Wednesday, May 28, 2008

Opacity, part III: What's 'easy'?

After a hiatus to take care of other business and read his recent Phonology paper, here are a few notes on Eric's third challenge (again full text at the bottom of the post):
3. SPE(-and-subsequent-developments) “could deal with [opacity] easily”.
Eric adds, in a comment on the last post:
The point of that is to say that derived environment effects are cases of opacity that can't be easily handled by rule ordering alone.
The person I originally paraphrased had given simple opacity examples, like Canadian Raising. But we can expand the notion to include derived environment effects (DEE). And they do go beyond what's directly predicted and handled automatically by rule ordering alone. I'm not sure that changes much from part II: Classic opacity from the interaction of postlexical rules is predicted by and falls out automatically from generative phonology. By contrast, Baković (2007:220) quotes McCarthy on this:
Unless further refinements are introduced, OT cannot contend successfully with any non-surface-apparent generalizations, nor with a reside of non-surface-true-generalizations.
But are these mere 'refinements'? Simple constraint ranking as a way of getting optimal forms is one thing, while allowing conjunction among constraints (like other similar refinements) vastly changes and complicates OT's basic mechanics. Some old-school linguists who are waiting for the theory bus to turn the corner are left scratching their heads as to why OT crucially relies on gymnastic moves precisely where one shouldn't.

In traditional phonology, it's when we get to how morphology interacts with phonology, as in prototypical DEE, that we need additional tools. You need nothing fancy to handle purely phonological opacity, but you might where phonology and morphology meet.

Having read "A revised typology of opaque generalizations", I see why Eric got so bent over a quip that opacity is easy in earlier models and hard for OT: His article aims to debunk precisely this view. It's a balanced account, where everybody has trouble with some types of opacity and can handle others. Still, the problems he hangs around the neck of traditional models are quite different from those OT faces (above). Eric concedes that Vaux's non-OT analysis of New Julka Armenian is "technically true" (p. 246), but Vaux "presents no specific arguments for this ordering analysis, empirical or otherwise". Vaux's ordering produces the right output, while the reverse ordering does not. It's not that a traditional analysis won't work, but that Eric doesn't find it sufficiently motivated?

Eric was right that I was going to skirt Challenge III, and I appreciate him prodding me. As hinted at above, I wonder if he isn't onto something about how phonology-morphology interactions appear more contortion-prone than plain phonology. When students gripe about the drivel they claim to get from their phonology or morphology profs (apologies to the peerage), often it is over such resultant interactions (or what happens to newbie sounds trying to squat on derivational territory occupied by old sounds at stem's edge). But what's odd about the phon-morph relation is that students often take a different approach from paid phonologists: They're content to load the lexicon, a Third Way of the Declarative Phonology sort. Correct me on this, not being a phonologist and all, but serial models of sound systems seem not to miss cross-lexicon generalizations brought about by loading the lexicon nor to create unnecessary problems by restructuring the nature of constraint evaluation.

With those questions tossed out there and with my flame retardant suit on, I'll leave this for now, but will return to Challenge IV soon, and maybe even a postscript after that. I'm just trying to figure this stuff out.

Challenge III:
3. SPE(-and-subsequent-developments) “could deal with [opacity] easily”.
Although it is conveniently hardly ever talked about in this way, ‘derived environment effects’ are precisely cases of opacity that cannot be “dealt with easily” in SPE, by which I mean that rule ordering on its own won’t do the trick. Consider the classic Finnish assibilation case (whether or not you believe it): t → s / __i, but “only in derived environments”. Let’s stick to phonologically-derived environments here: assibilation does not apply to underlying /ti…/, because no part of the structural description of assibilation is derived; assibilation does apply to /…te/, because there is another rule raising word-final /e/ to [i], which means that this part of the structural description of assibilation is derived by raising. (So: /ti…/ → [ti…], but /…te/ → |…ti| → […si].)

Several things to note here. First, this is a case of opacity of Kiparsky’s (1973) “Type (i)”: assibilation is not surface-true, because there are surface strings of the kind that match the structural description of assibilation (to wit, [ti…]). Second, although many examples of non-surface-truth can be accounted for with counterfeeding rule ordering, this kind of case cannot, which is why some separate principle(s) responsible for derived environment effects are necessary — hence my conclusion that SPE does not, in fact, “deal with [opacity] easily” in the way Mr. Verb’s summer-upper and many others claim. Third, how does the distinction between ‘derived’ and ‘non-derived’ require any less “gymnastics” than the one between ‘new’ and ‘old’ markedness constraint violations, the basic idea behind comparative markedness? I’d really like to know.
Image from here.

