Original researcher here. For context, people are not arguing that nothing is learned, but instead that children come innately prepared with abstract categories of "noun" and "determiner" even if they don't know which words are nouns or determiners or even exactly how these categories are combined.

Here's a paper that makes this argument: http://www.ling.upenn.edu/~ycharles/PNAS-2013-final.pdf

Hey, Michael. Thanks for posting. This is a more understandable claim than the one I read from Stanford News. That was I guess a misunderstanding on my part.

I'll take a look at the paper you linked, but I'm also curious, how does your technique make a distinction between rule-based grammar being fully learned vs mostly or partially learned? That is, how can you say that improved grammar over time is evidence against innate ability to understand/use grammar? It seems we have an innate ability to walk in that we seem to be "wired for it", but we are still terrible at it for a bit over a year on average. The fact that we get noticeably better over time doesn't seem to rule out innate ability. This of course assumes "innate ability" doesn't mean "completely unlearned".

Sorry for the slow response. Our model is - we think - an advance because it actually allows us to infer the amount of shared structure between different words. This is an index of how much generalization the child is doing. If they do no generalization, that's evidence for what's called "item specific learning" (no abstract grammatical rules). If they do complete generalization, that's the innateness hypothesis in the paper I linked. What we found was - of course - a mixture, but critically with something close to item-based learning in the very earliest data from the very youngest kids.

To simplify, the Chomskyan position is that there are certain shared characteristics of all human languages' grammars, and that these characteristics reflect structures in the human brain.

Fair enough. The article explained that poorly and made it sound as if there is a real contingent of psychologists who think that babies magically know grammar as opposed to learning it.

Given this reframing of the research, I fail to see how this study demonstrates anything, though. If the brain has special structures that allow humans to learn grammar, you would still expect to see the same pattern of "terrible, broken grammar -> mediocre grammar -> good grammar" that we see. We see the same learning pattern for things that the brain does have special structures for, such as... well... everything from sight to motor control.

Like most ticklish philosophical problems, I think it becomes less obvious the deeper in to the problem you look.

There are a number of considerations at play here, and a long history of violently disagreeing about them.

Is language tied to grammar? Or is grammar tied to language[1]? There has been a lot of work seeking to examine the common underpinnings of languages that seem otherwise disparate, and a somewhat flexible core grammatical structure has been proposed as one of these unifying components. The speculative step some take next (including people like Chomsky) is that the reason that all of these common patterns appear is that the brain has evolved some specific hardware that serves as the basis for a set of communication protocols that we call natural languages, and all natural languages instantiate this basic hardware in different ways[2].

If you think this kind of hardwiring is possible, then another obvious-but-not-obvious question follows. Can language really be learned[3]? Can you take a clean slate of neurons (some arbitrary configuration of connections) and, given enough input data, end up with a system that understands language (to frame it as a machine learning problem, which it may or may not be)? Or is it closer to a game of connect-the-dots; the bare framework is there, built into the way the brain arranges itself over the course of millions of years. Input data is just required to figure out exactly what the final arrangement of the picture is, whether it looks like English or Russian or Mandarin.

The current mood in cognitive neuroscience academic circles seems to be the former case, that you can 'simply' throw enough data at a general-purpose learning model and get a language-speaker out. Partly due to the success to deep learning approaches in what were previously very difficult problems in NLP. But I don't think it's fair to say there's nothing worthy of debate here.

---

[1] Or is this a false dichotomy and there's some even weirder relationship between the two.

[2] For a primitive example, languages with SVO ordering vs. SOV ordering. Different instantiations of the same triple with subjects, objects, and verbs. Ray Jackendoff took this another step with formulations like XBar grammar, which was trying to demonstrate how a universal, parameterized grammar might look.

[3] Grice had a fun theory that the observable signal of language is _so_ noisy that it's impossible to learn outright, regardless of the number of examples you're exposed to. And thus, everyone speaks a slightly different language, and some sets of these idiolects are, roughly, mutually compatible. I don't think he ever did any math to demonstrate this though.

> Grice had a fun theory that the observable signal of language is _so_ noisy that it's impossible to learn outright, regardless of the number of examples you're exposed to. And thus, everyone speaks a slightly different language, and some sets of these idiolects are, roughly, mutually compatible. I don't think he ever did any math to demonstrate this though.

I'm not sure how you'd reduce this to something mathematically demonstrably, but it seems to me that it's the only thing that could be true. The alternatives is that languages are real, distinct, concrete things that exist independent of people, rather than arbitrary lines drawn by people around the continuous and evolving variation in ways people communicate.

So, the only hunch I have is that there might be a kind of cryptographic/information approach to it. If you treat natural language as a noisy encoding of some underlying set of intents (whatever sentience is), then the question becomes "is there enough Information in the signal (natural language) to accurately model the underlying intents statistically?". If the answer is 'no', or 'yes, but very badly', then it seems likely to me that there's some shared hardware/firmware (in this framing, the functional equivalent of a decryption key) between people. If the answer is 'yes', or 'yes, mostly at least', then there's no need for any kind of universal grammar or such, and you can learn language from a blank state. Basically, is natural language something like a one-time pad, or is it more like a breakable form of encryption?

Of course, there are plenty of assumptions in that model (for one, can you even separate 'language' from 'sentience' like an information model would suggest. I'm not so sure you can, but then I've always been somewhat sympathetic to Sapir-Whorf). I'm also not a competent enough mathematician in these fields to see where all the holes in this line of thinking are.

Maybe stated in a different way, how good do your priors have to be to learn and understand a language? With an infinite amount of training data, any language can be learned. The question is what prior assumptions are required to learn a language in a reasonable amount of time?

Thank you. As a parent of a toddler my first thought was the same as the op's but after reading your post I get why it might require more examination.

Because you're assuming the grammar is tied to the language, rather than the structure of the human brain. There may be some grammatical primitives that the brain structure supports which enables the grammar of language to function on top of them. This is one of the common linguistic explanations for why you can teach a human to understand grammar, but not a dog, for instance.

Dogs can learn. They just can't learn grammar. Why? Because the human brain is primed with grammatical primitives that a dog brain hasn't got.

Personally, I think that's a bunk explanation, but this conclusion is not obvious either way.

Nothing is obvious. If this is obvious to you then well done - give yourself a little pat on the back and smile to yourself. But in general nothing is obvious and everything must be questioned.