Monday, 1 June 2015

Athens Day 3

Day 3 consisted of no new panel sessions--- just summaries and discussion.  In the morning, we tried to raise and discuss in more detail the most important issues that emerged from the meeting.

One interesting issue in the morning discussion was the idea of 3rd Factor principles and the extent to which they can, in  practical terms, constrain the theories we develop. This had been a recurring theme, and people clearly pulled in different directions here.  Hornstein had already spoken out strongly in favour of top down general principles that are used to constrain the questions we ask and the kinds of solutions we seek at the linguistic level.   3rd Factor principles such as general cognitive limitations and proclivities, and logical facts about memory, or learning, or the way symbolic systems configure themselves for maximum efficiency etc.  are the boundary conditions for what the grammar ends up looking like, and have contentful implications, apparently belied by their abstractness.  Rizzi expressed in a clear calm way what I take to be the source of resistance many working linguists feel in practice which is that we don’t have clear enough ideas to make  a priori  choices about what those 3rd factor principles are. There is an empirical dimension to these facts, and our choices of which are relevant where, that makes it hard to use them at a concrete level.  What is the `perfect solution’  to the problems imposed by the interfaces? And why do we think that what our human logic considers to be conceptually perfect, or simple,  is what biology considers to be perfect? Maybe biology thinks repetition and overlap and reuse is `perfect’.  We are all just speculating here.
I think Rizzi is right in this, but I also think Hornstein is right to emphasize the role of top down theorizing, and the drive for abstractness in keeping syntax healthy.

In the afternoon we attempted a much more positive contribution by trying to brainstorm as a collective to come up with a consensual list of  major results and mid level generalizations that had emerged within the generative linguistic tradition in particular.

This seemed important for a number of reasons.

A.   It is important because it reminds  us of what our current state of knowledge is and allows the next questions to be asked in a way that will be genuinely cumulative and productive. 

Generative syntax has often been accused of being narrow, jargonistic, or impenetrable, so that gives us a second reason.  

B.  It is important to try to state the important discoveries and generalizations of the field in terms that are general enough to transcend frameworks and particular implementations, as a healthy mental exercise for ourselves. 

Actual day to day syntactic analysis requires formal discipline and precision in practice, to ensure that the model one builds is indeed capturing the data one has described, and specialized terminology is unavoidable.   (If I understand Müller correctly, the substance of his complaint about our field was that much recent work seems to have lost the skills or inclination to pursue a syntactic analysis at this `algorithmic’ level. I don’t know if this is true or not. It might be. There’s unfortunately a fair amount of shoddy syntactic papers out there, but there’s good and bad work in every field). At the same time, we certainly do not want to say that the results of the generative enterprise are confined to those discovered in one particular framework or architecture.  Or that those discovered in different frameworks and in different eras are not commensurable.   Nobody in the room wanted to say that. And with respect to older versions of Chomskian grammar such as GB, it was explicitly affirmed that GB was responsible for establishing certain MLGs for the field that are a central part of the cumulative legacy of generative syntax.

 Our mathematical linguist on the ground here in Athens, Thomas Graf,  kept insisting and reminding us that he could not see that anything substantive could hinge on the differences at the level of  grammar architecture that the field often seems to be preoccupied with arguing about.  This came up strongly in the derivations vs. representations session, and again in the discussions of how the formal grammar interacts with the parser. 

In discussions over lunch and dinner, we syntacticians felt compelled to  pursue this with Graf in more detail.  There is a strong intuition among us, that some theories (with a little t) are just better at capturing particular generalizations  than others. We did not like to be told that these `small-t-differences’  were illusory, and not worth arguing about.   Graf, in turn, voiced a completely different perspective on small-t-Differences  that resonated strongly with me. In mathematics, the reality of a phenomenon transcends small-t-differences, but choice of small-t framework is a powerful determinant of how easy it is to solve a  particular analytic problem, whether generalizations emerge in a way that is intuitive to the human perceiver, and even the kinds of issues one even looks at. So translating into different small-t frameworks is actually a Thing--- a positive tool, that allows you to walk all the way around a particular problem and see its front side and its back side and its top side etc.   Some small-t-theories allow a particular pattern to emerge gracefully to our senses where their translation into a notational equivalent would not.  Being able to be precise about our small-t-theories but also in some cases switch between them should a be seen as a source of strength for problem solving, not a source of point-scoring fodder for article publications. 

