Hierarchical structure has long been considered the hallmark of human language syntax (Chomsky, 1957, Joshi et al., 1975, Kaplan and Bresnan, 1982, Steedman, 2000). Accordingly, it is thought to play a fundamental role in language comprehension by guiding the incremental interpretations that comprehenders assign to sentences, and has been the central focus of many theories of real-time language use (Dillon, 2011, Frazier and Clifton, 1996, Gibson, 1998, Gibson, 2000, Keshev et al., 2025, Kush, 2013, Levy, 2008, Phillips et al., 2011, Sturt, 2003, Wagers, 2008; cf. Frank & Bod, 2011).

Hierarchical structure constrains the interpretation of long-distance dependencies that may span multiple intervening words. For example, in (1) the anaphor each other must be interpreted in relation to its antecedent kids. Despite the presence of feature-matching (i.e., plural) referents in the sentence, i.e., seniors and neighbors, kids is the only available antecedent. This is due to its hierarchical position relative to the anaphor within the sentence’s syntax, expressed in (for example) Principle A of Binding Theory (Chomsky, 1981, Reinhart, 1976): Kids is local (e.g., clause-mate) to the anaphor and hierarchically prominent (e.g., in a c-commanding position) relative to it within the embedded clause.

Such hierarchical constraints are commonly defined in terms of structural relationships between pairs of linguistic items within a sentence, as in x is local to y or x c-commands y.

One question for theories of real-time dependency formation is how to integrate such relational constraints between items with theories of the psychological mechanisms that subserve real-time dependency formation (Alcocer and Phillips, 2012, Dillon, 2011, Kush, 2013, Wagers, 2008). On the one hand, there is an abundance of evidence that comprehenders activate numerous interpretations of linguistic input rapidly and in parallel (e.g., MacDonald et al., 1994). Such a view would suggest that any grammatically licensed interpretation of the input could be made available very early in the comprehension process.

On the other hand, theories of dependency formation often invoke a content-based, associative retrieval mechanism (e.g., Lewis and Vasishth, 2005, Lewis et al., 2006, McElree, 2000, McElree, 2006, Van Dyke and McElree, 2011). On this view, linguistic items are encoded in short term memory with their item-specific features, including structural (e.g., subjecthood, case) and morphological (e.g., number, gender) information. Dependency formation requires reactivating these encodings on demand in response to a set of retrieval cues. Dependency formation should be rapid when cued by the features of a target item, as this allows rapid retrieval from working memory (Lewis and Vasishth, 2005, Lewis et al., 2006, McElree et al., 2003). But relational notions such as c-command or relative linear order are less straightforwardly encoded in such an architecture (Alcocer and Phillips, 2012, Dillon and Keshev, 2024, Kush, 2013, Lewis and Vasishth, 2005, Wagers, 2008). If comprehenders must infer or reconstruct the relationship between two elements to establish that they are in an appropriate configurational relationship (Dillon, 2011, Dillon et al., 2014, Kush et al., 2015, Lewis and Vasishth, 2005), then dependency formation could be slower or less reliable (Lewis and Vasishth, 2005, McElree and Dosher, 1993; Ratcliff & McKoon, 1989; cf. Dosher, 1983).

In an attempt to address this question, much research has investigated the role of structural relations in the processing of anaphors, addressing their time-course and fidelity. This body of work has shown that comprehenders rapidly employ structural constraints to link anaphors with their antecedents. But it does not definitively establish whether comprehenders specifically use hierarchical or relational constraints to do so (Dillon et al., 2013, Kush et al., 2015, Kush and Phillips, 2014). This work has also examined whether structurally unavailable antecedents can interfere in dependency formation. In the present study, we focus primarily on the question of whether hierarchical information guides dependency formation in the earliest stages of comprehension, largely setting aside the question of interference from structurally unavailable distractors.

