Sunday, January 26, 2014

Three kinds of syntactic research

Robert Chametzky has some useful things to say about the absence of theoretical work within syntax that I labored to make clear in an earlier post (here). He distinguishes metatheoretical, theoretical and analytic work, the last being by far the predominant type of research. All three are valuable, but, as a matter of fact, the third is what predominates in syntax and is what is generally, inaptly, called “theoretical.” Here is Rob’s tripartite typology, pp. xvii-xix from his sadly under read book A Theory of Phrase Markers and the Extended Base available here. I transcribe, with some indicated omissions, the relevant two pages immediately below.

There are three sorts of work that can generally be distinguished in empirical inquiry. One is metatheoretical, a second is theoretical and the third is analytic. As is often the case, the boundaries are not sharp and so the types shade off into the other, but the distinctions are real enough for the core cases. I take up each in turn.

Metatheoretical work is theory of theory, and divides into two sorts: general and (domain) specific. General metatheoretical work is concerned with developing and investigating adequacy conditions for any theory in any domain. So, for example, it is generally agreed that theories should be (1) consistent and coherent, both internally and with other well-established theories; (2) explicit; and (3) simple…Specific metatheoretical work is concerned with adequacy conditions for theory in a particular domain. So, for example, in linguistics we have Chomsky’s (1964, 1965) familian distinctions between observational, descriptive and explanatory adquequacy. Whether such work is “philosophy” or, in this case “linguistics” seems to me a pointless question.

Theoretical work is concerned with developing and investigating primitives, derived concepts and architecture within a particular domain of inquiry. This work will also deploy and test concepts developed in metatheoretical work against the results of actual theory construction in a domain, allowing for both evaluating of the domain theory and sharpening of the metatheoretical concepts. Note this well: deployment of metatheoretical concepts is not metatheoretical work; it is theoretical work.

Analytic work is concerned with investigating the (phenomena of the) domain in question. It deploys and tests concepts and architecture developed in theoretical work, allowing for both understanding of the domain and sharpening of the theoretical concepts. Note this well: deployment of theoretical concepts is not theoretical work, it is analytic work. Analytic work is what overwhelmingly most linguists do overwhelmingly most of the time…

Linguists tend to confuse analytic work with theoretical work and theoretical work with metatheoretical work…

Linguists typically distinguish not among the three types of work described above, but rather between “theoretical” and “descriptive” work, where both of these are better understood as analytic work, with, respectively, more or less reliance on or reference to a specific framework and its concepts and architecture…This distinction between “theoretical” and “descriptive” is not only ill-conceived, but also, for reasons I do not fully understand, invidious. The tripartite distinction discussed above involves no evaluative component for or ranking of the three sorts of work other…


  1. Do you think that the balance between types of research is somehow off in linguistics? As a person with self-ascribed physics envy, do you think that the balance of research in physics is radically different?

  2. Interesting question. Personally, I think that there is so little real theoretical work in Chametzky's sense that I would like to see more. There is lots of excellent analytic work. But as for theory, I find that as a field we have largely off loaded the responsibility for this to Chomsky. So, yes, there is not enough in my view. Moreover, except for Chomsky's musings, most other theory is barely tolerated. We don't have ways of evaluating proposals except analytically.

    How's this compare with other fields? Well in the ones I envy, I think that theory is accorded a place. It's not the ONLY game but it's a legit one. In linguistics, at least in syntax, not so much. At least that's my opinion. You?

  3. I like this division: I think it draws the lines in the right place.

    Isn't the whole discussion of MCS languages, and the MCFG hierarchy etc a very solid bit of theoretical linguistics in this sense?
    I agree that this work is maybe not that familiar to most linguists of the non-computational persuasion.

    1. Part 1:
      Glad you asked this. I've been trying to figure this our myself. I've gone back and re-read Greg's thesis and have been trying to figure out what being MCS or not tells us about FL. In Greg's thesis, if I understood the last chapter, he argues that MCS is a diagnostic for copying operations as part of the FL rule inventory. The argument seems to be that allowing Copy means that Gs of FL are not MCS. If this is correct, what follows? I'm not sure. The reason is that I am not quite sure why being MCS is a good or bad thing to be. The way Greg and Ed (Stabler) put things, discussing these abstract string properties has two virtues: (1) that it allows theory comparison between formally apparently different kinds of theories (e.g. Tags vs MGs). One can compare these approaches wrt these abstract string properties and note how they differ and where they are the same. (2) It can constrain formalisms, e.g. if all Gs are MCS then a theory with Copy is out (I know that this is too easy, but it's good enough for here). Now, As Greg uses this in the thesis (1) dominates. What this gives us is the conclusion that most of the formalisms we are currently discussing are effectively equivalent using these metrics. But, for my interests, they clearly aren't. An MP with minimality but no phases is different from one with phases but no minimality. Which do we have and why is my question of interest for it tells me something about FL/UG. So, the methods available using this form of investigation, seem too blunt to address the questions I am interested in; i.e. the structure of FL/UG. As I also think that the only important theoretical questions in syntax concern the structure of FL/UG, though these concerns might have been insightful, they appear not to be (at least currently). This is a little like what has taken place, I am told in phonology. Jeff Heinz and Bill Idsardi have argued that phonology sits in the domain of regular languages. However, what's interesting is that it is a very particular trio of such languages, not just anyone. Ok, if one wants to know is why the phonology of FL is the way it is, then hearing that it is regular does not do much. So, is this enterprise theoretical? Well, it's abstract and hard but right now it seems form what I can tell not clearly advancing the central theoretical question: what's the structure of FL and why.

    2. Part 2:
      There is a second point to make. For Greg, the argument is from the structure of what grammars actually look like (e.g. do they motivate copy operations) to conclusions about the MCS/MCFG questions. In other words, the linguistics has implications for the formal questions rather than the formal questions having implications for the FL questions. This is less so for Ed who believes that it is a well attested FACT that languages are MCS and this does eliminate many kinds of grammars as options, e.g. MGs with COPY if Greg is right. That means that we need to reanalyze the phenomena that appear to motivate COPY and and this motivation is clearly theoretically driven IF languages are actually MCS. My problem with this, is that this does not make MCS etc different from any other fact that we believe to be true. So, say we believe that the evidence for islands is overwhelming (e.g. me) then any theory that cannot accommodate islands is ipso fact out. What to do if there are apparent counter-examples? Well, reanalyze them. This seems to me quite a bit different from the kinds of large regulative facts that I see prompting most theory: e.g. any kid can learn any natural language in more or less the same time on the basis of sparse degenerate data (Plato's problem), language evolved in humans (Darwin's problem), language is understood more or less as it is heard, etc. These seem to me like BIG facts that have theoretically interesting consequences for FL. I personally do not yet see that what people have conjectured about the complexity hierarchy as being of the same size.

      This said, I have an open mind here at the moment, which is why I welcome the discussions on the blog led by Thomas. Maybe these methods/problems will prompt more interesting theoretical developments. I hope they eventually say something about FL. If they do, and I am hopeful, that would be great. Yet another source for theoretical innovation.

    3. @Norbert: There's many points in those two posts that I feel differently about, but it's nice to hear how our work is perceived by other linguists.

      I don't have much time to reply as I'm on the road right now, but I think there is one big difference in what you are aiming for and what I am aiming for. I get the impression that you want your theory to nail exactly what FL looks like, even if this increases the risk of your theory being wrong. Whereas I prefer an account that carves out a class of possible FLs -- a weaker result --- but is more likely to "survive" new data or changes in how we think about existing data. The MCS result is one such result: you look at all the grammar formalisms compatible with it and conclude that the result is too weak to be interesting. I look at it and see all the grammar formalisms it rules out and conclude that it is a very profound, robust result that really narrows down the space where we should look for FL.

    4. Norbert wrote: In Greg's thesis, if I understood the last chapter, he argues that MCS is a diagnostic for copying operations as part of the FL rule inventory. The argument seems to be that allowing Copy means that Gs of FL are not MCS. If this is correct, what follows? I'm not sure.

      Yes, I think your understanding is basically right ... so wouldn't finding out that some natural languages are not MCS, and therefore that some states of the language faculty must correspond to grammars with that sort of copying operation, be telling us "something about FL/UG"? I get the feeling you're looking for it to "tell us" a different kind of thing, but I'm not sure what other kind of thing it is you're looking for. (Perhaps the answer is along the lines of what Thomas just added above.)

      I think the equivalence between MGs and (kinds of) TAGs is a bit of a red herring, and not really centrally related to Greg's thesis. The important point there is precisely that there are two formalisms that are not equivalent: PMCFGs are strictly more powerful than MCFGs, so it's possible to find a language whose set of strings can be described with a PMCFG but not with an MCFG. The basic claim is that Yoruba is such a language. The equivalence between MGs and TAGs (and MCFGs) means that you can't make that kind of argument (based on a language's set of strings) for MGs over TAGs or vice-versa, but there might be all sorts of other ways to make such an argument.

      One point that seems to me to be somehow overlooked in a lot of discussions about weak generative capacity is that making an argument based on a set of strings doesn't commit one to the belief that only arguments based on sets of strings are admissible, or anything deep and far-reaching like that. It's just one kind of argument, and it gets you as far as it gets you.

    5. @Thomas: I see that our research personalities differ across predictable lines; the impetuosity of old age versus the caution of youth! Let me add, however, one observation. I am not sure that the weaker conclusions are any less immune to the shifting tides of empirical insight. Take Greg's argument for example. It seems that the logic runs from the linguistic data, to the more general computational conclusions. Moreover, the argument is very indirect empirically, relying on quite a bit of "theory" if you will. The argument is that IF there is copy then there is no MCS. The argument for having copy, moreover, is not that we can discerns actual strings that native speakers reliably recognize but that the mechanisms required to *elegantly* analyze RCs in Yoruba would generate strings that do not obey constant growth. A reply to this, in fact the one that Joshi and colleagues regularly make, is that it is never allowed to apply unboundedly and so the premise fails. They point out (probably correctly) that native speakers probably would have quite a hard (impossible?) time accepting the "longer" empirically relevant cases and use this to argue that there is no MCS threat here at all. They did this for the other cases Greg reviews that challenged the MCS and would, I suspect, be happy to run the same argument against his. So, his conclusion rests on the same standard linguistic bases that ordinary arguments in my part of the world rest on: viz. that the best analysis of particular data implies the need for a copy mechanism. So the empirical foundations of his conclusions, and most others, are no different than those I rely on to make my, no doubt, extravagant claims. In other words, the belief that these weaker arguments are empirically more secure is, from what I can tell, an illusion.

      This said, I am happy with the conclusion that the work you provide sets a lower bound on adequacy of grammars. However, this has never been disputed, has it? The point has been that this is setting one's sights way too low. That the aim is to find out the detailed structure of FL and it seems that we agree that the methods you are advocating are too dull to cut to the bone here. Is this correct?

      So, in a moment of weak ecumenism: I am very interested in seeing how far your methods get us. What I am still somewhat unsure of (but eager to learn about) is how far these methods will go. IMO, we know a lot about FL already. It would be a shame to pretend we don't. That said, I am always open to new methods for investigating my favorite issues. Cannot wait.

    6. @Tim: As always your comments are spot on. I agree that looking at string properties does not prohibit one from making other arguments. However, two things are worth noting. First, that those who look at these, e.g. Thomas, seem skeptical concerning the other things we have discovered. After all, he thinks that this is the conservative position to take. Second, that as I noted in my comment to him, the conservatism is, it seems to me, illusory. Why? Because it is not actual string properties that are being used to argue for one position or another, but the properties of potential strings, those generable were the grammar with copy, say, allowed to run unboundedly. In other words, the argument is purely theoretical and the empirical/theoretical evidence for the conclusion is no more or less sturdy than the ones that people like me use all the time for what Thomas takes to be less secure conclusions. But, from where I sit, they are not less secure for they rest on the exact same empirical foundations, viz: the best analyses of this or that linguistic phenomenon.

      Last point: In LSLT Chomsky observed that we idealize to unbounded/infinitely sized sentences because we are interested in generative capacity. But the real important distinction is not between infinite and finite but between small and large. Say the actual grammatical set of sentences were in the 20 billion range, but finite. Would that really change the problem for me? Nope. The projection problem of how to form a grammar from PLD would be as acute. Would it change the weak generative capacity problem? Well finite is finite, right?

    7. Yes, good points.

      I think there is still sometimes a certain concreteness that comes from basing the arguments on sets of strings, but this is just a matter of practical convenience (some combination of factors about the nature of the observations and the well-understood mathematical properties of sets of strings), not because sets of strings have any particularly special status. I agree that there are still a number of idealizations of the usual sorts and it's still "indirect empirically", so I think the difference in approaches is just a matter of degree, about how much indirectness one tolerates versus how much concreteness one requires, rather than any difference in kind.

      There may be some who disagree and see weak generative capacity arguments as different in kind (fully concrete and not at all idealized or indirect), but I don't think Ed/Greg/Thomas/etc. are in that camp.

    8. I'd like to take the opportunity to publicly thank Norbert for including the material from my '96 book here. And I'm grateful, too, for Alex Clark's comment. If my work is, as Norbert suggests, sadly under read, it makes up for it by being exuberantly overwritten, if I might be somewhat self-Earnest (if not Important).

      As for the current discussion, which took off from Alex's question

      Isn't the whole discussion of MCS languages, and the MCFG hierarchy etc a very solid bit of theoretical linguistics in this sense?

      I guess I'd say that there's specific metatheoretical concepts being deployed and tested ("Specific metatheoretical work is concerned with adequacy conditions for theory in a particular domain." "the work you provide sets a lower bound on adequacy of grammars."--Norbert), so, yeah, it would count as theoretical work as far as my bit of taxonomic analysis (analytic taxonomizing?) goes. And as the ongoing productive discussion reestablishes, answering such a merely taxonomic question isn't what inquiring minds want to know. But if providing such terms helps spur such discussions, I'm more than happy, and my work here is done.


    9. @Norbert: I actually had a pretty long post ready, about the status of the unboundedness assumption, a comparison of the evidence against context-freeness versus the argument in Greg's thesis for copying, basically all kinds of things that make my previous robustness claim more clear. However, I realized that this wouldn't actually address what you're really interested in, it would only serve to make my own position a little bit more secure. So I came up with another super-long post instead ;)

      As far as I can tell, what you really want to see is some juicy results, something where the formal perspective i) directly addresses the questions syntacticians struggle with every day, and ii) tells us something we didn't already know through independent means. The original goal for the upcoming posts about derivation trees was exactly that. But having read your two posts above, I'm not sure if they will actually give you what you desire. The reason being that several such results have already been discussed on this blog before, but have apparently failed to fully convince you:

      1) Last week I explained why remnant movement is indispensable if you only have phrasal movement. Moreover, I pointed out that roll-up movement and smuggling do not increase strong generative capacity over having just Merge. Another way of putting this is that PBC-obeying movement alone does not cut it, no matter how many coding tricks you employ.

      Doesn't this tell us something about FL? The last time I mentioned the result to a group of syntacticians (sprinkled with one or two semanticists), they found it rather surprising. In particular that it holds no matter what your preferred analysis for phenomenon X is or whether movement is chain-formation, multidominance, etc. None of those contentious points factor into the claim.

      2) The Merge-MSO connection shows that Merge is far from the simple operation we usually view it as, it packs quite a punch. I know that you had some qualms about the validity of the result if the set of categories is fixed, but as I explained back then this may reduce the problem somewhat, but does not make it disappear. Whatever the solution to this problem will be, it will tell us something about the status of constraints and category features in FL.

      3) I don't remember if I mentioned my work on island constraints on this blog or if this is something we discussed in private. It is firmly grounded in a mathematical approach to language, but is uses this grounding to show how the adjunct island constraint follows from surface properties of adjuncts. A nice result about FL: the AIC follows from the machinery that creates adjuncts, whatever that machinery may look like.

      So my question is, do these results meet your criteria for interesting claims about FL? If not, where do they fall short?

    10. Hmm, I am surprised that I am asked to arbitrate what's theoretical and what isn't. But as you ask I will try. A distinctive mark of analytic work is that it argues from data to conclusion. This amounts to curve fitting in more or less a=elaborate ways. So, Kepler's first law fits an ellipse to the data sets of Mars, Galileo's acceleration law fit the inclined plane data. We defend the formula on the basis of its fit to the observables. This is important work, but it is more formal than theoretical.

      In linguistics one finds the same thing in GB: the binding theory looks the way it does because the data are the way they are. Were the data different, the principles would change and there would be no cost to changing them for there is nothing that hands on principle A/B being stated as they are except the observed facts.

      One gets to theoretical work, when we not only explain the primary data but when we also explain the explanations of the primary data: why do the laws look as they do. In the 17th century, this is what Newton should a honcho: his theory explained both Kepler's and Galileo's and did it in a unified manner. One derives Kepler from Newton and derives Galileo from Newton, and gets the tides in as a bonus. Nice work. The reasons their laws work is because they are special cases of his. He not only explains their facts but explains their laws. Nice.

      Anything analogous in linguistics? Less clear here. I would say that this is what MP should be aiming for: to derive the laws of GB from something more general. We get a taste of this by unifying case, control, agreement, binding etc via movement and unifying movement and phrase structure via a generalized version of Merge. Does this work? Well, in part but there is, ahem, controversy. At any rate, I think that this is theory for it explains the shape of the explanations. Theory has laws as explanada, or this I believe is a good rule of thumb.

      Ok, now for your stuff, how's it fit? You show that remnant movement adds new expressive power and that roll up and smuggling do not. Interesting, but does this tell us why the generalizations we observe are there? I am not sure, and probably am not the person to ask. However, even if true, this does not require that there be remnant movement (this is justified empirically) it does not explain the properties of remnant movement should it exist, does it?

      Similarly for the two other projects you mention. The way I see theory it constrains the fundamental laws and/or operations in some way. It adds another desideratum to our evaluations. It is not the application of formal tools to clarify difficult issues or the computing of consequences of analyzed data. Theoretical work aims to constrain lower level generalizations, absent these theoretical work is not possible.

      Note that on this ground assuming that languages are necessarily MCS does constrain UG, as Greg shows; it is inconsistent with Copy (or at least unbounded copy). The question is what are the reasons we have for assuming that NLs are MCS? It seems that the answer is that this is a good empirical generalization. It is like claiming that planets move in ellipses. Important if true, but not quite what I have in mind.

      There is little doubt that these conceptions of what is and isn't theoretical is idiosyncratic, in part (not quite, as many physicists think that theoretical physics really started with Einstein). But you asked.

      Last point, none of this should be taken to demean non-theoretical work. Here I agree with Rob C completely. But as you asked...

    11. @Norbert: No worries, I'm not miffed at all, just trying to figure out what kind of result you're looking for and if I know anything of this sort that I could write about. Now that I see where you draw the lines, I can start thinking about this more clearly.

      One quick and dirty attempt. How about the following, more speculative claim (I'm simplifying things here for the sake of argument, there's some pesky empirical issues that make the picture less appealing):

      1) Phonology is some kind of finite-state device over strings.
      2) Syntax is a combination of two finite-state devices, one computing derivation trees, the other one the mapping from derivation trees to output structures (maybe multidominance trees, maybe LF and PF representations, maybe something completely different)

      The second point implies through various theorems that language is MCS, and the limitation to finite-state computability links it to phonology. Now if the limitation to finite-state computability could somehow be derived from general cognitive restrictions, that would tie together quite a few things into a neat package.

    12. Bill Idsardi has recently mumbled similar things in my ear. I confess that this is above my pay grade. BUT, it sure sounds like a big idea, so I'm inclined to agree that this would fit even my idiosyncratic considerations.

      By the way, neither you nor I are or have been miffed. Just talking vigorously, and having fun and shedding lightness and joy in the process. Right?

    13. neither you nor I are or have been miffed.
      Well, there was that one time when... nah, just kidding. ;)

  4. The discussion of what these computational properties have to say about FL, and whether it's theory in RC's sense or not, Norbert referring to research I conduct with Bill Idsardi and Jim Rogers, plus the latest post on "being Edgy" are compelling me to chime in here. While I don't mind the fact that reading this blog has become a habit, I am concerned that writing posts to it could take more time away that I don't already have... but here goes.

    I would answer Alex's question above more generally: investigating the computational properties of the weak and strong generative capacities of natural language are absolutely solid theory in RC's sense. Norbert describes "the central theoretical question" as "What's the structure of FL and why?" In the "being Edgy" post, Norbert also makes the point that the properties of UG are going to be abstract. I completely agree with Norbert on these two points (w.r.t. the last Keenan and Stabler response to Levinson et al 2009 is a good read; dunno how to do the linking yet). And I would come out swinging that computational properties of natural language patterns can in fact provide strong, abstract, restrictive hypotheses about the nature of FL *and* they can help explain why.

    One example I can give for this comes from my own work in phonology. It appears that the attested phonological patterns (both phonotactic patterns modeled as formal languages--i.e. sets of strings---and phonological processes modeled as mappings from strings to strings) belong to very specific subregular regions defined by particular computational properties. The linguistic hypothesis is that these properties reflect the character of the FL. They are strong properties because they eliminate most logically possible patterns. They make strong, falsifiable empirical predictions about the kinds of patterns we expect to find cross-linguistically and the kinds of patterns learnable/processable in psycholinguistic experimentation. They are abstract properties because they are not about surfacey things at all like "all languages have vowels". As for why natural language exhibits these properties, I think one answer comes from learning. If the FL/UG generalizes in certain ways from its linguistic experience, then we expect the empirical data to look the way it does.

    I'm less familiar with all the syntactic generalizations, and it is true that the claim that all syntactic patterns are MCS is much weaker than the claims I am making about phonology. But noone is saying the story ends there. The observations that Thomas and Greg and others are making about the nature of derivation trees in MG could very well lead to similarly restrictive hypotheses. This line of research will inevitably lead IMO to interesting subclasses of the MCS languages or other restrictive classes that crosscut the Chomsky Hierarchy (and which may include copying). The work that Alex is doing on learning is a complementary approach focusing on computational properties of grammars and languages that are learnable in various senses and display natural-language like behavior.

    I'm very excited and optimistic about the future for 2 reasons: I think in my lifetime the theoretical/analytical linguistic work, the mathematical linguistic work, and the work in grammatical inference, all of which is ongoing, is going to make some very exciting connections. The fact that people from all those important corners are talking on this blog is the 2nd reason.

    1. Jeff, great to have you join the discussion. Here's hoping you'll give in to the temptation once in a while, time constraints notwithstanding.

      I agree with everything you say, even down to the two reasons for optimism. That's a good opportunity to thank Norbert again for welcoming such technical discussions on his blog. Also a tip of the hat to Alex C, who --- if I'm not mistaken --- was the first of us computational guys to comment on this blog and thus helped getting the ball rolling.

      PS: Linking is the standard html markup fare.

    2. Thanks. Here is a link to the Keenan and Stabler paper in Lingua (and a preprint version from Stabler's website since ScienceDirect appears down at the moment.)

  5. @Norbert: "Last point: In LSLT Chomsky observed that we idealize to unbounded/infinitely sized sentences because we are interested in generative capacity. But the real important distinction is not between infinite and finite but between small and large. Say the actual grammatical set of sentences were in the 20 billion range, but finite. Would that really change the problem for me? Nope. The projection problem of how to form a grammar from PLD would be as acute. Would it change the weak generative capacity problem? Well finite is finite, right?"

    This is a really good point. I think the standard way of presenting this property misses this point -- the size of the representation is crucially important. This is also important when we say things like MGs are the same as MCFGs -- well they aren't because the MCFGs may be massively bigger than the corresponding MGs. So there are important intensional differences between MCFGs and MGs that the weak (and strong) generative capacity arguments miss. But that is just an argument saying that we need a slightly better theory -- not an argument against having mathematically precise theories.

    1. Just in case any of you are not familiar with this paper which speaks to this issue: Savitch's 1993 paper defines a finite language L as "essentially infinite" with respect to a class of representations if (1) there is a rep R1 for L and (2) there is a rep R2 for an infinite language L' which is coextensive with L up to the length of L's longest string and (3) R2 is smaller than R1. He goes on to prove that whether finite languages are essentially infinite depends on the the class of representations.

  6. I agree with Rob's distinctions. It seems like an insightful way to divide things up. The way I understand what he is saying is that analytical work takes some system of concepts/mechanisms/principles and applies them to some language data (the "phenomena"), to try to make sense out of that data, and then, on the basis of that analysis to modify the concepts/mechanisms/principles. Almost all cartographic work falls into this category, I believe. Most of my books (with Paul Postal) "Imposters" and "Classical NEG-Raising", also fall into this category. "Analytical" work is highly addictive. I had originally (in 1993) thought that the Minimalist Program would stimulate much more theoretical work (in Rob's sense). An early clear example of this was Epstein's well-known paper on c-command. The whole point of the SMT (Strong Minimalist Thesis) is to try to look into the relations of various concepts, to see what their status is. And the SMT gives a concrete way of performing this task. One can ask, for each concept/mechanism/principle: "How does this fit into the SMT?" That is an exciting research program that really has not reached its potential. But this kind of work definitely takes a back seat to analytical work. The natural instinct of a syntactician is to approach such questions by immediately thinking of relevant data. We cannot even help ourselves. But a different path would be to approach them "theoretically", as Rob suggests. Discussing these issues goes right to the heart of our field, and even minor clarifications could lead to huge advances on the "analytical" level. In addition to c-command and labels, which have been discussed from the "theoretical" point of view on occasion, I would like to add that there is need of much more discussion of basic issues like occurrence, Merge, multi-dominance, chains, copies, NTC, inclusiveness, derivations, workspaces, the primacy of the C-I interface, interpretability, "late insertion", the format of lexical items, the nature of grammatical categories, semantic interpretation without indicies, what semantic interpretation actually is, etc.

  7. I have always found Rob's distinctions useful, but I think there are some practical reasons why most current work in Linguistics is analytical. One is that it is easier to evaluate analytical work - It adopts some premises, applies them within a domain, and analyzes the outcome. Theoretical work involves examination of premises, and it's much more difficult to convince people that their premises are wrong than to convince them that such-and-such data can be analyzed within their (perhaps slightly amended) premises. The other reason is the one referred to by Norbert as "offloading" to Chomsky. In fact, if someone who isn't Chomsky does try to do theoretical work and comes up with a proposal that isn't Chomsky's, the proposal tends to get sucked into a black hole. If Chomsky comes up with something similar, the proposal becomes Chomsky's, and if he comes up with something else, the proposal becomes tragically misguided. I don't think this phenomenon is malevolent - In many ways it's helpful to a field to have forces that cause theoretical developments to be generally shared across the field. This discussion is very interesting. Presumably there are other ways to foster shared assumptions while encouraging theoretical work in Rob's sense.

    1. I agree with Peggy's observations, but I think that there is something more going on. I believe that theoretical work makes sense when one tries to keep one's eyes on the big problems that animate the field; Plato's Problem, now Darwin's etc. These are all too often lost in the nitty gritty of research. What partly makes Chomsky such a font of theoretical ideas is that he never forgets what the whole thing is for; what the big questions are and why the technical stuff is interesting. We often fail to ask ourselves why anyone should care about our results, why they are important, what they tell us, if anything. This has the effect of shifting all of our attention to areas where we think the value is obvious. So it is not only ease, though I agree there is some of that, but what we value as well and what we take the object of inquiry to be. P's and D's problems focus attention on the structure of FL/UG. Chomsky keeps asking himself, I believe, what does this tell me about FL/UG. If the answer is "not much" then he moves on. If the answer is "possibly a lot" then he moves forward. This strikes me as a pretty good mantra, and it will foster some useful speculative thinking of a theoretical nature.

    2. I think it is right to keep focused on the central problems: Plato's problem and Darwin's problem: these are empirical constraints on the theories. We know that languages are learned (sorry acquired) and we know that UG evolved. But Chomsky has essentially given up on Plato's problem, and
      instead seems more interested in non-empirical constraints like the SMT. That is a bad development, and I don't think theoretical linguistics should follow him down those paths.

      Other theoretical concepts that RC mentions above also lack any empirical motivation -- NTC and inclusiveness for example.

    3. "Given up" is inaccurate on reflection: "declared victory and left the field" is better I think.

    4. @Alex Clark "Other theoretical concepts that RC mentions above also lack any empirical motivation -- NTC and inclusiveness for example."

      Actually, this was Chris Collins (and hi, Chris . . . been a while), not me.


    5. I think that Alex puts his finger on the central point. I have a post on this that I am in the middle of and that I hope to put up in a day or two. Where I might disagree with Alex is the status of Plato's problem in current theory. It's not that it's been abandoned or solved but that another problem has been added that once again reconfigures the problem, in ways, actually, that I though Alex would be sympathetic to, viz. there may be less linguistically dedicated acquisition mechanism than earlier contemplated. At any rate, Alex has, IMO, made the critical observation; in linguistics theory inevitably confronts the great meta-theoretical concerns and is evaluated in these terms. More soon, I hope.