Morphemic harmony as featural correspondence

Available online at www.sciencedirect.com

Lingua 119 (2009) 478–501 www.elsevier.com/locate/lingua

Morphemic harmony as featural correspondence§ Sara Finley * Johns Hopkins University, United States Received 18 September 2007; received in revised form 16 September 2008; accepted 18 September 2008 Available online 14 November 2008

Abstract This paper presents a classification of vowel harmony, separating prototypical phonological harmony from morphemic harmony, which is triggered by the presence of particular morphemes. Morphemic harmony is driven by featural correspondence constraints requiring the realization of a feature across an output domain; phonological harmony is driven by markedness constraints requiring featural identity across an output domain. Morphemic harmony must always impose a fixed featural value on the form, which cannot be accounted for using approaches to phonological harmony. Phonological harmony cannot be incorporated into an analysis of morphemic harmony because morphemic harmony is represented only in terms of triggers, while exceptions to phonological harmony are generally represented in terms of targets. Typological predictions to the correspondence approach to morphemic harmony are also explored, including gapped representations, edge-only effects, and the variability of the realization of multiple features in morphemic harmony. Extensions of the present proposal are provided, illustrating that general featural affixation can be accounted for with morpheme-feature correspondence. Alternatives are also discussed. # 2008 Elsevier B.V. All rights reserved. Keywords: Vowel harmony; Featural affixation; Mutation; Morphology

1. Introduction This paper presents a formalization of the morphological influences on vowel harmony, distinguishing between two types of vowel harmony: phonological and morphemic (Cole, 1991). These types of vowel harmony differ both in their empirical instantiation as well as their formal implementation, particularly via the markedness/ faithfulness distinction of Optimality Theory (OT) (Prince and Smolensky, 1993/2004). Phonological harmony results from general agreement of features within an output domain, and is induced by markedness constraints that require generalized featural agreement. Morphemic harmony is the result of a feature or subset of features functioning as a morpheme of their own, and follows from faithfulness constraints to the harmonic feature associated with the harmony-inducing morpheme. Drawing on the insights of previous work (Akinlabi, 1996; § This work would not be completed without the helpful advice, comments and suggestions of the following: Paul Smolensky, Luigi Burzio, Paul de Lacy, Jane Grimshaw, John Alderete, Jaye Padgett, Joe Pater, Matthew Wolf, and several anonymous reviewers. I would also like to thank the audiences at HUMDRUM 2004, HUMDRUM 2007, the 2005 LSA Annual Meeting, and the 2005 Harvard ISOKL for their attention, criticism and comments. All errors are my own. This work was supported by NSF IGERT and a Jacob Javits graduate fellowship to the author. * Correspondence address: Department of Brain and Cognitive Sciences, University of Rochester, Meliora Hall, Rochester, NY 14627-0268, United States. E-mail addresses: [email protected], [email protected].

0024-3841/$ – see front matter # 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.lingua.2008.09.009

S. Finley / Lingua 119 (2009) 478–501

479

Cole, 1991; Rose, 1997; Zoll, 1996), this paper presents a novel faithfulness-based correspondence approach to morphemic harmony that forms a typology of morphemic harmony. I also argue for a distinction between morphemic and phonological harmony. In morphemic harmony, a morpheme’s realization demands that the vowels in the domain agree in a feature value. Kanembu contains a classic case of morphemic harmony in which the incompletive is marked with the feature [+ATR] realized on all vowels in the word, as in /g n kI/ ([ ATR]) [gon ki] ([+ATR]) ‘take’ (Roberts, 1994). Such morphemes have been labeled featural affixes (Akinlabi, 1996) because the morpheme is realized not with segmental material, but with featural material. An instance of featural affixation is a case of vowel harmony if the feature associated with the affix is required to surface on all vowels in a domain, imposing agreement. In this paper, I argue that morphemic harmony is realized via a correspondence relationship between the feature and the segments of the output, in which the harmonic feature (e.g., [+ATR]1) is co-indexed all of the vowel in the output form (e.g., [go1n 1ki1]). This correspondence is governed via the constraints ANCHOR and O-CONTIGUITY. ANCHOR places the harmonic feature onto the left and right edges of the output and O-CONTIGUITY connects the correspondences between the left and right edges. The typology of interactions between these correspondence constraints with featural markedness and featural identity are borne out. First, featural anchoring (as opposed to morpheme alignment) predicts that morphemic harmony systems with multiple features may realize each of these features on different segments in the output, as found in Korean ideophones. Second, because O-CONTIGUITY, the correspondence-theoretic constraint preventing gapped representations, is violable, gapped representations should be attested, as found in Lena Asturian. Third, because the harmony-inducing constraints are morpheme specific, the present approach predicts that a language may have a morpheme-specific vowel inventory, with segments only appearing in the presence of a particular morpheme, as in Pasiego Spanish. This entails that no morpheme will induce general agreement of vowel features (e.g., leftward spreading of the [+/ ATR] feature of the stem). No counterexamples to this prediction have been found. The paper is organized as follows; I begin with an explication of how correspondence constraints (McCarthy and Prince, 1995) account for morphemic harmony, with Kanembu as a test case (section 2). This is followed by a presentation of the factorial typology of the correspondence approach to morphemic harmony. Included are discussions of Korean, Lena Asturian and Pasiego Montan˜es (section 3). The approach is extended to consonant-based featural affixation in Chaha and Terena (section 4). These analyses are followed by a discussion of why morphemic harmony must be analyzed differently from phonological harmony. First, the morpheme-specific nature of morphemic harmony is always couched in terms of triggers, while morpheme-specific effects of phonological harmony generally apply to undergoers, targets. This makes it impossible to extend the correspondence approach to morphemic harmony to phonological harmony (section 5). Second, morphemic vowel harmony requires both agreement of vowels as well as the realization of a specific feature value, while phonological vowel harmony only requires general agreement. This makes it impossible to incorporate morphemic vowel harmony into accounts of phonological harmony (section 6). Third, idiosyncrasies that are found in morphemic harmony, such as edge-only effects and arbitrary skipping of segments occur in morphemic vowel harmony but not phonological vowel harmony. Alternative proposals are also discussed (section 7). 2. Morpheme-feature correspondence This section provides an account of morphemic harmony using morpheme-specific correspondence constraints. I begin by examining the Kanembu, a Nilo-Saharan language (Hall et al., 1973; Jouannet, 1982) that displays vowel harmony in the alternation of incompletive and completive verb forms; completive verbs are [–ATR] ([a, , , I, ]) and incompletive verbs are [+ATR] ([ , o, i, e, u]) (Akinlabi, 1994; Jouannet, 1982; Roberts, 1994), illustrated in (1) below. While Kanembu does not appear to have a regular phonological system of vowel harmony,1 the incompletive alternation may be a relic of a harmony process that existed at one time.