15 comments:

Anonymous said...

Do I understand this? Classic OT cannot get opacity at all and yet this paper criticizes Vaux for merely getting the right answer on opacity.

That's like being the guy in logic class who can't get a proof to work, but criticizes others for workable proofs that aren't elegant enough.

Ed Keer said...

I'm not sure how you measure the level of gymnastics. You claim that constraint conjunction (and other refinements) "vastly" differ from classical OT. What does that mean? Is it a question of adding power? What's the distance between a theory that has undordered rules and one that has crucially ordered rules? Is that distance bigger or smaller than a the distance between Classical OT and OT with conjoined constraints? Inquiring minds want to know.

Mr. Verb said...

Thanks, both. Anon, I'm not sure what's what there. Certainly, within traditional generative phonology something like the Lithuanian data Eric treats do involve a bigger generalization (an OCP-type ban on adjacent identical Cs) which old-style phonology doesn't instantly get.

Ed, measuring gymnastics is obviously a massive undertaking, one I wish phonologists would pay more attention to. The key point in the current context might be that traditional generative phonology had ordering of rules built into the system (in actual SPE) as part of the system. Things like local conjunction and sympathy theory are add-ons and they change how evaluation works. It's never been clear to me how we constrain them. (Doesn't McCarthy say that we need to figure out how to constrain local conjunction somewhere like the Thematic Guide?)

Would you argue that conjunction doesn't add significant difference from classic OT? Could you lay out what level of modification you think it is to the theory?

Ed Keer said...

As I mentioned in an earlier comment, sympathy theory does rely on the classical OT mechanism of faithfulness. So while its not 'built into the system' it is consistent with classical OT.

To me, both sympathy and constraint conjunction are reasonable extensions of OT. The thing they seem to be doing is refining the notion of 'constraint'. Its not like they change the fundamental idea of ranked and violable constraints, or introduce some other mechanism, into the system. They seem parallel to non-linear phonology--which expanded the notion of what a rule could do.

It seems like the beef is that P&S didn't provide a theory of everything. It sounds like your saying that if sympathy or constraint cunjunction were in P&S then you wouldn't see them as gymnastics. But since they are not, then they have to be viewed as add-ons.

Mr. Verb said...

Only time for a quick note now:

--Sympathy Theory especially has proven profoundly problematic, for reasons implicit in much of the present discussion. So, is McCarthy's new Harmonic Serialism unnecessary, in your view? (Sympathy and HS are different creatures in some ways, but I'm wondering about your big-picture view of how the theory needs to work.)

--My point wasn't about expecting a 'theory of everything': The tacked-on feel of conjunction and such is not historical but architectural.

Eric said...

In traditional phonology, it's when we get to how morphology interacts with phonology, as in prototypical DEE, that we need additional tools. You need nothing fancy to handle purely phonological opacity, but you might where phonology and morphology meet.

Look at my "challenge 3" again: I specifically mentioned phonologically-derived environment effects, not morphogically-derived ones. The former have nothing whatsoever to do with morphology (at least in principle), and so are cases of "purely phonologial opacity" that rule ordering needs DEC-style gymnastics to handle.

Eric said...

Do I understand this? Classic OT cannot get opacity at all and yet this paper criticizes Vaux for merely getting the right answer on opacity.

Anon (if I may): I suggest reading the article -- or at least the relevant section -- before jumping to conclusions based on (your reading of) Mr. Verb's drive-by mention of this case.

For the truly curious, what's at issue is the following empirical question: are there languages in which epenthesis and assimilation are in a counterbleeding (assimilation before epenthesis) rather than a bleeding (epenthesis before assimilation) order? In his 1998 book, Vaux appears to answer this question in the affirmative, by suggesting a possible counterbleeding analysis of schwa epenthesis and laryngeal assimilation in New Julfa Armenian.

What I point out in my article (something that was also pointed out by Orhan Orgun in his review of Vaux's book) is that there is no independent evidence -- that is, other than that the analysis "works" -- that the relevant schwas are epenthetic, which (one would think!) is crucial to Vaux's argument. I also point out a set of relevant facts from the Urban Utrecht dialect of Dutch (courtesy of Marc van Oostendorp) showing that voicing assimilation applies across (crucially) non-epenthetic schwa. Therefore, my argument goes, Vaux's data are consistent with the other analysis suggested by Vaux -- the one reflected by his formalized statement of the assimilation rule -- that laryngeal assimilation applies between consonants whether or not there is a schwa between them: transparency of assimilation, not opacity of derivation.

