1.1. Total Reduplication

First we will look at total reduplication. With total reduplication, a root, stem, word or phrase is repeated in its entirety; thus shape and size vary with the size of the unit repeated. As one can see, the relevant constituent is copied in its entirety. Word reduplication in (1c) is distinguished from others by repeating the inflectional affix.

(1)

While reduplication commonly involves repetition of stems and roots, true cases of affix reduplication are rare, though there are numerous cases where affix material is copied as part of stem reduplication (see Inkelas & Zoll, 2005, for examples). In assessing examples, one needs to take care to tease apart cases of multiple occurrences of a particular affix such as the diminutive affix in Spanish gat-it-it-o ‘little little cat.’

The mechanism by which total reduplication occurs varies with different theoretical approaches. Early research achieved this by morpheme or stem repetition (Marantz, 1982), referred to as tautologous compounding by McCarthy and Prince (1988, 1990). More recently, Morphological Doubling Theory (Inkelas & Zoll, 2005, p. 6) has been proposed, in which “reduplication results when the morphology calls twice for a constituent of a given semantic description.” This model will be discussed in more detail. The following sections outline patterns of partial reduplication, starting with foot-sized reduplication.

1.2. Foot-Sized Reduplication

The range of what has been identified as foot-sized reduplication includes all categories of attested metrical feet, such as iambs and trochees. McCarthy, Kimper, and Mullin (2012) surveyed the range of bisyllabic reduplication patterns and found that, overwhelmingly, the pattern is to copy the first syllable exactly as in the base, but to have an open second syllable, as illustrated in Waalubal below. Notice that in the last example in (2a), even though the second syllable of the base ends in a consonant, it is open in the reduplicant. The pattern in Yidiny illustrates a case where the second syllable of the reduplicant is copied in its entirely, only if the second syllable of the base is closed, as the last two forms in (2b) illustrate. And the third pattern of foot reduplication is illustrated in Manam, where a moraic trochee is suffixed to the base. Moraic trochees contain two moras, which can be either two light syllables or a single heavy syllable, as the last three forms in (2c) show.

(2)

It appears then, that the pattern of foot-shaped reduplicants seems to be to copy the same prosodic structure as found in the first two syllables of the base, with a strong preference for an open final syllable. This differs from what has been observed for syllable-shaped reduplicants.

1.3. Syllable-Sized Reduplication

Some of the earliest typological research on reduplication noted that there is no syllable copying per se (Moravcsik, 1978). Instead, the reduplicant has an invariant shape to which base segments are mapped. For example, in Mokilese (Oceanic), the target shape is a heavy syllable. In (3a), this is satisfied by repeating the initial consonant of the second syllable (syllable boundaries of the base are indicated with a period). In (3b), the first syllable of the base contains a long vowel, satisfying the requirement for a heavy syllable and is repeated exactly. In (3c), the second syllable of the base is light and does not have an onset, so the reduplicant shape requirement is satisfied by lengthening the vowel of the first syllable.

(3)

The following example from Ilocano (Austronesian) illustrates that if syllable copying per se were found, we would expect to find patterns where the reduplicant is an exact copy of the first syllable of the base, as indicated in the starred examples. This doesn’t appear to be the case.

(4)

Findings like this, regarding having a target shape for syllable-sized reduplicants, are what prompted researchers to propose that reduplicative morphemes are templates (Marantz, 1982). Templates are segmentally empty phonological objects that are affixed to a stem, and they must be given segmental content. This research program developed alongside advances in autosegmental phonology (Goldsmith, 1976), in which different phonological content is expressed on different tiers. The extension of this insight to morphology, in which different morphemes are represented on different tiers in Semitic languages, allowed what appeared to be process-based word formation patterns to be analyzed as items (McCarthy, 1981). The segmental content of templates is provided by copying the base and filling the template, as illustrated by example (5).

(5)

Originally templates were composed of skeletal slots, like C and V (Marantz, 1982; McCarthy, 1981). However work in prosodic phonology and reduplication recast them as units of prosody (McCarthy & Prince,1986, et seq.). This has been formalized as the Prosodic Morphology Hypothesis.

