints Hypothesis. Building on interactive models of sentence processing (e.g., Ferreira & Henderson, 1990; MacDonald et al., 1994; Tanenhaus et al., 1995; Trueswell et al., 1993, 1994), this hypothesis predicts that structural and agreement constraints are uniformly available and can influence pronoun resolution from the outset. As a conceptual starting point, we further assume that structural and non-structural constraints are similarly weighted during early-stage processing,³ where weight refers to relative influence during online processing and does not denote constraint ranking within the grammar. Therefore, the Uniform Multiple Constraints Hypothesis predicts that gender-congruent antecedents, regardless of their structural accessibility, are considered during the initial processing stages in crossover configurations.

In contrast, the Early Filter Hypothesis posits that only structurally accessible referents are initially considered as potential antecedents, while structurally inaccessible referents are excluded at early stages but may become available later (e.g., Chang et al., 2020; Chow et al., 2014; Cunnings et al., 2015; Cunnings & Sturt, 2014, 2018; Dillon et al., 2013; Sturt, 2003). On this view, structural constraints function as early, defeasible filters. In terms of timing, they are deployed immediately to restrict the candidate set of antecedents, based on configurational relations; but in terms of strength, their effects may weaken over time. Importantly, this prediction extends to all configurational constraints, including the crossover constraint and Condition C, whose ultimate theoretical status may be syntactic, semantic, or pragmatic (for a stricter interpretation in which only purely syntactic constraints are prioritized, see Drummer & Felser, 2018). What matters for the predictions of the Early Filter Hypothesis is that coreference depends on structural relations (e.g., c-command) between the pronoun and its potential antecedents (e.g., Cunnings et al., 2014; Kazanina et al., 2007; Kush et al., 2015, 2017; Phillips et al., 2011). Applied to crossover, the Early Filter Hypothesis predicts that wh-phrases should not influence early pronoun resolution due to their structural inaccessibility, irrespective of gender congruency.

To illustrate these predictions, consider the WCO configuration in (3). Upon encountering the pronoun her, the parser searches for a structurally accessible antecedent matching its agreement features. For ease of exposition, we represent structural accessibility by {+Access}, abstracting away from the specific crossover constraint in question (e.g., A-binding). Gender features are represented as {+/–Masc/Fem}, with +/– indicating feature match or mismatch. Furthermore, following standard terminology (e.g., Dillon et al., 2013; Sturt, 2003), we refer to the structurally accessible DP as the target and the structurally inaccessible DP as the distractor.

In (3), the target Sarah is retrieved, because it is the only structurally accessible DP matching the gender of her. Both hypotheses make similar predictions here: the distractor which man is an unlikely antecedent for her due to both gender mismatch and structural inaccessibility. The predictions diverge, however, when the distractor matches the pronoun’s gender, as in (4).

Under the Uniform Multiple Constraints Hypothesis, the gender-congruent distractor which man may be considered at early stages. Furthermore, in cue-based implementations (Lewis et al., 2006; Lewis & Vasishth, 2005),⁴ it additionally predicts slower retrieval in (4) than in (3) due to competition between Peter and which man. This slowdown – inhibitory interference – arises when morphological and structural cues have similar weights. By contrast, the Early Filter Hypothesis predicts no such effect: as only the target is retrieved during the early processing stages, antecedent retrieval speed should not differ between (3) and (4) at the onset of the pronoun.

Examples (3–4) involve cases where the target fully matches the pronoun’s gender. Now, consider cases where the target mismatches the pronoun, as in (5–6). In these cases, neither of the potential antecedents are optimal. However, the distractor which man matches the gender of the pronoun in (5), but not in (6). In this case, will the gender of the distractor affect pronoun resolution?

The Uniform Multiple Constraints Hypothesis again predicts that in (5), the gender-congruent distractor may be retrieved early. However, compared to (6), this manifests as a facilitatory interference (see Jäger et al., 2017, p. 319 for a similar understanding of the concept) in prior sentence processing studies, shown by faster reading times in (5) compared to (6). The distractor interference effect in (5) is facilitatory, rather than inhibitory, because the sentence contains a gender-congruent distractor which man, in contrast to (6) where neither referent is semantically plausible. The Early Filter Hypothesis predicts no early interference in either case, as the distractor is excluded by the crossover constraint regardless of gender agreement.

In summary, the Uniform Multiple Constraints Hypothesis posits that gender congruency between the pronoun and the distractor causes some form of interference – facilitatory or inhibitory – during the initial processing stages. We call this kind of interference gender intrusion effects. By contrast, the Early Filter Hypothesis predicts no intrusion from the distractor during the early stages, with processing difficulties only arising from gender mismatch between the pronoun and the target. It should be noted that these two approaches to pronoun resolution primarily differ in their predictions for early-stage processing, as gender intrusion frequently occurs at later processing stages, including post-reading judgments (e.g., Chow et al., 2014; Cunnings et al., 2015; Sturt, 2003). Therefore, in our self-paced reading study, we not only examine gender intrusion at the pronoun (early stage) and the spillover (later stages) regions, we also assess readers’ post-reading acceptability judgments (global stage), in order to get a comprehensive understanding of pronoun resolution in crossover configurations.

2.2 The processing of SCO and WCO configurations

As mentioned above, both hypotheses predict similar processing patterns for SCO and WCO configurations. However, some prior studies of crossover processing present results that do not fully align with this prediction. In an eye-tracking study in German, Felser and Drummer (2017) found no gender intrusion in the SCO configuration, but observed early signs of gender intrusion in WCO configurations, as reflected in first-fixation and first-pass reading times. However, this unexpected asymmetry – where gender intrusion is present in WCO, but not in SCO – should be interpreted with caution, because the WCO constraint might not apply in German (e.g., Grewendorf, 1988; Haider, 1993) and the strength of WCO effects appears to vary cross-linguistically due to different strength relationships between hierarchical prominence and linear order across languages, as suggested by Bresnan (1998).

A more substantial challenge to these hypotheses is presented by Kush et al. (2017). They tested whether and when structurally inaccessible antecedents are retrieved in SCO and WCO configurations in English, as in (7a-b), and reported findings similar to Felser and Drummer (2017). Using self-paced reading, Kush and colleagues found that in SCO configurations like (7a), where the structurally accessible target does not match the gender of the pronoun, no reliable gender intrusion was observed, as gender mismatch between the wh-phrase and the pronoun did not cause reading slowdowns. However, in WCO configurations like (7b), a facilitatory interference effect was observed at the pronoun, as gender match between the wh-operator and the pronoun led to faster reading times. To explain the different processing patterns observed in SCO and WCO configurations, Kush and colleagues suggest that the crossover constraint, which applies to both SCO and WCO, may be less weighted compared to Condition C, which only applies to SCO. Thus, Kush and colleagues account is essentially a weaker version of the early filter approach as they posit that only certain structural constraints (e.g., Condition C) are prioritized at the early stage of processing, while others (e.g., the crossover constraint) are not.

