In the forty years since Chomsky (1981) first proposed the Binding Theory, there has been a wealth of research on the form, distribution, and interpretation of reflexive anaphors (henceforth reflexives). The 1981 Binding Theory consists of three syntactic principles. Two of the principles are stated in (1a), with some relevant terms defined in (1b).

Principle A requires reflexives and other anaphors to have a c-commanding antecedent in a designated local domain—a domain roughly equivalent to the clause. Principle B requires pronominals (ordinary personal pronouns) not have a c-commanding antecedent in this local domain. Together, Principles A and B paint a simple, appealing picture in which reflexives and pronominals are in complementary distribution: in structures in which a reflexive is grammatical, the corresponding pronominal with the same index is ungrammatical.

It quickly became clear that this picture is too simple. On the one hand, reflexives are heterogeneous, both crosslinguistically and within individual languages. The heterogeneity extends to their morphosyntax, the domains within which they must be bound, the types of DPs that can bind them, their (in)sensitivity to perspective, and so on (Reinhart & Reuland 1993; Cole et al. 2006; Kiparsky 2015; Déchaine & Wiltschko 2017; Reuland 2018; Charnavel 2021). These language-specific details must be stipulated. On the other hand, under the right discourse conditions, a pronominal can appear to have a c-commanding antecedent in its local domain. Such pronominals often have a different interpretation from the corresponding reflexive (Reinhart 1983; Heim 1998). In response to all this, numerous attempts have been made to update, revise, or replace the Binding principles in (1a) (Reinhart 1983; Burzio 1991; Levinson 1991; Pollard & Sag 1992; Reinhart & Reuland 1993; Levinson 2000; Wilson 2001; Safir 2004; Reuland 2011; Safir 2014; Kiparsky 2015). These attempts are conceptually and analytically diverse. What they share is that—unlike the 1981 Binding Theory—they are not exclusively syntactic. Instead, they account for the distribution of reflexives and pronominals by appealing to the interaction of syntax, semantics-pragmatics, and morphology. Most such approaches adopt some or all of the following assumptions:

One challenge for all approaches to Binding Theory is to explain why reflexives should be morphologically distinguished from pronominals in the first place. Safir (2014), who is one of the few to address this question directly, states, “One might reasonably speculate that the manifestation of this sort of morphological within-phase distinction, expressible in the syntax, might be reinforced by functional pressure for systems where semantic binding at the most local opportunity is distinguished interpretively from semantic binding that requires a longer store” (p. 119). The exact character of this functional pressure is unclear. Conceivably, it might be related to Levinson’s (2000: 273) claim that “where a coreferential interpretation is available it will generally provide an informationally richer, more informative, stereotypical, and ‘better’ interpretation” than a noncoreferential interpretation.

What functional pressure might privilege semantic binding at the most local level? This question could potentially find an answer in language processing. But previous research on the comprehension of reflexives has been limited to languages in which they are morphologically differentiated from pronominals, such as English and Mandarin (see Dillon 2014 for a review). It is unknown how comprehension proceeds in languages in which the two are not morphologically distinct. In two experiments, we probed the comprehension of reflexives in Chamorro, an Austronesian language that is spoken in the Mariana Islands. Chamorro has no morphologically distinct reflexive anaphors; instead, reflexive clauses are constructed with ordinary pronoun forms. We asked two questions. First, given that Chamorro has no morphologically distinct reflexives, what cues, if any, do comprehenders rely on when deciding whether to interpret a given pronoun form as reflexive? Second, when such cues are entirely absent, is there nonetheless ‘functional pressure’ that favors a reflexive interpretation when it is available (and grammatical)?

The rest of this introduction surveys the literature on the processing of reflexive pronouns (in 1.2) and then introduces relevant aspects of the morphosyntax of Chamorro (in 1.3). Section 2 discusses Experiment 1, a touch-tracking picture-matching experiment that examined the comprehension of transitive clauses whose object could be interpreted as reflexive. Section 3 discusses Experiment 2, a picture-matching experiment that examined the comprehension of clauses in which the possessor of the object could be interpreted as reflexive. Section 4 concludes.

In sentence processing, anaphora resolution is construed as a set of processes for identifying the antecedent of a reflexive anaphor, pronominal, or other anaphoric element and linking the two.2 This has proven to be a fruitful area of investigation because it demands theories of how the processor is able to coordinate representations across several levels—lexical, syntactic, semantic, and so on. Such coordinated representations are richly structured and Janus-like. On the one hand, they prospectively generate expectations about upcoming elements (Arnold & Tanenhaus 2011). On the other hand, these representations must be reactivated or navigated when an anaphoric element is encountered. An early suite of studies on the processing of reflexive anaphors and pronominals in English asked whether the principles of the Binding Theory act as “early filters” to generate the candidate antecedents for the anaphor or pronominal (Nicol & Swinney 1989; Badecker & Straub 2002; Sturt 2003). In the past decade, the questions addressed by this research have been recast as questions about memory retrieval (Dillon 2014) or a “mix of retrieval and expectancy-based processes” (Kaiser et al. 2009). What are those retrieval processes, and do they manifest a functional pressure for local binding?

