|
|
|
|
|
|
|
|
 |
 |
 |
 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
GG Not exactly. Your opening remark is misleading in a couple of ways. Firstly, I had always been interested in computational linguistics - I'd been a member of ACL since I was a graduate student and, as I outlined earlier, I had a pretty good knowledge of NLP before I got to Sussex in 1975. Secondly, my interest in ``computation in service to linguistics'' wasn't something that emerged at the end of the GPSG period. In fact, Roger and I had started work on ProGram right at the start of the GPSG enterprise. And I guest-edited a special issue of the journal Linguistics on ``computational tools for doing linguistics'' that appeared in 1983 though, of course, it had been initiated a couple of years before it appeared.
EJB It was a passive interest ...
GG But it was a very significant passive interest. It's true that the first time my interest in NLP became public, it was in connection with ``computation in service to linguistics''. But I certainly didn't think that that was the be-all and end-all of NLP. However, as far as the future health of linguistics was concerned, I thought it was an important topic. And it is really quite extraordinary that now, eighteen years later, the only use that most linguists have for a computer is word processing. Their students aren't taught how to make sensible use of them, either. Linguistics operates as a typical humanities subject, despite its scientific pretensions. Of course, there are exceptions: phoneticians have used computers for decades, as have corpus linguists, and a significant proportion of morphologists also use computational tools these days.
EJB Well, in 1983 you were certainly ahead of the game. In 1983, as a linguist, if you were lucky or had spent some of your own money, then you you might have had a BBC micro or something in this country.
GG There would have been University PDP11s and central machines on the end of wires at that time.
EJB But they were unlikely to have been available as a generally accessible and user-friendly service around campus.
GG One had to make an effort but, nevertheless, it was beginning to be the case that if linguists put their mind to it, they could get access to computers that could do useful things for them. The Linguistics special issue contained a variety of suggestions. But most linguists don't want to know.
EJB Well, the Alvey Natural Language Tools (Briscoe et al. 1987) were at one point being used at Edinburgh, Lancaster, York, and Cambridge to teach with. That died and I think it died as GPSG was perceived as no longer the hot thing to work on. But it's not completely true that there is nothing around now. What is interesting to me is that what's around tends to be software that is developed more from a pedagogical point of view. For example, there is teaching software which allows students to draw trees and check their well-formedness. Such a facility is rarely found embedded in full scale parsing systems. Arguably, the Alvey Tools system, as it was developed as an open source engineering tool, wasn't the perfect vehicle to teach people about syntax. But you're right, of course that we're talking about a small minority of linguists and syntax courses.
GG ProGram was also quite widely used but it was constantly misinterpreted as a bit of real computational linguistics rather than as a tool, which was kind of irritating.
EJB Yes, ProGram was largely taken up by people who were really doing NLP and then dismissed by them as a rather baroque way of deploying useful grammars.
GG The other thing was it was the 1982 version of GPSG and not the 1985 version. And it wasn't kept up to date - for good reasons.
EJB In 1985 you became professor of computational linguistics. And, in 1989, you produced the textbooks NLP in { Prolog , POP-11, LISP}: An Introduction to Computational Linguistics with Chris Mellish (Gazdar & Mellish 1989a,1989b,1989c). Obviously that had been in gestation for a while. Did this mark a turning point or was this a logical progression from what came before?
GG Well, I think it was a progression. The history of my appointment was that I had been a lecturer in linguistics and then I got promoted to reader in linguistics in 1980. But, in the early 1980s, I was looking for an opportunity to change my departmental allegiance. Anyway, there came a point where the linguistics department was having a hard time getting any students, which wasn't an uncommon problem during that period. We had a one-to-one staff-student ratio - I'm joking but it wasn't good.
EJB This was the era of Thatcherite reform of higher education?
GG I can't remember exactly when it was. It must have been 1983 or 1984, I suppose. Never slow to make a major self-sacrifice, I offered to relinquish half of my post in linguistics and move it to computing. As I anticipated, this turned out to be an extremely welcome offer which was immediately accepted. So, from 1984, I was half in computing and half in linguistics. Then John Lyons left Sussex for Cambridge and his chair was advertised. I didn't originally plan to apply. I didn't really want a chair in linguistics. I didn't think it was proper for me to apply since it I had written the the job specification for the post - I had stepped into John's administrative shoes after he left. However, my arms were severely twisted, both by John and by the then Pro Vice Chancellor, so I did apply. Once it was offered to me, I asked that it not be a Chair in Linguistics but a Chair in Computational Linguistics and that I retain my existing logistic position, i.e., half in linguistics, half in computing. And the VC agreed.
EJB And the books?
GG Given that I then had that job title, it seemed to me that it was time that I did something to deserve it. Chris and I had been alternating the teaching of the NLP course for a few years and we had written a little tutorial/survey paper (Gazdar & Mellish 1987) together for New Horizons in Linguistics 2 and that had gone well in terms of the two of us collaborating. So I guessed that collaborating on a book with Chris would be fine. As, indeed, it turned out to be. I can't imagine a less stressful person to collaborate with. I don't know what I was like but he gave me no grief at all. It was very different from the GPSG book. With Chris, it was just ``Okay, so you draft this chapter''', ``I'll draft this chapter'', ``I've done this one, could you go through it now?'', and so forth.
EJB Was that a reflection of a difference of personality or a difference of field and assumptions about how one behaves?
