Another day, another paper. This time a rumination on Chomsky’s Syntactic Structures arguments about the autonomy of syntax. I think, despite Fritz Newmeyer’s excellent attempts to clear this issue up over many years, it’s still reflexively misunderstood by many people outside of generative grammar. Chomsky’s claim that syntax is autonomous is really just a claim that there is syntax. Not that there’s not semantics intimately connected to that syntax. Not that syntactic structures aren’t susceptible to frequency or processing effects in use. Just that syntax exists.

Current alternatives to the generative approach to dealing with language still, as far as I can tell, attempt to argue that syntactic phenomena can be reduced to some kind of stochastic effect, or to some kind of extra-linguistic cognitive semantic structures, or to both. This paper attempts to look at the kinds of arguments that Chomsky gave back in the 1950s and to examine whether the last 60 years have given us any evidence that the far more powerful stochastic and/or cognitive semantic systems now available can do the job, and eliminate syntax. I guess most people that know me will be unsurprised by my conclusion: even the jazziest up-to-the-minute neural net processors that Google uses still don’t come close to doing what a 3 year old child does, and even appealing to rich cognitive structures of the sort that there is good evidence for from cognitive psychology misses a trick when trying to explain even the simplest syntactic facts. I look at recent work by Tal Linzen and colleagues that shows that neural net learners may mimic some aspects of syntactic hierarchy, but fail to capture the syntactic dependencies that are sensitive to such structure. I then reprise and extend an argument that Peter Svenonius and I gave a few years back about bound variable pronouns.

One area where I do signal a disagreement with the Chomsky of 60 years ago is in the semantics of grammatical categories. Chomsky argued that these lack semantics, but, since my PhD thesis back in the early 1990s, I’ve been arguing that grammatical categories have interpretations. Here I try to show that the order of Merge of these categories is a side effect not of their interpretations, but of whether the kind of computational task they are put to is more easily handled with one order or the other.

The idea goes like this (excerpted from section 4 of the paper).

“Take an example like the following:

(20) a. Those three green balls

b. *Those green three balls

As is well known, the order of the demonstrative, numeral and descriptive adjective in a noun phrase follow quite specific typological patterns arguing for a hierarchy where the adjective occurs closest to the noun, the numeral occurs further away and the demonstrative is most distant (Greenberg 1963, Cinque 2005). Why should this be? It seems implausible for this phenomenon to appeal to a mereological semantic structure. I’d like to propose a different way of thinking about this that relies on the way that a purely autonomous syntax interfaces with the systems of thought. Imagine we have a bowl which has red and green ping pong balls in it. Assume a task (a non-linguistic task) which is to identify a particular group of three green balls. Two computations will allow success in this task:

(21) a. select all the green balls

b. take all subsets of three of the output of (a)

c. identify one such subset.

(22) a. take all subsets of three balls

b. for each subset, select only those that have green balls in them

c. identify one such subset

Both of these computations achieve the desired result. However, there is clearly a difference in the complexity of each. The second computation requires holding in memory a multidimensional array of all the subsets of three balls, and then computing which of these subsets involve only green balls.

The first simply separates out all the green balls, and then takes a much smaller partitioning of these into subsets involving three. So applying the semantic function of colour before that of counting is a less resource intensive computation. Of course, this kind of computation is not specific to colour—the same argument can be made for many of the kinds of properties of items that are encoded by intersective and subsective adjectives.

If such an approach can be generalized, then there is no need to fix the order of adjectival vs. numeral modifiers in the noun phrase as part of an autonomous system. It is the interface between a computational system that delivers a hierarchy, and the use to which that system is put in an independent computational task of identifying referents, plus a principle that favours systems that minimize computation, that leads to the final organization. The syntax reifies the simpler computation via a hierarchy of categories.

This means that one need not stipulate the order in UG, nor, in fact, derive the order from the input. The content and hierarchical sequence of the elements in the syntax is delivered by the interface between two distinct systems. This can take place over developmental timescales, and is, of course, likely to be reinforced by the linguistic input, though not determined by it.

Orders that are not isomorphic to the easiest computations are allowed by UG, but are pruned away during development because the system ossifies the simpler computation. Such an explanation relies on a generative system that provides the structure which the semantic systems fill with content.

The full ordering of the content of elements in a syntactic hierarchy presumably involves a multiplicity of sub ordering effects, some due to differences in what variable is being elaborated as in Ramchand and Svenonius’s proposal, others, if my sketch of an approach to the noun phrase is correct, due to an overall minimizing of the computation of the use of the structure in referring, describing, presenting etc. In this approach, the job of the core syntactic principles is to create structures which have an unbounded hierarchical depth and which are composed of discrete elements combined in particular ways. But the job of populating these structures with content is delegated to how they interface with other systems.”

The rest of the paper goes on to argue that even though the content of the categories that syntax works with may very well come from language external systems, how they are coopted by the linguistics system, and which content is so coopted, still means that there is strong autonomy of syntax.

The paper, which is to appear in a volume marking the 60th anniversary of the publication of syntactic structures is on Lingbuzz here.