Psycholinguistic theory was re-attuned to the associative character of working memory by the influential language processing experiments of McElree (2000), McElree et al. (2003), and Van Dyke & Lewis (2003), and the accompanying computational proposals of Lewis & Vasishth (2005) and Lewis, Vasishth, & Van Dyke (2006). The core claim is that recently-encountered linguistic material can be accessed in two ways: (i) directly, as an item or feature in the focus of attention; or (ii) associatively retrieved, using a set of cues provided by the grammar/parser in the local context. The associative character of retrieval was revealed in contexts in which the matrix verb being processed was preceded by a complex subject phrase (Van Dyke & Lewis 2003; Wagers et al. 2009). Constituents within the complex subject which are syntactically inaccessible but, critically, partially match the requirements of the matrix verb can influence processing at the verb. This is unexpected if the verb’s dependents are identified by strictly traversing a hierarchical representation like a tree; however, it follows if the verb’s dependents are identified by a cue-based retrieval in which recently encountered elements are re-activated as a function of their similarity, or match, to the cues. Elements that partially match are prone to cause interference (Van Dyke & McElree 2011).

Against this background, reflexives have “stuck out”: unlike subject-verb agreement or even the thematic integration of the subject, the dependency between a reflexive anaphor and its antecedent is less vulnerable to interference. This generalization, which was suggested by earlier research (Phillips et al. 2011), was directly validated by Dillon et al. (2013) in a series of eye-tracking and computational modeling studies which compared the processing of subject-verb agreement and the processing of reflexive anaphors. However, the apparently special behavior of reflexives may be a matter of degree, rather than kind. For example, Jäger et al.’s (2017) metanalysis suggests that interference effects do exist in reflexive anaphora, but are considerably smaller than in subject-verb agreement. And Parker & Phillips (2017) provided evidence that, when a reflexive direct object pronoun mismatched the subject of its clause (the closest DP that could potentially bind it) in more than one feature, nonlocal potential binders could have an impact on comprehension and could ease the effects of ungrammaticality. Of course, this does not, on its own, explain why reflexive anaphora and subject-verb agreement show different degrees of resilience to distractors. What is responsible for the extra “protective factor” of reflexives?

Parker & Phillips (2017) relate the protective factor of reflexive pronouns to their (un)-predictability and how that affects the quality of the retrieval cues used by the comprehender. They conjecture that, because reflexive pronouns are simply less predictable than agreeing verbs, resolving the dependency between a reflexive and its antecedent will depend more on retrieval and less on prediction. Consequently, they reason, the comprehension mechanism will more strongly weight cues related to the dependency’s structural requirements. In contrast, the features of agreeing verbs can routinely be predicted from their agreement controllers. This means that retrieval might simply not be needed to resolve every subject-verb dependency, as pointed out by Wagers et al. (2009). And consistent with that view, interference, in the guise of “agreement attraction”, occurs most reliably in ungrammatical strings when the prediction made about the verb’s features fails.

Put in slightly different terms, the structural constraints that apply to subject-verb agreement are typically implemented prospectively, by predicting a verb of the right type and position. The structural constraints that apply to reflexive anaphora are typically implemented retrospectively, by incorporating them as part of the cue set used during retrieval.3

Structural cues have been considered challenging to implement using standard associative memory mechanisms, because structural constraints are often conceived of as relational and open-ended. For example, the c-command relation can be evaluated for any two heads or projections in the structure, but “being a c-commander” isn’t an inherent feature of any given head. Thus, it cannot be recruited as a feature to use during retrieval. This problem, however, is not insuperable, because there are inherent features that could serve as reasonable surrogates. For example, in resolving a reflexive-antecedent dependency, the parser may do well to first reactivate subjects, which are the classic antecedents of reflexives, even though these might not be co-extensive with the entire set of c-commanding DPs. And an attribute like nominative case, or a similar feature involving grammatical relations, could be used to directly identify elements in “subject position” (Arnett & Wagers 2017).

Another open-ended relation is locality: can the parser limit the retrieval of preceding elements to just the local ones, since “being local” is not a fixed feature of any given item? Surrogates are probably available here, too. For example, Wagers (2008) proposed that all elements are encoded with a global state vector which evolves slowly with the parser’s state (cf. Howard & Kahana 2002), but changes more rapidly at clause boundaries. This would allow retrievals to initially target more local elements or, at least, elements belonging to the same clause. And some evidence suggests that the search for a reflexive’s antecedent can first target just the subject of the local clause (henceforth, the local subject). Dillon et al. (2014) investigated the processing of the Mandarin long-distance reflexive ziji. This reflexive anaphor can have a c-commanding antecedent that is the local subject or the subject of a higher clause, and can also function as a logophor (cf. Huang & Liu 2001; Cole et al. 2006). At stake in their experiment was the speed of accessing the local subject versus a more distant subject in the next higher clause. (2) illustrates this contrast, which takes advantage of the fact that the antecedent of ziji must be animate and sentient: in (2a), the only acceptable antecedent, Coach Zhang (underlined), is the more distant subject, while in (2b), the only acceptable antecedent is the local subject. Dillon et al. (2014) used the MR-SAT response-signal technique (Wickelgren et al. 1980; Martin & McElree 2008) to measure the speed-accuracy trade-off function for judging the acceptability of sentences like (2a–b). The key result was a dynamics difference in the SAT function. Participants were faster at recognizing (2b) as acceptable compared to (2a), even though participants ultimately preferred sentences in which ziji was bound by the more distant subject, as in (2a).