1

Regardless of whether vowel harmony is a regular process in Kanembu, (though there are disharmonic items such as ‘[ (Jouannet, 1982), the completive/incompletive status of every verb is realized by the ATR value of the vowel.

]‘encircle’)

480

(1)

S. Finley / Lingua 119 (2009) 478–501

Completive/Incompletive Alternations in Kanembu (Akinlabi, 1996)

The Kanembu incompletive morpheme is realized via (i) realization of the feature [+ATR] and (ii) realization of that feature on all vowels in the lexical item.2 Morphemic vowel harmony can be straightforwardly analyzed if the harmony-inducing morpheme is realized only when the feature it is associated with is in correspondence with all vowels in a lexical domain. In morphemic harmony, the input feature is preserved in the out by spreading across the entire output domain. Morpheme-specific versions of L-ANCHOR, R-ANCHOR and O-CONTIGUITY3 (constraints (2)–(4)) achieve this effect. (2)

L-ANCHOR-[ F]-MORPH: The [ F] feature of MORPH in the input must be in correspondence with the leftmost edge of the lexical domain in the output (which therefore must be [ F]).

(3)

R-ANCHOR-[ F]-MORPH: The [ F] feature of MORPH in the input must be in correspondence with the rightmost edge of the lexical domain in the output (which therefore must be [ F]).

(4)

O-CONTIGUITY-[ F]-MORPH: The output vowel [F] features in correspondence with a MORPH [ F] feature must form a contiguous string.

These constraints are feature-based versions of the original correspondence constraints (McCarthy and Prince, 1995) as well as the proposals made by Akinlabi (1996) and Rose (1997). The proposal in this paper draws mainly from Rose (1997) in which morpheme realization is achieved via correspondence with the difference that the present proposal requires a correspondence between features and edges (as opposed to correspondence between morphemes and edges). Many researchers have taken advantage of correspondence relations between a segment and a feature. Zoll’s MAX[SUBSEGMENTAL] constraint (1996, 2001) requires correspondence between a feature and a segment. The index of the feature percolates from the feature to the root node of its host. Correspondence between a feature and a segment is also employed by existential faithfulness (Struijke, 2002), in which faithfulness is achieved if an input feature is preserved on an output segment. Klein (1995) uses feature-based ANCHOR constraints for floating autosegments in umlaut. The representation of feature-percolation for feature-segment correspondence is provided in (5) below. The first three lines represent the elements of the input. The featural affix is in correspondence with the ATR features of each of the vowels in the output. This correspondence relationship can be represented as a single, multiply linked autosegment, or as several individual autosegments, with equivalent constraint profiles.

The proposed constraints require a correspondence relation between the featural affix and the features in the lexical item. In principle, both consonants and vowels may serve as docking sites for a floating feature (as in Chaha and Terena, discussed in section 3). Because it is uncertain whether consonants are imperceptible targets of spreading, or

2 3

I assume that completive verbs also undergo morphemic harmony, but with the [ ATR] feature. Abbreviated versions may appear in some tableau to save space: ANCHOR: ANCH, and O-CONTIGUITY as O-CONT.

S. Finley / Lingua 119 (2009) 478–501

481

are simply skipped due to high-ranked constraints preventing vowel features from appearing on consonants, consonants are ignored in morphemic vowel harmony. The method for assigning violations (illustrated in the tableau in (7)) is based partially on insights from Klein (1995). Evaluations for L-ANCHOR (and analogously for R-ANCHOR) begin at the left edge of a candidate. If featural correspondence is found, no violations are assigned. If there is no correspondence, a violation is assigned, and the correspondence is checked on the next vowel to the right. If there are no correspondents in the word, ANCHOR is violated once for every vowel in the word. ANCHOR constraints are therefore gradient and realizational (Kurisu, 2001; Rose, 1997). The realizational component is based on Rose’s (1997) proposal that a morpheme must be realized on the lexical item in order to satisfy ANCHOR (ANCHOR cannot be vacuously satisfied by deleting the morpheme). If the morphemic harmony feature is not realized on the lexical item, all vowels in the lexical item incur violations of ANCHOR. I assume that all vowels in correspondence with the harmonic feature of the morpheme are linked to a single autosegment, and that gaps in this autosegment’s associations represent gaps in the correspondence. Because violations are assigned in terms of the vowel that breaks the contiguity of the correspondence, one violation is assigned for each vowel that breaks the contiguous string, as illustrated in the hypothetic example in (6) below, which incurs two violations of O-CONTIGUITY.

