Download Onset Transfer in Reduplication: A Comparative Study of Klamath, Gothic, and Attic Greek and more Papers German Philology in PDF only on Docsity!
Onset Transfer in Reduplication^1 Heidi Fleischhacker ([email protected])
- I NTRODUCTION
A frequently noted property of reduplication is that the reduplicant is generally a contiguous substring of the base (e.g. Marantz 1982; McCarthy and Prince 1986, 1995; Lamontagne 1996). For example, Lamontagne (1996) identifies [[ABC]R [ABCDE…] (^) B ] as a 'typical reduplication pattern', and [[ACD]R [ABCDE…]B ] as 'atypical'. In optimality theory, this property is enforced by the "no skipping" clause of the correspondence constraint CONTIGUITY (Kenstowicz 1994; McCarthy and Prince 1995), which prohibits base-reduplicant mappings of the type ABC → AC.
The focus of this paper is on ABC → AC mappings in reduplicative onset transfer: specifically, cases in which a [C 1 C 2 V 3 …] base corresponds to a prefixed [C 1 V 3 ] reduplicant, as in the Klamath distributive form [[t’ 1 a 3 ]R [t’ 1 w 2 a 3 ja]B ] 'dist. work for' (Barker 1964). I suggest that such mappings are actually characteristic in one case: when C 1 is an obstruent (O) and C 2 is a sonorant consonant (R). The typology of onset cluster simplification under reduplication, presented in §2, shows that O 1 R 2 V 3 → O 1 V 3 reduplication occurs even when other base clusters do not simplify under reduplication (§2.1), and even when other base clusters do not reduplicate at all (§2.2). I argue that O 1 R 2 V 3 → O 1 V 3 mappings have a privileged status in reduplication because the perceptual difference between O 1 R 2 V and O 1 V is smaller than the difference between C 1 C 2 V and C 1 V in the general case, and smaller than the difference between C 1 C 2 V and C 2 V. Several lines of evidence in support of the perceptual difference theory are presented in §3. An analysis is proposed (§4) in which correspondence constraints sensitive to the magnitude of perceptual differences between correspondent strings, interacting with phonotactic constraints and a violable constraint demanding reduplication, determine whether a particular cluster will reduplicate, and if so, exactly what portion of the cluster will be copied.
- PARTIAL ONSET TRANSFER PATTERNS
This section presents a survey of partial onset transfer, i.e. reduplication patterns in which, for at least one type of base-initial biconsonantal cluster, only one cluster member is copied. (Note that in every pattern presented, CV-initial bases take CV- reduplicants.) The data discussed extend slightly the typology of onset transfer presented by Steriade (1988).
In all of the data discussed below, the reduplicant is a prefixed CV or CCV syllable. Onsets of the base and reduplicant are underlined; a dash separates reduplicant from base. I make a distinction between obstruent + sonorant onsets (OR) and all other onset clusters (¬OR); and further divide OR into stop + sonorant onsets (TR), and sibilant fricative + sonorant onsets (SR). The transfer patterns are grouped into 3 classes, which I will call sufficient copy , selective copy , and blind criterion ; the meanings of these class labels are spelled out below.
(^1) This paper presents work in progress. For helpful comments, thanks especially to Bruce Hayes, Colin Wilson,
and Donca Steriade.
2.1 Sufficient copy
Under sufficient copy reduplication, all complex onsets other than OR are copied in full. Only OR is simplified, and always by failure to copy the sonorant.
In Gothic (Braune 1883; Wright 1910; Steriade 1988), a reduplicating C(C)V- prefix (with fixed vowel e ) marks the perfect for a subset of the strong verbs:
(1) Gothic (data from Braune 1883)
CV a. [he-het] 'called' TR b. [ge-grot] 'wept', [fe-fres] 'tried, tempted' SR c. [se-slep] 'slept' ¬OR d. [ste-stald] 'possessed', [ske-sked] 'separated'
All OR clusters are simplified under reduplication, with copy of the obstruent only (b, c). In contrast, /sp, st, sk/—the only ¬OR onset clusters of Gothic—are copied in full (d). This is "sufficient copy", in the sense that only as much of the base cluster is copied as is necessary to achieve the requisite degree of perceptual similarity between base and reduplicant: if less than full copy will satisfy the perceptual similarity requirement, then full copy is not necessary.