Dillon et al. (2014) constructed two computational models to explain these differences, both grounded in an associative memory. In their Local Search model, comprehenders initially attempt to retrieve only local antecedents—that is, antecedents that are local subjects. Only if this fails do they then proceed to cast a wider net. In the Unrestricted Search model, comprehenders do not specifically search for local antecendents, but a locality advantage can emerge due to the prominence or recency of the local subject. In both models, the characteristic that controls how well they fit the data is how often memory has to be sampled to attain a compatible antecedent (or, more precisely, the probability distribution over number of samples). In a model comparison analysis, Local Search outperformed Unrestricted Search with a 5:1 Bayes Factor advantage.

The question prompted by their study is why comprehenders engage in Local Search when multiple subjects are available that could serve as grammatical antecedents of ziji. If a locally-bounded search is necessary, or strongly incentivized for other reasons, then that could support the idea that Safir’s (2014) functional pressure for local binding is a processing pressure. What remains unknown is whether this pressure would be observed in the absence of a morphological distinction between reflexives and pronominals.

Chamorro is a verb-first language in which the word order of arguments after the verb is not fixed. Clauses can be Verb Subject Object (VSO) or Verb Object Subject (VOS), although VSO is the default. Subjects or direct objects (henceforth simply objects) that are personal names are preceded by the case marker si; otherwise, subjects and objects have no overt case marking; see (3). In this and other examples, CD identifies sentences taken from the electronic database for the Revised Chamorro-English Dictionary.

The verb agrees with the subject through a prefix or proclitic that also indicates transitivity and mood. When the verb is transitive, this agreement registers person and number. In (3), for instance, the proclitic ha indicates that the subject is third person singular, the verb is transitive, and the mood is realis.

Chamorro has both overt and null pronouns.4 Pronouns that are cross-referenced by agreement in person must be null. These necessarily null pronouns include the pronoun subjects of transitive verbs and pronoun possessors, such as the possessor of kålulut ‘finger’ in (3b). Otherwise, animate pronouns can be overt or null; see the object pronoun in (4).

The overt object pronoun forms are weak pronouns, which must immediately follow the verb (in the clauses discussed here). Thus, in a transitive clause, a pronoun subject must be null and a pronoun object must immediately follow the verb.

Chamorro has no morphologically distinct reflexive anaphors. Ordinary pronoun forms are used when a pronoun is interpreted as reflexive, that is, bound by a c-commanding antecedent in the local clause (= the clause most immediately containing the pronoun). Consequently, many reflexive clauses have no morphology at all that distinguishes them from nonreflexive clauses; see (5).

Pronouns that are interpreted as reflexive are generally overt. This is why the weak pronoun object in (6a) can be interpreted as reflexive, but the null pronoun object in (6b) cannot be. However, a null pronoun can be interpreted as reflexive when it is cross-referenced by agreement in person and therefore cannot be overt. For instance, the null pronoun possessor in (6c) can have a reflexive interpretation because it is cross-referenced by the 3sg. possessor-noun agreement suffix -ña.

Finally, verb morphology can signal that a pronoun object is interpreted as reflexive. In such cases, the reflexive adverb maisa ‘(by one)self’ combines with the verb to form a reflexive verb, as in (7).

The combination of the verb plus maisa forms a single phonological word. This is why reduplication for the progressive aspect, which doubles the stressed CV of the verb, doubles the first CV of maisa in (8).

The presence or absence of maisa seems to have no semantic effect; compare (5b) and (7).

The absence of morphologically distinct reflexive anaphors raises the issue of whether Chamorro is constrained by the Binding Theory at all. The issue is whether pronouns that are interpreted as reflexive differ syntactically from other pronouns, despite the absence of any morphological distinction. In fact, Chamorro grammar does indeed distinguish pronouns that are reflexive anaphors from other pronouns (which are pronominals). We show this here for pronoun objects.

The most straightforward evidence comes from the Chamorro restrictions that ensure that subjects are salient (i.e. topical or noteworthy). Some of these restrictions demand that the subject have a minimum level of salience; others ensure that the subject is at least as salient as the object. For instance, the person-animacy restriction (PAH) prevents the object from outranking the subject in person or animacy (Chung 2014, 2020; Clothier-Goldschmidt 2015). One subcase of the PAH prevents a transitive verb from having an animate pronoun as the object when the subject is a nonpronoun. The transitive clauses (9a–b) conform to the PAH, but the ungrammatical (9c) does not.

Another such restriction is the third plural restriction (TPR), which prevents a transitive verb from having a subject that is third person plural, but not a pronoun (Chung 2020). The transitive clause (10a) obeys the TPR, but the ungrammatical (10b–c) do not.

The PAH is sometimes violated in elicitation contexts in fieldwork, but is a hard constraint in production (Clothier-Goldschmidt 2015). The TPR is never violated. Both restrictions exhibit morphological quirks that are irrelevant here.5 When a clause would violate these restrictions, the typical remedy is to employ the corresponding passive, since passives are intransitive.

Significantly, the PAH and the TPR do not affect transitive clauses in which a pronoun object is interpreted as reflexive. These reflexive clauses systematically escape the restrictions. The reflexive clauses in (5b), (7), and (8) ought to violate the PAH, given that their object is an animate pronoun but their subject is a nonpronoun. Nonetheless, they are grammatical. Similarly, the reflexive clauses in (11) ought to violate the TPR, because their subject is third person plural but not a pronoun. These clauses are grammatical as well (Chung 1989).