We recognize that different structural constraints may show variability in their weights during processing. However, we consider it premature to conclude that the crossover constraint is weak and does not constrain early-stage processing. Alternative explanations for Kush and colleagues’ findings are conceivable. In their stimuli, the wh-operator and the pronoun occur in different clauses: the operator appears in a higher clause, while the pronoun is embedded in a subordinate clause. Huang (1993) notes that reconstruction effects, which are linked to crossover effects, are attenuated when the pronoun occupies a more deeply embedded position, and subsequent experimental work has shown that both linear and structural distance between the wh-operator and the pronoun modulates reconstruction effects (e.g., Adger et al., 2017; Salzmann et al., 2023; Wierzba et al., 2021). Relatedly, independent observations also suggest WCO effects are weaker when the pronoun is more embedded structurally (e.g., Agüero-Bautista, 2012, 2014; Ishii, 2006).

Moreover, the relatively high structural complexity of Kush and colleagues’ test materials may have increased the overall processing load, potentially encouraging shallow parsing strategies (e.g., Ferreira et al., 2002; Sanford & Sturt, 2002). Under shallow processing, readers may construct underspecified representations and rely more heavily on salient surface cues, such as linearly proximal gender features of the distractor, rather than fully compute hierarchical relations. Further contributing to this potential bias, in their tested sentences, the structurally accessible target consistently mismatched the pronoun in gender, a design choice that could potentially bias participants towards bound variable readings. These factors may particularly affect WCO configurations relative to SCO, since the crossover constraint itself is relatively more violable than Condition C and thus more susceptible to distance-related effects. Taken together, it is worth exploring whether early gender intrusion effects would still arise when the wh-phrase and the pronoun occur within the same clause and the overall sentence structure is less complex, as in (3–6).

To summarize, the processing of pronouns in crossover configurations presents an ideal case for assessing contrasting views of real-time pronoun resolution. While previous studies have yielded valuable insights, there remains space for further exploration due to factors related to experimental design. In this study, we follow in the footsteps of prior work by examining the processing of SCO and WCO configurations, in order to further evaluate psycholinguistic models of pronoun resolution.

5.1 Participants

Ninety-four students over age 18 from the Stony Brook University community who spoke English as their first language participated in Experiment 1. The participants received course credit for participation. Studies reported in this article have been approved by the institutional review board of the authors’ institutions.

5.2 Materials

We follow prior psycholinguistic work (e.g., Cunnings et al., 2015; Sturt 2003) by manipulating the gender congruency between the pronoun and both the structurally accessible target (target match/mismatch) and the structurally inaccessible distractor (distractor match/mismatch). This yields a 2 × 2 factorial design. See (8a-d) for an example. The subscripts indicate presentation regions in self-paced reading.

We kept the distractor phrase (e.g., which man) preceding the pronoun constant across conditions and varied the gender of the target and the pronoun. We also fully counterbalanced the use of gendered proper names and pronouns for each condition (i.e., half of the items contain he; half contain she). This gender counterbalancing across conditions ensures that the four conditions are comparable to each other regarding the use of gendered elements and helps minimize effects unrelated to our purposes (e.g., she differs from he in alphabetic length and frequency).

In the self-paced reading experiment, reading slowdowns due to gender mismatch between the pronoun and either the target or the distractor – referred to as the gender mismatch effect – are taken as evidence that readers attempt to link these two elements. The gender mismatch effect is often used as a diagnostic for detecting binding in anaphora resolution (e.g., Chow et al., 2014; Kush et al., 2017; Patil et al., 2016; Sturt, 2003). Thus, the gender mismatch effect allows for the examination of the accessibility of structural and gender cues during real-time anaphora resolution. Furthermore, we anticipate a main effect of Distractor or a Target × Distractor interaction at the earliest moments (e.g., pronoun region), if gender information is used immediately for pronoun resolution; if the structural constraint applies as an early filter, we should only observe a main effect of Target.

Twenty-four sets of target sentences, as in (8), were mixed with 48 filler sentences and randomly presented to the participants in a Latin Square design. Of the 48 fillers, 24 were syntactically and semantically/pragmatically well-formed; the other 24 fillers were composed of 12 syntactically and 12 semantically/pragmatically anomalous sentences. The purpose of including well-formed and ill-formed fillers is to check whether participants paid attention to the experiment. All items in this study are available on the Open Science Framework, linked in the Data Accessibility Statement.

5.3 Procedure

The experiment was hosted on PCIbex (Zehr & Schwartz, 2018), and consisted of a self-paced reading task and an acceptability judgment task. The participants first read the sentences region by region in a moving window paradigm. After reading each sentence, they judged how acceptable the sentence was on a 7-point Likert scale. Following Kush et al. (2017), the leftmost and rightmost ends of the scale were labeled “Bad” and “Good,” respectively. The participants could either click on the radio button associated with a numerical value or press a number on the keyboard to indicate the acceptability of a sentence. Before the main experiment, participants were familiarized with the task by means of 4 practice trials.

5.4 Data analysis

Before conducting data analysis, we removed data from participants who exhibited a higher mean rating for ill-formed fillers than for well-formed fillers, consistently selected a single number to indicate the acceptability of the sentence, or demonstrated overly rapid reading times (i.e., below 100 ms on average). Applying these criteria resulted in the exclusion of 9 participants, which left 85 participants for analysis.

For the analysis of acceptability ratings, we scaled the raw ratings of target items using z-transformation by participant. For the analysis of reading times (RTs), we first removed outliers falling outside the range of 100–4000 ms and 2.5 standard deviations beyond the mean by region and condition, which affected 2.99% of the original data. We applied a common two-step data screening procedure (see the review by Marsden et al., 2018), because while absolute cutoffs remove implausible values, SD-based trimming targets distribution-specific outliers that may still fall within absolute bounds, but exert disproportionate influence on condition means. Additionally, raw RTs were log-transformed before data analysis.

