3.1. Core proposal

I suggest that repeaters emerge post-syntactically, at the time of Vocabulary Insertion. Languages with repeaters have two ways to lexicalize (i.e., spell out or expone) the Cl node. The first method is to choose from the Exponent List a vocabulary item that is specified for insertion in the abstract Cl node and is permanently associated with a specific, invariable phonological form. This produces garden-variety classifiers and is available in all classifier languages.

(23)	List of Vocabulary Items that can be inserted to spell out the abstract Cl node
	a.	Cl $ \leftrightarrow $ khan/{ , etc.}__
	b.	Cl $ \leftrightarrow $ lûuk/{, etc.}__
	c.	…

(24)	Vocabulary Insertion into Cl
	a.	Cl $ \to $ [Cl-khan]
	b.	Cl $ \to $ [Cl-lûuk]
	c.	…

As shown in (23), the Exponent List of a typical classifier language like Thai contains several vocabulary items that can spell out Cl. As is usual for such cases, ‘the winner is determined by a morpheme local to the morpheme undergoing insertion’ (Embick Reference Embick2015: 92). In other words, the vocabulary items for classifiers are supplied with contextual information that restricts their occurrence to the context of specific roots. Thus in the context of the root rôm ‘umbrella’, Cl is spelled out as khan, while in the context of the root mánaaw ‘lemon’, Cl is spelled out as lûuk. Footnote ¹²

The second method, available only in repeater languages, is to copy the phonological matrix of the noun and use this melody to spell out the Cl functional head. This operation affects only N’s phonological features/melody; its syntactico-semantic features remain unaffected. The insertion targets the Cl head directly, without a null affix being present. This is schematized below. The subscript ‘x’ after ‘melody’ is meant to indicate the identity of the two melodies.

(25)

(26)

‘N’ is a shorthand for the root and its categorizer n. Whether roots are subject to late insertion is a matter of debate (see Alexiadou et al. Reference Alexiadou, Borer, Schäfer, Alexiadou, Borer and Schäfer2014) and it does not directly concern us here (though see the end of this section for some remarks). What is important is that Vocabulary Insertion proceeds bottom up (Bobaljik Reference Bobaljik2000), and so by the time Cl is vocabularized, N has already been associated with its own melody.

The copying of the phonological matrix of the noun is not a movement operation (whether syntactic or post-syntactic); it is a type of syntactic (i.e., non-affix creating) reduplication instead. Taking my cue from Travis (Reference Travis1999, Reference Travis, Kim and Strauss2001, Reference Travis and Somesfalean2003), I suggest to capture this via a special Vocabulary Item for Cl, one that is not permanently associated with a phonological matrix of its own but which is specified for reduplication in the list of vocabulary items (28).

	b.	k	sǎam	k
		island	three	cl:island
		‘three islands’ (Hundius & Kölver Reference Hundius and Kölver1983)		Thai

(28)	Vocabulary Item, a member of the list in (23)

(29)	Vocabulary Insertion into Cl with (28)
	a.	Cl $ \to $ [Cl-prathêet]
	b.	Cl $ \to $ [Cl-k ]
	c.	…

This analysis echoes Burling’s (Reference Burling and Spiro1965: 250) view that repeaters are all ‘allomorphic variants of the same morpheme’, in the sense that they all go back to the same vocabulary item/listeme of the Exponent List, namely (28).Footnote ¹³

As mentioned above, on the standard approach Vocabulary Insertion operates bottom up, targeting the lower terminals before the higher ones (Bobaljik Reference Bobaljik2000). Thus after the melody of N is copied into Cl:red, Vocabulary Insertion continues upwards. It first reaches Num, and inserts the melody of the numeral (30). Next, Vocabulary Insertion reaches and expones the instance of the NP that has been placed into the pre-numeral position (see (5)/(12)). At the same time, the exponence mechanism also registers that this instance of the NP and the instance at the bottom of the structure, in the Cl-complement position, are fully identical: they involve the same listeme of the Exponent List and share all of their formal (morphosyntactic) features as well. They are therefore identified as being syntactic copies, which form a movement chain. This chain then undergoes the usual chain reduction, which suppresses the surface-exponence of the tail of the chain (31). (Note that Cl:red is not part of a movement chain, so its exponence is not affected.)