It is important that the pronoun object must be bound by the local subject in order for the clause to evade the PAH and the TPR. Clauses in which the pronoun object is bound instead by a c-commanding antecedent in a higher clause, or merely has an antecedent in previous discourse or a nonlinguistic antecedent, conform to the PAH and the TPR as expected. This pattern reveals that Chamorro grammar differentiates object pronouns that are interpreted as reflexive from other object pronouns. These pronouns are reflexive anaphors, even though they are not morphologically differentiated from other pronoun forms.6

The fact that reflexive anaphors are not distinguished by their morphological form raises the question of how they are recognized in comprehension. What cues are comprehenders sensitive to as they decide whether to interpret a given object pronoun as reflexive, and how does comprehension proceed when those cues are absent? Experiment 1, which investigates these questions, was designed to test two initial predictions. The first prediction was that comprehenders would be biased toward the reflexive interpretation by indirect cues, such as the reflexive adverb maisa, or a subject-object combination that would otherwise violate the PAH. The second was that when these cues were removed, the choice between the reflexive and nonreflexive interpretations would be essentially free. In brief, the results confirmed the first prediction, but not the second. Participants preferred reflexive interpretations, even with no direct or indirect morphological signal to guide them.

96 Chamorro speakers participated in Experiment 1, which took place in the U.S. Commonwealth of the Northern Mariana Islands (CNMI) in September 2016. They grew up in Saipan (39), Rota (49), Guam (5), and Tinian (3). Participants ranged in age from 22 to 70, with a median age of 45 years, and an inter-quartile range of 15.5 years. 56 participants reported that they had taken part in a previous experiment.

The experiment followed a within-subjects and within-items design. Each itemset contained four sentences in which the object was the overt 3sg. object pronoun gui’ and the local subject could potentially serve as its antecedent. These sentences crossed two factors: (i) Local Subject (LclSubj)—whether the local subject was a (null) Pronoun or a Nonpronoun (overt DP); and (ii) Reflexive Adverb (ReflAdv)—whether the reflexive adverb maisa was present or absent.

In addition, each itemset contained two control sentences in which participants could only access a non-reflexive interpretation. The PAH-Only control contained just one overt DP and no overt pronoun. The PAH forces the single overt DP in such clauses to be parsed as the object. The PAH+Plausibility control also contained just one overt DP, which in most cases was constrained by world knowledge to be plausible as the Theme but not the Agent.7 (Throughout, we use Agent as a cover term for the thematic role assigned to the subject, and Theme for the thematic role assigned to the object.)

18 itemsets were constructed which each realized the six conditions. Thus 108 target items were created in total, and were distributed via a Latin Square so that each participant heard 3 target items per condition and no participant heard more than one target item from the same itemset. Table 1 illustrates an itemset using the verb patmåda ‘slap’. All verbs were progressive. 9 itemsets contained an oblique DP that appeared after the verb and before any (overt) local subject; in Table 1, the oblique DP is ni panak lålu’, ‘with a flyswatter’.8 Appendix A lists the verbs and any oblique DPs.

Each target sentence was preceded by a context sentence that served to introduce the two individuals in the scenario and thus to make a null pronoun felicitous. The context sentence was played simultaneously with the display of a context picture on the screen, as in Figure 1A. Contexts were constructed so that both the reflexive and disjoint interpretations of gui’ in the target sentence were plausible. For the example itemset illustrated in Table 1 and Figure 1, the context sentence was (12):

The target sentence was embedded in the task instruction, which began playing immediately following the context sentence. Participants were directed to push a small icon, or ‘puck’, over to a depiction of the target sentence. The picture selection screen, as in Figure 1B, was presented simultaneously with the onset of the target sentence audio. For the example in Table 1, the full task instruction for the critical ReflAdv Absent:LclSubj Pronoun condition was (13):

Both pictures depicted the same type of event, e.g. a ‘slapping’ event. In the refl picture, a self-directed version was depicted, and thus only one individual participated; in the disj picture, there were two distinct individuals participating. refl and disj pictures were left-right counterbalanced on the response selection screen. Among the pictures that depicted two distinct participants, the left-right orientation of Agent/Theme participants was also counterbalanced. In the set of 18 refl pictures, 13 depicted a single individual only and 5 depicted both individuals in the frame. Pictures were created by an illustrator working from photos and descriptions of life in the CNMI and Guam.

Participants could begin moving the puck at any point after the onset of the task instruction. Three pucks were counterbalanced across the experiment: a star, a heart, and a fish. A picture was registered as selected once the puck entered one of the two response selection boxes, at opposite corners of the screen. A small checkmark button then appeared below it. Participants had to press this button to advance to the next trial. However, they could also change their response at any time before confirming it. Once the task instruction finished playing, participants could replay both the context sentence and task instruction by pressing a small replay button which appeared in the bottom left-hand corner of the screen.

The experiment was programmed in the OpenSesame environment (Mathôt et al. 2012) and deployed on Samsung Nexus 10 tablets. All stimuli and instructions were delivered auditorily in Chamorro. For each participant, the 18 target trials were randomly interspersed with 12 filler trials with exactly the same structure—picture selection with context and target screens—using items adapted from another experiment (Wagers et al. 2018). Thus each participant heard 30 trials in total. The 12 filler trials contained target sentences with ambiguous relative clauses, each of which had two possible interpretations.