(6)

More recently, some approaches to reduplicative shape eschew templates, arguing that the patterns found are constrained by the same mechanisms that determine morpheme shape in general. This approach to reduplicative shape has been referred to as Generalized Template Theory (henceforth GTT; McCarthy & Prince, 1999; Urbanczyk, 2001, 2006). Downing (2006) presents one of the most detailed proposals within this approach and takes the proposals of GTT one step further, by deriving canonical shape from a constrained set of principles on morpheme shapes. She is able to derive a wide range of patterns from principles of grammar, such as requiring lexical heads (such as roots and stems) to meet minimal requirements. It is highly constrained and requires little extra technical machinery, as the shape of reduplicants makes use of constraints needed for morpheme shapes in general.

Kennedy (2008) takes a different approach to determining shape. Rather than proposing that there are constraints on the shapes themselves, he argues that the shape properties follow from independently motivated constraints on morphological and prosodic alignment of categories. He also employs a number of constraints on well-formed morphoprosodic categories, so his approach indirectly derives the shape of reduplicants via these. This approach makes several predictions about the number, size and location of reduplicants in any single language. Challenging this approach are reduplicative morphemes that do not coincide with any prosodic or morphological boundaries.

A not uncommon pattern of reduplication occurs where the shape of the reduplicant does not coincide with a prosodic category, such as –VC(C). This is illustrated in Lushootseed and Mangarayi below. The repeated portions are separated by dashes.

(7)

While the shape of the reduplicant does not coincide with any prosodic unit, it does add a syllable to each word. This supports approaches in which the reduplicative morpheme is a segmentally empty syllable template or is derived via constraints on morpheme shapes. A further interesting feature of this pattern is that the shape is found almost exclusively after the first consonant. Various different researchers propose analyzing this as either a prefix that subcategorizes to be located after the first consonant (Pivot Theory by Yu, 2007), or a suffix that migrates in to be after the first consonant (Urbanczyk, 2006).

An entirely different approach to partial reduplication can be described as copy and truncate (Inkelas & Zoll, 2005; Steriade, 1988). In these models, total reduplication is followed by truncation to achieve the target shape. More will be discussed about this section 5.4 “Repetition Mechanism.”

1.4. A-Templatic Reduplication

A final set of patterns do not fit neatly into total reduplication or having a consistent prosodic shape. These are cases of multi-pattern reduplication, where the reduplicant is associated with more than one shape. In one type of multi-pattern system, copying involves one or more segments. The following data from Temiar (Mon Khmer) illustrate two ways in which the repetition of segments does not result in a fixed templatic shape (syllable boundaries are indicated with a period). When the base is biconsonantal, the simulfactive is formed by copying the first consonant and having an [a] as the nucleus, as in (8a). When it is triconsonantal (8b), the vowel is infixed between the first two consonants. This illustrates that copying serves to supply an onset, rather than meet some target shape for what is repeated. The continuative is likewise a-templatic because when the base is biconsonantal, both consonants are copied. However, when the base is triconsonantal, the final segment of the base is repeated after the first consonant.

(8)

Cases like this have been presented as evidence that reduplication need not be triggered by a templatic affix in the input. Copying segments occurs in order to satisfy overall shape requirements of the whole stem (Gafos, 1998), or by requirements on morpheme alignment alone (Hendricks, 1999, 2001). A key debate surrounding patterns like this relates to whether or not the repetition of segments is morphologically driven as part of the reduplication process or is phonological in nature. This is discussed further in section 5.4 “Repetition Mechanism.” What can be seen from the discussion in this section is that there is a connection between deriving shape properties and the nature of the repetition mechanism.

A second type of multi-pattern system relates to what Spaelti (1997) refers to as alloduples in which there are multiple shapes associated with one meaning. The shapes are not driven by any template; for the most part, the variation in shapes can be predicated based on phonological properties of the base (see for example, the languages discussed in Spaelti, 1997). Hawaiian presents an interesting pattern where the different shapes –even monomoraic reduplicants– are analyzed by different ways to realize a minimal word (Alderete & MacMillan, 2015). We turn our attention next to survey the key patterns found related to the segmental quality of reduplicants.

2.1. Identity

Wilbur (1973a, 1973b) identifies situations in reduplication in which phonological rules either fail to apply with the proper context—under-application—or apply without the appropriate triggers—over-application. For example, in Madurese, a process of nasal place assimilation fails to apply only in the context of reduplication, as illustrated in (9).