(30)

(31)

In (28) the reduplicative classifier is endowed with a contextual specification, similarly to ordinary classifiers (23). Alternatively, to capture Enfield’s (Reference Enfield2004: 120) intuition that ‘the repeater strategy is the closest thing to a residue option in the numeral classifier system’, we could drop the contextual information from the Vocabulary Item of Cl:red, and make it the default or elsewhere option, suitable whenever the given root is not listed in the contextual information of other, non-reduplicative classifier listemes.

(32)	Vocabulary Item, a member of the list in (23)
	Cl $ \leftrightarrow $ RED

In order to choose between (28) and (32), we need to consider the distribution of repeaters. In Ersu (Sino-Tibetan; Sichuan Province, China), for instance, repeaters are licit if and only if the quantified noun is not compatible with any of the regular classifiers (Zhang Reference Zhang2014). This also appears to be the case in Lao (Goddard Reference Goddard2005: 98). In these languages, the lexical entry of the reduplicative classifier can be left without contextual specification; (32) can be employed as a true last resort or rescue option whenever no other classifier is appropriate. This way, the derivation can meet the requirement that every syntactic terminal must be associated with a spellout, and Vocabulary Insertion can proceed further up in the tree.Footnote ¹⁴

In other repeater languages there is no full complementary distribution between repeaters and ordinary classifiers. In Thai, repeaters are generally not used when a garden variety classifier is available for the noun (33).

Conversely, nouns that take a repeater are generally not compatible with the existing garden variety classifiers. In (34), substituting an (the general classifier for inanimates) for repeaters produces unacceptability.

	b.	k	sǎam	k/*an
		island	three	cl:island/cl.general
		‘three islands’ (Pornsiri Singhapreecha, p.c.)		Thai

For this reason, Jenks (Reference Jenks2011: 94) suggests that ‘the repeater construction seems to be a kind of last resort’.

With some nouns, however, either a repeater or a regular classifier is possible. Pornsiri Singhapreecha (p.c.) informs me that the alternants in (35) carry the same meaning.

	b.	bâan	sɔŋ	bâan/lǎŋ
		house	two	cl:house/cl.house
		‘two houses’ (Placzek Reference Placzek1978: 43)		Thai

It thus seems to me that assuming a vocabulary item such as (32) as a general elsewhere-case is not fully appropriate for Thai: this would constrain repeaters to those nouns that cannot occur with ordinary classifiers, inappropriately ruling out alternations such as (35). I suggest that in a language like Thai, the nouns that take repeaters should be included in the entry of the reduplicative classifier in the form of a contextual specification, as in (23). Nouns that allow the alternation, such as those in (35), are simply listed as a contextual specification for both one of the regular classifiers and for the reduplicative classifier. Nouns in numeral classifier languages are often compatible with more than one ordinary classifier (e.g., in colloquial Thai pàakka ‘pen’ can take the general classifier an instead of dâam and cháaŋ ‘elephant’ can take tua ‘Cl.animal’ instead of the more formal ch ak ‘Cl.elephant’). Thus including a noun in the contextual specifications of two different classifiers is necessary independently of repeaters as well.Footnote ¹⁵

The reduplicative approach to repeaters avoids the problems raised by Simpson’s (Reference Simpson, Cinque and Kayne2005) and Simpson & Ngo’s (Reference Simpson and Ngo2018) accounts. It captures the item-specific nature of repeaters in the Vocabulary, which is the usual repository for certain idiosyncratic, unpredictable properties of listemes, and it does not violate the LCA. While they house the same melody, N and Cl remain clearly distinct in their morphosyntactic feature specification and so they are not involved in a movement chain. Therefore the issue of chain reduction (or lack thereof) is not relevant to them: both positions are spelled out by the same melody, as is usual in cases of reduplication. No typologically unattested order is derived and no unusual morphophonological constituency between Cl and N is assumed either.

This view also affords insights into the phenomenon of semi-repeaters. Recall that for compounds headed by an autoclassified noun, only the head of the compound is repeated in the Cl position.