The primary analysis was two regressions of reflexive choices. First, we regressed reflexive choice on the two experimental factors (ReflAdv, LclSubj) for the four core conditions in the experiment. As it turned out, there were vanishingly few errors in the ReflAdv Present conditions and, for some itemsets, in the Controls as well. Therefore, we used Firth’s penalized likelihood method (Firth 1993), implemented in the R package logistf with intercept corrections (Heinze et al. 2020). This analysis is more appropriate than mixed-effects logistic regression in this application, where quasi-complete separation occurs due to (nearly) perfect performance in some conditions, participants, and itemsets. Confidence intervals over slope estimates were reported based on penalized profile likelihood. We then separately analyzed the three PAH-constrained conditions (the ReflAdv Absent conditions plus the PAH-Only control) by estimating a Bayesian multilevel logistic regression in the package brms (Bürkner 2017).

In a secondary analysis, we analyzed touch time (TT), or the time when participants first touched the tablet, as measured from audio offset. TTs were analyzed with a Bayesian multilevel linear regression on the experimental factors, with an LKJ(2) prior on the random effects correlation (Lewandowski et al. 2009). A targeted pairwise comparison was then performed on critical ReflAdv Absent:LocalSubj Pronoun conditions, contrasting TTs for refl and disj readings.

ReflAdv Present conditions gave rise to reflexive responses in nearly all trials: 544 of 556 valid responses were reflexive responses (98%). Control trials, where the interpretation was constrained to be disjoint, either by the PAH alone or in conjunction with Plausibility, gave rise to reflexive responses infrequently: 57 of 557 valid responses were reflexive responses (10%). Thus, participants could perform the task accurately, and were able to discriminate two classes of sentences with unambiguous interpretations.9 Table 2 gives the percentage of reflexive choices for each condition and a logistic regression on the ReflAdv × LclSubj design, showing that both the ReflAdv and LclSubj factors affected choices.

In the ReflAdv Absent conditions, comprehenders made fewer reflexive choices overall. In these conditions, no special morphology marks reflexivity: only by virtue of the use of gui’ and the PAH can the comprehender calculate when a reflexive interpretation is allowed or required. We predicted that the presence of a LclSubj Nonpronoun would force a reflexive interpretation, given that the PAH should exclude the disjoint interpretation. Reflexive interpretations did, in fact, predominate, but they were not as common as in the ReflAdv Present conditions. The PAH has no impact when the LclSubj is a pronoun, so in this condition both the reflexive and disjoint interpretations should be possible. Here, we found fewer reflexive choices compared to the corresponding LclSubj Nonpronoun condition, but comprehenders still preferred reflexive interpretations.

We further modeled reflexive choices on three related conditions: the two ReflAdv Absent conditions and the PAH-Only control. These three conditions are linked because their interpretations are not constrained by verbal morphology (and they did not give rise to near-perfect performance). We constructed a Bayesian multilevel logistic regression on these conditions.10 We represented the differences among them with two coefficients coded using Helmert contrasts. One coefficient, pah strength, compared the two unambiguous conditions (LclSubj Nonpronoun and PAH-Only); the other coefficient, reflexive preference, compared the critical ambiguous condition (LclSubj Pronoun) with the mean of the two unambiguous conditions. Estimates of these coefficients support the following three conclusions.

We measured when participants initiated their response by first touching the tablet. Figure 2 plots the overall distribution of this variable, the Touch Time (TT). When we first considered the four target conditions, we found two simple effects: TTs were much faster in refladv Present conditions (–1106 ms, 95% CrI [–1461 ms, –747 ms]) and in LclSubj Nonpronoun conditions (–742 ms, 95% CrI [–1106 ms, –386 ms]). In this way, faster TTs correlate with conditions in which more reflexive responses were given. This is unsurprising if those conditions are the ones in which multiple grammatical constraints combine to deliver a more confident reading.

If we consider just the critical ReflAdv Absent:LclSubj Pronoun condition, we can ask whether participants gave faster responses to the disj or the refl reading. Although the disj reading gave rise to slightly slower TTs on average (215 ms), the 95% credible interval of TT for disj responses was –879 ms to 1249 ms. While we often expect minority responses to give rise to longer response times, we may simply have not had adequate power to resolve a true effect. Nonetheless, we found no evidence from TTs to suggest that the disj reading is somehow less accessible than the refl reading.

Inspection of individual item rates suggests that considerable variation was linked to specific item properties. We believe that the source of this variability was the verb’s argument structure—specifically, whether its Theme argument (which was always realized as the direct object in our stimuli) must be animate. Seven of our verbs require their Theme to be animate (e.g. na’chotchu, ‘feed’, suda’i, ‘find for’). When we split the data by this feature, the average percentage of reflexive choices for the critical ambiguous sentences was only 57% for verbs that require an animate Theme (n = 7) versus 93% for verbs that do not require this (n = 11; e.g. litratu, ‘photograph’, tåmpi, ‘cover’). To further explore the strength of the relationship between argument structure and reflexive interpretation, we added verb type and its interaction with the experimental conditions as a covariate to our model. Estimates from this richer model are plotted in Figure 3B, which shows an overall decrease in reflexive choices when the verb requires its Theme to be animate. This is reflected in a simple effect of verb type, 1.54 (s.e.: 0.43; 95% CrI: 0.70–2.39). The interaction of verb type with pah strength – the comparison between the two unambiguous conditions – was nearly zero (0.01, s.e.: 0.99; 95% CrI: –1.93–1.99). The averages plotted in Figure 3 suggest that the effect of verb type is strongest for the ambiguous condition (ReflAdv Absent:LclSubj Pronoun). The interaction of verb type with reflexive preference, which would quantify this impression, was positive (0.64, s.e.: 0.52), but the 95% CrI includes many negative values (–0.38–1.66). Therefore, it would be premature to conclude that this effect is confined only to the ambiguous condition. We instead interpret it as an across-the-board effect.11

