DIRECTIONAL SYLLABIFICATION IN GENERALIZED ALIGNMENT Armin Mester & Jaye Padgett University of California, Santa Cruz This note[*] presents a reconstruction of directional syllabification (Ito 1989, 241-254; Ito 1986, 163-219; see also Broselow 1980, 1982 and Selkirk 1981 for earlier analyses) within Generalized Alignment (McCarthy and Prince 1993; henceforth: "GA"), a recent proposal in Optimality Theory (Prince and Smolensky 1993, and related work). The guiding idea of this line of investigation is the Optimality-Theoretic notion that constraints are ranked and minimally violable ("Harmonic ordering of forms", see Smolensky's (1988, 19) Best Fit Principle and in particular P&S (1993, 68-75)). Our proposal makes use of a central analytical strategy in GA which derives foot directionality effects by means of a constraint requiring the alignment of every foot edge (L/R) with a prosodic word-edge (L/R). As demonstrated in GA, the doctrine of minimal violation ensures the selection of the candidate displaying the desired directional footing pattern (a suggestion originally due to R. Kirchner). This is so in spite of massive violations of the alignment constraint incurred by all feet not directly located at the relevant word edge. In broadest outline, the basic strategy for applying the alignment idea to directional syllabification is clear enough; somewhat surprisingly, however, a certain degree of analytical subtlety is required to get even the simplest patterns to emerge, which is all we intend to do in this note, leaving all further details, and the inevitable complications of the empirical, for future exploration. Taking up the 'alignment' strategy of GA, we show that directional syllabification patterns are the result of alignment constraints requiring that every syllable edge coincide with a prosodic word-edge. The general schema is given in (1), where E(dge) is a variable ranging over {L(eft), R(ight)} and "S" abbreviates "syllable"[**]: (1) S-Align (S,Edge,PrWd,Edge) For clarification, we note that this statement is to be interpreted as quantifying UNIVERSALLY over tokens of S and EXISTENTIALLY over tokens of PrWd (see GA for a more formal development and for further exemplification): EVERY syllable must be Edge-aligned with SOME prosodic word. In most situations, (1) is massively violated by syllables not located at Edge. For a given input, the optimal candidate is the one with the fewest violations (modulo the overall ranking of constraints, as usual in OT). RL syllabification and epenthesis (as in Iraqi Arabic; see Ito 1989 and other refs.) shows the pattern in (2) (the "Coda"- pattern: the stray c is syllabified as a coda to the epenthetic vowel (whose site is marked by capital 'V'): (2) [...]cvcccv[...] -> [...].cv.cVc.cv.[...] Thus /gil-t-l-a/ appears as .gi.lIt.la. 'I said to him', etc. This pattern follows from the L-Edge version of the alignment constraint in (3), under the overall constraint ranking in (4). (3) S-ALIGN(L): Align (S,L,PrWd,L) (4) | ONSET | >> |S-ALIGN(L)| >> NO-CODA | PARSE | | FILL | |NO-COMPLEX| In order to see how this works, the reader is invited to inspect the tableau in (5). We leave out the undominated constraints, which can in fact be grouped together with Gen, for the purposes of the present discussion. Taking up a suggestion made to us independently by Junko Ito and John McCarthy, we hypothesize that violations of alignment constraints are always reckoned in terms of the prosodic unit immediately subordinate in the prosodic hierarchy to the unit referred to in the first part of the constraint (in line with the theory of prosodic layering of Ito & Mester (1992, 1993), cf. the constraint ClosestMother and the principle of hierachical locality). (4) quantifies universally over syllables (loosely: "is a constraint on syllables"), hence it makes sense for violations to be computed in terms of moras. Note that the evaluation of S-ALIGN violations can proceed by summing over all syllables, not syllable-by-syllable in some ranked fashion (see GA (p.15) for a similar observation regarding FT-ALIGN). (5) RL syllabification tableau (counting syllables LR, "m" = "mora"; (!)...(!) indicates indeterminacy of fatality of violation, in the case of unranked constraints; "$" marks the winner, in the notation of Grimshaw 1993): || FILL : S-ALIGN(L) | NoCoda ============||=======:================================|========== /cvcccv/ || : S1 S2 S3 S4 | ------------||-------:--------------------------------|---------- cv.cVc.cv || * : m mmm | * $ ------------||-------:--------------------------------|---------- cvc.cV.cv || * : mm! mmm | * ------------||-------:--------------------------------|---------- cv.cV.cV.cv || **(!): m mm mmm(!) | ------------||-------:--------------------------------|---------- LR syllabification and epenthesis (as in Cairene Arabic; see refs.) shows the pattern in (6) ("Onset"-pattern: the stray c is taken as the onset to the epenthetic vowel): (6) [...]cvcccv[...] -> [...].cvc.cV.cv.[...] Thus /?ul-t-l-u/ appears as .?ul.tI.lu. 