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 grows in the brain, as his accolyte Pinker claims, "like teeth grow in the mouth." You don't "learn" to have teeth, do you?, Pinker asks.
How can this be, when we see so obviously that kids raised by Greek speakers learn Greek while those raised by English speakers learn English?
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.
This quickly required the definition of "language" to exclude every aspect of language except syntax, because everything else so obviously was learned and wasn't universal, and even with syntax, they ended up having to exclude the syntax actually used by real speakers really speaking. That syntax become known as "performance" as distinct from "competence", which was the universal language people really knew deep down inside but which got corrupted in real use by various noise-inducing factors.
So "real language" became syntax only, and then, only carefully constructed written examples of proper, uncorrupted "competence" syntax. And with those, all native speakers could use their innate rules plus parameter settings to unanimously agree on which word sequences were valid and which were invalid. You had a rule-based syntax with universal rules, right?
Except that, with each passing year, researchers outside the linguistics departments found more and more examples where native speakers disagreed over validity, and validity decisions that weren't discrete (clearly valid or clearly invalid) but were often shades of gray. (Inside linguistics departments, conformity to Chomskyism orthodoxy was usually enforced, as academia tends to do, so "linguistics research" has supported Chomsky for decades, while cognitive science hasn't.)
There were so many differences among native speakers of a single language, not to mention the ever-growing catalog of diversity between languages and the way that languages gradually diverge instead of in clearly discrete jumps, and so many shades of gray in judgments among natives that the notion of parameters got more and more ridiculous, even with the notion of "language" pared down to almost nothing.
Sure, you could claim that all books are the same universal book, too, as long as each character is a parameter. If so, then all languages are the same language under the surface, too, plus or minus some parameter settings.
Chomsky's ridiculous language ideas would have been thrown out long ago if he hadn't been such a leftist "intellectual icon" in the leftist temples of academia and media. Instead, his theory just changed dramatically from version to version but remained unquestionably true throughout. Every few years he would significantly revamp his "program." In 2002 he seemed to abandon everything about syntax, too, except for recursion. What's unique about human language, as opposed to the general principles found in all human (and some animal) cognition, is just the innate, universal ability to handle recursion in syntax. (That paper apparently infuriated Pinker by cutting his "Language Instinct" position down to just a syntactical recursion instinct.)
What a bunch of nonsense. And this paper is just one more nail in the coffin of "modern linguistics".
This is factually wrong in a number of ways. Modern linguists don't throw out "every aspect of language but syntax" (as you yourself indicate later in your own post). There are falsifiability problems ("all books are the same universal book too, so long as each character is a parameter") with some theories, but you have to start with a theory at some point.
I can relate to your sentiment that he's favored because he's such a lefty, but it's not really an argument.
He revamps his "program" every few years because he recognizes it as wrong. I'm not sure why you cite that as a bad thing?
"He revamps his "program" every few years because he recognizes it as wrong. I'm not sure why you cite that as a bad thing?"
The "bad thing" is not the person who decides that he was wrong; the bad thing is the theory that was wrong and the "modern linguists" who still promote it. The notion of "universal grammar" came from the original version(s) of the theory: Everyone knows a fantastically complex rule system perfectly by some very young age, which can't possibly be learned from noisy, real-world experience in so short a time, so language must be innate, like teeth, and since humans are all the same species, the language must be a universal language, which must have some parameters to explain the illusion that they aren't the same.
By 2002, there were not enough legs of that original theory still standing (still believed by even Chomsky himself) to support the notion of a "universal grammar."
Yet despite the fact that nobody starting fresh with what we know today would ever propose a theory of innate, universal grammar, we still have most people calling themselves "modern linguists" claiming to believe it.
Universal grammar is nonsense, and modern linguists' failure to drop it is the "bad thing."
"...you have to start with a theory at some point."
Yes, and you have to drop it at some point when new evidence keeps making it less and less plausible. That point was years ago.
The alternative to Universal Grammar is Skinner style behaviorism; which is clearly wrong.
Think of Universal Grammar is a rule for building grammar rules.
It's pretty clear that humans have a distinct innate ability to learn language: no monkey can learn English no matter how much you try to teach it. You can teach animals all kinds of interesting behavior but you can't teach them language.
> Yet despite the fact that nobody starting fresh with what we know today would ever propose a theory of innate, universal grammar, we still have most people calling themselves "modern linguists" claiming to believe it.
Quite the contrary. Anyone starting fresh would probably start with an assumption about some innate ability to learn language.
The alternative to UG is not behaviorism; there are countless alternatives. There are all sorts of learning algorithms that are more plausible than UG or behaviorism.
