On the universality of intrusive resumption: Evidence from Chamorro and Palauan

The literature on resumptive pronouns (RPs) has given rise to a rich taxonomy of the phenomenon. Despite the fact that RPs invariably have the morphosyntactic form of ordinary pronouns, they vary widely in distribution and function. In some languages, RPs are grammatically licensed; depending on the language and the syntactic context, they might or might not realize traces, compete with gaps, exhibit reconstruction effects, and so on. In other languages, notably English, RPs are ‘intrusive’ (Sells 1984). Kroch (1981), Asudeh (2004), Morgan and Wagers (2018), and others have proposed that intrusive RPs in English are ungrammatical products of the performance system—productions that satisfy local well-formedness but not global well-formedness. This account predicts that in every language, regardless of whether it has grammatically licensed RPs, intrusive RPs could also be found. Here we test this prediction against evidence from Chamorro and Palauan. Previous accounts have maintained that Chamorro does not have RPs and Palauan has only RPs. On the basis of corpus and elicited production data from Chamorro, and a re-examination of the Palauan evidence, we argue that both languages have grammatically licensed RPs, as well as intrusive RPs. Their grammatically licensed RPs differ in form and distribution. At least in Chamorro, the distribution of intrusive RPs produced is similar to that in English.


Introduction
The literature on resumptive pronouns (RPs) has given rise to a rich taxonomy of the phenomenon. RPs invariably have the morphosyntactic form of ordinary pro-B S. Chung schung@ucsc.edu 1 Stevenson Academic Services, University of California, Santa Cruz, 1156 High St., Santa Cruz, CA 95060, USA nouns in the language (McCloskey 2002(McCloskey , 2006. Despite this, they vary widely in distribution and function. According to Sells (1984), RPs in some languages, notably English, are intrusive, but RPs in other languages are grammatically licensed (bound by an A-bar operator). Depending on the language and the syntactic context, these grammatically licensed RPs might or might not realize traces (e.g. Engdahl 1982 on Swedish;Koopman 1982 on Vata;McCloskey 1990 on Irish), compete with gaps (e.g. Sichel 2014, exhibit reconstruction effects (e.g. Aoun et al. 2001 on Lebanese Arabic), exhibit weak crossover effects (e.g. Hendrick 2005 on Tongan), and so on.
An idea with a long history is that intrusive RPs are ungrammatical products of the performance system (Kroch 1981;Erteschik-Shir 1992;McCloskey 2017b;Morgan and Wagers 2018; see Goodall 2017 for a different view). One particular version of this idea, which we subscribe to, is that sentences that contain intrusive RPs are productions that satisfy local well-formedness but not global well-formedness, in the sense of Asudeh (2004, 2012) (cf. Tyler and Warren 1987. A more general version of the idea is that speakers sometimes use sentences that the grammar provides no legal derivation for. And indeed, there is a persistently observed disconnect between individual speakers' introspective judgments about the grammaticality of sentences that contain intrusive RPs and their seemingly commonplace occurrence. Suppose we take this disconnect seriously and assume that, whatever aspects of realtime language processing are responsible for it-for example, differential demands on memory systems (Wagers 2013;Chacón 2019)-those aspects do not vary across languages. Then we would expect the grammars of individual languages to differ in whether, and in what contexts, they have grammatically licensed RPs, but we would not expect a dichotomy between "true RP languages" versus "intrusive RP languages." Instead, we would expect intrusive RPs to be found in the productions of every language.
Here we test this prediction against evidence from Chamorro and Palauan, two Austronesian languages of Micronesia whose closest relatives within the Western Malayo-Polynesian subfamily are unknown. Chamorro and Palauan are head-initial, verb-first languages with null and overt pronouns. Both languages have A-bar dependencies with the hallmarks of wh-movement. These dependencies are sensitive to certain islands, exhibit strong crossover effects, and are flagged morphosyntactically by wh-agreement (Chung and Georgopoulos 1988). Previous work has placed these languages on opposite ends of the spectrum as far as RPs are concerned. Chung (1998) maintains that Chamorro has no RPs, whereas Georgopoulos (1985, 1991 maintains Palauan has only RPs, meaning that its A-bar dependencies are not created by wh-movement but rather involve syntactic variable binding of a base-generated (null or overt) RP by a base-generated A-bar operator. On the basis of new Chamorro evidence and a re-examination of Georgopoulos' Palauan evidence, we contend that both languages have grammatically licensed RPs as well as intrusive RPs.
In Sect. 2, we first review the evidence that some RPs are intrusive. We then survey the A-bar dependencies of Chamorro and Palauan and the evidence that they are created by wh-movement. Next, zooming in, we show that Chamorro grammar licenses (null) RPs in exactly one syntactic context: as the realization of traces of wh-movement of the possessor (Sect. 3). Then we examine data from an elicited pro-duction experiment we conducted in 2014 in the U.S. Commonwealth of the Northern Mariana Islands (CNMI), and corpus data from the Chamorro New Testament (NT) (Sect. 4). 1 These types of production evidence reveal that over and above its grammatically licensed RPs, Chamorro has intrusive RPs-RPs that are overt or null, produced far less frequently than grammatically licensed RPs but in a wider range of syntactic contexts, and highly similar in their distributional profile to intrusive RPs in English. Finally, turning to Palauan (Sect. 5), we propose a reanalysis of its A-bar dependencies that splits them into two types. One type of dependency is created by wh-movement and leaves traces, which are spelled out as overt RPs in exactly one syntactic context: when the trace is in the oblique case. The other type of dependency involves intrusive RPs-null or overt RPs inside islands and other contexts traditionally associated with intrusive resumption. Our reanalysis of Palauan RPs, which is based exclusively on Georgopoulos's data, is intended as a proof of concept and cannot conclusively settle the status of these RPs. Nonetheless, we show that our proposal has slightly greater empirical coverage than her account. It therefore challenges the exceptionality of Palauan as an "RP-only" language.
What emerges from the investigation is that Chamorro and Palauan are languages in which RPs arise in two ways, via grammatical licensing or-separately-as intrusive productions. We speculate about the larger significance of this result in the conclusion (Sect. 6).

