Represent!

I’ve just started reading Phil Brannigan’s Provocative Syntax, which is very provocative. In that it provokes. Syntax!

OK, but seriously Dr. Brannigan, if you didn’t want these kinds of cheap puns made, you should’ve named your theory something less, erm, provocative.

OK, OK, I’m done. I swear.

I’m very much looking forward to this, mostly because my recent poster presentation at the LSA with Steve argues sort of the opposite. Brannigan seems to be defending bottom-up chain formation mediated through phases. We argued for top-down reconstruction, no explicit creation of chains, and no phases or multiple spellout.

Of course, since (Epstein and Seeley 2006), phase-based derivations have been under strong attack – largely because of the interminable EPP mystery. That is, a lot of languages seem to need a word in subject position for a sentence to be grammatical, even if that word is just a dummy word like it or there. Since no one can think of any reason why this should be so, most syntactic theories just stipulate it. That is, they say something like “the specifier of IP must be filled in language x.”

Minimalist theories – certainly not those that use Bare Phrase Structure – don’t really have the tools to state that kind of requirement, though. So, instead, EPP has been translated into a feature. Certain heads have a feature that requires that some suitable item occupy their specifier, and if they don’t get it, the derivation crashes. And of course, now that this is a feature (rather than a constraint on legal phrase structure rules), it gets used to fill other holes in the theory beyond just ensuring that subjects surface. As Epstein and Seely have pointed out, though, it isn’t clear that such a feature is needed in all the cases it’s been leveraged to address, that in many ways leaning on an EPP feature to provide an explanation raises more empirical questions than it answers, and that in any case the theoretical status of the feature is unclear. It’s entirely theory-internal; to the extent that it’s based on real evidence, it describes, rather than accounts for, the data.

Naturally a person like Brannigan who wants to save phases is going to have to deal with the EPP. Trying to explain why the EPP doesn’t overgenerate sentences like “*There seems a man to be in the room” is one of the primary motivations for proposing phases in the first place. So he’ll have to either come up with an answer to Epstein and Seely or find a suitable replacement for the EPP.

And he is definitely aware of this responsibility: it takes less than 4 pages in to the first chapter for the EPP to feature prominently.

Now, here’s my bone to pick. On page 4, Brannigan writes the following:

Notice that even in this model, the feature EPP feature remains representational, in the sense that it is to be deleted only when the phrase marker provides something to occupy the specifier position for the probe in question. In other words, the right structure must first be formed, and then it is examined to ensure that the probe head has a specifier. If the phrase marker (representation) fails this test, the EPP feature remains intact, crashing the derivation. As Roberts and Roussou (1998) observe, positing this type of feature to characterize movement is inherently non-explanatory, since the feature does little more than point to the effect of movement.

That word “representational” sticks in my craw a bit. I’m not sure that this is a fair way to characterize the issue.

It’s true that the original formulation of the EPP was “representational.” It was a constraint on the kinds of phrase structure rules that a theory was allowed to have, and it operated basically like a filter – saying that trees of a certain architecture (namely those that had a specifier-less IP projection) were illicit.

In that kind of narrow definition, I don’t think the Minimalist concept of the EPP as a feature is “representational.” It’s rather “algorithmic,” saying that if a certain operation does not succeed at a particular point in a derivation, the derivation will crash.

Does that amount to the same thing?

Well, in some ways it does. It’s a constraint on architecture in the sense that it predicts that every grammatical structure in a langauge that respects the EPP will have members of a certain lexical class at specific points in the hierarchy. It’s hard to say that this doesn’t act like a representational definition, even if it isn’t formulated like one.

But in other ways it doesn’t. A representational rule has a more global feel to it. That is, there are these and those structures, and because these look like that and those don’t, these are good and those aren’t. That isn’t how this is formulated. In fact, this does’t make reference to structure at all. It’s true that it has a very restricted kind of structural side-effect – which is interesting, and I’ll come back to that in a bit – but as stated, it isn’t really any different from a lexical requirement that a particular head have an external argument. Merge only happens to check features, and this is one of those cases.

The reason I’m splitting this particular hair is that I think the EPP is actually a good example of where the Minimalist Program functions as Chomsky intended: that is, where attempts to reduce the operations of narrow syntax to those that satisfy “bare output conditions” in the least computationally burdensome way lead us to new insights.

