I just finished Cognitive Linguistics: Basic Readings (ed. Dirk Geeraerts).  I wanted to get up to speed on this cognitive linguistics, as the kids call it these days.  I actually like it more than generative syntax, though I think its program is better than its progress.  It wants to be cognitive, even if mostly it’s still people noodling around on sketchpads rather than, say, in the genome.  But it’s still an improvement over Chomsky, who despite holding that grammar is innate, is aggressively uninterested in the brain, in the evolution of language, or in any way language might be integrated into general cognition.

I used a few of the essays as fodder for my upcoming Xurnese grammar; but I thought I’d highlight one here: “Usage-based linguistics”, an overview of developmental studies by Michael Tomasello.  What’s most striking about his research, to me, is how incompatible it is with Chomsky’s latest theories.

Briefly, Chomsky thinks syntax is innate– we’re born with a complete universal grammar, and all the child has to do is figure out a few switches to flip once she discovers that she’s learning English rather than Ojibwe.  He supports this idea with the “poverty of stimulus” argument, which claims that there is far too little linguistic input to learn the full complexity of syntax.

Here’s what Tomasello found: children at first learn single-meaning utterances (‘holophrases’).  Some are imitations of entire sentences (“Gimme-that”), but it’s pretty clear that they’re used as invariable units.  Next, they try two-unit phrases– usually a noun filling a slot (“More juice”), though sometimes the added element is a verb (“I-wanna walk”).  Later the slots get more complicated.

He had an unusually rich corpus for one 2-year-old, and took the last half-hour, counting 455 distinct utterances.  Then he compared each one to the previous weeks of data.  78% were word-for-word duplicates of previous utterances.  18% were copies of previous utterances with one minor change; just 4% had two changes, though in each case the particular changes were themselves already attested. 

That is: the vast majority of utterances were things the child had said before, or very minor variations on them.  It’s been noticed that children rarely learn a new pattern that’s demonstrated in front of them, which has been taken as meaning that they don’t imitate adult speech.  But now we see that they don’t do it because a single instance isn’t enough data for them.  They don’t venture to use a new construction till they’ve heard it many times and know how to use it.

A nice confirmation of this: children learning inflectional languages don’t learn the six person/number combinations at the same rate.  They first master the ones with the highest frequency in adult speech– e.g. 1st person singular, rather than 3rd person plural.  Again, they’re learning by imitation, and it takes a huge amount of repetition for them to learn something.  They also seem to learn each verb paradigm separately– it takes a long time before they start generalizing.

Another supposed bit of evidence that children don’t imitate adult speech is that they make errors like “Her open it.”  But Tomasello points out children hear plenty of expressions like “Let her open it” or “Help her open it.”  They’re reproducing part of an utterance that they’ve heard without understanding the whole thing.  They don’t make mistakes like “Mary hit I”, because that never occurs in what they hear.

A child may use what seems like a complex construction, but it can be an illusion.  For instance, he found that kids pretty quickly say “I think…”  But at that age they didn’t have other forms (“she thinks”, “I don’t think”, “I thought”, or even “I think that”).  So this is very likely not real subordination; rather, they’re using “I-think” to mean “maybe”.

Now, why is all this striking?  Because it doesn’t fit at all with an innate, complete understanding of syntax. If all the kid had to do was flip switches, she wouldn’t struggle like this.  They aren’t born knowing the patterns of language; they have to laboriously acquire them, adding new features only after they’ve been exposed to a load of data.