In Klamath (Barker 1964; Steriade 1988), a reduplicating C(C)V- prefix marks distributive action in verbs:
(2) Klamath (data from Barker 1964)
CV a. [so-sota] 'dist. light a fire'
TR b. [t’a-t’waja] 'dist. work for', [go-gmta] 2 'dist. get old'
TR c. [q’ja-q’japga] 'dist. lie on their sides', [p’na-p’nandila] 'dist. bury underneath' TR d. [qni-qnj’a] ~ [qi-qnj’a] 'dist. have an erection'
SR e. [sl’o-sl’q’a] 'dist. shed hair', [sn’o-snis] 'policeman'
¬OR f. [sti-stiq’a] 'dist. have a cramp', [pse-psejisap] 'dist. uncles, father's brothers',
[lwo-lwasga] 3 'dist. take off clothes', [wqe-wqew’a] 'dist. break plural objects in two with long instruments'
Stop + sonorant (TR) clusters are simplified in some reduplicated forms, with copy of the stop only (b). In other reduplicated forms, TR is copied fully (c), and there is at least one case of free variation between full copy and simplification of TR (d). All clusters other than TR, including
(^2) The base of [go-gmta] is /gmota/; for some C1(C2)V (^) 1C3V 2 stems, the first stem vowel deletes in reduplicated
forms (Barker 1964:84). This process also applies to the forms in (2)d,e (their bases, in order: /qnij’a/, /sl’oq’a/, /sn’ojs/). (^3) The base of [lwo-lwasga] is /lwosga/; the change in stem vowel quality is accounted for by Barker (1964:89)
as the result of a rule mapping /V (^) RED + CGV 1 CC/ → [V 1 + CGaCC], where G = /w, j/ — i.e. the base vowel is overwritten by [a], but its quality survives in the reduplicant.
(4) Sanskrit (data from Steriade 1982)
CV a. [tu-tud] 'pushed', [ru-rudh] 'obstructed' TR b. [pa-prac h] 'asked', [du-druv] 'ran'
SR c. [si-mi] 'smiled', [a-rat h^ ] 'slackened'
¬OR d. [tu-stu] 'praised', [pa-psa] 'devoured'
¬OR e. [ma-mna-u] 'noted'
Base clusters are simplified by copy of the less sonorous cluster member only (b, c, d). If there is no sonority difference between the two members of the cluster, as with nasal + nasal clusters (e), the leftmost segment is copied—although as Steriade (1982) notes, this form is prescribed by Sanskrit grammarians but not actually attested.
This pattern seems to be shown by ancient Greek nominal reduplication as well (Steriade 1988): [ka-skandiks] 'wild chervil', [ko-skulmat-ia] 'leather cuttings'; however, as there are no attested reduplicated forms for OR-initial bases, this cannot be stated conclusively.
In Old Irish (Thurneysen 1961; Kuryłowicz 1971), a reduplicating CV- prefix (with fixed vowel e ) marks the perfect (and future, not shown here). In the data below, reduplicated perfect forms are followed by unaffixed present forms, presented for clarification of the pattern with respect to SR clusters:
(5) Old Irish (data from Thurneysen 1961)
CV a. [me-mad-] 'broke' TR b. [be-brag-] 'farted' cf. [braigid] 'farts, bleats', [ge-glann] 'learned' cf. [gleinn] 'learns' SR c. [se-laig] 'felled' cf. [sligid] 'fells', [se-naig] 'dripped' cf. [snigid] 'drips' ¬OR d. [se-scann-] 'flew off' cf. Modern Irish [sceinnim] 'I spring off, fly off'
Base clusters are simplified by copy of the leftmost cluster member (b, c, d), although this becomes clear in the case of SR (c) only on inspection of morphologically related forms. Thurneysen (1961:132) notes that "after reduplication syllables –sn– , –sl– gave single n , l "; generally, underlying intervocalic sm , sn , sl are realized as mm , nn , ll respectively.
Ancient Greek present reduplication (Steriade 1982) is also characterized by copy of the leftmost base consonant: e.g. [ki-kh^ rmi] 'borrow', [i-nsk] 'know'.
Finally, note that a logically possible type of blind criterion pattern is apparently unattested (Steriade (1988) did not find such a case, and neither have I): one in which just the rightmost member of a base cluster is copied, as in the hypothetical system re-pre , li-sli , ta-sta. Such a pattern would parallel Old Irish, differing only in that C 2 , not C 1 , is the sole member of a base C 1 C 2 cluster that survives simplification.
2.4 Data not addressed by this analysis
Before concluding this section, a reduplicative pattern that will not be addressed further in this paper should be noted. This pattern gives the appearance of infixing C 2 (V) after an initial C 1 C 2 V string, for at least some type of C 1 C 2 cluster. I am aware of three such patterns: Pima (Riggle 2001; Marcus Smith, p.c.), Latin (Steriade 1988), and Old High German (Jasanoff 2001; Helfenstein 1870). It is because each of these three cases is evidenced by extremely sparse data, and because two of the three (Latin and Old High German) are open to alternative interpretations, that this data pattern is not given further attention.
In Pima plural and distributive reduplication (Riggle 2001; Marcus Smith, p.c.), the reduplicant is an infixed CV or bare C: e.g. [go-go-gs] 'dogs', [ce-c-mait] 'cakes'. Complex onsets appear only in three words known to Riggle and Smith: [trogi] 'truck', [trampi] 'tramp',
and [skait] 'rich ones'; the first two are obviously borrowed. [trogi] reduplicates as either [tro- ro-gi] or [tro-r-gi]: that is, with an infixed C 2 (V) string. [trampi] can reduplicate as [tra-ra-mpi], also with infixed C 2 V; however, Riggle and Smith's consultant also uses unreduplicated [trampi]
in plural constructions. Finally, [skait] does not allow reduplication; Riggle and Smith's
consultant explicitly rejects [skai-kai-t].
