Variation as accessing‘non-optimal’ candidates*Andries W. CoetzeeUniversity of MichiganThis paper argues that rather than just select the best candidate, EVAL imposes aharmonic rank-ordering on the full candidate set. Language users have accessto this enriched information, and it shapes their performance. This paper appliesthis idea to variation. The claim is that language users can access the full candidateset via the rank-ordering imposed byEVAL. In variation, more than one candi-date is well-formed enough to count as grammatical. Consequently, languageusers will access more than just the best candidate from the rank-ordering.However, the accessibility of a candidate depends on its position on the rank-ordering. The higher the position a candidate occupies, the more likely it is to beselected. In a variable process, variants that appear higher on the rank-ordering(i.e. are more well-formed) will therefore also be the more frequent variants. Thismodel is applied to variation in the phonology of Faialense Portuguese andIlokano.Most linguistic theories in the generative tradition are categorical innature – they map an input onto a single, grammatical output. Naturallanguage, on the other hand, is full of non-categorical phenomena. Forinstance, a single word can often be pronounced in more than one way.There is a large literature on variation, both in sociolinguistics and informal phonological theory. One fact on which all the literature agrees isthat variation is not random, but is strongly influenced by grammar. Thisposes a challenge to classical generative grammar, which is a categoricalfunction by design.Optimality Theory (Prince & Smolensky 1993) has an advantage overmost other generative theories in this regard. In other generative theories,the grammar generates only a single form. However, a basic design feature* I am grateful to Pam Beddor, John Kingston, John McCarthy, Michael Marlo,Barbara Partee, Joe Pater, David Silva, the participants of Linguistics 792(University of Michigan, Fall 2004) and the members of the UMass phonologygroup, as well as the audiences at the LSA Annual Meeting (2004, Boston) andMCWOP 10 (2004, Evanston) for discussion, suggestions and encouragement. Anassociate editor and several anonymous reviewers also contributed greatly to theimprovement of this paper. Special thanks to the editors and the associate editor fortheir patience, encouragement and help. I am ultimately responsible for all viewsexpressed in this paper.Phonology 23 (2006) 337–385. f 2006 Cambridge University Pressdoi:10.1017/S0952675706000984 Printed in the United Kingdom337of OT is that the grammar generates more than one potential output form.In standard OT, only one of these potential output forms gains the statusof an actual output. But the other potential outputs still exist. It is there-fore not necessary to add generative power to an OT grammar. Whatneeds to be added to the grammar is a mechanism that will allow, in somecircumstances, more than one of the already generated possible outputforms to become actual outputs.In this paper, I present a novel approach to variation in OT that doesexactly this. In this model, variation does not arise as a result of variationin grammar (ranking) itself. I argue thatEVAL imposes a well-formednessrank-ordering on the full candidate set. In speech production, thelanguage user then has access to the full candidate set via this rank-ordering. The likelihood of a candidate being selected as output dependson the position the candidate occupies on the rank-ordering – the highera candidate appears, the more likely it is to be selected. In most circum-stances, only the topmost candidate on the rank-ordering is well-formedenough to be selected as output. In some situations, however, the toptwo (or more than two) candidates are both well-formed enough and canboth surface as grammatical outputs. The output of the grammar istherefore invariant – the same rank-ordered candidate set is output everytime. But how the language user uses this output varies. Sometimes s/heaccesses the topmost candidate as output, and sometimes s/he accesses acandidate lower down on the rank-ordering. A fundamental differencebetween the model developed in this paper and other OT models ofvariation is the locus of variation. In other models (Reynolds 1994,Anttila 1997, Nagy & Reynolds 1997, Boersma 1998, Boersma & Hayes2001, Ja¨ger 2003, etc.), variation resides in the grammar itself. Theranking between the constraints can vary from one moment to the next,so that a different candidate can be selected as optimal on different oc-casions.This paper is structured as follows. In w1, I discuss the theoretical de-tails of the model that I propose. The next two sections are dedicatedto illustrating the application of this model to two sets of variable data.In w2, variable deletion of unstressed vowels in Faialense Portuguese(Silva 1997) is discussed, and in w3, variation in the phonology of Ilokano(Hayes & Abad 1989, Boersma & Hayes 2001). w4 considers outstandingissues and domains other than variation in which the model can beapplied.1 The proposalThe basic proposal of this paper is that EVAL does more than simplyselect the best candidate. It rather imposes a harmonic rank-orderingon the full candidate set. I will therefore refer to this model as the ‘ rank-ordering model ofEVAL ’, or ROE. This section explains the detailsof ROE.338 Andries W. Coetzee1.1 Imposing a harmonic ordering on the full candidate setIn classical OT,EVAL is assumed to distinguish the winner from the losers,but not to make distinctions within the set of losers between the more andthe less well-formed. As soon as a candidate has been eliminated from therace for optimal status, all of its violations in terms of lower-ranked con-straints are ignored – indicated by shading the cells for these constraints inOT tableaux. If the information in the shaded cells is not ignored, then itis possible to order even the non-optimal candidates in terms of their well-formedness. To illustrate the point, imagine a language that allows closedsyllables but that avoids tautosyllabic consonant clusters by deletion – i.e.a language with the grammar *COMPLEX6MAX6NOCODA. The tableau in(1) shows how this language will evaluate output candidates for an inputwith both a consonant cluster and a coda.(1)pakpaprakpra/prak/1.2.3.4.*Complex NoCodaMax***!****!*!In classical OT, the grammar will say that [park] is