Please post a comment on McCarthy (2007) “Derivations and levels of representation” by noon, Wednesday, October 17th.
7 thoughts on “Comments on McCarthy 2007; due 10/17 noon”
Sasha Raikhlina
This paper examines a potential failure of OT in the case of opacity. McCarthy presents the elegant SPE approach to opacity, ground in rule ordering and series of non-designated levels of analysis. While Stratal OT provides an elegant solution to some instances of opacity, it requires further modification. The assumption that the Stratal OT approach only operates within two stratum prevents us from analyzing a language such as Bedouin Arabic in which deletion precedes raising but follows palatalization. And set of operations: PALATALIZE then DELETE then RAISE presupposes 3 levels of analysis, whether in OT or another framework.
For Bedouin Arabic as seen through the examples featured in today’s handouts, the two strata analysis is possible only if the processes of palatalization and deletion are collapsed into a single stratum. So, wrecking the beautiful analysis in (12), we can posit a constraint *K that will rule out candidates (b) and (c) and have *NUC/[HI] rule out (a), in which case the winner [ha:kjmi:n] would surface as an optimum within just one grammar. Yet this weak attempt at alternative analysis fails in the case of /thakumin/ –> [thakmin] and can maybe be patched up with a stronger IDENT constraint but this would be obviously messy.
Constraint conjunction is really the only thing that made sense as a tool for tackling opacity within OT. McCarthy says that “the problem with local conjunction is that it rules out the cooccurrence of unfaithful mappings in close
proximity”. I am not sure what he means by “cooccurrence of unfaithful mappings in close proximity” but I take it that the problem with constraint conjunction is that it does not have a localization parameter, i.e. it can rule out candidates where the constraint violations are noncontiguous and really don’t interact in any way. Am I correct in this assumption? I would like to see constraint conjunction explained in more depth.
On page 108, McCarthy writes that counterfeeding and counterbleeding are the only evidence that support a language particular rule ordering. McCarthy also debunks two offers arguments in favor of sequential rule ordering on page 106. (I particularly liked the argument he made about generation X+1’s grammar). Rather than focusing on simple rule ordering, he suggests we follow the Stratal OT method, which would have different levels (strata). Each stratum has its own constraint rankings. The output of the initial stratum becomes the input of the next. I like that Stratal OT does not simply order rules, but also explains the derivations at the stratal level.
My only question after reading this is related to the topic he brought up on page 108. He wrote that analyses such as SPE’s /rixt/ – [ra:jt] (right) are dubious analyses based on excessively abstract underlying forms, however these same types of analyses can be made using Stratal OT. How can we determine if an underlying form is excessively abstract? and how do we restrict the use of such forms?
The derivational phonology approaches presented on this article seem to reflect in a clearer way how we move from the underlying to the surface representation, following a sequence of rules that apply in a specific order in each language. Considering the relations of feeding, bleeding and their counterparts, we can predict the order in which rules applied. Furthermore, it helps the fact that, whereas OT considers only two levels of representation, other approaches (Harmonic Phonology or Cognitive Phonology) account for at least three levels, the morphophonemic (M), the closest one to the surface level, the phonetic level (P), more related to the surface level, and an intermediate level (W) or even more than three, as it is the case for Stratal OT. The general idea, however, is that at least three levels of representation are required to show how different grammars apply in each level of representation. This explains why sometimes a word can be produced in a different way, depending on the context in which it appears, as in the case of the French “liaison†commented in previous classes, or in the case of derived words, that follow a pattern different from the expected one if rules had applied only from one level directly to another, without considering an intermediate level that constitutes the input for the derived word. To conclude, I would like to say that I really appreciate the fact that the start point for the overview of the different levels of representation considered by each theory is structuralist phonology. It helped me at seeing the connection between what we are doing and what I did before I arrived in the States.
Has anyone ever conducted a typological survey of opacity? Specifically, has anyone ever tried to find commonalities among the opaque interactions that are describable in any given system (e.g. HS, Stratal OT) and those interactions that are not?
With respect to formulating constraints for the purpose of dealing with opacity, has anyone ever explored the typological predictions of formulating markedness constraints that only target originally underlying segments (as with originally-present /i/’s in the Bedouin Arabic example)? Given some approaches to faithfulness that make reference to differences between root segments and affixal ones, it seems that there is some precedent for distinguishing among segments on the basis of their origin.