Latin perfect reduplication (Steriade 1988) is attested only by the three following forms,
all with initial s + stop clusters: [ste-t-i], base [ste-]; [spo-po-nd-i], base [spond-]; and [sci-ci-d-
i], base [scid-]. It is of course not obvious what reduplicative behavior OR clusters would show in Latin; however, Helfenstein (1870: 409) notes that "forms arose such as [ cêpi ] from *[ ca- capi ], [ fêci ] from *[ fa-faci ], [ frêgi ] from * [ fra-fragi ] or rather * [ fra-fagi ]" [emphasis added]. This statement seems to suggest that the Latin reduplicant is not an infixed C 2 V string, regardless of the content of the base cluster; but rather that Latin is essentially the infixing counterpart of Sanskrit, with copy of only the less sonorous member of the base cluster: the stop of s + stop clusters, but the obstruent of OR clusters.
Old High German (Jasanoff 2001; Helfenstein 1870) contains several relics of proto- Germanic perfect reduplication: [steraz] 'pushed', from *[ste-zaut], and [pleruz] 'sacrificed', from *[ble-lōt]. These forms give the appearance of infixing copy of the more sonorous member of the base cluster; however, this appearance may be misleading. Jasanoff and Helfenstein both
argue that *[ste-zaut] and *[ble-lōt] are derived from *[ste-staut] and *[be-blōt], respectively, and Jasanoff proposes that these forms reflect a strategy which "concentrate[s] lexically relevant
information in the reduplication syllable (e.g., *b…bl- > *bl…l-), while allowing all but the coda of the root syllable to become opaque through sound change, sporadic dissimilation, and irregular shortening."
Finally, note that there are of course many other reduplicative patterns which fall outside the scope of the present analysis. For example, in Ilokano (Hayes and Abad 1989), bases with an initial consonant + glide sequence often allow two reduplicated outputs: one with full copy of the base-initial cluster, and one in which the vowel of the reduplicant corresponds to the base glide (e.g. [bwája] 'crocodile' reduplicates as either [na-ka-bwaj-bwája] or [na-ka-bu-bwája]). In Nuxalk (Bella Coola) (Carlson 1997), the reduplicant is located before the first vowel or
demand for reduplication is sacrificed in order to satisfy the demand of sufficient similarity between base and reduplicant—if there is no way for a particular cluster to simplify and still be similar enough to its base, reduplication fails. Again, O 1 R 2 V maps to O 1 V, but no other clusters participate in reduplication.
The analysis presented in §4 formalizes these interpretations of blind criterion copy, sufficient copy, and selective copy as the interactions of correspondence, markedness, and reduplicative constraints.
- PERCEPTUAL SIMILARITY
To support the analysis of partial onset transfer developed below, the relationships of perceptual similarity asserted in (6) must be proved:
(6) ∆(C 1 C 2 V–C 1 V), ∆(C 1 C 2 V–C 2 V) > ∆(O 1 R 2 V–O 1 V) where ∆(X – Y) = perceived difference between X and Y, and (C 1 C 2 V–C 1 V) ≠ (O 1 R 2 V–O 1 V)
That is, it must be shown that the perceptual difference between O 1 R 2 V–O 1 V is smaller than that between C 1 C 2 V–C 1 V in the general case, and smaller than that between C 1 C 2 V–C 2 V.
Below I offer evidence in support of (6) for the case when C 1 C 2 ≠ O 1 R 2 is the set of sibilant fricative + stop (ST) clusters, providing the following result:
(7) {∆(S 1 T 2 V–S 1 V), ∆(O 1 R 2 V–R 2 V), ∆(S 1 T 2 V–T 2 V)} > ∆(O 1 R 2 V–O 1 V)
The evidence comes from alliterative verse and imperfect puns, both phenomena in which the rules governing the situation (the constraints of the verse system, or the principles of humor) require the language user to establish an imperfect correspondence relationship—i.e. correspondence between strings that are similar but non-identical. The evidence thus rests on the assumption that examination of which strings are placed in such intentionally imperfect correspondence relationships, and with what frequency, can provide insight into language users' judgments of relative similarity.
3.1 English imperfect puns
In an imperfect pun, such as Napoleon Blown-apart , the pun word (here, Blown-apart ) corresponds to a phonologically similar but non-identical target word (here, Bonaparte ).^8 I follow Zwicky & Zwicky (1986) in treating imperfect puns as a source of evidence bearing on phonological similarity.
