I've heard several times the terms 'serial' and 'parallel' in phonology. In one of them the context was SPE and OT, respectively. However, from my feeble understanding of OT, the output is determined in rounds, such that in each round a different constraint is considered (according to the language specific ranking). Could someone elaborate these terms?

2 Answers 2


For those linguistic theories that distinguish underlying from surface expressions by deriving surface from underlying using transformation rules of some sort, when several changes are required to get from underlying to surface, one must decide how many steps the derivation can have. If there are many changes involved and only two steps, that requires parallel processing, since all the changes must be made at once, but if many derivational steps are countenanced, perhaps serial processing is sufficient, with changes made one at a time (i.e., serially).

The usual standard of comparison is the SPE theory, because it is so carefully worked out (not necessarily because it is correct). In SPE, each single rule applies in parallel to all sound segments the rule is applicable to, but distinct phonological rules must apply serially, so that derivations ordinarily have many steps.

  • Thanks, so both SPE and OT are serial, since the number of steps equals equals to the number of rules/constrains?
    – dimid
    May 14, 2017 at 23:17
  • SPE is both serial and parallel. It's complicated. Each cyclic rule applies in parallel everywhere it is applicable within a cyclic domain, but distinct rules apply serially in accordance with their linear order. Theories other than SPE differ wildly. I have proposed a phonological theory in which all processing is parallel, outlined here: linguistics.stackexchange.com/questions/15148/…
    – Greg Lee
    May 15, 2017 at 2:09

The distinction is generally framed in terms of the notion of "time-invariant computation", whether the output of one computation requires the results from another computation to be available. In a serial derivation, as in the case of a standard pre-OT phonology, it is possible for one rule to apply first, and apply a second rule after that – in a series. Thus the plural "bushes" /bυʃ-z/ → bʊʃɨz to which devoicing cannot apply. You have to know if the vowel is to be inserted, in order to correctly not apply devoicing, which could apply to /bυʃ-z/. The standard account is then to say you first actually do epenthesis, and then do devoicing to whatever results (and insertion of a vowel prevents devoicing of /z/).

On the other hand, if a language has a rule of intervocalic voicing and one palatalizing velars before front vowels, the derivation of /iki/ → [igʲi] does not require serial application of the rules, because no information provided by applying one of those rules is part of the information required for the other rule to apply (changing place of articulation by palatalization doesn't affect voicing since voicing doesn't care about place of articulation; voicing doesn't affect palatalization, since palatalization applies to velars regardless of voicing). In that case, we might say that the rules can be applied "in parallel". There actually two views of what that would mean, one being that the two rules literally apply at the same time, the other being that the rules apply one after the other, but it doesn't matter what order they apply in.

In classical OT, the computation of stars with respect to a given constraint is independent of the computation of stars for any other constraint. All constraints are evaluated in parallel (can be done at the same time, or in any random order). The potential for seriality arises from the adjudication of a tableau full of stars: if there are two constraints A, B ordered s.t. A precedes B, then {*,**} beats {**,*}. The easiest way to conceptualize this is to process the table of stars serially from left to right, but that isn't how they do it. There is a dense discussion of recursive summing-up of stars in Prince & Smolensky 1993. If you happened to be familiar with signal processing and A-to-D conversion as the simultaneous parallel weighted comparison of voltages, this might make some sense, otherwise I grant that it looks like hand-waving.

Classical OT does have serial elements: GEN must precede EVAL, assignment of stars must precede the winnowing of stars. And then there are post-classical developments ("classical" lasted about 4 years) which added a number of serial computations – Sympathy theory (star assignment in one column requires knowledge of all of the stars and the filtering, of a subset of candidates); Output-Output constraints (you have to know the results of the computation of one word before you can determine the star-assignment for another related word). Stratal OT also have multiple cycles of evaluation, so is somewhat serial (imposing an order between levels, but not on the derivation within a level). In fact, "Candidate Chains" cleans all of this mess up, by writing derivational steps left-to-right separates by commas (e.g. deʃʔ,deʃeʔ,deʃe is one candidate, where the last item is what is pronounced. OT-CC essentially abandoned the pretense of meaningful parallelism.

  • Thanks, could you point me to the chapter in P&S 1993 where this is discussed?
    – dimid
    May 15, 2017 at 13:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.