I’m not claiming that we’re any closer to understanding what the EPP is or why languages have it (or don’t). It’s still quite mysterious; we still have to deal with it in ways that are frustratingly stipulative. But trying to rework the EPP as a local featural requirement (as opposed to global configurational requirement, which we’ve banned from the pallette of acceptable grammar construction tools) at least illuminates in more detail what it is, precisely, that’s mysterious about the EPP, and this is a good pointer to solutions.

First, the EPP is mysterious because it behaves like a selection requirement with no semantic import. This is something that’s been said about it for a long time, but restyling it as a feature on heads makes the point clearer, I think. Now that we can’t think about it as a global constraint on trees, we’re forced to think of it as a lexical selectional requirement. THAT is interesting. Why would languages have this kind of constraint on their lexicons? What purpose does it serve? Most of all, why is it so comprehensive? Usually for lexical phenomena, there will be exceptions, and yet there don’t seem to be any for the EPP. In this way, it casts some doubt on the whole Minimalist enterprise: should we really be attempting to state rules that appear to be global in local terms?

Second, the EPP is mysterious because it seems to be a constraint on pronounciation. This is where I think Minimalism has some unique new light to shed on the issue. Since in Minimalism we are able to state rules that are purely concerned with pronounciation order in theoretically satisfying ways, and since Minimalism has already given us a better understanding of the disconnect between structure and sequential order, this aspect of the EPP has a more natural expression. And it is interesting too. Why would there be a requirement that states, essentially, that a word of a particular class must precede a particular head in the final sequential output? And yet, that does seem to be what it says – because there is precious little evidence for the operation of the EPP in the covert component (that is, in fact, one of the more successful arguments against it). In more representational theories, the EPP was purely hierarchical; it was stated in terms of the hierarchy as a constraint on hierarchy. But it doesn’t seem to have anything to do with hierarchy – one of its hallmarks being that it has no semantic reflex. Before Minimalism, it would have been hard to notice this. Now that we think of structure-building patterns as universal, differing accross languages more in terms of whether they have a pronounciation reflex rather than whether they happen at all, it is easier to see the EPP is a pronounciation-directed operation.

On second thought, maybe Brannigan is doing everyone a great service by using the word “representational.” By doing so, he’s highlighting what is surely the central modern concern about the EPP. Namely:

Given that we’ve had so much success porting our representational theory to a non-representational framework, it is reasonable to hope that non-representational explanations are available for every phenomenon we wish to capture. The central (modern) question about the EPP becomes: what would cause a non-representational phenomenon to behave as though it were representational?

- Joshua

[cross-posted at The Only Winning Move]

Spell THIS out

The more I get in to Minimalism, the more I find I like it. I started with some pretty strong prejudices against it, but I realize now that those were issues I had with Government and Binding Theory, and that the Minimalist Program exists largely to fix those issues.

That’s not to say that I buy the program hook, line and sinker. It’s more that we dovetail. Minimalism’s issues with Government and Binding are my issues, but it’s still obvious in a lot of areas that what I understand a theory of Grammar to be and what the Minimalist Program understands by that are not the same thing.

One such example came up today in an email exchange with Joel Wallenberg. We were talking about how to adapt Pronounce (my contribution to formal Syntax – the Sequentialization algorithm that interprets the tree for PF) to deal with phases. And that’s just the thing – I’ve never been a big fan of phases. Partly for the reasons that Epstein and Seely spell out in Derivations in Minimalism (condensed version: phases are a purely theoretical construct designed to solve a problem that seems empirical but is in fact, on closer observation, a problem with the theory – basically, that Chomsky is creating them to solve a problem that he created as a side-effect of trying to solve another problem). But mostly just because I can’t really make sense of what I think most researchers mean by “Multiple Spell-Out.”

This is something that Juan Uriagereka came up with and made public in 1999. The idea, as far as I can tell, is to enforce a kind of no-tampering condition on parts of the derivation that are “finished.” So, in a bottom-up, word-by-word theory driven entirely by adding a lexical item to a partially-assembled structure – which is what Minimalism assumes – that allows one to copy and “re-Merge” items or phrases that have been combined at earlier stages (this is Minimalism’s version of a “transformation” – note how non-transformative it actually is), we sometimes find we want to limit the domain of choices for “re-Merge.” So, Juan invented “Multiple Spell-Out.”