The reason for this, I suggest, is that the pattern in (36) is the optimal way to meet the requirement that Cl should receive an exponent. Targeting just the phonological material of the compound head keeps the melody copying mechanism minimal but at the same time allows Cl to be spelled out. In other words, this is the necessary and sufficient condition to allow Vocabulary Insertion to proceed beyond Cl. Semi-repeaters are parallel to cases of partial reduplication. These place a ‘cap’ on the amount of material to be repeated in such a way that the reduplicant may be smaller than the base, and the operation may affect material that would not be able to move independently.Footnote ¹⁶ No understanding of semi-repeaters is readily forthcoming from the head movement analysis, however.

To summarize, I suggest that repeaters involve the kind of copying found in reduplicative structures rather than the kind of copying that has been hypothesized to occur in movement structures. As discussed in Travis (Reference Travis, Kim and Strauss2001: 461), ‘the copying found in reduplication is very different from the copying mechanism used for movement (as in Copy and Delete). In syntactic reduplication, subdomains of the complement XP are being copied that would not be able to independently move. The result appears to be, then, that the copying of reduplication cannot be collapsed with the copying of movement’. Thai semi-repeaters provide one example for copying of a domain that would not be able to move on its own. A further example will be discussed in Section 4.

3.2. Repeaters as instantiations of a general classifier

As mentioned above, the proposed analysis ties in with Burling’s (Reference Burling and Spiro1965: 250) view that repeaters are all ‘allomorphic variants of the same morpheme’. In the present analysis, this morpheme is Cl:red. This single vocabulary item underlies all repeater classifiers, which are just different surface realizations of this morpheme. Theoretically, this is a more economical and streamlined model of the Thai list of vocabulary items than a potential lexicalist approach assuming 500+ different, highly specialized classifier entries, each of which is restricted to occur with the single noun that it is form-identical with. In addition to being preferable on the basis of Occam’s Razor, there is also empirical support for the present proposal over the lexicalist alternative.Footnote ¹⁷

Cl:red can be said to be a general classifier in the sense that it is desemanticized: it does not impose shape, size, or function-based restrictions on the wide variety of nouns it co-occurs with. From this perspective, we expect Cl:red and the ‘regular’ general classifier an to show some parallels. On the lexicalist view, on the other hand, repeaters are predicted to exhibit parallels with so-called unique classifiers. A unique classifier is a classifier whose use is restricted to a single – often culturally highly significant – noun (Grinevald Reference Grinevald and Wright2001). The single defining criterion of unique classifiers is their highly narrow range of application; they may have an entirely different phonological shape from the noun they classify. In Thai, ch ak (lit. ‘rope’) is a good example: this classifier is uniquely applied to domestic elephants (Burusphat Reference Burusphat2007, Aikhenvald Reference Aikhenvald2021) but it is not form-identical to the noun it classifies (‘elephant’ is cháaŋ).Footnote ¹⁸ Similarly, the classifier law lit. ‘reed’ is applied only to the noun pìi ‘flute’ (Hundius & Kölver Reference Hundius and Kölver1983). On the lexicalist view, repeaters are a proper subset of unique classifiers (distinguished from the rest of the class only by their phonological identity to the classified noun), and so repeaters are expected to have similarities with unique classifiers such as chak rather than with the general classifier an. This prediction is not confirmed, however: repeaters show parallels to an rather than to unique classifiers, both during language acquisition and in language loss.

Starting with the differences between repeaters and other unique classifiers, Carpenter’s (Reference Carpenter1991) acquisition study found that the use of repeaters increased until age 5 (where they occurred in 18% of all responses), after which their use declined. By contrast, only one 9-year old child out of 204 (age range 1;8–11;3) used the unique classifier chak with cháaŋ ‘elephant’, in spite of the fact that this classifier-noun collocation is part of the input in the formal classroom setting. That repeaters and other highly specialized classifiers are not treated in the same way by children is also confirmed by Gandour et al.’s (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984) results. Gandour et al. (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984) found that children’s errors in NPs that in adult language feature muan, a classifier that is applied only to cigarettes and cigars, involved the substitution of a classifier used with a wider range of nouns. Thus while highly specialized classifiers are difficult for children, they overuse repeaters, which is the type of classifier that is the most specialized on the lexicalist account.Footnote ^19, Footnote ²⁰

Turning now to the similarities between repeaters and an, classic studies on the acquisition of classifiers such as Gandour et al. (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984) and Carpenter (Reference Carpenter1991) have shown that children overuse both the general classifier an and repeaters at the expense of what Grinevald (Reference Grinevald and Wright2001) calls specific classifiers, and this was the case for Gandour et al.’s (Reference Gandour, Buckingham and Dardarananda1985) conduction aphasic patient, too. Specific classifiers are classifiers which, unlike an, impose some shape, size, or function-based restrictions on the nouns they occur with, but which, unlike unique classifiers, are not restricted to use with a single noun (cf. khon for people and lêm for long and pointed or sharp-edged objects).

