Please place your comments and questions on Ch. 1 of McCarthy’s text as comments to this post, by the end of the day Sunday, Sept. 9th.
Please place your comments and questions on Ch. 1 of McCarthy’s text as comments to this post, by the end of the day Sunday, Sept. 9th.
I am bothered by the infinite nature of OT’s GEN component. Recognizing a distinction between theory and implementation, what are some ways that computational work deals with the unrestrictedness of GEN and its operations? For example, how does one limit the size and contents of the candidate set before it is submitted to EVAL?
I think that for an OT newbie the chapter is rather easy to follow. However, I share the concern about the free application of GEN. OT is a theory about constraint ranking, but is it true that we don’t want the theory to restrict GEN?
This reminds me the discussion about the nature of transformations in syntax. Very roughly, in syntax we say that “Merge” applies freely, just as GEN does, and that grammatical derivations are favored over ungrammatical derivations by independent output conditions. However, if we take the more or less standard view that merge is optional and it is only “triggered” by feature checking, then it is not strictly true that “merge” is able to combine any two possible forms. So, what about GEN? Isn’t there a way to restrict its application along these lines?
I found McCarthy’s presentation of the conspiracy problem in Yawelmani beneficial because it succinctly illustrates the descriptive power of violable constraints and OT. As an undergraduate, with my haphazard introduction to OT, I was never entirely sure about OT’s descriptive power over SPE, but, well, now I am.
However, I would still like a better understanding of OT in the context of historical linguistic change. Does change in the ranking of constraints and reanalysis of underlying forms account for everything?
Like Sean, I am also interested in how computational implementations of OT handle GEN. Clearly, GEN cannot generate an infinite set of candidates because humans lack such mental capacity. However, any limit on GEN’s ability to produce candidates seems to be a stipulation, which OT shies away from.
In the box spanning pages 6 and 7, McCarthy provides some background about the way that OT adopted many of the assumptions of SPE. I note that he says ‘The underlying representation includes all of the unpredictable phonological properties of a morpheme,’ but does not state whether it does/should/can/cannot include more. So he stops just short of mentioning underspecification or motivation(s) for it.
As I have heard it in the past, the major original motivation for underspecification came from linguists’ awareness of the state of early computers. Memory was not readily available, so anything that could be programmed was preferred over anything that required more data storage. But now that we have progressed to thinking of memory as cheap and plentiful, is there still compelling motivation for storing only unpredictable information? Or is there a sense in which storing predictable information is more compatible with a simple/standard view of OT and how learning would work under it?
I think that Optimality Theory is actually a pretty brilliant theory. The ranking of constraints and the concept of minimal violation give a lot of credit to the theory itself because they allow us to account for not just one language or two, but almost all possible human languages, given the universal properties of the constraints. To be more specific, take stress for example, adopting rewrite rules or parametric methods to explain stress systems may suffice in most cases. But one could also run into situations where some rewrite rules that work in some places no longer work in the current situations. Then, more rules have to be created to fill the gap. Wouldn’t that be a threat to the elegant and economic aspects of the theory itself, theory that incorporates rewrite rules? Like the case of “Canada†we discussed in class, an extra step of applying extrametricality to account for the shifting of stress could be a downside in terms of economy, as Yohei mentioned, and Occam’s razor also points out the same idea.
We want the theory to be as general as possible and able to be applied to as many languages as possible. OT captures that idea under the assumption that the constraints are all universal and the differences in every language are simply a matter of constraint ranking. The violability of constraints gives more room for explaining languages that use different stress systems or languages that do not even adopt stress, even the stress variations within a language. It is just the result of different constraint dominations in different languages. That, I think, is the most general yet powerful aspect of OT. What I really have a question about is how these universal constraints are formed. Is it that they are made after the observation on many languages, followed by a conclusion that the languages all share the constraints, or that we just make the constraints based on one language, assuming that they are universal, and the reason why some languages violate them is simply because those languages have to abide by other higher ranking constraints?
I’d also like to say something about the question raised by Joe and answered by Amanda. The question was, “Besides the five stress patterns that are attested in human languages, what are some other logically possible stress patterns?†Amanda’s answer was, “Given any odd-syllable word, the syllable in the middle is stressed.†As I was doing the reading for my syntax class, I came across a putative rule that was used to describe English interrogative formation. And it suddenly hit me that this actually had a lot in common with the logically possible but practically unattested stress pattern mentioned. The putative rule is as follows:
In a pair of related sentences as in (1), the question is formed by moving the second word to the sentence-initial position.
(1) a. John is home.
b. Is John home?
However, as one can easily tell, this rule will generate ungrammatical sentences in plenty of other situations:
(2) a. The man with a hat is home.
b.*Man the with a hat is home?
Therefore, the rule cannot go by simply counting the words. One has to know which words make a constituent before deciding which should be moved where. In the case of (2), “the man with a hat†makes a constituent, so it is the word that follows this constituent should be moved:
(3) Is the man with a hat home?