Yes, it IS clear that humans have an innate ability to LEARN languages, as you insist. Unfortunately, UG denies this, claiming that we CAN'T possibly learn anything as rich and complex as a human language in so short a time with so little, and such messy, input, and since humans have NO innate ability to LEARN human (first) languages, they must instead GROW them "like you grow teeth."
"The alternative to UG" isn't behaviorism, it's that languages are LEARNED.
>"Quite the contrary. Anyone starting fresh would probably start with an assumption about some innate ability to learn language."
You're so right, except that your claim is not contrary to me, it's contrary to UG. Now try to convince the modern linguists of your theory that humans have the innate ability to LEARN first languages and see how that goes.
You are substantially mis-characterizing modern linguistic theory/linguists. The 'grow them like you grow teeth' is meant to indicate that, given certain inputs/environment, a child will develop normal language function. You don't need to 'learn' it in the sense that you do need to learn e.g. how to read. The reason for the contrast (argues a linguist) is that we have some internal cognitive structures that react to certain kinds of input, namely linguistic input, and that that reaction is called 'learning your first language(s)'. They disagree with the 'common sense' approach, that learning your first language is just like learning anything else.
I would characterize the debate between linguists and a certain class of cognitive scientists like this:
CogSci: Hey, you keep talking about UG/Innate mechanisms! We don't like that/it seems implausible. Instead, we should just have general learning algorithms that can be utilized to learn language!
Ling: Cool! Show us! Show us!
CogSci: Well...Here's a machine learning model that can learn English past tense with the following training data.
Ling: Oh. Um. Hmmm. Yeah, the data is more complex than that. How far can you get with this data (unloads data by the truckful). Also: that looks like how adults learn things (the kinds of errors made), not really how kids learn language (they make different kinds of errors). Can you model that?
CogSci: It's a simple model! It can't handle that data. That's for a later paper! Also, we don't care about the error classifications, as long as it looks like learning.
Ling: Ok. Let us know when that paper comes out. Have you seen this Bantu data? It's pretty cool too.
CogSci: later: Ok, look. We didn't get the model to work, but we really think your multiplying entities. I mean, it's just crazy/biologically-implausible/ugly to postulate this innate knowledge.
Ling: Yup. But here's the deal. We can't manage to actually explain everything we want even if we postulate innate rules/knowledge like crazy. Maybe we have a fundamentally broken model. Maybe machine learning really will come and eat our babies (or maybe the kinds of things we're postulating will turn out to built on top of machine learning, as explanations at different levels). But so far, it's the best we've got.
Obviously, people write books on these arguments, so some massive simplification was done here. And there is some really cool work being done by general cognitive scientists in the language space.
I'm not contradicting UG. From what I took out of the linguistics course I took at university, the idea is that part of this innate ability is an innate understanding of grammar, or rather we expect to learn grammar and so when we receive all the "input" our mind tries to make sense of it.
Re Leftist: Many subscribe to the democratic theory of truth, since if everyone thinks it, it's probably true. We are conforming creatures, the iconoclast is the exception, and it also explains why paradigms in science last so long.
Re Universal Grammar: But I think parallel evolution is the rule, not the exception; when confronted by the same needs, with the same tools, similar solutions are often arrived at. In the case of human communication, the needs seem to be to describe things, changes and qualities. The tools are probably what amounts to the Universal Grammar: things like the sequence of things being noticeable; that you can flag one thing to relate to another (like cases, tenses, number). The specific rules then just fall out, depending on whatever gets a foothold in the community.
Going out on a limb here: I have doubts that recursion is really fundamental to human language: most people top out very quickly (e.g. 2-3 levels). especially if they actually want to think in terms of it. That is, it only has a (very) finite recursive structure, which you can simulate with a less powerful grammar (e.g. regular instead of CFG). What we do have is the ability to reference other things (which can be cyclic): "Like this sentence."
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.
The fact that native speakers of a single language disagree on the validity of certain utterances does not invalidate the concept of Universal Grammar. No two individuals learn their native languages in exactly the same way. The language that they pick up is shaped by their unique experiences with other people using the language, so it is unlikely that any two individuals have the exact same "instance" of the language (a more technical way to put this is that everyone has their own ideolect), and thus may differ in the judgements on whether particular utterances in that language are valid.
Even though individuals may speak the same language, because the concept of this language in each of their heads may differ slightly due to their individual experiences, one could say that a few of their parameters differ. If groups of individuals are separated from one another, these parameters can end up diverging between populations, leading to differing dialects and even separate languages over longer periods of time.