Mixed-effects linear regression was used to analyze the ratings and RTs with the lmer function in the lme4 package in R (R Core Team, 2025). The main effects were the two factors, Target and Distractor, which were sum-coded (Target-match: 0.5; Target-mismatch: –0.5; Distractor-match: 0.5; Distractor-mismatch: –0.5). The random effects were the slopes and intercepts grouped by participants and items. When fitting statistical models, we followed Bates et al. (2015) by selecting more economical models. Therefore, we report results from a simpler model, rather than a more complex model, if the two models do not differ significantly. For model comparisons, we adopted an alpha level of 0.2 (following Matuschek et al., 2017), considering that an alpha level of 0.05 may impose an excessive penalty on model complexity.

5.5 Results

5.5.1 Acceptability ratings

Figure 1 shows the 85 participants’ mean acceptability of target sentences by condition in both raw and z-transformed scores. The mean ratings for well-formed and ill-formed fillers (represented by the two dashed lines in the left panel) are 5.78 and 2.59, respectively. As shown in Figure 1, the target-match conditions exhibit higher acceptability relative to target-mismatch conditions, whereas no prominent difference is observed between distractor-match and distractor-mismatch conditions.

Figure 1: Mean acceptability of target sentences by condition in raw and z-transformed scores in SCO.

Statistical analysis on z-transformed scores, summarized in Table 1, confirms a main effect of Target (p < 0.001). There is no main effect of Distractor: gender congruency between pronouns and distractors does not modulate acceptability ratings. The Target × Distractor interaction is not significant, either. These results, consistent with the findings of Kush et al. (2017), suggest that native English speakers do not consider structurally inaccessible antecedents in their global, post-sentence interpretations.

Table 1: Summary of statistics for z-transformed ratings of target sentences in SCO (*: < 0.05).

	β	SE	t	p
Target	0.62	0.04	15.41	< 0.001*
Distractor	0.004	0.04	0.12	0.90
Target × Distractor	0.01	0.08	0.14	0.89

Model structure: lmer (Score ~ Target * Distractor + (1|Item) + (1|Participant), data)

5.5.2 Reading times

The mean RTs by region and condition are displayed in Figure 2. See Table 2 for statistics. Prior to the critical pronoun region, no effects are significant. At the critical region, we found a main effect of Target (p = 0.03): gender mismatch between the structurally accessible Target and the pronoun leads to reading slowdowns. As mentioned earlier, such gender mismatch effects act as a diagnostic suggesting that participants attempted to retrieve the matrix subject during processing. In contrast, gender mismatch between the structurally inaccessible distractor (e.g., which man) and the pronoun does not significantly impact reading times, as evidenced by the absence of a main effect of Distractor. The Target × Distractor interaction is not significant, either. Thus, at the critical region, we only see an early effect of Target gender mismatch, but no distractor-related gender intrusion effects.

Figure 2: Mean RTs (ms) by region and condition in SCO.

Table 2: Summary of statistics for log-transformed RTs in SCO (*: < 0.05).

	Target				Distractor				Target × Distractor
Region	β	SE	t	p	β	SE	t	p	β	SE	t	p
Peter/Sarah	–0.01	0.01	–0.85	0.39	0.003	0.01	0.26	0.79	0.02	0.02	0.63	0.53
told	–0.008	0.01	–0.66	0.51	0.01	0.02	0.65	0.52	0.02	0.03	0.75	0.45
us	–0.002	0.01	–0.20	0.84	0.004	0.01	0.31	0.76	0.03	0.02	1.12	0.26
which man	0.007	0.02	0.41	0.68	–0.02	0.02	–0.92	0.37	–0.01	0.03	–0.44	0.66
in	0.006	0.01	0.43	0.67	–0.007	0.01	–0.56	0.58	–0.007	0.02	–0.29	0.77
the office	–0.001	0.01	–0.06	0.95	0.006	0.01	0.46	0.65	0.03	0.03	0.98	0.33
he/she	–0.03	0.01	–2.23	0.03*	0.002	0.01	0.13	0.89	0.01	0.02	0.51	0.61
had	–0.07	0.01	–6.69	< 0.001*	0.005	0.01	0.44	0.66	–0.003	0.02	–0.13	0.90
promoted	–0.03	0.01	–2.61	0.009*	0.01	0.01	1.15	0.25	–0.007	0.02	–0.34	0.74
recently	–0.03	0.03	–1.02	0.32	–0.01	0.02	–0.60	0.55	–0.06	0.03	–1.77	0.08

Model structures for critical and post-critical regions:

Pronoun region: lmer (logRT ~ Distractor * Target + (1 | Item) + (1 + Target | Subject), data)

1^st spillover region: lmer (logRT ~ Distractor * Target + (1 | Item) + (1 | Subject), data)

2^nd spillover region: lmer (logRT ~ Distractor * Target + (1 | Item) + (1 + Distractor | Subject), data)

Final region: lmer (logRT ~ Distractor * Target + (1 + Target | Item) + (1 + Target | Subject),

For the spillover regions following the pronoun, we found main effects of Target at the 1^st (e.g., had) and 2^nd (e.g., promoted) spillover regions (ps < 0.01) but no main effects of Distractor. This strengthens the finding from the critical region: the parser only attends to structurally accessible antecedents at initial stages of processing SCO configurations. We find no main effect of Distractor and no significant Target × Distractor interaction at any region. These findings show that the structural constraints for SCO (crossover constraint + Condition C) strictly guide antecedent retrieval during all processing stages.

5.6 Discussion

Experiment 1 employs self-paced reading and acceptability judgments to probe whether structurally inaccessible distractors influence pronoun resolution in the SCO configuration, where both the crossover constraint and Condition C are violated. To this end, we manipulated gender congruency between the pronoun and the candidate antecedents (target and distractor), in order to observe whether the gender of the distractor impacts the acceptability ratings and reading times.

In the post-reading judgment task, we see no interference from the distractor (the wh-phrase). Only the gender of the structurally accessible target modulates acceptability ratings, as gender mismatch leads to significantly lower acceptance. This suggests that in an SCO configuration, native English speakers do not consider distractors in their offline judgments. This finding replicates the results of Kush et al. (2017), who similarly observed no gender intrusion from the distractor in SCO. We note that our conclusion is restricted to wh-phrases that are thematic arguments and is not intended to generalize to wh-phrases that appear inside arguments (e.g. Which picture of Mary_i did she_*?i/j like?), where coreference can be marginally acceptable to some speakers (e.g., Adger et al., 2017; Bruening & Al Khalaf, 2019; Salzmann et al., 2023).

In the self-paced reading task, where participants read sentences incrementally, we see similar results as in the acceptability judgment task. Specifically, as soon as the pronoun was encountered, the participants showed reading slowdowns in the Target-mismatch conditions relative to Target-match conditions, constituting evidence that targets were considered for retrieval. The absence of gender effects related to the distractor at any processing stage (early or late) in any form (facilitatory or inhibitory) matches the findings of the acceptability judgment task that the structural constraints involved in the SCO configuration strictly guide pronoun resolution. Thus, the reading time results do not align with the Uniform Multiple Constraints Hypothesis, but instead correspond to the predictions of the Early Filter Hypothesis.