More specifically, we ask: Does real-time antecedent retrieval rely on hierarchical relations between noun phrases (NPs) as in x c-commands y? We report three visual world studies that measured the moment-by-moment reactivation of potential antecedents in processing a local anaphor in Turkish. We employed stimuli carefully designed to isolate the role of c-command relations between NPs in a sentence from other sources information. We achieved this by manipulating the hierarchical c-command relations of the potential antecedents with respect to the anaphor while holding their clause relations and case marking constant. Across our studies, we tested targets in c-commanding subject and object positions. Furthermore, we manipulated linear order and feature-match of antecedents with respect to the anaphor to reveal the impact of recency and feature-based interference on retrieval, respectively. To preview our results, we found that abstract hierarchical relations between NPs in a sentence determine antecedent availability above and beyond other sources of information from the earliest moments of processing an anaphor. The finding that abstract hierarchical relations immediately guide interpretation constrains psycholinguistic theories of dependency formation, which we take up in discussion.

The processing of local anaphors such as reflexives (himself, herself, themselves) and reciprocals (each other) has been a focus of research because they are common linguistic elements that are subject to relatively well understood syntactic relational constraints on interpretation such as clause-mateness and c-command. The c-command constraint is defined on the basis of the hierarchical organization of individual NPs in the syntactic configuration of a sentence. Formally, a node α c(onstituent)-commands a node β if β is α’s sister or is contained within α’s sister in the tree structure (Büring, 2005, Reinhart, 1976, Reinhart, 1983). For example, in the sentence The seniors mentioned that the neighbors’ kids playfully teased each other (repeated from (1) above), the target subject kids c-commands the anaphor each other because the anaphor is contained within the verb phrase (VP) that is the sister of the target subject’s NP, while neighbors does not c-command the anaphor as it is embedded within this subject’s NP as a possessor and is not a sister to the same VP.1

Another way to formulate the hierarchical relational constraint between a local anaphor and its antecedent is via coargumenthood, which is a central concept in certain theories for anaphor licensing (Pollard and Sag, 1994, Reinhart and Reuland, 1993). Because target antecedents of local anaphors are typically both in a c-commanding position and coarguments of the anaphor, the two factors often co-occur, making it difficult to definitively isolate c-command. Nonetheless, we frame the rest of the discussion here as well as the current study with reference to c-command as it is a fundamental notion governing the distribution of many syntactic dependencies (but see the General Discussion).

Many empirical studies on the processing of anaphors have shown that comprehenders rapidly distinguish hierarchically available, e.g., c-commanding and clause-mate, targets from distractors, which is consistent with the use of relational information in early memory processes (e.g., Dillon et al., 2014, Dillon et al., 2013, Han et al., 2021, Kush et al., 2015, Runner and Head, 2014, Sturt, 2003, Xiang et al., 2009). However, these findings do not necessarily implicate hierarchical, relational processing, as individual studies typically admit multiple interpretations. For example, Dillon et al. (2014) argue that comprehenders use structural information of ‘clausemateness’ to constrain the initial stages of memory retrieval when processing the Mandarin Chinese anaphor ziji (see also, Dillon et al., 2016). While it is possible to consider ‘clausemate’ a relational constraint, this is not a logical necessity. Clause level information could be encoded as an item-level feature, which may have independent psychological motivation if a clause is considered a single temporal context (Howard & Kahana, 2002) in the form of events anchored by its verb (Wagers, 2008). This would allow a simple, feature-based encoding of clausal context that can be cued in retrieval (Balachandran, 2024, Polyn et al., 2009, Sederberg et al., 2008, Wagers, 2008).

More compellingly, several studies have tried to isolate the role of c-command in processing bound variable pronouns (Cunnings et al., 2015, Kush et al., 2015, Moulton and Han, 2018). For example, Kush et al. (2015) compare the processing of bound variable and coreferential pronouns. They report an eye-tracking-while-reading experiment with stimuli as in (2) crossing antecedent type, i.e., quantificational (any janitor) or referential (the janitor), and the c-command relation between the pronoun he and the antecedent using either when or but clauses. In (2-a) the bound variable reading for he is available with the when clause because the when clause is contained within the embedded verb phrase of liked, allowing he to be c-commanded by any janitor. The bound variable reading is unavailable with the but clause because the but clause is combined with the matrix clause via a conjunction phrase higher up, not allowing he to be c-commanded by any janitor. A coreferential interpretation between he and the janitor in (2-b), on the other hand, is available irrespective of the type of the clause. If antecedent retrieval for bound variable pronouns uses hierarchical relational information such as the c-command constraint, extra processing difficulty is expected only when the quantificational NP any janitor does not c-command the pronoun he in a but clause.