From reading the Ars Technica piece (I don't have access to the original Nature article), it seems like they've shown that language features are highly dependent on lineage, but I don't see how this result is incompatible with the concept of Universal Grammar. Because a language has to be relearned by every individual born into the language environment, there is a lot imperfect copying going on, leading to individuals having slightly different parameters of their Universal Grammars. There is a direct biological analogy here: imperfect copying of DNA leads to mutations which can lead to speciation over time. The values of the DNA base pairs are the parameters, but the mechanism by which the DNA is translated into proteins by every individual is still the same. This mechanism is analogous to Universal Grammar translating parameters into the actual spoken language.
It is true that Universal Grammar tends to look at statistical occurrences of features across languages in order to deduce universal statements about them. This is the weakest part of the theory, because the absence of a feature across all languages does not necessarily mean that such a feature cannot arise, nor does the presence of a feature across all languages necessarily mean that it must always be present. Perhaps one of the worst results of language extinction is the fact that there will be fewer data sets to test proposed language universals against.
To address your other assertion, the concept of Universal Grammar does not only apply to syntax. For instance, there are phonological universals, which are deduced based on phonotactics of known languages. For example, all known languages have at the very least the minimal vowel system of three vowels /i a u/ or slight variations thereof.
"...Because a language has to be relearned by every individual born into the language environment, there is a lot imperfect copying going on, leading to individuals having slightly different parameters of their Universal Grammars."
My point with respect to the parameters notion is that the more closely we examine every aspect of language, the finer the distinctions we discover. At some point, the number of necessary "learned parameters" grows so large that it's like claiming that all books are the same except for the character parameters. The characters ARE the book, with some near-universal exceptions such as paper and binding, and the learned language parameters ARE the language, with some near-universal exceptions that are general cognitive properties of human brain architecture that apply to much more than just language.
Just because the parameters may be diverse does not mean the rules need to be just as diverse. If I may use the biological analogy again, despite the myriad combinations of DNA base pairs, they are all still translated by a universal genetic code that's simple enough to describe in one table (http://en.wikipedia.org/wiki/DNA_codon_table ). I'm fully aware that a description of Universal Grammar would likely be more complex than this though.
The collection of language parameters in a person's head is not the language itself, just like the DNA of an individual is not the species itself. It's the translation of those parameters by a shared mechanism that gives rise to the spoken language. This is where I think your book analogy falls through.
Thank you for this succinct analysis. Put this way, Chomsky's model sounds downright Platonic in its insistence that the messy things we observe are merely instances of an ideal form.
How can this be, when we see so obviously that kids raised by Greek speakers learn Greek while those raised by English speakers learn English?
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.
This quickly required the definition of "language" to exclude every aspect of language except syntax, because everything else so obviously was learned and wasn't universal, and even with syntax, they ended up having to exclude the syntax actually used by real speakers really speaking. That syntax become known as "performance" as distinct from "competence", which was the universal language people really knew deep down inside but which got corrupted in real use by various noise-inducing factors.
So "real language" became syntax only, and then, only carefully constructed written examples of proper, uncorrupted "competence" syntax. And with those, all native speakers could use their innate rules plus parameter settings to unanimously agree on which word sequences were valid and which were invalid. You had a rule-based syntax with universal rules, right?
Except that, with each passing year, researchers outside the linguistics departments found more and more examples where native speakers disagreed over validity, and validity decisions that weren't discrete (clearly valid or clearly invalid) but were often shades of gray. (Inside linguistics departments, conformity to Chomskyism orthodoxy was usually enforced, as academia tends to do, so "linguistics research" has supported Chomsky for decades, while cognitive science hasn't.)
There were so many differences among native speakers of a single language, not to mention the ever-growing catalog of diversity between languages and the way that languages gradually diverge instead of in clearly discrete jumps, and so many shades of gray in judgments among natives that the notion of parameters got more and more ridiculous, even with the notion of "language" pared down to almost nothing.
Sure, you could claim that all books are the same universal book, too, as long as each character is a parameter. If so, then all languages are the same language under the surface, too, plus or minus some parameter settings.
Chomsky's ridiculous language ideas would have been thrown out long ago if he hadn't been such a leftist "intellectual icon" in the leftist temples of academia and media. Instead, his theory just changed dramatically from version to version but remained unquestionably true throughout. Every few years he would significantly revamp his "program." In 2002 he seemed to abandon everything about syntax, too, except for recursion. What's unique about human language, as opposed to the general principles found in all human (and some animal) cognition, is just the innate, universal ability to handle recursion in syntax. (That paper apparently infuriated Pinker by cutting his "Language Instinct" position down to just a syntactical recursion instinct.)
What a bunch of nonsense. And this paper is just one more nail in the coffin of "modern linguistics".