It is clear that, if we want to deal with opacity within OT, we need to enrich the system by some means. SOT seems to work great to some extent, but it will not always suffice to do the work. Furthermore, I am concerned with the fact that, in a change from OT to SOT, we expect that the same typology predicted by a set of constraints X in OT be a subset of X predicted by SOT, if only because the possible arrangement of constraints is increased. The logical problem I see to this respect is that the system is in fact predicting more forms while it is still not able to explain some of the attested cases (see the discussion on page 111).
Related to this, I wonder whether anyone has tried recasting a “SOT-like” effect in HS. If HS applies one constraint every time, what precludes a given constraint from applying more than once? (Please, be benevolent, as I know close to nothing about HS and this is just a mere speculation). That is, we could define “environment specific” constraints that can only be applied if some conditions are met. For example, defining some CON-A’ that can only exist if CON-B was applied after CON-A, where CON-A’ = CON-A but for the fact that the former is applied first. This looks like something that a finite state automata could not do, but, would that be a problem? To this respect, opacity will definitely be linked to some kind of cyclicity, as McCarthy suggests towards the end of the paper. I am afraid, however, that there is a high risk of defining CON-A’ in terms of rules, rather than in purely OT terms.
In order to solve the problem of phonological opacity, many phonological theories have proposed different levels of reoresentations, where different rules apply. The original framework of OT lacks this respect. That is why Stratal OT came along. By deviding the mapping process between underlying forms and surface forms into two different levels of candidate evaluation, it successfully solves the problem of opacity as other theories. Everything seems perfect now. However, if compared to the assumptions in the original OT framework, Stratal OT seems to have some discrepancies that kind of puzzle me.
In illustration, take Bedouin Arabic for example, Stratal OT solves the opacity of palatalization of velars by putting the application of palatalization in the first stratum, which is the word level of representation, and the application of deletion in the second stratum, which is the phrase level. The constraints are ranked differently in these two strata for the phonological processes to apply. How, then, can we tell which ranking of the constraints, or maybe both, is really the phonological grammar of the language? When we were discussing the unpredictable stress pattern and the underlying stress in class, the underlying form of the candidates was assumed to have stress, and a new constraint, Faith-Stress, was introduced to account for the unpredictable stress. I asked Joe a question about why we can’t use different rankings within the same language to account for the stress pattern but have to come up with a new constraint that was constructed by looking back from the output to the input to assume some underlying stress.
The answer I was given, if I remember correctly, was that different rankings within the same language only happen in cases where the input has more than on outputs. Can we say that is the situation in Stratal OT here, one input has two different outputs in two differet strata? If not, how can we explain the assumption that the ranking of the constraints is the grammar of the language when there are two rankings at hand?
　McCarthy (2007) is a nice summary paper which thoroughly reviews previous literatures on phonological representations and derivations, “opacity†problem and several types of cyclicity. Specifically, it systematically presents crucial data, theory outlines and their predictions, empirical challenges and conceptual advantages. Here are my questions and comments:
(i) In section 5.3, I realized that feeding/bleeding relationships in the traditional SPE system can be readily recaptured in terms of markedness/faithfulness constraints proposed within OT, with unnecessary intermediate levels eliminated, hence a theoretically desirable direction. At this point, I’m just wondering whether when the complexity of feeding/bleeding relations highly increases (e.g. feeding rule X → rule Y → potentially fed rule Z), translations with constraints set still work or not.
(ii) In the end of section 5.3, John McCarthy casts doubt on two conceptual arguments in favor of sequential rule application: mental computation and diachronic change. As for the former, he convincingly argues that since the goal of generative linguistics is to explore linguistic “competenceâ€, not “performanceâ€, it is not unnatural that phonological theories do not mirror actual mental computation in human brain/mind. In the latter, it is claimed that historical sound changes cannot be regarded as addition of a new rule to the given ordering, because of the famous words: poverty of stimulus (primary linguistic data). My question here is ‘How does OT deal with these aspects of languages, especially phonology?’ For example, computational implementations of OT constraint ranking are discussed in the textbook (McCarthy 2008: Ch.2), as well as learning perspectives. So if these implementations did not reflect mental computation at all, where are these computational aspects located within OT? How does OT resolve the problem of poverty of stimulus/absence of positive evidence (cf. Yang 2000)?
(iii) In section 5.5, three types of OT-implementation of cyclicity are introduced: asymmetrical, symmetrical, and alignment constraints. From the computational perspective, I believe that the first ‘asymmetrical’ approach (so-called “Strict Cyclicityâ€) is most promising. In other words, computational burden/load will considerably reduce if working space/memory can forget all information after calculation at the first cycle has completed and that cycle becomes invisible to the next cycle. This perspective is also supported by virtue of the recent syntactic theory (Chomsky 1995 et seq.), especially computational efficiency in syntactic derivations.