Candidate (a) fails to surface because the [+ATR] feature is not realized on any vowel in the word. This incurs three violations of L-ANCHOR because there are three vowels from the left edge of the word that do not bear correspondence with the [+ATR] feature. Candidate (b) incurs one violation of L-ANCHOR because there is one vowel between the left edge of the word that does not bear correspondence with the [+ATR] feature. Candidate (c) satisfies both L and R-ANCHOR because the [+ATR] feature is realized on both edges of the word. However, the gap in correspondence causes this candidate to incur a violation of O-CONTIGUITY. The optimal candidate (d) only violates the lowest ranked constraint ID[ATR]. I assume that featural correspondence can only be achieved via featural agreement. In other words, GEN does not produce candidates in which feature correspondence is achieved without sharing the feature value of the input. This tactic ensures that the featural material of the morpheme is realized on the vowels it shares correspondence with, and is equivalent to an undominated constraint requiring segments in correspondence to share features (as in Rose and Walker, 2004). INTEGRITY (which penalizes an input value realized on multiple outputs) is low-ranked for all cases of morphemic harmony presented in this paper, and thus will not be presented in any analyses here. The ‘gradient’ effects of ANCHOR (not in Rose’s proposal) arise from the fact that the correspondence relation in question is concerned with all vowels in the lexical domain and the morphemic feature. The ANCHOR constraints in principle could be replaced with categorical MATCH constraints because each individual output vowel requires a 4

I assume the incompletive is derived from the completive, but any underlying form gives the same result.

482

S. Finley / Lingua 119 (2009) 478–501

correspondence relation (match) to the featural morpheme (Kager, 1999; McCarthy, 2003). Such a MATCH constraint would account for morphemic harmony that applies to all vowels in a lexical item by assigning a violation to any vowel not in correspondence with the morphemic feature. Morphemic harmony that applies only to specific edges would require additional constraints targeting these locations. McCarthy and Prince (1995) note that left and right ANCHOR constraints serve to prevent deletion at morpheme edges. In essence, featural versions of ANCHOR serve this same function. Because floating features have no default place of realization, the ANCHOR constraints prevent the floating feature from being deleted from the representation altogether. If the feature is not realized on the output, then it is deleted from the input. With the basis for the correspondence-based approach to morphemic vowel harmony established, I will now address the typological predictions of this proposal. 3. Typology/predictions of morphemic harmony The featural correspondence constraints presented in (2)–(4): L-ANCHOR, R-ANCHOR, and O-CONTIGUITY, induce morphemic harmony to various degrees depending on each constraint’s ranking with respect to IDENT. The factorial typology yields four different types of interactions, listed in (8) below. While it may appear at first blush that the patterns presented below are rarely, if ever, found in prototypical harmony systems, this is precisely the prediction made by the present proposal. Such patterns are expected to be more likely to emerge in morphemic processes (as opposed to phonologically driven processes). (8)

Factorial Typology of ANCHOR Type 1: L-ANCHOR, R-ANCHOR, O-CONT IDENT (6 Rankings) Full harmony: Kanembu, Korean Type 2: IDENT L-ANCHOR, R-ANCHOR; O-CONT freely ranked (8 Rankings) No Harmony; no morpheme realization: default Type 3: L-ANCHOR IDENT R-ANCHOR; O-CONT freely ranked (4 Rankings) R-ANCHOR IDENT L-ANCHOR; O-CONT freely ranked (4 Rankings) Harmony on one edge only (Edge-only effects): Korean, Ngbaka, Chaha Type 4: L-ANCHOR, R-ANCHOR IDENT O-CONT (2 Rankings) Harmony on both edges, with gaps: Lena Asturian, Lokaa

Type 1, found in Kanembu, occurs when a morpheme is realized on all vowels in the lexical item. This occurs when both ANCHOR constraints and O-CONTIGUITY outrank featural identity. Type 2 occurs when IDENT outranks all three harmony-inducing constraints. This is the default case for any language without morphemic harmony. Type 3 occurs when only one ANCHOR constraint outranks IDENT5 and the morpheme is realized on one edge. In Korean ideophones, high vowels alternate on the leftmost syllable only, but in Nbaka, an Ubangian language of Zaire, the definite form is realized by the fronting the rightmost vowel (Roberts, 1994). Type 4 occurs when O-CONTIGUITY is ranked low enough to allow gapping to occur, as in the Lena Asturian dialect of Spain (Ibero-Romance) (Hualde, 1989, 1992; Walker, 2005), Mafa (Ettlinger, 2008), Harari (Rose, 2004), and Lokaa (Iwara et al., 2003). The following sections provide data and analyses of Types 3 and 4. I also explore Pasiego Montan˜es (Hualde, 1989; McCarthy, 1984; Penny, 1972), which displays a pattern that arises when high-ranked featural markedness constraints interact with the constraints in (8) above, producing a morpheme-specific vowel inventory. 3.1. Edge-only effects: Korean ideophones Morphemic harmony alternations in Korean ideophones provide evidence for the correspondence-based approach to harmony. The most controversial aspect of Korean vowel harmony is its harmonic feature, as the alternation cannot be described with a single harmonic feature, a requirement for traditional approaches to phonological harmony. The correspondence approach however, allows for multiple harmonic features, creating a straightforward analysis of the 5

This type of harmony can also occur under the ranking R-ANCH bear the harmonic feature, as well as for high-ranked INTEGRITY.