I make the following assumptions about the nature of imperfect puns. First, because the target word is usually not made explicit in the pun's context, the pun word must be sufficiently similar to the target that the target can be inferred—this is what makes the difference between an
(^8) Note that perfect puns, in which the pun and target are phonologically identical but lexically distinct, are of
course not relevant here.
amusing pun and one that is just puzzling. Further, there is a positive correlation between pun- target similarity and the goodness of the pun: although puns may be bad for a variety of reasons (objectionable subject matter, artificial context, winking delivery, etc.), truly funny puns are generally those in which the phonological relationship between pun and target is unforced, subtle but quickly recognizable on examination. Finally, I assume that most puns are good-faith attempts at humor, and that the goodness of a particular pun category can be roughly quantified by calculating its degree of representation in a large corpus of imperfect puns.
An imperfect pun corpus was constructed as follows. I obtained a set of 1,924 puns collected by Arnold and Elizabeth Zwicky, which had been preserved in the Arnold M. Zwicky papers at the Western Historical Manuscript Collection.^9 These puns appear to be exclusively from Crosbie (1977), a book of pun jokes organized in dictionary format.^10 605 of these 1, were eliminated on the grounds that they were perfect puns, stress puns, clever definitions as opposed to puns, etc.; I then added 645 imperfect puns that I collected from a variety of magazines, newspapers, novels, radio, television, and advertising materials, including two books of slogans (Sharp 1984; Urdang and Robbins 1984). Thus the corpus contains a total of 1, imperfect puns, coded as illustrated by the examples below:
(8) Corpus coding: examples
pun word target word pun segment target segment context pun type context Blown-apart Bonaparte l Ø b_o Ø~R/O_V medial surgeon sturgeon Ø t s_ Ø~T/S_V medial raise praise Ø p #_r Ø~O/#_R initial Stabitha Tabitha s Ø #_t Ø~S/#_T initial
The examples in (8) show the four pun types of interest for the analysis here: Ø~R/O_V ( Blown- apart–Bonaparte ), which bears on the similarity of O 1 R 2 V to O 1 V; Ø~T/S_V ( surgeon– sturgeon ), which bears on S 1 T 2 V–S 1 V; Ø~O/#_R ( raise–praise ), which bears on O 1 R 2 V–R 2 V; and Ø~S/#_T ( Tabitha–Stabitha ), which bears on S 1 T 2 V–T 2 V.
Following Frisch, Broe, & Pierrehumbert (1997), degree of representation in the pun corpus was determined by calculating the ratio of the frequency of a pun type in the corpus to the frequency that would be expected if pun types occurred at random.
Observed frequency in the pun corpus was calculated by dividing the number of instances of a particular pun type by the total number of instances of its general type. For example, the frequency of O 1 R 2 V–O 1 V puns was calculated by dividing the number of these puns by the total number of puns characterized by word-medial insertion or deletion of any segment:
(^9) WHMC, 23 Ellis Library, University of Missouri-Columbia, Columbia, MO 65201. Many thanks to Arnold
and Elizabeth Zwicky for allowing me to use their materials, and to Arnold Zwicky and WHMC reference librarian John Konzal for help in locating them. (^10) The puns' origins are not indicated in the materials I have, but in fairly extensive checking I have never failed
to find a pun in my copy of Crosbie (1977).
The last column in the table is the observed frequency divided by the expected frequency (O/E) for each pun type. These O/E values are used to establish overrepresentation and underrepresentation in the pun corpus, as follows. When O/E = 1, the proportion of a pun type in the corpus is equivalent to the proportion of relevant word-pairs in English: thus, the pun type occurs with the frequency that would be expected if pun-target pairs were selected randomly from among English word pairs. When O/E is greater than 1, the pun type is overrepresented in the corpus, with respect to the set of English word pairs of the relevant type—i.e. there are more of these puns in the corpus than expected, all else being equal. Finally, when O/E is less than 1, the pun type is underrepresented in the corpus—i.e. there are fewer of these puns in the corpus than would be otherwise expected. Note that zero is the lower limit on O/E values. The O/E values for each pun type are displayed graphically in (12):
(12) O/E values, by pun type
0
1
STV-SV ORV-RV STV-TV ORV-OV
O/E
Observe that the O/E values for O 1 R 2 V–R 2 V and S 1 T 2 V–T 2 V cluster around 1; this signals that these pun types are about as frequent as would be expected if pun choice is essentially random. In contrast, S 1 T 2 V–S 1 V is underrepresented (O/E = 0.33), and O 1 R 2 V–O 1 V is overrepresented (O/E = 1.34). If it is correct that degree of representation in the pun corpus correlates with pun goodness, and that more similar pun-target pairs are better than less similar ones, these results can be interpreted as supporting the following scale of relative similarity:
(13) ∆(S 1 T 2 V –S 1 V) > {∆(O 1 R 2 V –R 2 V), ∆(S 1 T 2 V –T 2 V)} > ∆(O 1 R 2 V –O 1 V)
That is, analysis of the imperfect pun corpus supports the relative similarity relationships asserted above in (7), if the assumptions spelled out above concerning the nature of imperfect puns are true.