(9)

A well-studied pattern of over-application occurs in Tagalog (Austronesian). A process of nasal substitution applies, in which a nasal plus stop coalesce; the resulting segment retains the nasal manner and the place of the following stop, as can be seen in the first two forms. This process over-applies in reduplication, where both reduplicant and base have the fused segment, even though the context for the process is not found in the base of the surface form. Unattested forms that illustrate normal application are indicated at the far right.

(10)

The majority of these cases can be analyzed by interleaving phonological processes with the reduplication mechanism. For instance, with the Madurese example, nasal place assimilation would apply first, then reduplication would apply. Even though reduplication potentially feeds the process of nasal place assimilation, because it occurs after the phonological process, it does not apply. With over-application, as in Tagalog, the rule applies before reduplication, and reduplication copies the modified initial segment of the stem.

However, there are a few cases that cannot be accounted for by rule ordering, as illustrated in the over-application of nasal harmony in Malay (11). There is no ordering that can derive nasality on the first syllable of the word, as there is no context prior to, or after reduplication for the nasality.

(11)

Patterns like these led researchers to propose that an Identity relation exists between the reduplicant and the base (McCarthy & Prince, 1995, 1999; Wilbur, 1973a, 1973b). The most recent model put forward to account for this is couched within Optimality Theory (Prince & Smolensky, 1993). McCarthy & Prince propose that for reduplicated words, there is a Correspondence Relation between base and reduplicant as well as between input and output. The basic Base-Reduplicant Correspondence Theory (BRCT) model is illustrated below.

(12)

A set of Faithfulness constraints evaluate the identity between the base and reduplicant (BR-Faith), as well as input and output (IO-Faith). Identity effects are achieved when BR-Faith is high-ranking and hence obeyed. This Correspondence relation is formalized as in (13) below, where the base and input are S₁ and the reduplicant and output are S₂.

(13)

This generalized Correspondence relation is proposed to exist between any morphologically related strings. The full BRCT model also includes a correspondence relation between the input and reduplicant (IR) and is motivated by patterns of reduplication in which the reduplicant contains input material not present in the base. See McCarthy and Prince (1999, p. 232) for discussion of the limiting effect of the IR-relation. Struijke (2002) is able to achieve similar effects to IR-Faith by modifying how Faithfulness constraints are formalized, proposing that they are existentially quantified rather than universally.

2.2. Non-Identity

There are several widely attested situations in which there is a lack of segmental identity between base and reduplicant. When phonological processes happen as expected, the pattern has been described as normal application. Nothing special needs to be said about these cases, as the phonological processes apply to the word as a whole. The remainder of this section outlines two further patterns of non-identity: when the reduplicant eliminates marked structure, termed ‘emergence of the unmarked’ (henceforth TETU), and when the reduplicant has some form of fixed segmentism, not analyzable as a TETU effect.

The ʔay̓aǰuθ‎əm (Central Salish) plural forms below illustrate two ways in which unmarked patterns emerge in reduplication. The pattern in (14a) illustrates that schwa, the default epenthetic vowel occurs in the reduplicant. The pattern in (14b) illustrates that glottalized sonorants, while occurring in the language as a whole, do not occur in the reduplicant.

(14)

In addition to segmental TETU effects, there are prosodic ones as well. Sanskrit illustrates an example where complex onsets are eliminated in reduplication, as can be seen below.

(15)

Instances of TETU show exceptional phonological patterning, in the sense that a process applies only to the reduplicant. This leads to a mismatch in identity between base and reduplicant.

In terms of mechanisms to account for TETU effects, McCarthy and Prince (1994, 1995, 1999) derive the pattern by interleaving markedness constraint(s) between IO-Faith and BR-Faith. The base permits the marked structure, hence violating the Markedness constraint. With low ranking BR-Faith, the reduplicant can diverge from the base, to obey the higher ranked Markedness constraint.

(16)

For the Sanskrit pattern, the relevant Markedness constraint would ban complex onsets. The reduction of the vowel to schwa in ʔay̓aǰuθ‎əm is an extreme case of TETU, where a default segment is found, rather than simply eliminating an unmarked feature.

A range of patterns has been found for fixed segmentism (Yip, 1992). In addition to the default segmentism illustrated above, another pattern of fixed segmentism is quite common, found to accompany what is known as echo-word formation. This can be seen in the English pattern of schm- reduplication. These involve total reduplication of the base and a portion of the reduplicant is replaced with schm-.