Resumptive pronouns
The distinction between grammatically licensed RPs and intrusive RPs goes back to Sells (1984). For him, 'true' RPs are base-generated pronouns that are syntactically bound by an A-bar operator. These grammatically licensed RPs, he claims, occur in some languages (e.g. Swedish, Hebrew, and Irish) but not others (e.g. English). Although Sells does not discuss this, grammatically licensed RPs can also arise in a different way, as the morphological realization of traces of wh-movement. Grammatically licensed RPs of this second type pattern like wh-traces for the purposes of island constraints and other constraints on movement (see e.g. Koopman 1982:368 on Vata). According to Georgopoulos's analysis, Palauan's grammatically licensed RPs should be classified as 'true' RPs in Sells' sense (but see Sect. 5). The grammatically licensed RPs we discuss in Chamorro are realizations of wh-traces.
The study of intrusive RPs was initiated by Ross (1967) and Prince (1990). Ross (1967:432-434) observed that in "a dialect of English," relative clauses that contain an RP rather than a gap are grammatical and can violate island constraints. Prince, who characterized intrusive RPs in English as "officially ungrammatical...but...in fact not uncommon in speech " (1990:482), documented a number of their properties in relative clauses. Her study, which drew on corpus data collected by Anthony Kroch from various sources, "from casual conversation to classroom discourse to radio and TV newscasts to the New Testament" (Prince 1990:487), gave a semantic-pragmatic account of intrusive RPs in some relative clauses and attributed RPs in the others to "extralinguistic reasons-memory lapse, distraction-whatever reasons lead to disfluency in simple sentences " (1990:494). Asudeh (2012) refers to intrusive RPs as "processor resumptives," a term which highlights the idea that they can occur in expressions without being generated by any grammatical derivation. He further suggests that processor resumptives "are an instance of a processing strategy that is generally available cross-linguistically: a resumptive element, typically but not necessarily a pronoun, is inserted where a gap would lead to ungrammaticality or processing difficulty" (Asudeh 2012:41).
Within both of these general classes-grammatically licensed RPs and intrusive RPs-finer distinctions are necessary (cf. Sichel 2014), but the key idea underlying the notion of an intrusive RP is the same: sometimes speakers produce RPs to preserve local well-formedness (Asudeh 2012;Morgan and Wagers 2018;Chacón 2019), but the productions are themselves ungrammatical.
We can find a closely related but distinct view expressed by Shlonsky (1992), who describes "last resort" RPs as those that rescue an "otherwise ungrammatical derivation " (1992:443). RPs in Hebrew and Northern Palestinian Arabic have this character, but nonetheless, according to Shlonsky, they are inserted via languagespecific rules-the language has to allow pronouns to function resumptively in the first place. Safir (1986) gives an English example (the guy who we wondered whether *(he) was sane) in which the version with a gap would be ungrammatical, and the insertion of the RP rescues the derivation.
However, the idea that intrusive RPs serve to rescue otherwise ungrammatical derivations has been challenging to validate using acceptability judgments. Generally speaking, sentences containing RPs receive low acceptability ratings, even if the RP appears in a context that supports intrusion/rescue. Alexopoulou and Keller (2007) used magnitude estimation to establish the acceptability of RPs across a variety of structures in English, German, and Greek. They found that sentences containing RPs were never more acceptable than variants containing gaps. Many researchers have validated or extended their findings (e.g. Heestand et al. 2011;cf. Ackerman et al. 2018). Han et al. (2012) found that, at least for some island violations, gap-containing sentences are rated lower than RP-containing ones (specifically, for subject positions inside islands, i.e. classical ECP-violation configurations)-but no studies have found that sentences 'rescued' by RPs receive ratings characteristic of grammatical sentences.
While it has proved consistently challenging to get comprehenders to judge sentences with a (non-obligatory) RP as acceptable, it has proved consistently easy to get producers to utter, or write, sentences with an RP-either when they would otherwise have to produce an island violation (Ferreira and Swets 2005) or in situations of production complexity, such as long dependencies (Morgan and Wagers 2018). Generally speaking, filler-gap dependencies require the processor to encode, maintain, and periodically re-access the filler while simultaneously coordinating other parts of the utterance (Wagers 2013;Chacón 2019). If, during production, the processor loses the ability to re-access the filler or, indeed, 'forgets' that there even is one, then instead of producing a licit structure hosting a gap, they will produce something else-a production which, by hypothesis, will not flout local selectional restrictions. This provides a convenient view of why speakers sometimes produce RPs that are not generated by the grammar (Asudeh 2012).
Whether or not one adopts the specifics of this view, the claim that intrusive RPs are extragrammatical makes the strong prediction that they should be possible in every language. This prediction is consistent with several recent studies that have shown that the same complexity factors that promote the production of intrusive RPs in languages like English, which putatively has only intrusive resumption, are at work in languages with grammatically licensed RPs, like Hebrew or Irish. For example, Fadlon et al. (2019) found that Hebrew speakers produced RPs in direct object position, where RPs are optional, at a considerably higher rate in contexts that would render a relativized argument less accessible in memory (cf. Ariel 1999). In their experiments, such contexts included nonrestrictive relative clauses or relative clauses in which multiple NPs share semantic features. And surprisingly, even in Hebrew, such optional RPs are rated lower than their gap counterparts (Meltzer-Asscher et al. 2015). Mc-Closkey (2017b) arrived at a comparable result for direct object RPs in Irish, which have been considered optional. His corpus analysis shows that direct object RPs occur very infrequently in Irish; moreover, in almost all of these infrequent occurrences, the direct object RP is linked to an animate filler, a context that is well known to be difficult to process (Gennari et al. 2012). Finally, Hladnik (2015) shows that some types of RPs in Slovene are obligatorily conditioned by the syntactic environment. But, in a coordinated corpus and acceptability study, he identifies a distinct class of optional RPs that show the hallmarks of intrusive resumption.
These results, while far from conclusive, begin to suggest a consistent broader picture. We claim that all languages, including languages which have grammatically licensed RPs, may also have intrusive RPs. These intrusive RPs are produced either when a grammatically licensed continuation is impossible, or when a grammatically licensed continuation is possible, but the complexity of the expression leads the producer to lose track of the filler in a filler-gap dependency (Asudeh 2012;Chacón 2019). Expressions containing intrusive RPs will be relatively infrequent, often judged unacceptable, and not necessarily accurately interpreted by the comprehender. 2 The current investigation contributes to this line of inquiry by adding to the evidence that languages with grammatically licensed RPs can also have intrusive RPs. 2 Our discussion proceeds from the assumption that productions that are consistently judged by speakers to be degraded or ungrammatical, such as sentences containing intrusive RPs, are not generated by the grammar. Other initial assumptions are conceivable. For instance, one could assume that the grammar generates sentences containing intrusive RPs, but identifies them as highly marked. Proceeding from that alternative starting point, our claim would be that all languages have RPs that are routinely produced (and, by assumption, generated by the grammar) but not validated by speakers' introspective judgments of grammaticality.
It is sometimes claimed that intrusive RPs can facilitate comprehension (Hofmeister and Norcliffe 2013). However, there is no clear evidence that intrusive RPs are routinely interpreted as bound by the filler, and some recent research indicates they are often wrongly interpreted (Morgan et al. 2020). We further emphasize that there is language processing evidence that speakers treat (obligatory) grammaticallylicensed RPs differently from intrusive RPs. For instance, Hebrew comprehenders actively predict RPs inside island domains when there is an A-bar operator (Keshev and Meltzer-Asscher 2017), but English speakers do not (Chacón 2015(Chacón , 2019.

Morphosyntactic profile
Chamorro and Palauan are languages spoken in the Western Pacific. Palauan is one of the official languages of the Republic of Palau; Chamorro is one of the official languages of the U.S. Commonwealth of the Northern Mariana Islands (CNMI) and the unincorporated U.S. territory of Guam. Both are Austronesian languages and usually classified as members of the Western Malayo-Polynesian subfamily (which consists of some 500-600 languages; see Blust 2013:31). However, they are isolates within that subfamily, so they are not particularly closely related to other Western Malayo-Polynesian languages or to each other. This subsection gives thumbnail sketches of their morphosyntax. For more detail, see e.g. Topping (1973), Chung (1998) on Chamorro, and Josephs (1975, Georgopoulos (1991), and Nuger (2016) on Palauan.
Chamorro is a head-initial language that allows predicates of all major category types. When the predicate is a verb, the default word order is VSO, but the order of arguments after the verb is flexible. 3 DPs are inflected for unmarked, oblique, or local case. The language has a small inventory of determiners and prepositions. Verbs agree in person and/or number with the subject; this subject-verb agreement (in bold in (1)) also encodes mood and transitivity, and is realized as a prefix, infix, or proclitic. Nouns agree in person and number with the possessor; this possessor-noun agreement (underlined in (1) Palauan is a head-initial language in which predicates can be verbal or nominal. When the predicate is a verb, the word order of the clause is VOS (Georgopoulos 1991;Nuger 2016). Nonpronominal DPs begin with a (unglossed in the examples) or with a demonstrative. Palauan is sometimes described as having no case marking and just 3 Predicates that are adjectives pattern like verbs for the purposes of word order and agreement, so we use the term verb to refer to both in the text. 4 The examples cited are from published works, the unedited database for the Revised Chamorro-English dictionary (abbreviated CD), Nuebu Testamento (the Chamorro New Testament, abbreviated NT), orif not attributed to any source-from our Chamorro fieldwork. The following abbreviations are used in the morpheme-by-morpheme glosses. For Chamorro: AGR = agreement, ANTIP = antipassive, COMP = complementizer, EMP = emphatic, FUT = future, L = linker, LCL = local case, NM = unmarked case, OBL = oblique case, PASS = passive, PL = plural, PROG = progressive, Q = question, WH[OBJ] = object whagreement, WH[OBL] = oblique wh-agreement, WH[SBJ] = subject wh-agreement. For Palauan: COMP = complementizer, FUT = future, IM = imperfective, IR = irrealis, L = linker, NEG = negative, P = preposition/oblique case, PF = perfective, R = realis, RECIP = reciprocal, as well as 3S = third singular, 1P = first plural, etc. (for other persons and numbers). Palauan inflectional morphology is so complex that we did not try to develop a uniform set of glossing conventions for the two languages. Instead, we reproduce Georgopoulos' (1985Georgopoulos' ( , 1991 morpheme-by-morpheme glosses for the Palauan examples, with adjustments in the glosses of a handful of examples for uniformity. Note that Georgopoulos' glosses do not conform to the Leipzig Glossing rules. Readers should also note that orthographic y in Chamorro is a voiced alveolar affricate; orthographic ch in Palauan is glottal stop. one preposition, er (Georgopoulos 1985(Georgopoulos , 1991. Just as plausibly, one could describe the language as having no prepositions, but a radically reduced case system with just one overt case marker, the oblique marker er (cf. Nuger 2016). 5 Verbs agree in person and number with the subject; this subject-verb agreement (in bold in (2)) also encodes mood, and is analyzed by Georgopoulos as a prefix. In the perfective aspect, transitive verbs also agree with a direct object that is human or definite singular; this object-verb agreement (underlined in (2)) is realized as a suffix. (2) Palauan clause Te-chillebed-ii R.3P-PF.hit-3S a bilis dog a rengalek. children 'The kids hit the dog.' (Georgopoulos 1991:30, (15a)) Both languages have null pronouns as well as overt pronouns, and regulate the distribution of these types of pronouns in similar ways. There are two basic generalizations. First, Chamorro pronouns that are objects of prepositions must be overt; the same is true of Palauan pronouns marked with er (which, depending on the analysis, are either objects of prepositions or else DPs in the oblique case). Second, pronouns that control agreement in person must be null (Georgopoulos 1991:26, 30-31, 43-44, 52-54, Chung 1998. In Chamorro, some types of agreement involve person and number, and others involve number alone. Subjects of transitive verbs, subjects of irrealis verbs, and possessors control person-and-number agreement, so pronouns occupying these positions must be null. In Palauan, all agreement involves person and number. Subjects, direct objects in the perfective aspect, and possessors control agreement, so pronouns occupying these positions must be null. 6 (Null pronouns are not represented in the language material in the examples, except that null RPs-like other null gaps-are represented by an underscore.) (3) a. Chamorro possessor pronoun Agupa' tomorrow i the kumplihaños-ña birthday-AGR (*guiya). s/he 'Tomorrow is his birthday.' (Chung 1998:31, (21b) a bilis dog (*ngii). s/he 'S/he hit the dog.' (Georgopoulos 1991:26, (7c)) Over and above this, Chamorro has a third requirement: inanimate pronouns that are not reflexive must be null. Palauan has no similar requirement; its inanimate singular pronouns can be null or overt. The fact that the distribution of null versus overt pronouns is regulated to this extent will become important later.

