How to make an EVOLANG argument

Bob Berwick recently sent me something that aims to survey, albeit sketchily, the state of play in the evolution of language (evolang) and a nice little paper surveying the current state of Gould and Lewontin’s spandrels paper (here) (hint: their warning is still relevant). There have also been more than a few comments in FOL threads remarking on the important progress that has been made on evolang. I believe that I have invited at least one evolang enthusiast to blog about this (I offered as much space as desired, in fact) so as to enlighten the rest of us about the progress that has been made. I admit that I did this in part because I thought that the offer would not be taken up (a put up or shut-up gambit) and also (should the challenge be accepted) because I would really be interested in knowing what has been found given my profound skepticism that at this moment in time there is anything much to find.  In other words, for better or for worse, right now I doubt that there is much substantive detail to be had about how language actually evolved in the species.[1] In this regard, we are not unlike the Paris Academy over a century ago when it called for a moratorium on such speculation.

That said, who can resist speculating? I can’t. And therefore, this post was intended to be an attempt to examine the logic of an evolution of language account that would satisfy someone like me. I wanted to do this, because, though close to vacuous most of the discussion I’ve seen is (like the fancy inversion here?), I think that Minimalism has moved the discussion one small conceptual step forward. So my intention had been to outline what I think this small step is as well as point to the considerable distance left to travel.  

As you can tell from the modal tenses above, I was going to do this, but am not going to do it. Why not? Because someone has done this for me and instead of my laying out the argument I will simply review what I have received. The text for the following sermon is here, a recent paper by Chomsky on these matters.[2] It is short, readable and (surprise, surprise) lays out the relevant logic very well. Let’s go through the main bits.

Any discussion of evolang should start with a characterization of what features of language are being discussed. We all know that “language” is a very complex “thing.” Any linguist can tell you that there are many different kinds of language properties. Syntax is not phonology is not semantics. Thus in providing an evolutionary account of language it behooves a proposal to identify the properties under consideration.

Note that this is not an idiosyncratic request. Evolution is the study of how biological entities and capacities change over time. Thus, to study this logically requires a specification of the entity/capacity of interest. This is no less true for the faculty of language (FL) than it is for hearts, kidneys or dead reckoning. So, to even rationally begin a discussion in evolang requires specifying the properties of the linguistic capacity of interest.

So, how do we specify this in the domain of language? Well, here we are in luck. We actually have been studying these linguistic capacities for quite a while and we have a rich, developed, and articulate body of doctrine (BOD) that we can pull from in identifying a target of evolutionary interest. Chomsky identifies one feature that he is interested in. He terms this the “Basic Property” (BP) and describes it as follows:

[E]ach language yields a digitally infinite array of hierarchically structured expressions with systematic interpretations at interfaces with two other internal systems, the sensorymotor system for externalization and the conceptual system, for interpretation, planning, organization of action, and other elements of what are informally called “thought.” (1)

So one evolang project is to ask how the capacity that delivers languages with these properties (viz. I-languages) arose in the species. We call the theory of I-languages “Universal Grammar” or UG as it “determines the class of generative procedures that satisfy the Basic Property” (1). We can take UG as “the theory of the genetic component of the faculty of language.” If we do, there is a corresponding evolang question: how did UG arise in the species?[3]

Note, that the above distinguishes FL and UG. FL is the mental system/”organ” that undergirds the human linguistic competence (ie. The capacity to develop (viz. “grow”) and deploy (viz. “use”) I-languages). UG is the linguistically specific component of FL. FL is likely complex, incorporating many capacities only some of which are linguistically proprietary. Thus, UG is a subpart of FL. One critical evolang question then is how much of FL is UG. How much of FL consists of linguistically proprietary properties, capacities/primitives that are exclusively linguistic?

Why is the distinction important? Well, because it sure looks like humans are the only animals with BP (i.e. nothing does language like humans do language!) and it sure looks like this capacity is relatively independent of (viz. dissociates with) other cognitive capacities we have (see here). Thus, it sure looks like the capacity to generate BP-I-languages (BPIs) is a property of humans exclusively. And now we come to the interesting evolang problem: as a point of evolutionary logic (we might dub this the Logical Problem of Language Evolution (LPLE)) the bigger the UG part of FL, the more demanding the problem of explaining the emergence of FL in the species. Or as Chomsky puts it (3): “UG must meet the condition of evolvability, and the more complex its assumed character, the greater the burden on some future account of how it might have evolved.”

