Wednesday, December 5, 2012

Minimalism and the Language Specificity of UG

For this post let’s assume that the Minimalist project has entirely succeeded. What implications does this have for the structure of FL/UG? In particular: if MP is right does this reduce the linguistic particularity of UG? Does it mean that FL is just general intelligence plus the super special Merge sauce (with even Merge not necessarily parochial to language)?  I believe the answer is no. Let me explain.

I understand the Minimalist project as follows: It starts from the observation that humans are specially built for language. When more and more details are considered, this suggests a dedicated, pretty articulated, innate faculty of language with its own peculiar principles and organization. However, this is a problem for it seems that linguistic facility is a recent evolutionary addition to the cognitive repertoires so there is not enough time for the kind of complexity that theories of UG (e.g. LGB) attribute to FL to arise.  This is a kind of “poverty of stimulus” reasoning applied to the emergence of UG in a background where evolutionary change is gradual and complexity only emerges slowly. At any rate, Minimalism embraces “Darwin’s Problem” and considers it together with Plato’s Problem and so concludes that there is less linguistically specific to FL/UG than we earlier believed.  Does this support the conclusion that FL/UG is not a linguistically dedicated device with scads of innate structure just ready to pounce on PLD to produce grammars of specific languages? No. Why not?

The reason is that the Minimalist answer does not preclude a richly structured linguistically dedicated UG in FL and the evidence is overwhelming, that it is richly structured and very language specific. To show this, let me give another version of the Minimalist project using terminology I stole from David Poeppel.

Linguistic cognition is different. Say that cognitive theories explain some domain of cognitive facility by specifying the kinds of circuits minds/brains use to execute that kind of cognition and showing how these circuits interact to produce that cognitive capacity (think circuit diagram): so, e.g. the visual system deploys specific circuits organized in a certain way to segregate scenes and identify objects. Similarly with auditory systems, causal identification systems, numerosity systems, etc.  Now, the fact that humans seem uniquely gifted when it comes to language suggests that they are endowed with circuitry specifically dedicated to linguistic tasks. This would, for example, explain why humans do language and nothing else does. The most famous Minimalist suggestion on offer is that the special circuit is Merge. Add that circuit to those that are already there and out pops the capacity to develop linguistic facility (note: on this view Merge need not be linguistically specific, just unique to humans, see here for discussion). The project on this view is to show how the circuits that cognition in general makes available plus the Merge circuit plus whatever background “laws” of biology, physics and computation we can avail ourselves of lead to the kinds of linguistic facility we see in humans. What facility is that? The one described to a decent first approximation in LGB, for example (though you can sub the operations and principles of GPSG, LFG, HPSG, RG if they are more congenial to you, without affecting the main point). So the Minimalist project amounts to finding how a few (in the best case, just one) special circuits can combine together with those available in general cognition to derive the operations and generalizations of LGB, which we can take to be an accurate description of FL/UG (at least for present discussion). Assume we can do this, show how the LGB framework can be coded in Merge plus a bunch of other all purpose cognitive circtuis. Doesn’t it imply that there is at most a very weak sense in which FL/UG has dedicated linguistically specific structure? Nope.

To see this, consider Chomsky’s analogy between mental and bodily organs.  There are stomachs and there are kidneys. Stomachs don’t do what kidneys do and kidneys don’t do what stomachs do. GI specialists are not nephrologists and vice versa (in fact medical friends tell me that they even have very different personalities!). However, this is not because there are special stomach cells that do stomaching and special kidney cells that do kidneying. Rather the same kinds of cells, with the same kind of biological, chemical and physical properties organized in one way give stomachs and in another kidneys. So stomachs are very different from kidneys and we assume that the “genes” code specifically for stomach organizations and specifically for kidney organizations and this drives the development of these specialized organs dedicated to digestion and waste removal respectively.

I hope the analogy to FL is obvious (if not, read on).  FL differs from other areas of cognition the way that stomachs differ from kidneys. The minimalist conceit is that this is not by and large due to possessing different circuitry. The guess is that there is at most one (or two) specialized circuit (e.g. Merge) with everything else the same. On this conception, FL can be quite distinct in its properties from those of the other mental just like there is a big difference between kidneys and stomachs. In sum, the genes build stomachs with their special features, kidneys with theirs and FLs with theirs.