This paper examines a potential failure of OT in the case of opacity. McCarthy presents the elegant SPE approach to opacity, ground in rule ordering and series of non-designated levels of analysis. While Stratal OT provides an elegant solution to some instances of opacity, it requires further modification. The assumption that the Stratal OT approach only operates within two stratum prevents us from analyzing a language such as Bedouin Arabic in which deletion precedes raising but follows palatalization. And set of operations: PALATALIZE then DELETE then RAISE presupposes 3 levels of analysis, whether in OT or another framework.
For Bedouin Arabic as seen through the examples featured in today’s handouts, the two strata analysis is possible only if the processes of palatalization and deletion are collapsed into a single stratum. So, wrecking the beautiful analysis in (12), we can posit a constraint *K that will rule out candidates (b) and (c) and have *NUC/[HI] rule out (a), in which case the winner [ha:kjmi:n] would surface as an optimum within just one grammar. Yet this weak attempt at alternative analysis fails in the case of /thakumin/ –> [thakmin] and can maybe be patched up with a stronger IDENT constraint but this would be obviously messy.
Constraint conjunction is really the only thing that made sense as a tool for tackling opacity within OT. McCarthy says that “the problem with local conjunction is that it rules out the cooccurrence of unfaithful mappings in close
proximity”. I am not sure what he means by “cooccurrence of unfaithful mappings in close proximity” but I take it that the problem with constraint conjunction is that it does not have a localization parameter, i.e. it can rule out candidates where the constraint violations are noncontiguous and really don’t interact in any way. Am I correct in this assumption? I would like to see constraint conjunction explained in more depth.
On page 108, McCarthy writes that counterfeeding and counterbleeding are the only evidence that support a language particular rule ordering. McCarthy also debunks two offers arguments in favor of sequential rule ordering on page 106. (I particularly liked the argument he made about generation X+1’s grammar). Rather than focusing on simple rule ordering, he suggests we follow the Stratal OT method, which would have different levels (strata). Each stratum has its own constraint rankings. The output of the initial stratum becomes the input of the next. I like that Stratal OT does not simply order rules, but also explains the derivations at the stratal level.
My only question after reading this is related to the topic he brought up on page 108. He wrote that analyses such as SPE’s /rixt/ – [ra:jt] (right) are dubious analyses based on excessively abstract underlying forms, however these same types of analyses can be made using Stratal OT. How can we determine if an underlying form is excessively abstract? and how do we restrict the use of such forms?
The derivational phonology approaches presented on this article seem to reflect in a clearer way how we move from the underlying to the surface representation, following a sequence of rules that apply in a specific order in each language. Considering the relations of feeding, bleeding and their counterparts, we can predict the order in which rules applied. Furthermore, it helps the fact that, whereas OT considers only two levels of representation, other approaches (Harmonic Phonology or Cognitive Phonology) account for at least three levels, the morphophonemic (M), the closest one to the surface level, the phonetic level (P), more related to the surface level, and an intermediate level (W) or even more than three, as it is the case for Stratal OT. The general idea, however, is that at least three levels of representation are required to show how different grammars apply in each level of representation. This explains why sometimes a word can be produced in a different way, depending on the context in which it appears, as in the case of the French “liaison†commented in previous classes, or in the case of derived words, that follow a pattern different from the expected one if rules had applied only from one level directly to another, without considering an intermediate level that constitutes the input for the derived word. To conclude, I would like to say that I really appreciate the fact that the start point for the overview of the different levels of representation considered by each theory is structuralist phonology. It helped me at seeing the connection between what we are doing and what I did before I arrived in the States.
Has anyone ever conducted a typological survey of opacity? Specifically, has anyone ever tried to find commonalities among the opaque interactions that are describable in any given system (e.g. HS, Stratal OT) and those interactions that are not?
With respect to formulating constraints for the purpose of dealing with opacity, has anyone ever explored the typological predictions of formulating markedness constraints that only target originally underlying segments (as with originally-present /i/’s in the Bedouin Arabic example)? Given some approaches to faithfulness that make reference to differences between root segments and affixal ones, it seems that there is some precedent for distinguishing among segments on the basis of their origin.
It is clear that, if we want to deal with opacity within OT, we need to enrich the system by some means. SOT seems to work great to some extent, but it will not always suffice to do the work. Furthermore, I am concerned with the fact that, in a change from OT to SOT, we expect that the same typology predicted by a set of constraints X in OT be a subset of X predicted by SOT, if only because the possible arrangement of constraints is increased. The logical problem I see to this respect is that the system is in fact predicting more forms while it is still not able to explain some of the attested cases (see the discussion on page 111).