O-CONT

R-ANCH

ID if O-CONTIG applies to consonants that cannot

S. Finley / Lingua 119 (2009) 478–501

483

alternations in Korean: [HIGH], which is realized on the initial vowel only and [ATR], which is realized through the lexical item. The vowel inventory of Korean is divided between ‘‘light’’ [(ø), o, , a] and ‘‘dark’’ [i, (y)6, i, u, e, ] vowels, with alternations illustrated in (9) below. These terms are used (as opposed to a more theoretically familiar term) because there is no distinct natural class that can easily describe the variation of ‘‘light’’ and ‘‘dark’’ vowels. The terms LIGHT and DARK are also used to describe the semantic variation found in ideophones, illustrated in (10) below. LIGHT forms are composed of all ‘‘light’’ vowels and have a fast connotation, while DARK forms are composed entirely of ‘‘dark’’ vowels, and have a slow connotation.

Korean ideophones are always fully reduplicated (e.g., [mindu mindu ] ‘bald’). For the sake of perspicuity, however, this reduplication will not be marked. Alternations involving high vowels do not apply after the initial syllable. (10) (a) (b) (c) (d) (e) (f)

Ideophones Alternations (Cho, 1994; Chung, 2000; Kim, 2000; Lee, 1998): Dark Light Gloss [mindu ] [m ndu ‘bald’ [t e ki ] [t kia ] ‘clanging’ [hinil] [hanil] ‘in an airy manner’ [k ubud ] [k obud a ] ‘bent’ [umull k] [omullak] ‘chewing’ [p sil] [pa sil] ‘smiling’

The alternations between LIGHT and DARK forms are strict in maintaining round/front/back features. Front vowels alternate with front vowels, central vowels alternate with central vowels and back vowels alternate with back vowels. This creates what appears to be a lowering process (McCarthy, 1983b; Sohn, 1987). However, the alternations involve changes in [ATR], height or both. For example, [i] alternates with [ ]—a change in both height and ATR. [u] alternates with [o]—a change in height only. [ ] (a [+ATR] vowel) alternates with [a]—a change in ATR only. Describing the harmonic feature as ATR (Chung, 2000; Kim, 2000) faces similar problems, as all round vowels in Korean are [+ATR]. In order to explain why [o] and [ø] are ‘‘light’’ one must force these vowels into having the feature [ ATR], which does not fit the facts of the language (Yang, 1992), and violates general markedness principles ([ HIGH, LOW] vowels are preferred to be [+ATR], and [ ATR, HIGH] vowels usually always have a [+ATR] counterpart in the inventory). Thus, the traditional feature system is unable to account for harmony in Korean with a single phonological feature. I assume that this reflects the unusual nature of Korean vowel harmony, and adopt both [ ATR] and [ HIGH] as the harmonic features (Finley, 2006). While some may argue against using multiple features in order to characterize the harmony of a system, Korean is not the only language that could be analyzed in this way. For example, van der Hulst and van de Weijer (1995) argue that some languages have mixed-harmony systems that behave similarly to the Korean system, as in Santrokofi and Manchu (Ard, 1983). Using multiple features to describe a single harmony system is a valid way to describe vowel harmony in general, and could possibly be extended to other languages. Further, the morphological nature of vowel harmony in Korean ideophones makes it possible to use multiple features without formal consequences. 6

[y] and [ø] are in parentheses because these vowels have been cited in dialects of Korean other than the Seoul dialect described here (Cho, 1994).

484

S. Finley / Lingua 119 (2009) 478–501

One might argue that the LIGHT morpheme, with multiple features, is actually two separate morphemes. However, there is no evidence to suggest this, as the features of the LIGHT morpheme always co-occur. The fact that a morpheme has multiple features is not unusual, as morphemes with multiple exponences are fairly common crosslinguistically (e.g., Lena Asturian which has segmental and featural material, among others (Kurisu, 2001)). Further, the LIGHT morpheme cannot be properly instantiated without both features realized on the lexical item. Because morphemes correspond to units of meaning, there is no way to separate out the high and the ATR features into separate units of meaning in this case. The formation of DARK vs. LIGHT forms is analyzed as a morphological process in which DARK and LIGHT are morphemes bearing floating phonological features ([+ATR] for DARK, [ ATR], [ HIGH] for LIGHT) that induce a correspondence with the vowels in the word indexed to these morphemes: the Korean grammar includes constraints (2)–(4). I assume that both LIGHT and DARK forms are marked with floating features, rather than with LIGHT forms derived from DARK forms, explaining the fact that not all ideophones undergo semantic alternations, as in (11), as well as the fact that DARK forms have a restricted [+ATR] inventory. (11) (a) (b) (c) (d)

Dark Form Only [m m ] [ulthu buldu ] Light Form Only [k amt ak ] [agit agi]

Gloss ‘barking’ ‘bumpy’ Gloss ‘astonishing’ ‘be full of interest’

The correspondence between the floating feature and the output domain is represented both graphically and with subscripts on the harmonic feature in (12) below. Correspondence with [ATR] is associated with the subscript 1, and correspondence with [HIGH] is associated with the subscript 2. (12)

Correspondence Relations in DARK and LIGHT Forms: (a) DARK [te1 g 1 ] [+ATR]1 (b) LIGHT: [ HIGH]2 [ta1,2 g a1,2 ] [ ATR]1