Evidence for wh-movement
The A-bar dependencies of Chamorro and Palauan have been discussed extensively (see e.g. Chung 1982, 1998and Georgopoulos 1985, 1991. They occur in constituent questions, relative clauses, and focus constructions in Chamorro, and in various Palauan constructions analyzed by Georgopoulos (1991) as involving relative clauses. As noted earlier, these dependencies satisfy many of the classic diagnostics for wh-movement-that is, for A-bar movement in constituent questions, relative clauses, focus constructions, and similar dependencies (see Chomsky 1977 and many subsequent works). We illustrate this here for constituent questions in Chamorro and relative clauses in Palauan.
To begin with, the dependency in these constructions holds between a displaced constituent, or filler, and the position from which it is displaced, which we refer to pretheoretically as a gap. Consider the following examples, in which the filler and gap are co-indexed, the filler is underscored if it is overt, and the gap is represented by an underscore if it is null. In the Chamorro embedded question in (4), the filler is the interrogative phrase kuåntu mås na tåotao 'how many more people' and the gap is the subject of the verb hunguk 'hear'.  'I don't know how many more people have heard it.' In the Palauan relative clause in (5a), the filler is the head NP chad 'man' and the gap is the subject of the verb omechar 'buy'. Palauan also has null-headed relative clauses-relative clauses in which the filler is a null NP that can be translated 'one, thing.' In the null-headed relative clause in (5b), the gap is the direct object of the verb meleseb 'burn'. (Null-headed relative clauses do not have an overt complementizer. In the null-headed relative clauses in the examples, the gap is annotated with an index but the filler is not represented.) garden-3s 'the flowers that my mother thought Mary was growing in her garden' (Georgopoulos 1991:91, (52)) Third, the dependency between filler and gap is sensitive to islands. In Chamorro, these islands include embedded questions and relative clauses (Chung 1998:208-219)-both ordinary relative clauses and so-called existential relative clauses, which modify the pivot of an existential sentence and have special complementizer forms (Chung 1987 In Palauan, the relevant islands include adjunct clauses and coordinate structures (Georgopoulos 1991:107-108, 115-117; see Sect. 5 on the status of relative clauses).
The null-headed relative clauses in (9) (Georgopoulos 1991:107, (7)) Fourth, the dependency exhibits strong crossover effects. In both languages, a gap can be co-indexed with a pronoun that it c-commands, but the reverse is not possible: a gap cannot be co-indexed with a pronoun that c-commands it. In (10a) and (11a), the gap is the matrix subject and the co-indexed pronoun is inside the embedded clause; in the ungrammatical (10b) and (11b), the gap is inside the embedded clause and the co-indexed pronoun is the matrix subject. (The gap in (11b) is realized as an RP; see Sect. 5. Evidence that the gap in this example is in the embedded clause as opposed to the matrix clause is provided by wh-agreement; see the discussion below.) 'Who i is the one who thinks that I respect her i ?' (Georgopoulos 1991:192 pro i ]? ('Who i is the one who she i thinks I respect?') (Georgopoulos 1991:192, (25b)) Taken together, these diagnostics make a good case that Chamorro constituent questions and Palauan relative clauses are created by wh-movement. 8 Over and above this, the two languages have special morphological agreement that signals the presence of a gap created by wh-movement. This agreement, which is known as wh-agreement, will play a key role in the discussion below. It is realized on the verb, preempts the normal subject-verb agreement, and registers the grammatical relation (Case) of the gap (Chung 1982(Chung , 1991a(Chung , 1994(Chung , 1998Georgopoulos 1985Georgopoulos , 1991Chung and Georgopoulos 1988;Baier 2017).
In Chamorro, wh-agreement employs morphology otherwise found in various nonfinite forms to register whether the gap is a subject, direct object, or oblique (Chung 1994). When the gap is the subject of a realis transitive verb, wh-agreement is realized as the infix -um-, which is also found in transitive infinitives. (Note that this overt realization of subject wh-agreement is obligatory in questions and the focus construction, but optional in relative clauses.) When the gap is the direct object of a transitive verb, wh-agreement is realized-optionally-by the combination of nominalizing the verb plus the infix -in-. When the gap is an oblique, wh-agreement is realized by simply nominalizing the verb. (Nominalized verbs agree with their subjects via possessor-noun agreement.) See (12), where the verbs that show wh-agreement are in bold. In Palauan, wh-agreement employs morphology that otherwise encodes mood to register whether the gap is a subject or nonsubject (Georgopoulos 1991:84-90). When the gap is the subject of a realis clause, wh-agreement is realized as the realis form of the verb minus the normal subject-verb agreement (Baier 2017). When the gap is a nonsubject of a realis clause, wh-agreement is realized as the irrealis form of the verb, including irrealis subject-verb agreement, but the clause is semantically interpreted as realis. 9 (13) Palauan wh-agreement a. 'the book that the man bought' (Chung and Georgopoulos 1988:255, (10)) An important property of wh-agreement is that it is local. Wh-agreement on higher verbs of a long-distance dependency does not register the grammatical relation (Case) of the gap, but rather the Case of the local CP argument that contains the gap. To see this locality at work, consider the Chamorro long-distance questions in (14). Both involve wh-movement out of the CP complement of the verb tungu' 'know'. The gap in (14a) is a subject, so the embedded verb (såkki 'steal') shows subject whagreement; the gap in (14b) is a direct object, so the embedded verb (hunguk 'hear') shows object wh-agreement. But the higher verb tungu' (in bold) shows object whagreement in both examples, because the CP argument that contains the gap serves 9 The semantic distribution of realis vs. irrealis mood in Palauan is regulated as follows. Certain types of clauses are always in the irrealis mood; namely, negated clauses that are complements of the negative verb diak 'not', antecedent clauses of conditional sentences, certain adverbial clauses, and imperatives (Josephs 1975: Chapters 18 and 19;Georgopoulos 1985:77, fn. 19). All other types of clauses, including e.g. future clauses, are always in realis mood, but their mood morphology can be overridden by wh-agreement. (The mood morphology of clauses that require the irrealis mood, such as negated clauses, cannot be overridden by wh-agreement; see Georgopoulos 1991:89-90 andJosephs 1975:375, (28a) and ( 'How many complaints about himself do you know that Juan heard?' The same pattern can be observed in Palauan, as can be seen from the longdistance relative clauses below. The null-headed relative clause in (15a) involves wh-movement out of the CP complement of omes 'see'; the headed relative clause in (15b) involves wh-movement out of the CP complement of omdasu 'think'. The gap in (15a) is a subject, so the embedded verb (menguiu 'read') shows subject whagreement; the gap in (15b) is a direct object, so the embedded verb (omekeroul 'grow') shows nonsubject wh-agreement. But once again, the higher verb (in bold) shows nonsubject wh-agreement in both examples, because the CP complement that contains the gap serves as this verb's direct object.

