|
|
|
|
|
|
|
|
 |
 |
 |
 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
GG No, it's basically accidental. I more or less skipped semantics. Okay, I wrote the paper with Geoff on truth-functional connectives and the paper on the semantics of coordination that we discussed, but I don't think I've ever done anything else that really deserves the description ``contribution to linguistic semantics''. I know it spoils the story but I can't really lay claim to any significant achievements in that area.
I think the order of my interests and activity has been largely fortuitous. We talked earlier about what led me into syntax. The move from syntax to the lexicon was driven by the nature of what had happened in syntax, but my decision to spend a lot of time on morphology was an accident of the character of the phenomena that seemed to illustrate DATR most effectively. The connection to morphophonology was really driven by the work on morphology. You really can't do morphology without doing at least some morphophonology. You have to have some view of phonology as well. There's no way of describing morphology other than via phonological objects. It's not like syntax where you have orthographic words. There are no interesting orthographic counterparts to morphological objects so you are forced to use phonology. And, once you start using phonological representations, then you have to confront morphophonological problems. Of course, you can always hand wave at them and say ``well, I've got some abstract phonological representations here that reflect the true nature of the morphology: the morphophonology is somebody else's problem''. But it is very hard to do convincing analyses if you haven't got at least a bit of the morphophonology spelled out otherwise things can look really weird.
EJB Right. One theme of your career has been defaults but another, probably more important, has been formalization: ``do it properly!''. Has your view on that and what that means changed at all as you've moved through the different topics? Or is it just: ``it's always the same methodology, it's just that the topics require a slightly different approach to it''.
GG I think it's the latter. I don't think my methodological views on on formalization have changed at all. I read Montague and never looked back really.
EJB Although the connection between Montague and, say, the work on DATR is not immediately obvious.
GG No, but the methodology is the same.
EJB If you were doing the pragmatics work again now, would you take a different approach?
GG Yes, I think so. I think I would have used an existing default logic if such things had been available at the time I was doing that work. I'm not a logician and I would rather use logical things that have been developed by other people if they exist, rather than try and invent my own. I'm not someone who invents things de novo for the sake of it on a kind of ``not invented here'' basis. But there wasn't anything around in the early 1970s so I had to invent something myself.
EJB What about implementation? Would you have taken more or less the same approach that you did with DATR and tried to develop an implementation more or less in parallel with doing the formal work?
GG If I had been able to use a current default logic, then implemented theorem provers exist for several of them and could also have been used. One needs to be able to do proofs in the presupposition projection area because predictions about the presuppositions of compound sentences are not self-evident. It would have been a tremendous asset to have had a theorem prover rather than having to grind through proofs by hand.
EJB Returning to the general themes, defaults run through much of your work, albeit in slightly different instantiations. What do we mean by default? Is there a unifying theme there of any kind?
GG No, I don't think so.
EJB It's just a topic dependent issue?
GG Yes. I'm not really a believer in a grand overall theory of defaults such that everything can be subsumed within it. There might turn out to be such a thing be but that's never been my enterprise. It's just that I have gotten used to thinking about defaults so it's very natural for me to introduce them into work I'm doing if they look as if they'd do something useful. I don't think there is anything more to it than that.
EJB I've heard you say that, unlike Geoff Pullum or Paul Postal , you have mostly chosen not to engage in academic debate with the Chomskyan mainstream, with the exception of some acerbic footnotes and asides. That said, there do seem to have been points where you did seem to be engaging a bit. The Royal Society in 1981 was one occasion where you came into direct contact (Gazdar 1981). Perhaps the engagement there was not of your choosing. It also seems to me that, at points, you have chosen very directly to engage cant and half-baked stuff: the paper on Bernstein (Gazdar 1979c) is one example. And a lot of those acerbic footnotes and comments in passing are really much more explicit than anything anybody would say these days. So I think your claim is slightly disingenuous.
GG There's just much less of it. You're probably not an avid reader of the collected works of Paul Postal but Paul has contributed far more than I have to that enterprise.
EJB I'm sure that's true of Geoff Pullum as well.
GG Yes.
EJB So, why not more, then? Isn't important to reveal the half-baked and the cant for what it is?
GG Well, yes, it is in a field that is capable of taking such revelations on board. But I came to the conclusion that linguistics wasn't such a field. So I gave up, I'm afraid. NLP just isn't like that - if somebody makes a mistake in a paper then, fine, in a subsequent paper you can point it out. Or if somebody advances a view that you disagree with, you can argue against it. But the chances are that the rubbish is not influential and falls by the wayside anyway. There aren't a large number of influential people in NLP doing worthless work. I think that would be right. They might not be doing work that one would choose to do oneself or work that one is particularly enthusiastic about, but that doesn't make their activity pointless.