In addition, we have a third reason for why attempting to state the midlevel generalizations is important.

C. It  is important because it allows us to scale up our claims to a level of granularity at which we are able to interface with our colleagues in related academic disciplines, potentially narrowing the commensurability gap.

Having said that, the exercise of trying to do this in a group of one hundred people, and syntacticians of different persuasions was difficult, frustrating and instructive in equal measure.  

It is easy to agree when the discussion takes abstract form.
But as soon as an actual list was attempted, disagreements started to emerge about what counted as a mid level generalization:  why/if  we should be doing this at all;  how theoretically charged the terms  should be; judgements of size and/or importance; level of crosslinguistic generality;  how much consensus does there need to be on something for it to make it onto that list.  It's a minefield.

After three days of stimulating discussion,  and No Shouting (unlike what I had predicted),  irritations started to emerge.  There was so much residual good will in the room, that most of us stuck at it and battled through. If we get a List that anybody likes at all or thinks is important, it will be a triumph of unity.  And I still have hope that such a list will emerge. A committee has been put in charge of taking the 50 or so bullet points on our communally generated first draft and making it into something that makes sense.  But I fear that the difficulties here are representative of deep internal disagreements/perspectives about the current state of the field. 

In my next post, I will continue with my summary of where I think the internal faultlines are. I will reiterate my support for the exercise in articulating MLGs, and I will give an example of my own version of the  list in one small subdomain of syntax that I actually know something about (Argument Structure).


  1. So Thomas and I had a discussion on Norbert's blog about exactly this issue, that small-t theoretical differences can spark off different ways of thinking about topics and so they can be massively useful methodologically, even if at a deeper level the different theories are all modelling the same phenomenon. (discussion especially in the latter comments at I think I actually agree quite a bit with him that much of our discussions about different frameworks (within generative grammar) aren't very important beyond that methodological issue. I have papers I wrote using categorial grammar and HPSG before I moved to minimalism, and have always felt that there are deeper commonalities masked by surface differences (sounds like syntax!). The big differences are between usage-based construction grammar type systems and generative grammar writ large, and I think that's where we should focus (see the recent paper by me and Peter on bound variable anaphora which provides an argument against tackling the core phenomenon in a cognitive grammar way (at, please forgive the type on one of the Gaelic examples, will fix when we do revisions). All the different types of generative grammar are so similar compared to usage-based approaches, that the real empirical grist needs to be milled at that junction, I think.

    1. Yep. I agree. But going all fuzzy and hand-wavey on implementational issues is not the solution of course. We all need to become a little more multilingual. I differ from you perhaps in being rather dissatisfied with the kinds of questions that have been flowing from the dominant syntactic `language', and frustrated at where syntacticians sometimes pick their fights and pitch their disagreements.

    2. And indeed, Thomas was simply reprising some of that earlier discussion with us, in the Athens forum.

    3. Yeah, not keen on hand wavey at the implementational level! So what are the issues that frustrate you in terms of where disagreements are pitched?

    4. I'm very much inclined to disagree with Graf (though I'm kind of shaky in this area). Take CCG. Two statements about CCG that I'm sure are true: (1) it was developed partly in response to MLGs which were hot topics at the time; (2) it has different formal properties from competitor theories at the time. So the MLGs led to the development of formally different objects. Cross-serial dependencies and the abandonment of GPSG is another example. The interplay of empirical discoveries in locality theory, Chomskyan theorizing c.1973-1995, and the fall-out of the Peters-Ritchie theorem should be yet another, except it's murkier in pretty much every respect. So if Graf is trying to say that our current range of "hot topic" MLGs can be intertranslatable into various frameworks, fine. But if he's saying that it's impossible in principle for empirical work on MLGs to eventually feed into the development of the mathematical underpinnings of the theory, surely he's wrong.

