Maybe I'm reading too much into it, but a genealogical analysis of language seems like a strange way to attack Chomsky's overall hypothesis. The fundamental point of the poverty-of-stimulus argument is that there must be strong inductive biases in whatever language-learning apparatus we possess, since we seem to quickly converge on something deemed grammatical in our language, with relatively few examples.
Showing that a particular property varies pretty randomly across languages, correlated only with lineage, is decent evidence that that property isn't universal. That would attack some stronger possible formulations of universal grammar (which maybe Chomsky has actually advanced; I don't recall) that claim not only that we have an inductive bias in language learning, but that: here is a list of specific language features that are universal.
But it doesn't show that there isn't some inductive bias, or even a generative-grammar-style inductive bias, where "human language, in all its possible variations" is a parameterized space significantly smaller than the space of all possible symbol protocols, and relatively easy for us to learn because we're hardwired with the parameterization (but not the specific parameter settings).
Admittedly I am familiar with machine learning and not very familiar with current work in linguistics, so I could be totally missing some reason that makes my field's methods inapplicable to theirs, inconceivable as that might be. ;-)
Showing that a particular property varies pretty randomly across languages, correlated only with lineage, is decent evidence that that property isn't universal. That would attack some stronger possible formulations of universal grammar (which maybe Chomsky has actually advanced; I don't recall) that claim not only that we have an inductive bias in language learning, but that: here is a list of specific language features that are universal.
But it doesn't show that there isn't some inductive bias, or even a generative-grammar-style inductive bias, where "human language, in all its possible variations" is a parameterized space significantly smaller than the space of all possible symbol protocols, and relatively easy for us to learn because we're hardwired with the parameterization (but not the specific parameter settings).
Admittedly I am familiar with machine learning and not very familiar with current work in linguistics, so I could be totally missing some reason that makes my field's methods inapplicable to theirs, inconceivable as that might be. ;-)