(17)

This pattern has been analyzed as affixation, where the fixed segments replace the segments of the reduplicant. Alderete, Beckman, Benua, Gnanadesikan, McCarthy, and Urbanczyk(1999) identify several criteria to distinguish default segmentism from these affixal segments. Key points are that the segments tend to be more marked, as the cluster in English shows. A second property is that they tend to over-write the segments of the reduplicant (as proposed by McCarthy & Prince, 1990) and inherit the length of the original base. Tamil illustrates the basic pattern with ki as the over-writing segments: when the base has a short vowel, the fixed segment is short (18a). However, when it is long, the fixed segment inherits the length from the base (18b).

(18)

A third property of over-writing segments is that they must over-write a unit, rather than being a pure affix. The Urdu-Hindi echo-words below show that the fixed segment is v-. Nevins (2005) points out that if the segments were purely affixal, then one would expect a form like *roti-vroti for ‘bread and the like.’ with the segment simply affix; instead we find it over-writes the base segment (19a). The form in (19b) shows that the v- can replace a segment before r, showing that the sequence vr- is allowed in the language as a whole. A fourth property of over-writing segments is that they can show dissimilation, similar in spirit to what is found in haplology (Alderete et al., 1999). If the base begins with v, then a different segment is used, as illustrated in (19.c).

(19)

The key issues facing any analysis relate to the transfer effect of the vowel length, the dissimilation, and why the initial segment(s) of the reduplicant must be over-written. This last point leads us to issues related to the morphological structure of reduplicated words.

3.1. Parallels With Other Morphological Categories

As noted above, there is a range of shapes from total reduplication to single segment reduplication. Total reduplication has been most often analyzed as some form of compounding, which in general is stem-combination. As pointed out, several shape properties of partial reduplication can be derived from reference to morphological structure, depending on whether or not a reduplicative morpheme is a root, stem, or affix. The question arises then as to whether or not reduplicative morphemes illustrate the same affixation patterns as segmentally fixed affixes.

If reduplication is a form of affixation, then we expect the same kinds of affixal patterns as we do with pre-specified affixes. However, there seem to be some interesting typological differences. For example, while suffixing in general is more prevalent than prefixing, the reverse pattern is found with reduplication. It is much more common to find prefixing reduplication than suffixing. Nelson (2003) proposes that all cases of reduplication are prefixing, and the apparent cases of suffixing reduplication are due to other factors such as affixation to a stressed syllable. Even more puzzling for a purely affixal approach to reduplication are languages that don’t have any prefixes, but do have reduplicative prefixing, as illustrated in ʔay̓aǰuθ‎əm (14), and in Nuu-chah-nulth (20a).

An additional pattern that is expected if reduplication is affixation is multiple exponence. This is a situation where reduplication accompanies affixation and has been described as affix-triggered reduplication. In this case, each affix has a reduplication pattern associated with it. Examples can be seen in Nuu-chah-nulth (Southern Wakashan) and Tagalog (Austronesian). Nuu-chah-nulth has a number of suffixes that trigger changes to the base, either by adjusting the length of the vowel, modifying the final consonant or by repeating all or part of the base. Note: [R] is used to indicate the reduplication-triggering suffix in (20a):

(20)

This pattern of reduplicative multiple exponence tends to occur in languages that have a great deal of affixation. While a thorough typological study of such patterns has yet to be conducted to support this claim, Nuu-chah-nulth has 600 suffixes, with approximately 90 that trigger reduplication, and Tagalog has roughly 160 verbal prefixes, with approximately 60 that trigger reduplication. Preliminary investigation into this pattern found that 80% of the reduplicative affixes in Nuu-chah-nulth have (near) homophonous counterparts.

There is relatively little theoretical work on the nature of affix triggered reduplication. Research on these languages tends to focus on deriving the shape of the reduplicants independent of the co-occurrence of the affixes. However, the overall pattern has been analyzed in word-and-paradigm approaches to morphology, where the reduplicated stem is a type of stem-formation operation required by classes of affixes (Lee, 2013; Saperstein, 1997). A key observation about the Nuu-chah-nulth pattern is that, when there are multiple affixes that trigger reduplication, at most two reduplicants are found (Lee, 2013; Stonham, 2007). This lack of multiple reduplications has interesting theoretical implications for understanding the nature of reduplication and provides evidence that the reduplication in such languages is what Yu (2007) calls “compensatory reduplication.”