We can further sharpen the evolvability problem by noting one more set of boundary conditions on any acceptable account. There are two relevant facts of interest, the first “quite firm” and the second “plausible” and that we refer to with “less confidence.”  These are:

1.     There has been no evolution of FL in the species in the last 50k years or more.

2.     FL emerged in the way it exists today about 75k years ago.

As Chomsky puts it (3): “It is, for now, a reasonable surmise that language –more accurately UG- emerged at some point in the very narrow window of evolutionary time, perhaps in the general neighborhood of 75 thousand years ago, and has not evolved since.”[4]

Why is (1) firm? Because there are no known group differences in the capacity humans have in acquiring and using a natural language. As the common wisdom is that our ancestors left Africa and their paths diverged about 50kya then this would be unexpected were there evolution of FL or UG after this point.

Why is (2) less firm? Because we infer it to be true based on material cultural artifacts that are only indirect indicators of linguistic capacity. This evidence has been reviewed by Ian Tattersal (here) and it looks like the conclusion he draws on these issues is a plausible one. Chomsky is here relying on this archeological “consensus” view for his “plausible” second assumption.

If these assumptions are correct then, as Chomsky notes (3)  “UG must be quite simple at its core” and it must have emerged more or less at once. These are really flip sides of the same claim. The evolutionary window is very narrow and so whatever happened must have happened quickly in evo-time and for something to happen quickly it is very likely that what happened was a small simple change. Complexity takes a long time. Simplicity not so much.[5] So, what we are looking for in an evolang account of our kinds of natural langauges is some small change that has BPI-effects. Enter Minimalism.

Chomsky has a useful discussion of the role of evolvability in early Generative Grammar (GG). He notes that the evolvability of FL/UG was always recognized to be an important question and that people repeatedly speculated about it. He mentions Lenneberg and Luria in this regard, and I think I recall that there was also some scattered discussion of this in the Royaumont conference. I also know that Chomsky discussed these issues with Francois Jacob as well. However, despite the interest of the problem and the fact that it was on everyone’s radar the speculation never got very far. Why not? Because of the state of the theory of UG.  Until recently, there was little reason for thinking that UG was anything but a very complicated object with complex internal structure, many different kinds of primitives, processes and conditions (e.g. just take a look at GB theory). Given the LPLE, this made any fruitful speculation idle, or, in Dwight Whitney’s words quoted by Chomsky: “The greater part of what is said and written about it is mere windy talk” (4) (I love this Ecclesiastical description: Wind, wind, all is wind!).

As Chomsky notes, minimalism changed this. How? By suggesting that the apparent complexity of UG as seen from the GB angle (and all of GB’s close relatives) is eliminable. How so? By showing that the core features of BPIs as described by GB can be derived from very a simple rules (Merge) applied in very simple ways (computationally “efficient”). Let me say this more circumspectly: if to the degree that MP succeeds to that degree the apparent complexity of FL/UG can be reduced. In the best case, the apparent complexity of BPIs reduces to one novel language specific addition to the human genome and out falls our FL.  This one UG addition together with our earlier cognitive apparatus and whatever non-cognitive laws of nature are relevant suffice to allow the mergence of the FL we all know and love. If MP can cash this promissory note, then we have taken a significant step towards solving the evolang problem.

Chomsky, of course, rehearses his favorite MP account (7-9): the simplest Merge operation yielding unordered merges, the simplest application of the rule to two inputs yielding PS rules and Movement, natural computational principles (not specific to language but natural for computation as such) resulting in conditions like Inclusiveness and Extension and something like phases, the simple merge rule yielding a version of the copy theory of movement with obvious interpretive virtues etc.  This story is well known, and Chomsky rightly sees that if something like this is empirically tenable then it can shed light on how language might have evolved, or, at the very least, might move us from windy discussions to substantive ones.

Let me say this one more way: what minimalism brings to the table is a vision of how a simple addition might suffice to precipitate an FL like the one we think we have empirical evidence for. And, if correct, this is, IMO, a pretty big deal. If correct, it moves evolang discussion of these linguistic properties from BS to (almost) science, albeit, still of a speculative variety.

Chomsky notes that this does not exhaust the kinds of evolang questions of interest. It only addresses the questions about generative procedure. There are others. One important one regards the emergence of our basic lexical atoms (“words”). These have no real counterpart in other animal communication systems and their properties are still very hard to describe.[6] A second might address how the generative procedure hooked up to the articulatory system. It is not unreasonable to suppose that fitting FL snugly to this interface took some evolutionary tinkering. But though questions of great interest remain, Chomsky argues, very convincingly in my view, that with the rise of MP linguistics has something non-trivial to contribute to the discussion: a specification of an evolvable FL.