The input to every LIGHT morpheme contains the features [ ATR] and [ HIGH], which are mapped onto the output forms via the correspondence constraints given for Kanembu. The difference is that for Korean ideophones, there are two sets of constraints: one for [ATR] and one for [HIGH]. Additionally, these constraints are stratum-specific. That is, they refer only to ideophones (as standard lexical items do not have a ‘light’/‘dark’ dichotomy). Because this discussion will not venture outside the ideophone stratum, I will not refer to the specific ideophone stratum in the constraints or tableaux. The association of the LIGHT morpheme with [ HIGH] (as alternations from /u/ to [o] only changes in the high feature) only applies to the first syllable. The correspondence of the [ HIGH] feature on the first syllable is achieved when L-ANCHOR-[HIGH] is ranked above ID[HIGH] (with R-ANCHOR-[HIGH] and O-CONTIGUITY-[HIGH] low ranked). The [ ATR] feature achieves correspondence with all vowels in the word because all associated correspondence constraints are ranked above ID[ATR]. The fact that [ ATR] and [ HIGH] must be realized on different segments is straightforward under the present approach, as each feature carries its own set of constraints. If the LIGHT morpheme itself carried one set of constraints for all features associated with it, there would be no way to separate the realization of these features. This is precisely the requirement of the morphological alignment approach, in which it is the LIGHT morpheme realized on the word’s

S. Finley / Lingua 119 (2009) 478–501

485

edges, rather than the features of associated with the morpheme. Thus, a morpheme-alignment approach cannot account for this pattern, as it requires both features to be realized on the same segments. The tableau in (13) illustrates the different realizations of [ ATR] and [ HIGH] in the LIGHT form. Note that alternations will always result in a change in height and/or [ATR] due to high-ranked faithfulness to the features round and back (not shown).

Candidate (a) fails because [ HIGH] is not realized at the left edge, violating L-ANCHOR-[HIGH]. Candidate (b) has no high vowels, but because ID[HIGH] outranks R-ANCHOR-[HIGH] and O-CONTIG-[HIGH], the additional violation of ID[HIGH] is fatal. Candidate(c) fails because / / is not realized as [ ATR] [a], violating R-ANCHOR-[ATR]. Because high vowels only alternate in word-initial position, whenever [o] in LIGHT forms is derived from /u/, [o] appears word-initially. All DARK forms only contain vowels that are [+ATR]. The status of [o] as a ‘light’ vowel even though it is [+ATR] can be accounted for by the fact that mid-round vowels are marked in Korean. In DARK forms, instances of [o] are ruled out by the high ranked constraint *[+ROUND, HIGH] (Kaun, 1995, 2004),7 as in (14) below.

The fact that *[+ROUND, HIGH] is high-ranked does not remove [o] altogether from the vowel inventory. Because *[+ROUND, HIGH] is ranked below L-ANCHOR-[HIGH] and a change to any other vowel in the inventory would involve a violation of ID[BACK] and ID[ROUND], an input /o/ will always surface as [o] in a LIGHT form. Changing /o/ to [ ] violates inventory constraints. This is illustrated in the tableau in (15).

7

*[+ROUND,

HIGH] vowels is ranked below ID in the general stratum.

486

S. Finley / Lingua 119 (2009) 478–501

The diagram in (16) illustrates the ranking for Korean. Identity to round and back features is highest ranked, followed by L-ANCHOR, ranked above *[+ROUND, HIGH]. ID[HIGH] is ranked above L and R-ANCHOR[ATR], which are all ranked above ID[ATR].

Morphemic vowel harmony in Korean ideophones is a case in which multiple features and edge-only effects are straightforwardly accounted for with the correspondence-approach presented in this paper. While these effects have posed challenges for more traditional approaches to vowel harmony, they are in fact predicted by the proposal given here. 3.2. Gapped representations: Lena Asturian The correspondence approach to harmony predicts that morphemic harmony may skip vowels in order to find its optimal realization. In Lena Asturian (vowel inventory presented in (17) below), raising of a stressed vowel is triggered by the masculine singular morpheme and count morphology, but is allowed to skip unstressed vowels in order to do so, as in (18)(d).

(18)

Lena (a) (b) (c) (d) (e)

Asturian Vowel Harmony (Hualde, 1989; Walker, 2004) kordı´r-u ‘lamb’ (m sg) korde´r-os (m pl) reu´nd-u ‘round’ (m sg count) reo´nd-o (mass) ge´t-u ‘cat’ (m sg) ga´t-os (m pl) burwı´ban-u ‘wild strawberry’ (m sg) burwe´ban-os (m pl) pe´ ar-u ‘bird’ (m sg) pa´ ara (m pl)

The harmonic feature associated with the masculine morphology cannot simply be [+HIGH] because low vowels raise to mid (rather than all the way to high), but as mid vowels raise to [+HIGH], the feature cannot merely be [ LOW]. Because there are traits of both high and low in the raising process, I assume that the morphological feature is both [+HIGH] and [ LOW], and that the morpheme for the masculine singular is the suffix [-u] plus the features [+HIGH] and [ LOW]. Because the segmental material already bears the features [+HIGH] and [ LOW], there need not be an additional floating feature associated with the morpheme, as the features of the segment may be realized on the stressed vowel. However, if we assume a floating feature associated with the morpheme, the segmental material can be underlyingly any value of [HIGH]. Both analyses yield the same result.