3.2 Alliterative verse
The scale of relative similarity in (13) is bolstered by evidence from the alliterative verse systems of early Germanic and Irish, described below, assuming that alliterative constraints require words in certain metrical positions to begin with sounds that are sufficiently similar to signal an alliterative pairing. If this is correct, then the facts of what alliterates with what reflect judgments of relative similarity. That is, if a C 1 C 2 V–C 1 V or C 1 C 2 V–C 2 V pair does not alliterate, that pair must be less similar than a pair that does, since the alliterative standard of "similar enough" rules out alliteration in the first case but not the second.
In early Germanic verse, stressed syllables of half-lines alliterate as follows. Vowel- initial words alliterate with any vowel-initial word. Consonant-initial words alliterate with any word beginning with the same consonant, except that initial s + stop (ST) clusters alliterate only with themselves (Kuryłowicz 1971).^12 For example, pr V - alliterates with p V - , pl V - , and pr V - , whereas st V - alliterates only with st V - and not with any other s -initial form—i.e. not with s V-, sp V-, sk V-, s m V-, sn V-, or sl V-. Assuming that alliterative possibilities bear on perceptual similarity as described above, Germanic verse provides support for the similarity scale in (14):
(14) {∆(S 1 T 2 V –S 1 V), ∆(O 1 R 2 V –R 2 V), ∆(S 1 T 2 V –T 2 V)} > ∆(O 1 R 2 V –O 1 V)
That is, because only O 1 R 2 V –O 1 V pairs alliterate, these must be more similar than all non- alliterating cluster-singleton pairs.
The alliterative system of Early Irish is quite similar to that of early Germanic, with one primary difference: sm- acts in early Irish like sp- , st- , sk- , allowing only self-alliteration (Murphy 1961). This provides at least suggestive evidence that sm V – s V is less similar than any other O 1 R 2 V –O 1 V pair—i.e. less similar than any of sn V – s V, sl V – s V, sw V – s V, or T 1 R 2 V – T 1 V. This is relevant to the observation that, in Klamath and Ancient Greek, SR onsets pattern with non-OR onsets, whereas in Gothic, all OR onsets pattern together: if the perceptual difference between S 1 R 2 V-S 1 V is greater than that between T 1 R 2 V –T 1 V, then the variable behavior of SR, but not TR, with respect to reduplicative C 1 C 2 V → C 1 V mappings is not unexpected.
3.3 Summary, and a planned experiment
Taken together, the evidence from intentional imperfect correspondence provides support for the similarity scale in (15):
(15) {∆(S 1 T 2 V –S 1 V), ∆(O 1 R 2 V –R 2 V), ∆(S 1 T 2 V –T 2 V)} > ∆(O 1 R 2 V –O 1 V)
However, as noted above, this conclusion holds only if certain interpretive assumptions about imperfect puns and alliteration are true: namely, if representation in the pun corpus correlates positively with the degree of similarity between pun and target; and if cluster-singleton pairs that do enter into alliterative relationships are more similar than pairs that do not. In addition to being open to interpretation, the evidence from imperfect puns and alliterative verse is also incomplete for present purposes: because it comes from languages in which the only ¬OR clusters are /sp, st, sk/, it cannot bear on the relative similarity of C 1 C 2 V –C 1 V and C 1 C 2 V –C 2 V pairings for ¬OR clusters other than ST, e.g. stop + fricative, stop + stop, nasal + nasal, etc.—i.e. clusters which are essential to the analysis of onset transfer in Klamath and Greek.
A perceptual experiment currently in the stimuli preparation stage will address these issues.^13 In a discrimination task following the methodology of Tserdanelis (2001), native English speakers are asked to identify whether the members of a stimuli pair are the same or
(^12) Note that /sp, st, sk/ are the only non-OR onset clusters of Germanic. (^13) This work is in collaboration with Keith Johnson of the Ohio State University, to whom I am deeply grateful
for generous advice and practical support.
of perceptual difference between correspondent strings.^14 These constraints are the subject of the following section.