In terms of the range of reduplicative affixes found, we have seen prefixes, suffixes, and examples of multiple exponence (or circumfixes). In addition, there are many cases of infixing reduplication. The location of reduplicative infixing should also mirror other infixes. Interestingly, in Yu’s (2007) examination of infixing, it appears that larger infixed reduplicants are located after the final vowel of a stem, while only the smallest ones are located near stressed syllables.

3.2. Morphological Structure

In many languages, the reduplicant has other morphological material associated with it, as found in Kinande in (21a) and (21b). As you can see below all reduplicants are vowel-final. Downing (1999, 2006) analyzes this as the inflectional final suffix vowel that occurs with other verbs. In Bantu, all verb stems must end in a final vowel, including reduplicated stems as well.

(21)

Downing proposes that the reduplication pattern above is accounted for with a stem template, which is composed of a root plus the inflectional final suffix. Crucially, this cannot be classified as copying a stem, because affixal material associated with the stem (-ir, -ir-an) is not copied. Rather, the reduplicative stem itself has internal morphological structure.

Notice that affix-triggered reduplication differs from the pattern found in Bantu languages. While one can describe the patterns in Nuu-chah-nulth and Tagalog as affixes being accompanied by reduplication, the Bantu pattern is more accurately described as the reduplication being accompanied by an affix.

3.3. Other Morphological Effects Unique to Reduplication

A key pattern that has been the topic of much research on reduplication regards the proper treatment of what has been termed ‘exfixation.’ This pattern was introduced in example (10) in Tagalog, in which an affix is copied alongside the root/stem material, under specific conditions. Its relevance to the proper treatment of reduplication was first pointed out by Bloomfield (1933) and has been given various theoretical analyses as models of phonology and morphology have evolved over the years (Aronoff, 1988; Inkelas & Zoll, 2005; Marantz, 1982; McCarthy & Prince, 1999; Wilbur, 1973, among others). In Tagalog, prefix and root segments fuse, so the resultant segment belongs to both root and affix segments. Other types of exfixation occur where affixes are copied alongside root material in order to meet prosodic requirements, such as having an onset, or minimality conditions on the shape of the reduplicant. The data from Kíhehe illustrate this pattern, where vowel-initial stems copy the prefix to supply an onset to the reduplicant in (22b). Odden and Odden (1985) argue that this presents evidence for phonological processes (such as glide formation) happening prior to reduplication.

(22)

One of the earliest proposals for how to analyze exfixation morphologically treats reduplication as a morphological head operation (Aronoff, 1988), where the repetition mechanism targets the head of the word. This allows the initial fused segment of the stem to be copied. A second approach has been offered, wherein the reduplicant is infixed to a prosodic stem, rather than a morphological one (Downing, 1999; Inkelas & Zoll, 2005).

A further interesting morphological effect that is unique to reduplication is what Mutaka and Hyman (1990) refer to as a Morpheme Integrity Constraint. In Kinande (Bantu), the reduplicant is bisyllabic. This target shape is met by a variety of means; by copying the root exactly if it is two syllables (23a), by incorporating prefix material if it is not (23b), and by doubling up the reduplicant if there is no prefix material to incorporate (23c). Interestingly, if the base exceeds two syllables, as in (23d), reduplication does not occur.

(23)

To account for patterns like (23d), where no reduplication is possible, Mutaka & Hyman propose the following constraint.

(24)

So, if the morpheme to be copied doesn’t match the template exactly, no copying can occur. Related to this is a set of patterns in which there is a different reduplicative allomorph if the base does not match the template exactly. This is often associated with foot-sized reduplication. For example, in Makassarese, bisyllabic bases are copied exactly as in (25a). However, with larger bases, the first two syllables are copied and then a glottal stop is epenthesized at the end as in (25b).

(25)

Notice that, regardless of whether or not the base ends in a consonant in (25a), it is copied exactly. The crucial factor in analyzing these patterns is that the base is exactly a foot. Downing (2000) proposes that the pattern is motivated by obeying a constraint that requires total copy of the base (BR-Max) conjoined with a constraint requiring every feature in the reduplicant to be in the base (BR-Dep-F).