GG I think it was partly personality. But also the fact that it was a textbook and neither of us had any ego involved in it. There were odd bits of our own research in it, but only in passing and it wasn't intended to market our own ideas or our own work. It was very much intended to be a certain sort of textbook and we completely agreed about that. We had the same common vision of textbooks where there was code that students could just pick and run with. That was the type of book we wanted to write. It was a lot of work, as books always are, but it was entirely an agreeable experience.
EJB And a book that was used quite substantially, I think. Like all such books, it had a limited shelf life. Although, actually, I still think it is strong in some areas and has not been surpassed or replaced.
GG I think the field moved on rather than the textbook choice, if you see what I mean.
EJB Obviously the statistical stuff was crying out for a decent textbook treatment and now has something approaching that in Manning & Schütze (1999). If you and Chris were writing your book now, you would be writing a very different kind of book. James Allen updated his book (1995) with some of the statistical work. Nevertheless, in terms of just describing in detail, say how you implement a chart parser - something that somebody coming into the field with not necessarily a lot of computing background would find useful - there is no newer alternative
GG I've always been surprised at the extent to which the code in the books, especially the Prolog book, has been used professionally rather than just pedagogically. There were quite a number of academic projects where people said ``this parser is basically the chart parser in Gazdar & Mellish only we have added the following bells and whistles''.
EJB Well, Chris's Prolog code would probably have been as good as anybody's and a whole lot better than the code that most people could have produced.
GG True, but Chris barely wrote any of the code in the Prolog version of the book. He wrote all the POP-11 and LISP code. The Prolog code was either written by me or, in a couple of cases, adapted from things I'd picked up. And a bit of it was done by Bob Carpenter who TAed for us when we taught a course based on the book at Stanford. The reason I wrote most of the Prolog code was because I was half competent in Prolog, whereas I was less than half competent in LISP and my POP-11 was rusty. Chris can program in pretty much any language so, paradoxically given his history, it made sense for me to do the Prolog and him the other two. Actually, the LISP code was largely created automatically: Chris wrote a POP-11 -to-LISP translator.
EJB So through the years of your passive interest - I don't mean that in a pejorative way - did you program then or did you start programming effectively when you started to teach NLP ?
GG Oh no, I'd been programming for years - ever since I took a Fortran course when I was a postgrad at Reading. And, when I arrived at Sussex in 1975, I was presented with a copy of ``the silver book'' (Burstall et al. 1968), pointed to a teletype, and pretty much told to teach myself POP-2 - on a batch machine. A year or so later we had a PDP11 running Unix and Steve Hardy had developed POP-11 to run on it interactively through VDUs. At that stage, I even taught POP-11 at an introductory level. I've never been a really proficient programmer in any language but, over the years, I have cobbled code together in a dozen different languages.
EJB The context at Sussex in the 1970s was that you were teaching in a cognitive science programme and the students were expected to do practical work and programming. So it was largely in a teaching context that you were programming, was it? Did it impinge on your research programme at all?
GG It didn't impinge on my research, narrowly construed. But Geoff , Ewan and I did a huge computer-based bibliography of linguistics in the 1970s (Gazdar et al. 1978) and I did all the shell and awk scripts that that project entailed.
EJB Do you think that is something that people coming into linguistics should be taught, even if they don't plan to work in computational linguistics? Is it the kind of skill that goes beyond efficient bibliography production and provides a productive intellectual mind set? Is it useful or essential or ...
GG Well, yes, but I would put a particular spin on it. I think that it would do no harm if every undergraduate in any discipline was taught Unix. Unix is an object of considerable beauty. It has an spare elegance which has stood the test of time. When you have learned the overall architecture and become familiar with the tools it provides, then you can do anything you want. So I think knowing your way round Unix is a tremendous asset for all kinds of things including linguistics. Just sorting through data that you've collected of some obscure language. Being able to to do arbitrary queries on your data.
EJB Okay, so you can use grep and awk and develop very quickly the script that fits your query. That's all still at a practical level. I guess I was thinking ...
GG No, it's a bit more than that. I do think that conceptually Unix is a very elegant object. If you are imbued with its ethos and aesthetics, then it stands as a shining example of how things should be done.
EJB I would have guessed you would probably have said Prolog, not Unix.
GG I love Prolog too, but Prolog is a much more restricted tool in a way.
EJB Yes, but it's kind of cleaner as well. And, if you think having a good understanding of logic is a fundamental skill, then learning to program in Prolog will help to provide that. Not necessarily writing practical algorithms but just learning in a rather pure way the sort of algorithms that you can implement in Prolog. There seems to be one route for people into that.
GG Yes, I wouldn't disagree with that.
EJB The title of the text book Natural Languages Processing in X: was that Chris's choice or was that a joint decision? Do you actually see much difference between calling it that or in calling it Computational Linguistics?
GG Well, it had a subtitle which was Introduction to Computational Linguistics. So we put both expressions in the title quite deliberately because, for things like string searches and indexing, we wanted to have both phrases there. I think we both thought that, at least in the anglomerican context, the two phrases are essentially synonyms. They are not quite synonyms in Germany, say, where there was a tradition of computational linguistics which was not about natural language processing in the sense which you and I understand it. But the ACL's own journal is called Computational Linguistics and that is obviously about natural language processing.
Copyright © Ted Briscoe & Gerald Gazdar, Wednesday 2 May 2001