There is a lot more in this little paper. For example, Chomsky suggests that much of the windiness of much evolang speculation relates to the misconceived notion that the natural language serves largely communicative ends (rather than being an expression of thought). This places natural languages on a continuum with (other) animal communication systems, despite the well-known huge apparent differences. 

In addition, Chomsky suggests what he intends with the locution ‘optimal design’ and ‘computationally efficient.’ Let me quote (13):

Of course, the term “designed” is a metaphor. What it means is that the simplest evolutionary process consistent with the Basic Property yields a system of thought and understanding [that is sic (NH)] computationally efficient since there is no external pressure preventing this optimal outcome.

“Optimal design” and “computational efficiency” are here used to mean more or less the same thing. FL is optimal because there is no required tinkering (natural selection?) to get it into place.  FL/UG is thus evolutionarily optimal. Whether this makes it computationally optimal in any other sense is left open.[7]

Let me end with one more observation. The project outlined above rests on an important premise: that simple phenotypic descriptions will correspond to simple genotypic ones. Here’s what I mean. Good MP stories provide descriptions of mental mechanisms, not  neural or genetic mechanisms. Evolution, however, selects traits by reconfiguring genes or other biological hardware. And, presumably, genes grow brains, which in turn secrete minds. It is an open question whether a simple mental description (what MP aims to provide) corresponds to a simple brain description, which, in turn, corresponds to a simple “genetic” description. Jerry Fodor describes this train of assumptions well here.[8]

…what matters with regard to the question whether the mind is an adaptation is not how complex our behaviour is, but how much change you would have to make in an ape’s brain to produce the cognitive structure of a human mind. And about this, exactly nothing is known. That’s because nothing is known about how the structure of our minds depends on the structure of our brains. Nobody even knows which brain structures it is that our cognitive capacities depend on.

Unlike our minds, our brains are, by any gross measure, very like those of apes. So it looks as though relatively small alterations of brain structure must have produced very large behavioural discontinuities in the transition from the ancestral apes to us…

…In fact, we don’t know what the scientifically reasonable view of the phylogeny of behaviour is; nor will we until we begin to understand how behaviour is subserved by the brain. And never mind tough-mindedness; what matters is what’s true.

In other words, the whole evolang discussion rests on a rather tendentious assumption, one for which we have virtually no evidence; namely that a “small” phenotypic change (e.g. reduction of all basic grammatical operations to Merge) corresponds to a small brain change (e.g. some brain fold heretofore absent all of a sudden makes an appearance), which in turn corresponds to a small genetic change (e.g. some gene gets turned on during development for a little longer than previously).  Whether any of this is correct is anyone’s guess. After all there is nothing incoherent in thinking that a simple genetic change can have a big effect on brain organization, which in turn corresponds to a very complex phenotypic difference. The argument above assumes that this is not so, but the operative word is “assume.” We really don’t know.

There is another good discussion of these complex issues in Lenneberg’s chapter 6, which is worth looking at and keeping in mind. This is not unusual in the evolution literature, which typically assumes that traits (not genes) are the targets of selection. But the fact that this is commonly the way that the issues are addressed does not mean that the connections assumed from phenotypic mental accounts to brains to genes are straightforward. As Fodor notes, correctly I believe, they are not.

Ok, that’s it. There is a lot more in the paper that I leave for your discovery. Read it. It’s terrific and provides a good model for evolang discussions. And please remember the most important lesson: you cannot describe the evolution of something until you specify that thing (and even then the argument is very abstract). So far as I know, only linguists have anything approaching decent specifications of what our linguistic capacities consists in. So any story in evolang not starting from these kinds of specifications of FL (sadly, the standard case from what I can tell) are very likely the windy products of waving hands. 

---

[1] Happily, I have put myself in the good position of finding out that I am wrong about this. Marc Hauser is coming to UMD soon to give a lecture on the topic that I am really looking forward to. If there are any interesting results, Marc will know what they are. Cannot wait.

[2] I’d like to thank Noam for allowing me to put this paper up for public consumption.

[3] Please observe that this does not imply that BP is the only property we might wish to investigate, though I agree with Chomsky that this is a pretty salient one. But say one were interested in how the phonological system arose, or the semantic system. The first step has to be to characterize the properties of the system one is interested. Only once this is done can evolutionary speculation fruitfully proceed. See here for further discussion, with an emphasis on phonology.

