A number of P&P’s empirical objections relate to the lack of a representation of variables in RM’s model. They seem to allude to this as a general problem for PDP models; how bad is it in actuality? Networks beyond a certain complexity can certainly do variables. Is there any obvious work to point to as a solution to this? As they point out, possible models don’t save PDP in general, and even other models won’t save RM in particular, but it seems like an important point. How well do structures like hidden nodes fill the gap?
I’m wondering how strong P&P’s problems with Wickelphones/Wickelfeatures and integration of phonology and morphology in one model ultimately are.
I think it should be possible to accommodate segment repetition into Wickelphone-style representations if prosodic structure is introduced (in the form of additional boundary symbols); for instance,
(‘bala)ba
would become
{#ba, baS, aSl, Sla, laF, aFb, Fba, ba#}
– where S stands for a syllable boundary and F for a (syllable and) foot boundary (I’m kind of adopting representational stuff from this paper:
).
If it is true that morphological or prosodic structure is needed to allow repetition of phoneme sequences, then we might explain why languages tend to avoid repetition of adjacent material in phonology – and why repetition is often a sign of morphology-driven reduplication.
Secondly, and I may be wrong here, it seems to me that P&P provide no conclusive argument as to why RM’s model is fundamentally unable to capture the interaction of phonology and morphology – it seems perfectly fine for the model to duplicate phonological regularities for every morphological derivation, as long as there is supervised learning. And it is true that there is morpheme-specific phonology, so perhaps it -is- true that phonological patterns are duplicated throughout morphology.
I don’t know if new morphological constructions will result in the application of word-bounded phonological processes. If they don’t, then it will suffice to enrich RM’s model with a syntactic level which feeds a network determining between-word phonology – there will still be phonological regularities that are enforced over entire sentences, and therefore, formation of new morphological constructions will take over sentence-level phonological patterns.
A number of P&P’s empirical objections relate to the lack of a representation of variables in RM’s model. They seem to allude to this as a general problem for PDP models; how bad is it in actuality? Networks beyond a certain complexity can certainly do variables. Is there any obvious work to point to as a solution to this? As they point out, possible models don’t save PDP in general, and even other models won’t save RM in particular, but it seems like an important point. How well do structures like hidden nodes fill the gap?
I’m wondering how strong P&P’s problems with Wickelphones/Wickelfeatures and integration of phonology and morphology in one model ultimately are.
I think it should be possible to accommodate segment repetition into Wickelphone-style representations if prosodic structure is introduced (in the form of additional boundary symbols); for instance,
(‘bala)ba
would become
{#ba, baS, aSl, Sla, laF, aFb, Fba, ba#}
– where S stands for a syllable boundary and F for a (syllable and) foot boundary (I’m kind of adopting representational stuff from this paper:
http://www.sciencedirect.com/science/article/pii/S0024384106000969
).
If it is true that morphological or prosodic structure is needed to allow repetition of phoneme sequences, then we might explain why languages tend to avoid repetition of adjacent material in phonology – and why repetition is often a sign of morphology-driven reduplication.
Secondly, and I may be wrong here, it seems to me that P&P provide no conclusive argument as to why RM’s model is fundamentally unable to capture the interaction of phonology and morphology – it seems perfectly fine for the model to duplicate phonological regularities for every morphological derivation, as long as there is supervised learning. And it is true that there is morpheme-specific phonology, so perhaps it -is- true that phonological patterns are duplicated throughout morphology.
I don’t know if new morphological constructions will result in the application of word-bounded phonological processes. If they don’t, then it will suffice to enrich RM’s model with a syntactic level which feeds a network determining between-word phonology – there will still be phonological regularities that are enforced over entire sentences, and therefore, formation of new morphological constructions will take over sentence-level phonological patterns.