The overuse of an involves overgeneralization, while the overuse of repeaters amounts to overspecialization among classifiers. These are two seemingly opposite strategies, and thus their simultaneous use by the same speakers presents a puzzle. On the present account, this is not a mystery, however. Repeaters expone a general classifier whose only significant difference from an is that it is not associated to a piece of permanent, stable melody. The speakers in question therefore use a single strategy: both an and Cl:red are classifiers with a broad range of associated nouns, which substitute for more specialized classifiers.

In their discussion of the simultaneous application of overgeneralization and overspecialization, Gandour et al. (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984, Reference Gandour, Buckingham and Dardarananda1985) appeal to Conklin’s (Reference Conklin1981) theory of semantic load. Conklin suggests that due to their role in categorizing on the basis of inherent perceptible characteristics, specific classifiers add new information to the noun phrase, e.g., +human, long and thin, small and spherical, etc. This makes them semantically heavy. The general classifier, on the other hand, is semantically light, as it adds no semantic information to the NP beyond providing unit reference. For Conklin, repeaters are also semantically light: since they apply to singleton sets, they do not provide any semantic information beyond what is already present in the noun itself. To account for their data, Gandour et al. (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984, Reference Gandour, Buckingham and Dardarananda1985) suggest that semantically light classifiers are both acquired early in childhood and retained longer in language loss. This appeal to semantic complexity thus finds parallels between an and repeaters without taking the latter to be different phonological variants of a general classifier. At the same time, this type of explanation does not account for why the unique classifiers that are not form-identical to their associated nouns, such as chak, are difficult for children. Similarly to repeaters, chak-type classifiers also apply to singleton sets. They, too, eschew noun categorization on the basis of inherent perceptible characteristics, and so they should be semantically as light as repeaters. However, Carpenter’s (Reference Carpenter1991) findings do not confirm this: her results place an and repeaters on one pile, and unique classifiers such as ch ak on another.

In overextending repeaters beyond the sphere of nouns they are usually used with, children and aphasics treat repeaters as a relatively productive category. Among Gandour et al.’s (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984) subjects, one 5-year-old used repeaters 100% of the time, while one 6-year-old used them in 76% of all responses. This was so in spite of the fact that none of the input nouns were classified by a full repeater in adult usage, and that only a few nouns allowed a partial repeater as one of several acceptable adult options. Gandour et al. (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984: 472) suggest that this overuse ‘may be attributed, in part, to syntactic relationships between classifiers and classified nouns’ (emphasis mine). The present proposal aligns with a non-lexical connection between repeaters and their associated nouns, and makes the repeater-producing rules alluded to in Gandour et al. (Reference Gandour, Petty, Dardarananda, Dechongkit and Mukngoen1984) fully explicit.

3.3. Consequences beyond repeater constructions

The analysis of repeaters as reduplication has consequences for the syntax of ellipsis. As shown below, the head noun can undergo ellipsis in NPs with both garden variety classifiers and repeaters.

Ellipsis of the nominal head in repeater constructions is preferred (Juntanamalaga Reference Juntanamalaga1988, Beckwith Reference Beckwith2007: 101; Piriyawiboon Reference Piriyawiboon2010), presumably for reasons of economy. Recovery of the noun in repeater phrases is even more straightforward than in ordinary classifier constructions, as the former ‘fully describe the head N’ (Bradley Reference Bradley2005: 224), while the latter combine with a wider range of nouns. In order to derive examples such as (38), the order of operations must be 1) exponing N, 2) copying melody into Cl, 3) ellipsis of N. In other words, the melody of N must be present at an early stage of the derivation even in cases of NP ellipsis, in order to allow for melody copying from N to Cl to take place. This falls out naturally if unlike functional heads, roots are not subject to late insertion, but is also compatible with the late-insertion of roots as long as ellipsis is deletion at PF (Merchant Reference Merchant2001, Aelbrecht Reference Aelbrecht2010). However, ellipsis as radical lack of vocabulary insertion (Kornfeld & Saab Reference Kornfeld, Saab, Bok-Bennema, Hollebrandse, Kampers-Manhe and Sleeman2004) or as a null pro form in the ellipsis site (Lobeck Reference Lobeck1995) is ruled out.