So, is there a rich innate structure to FL? Yes, the evidence is overwhelming.  Is this structure peculiar to linguistic cognition? Yes, the evidence is overwhelming. Does this mean that Minimalism is wrongheaded? No, new circuits added to old can produce qualitatively new things. Indeed old circuits combined in new ways can yield qualitatively different things.  Minimalism is not warmed over Piaget. There is a faculty of language with rich innate linguistically dedicated structure and even a complete realization of Minimalist ambitions does not in and of itself gainsay this.  What there may not be are many many linguistically specialized components. The Minimalist bet, indeed, is that there are very few, maybe just one. And, it is entirely consistent with a triumphant Minimalism that the structure of FL is comprised of innate linguistically dedicated structure. In other words, Minimalism does not imply that linguistic cognition it’s just (or largely) general intelligence applied to language and it's entirely consistent with rich linguistically dedicated innate conceptions of UG.


  1. Completely beside the point, but there are lots of different types of cells in the liver and the brain and the kidneys and so on and that is why they work differently, but leaving that aside (as I am not a biologist) --

    You seem to be squashing UG and FL together; and the acquisition of language (which seems to require rich domain specific UG) together with use and knowledge of language, (which don't).

    So I kind of buy into the idea that one can derive properties of grammars from principles of computation -- this is kind of the research program of mathematical linguistics -- but I don't see how you could explain learnability in this paradigm. In your discussion of the Gleitman paper you were talking about how you needed a restricted hypothesis space -- where does that come from?

    I think this is the part of the MP that I have most difficulty understanding -- it slightly feels that you want to have your cake and eat it -- have a rich UG that solves the learning problem, but have a minimal UG that is evolutionarily plausible -- and I just don't see how the details of this will work out.

  2. I'll also leave aside the biology.

    I am not squashing, I am distinguishing, or at least I hope I am. The way it worked pre Minimalism is that UG described the knowledge that FL brought to the problem of learning. This was the given that allowed for learning to take place and, as given, is not itself learned. Minimalism comes along and people say, look, what was given in GB and was language specific is still given but is now domain general. Therefore (this is the important step) FL/UG is largely domain general. My remark: there is a way of understanding minimalism where this argument does not go through (note the conclusion may be right, but it does not follow from the premises). Here's how: Cognitive domains exercise their powers through circuits wired up in certain ways. Minimalism asks whether there are any linguistic proprietary circuits. One anser is Yes: Merge (for example). Now say this is the ONLY proprietary circuit. Does this mean that the knowledge coded in the FL circuitry is largely domain general? No, it can be pretty linguistically specific if it is wired to detect linguisticy things using its circuitry. After all my stove uses the same basic circuits as my dryer but the former is not the latter. So there is domain specific knowledge based on domain general circuitry PLUS one newbie, i.e. Merge.

    Does this let me have my cake and eat it too? Hope so. It's always nice to know that you can re-eat what you've eaten, especially if very tasty.

    Now, is this how Chomsky understands the program? Don't really know. Sometimes I think so, sometimes not. However, that's not the interesting question, right? If this is a reasonable way of viewing things then Minimalism does not have the learning implications generally attributed to it by domain general learning aficionados. Take a bow here Alex.

  3. Once you have eaten your cake it's gone -- and you don't have it anymore. So the problem with
    this argument, which I now understand (thanks!) is that it seems impossible to do. Once you have a big domain-specific UG you can't just wave a magic wand and turn it into something else.

    To be specific -- take syntactic categories (or uninterpretable features or primitive types in CCG or whatever)
    This has always seemed to me to be an acid test: where do the features/categories themselves come from?
    Obviously the fact that the word 'from' is a preposition is learned, but where does the category of preposition itself come from (or whatever category one takes 'from' to be).

    (I was thinking about this after reading a review by Boeckx of a book by Corbett on features).
    There seem to be two stories,
    one where they are innate (and therefore universal) and one where they are learned from the data. Now if I take your claim here at face value, you want the latter:
    You say:
    "Does this mean that the knowledge coded in the FL circuitry is largely domain general? No, it can be pretty linguistically specific if it is wired to detect linguisticy things using its circuitry."
    So I completely buy this part of it, if I understand it correctly .. even a die hard empiricist like myself says that the adult knowledge of language is language-specific, the output of the learning process.
    The debate is over whether the wiring to detect linguisticy things is itself linguisticy: whether the learning algorithm itself is linguisticy. (This is what I meant about distinguishing FL and UG -- though my terminology may be non standard.)

    But I think I am misreading you here? -- do you want to take some intermediate position on categories?

    I think categories are a useful piece of knowledge to focus on since they are fairly concrete and very linguistically specific. Even if one takes the view that say nouns are projected from some domain-general semantic notion of an 'entity' or object, there are plenty of syntactic categories that don't correspond to semantic categories in any nontrivial way.

  4. I put together a Reply To Alex as a post as what I wrote was too long to post as a reply. See there.