Please put comments and questions on serial HG and its learning here.
3 thoughts on “Week 9 comments and questions”
Ivy
Questions and comments on the Mullin paper:
Is linking like syllabification in HS? Does it come for free? It’s a bit confusing because linking a segment to an existing feature is one change but inserting a segment then linking it is one change as well. Intuitively the latter would require two changes unless linking is “free”.
Why do some segments block harmony and not others? I can see a possible explanation for why voiceless obstruents tend to block vowel harmony instead of more sonorous segments – because they are less vowel-like? If, as in serial HG, the feature spreading must go through each segment then it would make sense that a non-vowel-like segment could potentially block spreading of vowel features. The examples in this paper were of these kind, where voiceless obstruents block harmony. Do we see other kinds of segments which block vowel harmony?
Also, why would some features be strong triggers and others weak? From the examples presented it seemed that coronal was the strong feature and dorsal was the weak feature in multiple cases. This parallels markedness of place of articulation. Is that a possible reason? The less marked POA is the strong trigger?
The constraint which differentiates strong and weak heads is the *dependent-head(W) constraint. Are there any restrictions on what features are allowed to be referenced by it? It seems to me that this constraint makes very broad typological predictions because any features can be weak and any can be strong. Surely this is not actually the case and there are patterns in which features can be strong and weak heads. If coronal strong triggers are the usual case (as in all the examples from the paper), then maybe our constraint set should encode this.
I wanted to follow up on a question I asked in class this week.
In a MaxEnt framework, the harmony scores and probabilities of occurrence are related. In this model, would it be impossible to have variation between one form with the highest harmony score and another form with a much lower harmony score (not the second best, but maybe the fifth best candidate). I realize we can always change the constraints to change the harmony scores and probabilities. It is the relationship between harmony scores and probability that I’m thinking about – how do we know it is necessarily correct? Do we observe any cases of variation (say 50/50 probability of each) between one form which would have a high harmony score and another form with a very low harmony score?
Questions and comments on the Mullin paper:
Is linking like syllabification in HS? Does it come for free? It’s a bit confusing because linking a segment to an existing feature is one change but inserting a segment then linking it is one change as well. Intuitively the latter would require two changes unless linking is “free”.
Why do some segments block harmony and not others? I can see a possible explanation for why voiceless obstruents tend to block vowel harmony instead of more sonorous segments – because they are less vowel-like? If, as in serial HG, the feature spreading must go through each segment then it would make sense that a non-vowel-like segment could potentially block spreading of vowel features. The examples in this paper were of these kind, where voiceless obstruents block harmony. Do we see other kinds of segments which block vowel harmony?
Also, why would some features be strong triggers and others weak? From the examples presented it seemed that coronal was the strong feature and dorsal was the weak feature in multiple cases. This parallels markedness of place of articulation. Is that a possible reason? The less marked POA is the strong trigger?
The constraint which differentiates strong and weak heads is the *dependent-head(W) constraint. Are there any restrictions on what features are allowed to be referenced by it? It seems to me that this constraint makes very broad typological predictions because any features can be weak and any can be strong. Surely this is not actually the case and there are patterns in which features can be strong and weak heads. If coronal strong triggers are the usual case (as in all the examples from the paper), then maybe our constraint set should encode this.
Re “Learning serial constraint-based grammars”
Do I understand correctly that the only part to be learned are the weights — all the rest follows from definitions and standard probability theory?
I wanted to follow up on a question I asked in class this week.
In a MaxEnt framework, the harmony scores and probabilities of occurrence are related. In this model, would it be impossible to have variation between one form with the highest harmony score and another form with a much lower harmony score (not the second best, but maybe the fifth best candidate). I realize we can always change the constraints to change the harmony scores and probabilities. It is the relationship between harmony scores and probability that I’m thinking about – how do we know it is necessarily correct? Do we observe any cases of variation (say 50/50 probability of each) between one form which would have a high harmony score and another form with a very low harmony score?