Newsgroups: sci.lang
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!oitnews.harvard.edu!purdue!lerc.nasa.gov!magnus.acs.ohio-state.edu!math.ohio-state.edu!howland.reston.ans.net!newsfeed.internetmci.com!news.mathworks.com!tank.news.pipex.net!pipex!uknet!newsfeed.ed.ac.uk!edcogsci!davidt
From: davidt@cogsci.ed.ac.uk (David Tugwell)
Subject: Re: Chomksy, Significance, and Current Trends
Message-ID: <DDB8GB.9L5@cogsci.ed.ac.uk>
Organization: Centre for Cognitive Science, Edinburgh, UK
References: <4084i9$dml@newsbf02.news.aol.com> <DD18rC.AK0@actrix.gen.nz>
Date: Mon, 14 Aug 1995 16:57:04 GMT
Lines: 109


In article <40albm$dt8@ruccs.rutgers.edu>,
Jay Rifkin <jirifkin@ruccs.rutgers.edu> wrote:
> zohrab_p@atlantis.actrix.gen.nz (Peter Zohrab) writes:
>
> >In fact, the bases of Generative Grammar (the Competence-Performance
> >distinction, and native-speaker intuitions) won't stand up to one minute's
> >intelligent analysis.
>
> I'd be very interested in seeing your analysis.  One minute of your time
> doesn't seem like too much to ask.  Please, post it.
>  - Jay                                (jirifkin@eden.rutgers.edu)

I don't think I can be the only who shares Peter Zohrab's misgivings
about the Competence-Performance distinction and is itching to have
his minute's worth of saying why. And as a practising syntactician, I hope
I will manage to do so without too many sidetracks into (undoubtedly
important) issues such as men's rights, PC-ness etc...

Right let's take a (hopefully non-controversial) look at why the
distinction was introduced in the first place, rather than how it was
later justified (ideal speaker business etc.).

In the 50's Chomsky was (laudably) trying to build a formal model
characterizing the set of English sentences. As the initial basis for
this set, he took sentences speakers accepted as good English. Any
formal grammar built to characterize these sentences, however, would also
characterize sentences that speakers would agree to be non-English. To
take the standard example, there is no way a rule-based grammar of
the kind Chomsky envisaged could  allow a sentence with one or two
centre-embeddings, but rule out one with three or more, and so
neither Chomsky's nor anyone else's generative grammar can distinguish

(i) The man the dog bit died.

and

(ii) The man the dog the cat scratched bit died.

etc etc ad infinitum.

Now, this isn't really too much of a problem: there are good
methodological arguments for saying that although the model doesn't
quite characterize the data and overgenerates, it does capture
an important section of the data and in a concise and
linguistically-interesting way. 

But, instead of considering the model as a useful approximation of the
data, Chomsky makes a virtue out of necessity and insists that is
in fact a model of something else (linguistic competence) and is
therefore not deficient. So this useful, perfectly respectable and (at
the time) necessary methodological decision to be satisfied with a
first approximation of the data of study is instead built in as one of
the cornerstones of the linguistic theory.

The circularity of argumentation that this allows can be shown thus:
suppose I come up with an alternative grammar of English which does not
distinguish between the grammaticality of the sentences:

(iii) The cat sat on the mat.

and

(iv) Cat the sat on mat the.

Now this is a ropey old grammar, but any objections to it can be
countered by claiming that in fact (iv) is perfectly grammatical and
within the competence of an idealized English speaker, but is
unacceptable for (unspecified) processing reasons.

This might sound an extreme example, but I think it reflects the basic
feeling of unease that I am sure many people feel upon first coming
across an exposition of this founding tenet of generative
grammar. This is not to say that a lot of valuable work has not been done
in generative grammar over forty years but I think it is true that the
distinction has had a stifling effect as regards looking at genuine
language data. This has too often been replaced by ultra fine
``grammaticality distinctions'' of dubious validity and interest,
including such gems as ``it seems to me that sentence X is grammatical
but not acceptable while sentence Y is acceptable but not fully
grammatical'' etc.

But the numerous problems of acceptability/grammaticality intuitions are
well-known. The reality is that there are alternatives. Probabilistic
models combining structural and statistical information trained from
real (hand-parsed) texts CAN give a probabilistic thumbs down to
heavily centre-embedded sentences, they CAN characterize
phenomena such as heavy-NP shift etc, do NOT need negative examples
NOR any competence-performance distinction. And the only intuitive
judgements necessary are those of ``how should I understand this
naturally occurring sentence'' in which agreement between native
speakers is pretty solid and dependable (for otherwise language would
not be much use).

The problem for generative grammar is that being constrained at birth
to deal only with a nebulously and circularly defined linguistic
competence, the question of how to cope with real language data does
not even have to be addressed never mind solved. 

Well, I've got onto my own hobby-horse a bit, but I hope this will
help to keep the discussion going. I really hope some generative
grammarians out there will take the time (it does take rather more
than a minute it must be said) to defend the foundational honour of
their discipline and lay into me--I must say that I have hitherto found
it disappointingly difficult to provoke them on this rather
fundamental point.


David Tugwell                    