(15)
Palauan wh-agreement in long-distance relative clauses a. a [l-ilsa garden-3s 'the flowers that my mother thought Mary was growing in her garden' (Georgopoulos 1991:91, (52)) We claim that in both languages, wh-agreement realizes agreement of T, or perhaps C, with an argument that is, or contains, a wh-trace (see the references cited earlier for specific proposals). The idea that wh-agreement is agreement with a wh-trace will be crucial in what follows.

Grammatically licensed RPs in Chamorro
Chamorro grammar does not license overt RPs (Chung 1998:210, 217, 221). The A-bar dependencies in constituent questions, relative clauses, and focus constructions have a gap that must be null; see the examples cited earlier and (16a). This gap cannot be filled by an overt pronoun, whether it is an independent pronoun (as in (16b)) or a weak pronoun (as in the relative clause in (16c)). However, given that the language has null as well as overt pronouns, it is reasonable to ask whether Chamorro grammar countenances RPs that are null. (Following a suggestion by Boris Harizanov, we assume that null RPs differ from wh-traces in that a null RP is a syntactic object distinct from the filler, whereas a wh-trace is (a different instance of) the same syntactic object as the filler.) Here we show that null RPs are grammatically licensed, and must occur, in exactly one syntactic context: when the gap is created by wh-movement of a possessor.

Wh-movement of possessors
In Chamorro, DPs formed from common nouns consist of a determiner, its NP complement, and an optional possessor DP, which are linearized in that order. We assume that the possessor occupies the specifier of D, so that the hierarchical structure of DP, ignoring left-to-right order, is as shown below.
Possessors can be realized morphosyntactically in two ways. Either they appear in the unmarked case and control person-and-number agreement on the possessed noun, as shown in (18a), or else they are 'joined' prosodically to the possessed noun by the morpheme called the linker, as in (18b). Possessors that are pronouns must control person-and-number agreement on the possessed noun; they cannot make use of the linker option. Since pronouns that control agreement in person must be null (see Sect. 2.2.1), it follows that possessor pronouns are always null. 10 Chamorro grammar allows A-bar dependencies in which the gap is a possessor. These A-bar dependencies are created by wh-movement. For instance, the distance between the filler and the gap is apparently unbounded, but sensitive to islands, as can be seen from (19) A-bar dependencies of type (19) also exhibit wh-agreement. In (19a), the higher verb malagu' 'want' shows oblique wh-agreement, as expected, because this verb's oblique complement is the CP that contains the possessor gap. 11 Although Chamorro allows wh-movement of possessors, this movement is also tightly constrained. In order for it to occur at all, the possessive DP-the DP that has the possessor as its specifier-must satisfy three conditions (Chung 1991b(Chung :107-110, 1998. We call these conditions the argument structure condition, the locality condition, and the D condition. First of all, the possessive DP must be an internal argument. More specifically, it must be the direct object of a transitive verb or the subject of a passive verb, unaccusative verb, or nonverbal predicate (i.e. AP, DP, or PP). This argument structure condition can be seen at work in the questions of the possessor in (20). These questions are grammatical when the possessive DP is a direct object (as in (20a)) or the subject of an unaccusative verb (20b), but not when it is a locative DP (20c), even when the locative DP is the argument of an existential verb (20d). (20) Wh-movement of Chamorro possessors: argument structure condition a. Second, the possessor must be the specifier of this internal argument; it cannot be more deeply embedded within that specifier. This locality condition lies behind the contrast between the questions of the possessor in (21). In the grammatical (21a), the possessor is the specifier of the internal argument of the existential verb. But in (21b-c), the possessor is the specifier of the specifier of the internal argument, and the result is ungrammatical. These conditions arguably follow from the hypothesis that the A-bar dependencies in (19)-(22) are created by wh-movement. 13 For instance, wh-movement respects the Phase Impenetrability Condition (Chomsky 2001). The locality condition illustrated in (21) follows from this plus the fact that DPs in Chamorro are phases (Chung 1991b(Chung :107-110, 1998. Wh-movement out of DPs is known to exhibit a specificity effect, meaning that movement out of specific DPs is degraded (e.g. Chomsky 1977;Fiengo and Higginbotham 1981); some language-specific versions of this effect also distinguish between overt and null determiners (Karimi 1999). The D condition illustrated in (22) is a specificity effect of this second type. Finally, it is widely believed that complements and their subconstituents are more freely accessible to whmovement than are noncomplements and their subconstituents (e.g. Chomsky 1981;Huang 1982). This broad generalization, however it is accounted for, lies behind the argument structure condition illustrated in (20). 14 All this could be taken to suggest that the argument structure condition, the locality condition, and the D condition flow from properties of universal grammar and therefore do not have to be stated separately as part of Chamorro grammar. Nonethe-12 The D condition seems to be enforced less strictly for the definite article i than for other overt articles, demonstratives, and quantifiers. We have encountered some speakers who allow i in examples like (22b); see also Sect. 4.2. We note without further comment that the null indefinite article can be used felicitously in contexts in which it is common knowledge the possessee is unique relative to the possessor (Chung 2018). 13 In versions of Principles and Parameters theory that include the ECP or a head government requirement, all three conditions would follow directly from the hypothesis that these dependencies are created by wh-movement. In minimalist syntax, the locality condition follows directly, but it is not clear how the specificity effect and the restrictions on extraction of, and out of, specifiers are best handled. 14 Note, though, that sentential subjects, and subjects more generally, are not islands in Chamorro (Chung 1991a(Chung , 1991b). less, we will continue to refer to them as if they were separate conditions, because that will facilitate the exposition in Sect. 4. What is important here is that speakers have firm introspective judgments that sentences of the types shown (20c-d), (21bc), and (22a-c) are ungrammatical (though see fn. 12 on (22b)). 15 In other words, the possessors that can legally undergo wh-movement in Chamorro are restricted to a relatively narrow distributional slice of the possessors in the language.

Possessor gaps as null RPs
We turn next to the evidence that these possessor gaps, though created by whmovement, are realized as null RPs .
Chamorro grammar systematically bans transitive clauses with certain combinations of subject and direct object (Chung 1998(Chung , 2014Clothier-Goldschmidt 2015). Among the excluded clause types are transitive clauses whose direct object is higher than the subject on the person-animacy hierarchy shown in (23). We refer to the restriction that enforces this ban as the PAH. 16