Repeater constructions also weigh in on an old debate about the structure of classifier expressions. Whether the classifier forms a constituent with the noun or the numeral is an ongoing lively debate in the literature, with no consensus in sight for the time being. The proposal advanced here is only compatible with the [Num [Cl N]] constituency, and thus can be viewed as an argument corroborating this structure. This is because as argued in Travis (Reference Travis, Kim and Strauss2001: 461), a structural restriction on syntactic reduplication is that ‘subdomains of the complement XP are being copied’. Differently put, the base must be in the complement of the reduplicant, which holds only on the [Num [Cl N]] structure. Repeaters therefore provide a new, so far overlooked piece of argument in the debate on NP constituency as well.Footnote ²¹

4.1. Reduplication into n

Wang & Holmberg (Reference Wang and Holmberg2021) observe that monosyllabic common nouns in Traditional Xining Chinese (Northwest China) are always reduplicated. Reduplication has no semantic effect, it is just a formal requirement on common nouns. (See also Wang & Holmberg Reference Wang and Holmberg2023 on similar data involving Mandarin Chinese given names.)

Nouns do not come from the lexicon in a reduplicated form: reduplication is triggered in certain environments only. Thus in (39-a), where xiōng is an independent noun, reduplication is obligatory, but in (40), where it is the non-head of the compound xiōng-bóng, reduplication is not possible.Footnote ²²

Wang & Holmberg (Reference Wang and Holmberg2021) suggest that this pattern is due to a requirement in Traditional Xining Chinese that a free noun must have at least two syllables. Chinese displays a close correspondence between syllable and morpheme, but there are some polysyllabic loanwords as well. As shown below, these do not undergo reduplication because the minimum syllable count is already met.

In order to formalize their findings, Wang & Holmberg (Reference Wang and Holmberg2021) propose that Traditional Xining Chinese has a language-specific output condition (or filter) that applies after Vocabulary Insertion.

In cases where (42) would not otherwise be met, the phonological features of the root are copied into the categorizer n. As the target of melody insertion is a sister null nominalizer, reduplication cannot take place when such a sister is absent. This rules out reduplication for the non-head of compounds (40-b) (which are not sisters to n), for pronouns (which are merged higher in the DP) and for adjectives and verbs (which employ a different categorizer).

Reviewing the full details of Traditional Xining Chinese reduplication would take me too afield; the interested reader is referred to Wang & Holmberg (Reference Wang and Holmberg2021) for a full explication of the facts and the analysis. What is important for our purposes is the parallels that this reduplication displays with repeater classifiers. In both cases, the phonological matrix of the bottommost element of the extended NP is copied into a syntactically independently generated head position along the functional spine. The two phenomena are also similar in that reduplication itself has no semantic import (such as plurality, iterativity, or intensification); it only serves to satisfy a formal criterion (bisyllabicity of free nouns in Traditional Xining Chinese and overt filling of the Cl position in repeater languages). They are also similar in that the copying process mainly or exclusively targets the heads of compounds.

Where the two processes differ is the element whose requirements are met by reduplication. In Traditional Xining Chinese, the trigger of reduplication is N/the base (it cannot be a free word with less than two syllables), while in repeater languages the trigger of reduplication is Cl/the reduplicant (this position must receive a spellout). In other words, while n in Traditional Xining Chinese can receive a phonologically zero spellout (as long as its complement is already bisyllabic), there is no phonologically zero spellout (i.e., null classifier) in repeater languages, and thus an overt spellout of Cl is forced.

4.2. Melody copying into Cl

In this section, I turn to repeaters in Murui (Witotoan). In order to provide the relevant background for the discussion, I will first introduce the phenomenon of ‘multiple classifier’ languages and survey the basic facts about the Murui NP.