S. Finley / Lingua 119 (2009) 478–501

487

The fact that low vowels cannot raise to [+HIGH] (by violating both ID[HIGH] and ID[LOW]) is accounted for via the local conjunction of ID[HIGH] and ID[LOW], as in Kirchner’s (1996) analysis chain shifts (as well as Walker’s (2005) analysis of Lena). I assume that all prominent positions (edges of words, stressed vowels) may be subject to morphemic anchoring, ´ -ANCHOR to the set of featural correspondence constraints: making it possible to add the constraint V (19)

´ -[ F]-MORPH8: The [ F] feature of MORPH in the input must be in V correspondence with all stressed vowels in output of the lexical domain (which therefore must be [ F]).

´ -ANCHOR[HIGH] is violated any time the [+HIGH] feature is not in correspondence with the stressed vowel. There are as V ´ -ANCHOR as there are stressed vowels in the word. This differs from R/L-ANCHOR, in which there many violations of V can be as many violations as vowels in the word because an unstressed vowel adjacent to the stressed vowel does not make that vowel any more stressed than an unstressed not adjacent to the stressed vowel, but a vowel one vowel away from the right edge is closer to the right edge than a vowel three vowels away from the right edge.9 I assume separate ANCHOR constraints for each feature (e.g., one for [HIGH] and one for [LOW]). However, to save space, and because each feature has the same constraint ranking, I collapse the two constraints. Candidate (a) in (20) violates R-ANCHOR because the harmonic constraint is not realized on the right edge. Candidates (c) and (d) violate ´ -ANCHOR because the stressed vowel does not raise. Candidate (d) realizes the harmonic feature on the left edge, but V not the stressed vowel, incurring a violation of O-CONT.

Candidate (20)(b) satisfies both ANCHOR constraints with the fewest ID violations. Candidate (d) realizes the harmonic ´ -ANCHOR. Candidate (e) fails feature on the left edge, satisfying L-ANCHOR, but not the stressed vowel, violating V because L-ANCHOR is ranked below ID[HIGH], the leftmost vowel will only be raised if it is stressed. Tableau (21) below illustrates raising of a low vowel to a mid vowel.

8 9

It is also possible to analyze the Lena facts as a conjunction of ANCHOR and *STRUCTURE ( ´ -ANCHOR would produce the same results. Note that gradient assignment of violations of V

).

488

S. Finley / Lingua 119 (2009) 478–501

Candidate (21)(b) satisfies R-ANCHOR with the fewest ID violations. Because ID outranks O-CONTIGUITY, the more harmonic candidate (a) cannot win. Because low vowels can only raise to mid, candidate (e) cannot win. A low vowel ´ -ANCHOR because raising only to mid is better than not raising at all. Candidates (c) and (d) have two violations of V neither [+HIGH] or [ LOW] are realized on the stressed vowel. ´ -ANCHOR outranks L-ANCHOR, when the initial vowel is already [+HIGH], as in [burwe´banos] ‘wild Because V strawberry’ the morpheme is still realized on the stressed vowel, illustrated in (22) below.

While I have assumed a morphemic analysis of Lena, Hualde (1989) argues against a morphological analysis. First, masculine singular is generally considered the default morpheme. Treating harmony as a morphological process requires the masculine singular morpheme to be marked, but there are reasons to assume that count nouns should be basic as mass nouns are more complex than count nouns. Second, only masculine singulars that end in a high vowel trigger harmony (e.g., [fe´re] ‘type of hawk’ (m sg) is the same as its plural counterpart [fe´res]). While these arguments have merit, none of them rule out a morphemic analysis of Lena. First, harmony in Lena Asturian fits the profile of morphemic harmony: it is triggered by the presence of a specific morpheme, and bears a specific harmonic feature value. Even if count and masculine singular morphology is more basic than mass or plural morphology, one must still account for the fact that vowel harmony only applies in the presence of these morphemes. Any economy of analysis gained by treating count and masculine singular morphology as basic is lost in accounting for why harmony does not apply elsewhere. Second, it is possible to account for the fact that only nouns ending in high vowels trigger vowel harmony if count nouns bear a floating [+HIGH] feature that is realized on the final vowel and the stressed vowel via morpheme-specific ANCHOR constraints. If it is acknowledged that allomorphy exists within the realization of the masculine singular morpheme, and that only some forms are marked with the floating, morphemic [+HIGH] feature. Those that are not marked with the floating feature will not undergo morphemic harmony. While the proposal presented here suggests that morphemic harmony is realized via floating features, it does not exclude the possibility of allomorphy in which some forms realize the morpheme without floating features. It is also possible that these exceptions have high-ranked (lexically-specific) faithfulness to high features, making it impossible for them to undergo harmony. These exceptions are not phonologically conditioned, as they apply to stems ending in non-high vowels as well as high vowels (e.g., [aseme´ju] ‘resemblance’ (pl. [aseme´jos]) (Neira Martinez, 1955)). Third, if metaphony in Lena Asturian were governed purely phonologically, one should expect that any post-tonic high vowel should trigger raising. In (23)(a), the stressed vowel does not raise in the feminine and masculine-singular forms, despite the fact that there is a high vowel ([i]) following the stressed vowel. Further, even when the final vowel is high, stressed low vowels may occur, as illustrated below in (23)(b). (23)

Exceptions to metaphony (Campos Astorkiza, 2007; Walker, 2005) (a) masc. sg. fem. sg. masc. pl. gloss siliku´tiku siliko´tika siliko´tikos ‘suffering from silicosis’ (b) Adj. gloss aba´xu ‘down’