The results reveal no evidence for a difference between the two conditions with a referential NP, but increased difficulty in reading the post-pronoun region when any janitor does not c-command he (i.e., in the but clause condition). In a follow-up experiment, they report that feature match of quantificational NPs in a non-c-commanding position did not affect early reading measures, with only suggestive evidence for delayed interference effects from gender-matching, non-c-commanding quantificational NPs in later measures (for similar findings, see Cunnings et al., 2015). In a later study, Moulton and Han (2018) disentangle the syntactic notion of c-command from the semantic notion of scope, finding that it was the syntactic c-command relation that determined antecedent availability in processing bound variable pronouns. Overall, these results show that comprehenders use hierarchical relational processing to distinguish available from unavailable quantificational NPs when interpreting bound variable pronouns.

Still, it is possible that Kush et al.’s (and Cunnings et al.’s and Moulton and Han’s) findings are driven by the hierarchical position of an entire clause (e.g., when vs. but clauses in Kush et al.) relative to the critical antecedent in a different clause. That is, the structural manipulation in those studies concerns items across different clauses. Thus, although their findings do provide evidence for the use of clausal relations in structuring memory retrieval, they do not show evidence for relational processing within a single clause.

There is also evidence that the hierarchical organization of NPs impacts the availability of their memory representations for other types of processing. For example, testing subject-verb agreement cases in a probe recognition task, Franck and Wagers (2020) show that attractor nouns that c-command the verb in the surface order are more readily available in memory than those that do not (for similar findings in judgment tasks, see Franck et al., 2006, Franck et al., 2020, Franck et al., 2010). However, such findings do not necessarily indicate that c-commanding nouns are represented or accessed from memory on the basis of relational processing. Rather, as discussed by Franck & Wagers, these findings may reflect increased activation levels for c-commanding nouns in memory.

To date, dozens of studies have explored the real-time processing of anaphors, potentially establishing them as one of the most commonly studied dependencies in psycholinguistic research. On the surface, there appears to be a considerable amount of evidence that memory retrieval in real-time processing of anaphoric dependencies is sensitive to hierarchical information (e.g., Dillon et al., 2013, Nicol and Swinney, 1989, Sturt, 2003, Xiang et al., 2009). Studies employing different methodologies show that comprehenders rapidly differentiate the hierarchically available target antecedent from the hierarchically unavailable distractor. This is evidenced by the priming of semantic associates of targets in cross-modal priming tasks (Nicol & Swinney, 1989), increased difficulty in reading time measures in the absence of a feature-matched target antecedent (e.g., Dillon et al., 2013, Sturt, 2003), increased positivities in event-related potentials upon processing an anaphor with a feature-mismatching target (e.g., Xiang et al., 2009), and increased looks to targets over distractors in visual world studies (e.g., Han et al., 2021, Runner and Head, 2014). However, while these results provide clear evidence that comprehenders do use some type of structural information in retrieval, they do not unambiguously show that a hierarchical or relational notion like c-command per se is used to accomplish this.

Against the rich empirical background on anaphoric dependencies, it is perhaps surprising that the role of hierarchical relations in resolving anaphoric items remains unclear. But one key observation about existing studies is that most used stimuli that confounded multiple sources of structural information in defining the target of retrieval, failing to uniquely isolate the role of c-command or similar hierarchical notions in retrieval. Instead, a range of structural, item-level information could be used to retrieve the target antecedent instead of a relational notion like c-command. To substantiate this observation, we conducted a review of the literature on processing local anaphors, summarized in Table A.10 in Appendix A. Our crosslinguistic review includes 45 experiments from 28 studies on local anaphors that included structurally unavailable distractors. The studies are predominantly in English, with some in Chinese, German, Hebrew, Hindi, Korean, Russian, Swedish, Tagalog and Turkish. We examined the experimental stimuli in these studies to determine what information was used to distinguish the target and distractor. Our review also includes details about the studies (e.g., language under investigation, number of participants) and summarizes the findings as originally presented.