Eric said...

That's like being the guy in logic class who can't get a proof to work, but criticizes others for workable proofs that aren't elegant enough.

As just pointed out in my previous comment, the ultimate point is an empirical one. (And, if you're a working phonologist, an important one that happens to have deep theoretical stakes.)

By contrast, Baković (2007:220) quotes McCarthy on this:

Unless further refinements are introduced, OT cannot contend successfully with any non-surface-apparent generalizations, nor with a reside of non-surface-true-generalizations.


It'd be nice if you quoted what I write immediately after that quotation:

I argue that this conclusion is justified only to the extent that overapplication opacity consists solely of interactions that involve what in an OT analysis would be a gratuitous violation of a faithfulness constraint.

I understand that this is an anonymous, mostly non-academic blog. But I don't think that the things I address in my previous two comments were so unclear or inaccessible when I originally wrote them that they justify the level of misunderstanding I'm seeing here. Do I expect too much?

The Stranded Preposition said...

From a very casual, lunch break reading of the responses to Mr. V's answer to Eric's challenge it seems that the OTers and nonOTers are on different planes (with different scientific requirements) and hence pass in the cyber-night.

So, before punching the clock and getting back to work I'm wondering what role old fashioned Elliott Sober-type Simplicity (1975, OUP) is at plays here. On one hand you have a theory with additions (adding a sympathy candidate and local conjunction to the list of existing OT tools and tricks to account--but not predict--for opacity of the post-lexical, in-stratum sort). On the other hand is a theory Mr. V says allegedly predicts post-lexical opacity without any additions (put one rule in front of another).

Cassaday Rassmussen said...

Asbestos underwear on!

So in order...

(1) Anonymous, definitely read Eric's comment above mine because you do not get the full scope of Eric's argument in his article. More on this below.

(b) Ed, I think a relatively neutral measure of 'vastly' different for theories is how much you can carry forward from the 'old' theory. This is where Mr. V has a point in that constraint conjunction is not a trivial modification. Constraint conjunction is not trivial because it makes a major impact on (at least) the theory of CON and language acquisition.

CON is majorly impacted by conjoined constraints because the question of the universality of constraints must now be addressed. Are all conjoined constraints in UG and thus part of CON? Or, does the child have to learn conjoined constraints as part of the language acquisition process (more on this below). Another question about the structure of CON is the typology that is produced from the content of CON. What happens to typological predictions when conjoined constraints are thrown into the mix? I thought OT being typologically based was supposed to be an advantage but noone has kept track of what conjoined constraints do to typological predictions. I think the math is pretty simple there though... typological predictions go to crap with conjoined constraints!

The question of what conjoined constraints do to the theory of language acquisition in OT is another big question. Basically as I see it, all of the work on learnability in OT does not make reference to conjoined constraints. This means that the results from OT learnability studies are lost once you've added conjoined constraints. There's not much one can do about this because either the conjoined constraints are in CON which is not even close to the models of CON used in learnability studies in OT or you have to learn the conjoined constraints and again the learnability studies don't contain that crucial assumption.

A beautiful final question for Ed here. Hypothetically speaking, if we could find two empirically adequate analyses of some phonological phenomena one using conjoined constraints and one using sympathy which would you prefer and why?

(6) I bow down and worship the almighty Mr. V who only speaks gutter phonology.

(a) Back to Ed, see comments above and the 'beef' with OT is not that P&S 93 did not provide a 'theory of everything' but that if the core aspect of OT is only 'ranked and violable constraints' then its pretty close to a 'theory of nothing' (note connection between everything and nothing). The original rolling out of OT proclaimed all of these advantages that it had over whatever model was present in the early 90's (again I'm confused as to whether this was SPE or Lexical Phonology or some love child in between...). Typological predictions, parallelism, learnability results, serial derivation couldn't work, etc. Now that its 15 years later, people are wondering where these advantages are but the common response that I see is
'oh, yeah, we actually do need strata',
'oh, yeah, we actually do need an independent theory of morphology',
'oh, yeah, there does appear to be some connection between constraint and repair'
If everything about phonology (and its interfaces) is 'up for grabs' except for ranked violable constraints I don't really see what the advantage of OT is...

(=) On to my friend Eric, as you can see 'anon' at the top presents an ad hominem attack on you which I don't believe I have. There is a contrast here...