(23)
Chamorro person-animacy hierarchy 2 > 3 animate pronoun > animate > inanimate The PAH is reminiscent of constraints found in languages with direct-inverse systems (Aissen 1997(Aissen , 1999, and much literature since). At the same time, it exhibits arbitrary quirks that are arguably more characteristic of post-syntactic morphological spellout than of operations in narrow syntax. First person pronouns, both exclusive and inclusive, are unaffected by the PAH; so are clauses whose verbs show any type of agreement other than the default subject-verb agreement. We gloss over these details; see Chung (2014) for discussion. What is relevant here is how the PAH responds to (second and third person) pronouns. When the direct object of the clause is a second person pronoun, the subject must be first or second person as well. The subject cannot be third person, whether a pronoun or nonpronoun, because then it would be lower than the direct object on the person-animacy hierarchy. See (24) 15 The data discussed in this section were provided over the years by nine Chamorro speakers-six in Saipan and three (originally from Guam) in California. Their current age range would be 60-77 years. 16 The PAH is occasionally violated by speakers in elicitation contexts and can be outcompeted by other pressures in comprehension ). However, Clothier-Goldschmidt's (2015) corpus study of connected discourse has shown that it is a hard constraint in production. Similarly, when the direct object is a third person animate pronoun, whether overt or null, the subject must be an animate pronoun as well; see (25a). The subject cannot be a nonpronoun, because then it would be lower than the direct object on the person-animacy hierarchy; see (25b). 17 However, when the direct object is an inanimate pronoun, the PAH imposes no restrictions on the subject, as (25c) is intended to suggest. (The object pronoun in (25c)  Although the PAH is sensitive to the featural content of the subjects and objects it compares, it is generally insensitive to the features of their subconstituents. This can be seen from the clauses below, each of which contains a DP with a second person possessor. The PAH excludes (26a) because the direct object is second person but the subject is third person; it has no effect on (26b) because the direct object is inanimate. Had the features of the possessor been taken into account, the results would have been reversed. There is, however, one curious exception to this generalization. Exactly when the direct object DP has the null indefinite (nonspecific) article as its D and a possessor that is a pronoun, the PAH ignores the direct object's features, and instead compares the features of the subject with the features of the direct object's possessor (Chung 2014). Consider the clauses in (27), which have direct objects of this type. The PAH excludes (27a) because the possessor of the direct object is second person but the subject is third person; it excludes (27b) because the possessor of the direct object is an animate pronoun but the subject is a nonpronoun. Crucially, this pattern emerges only when the D of the direct object is the null indefinite article; compare (27)  These rather intricate patterns could be handled in various ways. The account that we sketch here relies on the following assumptions, most of which are uncontroversial. First, the PAH relies on agreement operations-either the Agree of narrow syntax or comparable morphological agreement operations in the post-syntax-that cause T to be valued for features of the subject and small v to be valued for features of the direct object. . Second, we claim that T and small v probe for the feature [Specific]-more precisely, when they encounter a head that is [Specific], they record this feature and as many of the other features just mentioned as that head bears. Subjects of transitive clauses in Chamorro must be specific (see fn. 18). This means that when T probes, it will invariably be satisfied by the subject's [Specific] feature and will be valued for the subject's other relevant features as well. But when small v probes, it may encounter a direct object that is nonspecific. In such cases it 18 For reasons of space, we observe but do not discuss the following. First, the PAH never accesses the features of the possessor of the subject. This is because external arguments must be specific, so the subject DP of a transitive clause cannot have the null indefinite (nonspecific) article as its D (Chung 1998:104). Second, although clauses like (27a) are occasionally accepted by speakers in elicitation contexts, a corpus search of the Chamorro New Testament reveals that clauses of this type-with a third person subject and a DP object with a null indefinite D and a second person possessor-are systematically absent. As is usual with PAH violations, the corresponding passive clauses occur instead. In other words, even in its interaction with possessors, the PAH is a hard constraint in production; see fn. 16. will probe further, into the direct object's specifier (= the possessor), and-if that DP is specific-will be valued for its relevant features. 19 The result of these operations is that all the featural information required by the PAH is encoded on two heads in the clausal spine-T and small v. We propose that these two heads share this featural information, in such a way that the features of the subject, on the one hand, and the features of the direct object or its possessor, on the other, are kept distinct (Anderson 1992). The PAH can then be implemented as a set of well-formedness conditions on the complex featural make-up of one or the other of these heads; see Chung (2014) for one specific proposal.
We are now ready to return to the main point. The subcase of the PAH illustrated in (25), which involves pronoun direct objects and nonpronoun subjects, can be used as a diagnostic for whether the gaps in various types of Chamorro A-bar dependencies are pronouns-more precisely, null RPs. Consider wh-questions of the direct object, such as (28). The gap in this example is construed as animate, so if it were also a pronoun, the PAH should require the subject to be an animate pronoun as well. The fact that no such requirement is imposed argues that this gap is not a null RP, but rather a wh-trace (see e.g. Chomsky 1981 on the nonpronominal character of whtraces). The situation is different when the gap is the possessor of the direct object (and therefore the D of the direct object is the null indefinite article), as can be seen from the whquestions of the possessor in (29). Then the pattern attributed to the PAH re-emerges: when the gap is construed as animate, the subject must be an animate pronoun (see (29a-b)), but when the gap is construed as inanimate, no such restriction is imposed (see (29c)). Speakers unequivocally judge questions of type (29b) to be ungrammatical, but questions of type (29c) to be well-formed. We take this as evidence that in wh-movement of the possessor, the possessor gap is a null pronoun. 20 The conclusion can be sharpened a bit. Recall that the PAH is quite likely imposed in the post-syntax, as part of morphological spell-out. Accepting this, what the evidence reveals is that the possessor gap is realized as a null RP at this relatively late stage of the derivation.
At this point, the question arises of how the wh-trace of a Chamorro possessor is realized as a null RP. The more general theoretical question is how wh-traces can come to be spelled out as RPs in any language. Although we cannot answer this question definitively, the Chamorro material just discussed suggests the following scenario (see McCloskey 2017a for a different view). In general, the wh-traces that are realized as RPs form the tail of an A-bar dependency and can be identified by grammatical relation (Case; see fn. 26). Let us assume that wh-traces are instances of the same syntactic object as the moved constituent, but pronouns-including RPs-are distinct syntactic objects. Then for a wh-trace to be realized as an RP, some operation in the post-syntax must detach the lowest instance of the moved constituent-the instance in the position where the moved constituent's Case is checked-so that it forms a distinct syntactic object. We conjecture that the syntactic object so formed is a 'minimal' copy of the moved constituent: it consists of just a category feature, a referential index, and the features otherwise associated with pronouns in the language (cf. Harley and Ritter 2002). In Chamorro, the detachment operation affects all and only moved constituents that are possessors, i.e. bear genitive Case. Suppose now that when T and v probe for the feature [Specific], as we proposed above, they do so in the post-syntax, where they recognize 'minimal' copies as pronouns. Then when the PAH is imposed, as part of morphological spell-out, it will produce the pattern seen in (29). Otherwise, 'minimal' copies will be spelled out in the same way as other pronouns. In Chamorro, this spell-out will have the result that possessor RPs-like all possessor pronouns-are null.
To sum up: Chamorro grammar limits the possessors that can undergo whmovement to a small distributional slice of the possessors in the language. We claim that the possessor gaps left by this movement are realized as null RPs.

Chamorro RPs in performance
We are now in a position to ask whether-over and above these grammatically licensed null RPs-Chamorro has intrusive RPs. Speakers' introspective judgments do not provide a good way of getting at this question, so we turn to evidence from production. Section 4.1 discusses evidence from an elicited production experiment conducted in the U.S. Commonwealth of the Northern Mariana Islands (CNMI) in 2014; Sect. 4.2 discusses corpus evidence from the Chamorro New Testament (Camacho et al. 2007). Both types of evidence reveal that Chamorro speakers do indeed produce RPs even when they are not grammatically licensed.