Taken together, Experiment 1 replicates the grammatical intuitions and experimental findings from previous studies regarding the unavailability of bound pronoun readings in SCO configurations (e.g., Kush et al., 2017; Wasow, 1972). Furthermore, it presents reading time data that support the structurally constrained approach (e.g., Chow et al., 2014; Cunnings & Sturt, 2014; Cunnings et al., 2015; Dillon et al., 2013). We now move on to our other research questions: (i) How acceptable are bound pronoun readings in WCO sentences when the wh-operators are D-linked? (ii) Does the crossover constraint apply as an early filter in WCO configurations? To address these questions, we turn to Experiment 2.

6.1 Participants

Ninety-eight native English speakers over age 18 from the Stony Brook University community participated in exchange for course credit. None had participated in Experiment 1.

6.2 Materials

Experiment 2 adopts a similar design to Experiment 1, fully crossing the factors Target (target match/mismatch) and Distractor (distractor match/mismatch) in a 2 × 2 factorial design. The materials used for Experiment 2 were minimally adapted from those in Experiment 1, with a complex NP as the embedded subject (e.g., his supervisor) to create a week crossover configuration, and the matrix predicate (e.g., wondered) presented as one region without any object to keep the number of regions in Experiments 1 and 2 identical. Four distinct matrix predicates (wondered, was curious, asked, knew) were used and evenly distributed among the 24 target items. See (9) for an example set.

As in Experiment 1, the gender of proper names and pronouns was fully balanced for each condition. The 24 sets of target sentences were mixed with the same 48 fillers from Experiment 1 and randomly distributed to the participants using a Latin Square design.

6.3 Procedure

The procedure for Experiment 2 was identical to Experiment 1.

6.4 Data analysis

The same screening procedure described in Section 5.4 was followed. Participants who did not pay close attention to the experiment task (assigning higher scores to ill-formed fillers than to well-formed fillers, giving identical scores to all sentences, or whose mean RTs were below 100 ms) were removed prior to analysis. This step removed 9 participants, leaving 89 participants for final analysis. In addition, for the analysis of reading times, data points outside the range of 100–4000 ms and 2.5 standard deviations above the mean by region and condition were also removed, which affected 3.44% of the original data. Mixed-effects linear regressions were used for data analyses.

6.5 Results

6.5.1 Acceptability judgments

Figure 3 shows the 89 participants’ mean acceptability of target sentences by condition in raw and z-scores. The acceptability ratings for well-formed and ill-formed fillers (represented by the two dashed lines in the left panel) are 5.87 and 2.86, respectively. See Table 3 for statistics.

Figure 3: Mean acceptability of target sentences by condition in raw and z-transformed scores in WCO.

Table 3: Summary of statistics on z-transformed ratings of target items in WCO (*: < 0.05).

	β	SE	t	p
Target	0.31	0.04	7.58	< 0.001*
Distractor	0.01	0.04	0.24	0.81
Target × Distractor	0.22	0.08	2.73	0.006*

Model structure: lmer (Score ~ Target * Distractor + (1|Item) + (1|Participant), data)

The mixed-effect linear model over z-transformed scores reveals a main effect of Target (p < 0.001), as Target-mismatch conditions were less acceptable than Target-match conditions, which suggests that the participants often treated the target and the pronoun as coreferential. There is no main effect of Distractor, which means that the structurally inaccessible distractor was not consistently considered in post-reading judgments.

However, the Target × Distractor interaction is significant (p = 0.006), demonstrating the influence of the distractor. Despite the modest size of the gender intrusion effects, planned comparisons indicate that, unexpectedly, distractor gender mismatch made the sentence significantly less acceptable within the Target-match conditions (β = 0.12, SE = 0.06, t = 2.11, p = 0.04), whereas distractor gender match marginally lowered the acceptability of the sentence within the Target-mismatch conditions (β = –0.17, SE = 0.09, t = –1.85, p = 0.07) (see similar processing patterns in, e.g., Cunnings & Felser, 2013; Cunnings & Sturt, 2014; Cunnings et al., 2015; Sturt, 2003; and see Engelmann et al., 2019, for discussion of such reversed intrusion effects).

Focusing on the two conditions that speak to our theoretical interest – conditions with gender-matching antecedents in structurally accessible and inaccessible positions (Target-match/Distractor-mismatch in (9b) and Target-mismatch/Distractor-match in (9c)) – we see that violating the crossover constraint incurs a cost (β = –0.68, SE = 0.01, t = –5.82, p < 0.001), even though the wh-phrase is D-linked. This is not predicted by grammar-based D-linking accounts, which rule in bound pronoun readings in WCO (e.g., Falco, 2010; Ishii, 2006).

Thus, unlike the results for the SCO configuration, distractors in the WCO configuration impact the final stages of pronoun resolution. However, we note that the distractors only play a minor role relative to the targets. The primary significance of the main effect of Target, especially given the contrast between (9b) and (9c), crucially suggests that the WCO effect cannot be reduced entirely to processing factors.

6.5.2 Reading times

Figure 4 displays the participants’ mean RTs by region and condition. See Table 4 for statistics. No effect reaches significance prior to the critical pronoun region. At the pronoun (e.g. his/her), a main effect of Target emerges (p = 0.009), as gender mismatch between the target and the pronoun leads to reading slowdowns. By contrast, the gender of the distractor does not significantly modulate RTs. The interaction is not significant, either. This means that upon encountering the pronoun, the parser only retrieves the structurally accessible antecedent. No gender intrusion effects appear at this early stage, which is more in line with the predictions of the Early Filter Hypothesis.

Figure 4: Mean RTs (ms) by region and condition in WCO.

Table 4: Summary of statistics for log-transformed RTs in WCO (*: < 0.05).