There is no doubt that there is a phonological component to the harmony in Lena Asturian, which can be captured in a morphemic analysis of Lena, as the constraints governing harmony are rooted in phonological principles. Further,

S. Finley / Lingua 119 (2009) 478–501

489

purely phonological approaches require additional representational assumption in order to account for gapped representations (such as those proposed in Walker (2004)). Lena Asturian is one of several cases gapped representations in featural affixation. In Mafa (Chadic) (Ettlinger, 2008), the imperfective is realized via leftward spreading of the [+PAL] feature (palatalization) onto all vowels and consonants that allow realization of [+PAL]. The inventory of palatal consonants in Mafa includes [nd , , , t , d , j] and front vowels (which are assumed to be [+PAL]) in Mafa include [i, y, e, ø]. Only segments with a [+PAL] counterpart undergo palatalization. (24)

Palatalization in Mafa (Ettlinger, 2008) (a) dad- ‘add water to’ ded(b) tsap- ‘spackle’ t ep (c) guts- ‘squirt’ gut (d) suwd k ‘miss’ uwdik

‘is ‘is ‘is ‘is

adding water to’ spackling w/ clay’ squirting’ missing’

In (24)(d) palatalization applies to the final vowel, and initial consonant but skips over the medial consonants, necessitating the use of a constraint against gapped configurations (such as O-CONTIGUITY) in CON rather than GEN. This example supports the predictions of the present account that gapped configurations are a part of featural affixation. The present proposal predicts the possibility that certain featural affixes may be realized at the left and right edges of a lexical item, and with no realization in between, no matter how many segments intervene. For example, the feature [+PAL] might be realized on the first vowel and the last vowel (e.g., / – – – / [+ – – – +]). This type of morphological process would simultaneously realize a feature as both a prefix and suffix as featural circumfixation. Because featural affixes are relatively rare, and circumfixes are also quite rare, it is not surprising that the inventory of featural affixes presented in this paper does not include a featural circumfix. There may also be a universally high-ranked constraint against circumfixes that contributes to this gap in the typology. One case of featural circumfixes is found in Lokaa, a Benue-Congo language spoken in Nigeria, the future tense is marked with a low tone on the final syllable and a prefix containing a low tone [n`-] (Iwara et al., 2003). This circumfix allows gaps between the initial low tone and the final low tone. The forms in (25) below all have high tones in the stem that do not undergo tonal spreading.

While there is no locality restriction in the tonal circumfix of Lokaa, all stems are bi-syllabic, making it impossible to find cases of gapped representations greater than a single vowel, even if they were possible. 3.3. Morpheme-specific consonant inventories: Pasiego Montan˜es The correspondence-based analysis of morphemic harmony predicts that a language can have a morphemespecific vowel inventory. While there are languages in which segments can only appear if they are derived from harmony, phonological harmony does not restrict these segments to particular morphemes. The Pasiego dialect of Montan˜es (spoken in the province of Santander in north-central Spain) exhibits a morphologically controlled [ATR] harmony in which the masculine singular suffix ([- ]) forces all vowels in the lexical item to be [–ATR]. The interesting point is that ‘‘lax vowels may not occur in a word without this suffix’’ (McCarthy, 1984:294).

The exceptionless morphemic harmony found in Pasiego is illustrated in (27) (note that [e] is transparent to harmony, as expected by the [ ATR] vowel inventory).

490

(27)

S. Finley / Lingua 119 (2009) 478–501

[ATR] Harmony in Pasiego Dialect of Montan˜es (McCarthy, 1984)

The analysis of Pasiego [ATR] harmony is roughly the same as the analysis of Korean, Lena and Kanembu above. The major difference is that in order to account for the transparency of [e], a constraint against [ ] must be ranked higher than the correspondence constraints, and O-CONTIGUITY is crucially ranked below the two ANCHOR constraints. This analysis reflects the interaction of inventory constraints with morphemic harmony. Inventory constraints ranked between the harmony constraints and the faithfulness constraint induce a morpheme-specific vowel inventory.

Tableau (29) illustrates that the harmony constraints permit [+ATR] [e] between two [ ATR] vowels.

If a potential undergoer is flanked by two transparent vowels, [e] will still be transparent because more violations of O-CONTIGUITY are incurred if [e] undergoes harmony. This is illustrated in (30). The medial vowel undergoes harmony in candidate (a) but not in candidate (b). Both candidates violate O-CONTIGUITY, by each vowel intervening between the right and left edge. Candidate (b) has three intervening non-undergoers, while candidate (a) only has two.

S. Finley / Lingua 119 (2009) 478–501

491

Pasiego also has a height harmony system in which a final high vowel induces raising of the stressed vowel (Hualde, 1989). This harmony only affects the [ATR] system presented here in that underlying stressed /e´/ surfaces as [ ATR] /I/ (e.g., [k rdIr ] ‘lamb’ (Hualde, 1989)). The raising of the stressed mid vowel allows it to participate in ATR harmony. As there are no other interactions with ATR harmony, the height harmony will not be discussed further. Pasiego is an example of a morphemic harmony system that creates vowels that only appear in the presence of a particular morphemic, another prediction borne our by the correspondence-approach to morphemic vowel harmony. 4. General predictions While the discussion in this paper has focused on morphemic vowel harmony, the proposal presented here easily accounts for cases of featural affixation involving consonants. This section illustrates this versatility through analyses of Chaha labialization and Terena nasalization. 4.1. Chaha labialization In Chaha, a Gurage language of Ethiopia, the third masculine singular object is realized via labialization of the rightmost labializable (non-coronal) consonant in the base, followed by the suffix [-n] (McCarthy, 1983a; Rose, 1994; Zoll, 1996). If the rightmost consonant is coronal, the labial feature will be realized on the closest labializable consonant. If no labializable consonants are present, the feature is not realized. (31)