4.1. Morphological Repetition

There seem to be two basic approaches to reduplication as morphological repetition. In one, reduplication is seen as a type of morphological operation, akin to other types of operation (Aronoff, 1988). The other involves repetition of morphological units (Inkelas & Zoll, 2005; Pulleyblank, 2009).

As discussed above, Aronoff (1988) accounts for cases of exfixation as a word-level head operation. This is motivated by the existence of word-level head operations with non-reduplicative morphological operations.

Morphological repetition is when the morphology calls for something twice. The details of how this works have evolved alongside developments in morpho-syntactic theory. One key model is Morphological Doubling Theory (MDT; Inkelas & Zoll, 2005). In MDT the morphology calls for semantic features twice. The basic operation associated with reduplication is illustrated below, where semantic features are doubled, not a stem, root, or lexical item.

(26)

Lexical insertion applies during spell-out to provide phonological content. A variety of operations can apply to one or both daughters of the compound to account for partial reduplication and segmental changes. Evidence to support this semantic approach is presented in the form of synonymy compounding, as found in Khmer and Vietnamese.

(27)

Inkelas & Zoll provide a number of other arguments to support semantic doubling such as patterns of divergent stem allomorphy and a number of case studies that they re-analyze. Crucially, there is no phonological identity relation. Notice that in this framework, nothing is unique about reduplication except semantic doubling. It should be noted that not all researchers recognize this as reduplication. An important aspect of MDT is to also admit phonological copying, the details of which will be presented below.

4.2. Phonological Repetition

Phonological repetition, on the other hand, includes a plethora of different mechanisms, depending on the other aspects of the model of reduplication.

Some of the earliest research in reduplication proposed a copy mechanism, in which the base was literally copied (Marantz, 1982). In segment copying approaches, various principles ensure that segments are associated to templates either according to universal tendencies or by language-specific means (Marantz, 1982). For example, adjacency of base and reduplicant follow from association conventions being left-to-right for prefixes and right-to-left for suffixes.

Many variations on this copy-and-associate model have been proposed, and it is compatible with most templatic approaches. This type of template-driven copying has re-emerged in Harmonic OT, in the form of Serial Template Satisfaction (STS; McCarthy, Kimper, & Mullin, 2012). Only phonological or morphological constituents are copied.

Some phonological approaches do not have a copy mechanism per se. Within some derivational frameworks, reduplication is accomplished by a separate linearization process. In one version, the reduplicative template is a parafix, attached above the word on a separate tier (Clements, 1985; Mester, 1986). The following illustrates how this accounts for exfixation pattern in Tagalog. Linearization occurs after nasal substitution.

(28)

Evidence to motivate parafixation includes syllable transfer effects and segmental identity effects. Linearization can take place either before or after phonological rules apply. If linearization occurs after, then the identity effect is achieved.

A second type of linearization model uses special notations to indicate precedence relations among segments (Frampton, 2009; Raimy, 2000). The existence of a loop is what leads to repeating a sequence of segments, and reduplication is a type of readjustment processes.

(29)

The loop identifies where the repetition occurs. In the example above, the fixed segments schm- follow the final consonant of the base word, indicated by the downward facing arrow. The ʃ precedes m, indicated with a leftward arrow. The loop goes from the m to the a, which is then repeated when the linearization process happens.

In direct opposition to this model is the Correspondence Theory model of reduplication (BRCT; McCarthy & Prince, 1995, 1999). As introduced at example (12), BRCT is couched in Optimality Theory and proposes that there is a set of relations that exist between morphologically related strings. The phonological content of reduplicants is therefore morpheme-driven, with various proposals regarding what the input reduplicative morpheme is. In this model, the only difference between reduplicative morphemes and other morphemes is that there is (for the most part) no segmental information associated with the input. Correspondence Theory has been generalized to include other relations for different morphological domains, such as between output words (Benua, 2000) or between other prominent positions in words such as roots and stressed syllables (Beckman, 1998).

Constraints evaluate the identity of morphologically related strings via a number of Faithfulness constraints that are defined according to a Correspondence Relation. The different patterns found with reduplicative morphemes are due to the different rankings of BR-Faith constraints, with Markedness and IO-Faith constraints resulting in different grammatical systems. The key feature of BRCT is that reduplication is a matter of identity, with constraints demanding identity between all aspects of reduplication and its base. Any divergence from total reduplication or identity must be compelled by some higher ranked constraint. In this model, there is no real copy mechanism per se, but rather there is a relation that exists between strings in words.