    5. I think we need to ask him directly, but I gather what he is saying is not that there is no feeding from MLGs to the mathematical level, but that some of the things we argue about with respect to the way in which MLGs are expressed are not relevant to that mapping. I guess this is because as non mathematicians we underestimate how distorting homomorphisms can be. I was also hoping that the MLG level could be one where we try to state things more framework neutrally in the first place.

    6. Rob, ask Thomas, but I thought the point that was being made, at least in the discussion that he and I had on Norbert's blog, is just the former (that much of what we do is really intertranslatable). One could just code all the MLGs up in some formal description language, that would be framework neutral, as Gillian suggests, but I think that would actually have to be an add-on to the framework-dependent work, rather than a substitute. That's because the development of restrictive frameworks of different sorts is, I think, crucial in the general discovery process.

    7. Thomas can speak for himself, but I felt the argument was that some arguments about MLGs are vacuous because one can translate freely between one option and the other: for some subset of derivations versus constraints,
      whether grammars are fully lexicalised or not (I don't know what that is called in minimalism), whether features are monovalued or binary valued, whether Merge is binary or n-ary etc etc.
      Speaking as a mathematician I have some sympathy for this view, though I think on occasion it neglects some important factors, e.g. some changes in the size of the representation as one moves from one to the other.

    8. I'm a little late to the discussion, I didn't see that some new comments had been added (Gillian, may I suggest adding a widget for recent comments like the one on FoL?). My claim was indeed about intertranslatability, in particualr with respect to proposals within the Minimalist framework.

      Of course not all of Minimalism can be squeezed into GPSG, and strictly speaking HPSG can do a lot more than Minimalism (though it is interesting that there is a translation from HPSG to TAG that looks like it covers most HPSG analyses; but I'm not very well-read in the HPSG literature, so take that with a grain of salt). But this is not what the discussions in the literature are about, there the hot-button issues are things like features VS constraints, what kind of features, multi-dominance VS copies, and so on. To the extent that there is an empirical core to these debates, couching it in these technical terms often obfuscates the essential issues rather than illuminating them.

      Alex is also right that the choice of representation can matter in various respects that are not captured by the standard tools like generative capacity. Bob Berwick had a comment on FoL along those lines where he explained why GPSG recognition is much harder than CFG recognition even though the two formalisms are weakly and strongly equivalent.

      I think the optimal solution is to have as many equivalent views of Minimalism as possible, each one with its unique advantages. People should use the representation that works best for their purposes. Sometimes feature-based stories are easier to understand, sometimes constraints are more illuminating. Sometimes you really need to use features for practical reasons even if you don't want to (e.g. because the parser you implemented cannot handle constraints). But then you run your automatic translation algorithm from constraints to features and everybody is happy. And if it turns out that the converted grammar is too big to be efficiently parsed, well, that shows that we should either have a parser that can handle constraints directly or there is something wrong with the constraints we use.

  2. Let me add another voice saying thank you for writing up these summaries, Gillian, as another person who's very interested in the discussions that happened in Athens, but who wasn't able to attend this weekend.

    What you say about the challenges of generating an agreed-upon list of MLGs resonates surprisingly with the experience I had this weekend participating in on of the linguistics wikipedia editathons. It's not only that many (though certainly not all) linguistics articles on wikipedia are in dire condition, its that many reflect very different perspectives and choices about what counts as an interesting generalization, or how to talk about data -- sometimes just in terms of terminology, but sometimes the terminological gaps are more terminological crevasses.

    This is different in many ways from the issue of agreeing on even a short list of MLGs -- but maybe not so much, since presumably one goal of better "outreach" is that we be able to communicate our major findings, as a field, to the kind of interested non-linguist who is willing to fall down a rabbit-hole of Wikipedia links.

    In any event, I'll be very interested to see the documents that come out of the discussions in Athens -- though I certainly don't envy anyone the task of putting those documents together.

    Looking forward to your next (final?) post.

    1. Thanks! and glad you find it useful. Yes, tomorrow will be the final wrap up.:)