You say tomato and Vaux says tomahto. Armenenian says opacity and Dutch says transparency. Do Armenian and Dutch have to be the same? Maybe you're right about Armenian, maybe you're wrong about Dutch. I don't really see either you nor Vaux presenting hard evidence that either analysis must be the way that it is. Your right that there are likely (actually probably definitely) cases where schwa is underspecified (is that ok in OT?) and thus assimilation can pass through it. This doesn't make your argument go through though. You have to show that the ordering analysis is IMPOSSIBLE which you haven't done. As long as Vaux's analysis is possible, the typological claim that there are languages where epenthesis and assimilation are in a counterbleeding stands.

Not all answers about phonology can be settled in theoretical pissing matches. We probably need to get some LabPhon folk or brain imagers in here to settle this...

Love
Cassaday

Ed Keer said...

Cassaday,

I disagree that conjoined constraints play havoc with CON. First, the question of universality is simply answered--conjoined constraints are universal. What's the problem? To admit conjoined constraints isn't saying that there is a method of constructing constraints on the fly. It's simply saying that there is a class of constraints that have this property that they behave like conjunctions of independent constraint. Each conjoined constraint still needs to be justified on it's own. Which brings up typology...

Second, admitting conjoined constraints as part of CON doesn't affect typology any more than admitting *NC. Obviously, the theorist needs to do the work and show how the proposed constraint predicts typologies. Beware of thinking that simply having a large set of constraints means that you will have a large number of predicted languages. It's pretty easy to show that the number of predicted languages from a constraint set is smaller than the n! of predicted constraint rankings. Many constraint rankings are indistinguishable wrt the input-output mappings they produce.

As for the learnability argument, that only holds if conjoined constraints are markedly different from garden variety markedness constraints. I don't see that they are. Do you mean that every time someone proposes a new constraint all the learnability results based on previous constraint sets are lost? That's odd, especially since the learnability results are usually based on a small subset of all the proposed constraints in the literature.

And to answer your question--I have no idea.

Ed Keer said...

If only there was some sort of searchable database of OT papers, where we could check claims like "but noone has kept track of what conjoined constraints do to typological predictions." Do de do de do. Oh what's this?

Typological Consequences of Local Constraint Conjunction
Phonological opacity has become the subject of several competing analyses in both rule- and constraint-based phonology. One, local constraint conjunction in Optimality Theory, is shown here to predict a typology of possible and impossible synchronic chain shifts involving epenthesis and deletion. The predictions, which do not follow from the other analyses, appear to be borne out by language data. In local conjunction, two Optimality-Theoretic constraints C and C' are combined into a new constraint (C & C')D which is violated if there is a representational domain D within which both C and C' are violated. This mechanism has been used to account for two kinds of opacity: chain shifts (Kirchner 1996) and derived-environment effects (Lubowicz 2002). This paper is about the chain shifts. In order for C and C' to be conjoined, there must be a common domain D within which both can be violated. In the framework of Correspondence Theory (McCarthy & Prince 1995), certain constraint families inherently cannot share domains, and hence cannot be locally conjoined. The phonological processes corresponding to the impossible conjunctions are predicted not to occur. Segmental DEP constraints are violated by a surface segment with no underlying correspondent; segmental MAX constraints, by an underlying segment with no surface correspondent. Therefore, no domain can contain both a DEP and a MAX violation, and the conjunction (DEP & MAX) is ruled out. It is shown that this entails the impossiblity of chain shifts of the the form AxB->AB-> AzB. A review of 35 chain-shift cases to date has found all of them to conform. An apparent counterexample (Donegan & Stampe 1979) is examined and refuted. The typological pattern is taken to support the local-conjunction account, which is the only one which predicts it.
WTF!

Eric said...

My good friend Ed appears to be on to something here: checking on claims before making them. Imagine that.

On to my friend Eric, as you can see 'anon' at the top presents an ad hominem attack on you which I don't believe I have. There is a contrast here...

Indeed there is. I attended my anger management class today, hence the unexpected response.

I don't really see either you nor Vaux presenting hard evidence that either analysis must be the way that it is.

I agree -- in fact, that was all I was saying (in my article and in my comment above).

You have to show that the ordering analysis is IMPOSSIBLE which you haven't done. As long as Vaux's analysis is possible, the typological claim that there are languages where epenthesis and assimilation are in a counterbleeding stands.

I think the claim is severely weakened by (a) the alternative, transparent-schwa analysis that Vaux himself suggests, and (b) the existence of at least one other language in which the transparent-schwa analysis must be right (because the counterbleeding analysis can be independently shown to be unavailable). But hey, it's just tomato vs. tomahto, right?