4.3. Mixed Approaches

A growing body of research on reduplication has proposed that there are multiple mechanisms for repeating segments, depending on the properties of the reduplicative morphemes. Various criteria have been proposed for identifying the appropriate copy mechanism (Inkelas & Zoll, 2005; Pulleyblank, 2009; Saba Kirchner, 2010). While the details of these approaches differ, they all share the property that there is one component that copies segments to satisfy conditions on phonological well-formedness and another component that doubles up some form of morpho-syntactic unit. They also eschew approaches that have an explicit base-reduplicant identity relation between the segments.

Saba Kirchner’s Minimal Reduplication (MR) model (2010) differs from MDT in a few key ways. First he proposes that reduplication is not maximizing, but that repetition only occurs when it is compelled, from the phonology, morphology, or syntax. A representation of the model is presented below. Phonological and morphological reduplication are achieved by violating the Faithfulness constraint Integrity, which requires an input segment only to have a single output correspondent. Therefore the repetition mechanism is one in which an input segment has two output correspondents.

(31)

Notice that both phonological and morphological reduplication are triggered when markedness and faithfulness constraints { C, F } dominate Integrity. Repetition comes at a cost, so is minimal (see also Pulleyblank, 2009). This differs from syntactic reduplication, which results from two syntactic operations: copying a target constituent and fusing the copy with a trigger morpheme into a single morphosyntactic unit. Because the repetition of segments is via Integrity violations or copy/merger, BR-Correspondence is superfluous and is not a component of the model.

A key feature of approaches with mixed methods of repeating segments is that each model is accompanied by diagnostics of when the different methods are employed. As such, they make important contributions to defining what is considered reduplication vs. other forms of segmental repetition. Inkelas and Zoll (2005) identify four criteria for when phonological copying occurs in MDT. These include a phonological requirement for copying (such as requiring an onset), proximity of the repeated segments, targeting segments only (rather than larger constituents), and a requirement for phonological identity. In MDT this is achieved via segmental identity relation between similar output segments (Hansson, 2001; Inkelas, 2008; Rose & Walker, 2004). Again, this differs from MR, which accomplishes phonological copying via Integrity violations. The final section summarizes some of the critiques that have been leveled against key models of reduplication.

5.1. Shape

In terms of accounting for reduplicative shape, models should be able to account for gaps in the patterns. That is to say, models that generate unattested patterns have too much power and are faced with having to impose conditions or constraints on why these patterns do not occur.

First, there are no reduplicative patterns that are three syllables. This is predicted by models in which the shape is determined, either directly or indirectly, by the prosodic hierarchy. Because there are no prosodic units that are exactly three syllables long, this is as expected. However, models that involve total reduplication and truncation could derive this pattern by operations that truncate to a foot plus one syllable. Japanese loanword truncations have several different forms, and one of these is trisyllabic (Itô, 1990).

A second gap relates to the size of infixed reduplicants: they never occur with total reduplication and are generally only a syllable or segment in size. This is predicted by those models in which total reduplication is seen as a type of compounding. However, the copy-and-truncate approaches predict that this should be possible, as the location of the infix is independent of the mechanism that truncates material.

Third, with models that derive shape via a truncation mechanism, we expect to find the full range of truncation patterns in reduplication. For example, Koasati (Muskogean) forms plurals in part by stem-final consonant deletion and rhyme deletion (Horwood, 2001). No patterns of reduplication parallel to these truncations patterns have been attested yet.

Finally, an important gap was claimed to occur in which the base back-copies templatic features. For example, a word like badupi could reduplicate as badu-badu. This is predicted to occur in BRCT if reduplicative shape is derived via templatic constraints (the ‘Kager-Hamilton problem’ as discussed by McCarthy & Prince, 1999). Because identity between reduplicant and base is essential to BRCT, making it is possible to back-copy segmental qualities, one also expects to back-copy shape properties as well. GTT is a way to address this issue, as the shape properties are derived via independent constraints on morpheme shape rather than via templates. However, Caballero (2006) discusses precisely this type of pattern in Guarijio (Uto-Aztecan), of abbreviated reduplication. She develops an analysis within MDT of the pattern, showing that GTT cannot account for the facts.