Experimental evidence
Together with our Chamorro team member, Manuel F. Borja, we conducted an elicited production experiment on Chamorro in the three inhabited islands of the CNMI (Rota, Saipan, and Tinian) in September 2014. Our aim was to learn which types of relative clauses speakers would produce when performing a simple description task. Relative clauses have various types of realizations in Chamorro: the relative clause can be prenominal or postnominal, the head NP can be overt or null, the verb of the relative clause might or might not show overt wh-agreement, and so on .
The method we used involves a guided conversation in Chamorro between the participant and the experimenter (Borja). Each participant is seated across from the experimenter and presented with a sequence of pairs of pictures-14 pairs in all. Each pair of pictures depicts actions involving two characters: one picture shows character X acting on character Y, while the other picture shows Y acting on X. One of the characters wears different identifiers-for instance, a lei or a hat-in the pictures. In Fig. 1, for instance, one picture shows a gecko jumping over a frog wearing a hat, and the other picture shows a frog wearing a lei jumping over a gecko. This feature is adapted from other production experiments which were designed to elicit relative clauses (e.g. Gennari et al. 2012).
In dialogue with the experimenter, who cannot see the pictures, the participant describes the picture in which the character is wearing a particular identifier-for instance, the picture in which the frog is wearing a lei. Participants were instructed not to point at the pictures and not to use location or direction terms (e.g. kåttan 'north', akakgui 'left'). The dialogue was digitally recorded. 21 Forty-four Chamorro speakers served as participants: 24 from Rota, 19 from Saipan, and 1 from Tinian. Twenty-six of the participants were female and 18 were male; they ranged in age from 18 to 70 years. Data from one participant who was not a native speaker was excluded; the median age of the other participants was 44 years.
Among the results that emerged from the experimental data was one we did not expect. Of some 728 relative clauses that participants produced, 38 (5.2%) are not grammatically licensed, meaning that they contain an RP that violates the conditions on wh-movement in Chamorro. These productions plausibly involve intrusive RPs, either overt or null. More specifically, they include: -One relative clause with an overt RP (in bold) in the local case, cited in (30). 22  beside-AGR 'that woman i that there's a shell on her i or beside her' (R002/B1) -Nine relative clauses with a null RP inside an island, where the island is either a relative clause (1 production), an existential relative clause (6 productions), or a coordinate structure (2 productions). Representative examples are cited in (31). We note that given the restrictions on how Chamorro pronouns are realized (see Sect. 2.2.1), just one of these nine RPs could potentially have been overt. Seven must be null, because they are cross-referenced by agreement in person; one is null because it refers to an animal (see (31c)), and animals that are not personified are treated as inanimate.  , and allows null RP possessors only in the narrow range of contexts delimited by the argument structure condition, the locality condition, and the D condition. Therefore, the types of RP productions illustrated in (30)-(32) are not grammatically licensed (i.e. they are ungrammatical). Their frequency in our data was initially surprising to us. We think what may have happened is that our experimental method unintentionally replicated certain aspects of the method of Ferreira and Swets (2005), whose production experiments were explicitly attempting to elicit island violations and RP repairs. In their case, this involved explicitly triggering the production of embedded wh-questions. In our case, the discriminating features of our illustrations may have been spatial in nature, and thus locative PPs may have been frequently used to plan a response (even though, for each scenario, we had transitive verbs in mind that we thought could have been used). Ferreira and Swets (2005) found that, when speakers produced RPs in islands, production latencies increased (speakers produced the utterance at a slower rate). They concluded that the production of such utterances was correlated with greater processing demands. While the design of our experimental task did not allow us to systematically measure production latencies, we were able to code speakers' productions for disfluencies, e.g. pauses, incomplete words, or hesitations followed by a restart. Interestingly, we found that disfluencies were more common in utterances in which at least one relative clause contained an intrusive RP. Table 1 gives a crosstabulation of utterances containing relative clauses by the presence vs. absence of apparent disfluency and the presence vs. absence of (at least one) intrusive RP. (Utterances with an 'apparent disfluency' included a noticeable pause or restart, indicated by . . . in the examples cited above.) Disfluencies were more likely when an intrusive RP was produced: a (slim) majority of utterances containing an intrusive RP included at least one disfluency, whereas there were relatively fewer disfluencies in utterances that contained no intrusive RPs (log-likelihood ratio test of independence, G(1): 4.3, p < .05).
This correlation provides us with indirect evidence about the status of the utterances that we claim involve intrusive RPs. Even though judgments from fieldwork establish that wh-movement in Chamorro is island-sensitive, and wh-movement of possessors is quite restricted, it remains a possibility that our participant sample included individuals with more permissive grammars. We do not know how our participants would have judged the acceptability of their own productions, as this was not part of the experimental design. But the fact that they produced more disfluencies in utterances containing an intrusive RP suggests that they found such utterances unnatural or complex in some way, or perhaps more processing resources were required to produce them.
A different kind of indirect evidence suggests that the null intrusive RPs illustrated in (31)-(32) are indeed pronouns, not traces of a more permissive wh-movement. Recall from Sect. 3.2 that wh-traces count as nonpronouns, but RPs count as pronouns, for the PAH. If the gaps illustrated in (31)-(32) were wh-traces, they should be able to occur as animate (possessors of) direct objects even when the subject is a nonpronoun. Interestingly, productions of this type are absent from the experimental data. For instance, the data include 17 productions in which an animate null RP is the possessor of the internal argument of a transitive verb root. All these productions conform to the PAH as if the RP is a pronoun. Either the clause is passive, so the PAH is evaded, or else the clause is transitive and its subject (the external argument) is also a pronoun. In the rightmost clause of the relative clause in (31d), for instance, the null RP is the possessor of the direct object, but the subject is also a pronoun, so the PAH is respected whether the RP (which refers an animal) is treated as animate or inanimate.
Finally, the distributional profile of the intrusive RPs in the experimental data closely tracks what previous research has established about the distribution of intrusive resumption in English and other languages. This research indicates the following four generalizations. First, intrusive RPs are commonly found in relative clauses, especially nonrestrictive relatives (Prince 1990). (Whether intrusive RPs occur to the same extent in other A-bar dependencies, such as constituent questions, has not been systematically investigated.) Second, intrusive RPs are more common inside islands (Ross 1967;Morgan and Wagers 2018) and in other contexts in which the RP is syntactically 'distant' from the filler (Erteschik-Shir 1992; Hofmeister and Norcliffe 2013). Third, intrusive RPs occur when the RP is focused (Sichel 2014); see the discussion in Sect. 4.2 below. Fourth and finally, an intrusive RP can be replaced by a DP with descriptive content which is referentially related to the filler, and which can be interpreted as containing a semantic variable bound by the filler (Prince 1990;McCloskey 2017b). Cf. Prince's (1990:492) example You assigned me to a paper which I don't know anything about the subject.
All 38 of the ungrammatical RP productions in the data from this experiment are found in relative clauses. Although this fact is consistent with the overall profile of intrusive resumption, we note that our experimental method did not elicit other types of A-bar dependencies. More significantly, some of these RP productions are embedded in an island that separates them from the filler; many occur in other contexts in which they are syntactically distant from the filler. Finally, consider the production in (33), a null-headed relative clause that describes a picture in which the identifier-a shell-appears among taros displayed for sale in a shed. A young man is sitting on the ground below the shed; the taros and the shell are visible above him. One taro has fallen from the display and is about to hit his head. In the relative clause, the filler (a null NP) refers to taros; if there were a gap or RP, it would be the intrusive null RP possessor of the locative DP påpa' 'bottom, under, below'. This intrusive RP is replaced by a DP with descriptive content (soko' 'shed, storage place', in bold) that is referentially related to the filler. Just as in Prince's (1990) and McCloskey's (2017b)  shed '[the taros that the shell is near are] those (taros i ) that a boy is sitting below the (taros i ) shed' (S002/Q1) These similarities strengthen the case that the ungrammatical RPs in (30)-(32) are intrusive RPs-not grammatically licensed but rather artifacts of performance.