In 2003, Vaux gave a talk at the LSA ("Why the Phonological Component must be Serial and Rule-Based", recently published in the book Vaux co-edited with Andrew Nevins; haven't seen the pub'd version yet) in which he claimed that opacity in rule-based serialism is easy to learn: each rule is learned independently, and their opaque interaction is used as evidence for their ordering. If that's true, then the Vaux's suggested counterbleeding analysis Armenian case is instructive: one of the rules is not independently learned. (I've found several cases in the literature of proposed opaque rule orderings with no independent motivation for one of the rules, though -- in fact, Vaux & Nevins presented such an analysis of Pig Latin the very next day at the 2003 LSA -- so maybe the relevant claim is just false.)

Your right that there are likely (actually probably definitely) cases where schwa is underspecified (is that ok in OT?) and thus assimilation can pass through it.

I never said that schwa is underspecified (and neither did Vaux, incidentally). There are many ways to analyze transparency to assimilation, no matter what overarching framework you subscribe to; underspecification is but one of those ways.

But besides: why would underspecification (a matter for the theory of representations) not be "ok in OT" (a theory of constraint interaction)? It's true that some (certainly not all) work within OT has aimed at ridding the theory of representations of underspecification, but this is a reflection of the trend already beginning in work from the early '90s that nothing to do with OT: Mohanan 1991, McCarthy & Taub 1992, Steriade 1995, and others.

(A similar confusion between the theory of representations and the theory of rule/constraint interactions was noted in the second-to-last paragraph -- and the associated footnote -- here.)

Perhaps I can appeal to you this way: was rule-based serialism fundamentally changed when non-linear representations were introduced in the mid-to-late '70s? In other words, is it reasonable to ask "are syllables ok to use in rule-based serialism?" While it's true that there was a flurry of work in the late '70s and early '80s attempting to reduce rules to representations, I think it's safe to say that the effort failed (or more diplomatically, failed to be widely accepted) and that rules and non-linear representations were co-existing peacefully throughout the rest of the '80s.

Not all answers about phonology can be settled in theoretical pissing matches. We probably need to get some LabPhon folk or brain imagers in here to settle this...

Someone once wrote to me: "The Lab Phono people are not working on showing that 'badu' is underlyingly /badw/. This doesn’t entail that opaque analyses are wrong. But it does show that knock-down evidence of the Copernican sort has not yet, shall we say, emerged from scrutiny of the facts".

Eric said...

I also have to respond to two comments directed at my friend Ed:

A beautiful final question for Ed here. Hypothetically speaking, if we could find two empirically adequate analyses of some phonological phenomena one using conjoined constraints and one using sympathy which would you prefer and why?

A beautiful final question for Cassaday here. Hypothetically speaking, if we could find two empirically adequate analyses of some phonological phenomena one using the Elsewhere Condition and one using the Strict Cycle Condition which would you prefer and why?

The original rolling out of OT proclaimed all of these advantages that it had over whatever model was present in the early 90's (again I'm confused as to whether this was SPE or Lexical Phonology or some love child in between...).

I would really like to see a specific piece of work (preferably by some reasonably influential phonologist, but almost anything would do) in which OT is proposed to supplant the entire edifice of phonological theory from 1968 to 1993. The only ideas specifically attacked are (i) that phonological processes result from string rewrite rules and (ii) that those rewrite rules are crucially ordered (within a component; that is, extrinsically). Just as non-linear representations and a Lexical-Phonological architecture are compatible with (i) and (ii), so are they compatible with the OT view that phonological processes result from the interaction of ranked and violable constraints.

There was never an explicit denial of evidence for strata (though there were attempts to do things without strata), nor was there ever explicit denial of the need for an independent morphological component. Again, if you can provide evidence of such explicit denials, I'd like to see them.

I'm sorry that you view the introduction of OT as so revolutionary (by which I mean 'an attempt at a revolution'), because to me and many others it was nothing of the sort. It was just progress along a particular line of thinking that already existed (recall Harmonic Phonology, Declarative Phonology, and so on). You can disagree with whether it was worth pursuing, of course, but crying foul because some elements of pre-existing, compatible theories stuck around just isn't getting us anywhere in this discussion.

Bill Idsardi said...

Two brief points:
1. Self-conjoined constraints add significantly to the computational complexity of OT; this is the essence of the proof in Idsardi 2006. Nobody has any idea what the computational complexity of (multiple) sympathies is.
2. Sober has given up his 1975 general theory of simplicity in favor of only domain-specific parsimony, see his "Let's Razor Ockham's Razor" in particular. A good discussion of this can be found in the Stanford Encyclopedia of Philosophy entry: http://plato.stanford.edu/entries/simplicity/.