In certain languages, the same set of classifiers can be used in more than one morphosyntactic environment. For instance, there are several languages in which the same set of classifiers occurs both with numerals and demonstratives (e.g., Mandarin Chinese), and sometimes classifiers occur with numerals, demonstratives as well as adjectives (Thai). In the case of Mandarin Chinese and Thai, researchers agree that one environment, namely that of numerals, is ‘primary’. This is because in the context of numerals classifiers are obligatory, while with demonstratives and adjectives they are not.Footnote ²³ These languages are therefore classified as numeral classifier languages, similar to Ch’ol (Mayan), which allows classifiers only with numerals and the interrogative ‘how many’ (Bale et al. Reference Bale, Coon and López2019).

In other cases, however, there is no reason to think that one classifier environment is ‘primary’ and the others are ‘secondary’ (e.g., classifiers are obligatory with numerals as well as demonstratives and adjectives). In these cases we are talking about a ‘multiple classifier’ language (Aikhenvald Reference Aikhenvald2000). An example of this type of classifier language is Murui. Classifiers in Murui are mono- or disyllabic suffixes that categorize nouns based on physical properties (shape, size, consistency, functional properties), animacy, and other characteristics (Wojtylak Reference Wojtylak, Körtvélyessy, Štekauer and Valera2016). In the DP, they can occur on bare nouns and on nominal modifiers.

Murui nouns come in two types: bound noun roots, which must occur with a classifier to form an independent word (43), and free nouns, which optionally occur with classifiers (44).

Some nouns, for instance, ñeki ‘chambira palm’, are compatible with more than one classifier (45). There are also free nouns for which there is no classifier, e.g., nokae ‘canoe’, kuiada (a fish type), or jaziki ‘jungle’.

Adjectives, numerals, interrogative words, demonstratives and other anaphorics in Murui are bound forms; they must combine with a classifier to form independent words.Footnote ²⁴ In adnominal position, these modifiers take the general classifier -(j)e, irrespective of whether the head noun is classifiable, and if so, which specific classifier(s) it takes. For the sake of brevity, I illustrate this only with numerals and demonstratives, but the other modifiers mentioned above also exhibit this pattern.

Full NPs like those in (46)–(49) are rare, however. More frequently, the head noun is elided, and the modifier takes on the classifier that would be borne by the noun. (This aids recovering the reference of the ellipted noun; the semantic load of the general classifier is so weak that it would not be a useful anaphoric/reference-tracking device.)

Crucial to us is what happens when the elided noun is such that there is no classifier for it. Filling the classifier slot is obligatory, and Murui meets this formal requirement by employing a repeater. In the default case, repeaters are partial copies of the corresponding noun: nouns with more than two syllables give rise to disyllabic repeaters (52), while nouns with two syllables give rise to monosyllabic repeaters (see -ko in (53-a)).Footnote ²⁵ (52) and (53-b) show repeaters for Spanish loanwords, while (53-a) illustrates repeaters with native words.

Repeaters have the same phonological features and prosodic status as garden variety classifiers (Wojtylak Reference Wojtylak2017: 200).Footnote ²⁶

Classifiers in Murui clearly have a different distribution from the classifiers of East and Southeast Asian languages (‘multiple classifier’ language vs. numeral classifier languages), and I do not claim that the two systems can be collapsed. There are, however, some striking parallels between repeaters in Thai (Lao, etc.) and Murui. In both cases, there is a formal requirement to include a classifier in a certain syntactic slot. When the noun does not combine with the existing classifiers, (part of) the phonological matrix of the noun is borrowed to ‘fill in the blank’ (Wojtylak Reference Wojtylak, Körtvélyessy, Štekauer and Valera2016: 27), in an almost expletive-like manner. Repeaters form a residual class in both cases (in Murui, applying only to nouns that are not classifiable, and in Thai, mostly but not exclusively applying to otherwise unclassifiable nouns), and they fill a slot that is otherwise taken up by garden-variety classifiers.

Murui clearly shows that it is possible to access the noun’s lexical entry to fill a classifier position without involving syntactic movement. The partial repeaters in (52) and (53) are not eligible for movement on their own because they correspond to strings on the sub-morphemic level; they involve (partial) copying of the phonological matrix of the nominal root as a last resort. The Murui facts therefore provide strong independent evidence for a crucial ingredient of my analysis of repeaters in East and Southeast Asia.