5.2. Segmentism

While some models of reduplication, like BRCT, have segmental identity built into them, the majority of models do not. The question then arises as to how these models compare when it comes to accounting for some patterns of segmentism. Leaving aside fixed segmentism, the two patterns that will be discussed are TETU and the application of phonological processes.

TETU effects seem to be found primarily in monosyllabic reduplicants. For example, it is rare or unattested for total or foot-shaped reduplication to eliminate marked prosodic structure in just the reduplicant, such as *ka.ni.ba-kram.ni.blan, which eliminates codas and complex onsets in the reduplicant. It would appear that most other models can derive this unattested pattern of TETU effects outside of syllable sized. For example, BRCT could easily delete all codas in the reduplicant as a TETU effect. Also, ‘repeat & truncate’ models can do separate operations to separate daughters, so it would be possible to apply the cluster simplification independently of truncation. One model that is able to account for this pattern is Serial Template Satisfaction (STS; McCarthy et al., 2012). STS copies entire strings, without skipping, so could never produce a form that deletes every coda in the base. This raises the question as to how STS can achieve the Sanskrit pattern, which eliminates complex onsets in the reduplicant. McCarthy et al. do so by copying the full syllable in the first step and then simplifying the onset in a second step, treating the reduplicant like any other affix in the language as a whole. The analysis is supported because Sanskrit affixes in general also lack complex onsets. It seems, therefore, that STS is the most constrained, predicting that TETU effects are only found in reduplication, when the same patterns are found in the language as a whole.

Models without an identity relation achieve segmental identity effects by having separate sets of constraints to evaluate separate aspects of the word, and often involve some form of cyclic application of rules. However, MDT differs from most of these in having separate phonological operations specified for each daughter. The question arises as to why processes need to apply to both. One could easily derive a pattern in which one daughter undergoes no processes at all. This problem does not occur with other models because processes apply to the whole word.

5.3. Morphological Aspects of Reduplication

While a great deal of research has aimed to reduce the number of reduplication-specific operations or relate them to other attested morphological or phonological operations in language, reduplication is in fact a unique morphological operation. In terms of providing a critique of the scholarship on this, the focus will be on whether the models can account for the unique morphological aspects of reduplication.

Recall there are several cases in which the pattern of reduplication depends on whether the templatic shape of the reduplicant matches the size of the base. There are a number of different options, depending on whether the base is smaller or larger than the reduplicant. If it is too small, the base can duplicate more than once. However if it is too large, then either reduplication does not occur at all, or the reduplicant has a different allomorph. Analyses have been presented within BRCT that account for this effect (Downing, 2000; McCarthy & Prince, 1994). However, it is unclear how models that don’t have an identity relation do this without simply stipulating that there is an MIC. In particular, the double-repetition pattern as illustrated in Kinande (23c) is particularly challenging for models in which repetition is of morpho-syntactic or semantic units, rather than stems. How does double reduplication occur to satisfy one meaning? The option of triplication is available in MDT, but only of semantic units; it is entirely unrelated to the size of the morphological units being doubled, and is not phonologically conditioned.

5.4. Repetition Mechanism

Finally, it is clear from the range of patterns that are found and the growing body of research advocating multiple methods of repetition, that minimally, two mechanisms are needed to account for the entire range of patterns found. However, the mechanisms by which this repetition is achieved vary significantly from theory to theory. In MR, phonological and morphological reduplication are repair mechanisms. Syntactic reduplication calls twice for some unit. Most models of reduplication treat the repetition mechanism as separate from other aspects of reduplication. BRCT seems to be the one model that can look for correlations between aspects of reduplicant form. Especially within GTT, there is a prediction that different morphological categories of reduplicants will exhibit correlations between shape and segmentism, which appears to be the case for Lushootseed (Urbanczyk, 2006). Given that there is a body of research that eschews BRCT, there is a minor void in proposing what can replace it.

In MDT, all reduplication is compounding followed by truncation. Interestingly, this makes predictions about what operations are available to languages—if the primary mechanism to achieve partial reduplication is compounding followed by truncation, then there are predictions about whether or not these operations occur in the same language. A question arises as to why they might not all co-occur. For example, Salish languages have extremely rich and diverse patterns of reduplication, yet do not have productive compounding or truncation.

Finally, because the basic mechanism is semantic duplication in MDT, the same range of patterns should be found for synonymy compounding as reduplication, which so far are unattested.