4.1 Enforcing perceptual similarity through correspondence constraints
As argued in §3, there is evidence in support of the following scale of relative similarity:
(16) {∆(S 1 T 2 V –S 1 V), ∆(O 1 R 2 V –R 2 V), ∆(S 1 T 2 V –T 2 V)} > ∆(O 1 R 2 V –O 1 V)
For the purposes of this analysis, several extensions to (16) must be accepted (with the promise of support or contradiction from the forthcoming experimental work described in §3.3). First, although the evidence given in §§3.1 and 3.2 above only covers OR and ST clusters, it must be assumed that the pattern is more general, such that O 1 R 2 V–O 1 V is more similar than any other C 1 C 2 V-C 1 V or C 1 C 2 V-C 2 V pair—that is, not just more similar than S 1 T 2 V –S 1 V and S 1 T 2 V –T 2 V. Second, it must be assumed that there is an internal division among OR clusters with respect to the C 1 C 2 V → C 1 V map: namely, T 1 R 2 V–T 1 V is more similar than S 1 R 2 V–S 1 V. As noted in §3.2, this assumption receives support from the alliterative system of Old Irish, in which sm - only self- alliterates. These extensions to (16) are summarized in (17):
(17) {∆(C 1 C 2 V–C 1 V), ∆(C 1 C 2 V–C 2 V)} > ∆(S 1 R 2 V–S 1 V) > ∆(T 1 R 2 V–T 1 V) where (C 1 C 2 V –C 1 V) ≠ (O 1 R 2 V–O 1 V), (S 1 R 2 V–S 1 V)
The similarity scale in (17) can be rewritten as follows, separating difference from context:
(18) { ∆(C 1 C 2 V–C 1 V) , ∆(C 1 C 2 V–C 2 V) } > ∆(S 1 R 2 V–S 1 V) > ∆(T 1 R 2 V–T 1 V) = = = = { ∆(C-Ø)/C_V ∆(C-Ø)/#_C } > ∆(R-Ø)/S_V > ∆(R-Ø)/T_V
That is to say, the difference between a sonorant and nothing in the T_V context is smaller than the difference between a sonorant and nothing in the S_V context, and so on. Adopting Steriade's (1999) P-map proposal, I assume that similarity scales such as the rewritten one in (18) project correspondence constraints and their rankings, as shown by the diagram in (19):
(19) Projection of correspondence constraints
similarity scale ∆(C-Ø)/C_V ,^ ∆(C-Ø)/#_C^ >^ ∆(R-Ø)/S_V^ >^ ∆(R-Ø)/T_V correspondence constraints M^ AXBR^ -C/C_V^ ,^ M^ AXBR^ -C/#_C^ »^ M^ AXBR^ -R/S_V^ »^ M^ AXBR^ -R/T_V
The context-sensitive MAX-C constraints shown in (19) penalize base-reduplicant correspondence relationships between consonants and zero—i.e., they penalize failure to copy base segments—in specific segmental contexts, with the penalty proportionate to the perceptual difference resulting from failure to copy. The table in (20) illustrates violation and satisfaction of each context-sensitive M AX-C constraint:
(^14) Note that CONTIGUITY also regulates splitting maps; I suggest that this portion of CONTIGUITY should be
replaced with context-sensitive D EP constraints, although this is not addressed further here.
(20) Violation patterns: context-sensitive M AX-C constraints
Outputs M AXBR -C_V M AXBR -C/#_C M AXBR -R/S_V M AXBR -R/T_V a. [[p 1 a 3 ]R [p 1 r 2 a 3 ]B ] * b. [[r 2 a 3 ]R [p 1 r 2 a 3 ]B ] * c. [[s 1 a 3 ]R [s 1 l 2 a 3 ]B ] * d. [[l 2 a 3 ]R [s 1 l 2 a 3 ]B ] * e. [[s 1 a 3 ]R [s 1 t 2 a 3 ]B ] * f. [[t 2 a 3 ]R [s 1 t 2 a 3 ]B ] *
M AXBR -R/T_V is violated only by the form in (a), which maps base T 1 R 2 V into reduplicant T 1 V. M AXBR -R/S_V is violated only by the form in (c), which maps S 1 R 2 V into S 1 V. M AXBR -C_V is violated by every form which maps general-case C 1 C 2 V (i.e. C 1 C 2 V other than S 1 R 2 V and T 1 R 2 V) into C 1 V; the only such form shown in (20) is (e). Finally, M AXBR -C/#_C is violated by any form that maps C 1 C 2 V onto C 2 V: here, these are (b), (d), and (f).
Note finally that the constraints in (19) represent only the fragment of the context- sensitive MAX family relevant to the analysis of partial onset transfer; definition of context- sensitive MAX constraints on other sound sequences, and their rankings, is left for further work.
4.2 Enforcing reduplication
Recall that in Attic Greek, base OR clusters reduplicate with a fixed vowel, [e]—as in [ge-graph^ a]—while base clusters other than OR do not reduplicate, and the output surfaces with a bare [e]—as in [e-ktona]. This raises the question of whether forms like [e-ktona] are reduplicated, i.e. analyzable as [[ e 3 ]R [k1 t 2 o3 n 4 a 5 ]B ]^15 , or simply prefixed, i.e. analyzable as [[ e 6 ][k 1 t 2 o3 n 4 a 5 ]]. The present analysis is constructed such that the constraints answer this question: [[ e 3 ]R [k1 t 2 o3 n4 a 5 ]B ] and [[ e 6 ][k1 t 2 o3 n4 a 5 ]] are both candidates for the output of perfect- inflected /ktona/.
Among the constraints choosing among these candidates are two morphological realization constraints, one favoring reduplication and one favoring prefixation:
(21) REDUPLICATE An output of morphological category X contains some pair of segments in base- reduplicant correspondence.
(22) PREFIX e- An output of morphological category X has the prefix e-.
(^15) Note that, following the guidelines established by Alderete et al (1997) for diagnosing phonological versus
morphological fixed segmentism, I assume that fixed segmentism in Greek reduplication (as well as in Gothic and Old Irish) is correctly analyzed as phonological—i.e. produced by a ranking in which IDENT-BR constraints on vowel features are interleaved with vowel markedness constraints such that [e] emerges as the optimal reduplicative correspondent for any base vowel.