Related to this, I wonder whether anyone has tried recasting a “SOT-like” effect in HS. If HS applies one constraint every time, what precludes a given constraint from applying more than once? (Please, be benevolent, as I know close to nothing about HS and this is just a mere speculation). That is, we could define “environment specific” constraints that can only be applied if some conditions are met. For example, defining some CON-A’ that can only exist if CON-B was applied after CON-A, where CON-A’ = CON-A but for the fact that the former is applied first. This looks like something that a finite state automata could not do, but, would that be a problem? To this respect, opacity will definitely be linked to some kind of cyclicity, as McCarthy suggests towards the end of the paper. I am afraid, however, that there is a high risk of defining CON-A’ in terms of rules, rather than in purely OT terms.
In order to solve the problem of phonological opacity, many phonological theories have proposed different levels of reoresentations, where different rules apply. The original framework of OT lacks this respect. That is why Stratal OT came along. By deviding the mapping process between underlying forms and surface forms into two different levels of candidate evaluation, it successfully solves the problem of opacity as other theories. Everything seems perfect now. However, if compared to the assumptions in the original OT framework, Stratal OT seems to have some discrepancies that kind of puzzle me.
In illustration, take Bedouin Arabic for example, Stratal OT solves the opacity of palatalization of velars by putting the application of palatalization in the first stratum, which is the word level of representation, and the application of deletion in the second stratum, which is the phrase level. The constraints are ranked differently in these two strata for the phonological processes to apply. How, then, can we tell which ranking of the constraints, or maybe both, is really the phonological grammar of the language? When we were discussing the unpredictable stress pattern and the underlying stress in class, the underlying form of the candidates was assumed to have stress, and a new constraint, Faith-Stress, was introduced to account for the unpredictable stress. I asked Joe a question about why we can’t use different rankings within the same language to account for the stress pattern but have to come up with a new constraint that was constructed by looking back from the output to the input to assume some underlying stress.
The answer I was given, if I remember correctly, was that different rankings within the same language only happen in cases where the input has more than on outputs. Can we say that is the situation in Stratal OT here, one input has two different outputs in two differet strata? If not, how can we explain the assumption that the ranking of the constraints is the grammar of the language when there are two rankings at hand?
　McCarthy (2007) is a nice summary paper which thoroughly reviews previous literatures on phonological representations and derivations, “opacity†problem and several types of cyclicity. Specifically, it systematically presents crucial data, theory outlines and their predictions, empirical challenges and conceptual advantages. Here are my questions and comments:
(i) In section 5.3, I realized that feeding/bleeding relationships in the traditional SPE system can be readily recaptured in terms of markedness/faithfulness constraints proposed within OT, with unnecessary intermediate levels eliminated, hence a theoretically desirable direction. At this point, I’m just wondering whether when the complexity of feeding/bleeding relations highly increases (e.g. feeding rule X → rule Y → potentially fed rule Z), translations with constraints set still work or not.
(ii) In the end of section 5.3, John McCarthy casts doubt on two conceptual arguments in favor of sequential rule application: mental computation and diachronic change. As for the former, he convincingly argues that since the goal of generative linguistics is to explore linguistic “competenceâ€, not “performanceâ€, it is not unnatural that phonological theories do not mirror actual mental computation in human brain/mind. In the latter, it is claimed that historical sound changes cannot be regarded as addition of a new rule to the given ordering, because of the famous words: poverty of stimulus (primary linguistic data). My question here is ‘How does OT deal with these aspects of languages, especially phonology?’ For example, computational implementations of OT constraint ranking are discussed in the textbook (McCarthy 2008: Ch.2), as well as learning perspectives. So if these implementations did not reflect mental computation at all, where are these computational aspects located within OT? How does OT resolve the problem of poverty of stimulus/absence of positive evidence (cf. Yang 2000)?
(iii) In section 5.5, three types of OT-implementation of cyclicity are introduced: asymmetrical, symmetrical, and alignment constraints. From the computational perspective, I believe that the first ‘asymmetrical’ approach (so-called “Strict Cyclicityâ€) is most promising. In other words, computational burden/load will considerably reduce if working space/memory can forget all information after calculation at the first cycle has completed and that cycle becomes invisible to the next cycle. This perspective is also supported by virtue of the recent syntactic theory (Chomsky 1995 et seq.), especially computational efficiency in syntactic derivations.