Anyone familiar with Minimalism knows the drill: the central question is to what extent can thinking about grammar as a system that it only concerned with legibility at the interfaces lead us to rethink the theory in ways that reduce the computational burden. So we start by assuming that there are only the motor-articulatory interface and the logical-semantic interface, and we go from there. Building up a syntactic structure means building up a structure that will serve as input – “instructions” – to both the articulatory and the semantic systems – and in this way the pairing of sound and meaning that’s central to language is accounted for.

Now, one’s first thought would be that you first assemble a complete structure, and then, to the extent that that structure obeys the rules of the language (= the principles of Universal Grammar parameterized in the particular way that is the language we’re talking about), PF (the articulatory interface) and LF (the logical-semantic interface) interpret the structure in their own terms. Sending this structure to PF and LF for interpretation is called “Spell-Out.”

Now, that much makes sense to me. It’s not exactly how I think about grammar, though. Rather this is a way of thinking about grammar that speaks to Chomsky’s Biolinguistic concerns.

I myself think of grammar in much more traditional/computational terms. For me, a language is a set of strings – in the case of natural language, it’s a set of strings of morphemes. So, a given language has a bag of morphemes (words, endings, etc.), and the potential members of the language are any sequence/assembly of these morphemes of any length. Obviously, trying to list the legal sequences is hopeless: there is reason to believe that not only is the number potential/candidate members of this set infinite, but the actual membership in the set is infinite too. So, having despaired of trying to simply list them, we try instead to come up with a computer program that generates them. Specifying the rules according to which the generative engine operates is the job of Syntax.

Unfortunately – thinking of Syntax as trying to characterize membership in a(n infinite) set of possible strings of morphemes turns out to be a handicap for understanding multiple spell-out. Or, rather, understanding it the way that it’s normally talked about.

Normally, when people write about multiple spell-out, they don’t really spell out the mechanics, but what they all seem to have in mind is that at not one but several stages in the derivation narrow syntax sends information to the interfaces. But I don’t understand what this means!!!

For me, the job is – given a sequence of morphemes – to say (a) whether it is a member of the set(= the language in question) and possibly also (b) whether this sequence is paired with one or multiple interpretations. Ok, well, given that that’s the case, and given that the question is about the complete sequence, and given that for at least some sequences (those with long-distance wh-movement, for example) characteristics from all parts of the sequence need to be inspected to answer the question, then surely it’s nonsensical to talk about part of the sequence having been sent to the interfaces for interpretation!

The way I see it, sequences are atomic with respect to the narrow question of whether they belong in the set. That is, the question is being asked of the whole sequence, not parts of it. We break the sequence into parts to analyze it because this is the only way that we know to account for the discrete infinity of the set. But it’s still the entire sequence that has a pronounciation, and it’s still the entire sequence that has an interpretation.

I suppose the idea is that the bits that get sent to the interfaces at various parts of the derivation are taken to be complete, and that they compose with each other to make the whole. That I can buy, but only in the narrow syntactic sense. These parts do not have self-contained interpretations and pronounciations. They may (in fact there is strong evidence they do) have self-contained syntactic life.

So I prefer to think of multiple spellout like this: spellout in this case just means that a particular subcomponent of a derviation is not (fully) accessible to the derivation after a certain point. After a certain point, the system can recognize that a particular phrase is no longer going to change. Or, in Minimalist terms, elements from it will no longer feed extensions of the tree being assembled. Any elements it contains are now syntactically inert. It is ready for interpretation, though how it combines with the rest of the structure might not be (it might re-merge at some later stage, thus participating in multiple places in the hierarchy). So for me – and man is this anticlimatic after all that background – multiple spell-out only makes sense if I think of it as the treatment of a complete phrase as though it were a single lexcial item for the remainder of the derivation. Itis, however, still present for interpretation at tne end of the derivation. That is, PF and LF don’t “read” it, and Pronoucne doesn’t try to sequentialize it (crucial!!!) until after all syntactic operations required to assemble this particular candidate member of the set have stopped. Sequentialization (linearization) and semantic interpretation wait until it’s all through.

- Joshua