	Target				Distractor				Target × Distractor
Region	β	SE	t	p	β	SE	t	p	β	SE	t	p
Peter/Sarah	–0.02	0.01	–1.32	0.19	–0.01	0.01	–1.24	0.21	–0.01	0.02	–0.43	0.67
wondered	–0.01	0.01	–1.22	0.22	0.01	0.01	0.89	0.37	0.004	0.02	0.17	0.87
which man	–0.006	0.02	–0.39	0.70	0.02	0.02	1.49	0.14	0.02	0.03	0.55	0.58
in	–0.01	0.01	–1.21	0.23	–0.006	0.01	–0.55	0.58	–0.0003	0.02	–0.01	0.99
the office	–0.009	0.01	–0.74	0.46	–0.006	0.01	–0.48	0.63	–0.01	0.03	–0.43	0.67
his/her	–0.03	0.01	–2.61	0.009*	–0.01	0.01	–1.26	0.21	–0.003	0.02	–0.12	0.91
supervisor	–0.07	0.01	–4.46	< 0.001*	–0.009	0.01	–0.70	0.49	–0.03	0.02	–1.26	0.21
had	–0.02	0.01	–1.69	0.09	-0.006	0.01	–0.59	0.56	–0.02	0.02	–0.89	0.38
promoted	–0.02	0.01	–1.54	0.13	0.01	0.01	1.10	0.27	–0.05	0.02	–2.37	0.02*
recently	0.06	0.02	3.20	0.001*	–0.03	0.02	–1.28	0.21	–0.05	0.03	–1.33	0.18

Model structures for critical and post-critical regions:

Pronoun region: lmer (logRT ~ Distractor * Target + (1|Item) + (1|Subject), data)

1^st spillover region: lmer (logRT ~ Distractor * Target + (1|Item) + (1 + Target|Subject), data)

2^nd spillover region: lmer (logRT ~ Distractor * Target + (1|Item) + (1|Subject), data)

3^rd spillover region: lmer (logRT ~ Distractor * Target + (1|Item) + (1 + Target|Subject), data)

Final region: lmer (logRT ~ Distractor * Target + (1 + Distractor|Item) + (1 + Distractor|Subject), data)

At the 1^st spillover region (e.g., supervisor), we found similar results: only the main effect of Target is significant (p < .001). There is no main effect of Distractor and no significant Target × Distractor interaction. However, visual inspection suggests that within the Target-match conditions, gender-congruent distractors are associated with faster reading times than incongruent ones, at least numerically. However, this numerical trend is not significant. Thus, the gender of the distractor may be accessible to the parser to a slight degree, but only at a later processing stage.

The 2^nd spillover region (e.g., had) does not show any significant effect. The 3^rd spillover region (e.g., promoted) does not show any main effect, but shows a significant Target × Distractor interaction (p = 0.02). Planned comparisons revealed a gender intrusion effect within the Target-mismatch conditions: distractor gender match leads to longer RTs (β = 0.04, SE = 0.02, t = 2.38, p = .02), indicating processing difficulty. This inhibitory interference effect within the Target-mismatch conditions – contrary to the facilitatory interference effect predicted by the Uniform Multiple Constraints Hypothesis – has been identified in some prior studies (e.g., Chow et al., 2014; Cunnings et al., 2015; Jäger et al., 2015; Sturt, 2003) and has been taken to suggest that “readers sometimes attempted to coerce an interpretation in which the pronoun was linked to a syntactically illicit but gender matching antecedent” (Cunnings et al. 2015, p. 15).

At the final region, there is only a main effect of Target (p = 0.001).⁶ This main effect is due to overall reading slowdowns in the Target-match conditions, which seems different from the main effects of Target in preceding regions. However, visual inspection suggests that this unexpected reversed main effect of the target is due to a distractor-related gender intrusion effect within the Target-match conditions. Planned comparisons support this observation, indicating that gender-congruent distractors lead to faster RTs than gender-incongruent distractors within the Target-match conditions (β = –0.05, SE = 0.02, t = –2.20, p = .03), but not within the Target-mismatch conditions. This facilitatory interference effect within the Target-match conditions is consistent with the similar pattern discovered at the 1^st spillover region.

6.6 Discussion

Experiment 2 on WCO aims to examine whether the gender of a structurally inaccessible but D-linked wh-phrase modulates real-time pronoun resolution. Recall that previous accounts make divergent predictions regarding these two questions. On the grammatical side, prior crossover theories present divergent predictions, including full ungrammaticality (e.g., Koopman & Sportiche, 1983; Reinhart, 1983), partial ungrammaticality (e.g., Wasow & Clausen, 2011), and full acceptability (e.g., Ishii, 2006). On the processing side, the Uniform Multiple Constraints Hypothesis predicts early gender intrusion effects, while the Early Filter Hypothesis predicts no gender intrusion from the structurally inaccessible distractor.

First, consider the results of the post-reading acceptability judgment task, which speak to the theoretical question regarding specificity/D-linking and the crossover constraint. A key finding of our study is that native English speakers are sensitive to the crossover constraint, regardless of the specificity of the wh-phrase. WCO configurations are rated less acceptable when the gender of the target mismatches the gender of the pronoun, which indicates that native speakers attempted to link the pronoun to the target, adhering to the crossover constraint. Comparing sentences like (9b) and (9c) with semantically plausible antecedents in structurally accessible and inaccessible positions, we see that Target-match sentences were judged to be significantly more acceptable than Distractor-match sentences, even though the wh-phrase (distractor) is linearly closer to the pronoun (and therefore less subject to memory decay). We interpret this as evidence that structural accessibility outweighs both specificity and linear distance in offline judgments. While these results do not adjudicate between syntactic and non-syntactic explanations of the crossover constraint and our work does not aim to do so, they pose a challenge for specificity-based accounts that predict full acceptability under D-linking (e.g., Falco, 2010; Ishii, 2006).

That said, unlike our Experiment 1 on SCO, our findings regarding WCO point to the possibility that the distractor is considered at later stages of comprehension, as evidenced by the lower ratings of distractor mismatch within the Target-match conditions.⁷ On one account, this may stem from some participants considering the distractor to be the binder of the pronoun due to the specific syntactic representations they have constructed (e.g., the wh-phrase first moves to an A-position, in line with Ishii, 2006). On another account, this may be due to processing reasons (e.g., Culicover, 2013; Wasow & Clausen, 2011) as the wh-phrase is linearly closer to the pronoun. We refrain from choosing between these two possibilities and simply note that accounts excluding the role of the crossover constraint cannot adequately explain the main effect of Target. It is plausible that a more explanatorily satisfactory theory would incorporate both syntactic and processing factors, as suggested by Wasow and Clausen (2011).