(25) LEFT -ANCHOR-BR (= L-ANCHOR) (McCarthy and Prince 1995) Any element at the left edge of the base has a correspondent at the left edge of the reduplicant.
L-ANCHOR assesses a violation when the leftmost elements of base and reduplicant are not in correspondence. Thus, L-ANCHOR is violated by [[C 2 V 3 ]R [C 1 C 2 V 3 ]B ] and [[V 3 ]R [C 1 C 2 V 3 ]B ], but not by [[C 1 V 3 ]R [C 1 C 2 V 3 ]B ] or [[C 1 C 2 V 3 ]R [C 1 C 2 V 3 ]B ]; and of course, not by non-reduplicated forms.
(26) BEST ONSET For two syllables σ (^) α, σ (^) β with onset consonants standing in base-reduplicant correspondence, the onset of the reduplicant must contain a consonant in correspondence with the least sonorous consonant of the onset of the base.
BEST ONSET is violated by [[C 1 V 3 ]R [C 1 C 2 V 3 ]B ] when C 1 is more sonorous than C 2 , and by [[C 2 V 3 ]R [C 1 C 2 V 3 ]B ] when C 2 is more sonorous than C 1 ; BESTONSET is satisfied by [[C 1 C 2 V 3 ]R [C 1 C 2 V 3 ]B ], because by necessity the less sonorous consonant of the base onset has a correspondent in the onset of the reduplicant. Note that BEST ONSET is employed instead of previous proposals for the analysis of sonority-driven reduplication in Sanskrit, e.g. Gnanadesikan (1995), Morelli (1999). The necessity of this move is discussed in §4.5.1.
The final constraint figuring in the analysis is the phonotactic O NSET :
(27) ONSET (cf. Prince and Smolensky 1993) Every vowel is preceded by a consonant.
ONSET is violated by the vowel-initial forms [[V 3 ]R [C 1 C 2 V 3 ] (^) B ] and [[V 4 ][C 1 C 2 V 3 ]].
Finally, note that all constraints not specifically mentioned above are presumed inviolable: for example, I do not consider candidates in which consonant clusters are simplified in the base, as these would be ruled out by undominated M AXIO-C.
4.5 Factorial typology
The constraints defined in the sections above were submitted to a factorial typology calculation, using software (Hayes n.d.) that allowed the a priori rankings of the context-sensitive M AX-C constraints to be specified. The inputs and output candidates included in the calculation are in (28):
(28) The candidate set a. Inputs: /C 1 C 2 a 3 /, [+X] — where / pra / represents C 1 C 2 = TR, / sla / represents C 1 C 2 = SR, and / sta / represents C 1 C 2 = ¬OR b. Outputs: full transfer [C 1 C 2 a 3 -C 1 C 2 a 3 ], C 1 -copy [C 1 a 3 -C 1 C 2 a 3 ], C 2 -copy [C 2 a 3 -C 1 C 2 a 3 ], vowel copy [a 3 -C 1 C 2 a 3 ], and prefixation [e 4 -C 1 C 2 a 3 ]
Note that all reduplicated outputs have a copy vowel, [a], while the non-reduplicated prefixed outputs have the affixal vowel [e]. Vowel quality is not at issue here; the [a]/[e] distinction is
employed only to make it easier to distinguish on the page the vowel copy forms ([a 3 -C 1 C 2 a 3 ]) from the prefixed forms ([e 4 -C 1 C 2 a 3 ]). Also included in the typology was an input with a singleton onset, /C 1 a 2 /, with three possible outputs: copy of the base consonant [C 1 a 2 -C 1 a 2 ], vowel copy [a 2 -C 1 a 2 ], and prefixation [e 3 -C 1 a 2 ]. Only [C 1 a 2 -C 1 a 2 ] ever surfaces as the optimal candidate, regardless of constraint ranking; this is the desirable result, given the cross-linguistic data presented above in §2.
The factorial typology calculation produced eleven outcomes, which are summarized in the sections below.
4.5.1 Blind criterion
The constraint rankings for the blind criterion patterns, Old Irish and Sanskrit, are shown below in (29) and (30). Constraints belonging to a single ranking stratum are in boxes; stratum- internal ranking is non-crucial.