While this analysis is exploratory, we tentatively conclude that the verb’s argument structure modulates how the comprehender interprets the object pronoun gui’. Further research would be needed to establish whether this effect is independent of the reflexive adverb.

The results from Experiment 1 show that comprehenders can accurately parse sentences whose object is unambiguously reflexive, even though Chamorro has no morphologically distinct reflexive pronouns. Sentences in which the reflexive adverb maisa appeared were almost invariably interpereted as reflexive. Sentences in which the local subject was a nonpronoun, and therefore a disjoint interpretation was excluded by the PAH, were interpreted as reflexive almost as frequently. The PAH was not as effective a cue as the reflexive adverb. We attribute this to the fact that maisa is a direct morphological signal of reflexivity, whereas the PAH’s impact on interpretation can only be calculated once the subject and object DPs are identified and their lexico-semantic content is compared. It is not surprising that comprehenders might make an interpretive choice before the evidence provided by these calculations is available.

When the local subject was a pronoun and the reflexive adverb was absent, the object pronoun was technically ambiguous between the reflexive and disjoint interpretations. However, comprehenders’ choices revealed a strong preference to interpret these sentences as reflexive. We attribute this reflexive advantage to two morphosyntactic factors. First, the object pronoun gui’ contrasts with the null object pronoun, which is never reflexive and, therefore, always disjoint from the local subject. Armed with this knowledge, comprehenders might reasonably infer that had a disjoint interpretation been intended, the speaker would have chosen the null object pronoun. The second factor is the PAH. The verbs in the ReflAdv Absent conditions show 3sg. subject-verb agreement and are immediately followed by the 3sg. object pronoun gui’. The subject, if overt, is encountered only later. Consequently, an early commitment to a reflexive interpretation of gui’ is a safe choice, because the result will always be grammatical. In contrast, an early commitment to a disjoint interpretation comes with the risk that the subject might be a nonpronoun, in which case the PAH would be violated and reanalysis would be required.

How can we be sure that the reflexive advantage is a real preference, and not merely response bias? If participants had an overall bias to make reflexive responses, we should have found fewer errors in the ReflAdv Absent:LclSubj Nonpronoun condition, and more errors in the PAH-Only Control condition. In fact, the error rates were identical: 13% (see Table 2). This suggests that, in Experiment 1, neither reflexive nor disjoint interpretation was favored.

We were surprised to find a correlation between the proportion of reflexive interpretations and the verb’s argument structure. When the verb required its Theme to be animate, the reflexive advantage was very slight indeed. When the verb did not require this, the reflexive advantage was as strong as in the ReflAdv Present conditions and the PAH-Only controls, when the sentence had only a reflexive interpretation. We speculate that the reflexive advantage is generally strong in sentences of this type, but can be depressed by a semantic-pragmatic factor—the verb’s argument structure—as follows. Recall that a predicate’s arguments are “intended to be disjoint, unless marked otherwise” (Farmer & Harnish 1987, p. 557). Animate entities are known to be more salient in general than inanimate entities (Ferreira 1994; Prat-Sala & Branigan 2000; cf. Nairne et al. 2013), and animacy requirements have been shown to be especially prone to trigger predictive processing (Wagers & Pendleton 2016; Wang et al. 2020). We conjecture that when the comprehender encounters a verb that requires its Theme to be animate, a robust prediction develops for the features of the upcoming DP that will realize this argument. This raises the salience of the two animate referents in the discourse and, by doing so, enhances the default for the predicate’s arguments to be disjoint. As a result, the reflexive advantage is largely neutralized. Put more succinctly, the verb’s argument structure jump-starts the default disjoint interpretation of the Theme before the object pronoun is encountered, preempting any processes that favor a reflexive interpretation.

Experiment 1 shows that Chamorro comprehenders can easily access the reflexive interpretation of ambiguous sentences containing an object pronoun, even in the absence of direct morphological signals of reflexivity. Reflexive interpretations do have at least two indirect cues. First, a reflexive interpretation forecloses the possibility of a PAH violation triggered by combining a pronoun object with a nonpronoun subject. Second, only overt object pronouns can be reflexive, so the use of the overt 3sg. object pronoun gui’ as opposed to the null object pronoun is possibly informative. A comprehender who tries to predictively interpret a sentence based only on a transitive verb and the object pronoun gui’ will find themselves effectively hemmed in by the unspoken alternatives.