Next, we turn to the reading times in the self-paced reading task, which pertain to psycholinguistic models. As explained in Section 2, the Uniform Multiple Constraints Hypothesis predicts gender intrusion effects related to the distractor: a facilitatory interference effect in the Target-mismatch conditions and an inhibitory interference effect in the Target-match conditions. While we did observe gender intrusion effects in RTs in Experiment 2, the results are inconsistent with the predictions of the Uniform Multiple Constraints Hypothesis for two reasons. First, gender intrusion only reliably occurs at the final wrap-up region and numerically at the 1^st spillover region. Second, the direction of the gender intrusion effects is opposite to the predictions. We observed facilitatory interference in the Target-match conditions and inhibitory interference in the Target-mismatch conditions (see, e.g., Chow et al., 2014; Cunnings et al., 2015; Cunnings & Felser, 2013; Sturt, 2003, for similar findings). Furthermore, the RTs reported in Experiment 2 differ from Kush et al.’s (2017) findings and thus do not align with a weaker version of the Early Filter Hypothesis mentioned previously in Section 2, which posits that the crossover constraint is less weighted during initial processing compared to other structural constraints (e.g., Condition C). This variant version is not supported by our study due to the late emergence of gender intrusion effects. Instead, our findings are more consistent with the standard Early Filter Hypothesis, which views configurational constraints as early filters to pronoun resolution, regardless of the ultimate theoretical source of the constraint. Therefore, regardless of the type of constraint involved, WCO or SCO, our experimental findings suggest that native English speakers only consider structurally accessible referents during initial stages of processing.

7.1 Summary of results

The two experiments reported in this article aim to achieve two goals. Our primary goal is to examine two competing views of pronoun resolution: while the Uniform Multiple Constraints Hypothesis predicts early gender intrusion effects during pronoun resolution, the Early Filter Hypothesis predicts that structurally inaccessible antecedents should not be considered at early stages of processing. To adjudicate between these views, we examined how gender congruency between pronouns and potential antecedents affects offline acceptability judgments and online processing in SCO and WCO configurations. A secondary goal is to assess the acceptability of bound pronoun readings in WCO configurations, thus contributing to theoretical discussions concerning D-linking and the crossover effect.

In Experiment 1 on SCO configurations, we discovered clear evidence that structurally inaccessible distractors (i.e., wh-phrases) are not considered as antecedents in either incremental processing or in global interpretation. In post-reading acceptability judgments, our participants only showed sensitivity to the gender of the structurally accessible target, since gender mismatch between the pronoun and the target led to lower acceptability ratings. No interference from the distractor was observed. Reading times from self-paced reading yielded similar results: a gender mismatch effect, serving as a diagnostic for detecting binding attempts in processing, emerged immediately at the pronoun when the target mismatched in gender, but no such effect was observed for the distractor at any region. Together, these results replicate previous findings obtained with different experimental designs/paradigms (Felser & Drummer, 2017; Kush et al., 2017) and indicate that pronoun resolution in SCO configurations is strictly constrained by structural principles (Condition C plus the crossover constraint). Experiment 1 therefore establishes a clear baseline against which the processing of WCO configurations in Experiment 2 can be evaluated.

In Experiment 2 on WCO configurations, we found a more nuanced pattern compared to Experiment 1 on SCO. In the acceptability judgment task, participants primarily linked the pronoun to the structurally accessible target, as indicated by higher ratings for Target-match than Target-mismatch conditions. However, unlike in SCO, the distractor also had an (albeit limited) influence: within Target-match conditions, gender mismatch with the distractor led to lower acceptability. This pattern suggests that the crossover constraint is weaker in WCO than in SCO, where it is reinforced by Condition C. Nevertheless, the crossover effect was observed in WCO, even with D-linked wh-phrases, which does not align with accounts predicting full grammaticality for WCO configurations under D-linking (e.g., Falco, 2010; Ishii, 2006).

Pertaining to the evaluation of sentence processing models, the results of the current study are more in line with the Early Filter Hypothesis than the Uniform Multiple Constraints Hypothesis. As explained in previous sections, the latter predicts early gender intrusion, while the former predicts the absence of such intrusion. As we saw in Experiment 2, the gender intrusion effects during incremental processing are absent at the critical pronoun region, which lends support to the Early Filter Hypothesis. Gender intrusion only emerges at later spillover regions, and the direction of the gender intrusion may present further challenge to the Uniform Multiple Constraints Hypothesis, since our finding (i.e., facilitatory interference in the Target-match conditions and inhibitory interference in the Target-mismatch conditions) is the opposite to its predictions. By contrast, delayed gender intrusion effects do not pose a challenge to the structurally-constrained approach because structural constraints can be ‘defeasible’ at later stages (e.g., Sturt, 2003).

To summarize, the findings from the current study suggest that the crossover constraint is arguably not reducible to processing factors alone and that it strictly guides antecedent retrieval at the early processing stages in WCO configurations. Moving forward, we discuss some broader implications of our study for understanding the similarities and differences between the SCO and WCO configurations. Given our priority, we first focus on models of sentence processing and potential alternative explanations for our results, before turning to linguistic discussions.

7.2 The structural primacy effect and the processing of crossover configurations

One major finding of our study is that the crossover constraint applies early in incremental pronoun resolution, albeit being defeasible during later processing stages. In this section, we focus on two patterns that have implications for understanding pronoun resolution: (i) the absence of gender intrusion at the pronoun region and (ii) the contrastive processing profiles of SCO and WCO configurations.

We begin with the absence of gender intrusion at the pronoun region in real-time processing. The overall reading-time patterns observed in our experiments closely mirror those reported in a large body of prior work investigating the accessibility of structurally accessible versus inaccessible antecedents (e.g., Chow et al., 2014; Cunnings & Sturt, 2014; Cunnings et al., 2015; Dillon et al., 2013; Felser et al., 2009; Kush et al., 2015; Sturt, 2003). These studies converge on the conclusion that, at early stages of processing, structural cues are prioritized, or at least heavily weighted, relative to non-structural cues (e.g., Van Dyke & McElree, 2011). This pattern is commonly taken to support a strong interpretation of the Early Filter Hypothesis. Our findings are consistent with this view, suggesting that the crossover constraint, like other configurational constraints, is operative early on and can effectively filter out structurally inaccessible antecedents during the initial stages of pronoun resolution.

At the same time, our results differ from those of Kush et al. (2017), who reported early gender intrusion in WCO configurations. As discussed earlier, one plausible source of this discrepancy lies in differences in structural complexity between the sentence structures tested. In our stimuli, the wh-phrase and the pronoun occur within the same clause, whereas in Kush and colleagues’ study, the pronoun is more embedded in a subordinate clause. It is well established that increased linear and structural distance between dependent elements increases processing demands. Under such conditions, comprehenders may be more likely to misrepresent the syntactic position of the wh-phrase (e.g., as in an A-position) or to establish ungrammatical anaphoric dependencies as a way of alleviating processing load (see, e.g., Chacón, 2019). From this perspective, early gender intrusion effects in more complex structures may reflect processing pressure rather than the early accessibility of non-structural constraints.