We find that no study in our review directly isolates the role of c-command per se without confounding it with other potential structural cues such as clause-mateness and/or case marking. Importantly, these are cues that are plausibly construed as item-level features, obviating the need to implement hierarchical or relational structure at retrieval. In our analyses of the stimuli from previous studies we use case marking to refer to the grammatical role of a noun phrase in the sentence or its morphophonological marking, as we consider both types of case marking to be plausibly relevant—and both can be construed as item-level features. Accordingly, we consider the target and distractor to be differently case-marked when they have either distinct (abstract) structural case marking, as this points to differing grammatical roles in the sentence (e.g., subject vs. object),2 or different overt morphophonological case marking.

A summary of our review is illustrated with a Venn diagram in Fig. 1. The diagram shows the structural/relational cues (i.e., C-command, Clause, Case) that distinguish the target from the distractor.

To illustrate further, 12 experiments from 10 studies create a distractor by manipulating all three structural/relational features. An example is (3-a). The distractor (middle manager) is in a non-c-commanding position, in a different clause from the target (new executive), and has a different case (i.e., grammatical role) from the target. Fifteen experiments from 12 studies have a non-c-commanding, non-clause-mate distractor (e.g., Fred in (3-b)). Five experiments from 4 studies have a non-c-commanding, differently-case-marked distractor (e.g., fireman in (3-c)). One study has a non-clause-mate, differently-case-marked distractor, (e.g., janitor in (3-d)). Lastly, 15 experiments from 12 studies have a non-clause-mate distractor (e.g., John in (3-e)). In (3) anaphors are given in bold, target antecedents are underlined and distractors are in italic.

As discussed above, clause-mateness could be driven by item-level features indexing temporal/contextual information (Balachandran, 2024, Wagers, 2008). Likewise, in all the studies that we review the target antecedent was the subject of the clause containing the anaphor. Comprehenders could thus rely on a composite, item-level cue such as subject-of-the-current-clause in accessing the target (e.g., Arnett and Wagers, 2017, Kush and Phillips, 2014, Schoknecht et al., 2025, Xiang et al., 2009). In either case, retrieval could succeed without appealing to hierarchical or relational processing among elements within a clause.

While many acknowledge the importance of testing relational notions like c-command in retrieval, our review shows that it has proven difficult to uniquely isolate their role in the processing of anaphors. Our study here aims to fill this gap by testing whether hierarchical relations between NPs within a clause can be used to guide antecedent retrieval. We investigate the preverbal, local reciprocal birbirleri in Turkish, which must be bound by a c-commanding, local (e.g., clause-mate) antecedent3 (Akkuş, 2021, Kornfilt, 1997). Turkish provides an ideal empirical ground to test this: (i) Its rich morphological system allows us to manipulate the hierarchical availability of antecedents while eliminating other confounding cues (i.e., clause, case and subjecthood), and (ii) its head-final structure allows us to investigate a preverbal anaphor, which is crucial for eliminating any confounds related to reactivation of other arguments at the verb (for a discussion of the importance of this point, see Dillon et al., 2013, King et al., 2012).

We examine the availability of c-commanding subjects (Subj) in Experiment 1 (as simplified in (4-a)) and c-commanding indirect objects (IOs) in Experiment 2 (as in (4-b)), both in contrast to non-c-commanding distractors located within the same embedded clause as the anaphor. Experiment 3 is a high-power, pre-registered replication of the core contrasts in Experiments 1 and 2 (as in (4-c)). We control for case marking and linear order and manipulate the number (mis-)match between the anaphor and the critical referents to test the advantage of c-commanding antecedents above and beyond these factors.

We use the visual world paradigm for spoken language comprehension in our experiments (Cooper, 1974, Tanenhaus et al., 1995). This choice contrasts with many studies on processing anaphors, which generally employ reading methodologies. However, reading measures only generate indirect inferences about memory retrieval processes through variations in reading times (e.g., slow-downs or speed-ups). In contrast, the visual world paradigm offers a more direct window into the competition between the target antecedents and distractors by tracking overt looks to their visual referents, as there is a well-documented relationship between fixation behavior and attention to the referents of visually presented objects (Allopenna et al., 1998, Magnuson, 2019). It is also highly suitable for investigation of real-time anaphor resolution due to its high temporal resolution (e.g., Han et al., 2021, Runner et al., 2003).