Corpus evidence
The claim that the Chamorro RPs discussed in Sect. 4.1 are extragrammatical productions leads us to expect that they will also be produced in contexts outside the experimental setting. Corpus evidence reveals that this expectation is borne out.
The largest searchable corpus of Chamorro that we have access to is the Chamorro New Testament (Nuebu Testamento, henceforth NT), which was published in 2007. NT is a translation into Chamorro of the New Testament chapters of the New American Bible. It is the result of a multi-year project undertaken by a small group of fluent Chamorro speakers in the CNMI, led by the late Bishop Tomas A. Camacho, who was a prolific Chamorro orator, writer, and translator. NT is an excellent example of a somewhat formal register of the contemporary Chamorro language.
We mined the electronic version of NT for instances of RPs, using a two-pronged strategy that combined two systematic searches for null RP possessors with the collection of other RPs encountered by chance in our searches for various types of relative clauses. The searches for null RP possessors were made possible by the fact that possessor-noun agreement is written with a hyphen in standard Chamorro orthographies. We searched systematically for the third person possessor agreement suffixes -ña '3SG' and -ñiha '3PL', and then identified by hand the subset of possessors that were null RPs. We were particularly interested in null RP possessors because they are grammatically licensed in some syntactic contexts but not others. Our first search for RP possessors was limited to relative clauses, which are more frequent than other Abar dependencies in this corpus. In a second search, we rechecked the relative clause data and extended our data collection to null RP possessors in constituent questions and focus constructions. 24 24 In the focus construction, a syntactically focused DP or PP appears at the left edge of CP. Most instances of the focus construction in Chamorro have two possible analyses, one involving wh-movement of the syntactically focused constituent and the other involving a cleft construction in which the focused constituent is a predicate whose subject is a null-headed relative clause. However, in some cases, only one analysis is possible (Chung 1998:268-275). The example of the focus construction cited later in (36) has only a wh-movement analysis. A cleft analysis is not possible for this example, because the focused constituent, maseha håyi 'whoever', does not occur independently as a predicate.
Our systematic search for null RP possessors uncovered 164 instances of restrictive or nonrestrictive relative clauses with possessor gaps. The null RPs in 147 of these relative clauses are grammatically licensed; that is, their possessive DP satisfies the argument structure condition, the locality condition, and the D condition The null RPs in 17 relative clauses are not grammatically licensed, because their possessive DP violates at least one of the three conditions. Four violate the argument structure condition, because the possessive DP is a locative (3 instances) or an external argument (1 instance); two violate the locality condition, because the gap is embedded within the specifier of the possessive DP; eleven violate the D condition, because the possessive D is the indefinite article un (1 instance) or the definite article i (10 instances). The fact that there are speakers who allow i to satisfy the D condition (see fn. 12) suggests that relative clauses of this last type should be set aside. We claim that the null RPs in the remaining 7 relative clauses, 4 of which are cited below, are intrusive. 'a man...who had a withered hand (lit. a man i that a hand of his i was dead)' (NT, Mark 3:1) These intrusive RPs are highly similar in distribution to the intrusive RP productions in our experimental data. Our second systematic search for null RP possessors also uncovered 31 instances of focus constructions with possessor gaps. The null RPs in 29 of the 31 focus constructions are grammatically licensed: their possessive DP satisfies the argument structure condition, the locality condition, and the D condition. The null RP in one focus construction, cited in (36), violates the argument structure condition, because the possessive DP is a locative; we claim that this null RP is intrusive. The null RP in another focus construction violates the D condition, because the possessive D is the definite article i; we set it aside for reasons described in the preceding paragraph. No instances of constituent questions with possessor gaps were encountered. 'Whoever believes in the Son of God has this testimony within himself (lit. [Whoever that believes in the Son of God] i , this testimony is in his i heart).' (NT, 1 John 5:10) Furthermore, in this and other corpus searches of NT, we randomly encountered a number of other types of extragrammatical RPs that amplify the picture considerably. These other RPs, though impressionistically infrequent, have a wider syntactic distribution than the RPs in our experimental data: they can be null or overt, have a range of grammatical relations, and occur in argument and adjunct positions-essentially, wherever ordinary pronouns can occur. As might be expected by now, these RPs have the profile of intrusive resumption. Specifically, they are found: -In relative clauses, both restrictive and nonrestrictive. Consider the overt RPs (in bold) in the nonrestrictive relative clauses in (37) -Inside islands and in other contexts involving syntactic distance of the RP from the filler, such as relative clauses in which a syntactically focused constituent intervenes between the filler and the RP. This is illustrated by the null and overt RPs in the relative clauses in (38). Once again, the PAH provides indirect evidence that the intrusive RPs illustrated above are pronouns, not traces of a more permissive wh-movement. There are no null RPs in the corpus data which are (possessors of) direct objects and which pattern like wh-traces (nonpronouns), as opposed to pronouns, for the purposes of the PAH. Our systematic searches for null RP possessors uncovered 16 relative clauses or focus constructions in which an animate null RP is the possessor of the internal argument of a transitive verb root. All these constructions conform to the PAH as if the null RP is a pronoun. Either the clause is passive, or else the clause is transitive and its subject (the external argument) is also a pronoun. The corpus data also include 3 relative clauses of type (38) in which a transitive verb root has an animate intrusive RP as its internal argument and a wh-trace (a nonpronoun) as its external argument. These clauses too conform to the PAH as if the RP is a pronoun. One clause is passive; in the other two, the verb shows overt wh-agreement, so the PAH-which takes effect only when the verb shows the default subject-verb agreement-is evaded. See (38b), (38d), and the discussion in Sect. 3.2. Finally, in the relative clause in (40), the expected gap or RP is replaced by a DP with descriptive content (i dos 'the two') that can be interpreted as containing a semantic variable bound by the filler, as in 'the group consisting of father and son with the property λx [[the two individuals in x] work for the gospel]'. Socioculturally, it is a very delicate matter to attempt to elicit speakers' introspective judgments on the grammaticality of sentences from a sacred text. We have not attempted this sort of elicitation. However, we did ask one speaker to help us with the English translation of the sentence from NT that is cited in (41a). In this sentence, a null-headed relative clause contains a null RP that is the possessor of the possessor of the subject of an unaccusative verb-a violation of the locality condition on wh-movement of possessors. We provided the speaker with the Chamorro sentence, the information that it was from NT, the corresponding English sentence from the New American Bible, and some candidates for a more literal English translation. The speaker ignored the English translations and instead responded, "To me, the Chamorro sentence is a bit awkward to begin with. The difficult meaning here is u fåttu chi-ña because it does not go well with baban kurason-ña. I can't translate it comfortably, actually." The speaker then suggested two Chamorro sentences as possible alternatives to (41a). One contains a relative clause with a subject gap; the other, cited in (41b), contains relative clause with a null RP possessor that does not violate the conditions on wh-movement of possessors. . thought 'Take care and let it not be that there is (one i ) among you that his i limit will be reached in evil thought.' This response strongly suggests that (41a) is degraded. The speaker's intuition that the sentence is awkward because the parts of the relative clause do not go well together is typical of Chamorro speakers' reactions to e.g. island violations. For us, this is one small piece of independent evidence that the null RP possessor in (41a) is intrusive.

Interim summary
The experimental evidence and the corpus evidence converge to make a strong case for the presence of two types of RPs in Chamorro: grammatically licensed RPs, on the one hand, and intrusive RPs, on the other. The grammatically licensed RPs are null and limited to exactly one syntactic context-possessors whose possessive DP satisfies the three conditions on wh-movement of possessors. The intrusive RPs can be null or overt; they have a broader syntactic distribution than the grammatically licensed RPs, but are far less frequent, as is clear from the corpus data. Finally, the intrusive RPs have the distributional profile associated with intrusive resumption in other languages. All this is what we would expect if the grammars of individual languages differ in whether, and in what contexts, they have grammatically licensed RPs, but intrusive RPs-RPs that we view as extragrammatical artifacts of performanceare found in every language.

Grammatically licensed and intrusive RPs in Palauan
Our investigation of Chamorro raises the question of whether languages at other end of the RP spectrum distinguish between grammatically licensed RPs and intrusive RPs. Here we motivate an account of Palauan's A-bar dependencies that relies on this distinction. Our analysis, which differs significantly from that offered by the late Carol Georgopoulos, is based exclusively on her fieldwork data (see Georgopoulos 1985 and1991).

Georgopoulos' analysis
Georgopoulos observed that the gaps in A-bar dependencies in Palauan are either null or filled by an overt RP. Overt RPs occur as objects of er, which she analyzed as Palauan's only preposition, but which could equally well be treated as Palauan's only overt case marker (see Sect. 2.2.1). Otherwise, the gaps are null. For reasons to be discussed immediately, Georgopoulos proposed that Palauan's A-bar dependencies are not created by wh-movement, but instead always involve a base-generated RP that is syntactically bound by a base-generated A-bar operator. Specifically, she claimed that the gaps in examples like (5), (7), (13)  First, she pointed out that the distribution of null gaps and overt RPs in Palauan closely parallels the distribution of ordinary null and overt pronouns in the language. Overt RPs occur in exactly the contexts where ordinary pronouns must be overt; null gaps occur in the contexts where ordinary pronouns can be null. Therefore, she claimed, the simplest analysis is for all null gaps and overt RPs to be base-generated as pronouns that are syntactically bound by their fillers. Second, she showed that relative clauses, constituent questions (which she analyzed as structures that contain a relative clause), and sentential subjects are not islands for A-bar dependencies in Palauan. Her examples include the following.  (Georgopoulos 1991:80, (36b)) The non-islandhood illustrated in (43) might seem to be decisive evidence that Palauan's A-bar dependencies are not derived by wh-movement. However, Georgopoulos (1991:107-108, 115-117) also showed that adjunct clauses and coordinate structures are islands for A-bar dependencies in Palauan (see Sect. 2.2.2), so the island facts are, at best, equivocal.
As far as we know, Georgopoulos' analysis of Palauan's A-bar dependencies is unique. No other language has been claimed to have syntactic variable binding as its one and only strategy for deriving A-bar dependencies. Moreover, if her analysis is correct, it would pose a serious challenge to our claim that intrusive resumption can be found in every language. The reason is that under her analysis, every putative instance of intrusive resumption in Palauan could equally well be analyzed as involving syntactic variable binding of a grammatically licensed, base-generated RP. The reanalysis we propose below avoids the exceptionalism of Georgopoulos' analysis. It also makes room for Palauan to draw a distinction between grammatically licensed RPs and intrusive RPs-a distinction which, we suggest, finds some empirical support.