Despite the robust structural primacy effect observed in our study, we caution against overgeneralizing our findings to suggest that all structural constraints are invariably prioritized, regardless of the specific competing constraints and the language under consideration. In fact, crosslinguistic evidence suggests that discourse-level factors, such as discourse topicality and logophoric prominence, can influence pronoun resolution during early stages of processing, especially in languages (e.g., Chinese, Korean) and structures (e.g., logophoric contexts) where discourse prominence plays a crucial role (e.g., Han et al., 2015; Jäger et al., 2015; Kaiser et al., 2009; Lyu & Kaiser, 2021, 2024; Sloggett, 2017). Therefore, our conclusions regarding the structural primacy effect during incremental processing are restricted to the specific constraint of crossover in English. We also note that the structural primacy effect should be understood as a relative, rather than an absolute, notion; our results only suggest that the target is retrieved earlier relative to the distractor at the point when the pronoun is encountered. As self-paced reading methods do not provide distinct measures of early-stage versus late-stage processing, further research using eye-tracking methods is needed before stronger conclusions can be drawn.

Next, we turn our attention to the comparison between the processing patterns of SCO and WCO, which speaks to how we understand different approaches to pronoun resolution. Recall that some previous studies found no gender intrusion effects from structurally inaccessible distractors in SCO, but clear evidence of gender intrusion in WCO (Felser & Drummer, 2017; Kush et al., 2017). These results align more closely with a weaker interpretation of the Early Filter Hypothesis which suggests that only certain structural constraints are applied early on and are given precedence over non-structural constraints. As mentioned above, we are open to this nuanced position, particularly when discourse-based constraints are at play, but our findings seem to indicate that the crossover constraint does not fall into the category of “weak” structural constraints. Instead, given similar processing patterns for SCO and WCO at the pronoun region in our study, it appears that the crossover constraint acts as a robust early filter in pronoun resolution.

How, then, do we make sense of the different processing patterns of SCO and WCO at the later stages, and how can our results be reconciled with previous studies that present different findings? We hypothesize that although individual structural constraints may be overridden at later stages, the cumulative effect of multiple constraints can mitigate against distractor interference at later stages. Indeed, other structural constraints examined in isolation indicate similar delayed gender intrusion effects in relation to Condition A (e.g., Cunnings & Sturt, 2014; Sturt, 2003), Condition B (e.g., Cunnings & Sturt, 2018), and even Condition C (e.g., Kazanina et al., 2007), which contrast with our findings for SCO configurations. Viewed from this perspective, the different processing patterns of SCO and WCO are not necessarily related to the differences between the crossover constraint and Condition C, as suggested by Kush et al. (2017), but rather to the number of structural constraints at work. Additionally, pragmatic factors such as discourse plausibility (see more in Section 7.4) may also play a disproportional role in WCO configurations relative to SCO configurations during the later/global stages of processing (e.g., Gor, 2020; Gor & Syrett, 2019).

7.3 Potential alternative explanations for the structural primacy effect

We now consider four alternative explanations for the structural primacy effect observed during the early-stage processing of crossover configurations. While some of these accounts can capture certain aspects of our results, they do not cover the full range of facts; other alternative accounts remain viable in principle, but would require additional empirical support to offer a strong alternative to a structurally constrained account.

The first alternative account relates to the potential preference for coreferential readings. As the target referent in crossover configurations is a referential DP, whereas the distractor is a wh-operator, one potential alternative explanation is that the target advantage reflects a general bias for coreferential over bound variable readings. It is possible that bound variable readings may incur greater processing costs and that this processing burden, instead of the crossover constraint, leads to the absence of gender intrusion by the distractor.

However, this explanation is hard to reconcile with existing psycholinguistic evidence for two reasons. First, when both bound variable and coreferential options are available, native speakers often prefer bound variable readings (e.g., Cunnings et al., 2014; Frazier & Clifton, 2000; Koornneef, 2008; Koornneef et al., 2006, 2011; Puebla & Felser, 2024). This pattern challenges the view that bound variable readings are generally disfavored due to higher processing costs. Second, pronoun resolution is strongly influenced by linear proximity (e.g., Cunnings et al., 2014; Omaki et al., 2019). In our materials, the distractor (wh-phrase) is linearly closer to the pronoun. If structural constraints were not operative at early stages, the bound variable reading should enjoy a clear processing advantage. However, contrary to this prediction, we observed no gender intrusion from the distractor during the initial stages of processing, suggesting that the structural primacy effect is plausibly related to the crossover constraint.

The second alternative invokes the parallel function constraint (e.g., Chambers & Smyth, 1998; Sheldon, 1974; Smyth, 1994): could the preference we observed for the (subject) target stem from a bias toward antecedents that share the same grammatical function as the (subject) pronoun, rather than the operation of the crossover constraint per se? This account seems untenable when applied to WCO sentences. While the parallel function constraint is plausibly relevant for SCO sentences (e.g., Peter wondered which man in the office he…), it is less straightforwardly applicable in WCO sentences, where the pronoun appears as a genitive modifying a subject (e.g., Peter wondered which man in the office his supervisor…). Given the lack of grammatical role parallelism in WCO, it seems that the parallel function constraint alone cannot account for the structural primacy effects observed in the WCO experiment.

A third alternative concerns the potential role of Condition B. A reviewer suggests that comprehenders may underspecify wh-movement during incremental processing and thus may only consider coreferential readings at the point when the pronoun is initially encountered. On this view, Condition B would rule out the most local referent in favor of a non-local one, i.e., the target, thereby producing the observed preference for the target referent without appeal to the crossover constraint. However, this proposal seems inadequate in light of existing psycholinguistic evidence and in terms of its empirical coverage. First, previous work indicates that native speakers do compute wh-movement during online processing, as revealed by evidence for immediate gap positions (e.g., Felser & Roberts, 2007; Gibson & Warren, 2004; Keine, 2020). In light of this, it is reasonable to assume that participants in our study had already constructed a representation of wh-movement when the pronoun was encountered. Second, this account does not extend to our WCO sentences: since Condition B does not rule out coreference between the wh-phrase and the possessive pronoun (as they are not co-arguments), we should observe coreferential effects, contrary to our findings. Given these reasons, we think that Condition B cannot fully account for our findings.