(29) Old Irish = [ s a- st a], [ s a- sl a], [ p a- pr a]
REDUPLICATE , ONSET , C/V, L-ANCHOR , M AX -C/#_C » PREFIX e- , M AX-C, BESTO NSET , M AX -C/C_V , M AX-R/S_V, M AX-R/T_V
(30) Sanskrit = [ t a- st a], [ s a- sl a], [ p a- pr a]
REDUPLICATE , ONSET , C/V, BESTO NSET , M AX (^) BR-C/C_V » PREFIX e- , M AX-C, L-ANCHOR , M AX -C/#_C , M AX-R/S_V, M AX-R/T_V
The differences in ranking that crucially distinguish Old Irish and Sanskrit are highlighted. In both cases, C/V, REDUPLICATE , and ONSET are all undominated by any relevant constraints. Thus, the only viable candidates are reduplicated, with a singleton onset in the reduplicant: that is, the only candidates to satisfy the top-ranked constraints are [C 1 a-C 1 C 2 a] and [C 2 a-C 1 C 2 a]. Exactly which member of the base cluster is copied is determined by the relative rankings of L- ANCHOR and BEST ONSET (and by necessity, the relative rankings of MAX-C/#_C and M AX- C/C_V). When L-ANCHOR is ranked above B ESTONSET , the leftmost member of the base cluster is copied, as in Old Irish:
(31) Old Irish
L-ANCHOR M AX-C/#_C BEST ONSET M AX-C/C_V ) sa-sta * * /sta/, [+X] ta-sta *! *! /sla/, [+X] )^ sa-sla la-sla *! *! * ) (^) pa-pra /pra/, [+X] ra-pra *! *! *
below containing the relevant portion of the constraint hierarchy shows how the outcome set is generated:
(34) Language Q = [ t a- st a], [e- sl a], [ p a- pr a], [ t a- t a], [e- l a]
μ/liquid μ/fricative ONSET μ/stop L-ANCHOR sa-sta **! * ) ta-sta * ** * a-sta * *! * * /sta/, [+X}
e-sta * *! * sa-sla * **! la-sla **! * * a-sla * * * *!
/sla/, [+X]
) e-sla * * * ) ta-ta ** /ta/, [+X] a-ta *! * e-ta *! * sa-sa **! /sa/, [+X] a-sa * * *! ) (^) e-sa * *
With O NSET ranked above μ/stop and L-ANCHOR, but below μ/liquid and μ/fricative, inputs containing a stop in the onset reduplicate, by copy of the stop; but inputs lacking a stop in the onset do not reduplicate, and instead are prefixed (which avoids violation of L-ANCHOR).
The outcome set discussed above is one of many unattested patterns generated by the constraint set containing the μ/X family; similar results obtain if the μ/X family is replaced by *SONORANT SIMPLE ONSET » *FRICATIVE SIMPLE ONSET » *STOP SIMPLE ONSET , or by a constraint requiring singleton onsets to be non-sonorous, and assessing increasing numbers of violations as sonority increases. In short, unless the choice of which segment to copy in order to minimize onset sonority is relativized to those segments actually present in the base, the analysis predicts extremely unusual patterns in which forms reduplicate or do not, or simplify or do not, based solely on whether to do so would result in a reduction in the number of "undesirable" onsets.
4.5.2 Sufficient copy
The constraint ranking for the sufficient copy pattern exemplified by Klamath is shown below:
(35) Klamath = [ st a- st a], [ sl a- sl a], [ p a- pr a]
REDUPLICATE , ONSET , BEST ONSET , L-ANCHOR, M AX -C/#_C , M AX -C/C_V , M AX -R/S_V » PREFIX e- , C/V » M AX-C, M AX -R/T_V
With REDUPLICATE and ONSET undominated, the only realistically possible outputs are reduplicated, and with at least a singleton onset: i.e., the viable candidates are [C 1 a-C 1 C 2 a], [C 2 a-C 1 C 2 a], and [C 1 C 2 a-C 1 C 2 a]. The question of which complex onsets emerge in the reduplicant, and which complex onsets are simplified, is settled by the relative ranking of C/V with respect to the context-sensitive MAX-C constraints (these constraints are highlighted in the diagram above). When C/V dominates only MAX-R/T_V, as in Klamath, only TR clusters are compelled to simplify under reduplication, while all other onsets are fully copied:
(36) Klamath
M AX-C/#_C M AX-C/C_V M AX-R/S_V C/V M AX-R/T_V
) sta-sta ** /sta/, [+X] sa-sta *! * ta-sta *! * ) (^) sla-sla ** /sla/, [+X] sa-sla *! * la-sla *! * pra-pra **! /pra/, [+X] ) pa-pra * * ra-pra *! *
High-ranking context-sensitive MAX-C constraints protect all clusters other than TR from the demands of C/V. However, because C/V dominates M AX-R/T_V, it has the power to assert its preference for simple onsets in just the case of base TR; C/V is satisfied at the cost of violating only the low-ranked correspondence constraint regulating preservation of a sonorant in the T_V context.
The difference between Klamath and Gothic is only the reranking of C/V with respect to M AX-R/S_V:
(37) Gothic = [ st a- st a], [ s a- sl a], [ p a- pr a]
REDUPLICATE , ONSET , BEST ONSET , L-ANCHOR, M AX -C/#_C , M AX -C/C_V
PREFIX e- , C/V » M AX-C, M AX -R/S_V , M AX -R/T_V
With M AX-R/S_V now ranked below C/V, all OR clusters—not just TR—are simplified under reduplication:
