The SMT again

Two revisions below thx to Ewan spring a thinko, a slip of the mind.

I have recently urged that we adopt a particular understanding of the Strong Minimalist Thesis (SMT) (here).  The version that I favor treats the SMT as a thesis about systems that use grammars and suggests that central features of the grammatical representations that they use will be crucial to explaining why they are efficient. If this proves to be doable, then it is reasonable to describe FL and the grammars it makes available as “well designed” and “computationally efficient.” Stealing from Bob Berwick (here), I will take parsing efficiency to mean real time parsing and (real time) acquisition to mean easy acquisition given the PLD.  Put this all together and the SMT is the conjecture that the grammatical format of Gs and UG is critical to allowing parsers, acquirers, producers, etc. to be very good at what they do (i.e. to be well-designed). On this view, grammars are “well designed” or “computationally efficient” in virtue of having properties that allow their users to be good at what they do when such grammars are embedded transparently within these systems.

One particularly attractive virtue of this interpretation (for me) is that I understand how I could go about empirically investigating it.  I confess that this is not true for other versions of the SMT that talk about neat fits between grammar and the CI interface, for example. So far as I can tell, we know rather little about the CI interface and so the question of fit is, at best, premature. On the other hand we do know a bit about how parsing works and how acquisition proceeds so we have something to fit the grammar to.[1]

So how to proceed? In two steps I believe. The first is to see if use systems (e.g. parsers) actually deploy grammars in real time, i.e. as they parse. Thus, if it is true that the basic features of grammatical representations are responsible for how (e.g.) parsers manage to efficiently do what they do then we should find real time evidence implicating these representations in real time parsing. Second, we should look for how exactly the implicated features manage to make things so efficient. Thus, we should look for theoretical reasons for why parsers that transparently embody, say, Subjacency like principles, would be efficient.  Let me discuss each of these points in turn.

There is increasing evidence from psycho-ling research indicating that real time parsing respects grammatical distinctions, even very subtle ones.  Colin Phillips is a leader in this kind of work and he and his (ex) students (e.g. Brian Dillon, Matt Wagers, Ellen Lau, Dave Kush, Masaya Yoshida) have produced a body of work that demonstrates how very closely parsers respect grammatical conditions like islands, c-command, and local binding domains. And by closely I mean very closely.  So, for example, Colin shows (here) that online parsing respects the grammatical conditions that license parasitic gaps. So, not only do parsers respect islands, but they even treat configurations where island effects are amnestied as if they were not islands. Thus, parsers respect both the general conditions that grammars lay down regarding islands and the exceptions to these general conditions that grammars allow. This is what I mean by ‘close.’

There is a recent excellent demonstration of this from Masaya Yoshida, Lauren Ackerman, Morgan Purier and Rebekah Ward (YLPW) (here are slides from a recent CUNY talk).[2] YLPW analyzes the processing of backward sluicing constructions like (1):

(1)  I don’t recall which writer, but the editor notified a writer about a new project

There is an ellipsis “gap” right after which writer that is redeemed by anchoring it to a writer in the following sentence. What YLPW is looking to determine is whether the elided gap site is sensitive to online parsing effects. YLPW uses a plausibility effect as probe as follows.

First, it is well known that a wh in CP triggers an active search for a verb/gap that will give it an interpretation. ‘Active’ here means that the parser uses a top down predictive process and is eagerly looking to link the wh to a predicate without first consulting bottom information that would indicate the link to be ill-advised. YLPW show that the eagerness to “fill a gap” is as true for implicit gaps within ellipsis sites as it is for “real” gaps in regular wh sentences.  YLPW shows this by demonstrating a plausibility effect slowdown in sentences like (2a) parallel to the ones found in (2b):

(2)  a. I don’t remember which writer/which book, but the editor notified a writer about a new book

b. I don’t remember which writer/which book the editor notified GAP about a new book

When the wh is which book then there is a significant pause at notified in both sentences in (2), as contrasted with the same sentences where which writer is the antecedent of the gap.  This is because parsers, we know, greedily try and relate the wh to the first syntactically available position encountered and in the case of which book the wh is not a plausible filler of the gap and the attempted filling results in a little lingering about the verb (*notify this book about…). If the antecedent is which writer no such pause occurs, for obvious reasons.  The plausibility effect, then, is just a version of the well-known filled gap effect, with a little semantic kicker to add some frisson. At any rate, the first important discovery is that we find the same plausibility effect in both (2a) with the gap inside a sluiced ellipsis site, and (2b) where the gap is “overt.”

The next step is to see if this plausibility/filled gap effect slowdown occurs when the relevant antecedent for the sluiced ellipsis site is inside an island. It is well known that ellipsis is not subject to island restrictions. Thus, if the parser cleaves tightly to the distinctions the grammar makes (as the SMT would lead us to expect) then we should find plausibility slowdowns except when the gap is inside an ellipsis site for the latter are not subject to island restrictions [and so should induce filled gap/plausibility effects (added: thx Ewan)].  And that’s exactly what YLPW find. Though plausibility effects are not found at notified in cases like (3) they are found in cases like (4) where the “gap” is inside a sluice sight.

(3)  I don’t remember which book [the editor who notified the publisher about some science book] had recommended to me

(4)  I don’t remember which book, but [the editor who notified the publisher about some science book] recommended a new book to me

This is just what we expect from a parser that transparently embeds a UG like grammar that treats movement but not ellipsis as a product of (long) movement.

The conclusion: it seems that parsers make just the distinctions that grammars make when they parse in real time, just as the SMT would lead us to expect.

So, there is growing evidence that parsers transparently embed UG like grammars.  This readies us for the second step. Why should they do so?  Here, there is less current research that bears on the issue. However, there is work from the 80s by Mitch Marcus, Bob Berwick and Amy Weinberg that showed that a Marcus style parser that incorporated grammatical features like Subjacency (and, interestingly, Extension) could parse sentences efficiently (effectively, in real time).  This is just what the doctor ordered. It goes without saying (though I will say it) that this work needs updating to bear more directly on the SMT and minimalist accounts of FL. However, it provides a useful paradigm of how one might go about connecting the discoveries concerning online parsing with computational questions of parsing efficiency and their relationship to central architectural features of FL/UG.

The SMT is a bold conjecture. Indeed, it is likely false, at least in fine detail. This does not, however, detract from its programmatic utility.  The fact is that there is currently lots of research that can be understood as bearing on its accuracy and that fruitfully brings together work in syntax, psycholinguistics and computational linguistics.  The SMT, in other words, is a terrific hypothesis that will generate fascinating work regardless of its ultimate empirical fate.  That’s what we want from a research program and that’s something that the Strong Minimalist Thesis is ready to deliver. Were this all that the Minimalist Program provided, it would have been enough (dayenu!). There is more, but for the nonce, this is more than enough. Yay, for the Minimalist Program!!!

---

[1] Let me modulate this: we know something about some other parts, see here for discussion of magnitude estimation in the visual domain. Note that this discussion fits well with the version of the SMT deployed here precisely because we know something about how this part of the visual system works. We cannot say as much about most of the other parts of CI. Indeed, we don’t really know how many “parts” CI has.

[2] They are running some more experiments, so this work is not yet finished. Nonetheless, it illustrates the relevant point well, and it is really fun stuff.

DTC redux

Syntacticians have effectively used just one kind of probe to investigate the structure of FL, viz. acceptability judgments. These come in two varieties: (i) simple “sounds good/sounds bad” ratings, with possible gradations of each (effectively a 6ish point scale ok, ?, ??, ?*, *, **), and (ii) “sounds good/sounds bad under this interpretation” ratings (again with possible gradations). This rather crude empirical instrument has proven to be very effective as the non-trivial nature of our theoretical accounts indicates.[1] Nowadays, this method has been partially systematized under the name “experimental syntax.” But, IMO, with a few important conspicuous exceptions, these more refined rating methods have effectively endorsed what we knew before. In short, the precision has been useful, but not revolutionary.[2]

In the early heady days of Generative Grammar (GG), there was an attempt to find other ways of probing grammatical structure. Psychologists (following the lead that Chomsky and Miller (1963) (C&M) suggested) took grammatical models and tried to correlate them with measures involving things like parsing complexity or rate of acquisition. The idea was a simple and appealing one: more complex grammatical structures should be more difficult to use than less complex ones and so measures involving language use (e.g. how long it takes to parse/learn something) might tell us something about grammatical structure. C&M contains the simplest version of this suggestion, the now infamous Derivational Theory of Complexity (DTC). The idea was that there was a transparent (i.e. at least a homomorphic) relation between the rules required to generate a sentence and the rules used to parse it and so parsing complexity could be used to probe grammatical structure.

Though appealing, this simple picture can (and many believed did) go wrong in very many ways (see Berwick and Weinberg 1983 (BW) here for a discussion of several).[3] Most simply, even if it is correct that there is a tight relation between the competence grammar and the one used for parsing (which there need not be, though in practice there often is, e.g. the Marcus Parser) the effects of this algorithmic complexity need not show up in the usual temporal measures of complexity, e.g. how long it takes to parse a sentence. One important reason for this is that parsers need not apply their operations serially and so the supposition that every algorithmic step takes one time step is just one reasonable assumption among many. So, even if there is a strong transparency between competence Gs and the Gs parsers actually deploy, no straightforward measureable time prediction follows.

This said, there remains something very appealing about DTC reasoning (after all, it’s always nice to have different kinds of data converging on the same conclusion, i.e. Whewell’s consilience) and though it’s true that the DTC need not be true, it might be worth looking for places where the reasoning succeeds. In other words, though the failure of DTC style reasoning need not in and of itself imply defects in the competence theory used, a successful DTC style argument can tell us a lot about FL. And because there are many ways for a DTC style explanation to fail and only a few ways that it can succeed, successful stories if they exist can shed interesting light on the basic structure of FL.

I mention this for two reasons. First, I have been reading some reviews of the early DTC literature and have come to believe that its demonstrated empirical “failures” were likely oversold. And second, it seems that the simplicity of MP grammars has made it attractive to go back and look for more cases of DTC phenomena. Let me elaborate on each point a bit.

First, the apparent demise of the DTC. Chapter 5 of Colin Phillips’ thesis (here) reviews the classical arguments against the DTC.  Fodor, Bever and Garrett (in their 1974 text) served as the three horsemen of the DTC apocalypse. They interned the DTC by arguing that the evidence for it was inconclusive. There was also some experimental evidence against it (BW note the particular importance of Slobin (1966)). Colin’s review goes a very long way in challenging this pessimistic conclusion. He sums up his in depth review as follows (p.266):

…the received view that the initially corroborating experimental evidence for the DTC was subsequently discredited is far from an accurate summary of what happened. It is true that some of the experiments required reinterpretation, but this never amounted to a serious challenge to the DTC, and sometimes even lent stronger support to the DTC than the original authors claimed.

In sum, Colin’s review strongly implies that linguists should not have abandoned the DTC so quickly.[4] Why, after all, give up on an interesting hypothesis, just because of a few counter-examples, especially ones that when considered carefully seem on the weak side? In retrospect, it looks like the abandonment of the strong hypothesis was less a matter of reasonable retreat in the face of overwhelming evidence than a decision that disciplines occasionally make to leave one another alone for self-interested reasons. With the demise of the DTC, linguists could assure themselves that they could stick to their investigative methods and didn’t have to learn much psychology and psychologists could concentrate on their experimental methods and stay happily ignorant of any linguistics. The DTC directly threatened this comfortable “live and let live” world and perhaps this is why its demise was so quickly embraced

by all sides.

This state of comfortable isolation is now under threat, happily.  This is so for several reasons. First, some kind of DTC reasoning is really the only game in town in cog-neuro. Here’s Alec Marantz’s take:

...the “derivational theory of complexity” … is just the name for a standard methodology (perhaps the dominant methodology) in cognitive neuroscience (431).

Alec rightly concludes that given the standard view within GG that what linguists describe are real mental structures, there is no choice but to accept some version of the DTC as the null hypothesis. Why? Because, ceteris paribus:

…the more complex a representation- the longer and more complex the linguistic computations necessary to generate the representation- the longer it should take for a subject to perform any task involving the representation and the more activity should be observed in the subject’s brain in areas associated with creating or accessing the representation or performing the task (439).

This conclusion strikes me as both obviously true and salutary, with one caveat. As BW has shown us, the ceteris paribus clause can in practice be quite important.  Thus, the common indicators of complexity (e.g. time measures) may be only indirectly related to algorithmic complexity. This said, GG is (or should be) committed to the view that algorithmic complexity reflects generative complexity and that we should be able to find behavioral or neural correlates of this (e.g. Dehaene’s work (discussed here) in which BOLD responses were seen to track phrasal complexity in pretty much a linear fashion or Forster’s work finding temporal correlates mentioned in note 4).

Alec (439) makes an additional, IMO correct and important, observation. Minimalism in particular, “in denying multiple routes to linguistic representations,” is committed to some kind of DTC thinking.[5] Furthermore, by emphasizing the centrality of interface conditions to the investigation of FL, Minimalism has embraced the idea that how linguistic knowledge is used should reveal a great deal about what it is. In fact, as I’ve argued elsewhere, this is how I would like to understand the “strong minimalist thesis,” (SMT) at least in part. I have suggested that we interpret the SMT as committed to a strong “transparency hypothesis” (TH) (in the sense of Berwick & Weinberg), a proposal that can only be systematically elaborated by how linguistic knowledge is used.

Happily, IMO, paradigm examples of how to exploit “use” and TH to probe the representational format of FL are now emerging. I’ve already discussed how Pietroski, Hunter, Lidz and Halberda’s work relates to the SMT (e.g. here and here). But there is other stuff too of obvious relevance: e.g. BW’s early work on parsing and Subjacency (aka Phase Theory) and Colin’s work on how islands are evident in incremental sentence processing. This work is the tip of an increasingly impressive iceberg. For example, there is analogous work showing that that parsing exploits binding restrictions incrementally during processing (e.g. by Dillon, Sturt, Kush).

This latter work is interesting for two reasons. It validates results that syntacticians have independently arrived at using other methods (which, to re-emphasize, is always worth doing on methodological grounds). And, perhaps even more importantly, it has started raising serious questions for syntactic and semantic theory proper. This is not the place to discuss this in detail (I’m planning another post dedicated to this point), but it is worth noting that given certain reasonable assumptions about what memory is like in humans and how it functions in, among other areas, incremental parsing, the results on the online processing of binding noted above suggest that binding is not stated in terms of c-command but some other notion that mimics its effects.

Let me say a touch more about the argument form, as it is both subtle and interesting. It has the following structure: (i) we have evidence of c-command effects in the domain of incremental binding, (ii) we have evidence that the kind of memory we use in parsing cannot easily code a c-command restriction, thus (iii) what the parsing Grammar (G) employs is not c-command per se but another notion compatible with this sort of memory architecture (e.g. clausemate or phasemate). But, (iv) if we adopt a strong SMT/TH (as we should), (iii) implies that c-command is absent from the competence G as well as the parsing G. In short, the TH interpretation of SMT in this context argues in favor of a revamped version of Binding Theory in which FL eschews c-command as a basic relation. The interest of this kind of argument should be evident, biut let me spell it out. We S-types are starting to face the very interesting prospect that figuring out how grammatical information is used at the interfaces will help us choose among alternative competence theories by placing interface constraints on the admissible primitives. In other words, here we see a non-trivial consequence of Bare Output Conditions on the shape of the grammar. Yessss!!!

We live in exciting times. The SMT (in the guise of TH) conceptually moves DTC-like considerations to the center of theory evaluation. Additionally, we now have some useful parade cases in which this kind of reasoning has been insightfully deployed (and which, thereby, provide templates for further mimicking). If so, we should expect that these kinds of considerations and methods will soon become part of every good syntactician’s armamentarium.

---

[1] The fact that such crude data can be used so effectively is itself quite remarkable. This speaks to the robustness of the system being studied for such weak signals should not be expected to be so useful otherwise.

[2] Which is not to say that such more careful methods don’t have their place. There are some cases where being more careful has proven useful. I think that Jon Sprouse has given the most careful thought to these questions. Here is an example of some work where I think that the extra care has proven to be useful.

[3] I have not been able to find a public version of the paper.

[4] BW note that Forster provided evidence in favor of the DTC even as Fodor et. al. were in the process of burying it. Forster effectively found temporal measures of psychological complexity that tracked the grammatical complexity the DTC identified by switching the experimental task a little (viz. he used an RSVP presentation of the relevant data).

[5] I believe that what Alec intends here is that in a theory where the only real operation is merge then complexity is easy to measure and there are pretty clear predictions of how this should impact algorithms that use this information. It is worth noting that the heyday of the DTC was in a world where complexity was largely a matter of how many transformations applied to derive a surface form. We have returned to that world again, though with a vastly simpler transformational component.

Complexity redux redux

Alex has had some interesting comments on an earlier post. I thought that I would move the discussion from the comments to the main text.  Thanks Alex.

I think that I have found the nub of my disagreement with Alex. It can be localized to the first sentence of paragraph 2 of his last comment. 

We don't know what the grammars are so *of course* we have to abstract away from the properties of grammars.

No, No, No!!! We know a great deal about grammars. What I want from computational theories/considerations is an explanation of why what I know to be the case has the properties it does.  For Alex, grammatical properties have to earn their computational keep by solving antecedent problems of various sorts.  For me the shoe is on the other foot: I am interested in computational theories to the degree that they earn their linguistic keep by enlightening me as to why grammars have the properties I know that they have. For example: Why are there locality restrictions like minimality?  Why are the phase heads v, C and D?  Why do antecedents of moved elements c-command their launch sites?  Note that these why questions all presuppose that there are locality restrictions like minimality, and "subjacency." That there is displacement and that c-command describes a required grammatical restriction etc.  I go further: I think that we KNOW that the structure of FL/UG are roughly described by the generalizations outlined in LGB. This I know. What I don't know is why, and here is where I look to computational considerations to possibly enlighten me. If such enlightenment is forthcoming, great. If not, I look somewhere else. This opportunism has led me to the tentative conclusion that much of what goes on in CS is beside the point.  Oddly some of what goes on in software design is not.  So, maybe the right way of thinking of computation in linguistics should look to what they are doing in software engineering rather than in CS departments.  I am not dogmatic about this for after all I am being deliberately opportunistic (a variant on Tom Lehrer’s Lobyachevsky strategy).  So were there to be something interesting coming from core CS considerations, I would be happy to embrace the results. Here I am open minded. Where I am not open-minded concerns whether we know what FL/UG (roughly) looks like. We do. There’s more to discover, but we know a lot and should NOT set this knowledge aside in doing our computational investigations.

My impression is that this is not how Alex see things. What he wants is a computational theory able to handle linguistic data without assuming that these properties of grammar hold. He is skeptical about them and so wants to finesse them. I am convinced they are true and so they form boundary conditions on further discussion.  We have made different judgments and they determine what we are looking for. Much of our disagreement stems from this, I believe.

There is one further source: we have very different conceptions concerning the utility of formalization.  I am not sold on the view that without formalization science grinds to a halt. Indeed, I might go further, it often impedes. I was recently reminded of the fact that physics used the calculus quite effectively for about 200 years before it was put on solid foundations (Cauchy-Wierstrass). Till then it was riddled with all sorts of foundational problems that philosophers and mathematicians were eager to parade and probe.  Doing so was, I have been told, very central to work on physics in England post Newton.  The problems were ignored on the continent. Consequently, physics thrived on the continent and died in England for about 100 years.  It seems that despite the foundational flaws lots of good work could be done and was done. 

The tools we use in syntax are clear enough in my view and formalization has (putting a few exceptions aside) yet to reveal anything very deep, at least in syntax.  I have nothing against people working on doing this, but I don't really see this as a pressing need. Our problems in linguistics don't arise, I believe, because we misunderstand our tools.  Nor, have I so far seen many deep results arise from the various formalizations that have been wrought. Formalizations can be useful, but are by no means required. Indeed, my impression from Stabler’s work is that should one want to formalize GB or Minimalism or GPSG or whatever it seems doable.  The question then is not if it can be done, but why, and here I am less sanguine than Alex seems to be that it is worth the effort.

Very last point: Alex notes that the ease of parsing long sentences needs explanation. Yup. In fact, Berwick and Weinberg in other work provide one. Certain left corner parsers seem to do the trick. However, as BW show, these require bounded left contexts. Interestingly, as they show, grammars that embed something like phases/subjacency/barriers deliver a grammar format of the right type, one that snugly fits into an efficient left corner parser. Now, I don't know if left corner parsers are the "right" parsing theory. BW’s was deterministic for example and whether grammars are deterministic is contentious, I am told. But I do like the style of argument: it provides a nice computational reason for a certain kind of grammatical code. One might be tempted to say that a grammar would be well designed (i.e. the code would be “efficient,” “optimal”) to efficiently parse language, if it had  subjacency/phases/barrier like restrictions as a design characteristic. I like this kind of computational explanation.  It aims to explain a known design feature of FL/UG in computational terms. We need more results like this. Get to work Alex!