3 ms·
Chomsky's idea wasn't that there was a strong inductive bias in language learning, it was that language isn't learned at all [...] It's because, acc. to Chomsky
by _delirium 16y ago
Chomsky's idea wasn't that there was a strong inductive bias in language learning, it was that language isn't learned at all [...] It's because, acc. to Chomsky and his followers ("Modern Linguists"), Greek and English are really the same language, and kids don't LEARN this language, called "Universal Grammar," they GROW it. What they DO learn is the values of some "parameters" that cause different instantiations of UG to appear superficially different, like Greek and English.
At least from the perspective of machine learning, those seem like similar claims to me! A very common form of strong inductive bias is that the learner has a parameterized model class, and is learning model parameters. For mathematical simplicity the models are usually simpler than anything towards the upper ends of the Chomsky hierarchy, but there's no particular reason the model class can't involve a regular expression or context-free or context-sensitive grammar.
My impression of the UG proponents is that they think the size of the parameter space is quite small relative to the total size of the symbol-protocol space, so the language-learning problem is greatly simplified, perhaps so much that it's not even worth calling it learning (but that seems like a philosophical dispute with a gray area). The way to argue against that afaict would be to somehow show that in fact the minimal size of the parameter space is quite large. Perhaps computational methods don't make that feasible to do currently though, hence the focus on a genealogical analysis of one particular language feature.