EJB I can certainly think of work that in different ways had some of the flavour of Chomsky in linguistics. There was Schank . More recently, there was Berwick , Barton and Ristad .
GG I'll grant you Schank's rhetorical style, but the work itself was no less well baked than other NLP of its period. And some of the Berwick et al. work seemed to me to be quite interesting.
EJB What is it about computational linguistics or computational linguists that makes them more impervious than linguists?
GG I don't know. It would be nice to think that it was the testing inherent in a task. But NLP often isn't tested in that way. More now than in the past. But lots of NLP has never really been tested in practice. So, although it is tempting to give that as an answer, I don't really think that would be a good answer.
EJB I guess that NLP work is more often implemented than not, even if it doesn't get deployed in a serious application.
GG Yes, it gets implemented usually. But I meant actually tested doing something useful. There are now particular components like taggers and parsers which compete against each other. If something doesn't work well, then it becomes visible. But that is currently true only of a minority of NLP components.
EJB So it's not really something about the training or the background of the people involved in the field. It's something about what counts as an adequate test and what you have to before you publish something.
GG It's clearly the sociology of the field but why the sociology is the way it is, I don't know.
EJB It's not a view I subscribe to, but I've heard others say that you are somebody who has formalized ideas that were around, but not contributed to linguistics.
GG I think that charge could only be made by somebody who really didn't know much about the fields I've worked in.
The treatment of presupposition projection I developed in 1974 was quite unlike that of Karttunen or any of the less formal work on the topic published in that era.
When Geoff and I started reviewing the arguments against the context-freeness of natural languages, we weren't working in a field that was full of people who thought the existing arguments were invalid. On the contrary - most linguists were teaching their students that the facts of English subject-verb agreement showed that English was a context-sensitive language.
Turning to GPSG , many of the ingredients of that framework were very standard things that had been around for some time. For example, replacing monadic nodes with feature bundles in syntax is pervasive in the Stockwell , Schachter & Partee work from the 1960s that I referred to earlier, and combining that approach with unification was already a feature of NLP work, notably Martin Kay 's. So there was nothing new there. What was new was a realization that that formal technology could provide interesting solutions to significant linguistic problems like the description of unbounded dependencies and coordinate structure constraint facts. Slashed categories weren't a formalization of an idea that was already around in 1978.
When Roger and I hatched DATR in 1989, we were very familiar with the existing formal work by Daelemans , Flickinger and Shieber on lexical knowledge representation, and by Touretzky on inheritance networks, and all four of them influenced different aspects of what we did. But there is no way that DATR can be construed as just a formalization of a bunch of informal ideas that were already around in linguistics or NLP .
Notwithstanding the case for the defence, it's worth attending to the charge itself for what it reveals about linguists, namely the notion that formalizing an idea is not itself a contribution to linguistics. Imagine, in computer science, that Razborov or someone manages to get a proof that P does not equal NP. Would computer scientists say ``well, all Razborov has done is prove something we already knew, that's not a contribution to computer science''? I think not.
EJB I've heard you say that the typical linguist knows a bunch of cute facts, but that knowing cute facts isn't a sufficient goal for linguistics as a field. Why not?
GG Linguists get brownie points for cute facts. A linguist like Ivan has high prestige amongst syntacticians because he does have this wonderful skill of being able to produce relevant data. Postal likewise. And, of course, there are many people within the MIT tradition who have similar skills. Such people are tremendously useful if, for example, you are developing a computational grammar of English. To have someone like Ivan on tap is a huge short cut. You don't have to go to the library, you just ask Ivan and he will produce five relevant examples on the spot. So the ability to deal in cute facts is not something to be dismissed lightly. On the other hand, for that skill to be what makes the field tick seems to me to be unsatisfactory from an external viewpoint. Imagine what mathematics would be like as a discipline if the key to career advancement was an ability to do complex mental arithmetic quickly. If you think about funding linguistics departments in universities so that they do the work necessary to describe natural languages and to theorize about the nature of natural languages in a way that constitutes a contribution to scientific knowledge, then linguistics departments patently don't do that. That's not where the brownie points are. A lot of it is hermetic, it's not actually of any use to anyone else. It's a closed internal system.
If I was charged with university funding, I would pull the plug.
EJB On a personal note, if you were starting over again, is there anything you would do differently in terms of the training that you had or the career choices that you've made along the way?
GG There's only one career choice that I sometimes wonder if I regret. After I graduated at UEA , I was offered the opportunity to do an MA in economics. I turned the offer down because I'd only done economics as a space filler since I couldn't do philosophy as a single subject. I do sometimes wonder about how very different my life would have been if I'd done an MA in economics and become a city economist.
Copyright © Ted Briscoe & Gerald Gazdar, Wednesday 2 May 2001