What would happen if this conspiracy of constraints and unspoken alternatives was removed? In Experiment 2, we investigate pronoun possessors of the object. Possessor pronouns, like object pronouns, have no dedicated reflexive forms. In addition, because of Chamorro’s flexible word order, a possessor pronoun within the object can precede or follow the local subject. Possessor pronouns differ from object pronouns in that they (i) are not arguments of the verb, (ii) must be null, and (iii) are generally not constrained by the PAH.12 They therefore provide a neutral testing ground to assess whether the reflexive interpretation, in and of itself, has any privileged role in the comprehension of Chamorro pronouns.

Nouns in Chamorro can be preceded by an article and followed by a possessor. A possessed noun either agrees with the possessor in person and number (see (14a)) or else is joined to the possessor by a morpheme called the linker (14b).

When the possessor is a pronoun, the possessed noun must agree with it; there is no linker option. This means that possessor pronouns are cross-referenced by agreement in person, and therefore must be null; see (15).

Like object pronouns, possessor pronouns can serve as reflexive anaphors or as pronominals. Evidence for this is provided by the TPR. Recall that the TPR prevents a transitive verb from having a subject that is third person plural, but not a pronoun (see 2.1). This restriction is systematically evaded by clauses in which the object is a reflexive anaphor. In addition, the TPR is evaded by clauses in which a pronoun possessor of the object is bound by the local subject. The transitive clauses in (16) are grammatical, despite the fact that their subject is third person plural but not a pronoun, because their object’s possessor is a (null) pronoun that is bound by the subject.

Reflexive clauses of this type are grammatical whether the subject precedes or follows the object. As long as the object’s possessor is a pronoun bound by the local subject, the clause escapes the TPR. But if the possessor pronoun had been bound by a c-commanding antecedent in a higher clause, or had had a discourse antecedent or a nonlinguistic antecedent, the TPR would take effect as expected. This is evidence that the possessor pronouns in (16) are reflexive anaphors; they differ syntactically from possessor pronouns that are pronominals, even though the two are morphologically indistinguishable (since all possessor pronouns are null).

114 Chamorro speakers participated in Experiment 2, which took place in the CNMI in June 2019. They grew up in Saipan (58), Rota (51), Guam (5), and Tinian (1). Participants ranged in age from 18 to 76, with a median age of 47 years, and an inter-quartile range of 20 years. 37 participants reported that they had taken a previous experiment.

The experiment followed a within-subjects and within-items design. Each itemset contained four complex sentences in which a transitive verb was accompanied by overt subject and object DPs. The object contained a possessor pronoun that could either be interpreted as reflexive or else anteceded by the subject of a different clause. These sentences crossed two factors: (i) Local Subject (LclSubj)—whether the local subject preceded or followed the object; and (ii) Distant Subject (DistSubj)—whether the nonlocal subject, which always preceded the possessor pronoun, was in a higher clause or an adverbial clause. The local subject is always a potential binder for the object’s possessor, since it c-commands the object (and its possessor) whether the clause is VSO or VOS. A distant subject in a higher clause c-commands the possessor as well, and can therefore serve as a potential binder for it. A distant subject in an adverbial clause does not c-command the possessor, but it could serve as its discourse antecedent in our examples, because it always preceded the possessor. We included these two types of Distant Subjects in an attempt to tease apart whether the reflexive advantage we found in Experiment 1 was due to the presence of a local antecedent or a c-commanding antecedent. These two factors were confounded in Experiment 1: the local antecedents were exactly the c-commanding antecedents. In Experiment 2, Distant Subjects in a higher clause provided the comprehender with potential antecedents for the possessor that were nonlocal but c-commanding. If the reflexive advantage in Experiment 1 stemmed exclusively from the presence of a c-commanding antecedent, then these Distant Subjects should be more effective antedecents of the possessor than a Distant Subject in an adverbial clause.

In addition, each itemset contained matched control sentences in which participants could access only the interpretation in which the distant subject was the antecedent of the possessor. In these controls, the local clause was passivized, so that the DP containing the possessor was now the local subject. A possessor pronoun in this configuration cannot be bound by the passive agent (PassAgt), because the passive agent does not c-command it. (A separate Chamorro restriction prevents a pronoun in this configuration from having the passive agent as its discourse antecedent; see Chung 1989, 2020.)

32 itemsets were constructed which each realized the eight conditions. Thus 256 items were created in total, and were distributed via a Latin Square so that each participant heard 4 items per condition and no participant heard more than one item from the same itemset. Table 3 illustrates the itemset constructed from the verb phrase sugun i karetå-ña ‘drive her/his car’. All verbs were in the realis mood. The verbs in the adverbial clauses were either progressive (e.g. mamaigu’ ‘sleep (prog)’) or stative (e.g. gaigi ‘be (at a location)’); all other verbs were nonprogressive. The LclSubj and DistSubj were personal names that differed in their conventional gender association. Each pair of names was used twice, across two itemsets, and their order was swapped between these itemsets, so that LclSubj in item N became DistSubj in item N+16). Appendix A contains a list of the names and verb phrases.

Each trial began with a context sentence that served to introduce the two individuals named in the target sentence. The context sentence was played simultaneously with the display with a pair of “portraits” on the screen, as depicted in Figure 4. Contexts were constructed so that the possessor pronoun in the target sentence could plausibly refer to either individual. For the itemset shown in Table 3 and depicted in Figure 4, the context sentence was (17):

The target sentence was immediately followed by a question, which asked (i) which individual was the possessor (an Antecedent Question) or else (ii) which individual was performing the action denoted by the verb (a Thematic Role Question). We included Thematic Role Questions to ensure that participants were attending to the meaning of the sentence and not just to the identity of the possessor. Antecedent Questions and Thematic Role Questions were equally balanced across trials. The Antecedent Question for the itemset shown in Table 3 is given in (18a) and the Thematic Role Question is given in (18b).