This rule accounting for this fact is said to be structure-dependent. This is actually my point. It is all about structure. If we apply this thought of structure to the logically possible stress pattern above, we can probably tell why it is unattested in human languages. We cannot simply count the number of syllables and decide that the one in the middle should be stressed. It would end up being the same situation as that of the putative interrogative-formation rule. To account for where the stress should land, we need structure. I believe that is why foot and directionality of its placement are introduced to do the job. They are a kind of structure. Now we have OT as another kind of structure. Which kind of structure works better in solving phonological problems is one thing. There should always be structure, regardless. And I guess, being relatively distinct subjects, phonology and syntax having structure in common probably gives us a little idea of how our brain works in terms of language.
I post one further question, this one more practical. It concerns example (4) in the first chapter of McCarthy’s book. There is a chart with all the possible outputs for the input /xat-ka/. But, if the rules of the transformational component are all optional and there is no priority with respect to which rule applies firs, aren’t there missing two possible outputs? I mean, wouldn’t [xa.ti.ka] and [xa.ti.k] be also possible outputs? ([xa.ti.ka] by syllabification epenthesis plus final vowel deletion; [xa.ti.k] by syllabification epenthesis plus final vowel deletion).
I have a question about syllable structure and how it interacts with OT. Yewelmani is a language that favors a CVC structure, however throughout the chapter we’ve seen a variety of syllabifications. For instance, [ta.xak], the optimal candidate to the underlying form /taxa:-k?a/ can be syllibified in two ways: [ta.xak] and [tax.ak]. Why the preference for the for variant, over the second? Similarly, why [li.him.hin] over [lih.in.hin]? There isn’t a rule in the SPE interpretation or a constraint in the OT that would favor a CV.(CVC) formation over a CVC.CV.CVC or a CVC.CV formation. Although having looking at the previous sentence, I can see that even by description that CV.(CVC) variant appears more elegant. So perhaps this constraint would be a helpful addition to our interpretation of the Yewelmani data we’ve been presented with through the chapter.
Another instance that puzzled me was the discussion of filter model as demonstrated in table (4). The argument was that the filter model is not sufficient because it does not determine the interaction between the rules within the transformational component before its outputs encounter the filter component. Table (4) shows [xa.tik?] as the ultimate output, while noting that this is also a wrong output, making the filters model an inaccurate description of the language. However, if the transformational component is still described in terms of rules, I do not see that [xa.tik?] would be the ultimate output. The epenthesis rule is described as 0 –> i / C_CC, where an /i/ is epenthesized after the first consonant of a three-letter cluster consonant. In my understanding, if the underlying form /xat-k?a/ had undergone deletion, the epenthesis rule would not apply, because [xat.k?] does not have a three-letter consonant cluster described by the epenthesis rule.
I realized that Optimality Theory (OT) is an important achievement within the generative tradition especially in that it focuses on explanatory power (e.g. economy, simplicity) without losing empirical coverage. And interestingly, as pointed out in the textbook, OT has many significant analogies with the history of the generative syntax. So I would like to add some other parallelisms and differences with/from the development of syntactic theory.
In the first generation (1950s-1960s), both syntax and phonology crucially depend on the rewriting rule systems (A→B/C_D) with an optional/obligatory distinction or cyclicity (Chomsky 1957, 1965). As much research proceeds, however, it is revealed that the rule system doesn’t work because of its empirical failure (e.g. passive transformation) and theoretical complexities caused by explosion of language-/construction- specific rules. The same is true of the phonological enterprise according to the introduction.
Next, in 1970s-1980s, theories change from rule-based to constraint-based. Specifically, the Principles and Parameters (P&P) approach emerges in the syntactic research, so that universal constraints and some parametric devices are proposed (Chomsky 1981, 1986). For example, the huge list of transformation is reduced to the “Move α†operation, and universal constraints (e.g. Empty Category Principle) appear. Phonology and syntax are different in that phonological operations seem not to be reduced to the unique one, but they basically share universality of constraints. Furthermore, parameter setting in the P&P approach is very similar to ranking of constraints (violable/inviolable) in OT.
Finally, the current syntactic research is conducted under the working hypothesis: “minimalismâ€. From this standpoint, some syntacticians cast doubt on the existence of parameters because they are highly stipulative and cannot be explained furthermore (In a given parameter, some languages are ON, other languages are OFF. That’s it.). Given this, the same question should also rise in phonology; that is, why A>>B in some languages, why B>>A in other languages? Can we derive parametric differences among languages in a non-arbitrary way? The attempt to answer this question will leads to “minimalist phonologyâ€.
I think that phonology and syntax are closely interrelated, and in fact syntax owes the important notion “cyclicity†to generative phonology (Chomsky and Halle 1968). Therefore, central concepts in OT (violability, ranking of constraints) should also be useful in the syntactic research. I would like to learn the details of conception in generative phonology, especially OT, to contribute to both phonology and syntax.