1. What role does the social network play in the simulations performed here? Is it the case that the social network could be replaced by running the simulation many times on a sequence of one-agent generations (i.e., one adult talking to one child)?
2. This paper made me think about the difference between Maximum Entropy models of learning, where notable absence of certain forms leads to promotion of a constraint against these forms, and the kind of model espoused here, which appears to be a kind of Bayesian model that operates over exemplars. What predictions could tease these two classes of models apart?
Here again in this paper, we have the authors selecting where in the word (the stem-final consonant) that the morphophonological similarity measure will be focused on. Footnote 8 says that stem-final consonants are important for other predicting other things. Since the focus location of the similarity measure must be learned, one wonders how successfully algorithms will learn this and not some other spurious generalization. Because this is suffixation, it seems natural the context closest to the suffix (the stem-final segment) would be more important than more stem-internal segments. But, if the morphological operation was something else like internal vowel ablaut, would the context location be just as easily determined?
I am confused over author’s assumptions. On p. 942 they state that the information about non-usage of certain form blocks generalizations (over these forms). Do they assume that for all other forms the child hears ALL possible inflections?
Also, since according to the authors the gaps exist only for 10.4% of dental stems (more than for other stem types though), I wonder how the Russian speakers would do (or already do) on a wug-test…
I have two questions:
1. What role does the social network play in the simulations performed here? Is it the case that the social network could be replaced by running the simulation many times on a sequence of one-agent generations (i.e., one adult talking to one child)?
2. This paper made me think about the difference between Maximum Entropy models of learning, where notable absence of certain forms leads to promotion of a constraint against these forms, and the kind of model espoused here, which appears to be a kind of Bayesian model that operates over exemplars. What predictions could tease these two classes of models apart?
Here again in this paper, we have the authors selecting where in the word (the stem-final consonant) that the morphophonological similarity measure will be focused on. Footnote 8 says that stem-final consonants are important for other predicting other things. Since the focus location of the similarity measure must be learned, one wonders how successfully algorithms will learn this and not some other spurious generalization. Because this is suffixation, it seems natural the context closest to the suffix (the stem-final segment) would be more important than more stem-internal segments. But, if the morphological operation was something else like internal vowel ablaut, would the context location be just as easily determined?
I am confused over author’s assumptions. On p. 942 they state that the information about non-usage of certain form blocks generalizations (over these forms). Do they assume that for all other forms the child hears ALL possible inflections?
Also, since according to the authors the gaps exist only for 10.4% of dental stems (more than for other stem types though), I wonder how the Russian speakers would do (or already do) on a wug-test…