Thematic Role Questions were counterbalanced across itemsets to ask about the event denoted by the local clause (e.g. driving in Table 3) or the event denoted by the adverbial clause (e.g. sleeping in Table 3). Participants selected their answer by touching the appropriate portrait.

Experiment 2 departed from Experiment 1 in the structure of its trials and its response selection method. In Experiment 1, the different potential antecedents for the object pronoun corresponded to different events that could be easily distinguished by means of line drawings (e.g. self-slapping vs. other-slapping), so it was straightforward to present side-by-side depictions of these events and ask participants to select the correct depiction. In Experiment 2, the different potential antecedents for the possessor pronoun often corresponded to different events that could not be reliably distinguished in this way. Although the events of Maria driving her own car and Maria driving Isidro’s car are different, it would be challenging for a single line drawing to capture the driving event and the car’s ownership simultaneously. For this reason, we defaulted to a simplified response style, in which participants merely responded to a question that probed the referent of the possessor directly. We would have preferred a response selection method which provided data about the time course of processing the target sentence and which was (arguably) less taxing for participants. But we were unable to come up with a feasible version of such a method.

The experiment was programmed using the OpenSesame environment (Mathôt et al. 2012) and deployed on Samsung Nexus 10 tablets. All linguistic stimuli and instructions were delivered in Chamorro, while participants listened over headphones. Each participant responded to 32 trials in total. There were no filler sentences in this experiment, given the variety of sentence structures in the target trials and given the different question types.

The primary analysis was two sets of regressions of reflexive possessor responses to the Antecedent Question, i.e. responses in which participants selected the portrait depicting the local subject of the target sentence. We also analyzed participants’ accuracy on the Thematic Role Question. In each of these analyses, we estimated a Bayesian multilevel logistic regression in the package brms (Bürkner 2017). Fully-nested random effects were estimated with an LKJ(2) prior on the random effects correlation (Lewandowski et al. 2009).

For analyses of the Antecedent Question, we first regressed reflexive choice on the two experimental factors (DistSubj, LclSubj) for the four core conditions in the experiment, in which the target clause was in the active voice. We then separately analyzed the four passive conditions, defined by the cross between DistSubj and PassAgt.

Finally, we made a programming error that led to unequal distribution of the lists, with the consequence that we did not collect equal amounts of data for all items in all conditions. In particular, for some items, we did not get any responses to the Antecedent Question for Higher Clause conditions. We took two precautions in the analysis, to make sure that the imbalance does not misleadingly influence our conclusions. First, we always report three rates when we report the Reflexive Interpretation rates (Table 4): the aggregate reflexive interpretation rate, the average of the reflexive interpretation rate over participants, and the average of the reflexive interpretation rate over items. For each average, we report the number of participants or items that went into that average, therefore making it clear how much data was lost. Second, we supplemented our analysis with a composite analysis, which split the dataset by the Distant Subject factor and regressed on the Local Subject factor separately in both cases.

Data from two participants were excluded from further analysis because they attained very low accuracy on Thematic Role Questions (38% and 42%). Accuracy was otherwise high (see Tables 5 and 6).

There were two main results from Experiment 2. The first is that comprehenders have a slight preference for the local subject as antecedent, but it is slight indeed. If we compare across the two experiments, we find that the preference for the local subject as antecedent in the active target conditions of Experiment 2 is comparable to the weakest preference for the local subject as antecedent shown in Experiment 1’s ambiguous Lcl Subj Pronoun conditions, i.e., when the verb required an animate Theme (Exp. 1: 62%, 95% CrI, 45%–76%; Exp. 2: 57%, 95% CrI, 50%–64%). Overall, we interpret this as evidence for a local antecedent advantage, albeit one qualified in strength and clearly defeasible in the face of other constraints.

The second result from Experiment 2 is the lack of convincing evidence that there is privilege for c-commanding antecedents over non-c-commanding antecedents – or, put another way, a preference for bound variable anaphora over discourse anaphora. Such evidence would come from a positive difference in reflexive interpretation rates between Adverbial Clause and Higher Clause conditions. The distant subject in the Higher Clause conditions c-commands the possessor pronoun, whereas the distant subject in Adverbial Clause conditions does not. If there were a preference for c-commanding antecedents over non-c-commanding antecedents, that should have rendered the distant subject a stronger competitor to the local subject in Higher Clause conditions, and we should have observed fewer choices of the local subject as antecedent in those conditions. (No such effect should have been observed in the Adverbial Clause conditions, given that the distant subject in those conditions does not c-command the possessor pronoun.) Average reflexive interpretation rates did conform to this prediction numerically: 58% in Adverbial Clause conditions v. 53% in Higher Clause conditions. However, the uncertainty around this difference reveals enough imprecision in the estimate to be circumspect (95% CrI, difference ranging from –3% to 14%). In a Bayes Factor analysis we directly compared the logistic regression reported above – which included Distant Subject as a factor – to a null model without it. This revealed moderate support for the null model (BF: 0.22).