Is it just me, or does the “phonological filter” that Albright and Hayes use to rule out forms like [nid+d] look a bit like sweeping morphophonological dust under the carpet?
First, I find it a little odd that Albright and Hayes need a morphology component that consists of a set of probabilistically reliable rules, and a phonology component that consists of (what appears to be) a set of exceptionless, categorical constraints/rules.
Second, does this partitioning hold up typologically? Or do there exist examples of apparent cyclicity/parallelism that would require the phonological filter to apply before/at the same time as the morphological component?
Third, if we assume that morphology works the way Hayes and Albright says it does, but the phonological filter is still OT-like, then do speakers ever learn phonological rankings that can only be established by morphological alternation?
For example, consider a language that allows adjacent consonants with opposing [voice] specifications root-internally. Imagine that plurals in this language are like a completely regular (simplified) version of English — roots that end in a voiceless sound take /s/, while roots that end in a voiced sound take /z/. In OT, this alternation would be produced by the ranking:
However, because the realization of the plural is completely regular in this language, we don’t ever need to establish this ranking in the phonological filter, because Albright and Hayes’ morphology component will take care of it. We’ll just posit one morphological rule that realizes the plural as /s/ after voiceless sounds, and one that realizes it as /z/ after voiced sounds, and we’ll eventually learn that those rules each have 100% reliability, since there are no competing forms. As a result, the three constraints given above remain unranked with respect to each other in the phonologicla filter, and there’s no consequence to the learner, because what s/he knows about morphology will get the right result every time.
It would be interesting to find a language like this, and then to test whether something that can be learned as a 100% reliable morphological rule behaves more like a morphological rule, or a phonological rule. For example, we could teach spekaer of this language a novel inflectional suffix for a morphological feature that their own language doesn’t realize as a suffix, give it the same forms and distribution as the plural, and see how fast they learn it.
If they pick up on it right away, then they must have phonologized the right ranking base on the native plural forms– but why would they have done that, if the morphology already gave them the correct result?
If it takes time for them to get the distribution right (or they never do), then maybe we really do need Albright & Hayes’ morphology module.
The article made me think of three things:
1. I realized that I’m not 100% clear on what exactly Albright’s and Hayes’ algorithm tries to model. Do they only check on whether their model makes better predictions regarding the works of the already established language module? Or do they also make any claims about how the language is acquired or learned?
2. Do analogical models make any claims about being able to model how the language is acquired? Or do they only claim to imitate the state of the language module AFTER it’s been established? I was thinking that maybe Albright and Hayes could use an additional argument in the favor of the rule-based model, by saying that language acquisition can (as far as I know) be usually explained by deriving more abstract generalizations over sets of data. (It seems to me) that it would be much harder to explain acquisition in terms of analogy, since it is unclear how the “baseline” form would be established.
3. I was also thinking whether a model that works only on pairs of base-form and final state could account for cases of paradigm contrast, as in Russian and Hungarian, which have additional factors of pure surface similarity to other members of paradigm that have to be taken into consideration. From what I saw, analogical models (such as David Eddington’s on Spanish stress) can include the factors of grammatical class to affect the final output…however, in cases where the whole paradigm is involved, the model will probably have to have access to all the paradigm, before selecting the appropriate suffixes.
(On another note, is it possible to wug-test a whole paradigm? Meaning, do people physically capable of holding a new paradigm of new words in their memory for 1/2 hour or so? I was thinking that it might be interesting to see how productive the paradigm contrast is, but I can’t think of a way how it could be modeled…)
Is it just me, or does the “phonological filter” that Albright and Hayes use to rule out forms like [nid+d] look a bit like sweeping morphophonological dust under the carpet?
First, I find it a little odd that Albright and Hayes need a morphology component that consists of a set of probabilistically reliable rules, and a phonology component that consists of (what appears to be) a set of exceptionless, categorical constraints/rules.
Second, does this partitioning hold up typologically? Or do there exist examples of apparent cyclicity/parallelism that would require the phonological filter to apply before/at the same time as the morphological component?
Third, if we assume that morphology works the way Hayes and Albright says it does, but the phonological filter is still OT-like, then do speakers ever learn phonological rankings that can only be established by morphological alternation?
For example, consider a language that allows adjacent consonants with opposing [voice] specifications root-internally. Imagine that plurals in this language are like a completely regular (simplified) version of English — roots that end in a voiceless sound take /s/, while roots that end in a voiced sound take /z/. In OT, this alternation would be produced by the ranking:
Ident[voice]-Root >> *[a voice][b voice] >> Ident[voice]
However, because the realization of the plural is completely regular in this language, we don’t ever need to establish this ranking in the phonological filter, because Albright and Hayes’ morphology component will take care of it. We’ll just posit one morphological rule that realizes the plural as /s/ after voiceless sounds, and one that realizes it as /z/ after voiced sounds, and we’ll eventually learn that those rules each have 100% reliability, since there are no competing forms. As a result, the three constraints given above remain unranked with respect to each other in the phonologicla filter, and there’s no consequence to the learner, because what s/he knows about morphology will get the right result every time.
It would be interesting to find a language like this, and then to test whether something that can be learned as a 100% reliable morphological rule behaves more like a morphological rule, or a phonological rule. For example, we could teach spekaer of this language a novel inflectional suffix for a morphological feature that their own language doesn’t realize as a suffix, give it the same forms and distribution as the plural, and see how fast they learn it.
If they pick up on it right away, then they must have phonologized the right ranking base on the native plural forms– but why would they have done that, if the morphology already gave them the correct result?
If it takes time for them to get the distribution right (or they never do), then maybe we really do need Albright & Hayes’ morphology module.
The article made me think of three things:
1. I realized that I’m not 100% clear on what exactly Albright’s and Hayes’ algorithm tries to model. Do they only check on whether their model makes better predictions regarding the works of the already established language module? Or do they also make any claims about how the language is acquired or learned?
2. Do analogical models make any claims about being able to model how the language is acquired? Or do they only claim to imitate the state of the language module AFTER it’s been established? I was thinking that maybe Albright and Hayes could use an additional argument in the favor of the rule-based model, by saying that language acquisition can (as far as I know) be usually explained by deriving more abstract generalizations over sets of data. (It seems to me) that it would be much harder to explain acquisition in terms of analogy, since it is unclear how the “baseline” form would be established.
3. I was also thinking whether a model that works only on pairs of base-form and final state could account for cases of paradigm contrast, as in Russian and Hungarian, which have additional factors of pure surface similarity to other members of paradigm that have to be taken into consideration. From what I saw, analogical models (such as David Eddington’s on Spanish stress) can include the factors of grammatical class to affect the final output…however, in cases where the whole paradigm is involved, the model will probably have to have access to all the paradigm, before selecting the appropriate suffixes.
(On another note, is it possible to wug-test a whole paradigm? Meaning, do people physically capable of holding a new paradigm of new words in their memory for 1/2 hour or so? I was thinking that it might be interesting to see how productive the paradigm contrast is, but I can’t think of a way how it could be modeled…)