[4] It is worth noting that this is very fast in evolutionary terms and that if the time scale is roughly right then this seems to preclude a gradualist evolutionary story in terms of the slow accretion of selected features. Some seem to identify evolution with natural selection. As Chomsky notes (p. 11), Darwin himself did not assume this.

[5] Furthermore, we want whatever was added to be simple because it has not changed for the last 50k years. Let me say this another way: if what emerged 100kya was the product of slow moving evolutionary change with the system accreting complexity over time then why did this slow change stop so completely 50kya? Why didn’t change continue after the trek out of Africa? Why tones of change before hand and nothing since? If the change is simple, with not moving parts, as it were, then there is nothing in the core system to further evolve.

[6] I’ll write another post on these soon. I hope.

[7] If this reading of Chomsky’s intention here is correct, then I have interpreted him incorrectly in the past. Oh well, won’t be the last time. In fact, under this view, the linguistic system once evolved need not be particularly efficient computationally or otherwise.  On this view, computationally efficient seems to me “arose as a matter of natural law without the required intervention of natural selection.”

[8] The relevant passage is

May at the NSF

At the end of May I attended a terrific workshop organized by David Poeppel for the NSF.  Here’s a link to the roster and the background papers we were given. The confab was fun, largely because the free wheeling discussion was based on some uncontroversial givens, first and foremost among these being that all accepted that something like Marr’s view was a reasonable idealization of the kinds of levels required to link brain and mind. In particular, all bought into the view that cognitive neuroscience needs good high level descriptions of the computational competences that the brain has in order to understand how it is organized at the neural level.[1]  Indeed, Randy Gallistel (to only a very few grumbles) proposed a much stronger version of the Marr thesis: that brain architectures cannot be fruitfully studied at all in the absence of good computational level accounts. It was further accepted that models of linguistic competence of the generative variety are paradigmatic instances of such Marrian computational level theories. 

So, given this wonderful coming together of minds, what did I learn? Here are some recollected personal highlights.

First, We don’t actually know much about the neural bases of mental computation and so it is unreasonable to give neuronal level accounts a privileged status.  One of our pre-conference readings (c.f. Mausfield here) has the following juicy quote:

Given that we presently know next to nothing about the physical principles underlying mental phenomena and achievements, there is no reason to assign the level of neurons a privileged explanatory role. (p.4)

This was not a point avidly disputed by the neuroscientists present, despite the acknowledgment all round that neuroscience has made impressive progress over the last 25 years.  This is not incompatible with the realistic appraisal that there is a very long way to go and that, at this time, what we know about the brain offer few constraints on possible computational level theory (e.g. on linguistic proposals about the structure of FL/UG). We might wish that things were different (I know that I do) but they aren’t.

Evidence? Well, here’s one: it seems that for much simpler systems e.g. C. elegans with all of 302 neurons (whose wiring we know), why it does what it does is still a mystery.  As Mausfield observed: “In the case of C.elegans, the complete knowledge of the components of its biological hardware would constitute a particularly favorable situation for understanding its complex behavior…” Nonetheless, Mausfield quotes a recent review that notes that despite knowing all we might wish to know about the 302 neurons in these nematode brains these reductive efforts have proven quite unsatisfactory: “C.elegans responds behaviorally to the presence or absence of food in a plethora of ways…Surprisingly little progress has been made in understanding these responses” (p.3 note 2)).

Second, this relative ignorance is nothing new.  It seems that the goal of understanding mental phenomena in neurological terms has been a long-standing project, at least since the 18th century.  In other words, this is not a bold new surprising thesis, despite what some hyperventilating philosophers might suggest.  Again as Mausfield put it:

For over 200 years, the premise that mental processes must be considered a function of the brain has been more or less commonplace. This has deluded us into overlooking the fact that…our theoretical understanding is next to nil of what exactly…this function might actually be taken to be. (p.3)

It seems that Priestly (of chemical fame) already thought that this was the obvious scientific position to take (LaMettrie preceded him by a century or so). However, Priestly was considerably more modest than many of our current neuro-philosophers. His position was described as follows by the London Encyclopedia (1829):

Dr Priestly apprehends that sensation and thought necessarily result from the organization of the brain…but he professes to have no idea at all of the manner in which the power of perception results from organization and life.[2]

A becoming modesty, I think!