Our reanalysis
Our alternative account of Palauan's A-bar dependencies proceeds from the following considerations. It is well known that there are languages with wh-movement in which certain types of relative clauses are weak islands, or not islands at all (e.g. Engdahl 1997 on Scandinavian languages; Sichel 2014, 2018 on Hebrew). Moreover, sentential subjects generally are not islands in verb-first languages (Chung 1991a). Together, these observations suggest that the putative non-islandhood of relative clauses and sentential subjects is not a particularly good indicator that Palauan lacks whmovement.
We take it for granted that sentential subjects are simply not islands in Palauan, which is a verb-first language. What about relative clauses? If, contra Georgopoulos, Palauan does have wh-movement, can this movement operate out of a relative clause? Wh-agreement ought to shed light on this question. If the A-bar dependencies in (43a-b) are derived by 'long' wh-movement out of a relative clause, the verb of the relative clause should, in principle, show wh-agreement with the wh-trace left by that 'long' movement. What is interesting is that this never happens. Instead, the verb of the relative clause agrees with the wh-trace of the more local movement that created the relative clause in the first place. In (43a), for instance, the verb of the relative clause (odurokl 'send') shows wh-agreement with the subject wh-trace left by local movement of the head NP redil 'woman'. It does not show wh-agreement with the nonsubject wh-trace left by long movement of the null NP ('one, thing') out of relative clause. The situation in (43b) is similar. One way of describing this pattern is to say that when a verb has two arguments that are wh-traces, it must show wh-agreement with the wh-trace whose local antecedent is closer-in derivational terms, the wh-trace created first. If we accept this, then wh-agreement does not help to resolve the issue of whether relative clauses in Palauan are islands.
These indeterminacies open up space for a reanalysis of Georgopoulos' data that brings Palauan's A-bar dependencies more in line with crosslinguistic expectations. Our proposal runs as follows. To begin with, we maintain that Palauan does indeed have wh-movement (recall the evidence presented in Sect. 2.2.2). Wh-movement affects DPs that have no overt case marking, as well as DPs that are case-marked with er, which we take to be the language's only overt case marker (see fn. 5 and cf. Nuger 2016). Wh-traces that are case-marked with er are spelled out as overt RPs. One could attribute this to a requirement that overt morphological case be recoverable, along the lines suggested by Hladnik (2015). 26 Otherwise, wh-traces are unrealized. Crosslinguistically, this is a familiar pattern (see e.g. Bauer 1997 on Māori; Hladnik 2015 on Slavic languages). We illustrate it by annotating the A-bar dependencies presented earlier in (42)  The idea that the RPs in (44b-d) are the realizations of wh-traces is supported by the fact that they control wh-agreement, which we have taken to be agreement with a wh-trace (see Sect. 2.2). We claim further that Palauan has a standard collection of islands that includes adjunct clauses and coordinate structures, but not relative clauses or sentential subjects. Wh-movement cannot operate out of islands; that accounts for the ungrammaticality of the A-bar dependencies in (9). We take the non-islandhood of relative clauses and sentential subjects to have been established by Georgopoulos (1985Georgopoulos ( , 1991. Significantly, her findings were later confirmed by Justin Nuger in his field research (Justin Nuger, p.c.).
Finally, we propose that when A-bar dependencies in Palauan do cross island boundaries, they involve intrusive RPs that may be overt or null. Our overall approach to intrusive resumption leads us expect these intrusive RPs to be found inside islands (i.e. adjunct clauses and coordinate structures) as well as in other Palauan structures that involve production complexity (because a wh-trace inside them counts as syntactically distant from a filler that lies outside). Tantalizingly, there are a very few examples of A-bar dependencies in Georgopoulos's works that suggest that this expectation might be realized. These examples are not accounted for by her analysis, but have a natural explanation as instances of intrusive resumption. In other words, they constitute a small but crucial amount of evidence in favor of our proposal.
Consider first the A-bar dependencies in (45), in which the gap occurs inside an embedded CP headed by one of the complementizers e or el kmo. These Cs, which are selected by higher verbs of communication, select a realis TP complement (Chung and Georgopoulos 1988:261-262). The key point for current purposes is that these gaps evidently do not control wh-agreement. In (45a), the gap is a nonsubject but the verb does not show nonsubject wh-agreement, which-recall-is realized by an irrealis form of the verb (see Sect. 2.2). In (45b-c), the gap is a subject but the verb does not show subject wh-agreement, which is realized by a realis form of the verb plus the absence of subject-verb agreement. Japan 'Who (is the one who) they believe that we think will go to Japan?' (Georgopoulos 1991:92, ( 'Mary (is the one who) it's clear that _ loves John.' (Georgopoulos 1991:80, (36a)) The fact that the embedded verb in (45a) (menguiu 'read') is not in a wh-agreeing form is not particularly telling, since the normal realization of wh-agreement might simply be overridden by the embedded C's demand that its TP complement be realis (see fn. 9 for some of the issues). The absence of wh-agreement in (45b-c) is more significant. In these cases the embedded C's demand for a realis TP would be satisfied whether or not subject-verb agreement also appears. Given that absence of subject-verb agreement is part of the morphological signature of Palauan whagreement (Baier 2017), the presence of subject-verb agreement (in bold) in (45b-c) suggests that there is no wh-agreement with the gap in these examples. That in turn suggests that the gap is not a wh-trace, but rather a null pronoun-possibly, an intrusive null RP. More generally, we conjecture that CPs headed by e or el kmo are islands in Palauan. If the conjecture is correct, then the gaps in (45) could well be instances of intrusive RPs. 27 Consider next the A-bar dependency in (46), which appears in a footnote to Georgopoulos' (1985:86-89) demonstration that coordinate structures are islands for Abar dependencies in Palauan. Here, an overt RP serves as one conjunct of a coordinate DP.

(46)
Palauan overt RP as one conjunct of a coordinate structure A Naomi Naomi a [kau-sechelei RECIP-friend ngii i she me and a Chris]. Chris 'Naomi i (is the one who she i ) and Chris are friends.' (Georgopoulos 1985:87, fn. 27, (i)) Observing that an overt RP is possible in this example but a gap is not, Georgopoulos (1985:87, fn. 27) attributed the impossibility of the gap to a failure of proper government. What remains a mystery in her analysis, given her approach to A-bar dependencies and her demonstration that coordinate structures are islands for them, is the fact that the overt RP is possible at all. For us, this overt RP is another instance of intrusive resumption.
To recapitulate, we propose that Palauan has wh-movement, a collection of islands that includes adjunct clauses and coordinate structures, and two types of RPs: grammatically licensed RPs and intrusive RPs. The grammatically licensed RPs are overt pronouns that realize wh-traces; they are limited to exactly one syntactic context-DPs case-marked with er, Palauan's only overt case marker. The intrusive RPs can be null or overt; they are found in islands, including coordinate structures (as in (46)), CPs headed by e or el kmo (as in (45)), and presumably in other contexts that involve production complexity. Our proposal encompasses the A-bar dependencies in (45)-(46), and in this sense has slightly better empirical coverage than Georgopoulos' account.
The idea that intrusive resumption is involved in (45)-(46) leads to the expectation that these sorts of dependencies will be produced readily, but judged as degraded or ungrammatical, by native speakers. The literature on Palauan syntax contains no discussion of speakers' reactions to these dependencies. Until such information is supplied, our proposal must be considered tentative. Nonetheless, the proposal, speculative though it is, makes a point. If languages exist in which every A-bar dependency is formed by syntactic variable binding of a grammatically licensed, base-generated RP, such languages could counterexemplify our claim that intrusive resumption can be found in every language. Georgopoulos maintained that Palauan was a language of just this type. What we have shown is that an equally plausible analysis of Palauan's A-bar dependencies can be given which does not involve syntactic variable binding of a grammatically licensed, base-generated RP. Moreover, even in Georgopoulos' works, data can be found which suggest that this language may also have intrusive RPs.

Conclusion
In our view, intrusive RPs are ungrammatical products of the performance systemproductions that satisfy local well-formedness but not global well-formedness. As such, they could in principle be found in any language, not just those languages that lack grammatically licensed RPs.
We investigated the A-bar dependencies of Chamorro and-more briefly-Palauan. Based on evidence from fieldwork, an elicited production experiment, and the Chamorro New Testament, we showed that Chamorro has grammatically licensed null RPs in one narrow syntactic context, as well as null and overt RPs that are intrusive. These intrusive RPs are far less frequent, but more widely distributed, than Chamorro's grammatically licensed RPs. Based on our reanalysis of Georgopoulos' (1985Georgopoulos' ( , 1991 Palauan data, we tentatively proposed that Palauan has a similar divide between grammatically licensed overt RPs and intrusive (null and overt) RPs.
The idea that intrusive RPs should be found in every language is not the typical sort of universal claim, in that it applies to performance (E-language) as opposed to competence (I-language). At the same time, it opens the door to a reshaping of some familiar assumptions about the taxonomy of RPs. It is sometimes assumed that languages with grammatically licensed RPs allow island violations as a matter of course (e.g. Borer 1984 on Hebrew; McCloskey 1990 on Irish). Our reconception suggests that such languages may have both grammatically licensed RPs and-separatelyintrusive RPs. If so, one might expect there to be differences in the frequency, distribution, and acceptability of the two types of RPs, as has been shown by Alexopoulou and Keller (2007) for Modern Greek.
What is needed to take these speculations further is more extensive production evidence from other languages-especially languages with grammatically licensed RPs. We offer the current study in the hopes that others will join us in pursuing this investigation.