The last alternative account we consider is topicality. The idea is that the target referent, a proper name no less, may be construed as a topic. Since topicality boosts referential accessibility (e.g., Ariel, 1990; Arnold, 2010; Kaiser, 2011), the preference for the target may stem at least partly from a preference for a topical referent. Relatedly, an anonymous reviewer points out that the structural primacy effect observed in SCO may plausibly reduce to Condition C effects, whereas the corresponding effect in WCO could instead arise from discourse-pragmatic factors (e.g., topicality), without requiring an independent crossover constraint. Under such an interpretation, the dispreference for the distractor antecedent in WCO may reflect differences in discourse rather than structural accessibility.

We acknowledge that this alternative interpretation has important implications for how the WCO findings should be understood, and we therefore adopt a more cautious interpretation of the results. At the same time, we note that although the matrix subject can be theoretically analyzed as a discourse topic even in a stand-alone sentence, topicality was not independently established or manipulated in our experimental design. Therefore, whether and to what extent participants actually treated the matrix subject as a discourse topic is unclear (also see Kaiser, 2011, on the dissociation between subjecthood and topicality in English). Thus, while discourse accessibility could potentially account for the crossover effect, we remain cautious about to what extent this account affects our core conclusions regarding the crossover constraint as an early filter.

In sum, although these alternatives may account for parts of our findings, none provide a fully satisfactory account of the early-stage structural primacy effects. We thus conclude that the Early Filter Hypothesis offers a more plausible account of our results, but acknowledge that potential effects of discourse-level factors merit further investigation.

7.4 Implications for theoretical accounts of crossover

For the remainder of our discussion, we focus on the theoretical implications of our findings. In particular, the current study speaks to ongoing debates about the nature of crossover configurations and the role of D-linking in licensing bound variable interpretations.

First, our results provide empirical evidence that challenges claims that D-linking bleeds the crossover constraint (e.g., Falco, 2010; Ishii, 2006). Recall that Ishii (2006) proposes that a D-linked wh-phrase initially moves to an A-position, rendering the crossover constraint vacuous (e.g., Mahajan, 1990; Reinhart, 1983). Under this analysis, a bound variable interpretation of the pronoun in a WCO configuration should be available from the earliest stages of the syntactic derivation. Similarly, Falco’s (2010) feature-matching account, which derives binding via agreement feature correspondence between the wh-operator and the pronominal variable, predicts full availability of bound variable readings in such configurations.

The results of our acceptability judgment study concerning WCO sentences with D-linked wh-phrases indicate that native English speakers prefer coreference with gender-congruent targets over gender-congruent distractors. Crucially, the online reading-time data further indicate that native speakers initially retrieve the target referent and only consider the distractor at later stages, and, even then, to a limited extent. If D-linking were to bleed the crossover constraint from the outset, one would expect the wh-phrase to be immediately accessible during real-time parsing. This prediction is not borne out by our results. Instead, our findings align better with alternative linguistic accounts that (i) predict crossover effects regardless of the referential complexity of wh-operators (e.g., Agüero-Bautista, 2012; Koopman & Sportiche, 1982; Reinhart, 1983; Shan & Barker, 2006) and (ii) attribute the D-linking effect, characterized by the increased acceptability of D-linked wh-phrases relative to bare ones (e.g., Wasow & Clausen, 2011), to non-syntactic factors (e.g., informativity, plausibility, processing).

Our results are more compatible with a “non-reductionist” view that assigns an independent role to the crossover constraint. At the same time, it is worth considering Culicover’s (2013) processing-based account as a potential complementary perspective. Culicover proposes that crossover effects arise from the low accessibility of wh-phrases at the point when the pronoun is encountered, due to their underspecified grammatical function and thematic role. However, this factor does not fully account for the present findings. In particular, the wh-phrases in our materials are highly referential – being D-linked and further modified by a prepositional phrase – which should mitigate accessibility-based limitations. Indeed, WCO configurations with heavily modified wh-phrases have been taken to be fully grammatical (see Culicover, 2013, p. 130). Thus, for Culicover’s account to be reconciled with the early structural primacy effects, a more explicitly articulated model that quantifies the interaction of various processing factors is needed.

We acknowledge a limitation in our study regarding specificity, as noted by an anonymous reviewer. D-linking was not explicitly manipulated in our design, and a bare wh-phrase baseline is unavailable, since bare wh-phrases do not inflect for gender agreement in English. For this reason, our results only constitute indirect evidence against analyses that predict early availability of bound variable readings with D-linked wh-phrases. That said, we note that the predictions of D-linking accounts should remain applicable to the structures tested in our study even in the absence of a non-D-linked comparison condition; such a baseline would only serve to quantify the magnitude of any D-linking effect. Additionally, we also acknowledge that Falco’s account could potentially accommodate the present findings, if the proposed feature-matching mechanism were assumed to operate at a later stage of processing, even though this extension lies beyond the original proposal. Nevertheless, we believe that our findings contribute meaningfully to theoretical discussions by providing experimental evidence on the interaction between D-linking and crossover in both offline judgments and online processing.

Second, this study highlights the similarities, in addition to the differences, between WCO and SCO. While it has been noted since Wasow (1972) that bound variable readings are considerably less acceptable in SCO than in WCO, it remains unclear how we should interpret this difference in acceptability. A common view held by many researchers is that the crossover constraint is violable, and that the stronger crossover effect observed in SCO is due to the (additional) violation of Condition C, which effectively rules out bound pronoun readings (e.g., Bruening & Al Khalaf, 2019; Lasnik & Stowell, 1991). However, our experimental results refine this picture by showing how the crossover constraint operates over time. We find that the crossover constraint influences interpretation in both SCO and WCO configurations, and that the two constructions show similar processing profiles at early stages. In both Experiments 1 and 2, early reading-time measures reveal no gender intrusion effects and only a main effect of the structurally accessible target, suggesting that SCO and WCO are processed similarly during initial pronoun resolution.

Our findings demonstrate that the differences between SCO and WCO only emerge during later stages of processing. In this respect, our study provides preliminary insight into when, rather than merely whether, the two configurations diverge. As for the source of these later-stage differences, prior experimental work suggests that pragmatic factors, such as plausibility, can modulate the strength of Condition C effects (e.g., Gor, 2020; Harris & Bates, 2002; Salzmann et al., 2023). Indeed, plausibility conceivably may have played a more prominent role in SCO. For example, while it is plausible for someone to be affected by the actions of others (e.g., being promoted by one’s supervisor), as in WCO sentences like (9), it is less plausible for someone to perform an action on themselves (e.g., promoting themselves), as in SCO sentences like (8). To closely examine the independent role of Condition C as a structural effect and the extent to which it constrains the bound pronoun readings in SCO, future studies should carefully tease apart the factors of plausibility and Condition C. We regard this as a valuable direction for future research.