Third, historically, computational level theories have generally laid the groundwork for neural explorations rather than brain functions constraining higher level accounts. Once again here’s Mausfield:

 …advances in our psychological understanding of perceptual phenomena have in the first place benefited and fostered neurophysiology rather [than-sic] the other way around. (p.2)

The implications of this for Generative Grammar (GG) are pretty clear. GGs provide computational level accounts of the mental powers of native speakers; a descriptively adequate grammar of L describing the mental states of a competent speaker/hearer of L and an explanatory adequate theory of L describing how L derives from FL/UG given the PLD of L.  These computational level accounts, one hopes, will serve to guide neuro-scientific research. How? Well, in much the way that Barlow (quoted in Mausfield) envisioned for work in the psychology and physiology of perception:

As to the claim that a theoretical understanding of visual perception derives from neurophysiological investigations, Barlow (1983, p.11) emphasized: “Nothing could be more misleading, for all the important properties of the visual system were first established by psychophysical and psychological observations made on the system working as a whole. […] physiologists need to be told what the visual system does before they can set about the difficult task of finding out how it does it. (p.2)

Substitute FL for ‘the visual system’ above and you have more or less the current state of play in the cognitive neuroscience of language, at least as seen by the cohort of people that David managed to get to sit down together to discuss these matters.

Fourth, to the problem that Poeppel and Embick (P&E) (here and here) identified as the “granularity mismatch problem” is still with us.  In my presentation, I suggested that one of the virtues of the Minimalist Program (MP) is that it offers a way of bridging the divide that P&E identify. In particular, if we can really unify grammatical phenomena and reduce them to a common Merge-like core, then this will provide a convenient target for neurophysiological investigation: find a Merge-like circuit.[3] Thus, the grammatical project described (e.g. here) was not dismissed as irrelevant to finding ways of incarnating minds in brains.

Fifth, I got a great peek into how neuro/psycho types are thinking of basic operations in their respective domains, c.f. Dave Heeger (here) and Greg Hickok (here) for a pretty good taste of what they are doing. Dave Heeger’s talk had, what to my ear, was a real minimalistic theme: how “ set of canonical neural computations” shared across different “brain regions and modalities” could apply “similar operations to different problems” (51). In the best of all possible worlds, we would love to find something similar in cognitive domains including language. Greg’s talk showed how we might integrate higher level psych-linguistic descriptions of speech, with lower level articulatory motor control.  What to me was very exciting were the analogies between headedness in syntax and similar notions in motor plans, which appear to have similar kinds of structures. All of this was very speculative, and hence extremely interesting.

Sixth, Elisa Newport gave a fascinating presentation focusing on brain specialization for speech. She made (at least) two fascinating points. First, she observed that brains in which the language areas are compromised can redirect this function to other parts.  However, not to just any other part. Rather, a brain can redirect linguistic capacity from a left language impaired hemisphere to the very same place in the other hemisphere. This suggests two things: (i) that either hemisphere can adequately subserve language and (ii) that the same regions in both hemispheres are particularly well suited (specialized?) for the kinds of operations language demands. In other words, though young brains rearrange the cognitive furniture, not all parts of the brain are equally adept at filling in for missing linguistic capacity, viz. brains are labile but not arbitrarily so. Second, Elisa asked a terrific question: why is the language area located where it is? And she shot down one plausible answer: it sits between the perceptual regions that care about audition and the motor regions that move lips and tongue. This, Elisa noted, cannot be the whole answer for the exact same regions subserve ASL speakers (a point also made by Helen Neville), and ASL speakers don’t much worry about audition or lip/tongue movement.   

Last, Bob Berwick gave a great presentation on genetic differences between us and our Neanderthal cousins.  The main finding is that we are almost identical! There were very very few differences. In effect, the minimalist assumption (that whatever happened that allowed language to emerge was both rapid and genetically “minor”) seems on the right track. I am in the process to trying to convince Bob to post on this, so stay tuned.

There was much much more; great presentations, excellent lunches, a terrific supper, lots of discussion, jokes, arguments, speculations and general good cheer. However, most heartening of all was the realization that there is a reasonable group of people out there that have no problem with the standard (Chomskyan) Generative view that linguistics is an important part of the cognitive neurosciences. I can only only hope that this view becomes even more widespread. If it does, it suggests the coming of a golden age.

---

[1] Bob Berwick made the reasonable point that restricting matters to three levels is likely a radical simplification, apparent if one considers how many levels computer engineers postulate to get one from programming languages to machine code. So, the Marr perspective is best stated that there are at least three levels worthy of serious consideration.

[2] Quoted in Mausfield p.3.

[3] I actually proposed that we should look for two circuits: one that combines elements and one that labels the resulting combination. The former is plausibly cognitively generic while the second is my candidate for the real distinctive linguistically special operation.  However, the logic of MP does not require that my specific proposal be the right one (though, of course, I have no doubt that it is).