Chaha Labialization (Akinlabi, 1996; McCarthy, 1983a; Piggot, 2000; Rose, 1997; Wolf, 2007) (a) Rightmost consonant of the stem is labializable (i) [da¨na¨g] [da¨na¨gwn] ‘hit’ (ii) [na¨da¨f] [na¨da¨fwn] ‘sting’ (b) Medial consonant of the stem is labializable; final is not (i) [na¨ka¨s] [na¨kwa¨sn] ‘bite’ (ii) [ka¨fa¨t] [ka¨fwa¨tn] ‘open’ (c) Only leftmost consonant is labializable (i) [qa¨ta¨r] [qwa¨ta¨rn] ‘kill’ (ii) [ma¨sa¨r] [mwa¨sa¨rn] ‘seem’ (d) No labializable consonant (i) [sa¨da¨d] [sa¨da¨dn] ‘chase’

The lack of labialized consonants is handled by the constraint against labialized coronals: *COR/LAB (Akinlabi, 1996). (32)

*COR/LAB: If [+CORONAL] then not [+ROUND]

This constraint is ranked above R-ANCHOR, which forces the featural affix to be docked on the right edge. (33)

R-ANCHOR-[+ROUND]-3MG-SG: The [+ROUND] feature of 3MG-SG must be in correspondence with the rightmost consonant of the output domain of the lexical item (which therefore must be [+ROUND]).

These constraints are ranked above ID[ROUND], and give rise to labialization given in (31). Note that because L-ANCHOR is ranked below ID, labialization will only be realized on one consonant if there are multiple labializable consonants in the input.

492

S. Finley / Lingua 119 (2009) 478–501

In the tableaux in (34) and (35), the [+ROUND] feature is realized on the rightmost labializable consonant. Any labialized coronal consonant violates *COR/LAB. As the realization moves from right to left, the number of R-ANCHOR violations increases, enforcing realization on the rightmost consonant.

The above analysis illustrates a clear and straightforward account of featural affixation involving consonants using the present approach to morphemic harmony. 4.2. Terena nasalization In Terena, an Arakwakan language of Brazil, the first person pronoun is marked through long-distance nasalization. Nasalization applies in a left-to-right fashion, and affects both consonants and vowels, but is blocked by obstruents, which become pre-nasalized (Akinlabi, 1996; Eastlack, 1968; Piggot, 2000; Rose, 1997; Wolf, 2007).

The basic realization of Terena is accounted for with the ranking L-ANCHOR O-CONTIGUITY R-ANCHOR. A constraint against nasal obstruents blocks spreading through obstruents. Following Steriade (1993)10 I assume that prenasalized obstruents are represented in terms on an initial closure associated with [+NASAL] followed by a release, associated with [ NASAL], as illustrated in (37).

10

It is also possible to analyze prenasalization as a homorganic sequence of root nodes, as in Padgett (1995) with the same result.

S. Finley / Lingua 119 (2009) 478–501

493

Because the primary specification for prenazalized stops is [ NASAL], these stops do not violate *NASAL-OBST. However, secondary [+NASAL] representation allows prenasalized stops to satisfy ANCHOR. Because prenasalized stops contain a [+NASAL] representation followed by a [ NASAL] representation, a gapped representation may result if a nasalized segment follows the prenasalized stop, incurring a violation of O-CONTIGUITY, as illustrated in candidate (b) in the tableau in (38) below.

The nasal feature spreads from the left to the right, but stops at an obstruent. Spreading through the obstruent incurs a violation of O-CONTIGUITY (candidate (b)). Partial spreading incurs violations of R-ANCHOR. Candidate (c) surfaces because it has the fewest R-ANCHOR violations without violating O-CONTIGUITY or *NASAL-OBST. The extension of the present proposal to Chaha and Terena illustrates the ability of the present analysis to account for general cases of featural affixation. The next sections discuss alternative proposals, beginning with possibilities for integrating morphemic and phonological vowel harmony. 5. Phonological harmony as morpheme-correspondence Previous accounts of vowel harmony have assumed morphemic and phonological harmony to be indistinguishable. For example, Akinlabi’s (1994) analysis of ATR harmony uses the morphemic harmony language Kanembu, as a basis for a more general account of phonological harmony. However, there are three main arguments supporting a distinction between morphemic and phonological harmony. First, analyzing phonological harmony in the same manner as morphemic harmony would resemble the root-marker theory of harmony (Lightner, 1965), requiring all harmony-inducing morphemes to bear a floating underlying feature. Unfortunately, this theory suffers from duplication resulting from phonological features and a morphological diacritic (Kenstowicz, 1994; Kiparsky, 1973). In phonological harmony, phonological features would do the same work as a morphological diacritic, but in morphemic harmony, the additional floating feature serves as a morpheme as well as a trigger of harmony. Second, analyzing phonological harmony as featural correspondence reduces phonological harmony to a morpheme-specific process, losing the insight of the connection of phonetics and markedness to vowel harmony. This misses the point that agreement of vowel features is an unmarked state, stemming from phonetic and acoustic factors such as co-articulation and perception of contrasts (NiChisosain and Padgett, 2001). A theory in which harmony is induced by a floating feature associated with a morpheme predicts that languages should be much more idiosyncratic