Remarks on URs-as-violations

Prosodic shape requirements for lexical items

These are taken to motivate 'underlying' syllabification, pedification etc. across the board. Let's examine some ...

Roots are monosyllabic: instead of reference to the syllable, require 'forall roots: exists exactly one sonority peak'.
Roots are bimoraic: assuming the previous shape req., 'forall roots:exists >= 1 (virtual) sonority fall', with the concept of virtual sonority resolving sonority plateaus such as long vowels-as-XX positions (cf. Walther '92,'93,'95, found here).

The advantage of the above reformulations is that they steer clear of non-monotonic resyllabification ("no non-structure-preserving behaviour").

Requiring lexical items to have a certain (minimal/maximal) shape is what linguists are after. A postulated dichotomy between Underlying Representations (URs) and Surface Representations is not necessarily connected to that issue!
The lexical representation of (vowel) length by means of two X/segmental positions does the job. Invoking moras for this purpose is not necessary. The latter concept has been extensively criticized by Noske (1992).

Given unordered syllable structure, strings are not always predictable

German examples follow (additional linear precedence statements in brackets):

t'icktack 'clock noise' vs T'aktik 'tactics' [tIktak],[taktIk] ( i </> a)
m'ono- 'deriv.prefix' vs. n'omo- 'deriv.prefix' (n </> m) ; eg. in Monographie vs Nomographie
K'ilo 'kilogramm' vs L'oki 'first name of wife of former chancellor Schmidt' (o </> i, k </> l)

Also non-redundant linear precedence within coda (eg. Engl. Max vs task) and onset (eg. Germ. Psalm vs Spitz, Northern NHG pronunciation). Although finer classification of peripheral syllable roles using pre/appendix constituents is possible, this looks artifical and weakens the elegance of a unordered prosodic approach.

Prosodic underspecification and lexical shape are incompatible

To avoid resyllabification, one is tempted to leave alternating 'features'/prosodic roles underspecified or 'floating'. But to ensure bimoraicity, say, in roots like aj, gam, paek the alternating final consonant C (which will become an onset when V-initial suffixes are added) must not be underspecified. This is because bimoraicity needs to know C is in the rhyme, not in the onset, to add to syllable weight. Morale: mononicity and lexical prosodic shape constraints using prosodic constituents cannot both be satisfied.

Shape constraints may well be a complex statistical tendency

A parallel case involves alleged robust OCP effects on permissible root consonant sequences in Arabic. Pierrehumbert(1993) has shown this to be a statistical tendency only, revealing several subregularities and exceptions. Hence it is most probably wrong to reflect this directly in phonological grammar. So where are the large-corpora-based statistics of shape constraints like the above, and what do they reveal?

Morphemes as constraints are needed anyway

Bolozky (1978:17-19) describes Modern Hebrew C/0 alternations confined to single morphemes and minor rules affecting very small groups of morphemes. These morphemes are regular w.r.t. to other grammatikal processes such as inflection etc. Now, in all-constraint theories like DP and OT it is natural to try to cover all alternations with constraints, disregarding questions of universality for now. OT is especially suited for this, since it was expressly designed to cope with the except when case that arise often in linguistic analysis (Prince & Smolensky:ch.4). The above exceptional alternations in an otherwise regular setting is the prototypical "except when" situation, hence should be dealt with the mechanism that's already there, violable constraints plus ranking. In effect, contra Prince (p.c.,optimality-list), OT being a default-based formalism (Ellison 94) already has the mechanism to deal with exceptions.

So what's so problematic about URs-as-violations?

Most importantly,violation marks may simply not be available! Prince & Smolensky do not clearly state them to be a ontological primitive of OT like violable constraints, ranking or defining solutions to be most harmonic. A very reasonable view takes them to be mere graphical devices to convey constraint computation to the linguist. Demanding violation marks to be "real" onotological primitives forces specific technical implementation detail into OT, hence pertains to OT performance or actual computation rather than OT competence. Theoretical linguists commonly concentrate on compentence and leave performance issues to implementors/CS people.
The fatal violation mark (!) in tableaux already is a device to signal that there is no purpose in more constraint evaluation. Optimizing implementations would take this into account or even develop radically different modes of computing with OT, with no guarantee of literal violation marks present at any level of the system. URs-as-violations would prematurely preclude all such advanced computational models of OT. It is difficult to see whether existing proposals for OT computation (Ellison 1994, Tesar 1995) are compatible with URs-as-violations: most likely they are not.
"Grammars do not count" - so what about 7 *ONS or 13 *FILL as crucial parts of URs-as-violations?
If memorized violation patterns determine the identity of morphemes, morpheme combination will normally create new and different patterns, due to new syllabic roles, foot restructuring etc. Morpheme recoverability or compositionality demands that the new violation patterns be properly distinguished from UR-defining ones. This is a diacritical distinction, probably for every mark, which might as well be named +/- underlying or +/- derived. In fact, it's completely unclear how morpheme combination should actually work in the technical sense.
Speech is transmitted over a serial communications channel that operates in space-time (this includes sign language), so linear order comes for free in daily speech usage and acquisition. The most natural strategy would seem to exploit this fact, matching performance linear order with compentence representational content. Why deviate? There would have to be extremely good reasons to not exploit linear order.
Assuming one could elaborate on present proposals for URs-as-violations to always recover string-based traditional UR's including complete linear order from sets of violation patterns -- what would be the decisive advantage that cannot be achieved with present OT machinery? That's hard to see given that OT is already a powerful default system.

Markus Walther walther@ling.uni-duesseldorf.de