'I said to him', etc. This pattern follows from the R-Edge version of the alignment constraint (7), under the overall constraint ranking in (8). (7) S-ALIGN(R): Align (S,R,PrWd,R) (8) | ONSET | >> |S-ALIGN(R)| >> NO-CODA | PARSE | | FILL | |NO-COMPLEX| Tableau (9) illustrates this point. (9) LR syllabification tableau (counting syllables RL) || FILL : S-ALIGN(R) | NoCoda ============||=======:================================|========== /cvcccv/ || : S1 S2 S3 S4 | ------------||-------:--------------------------------|---------- cv.cVc.cv || * : m mmm! | * ------------||-------:--------------------------------|---------- cvc.cV.cv || * : m mm | * $ ------------||-------:--------------------------------|---------- cv.cV.cV.cv || **(!): m mm mmm(!) | ------------||-------:--------------------------------|---------- The major result of Ito's (1989) analysis is its ability -- without invoking any additional constraints -- to account for the convergence of the two opposing patterns (RL: Iraqi Arabic; LR: Cairene Arabic) in the case of CCCC clusters: Both languages syllabify this case as in (10). (10) [...]cvccccv[...] -> [...].cvc.cVc.cv.[...] Thus RL Iraqi Arabic syllabifies /gil-t-l-ha/ as .gil.tIl.ha. 'I said to her', and LR Cairene Arabic answers by syllabifying /?ul- t-l-ha/ as .?ul.tIl.ha. "I said to her', etc. Our alignment analysis replicates Ito's result, as demonstrated in (11) and 12). (11) RL syllabification of /...cccc.../ (counting syllables LR): || FILL : S-ALIGN(L) | NoCoda ============||=======:================================|========== /cvccccv/ || : S1 S2 S3 S4 | ------------||-------:--------------------------------|---------- cvc.cVc.cv || * : mm mmmm | ** $ ------------||-------:--------------------------------|---------- cvc.cV.cV.cv|| **(!): mm mmm mmmm(!) | * ------------||-------:--------------------------------|---------- cv.cV.cVc.cv|| **(!): m mm mmmm(!) | * ------------||-------:--------------------------------|---------- (12) LR syllabification of /...cccc.../ (counting syllables RL) || FILL : S-ALIGN(R) | NoCoda ============||=======:================================|========== /cvccccv/ || : S1 S2 S3 S4 | ------------||-------:--------------------------------|---------- cvc.cVc.cv || * : m mmm | ** $ ------------||-------:--------------------------------|---------- cvc.cV.cV.cv|| **(!): m mm mmm(!) | * ------------||-------:--------------------------------|---------- cv.cV.cVc.cv|| **(!): m mmm mmmm(!) | * ------------||-------:--------------------------------|---------- This approach derives (some version of) syllabic compression (the Minimization Principle of Selkirk 1981) as a theorem: the more syllables, the more Fill violations, and the more alignment violations (sharing the first part of the consequent with other current proposals). Hence Ito's directionality-driven convergence effect between LR- and RL-systems in CCCC clusters follows directly, without additional legislation. Note the redundant overkill by Fill and Align; but note also that syllabic compression effects can be obtained, in this approach, INDEPENDENT of dominated FILL (i.e., independent of epenthesis). Our more disturbing findings include the fact that, in the current version of GA theory, the LR and RL patterns can also be derived under an OPPOSITE Edge setting of the alignment constraints, i.e. as Align(S,Edge,PrWd,Edge'), as the reader may verify for her/him-self by recasting the tableaux given above in the appropriate way. And they can be derived by counting segments instead of counting moras. They can even be derived with a strictly segmental setting, requiring each SEGMENT to be as closely aligned to one edge as possible (this is sufficient to push the epenthetic vowel as far in the opposite direction as possible, with multiple epenthesis discouraged as before). Besides hierarchical locality considerations discussed above that would militate against the segment counting version, this indicates to us that the theory of GA needs to be tightened up. The task is to isolate, among the set of statements of the form "Align(x,Edge(i),y,Edge(j))" made available by the general Alignment scheme, those that have true linguistic significance. We anticipate that ROW-1 will constitute a major step in this direction. Our proposal has barely scratched the surface of directional syllabification patterns, and there are many further issues to be dealt with. Junko Ito and John McCarthy have both pointed out that all the proposals presented above fail to adequately deal with the way in which INITIAL #cc clusters are resolved in the various dialects of Arabic (a topic investigated in the forthcoming U Utah dissertation by Samira Farwaneh). Given our ignorance about the empirical details, we refrain from fabricating hypotheses, and leave the issue unresolved. We nevertheless take comfort in the fact that SOME cases of special behavior in PrWd-initial position, in non-Arabic languages, have a straightforward account in our scenario. Thus Ito (1989, 252-254) analyzes Temiar epenthesis as basically RL (Coda-preference; in our terms, ALIGN-L), in spite of the fact that INITIAL #ccv-configurations are resolved as #.cV.cv... and not as #.Vc.cv... This follows directly in our approach: In #.cV.cv..., the second syllable is separated from the initial word-edge by one mora; in #.Vc.cv..., the mora count is two. More importantly, ONSET is a dominating constraint in Temiar, as pointed out by Ito (1989), whose gradiently fulfillable principle "Avoid Onsetless Syllables" is among the precursors of current Optimality-Theoretic reasoning. [*] We gratefully acknowledge significant help from discussions and correspondence with Junko Ito, John McCarthy, Alan Prince, and Cheryl Zoll; it was Cheryl's work on Yawelmani Yokuts that first inspired us to confront the issues dealt with in this note. As usual, all remaining shortcomings are our responsibility alone. [**] John McCarthy and Alan Prince independently pointed out the feasibility of such a 'same-edge' analysis. Our original proposal was stated in opposite-edge terms.