Newsgroups: sci.lang
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!newsfeed.internetmci.com!EU.net!uknet!newsfeed.ed.ac.uk!edcogsci!steve
From: steve@cogsci.ed.ac.uk (Steve Finch)
Subject: Re: Chomksy, Significance, and Current Trends
Message-ID: <DDGMxA.B3E@cogsci.ed.ac.uk>
Organization: Centre for Cognitive Science, Edinburgh, UK
References: <4084i9$dml@newsbf02.news.aol.com> <DD5CLH.2nJ@cogsci.ed.ac.uk> <40flg8$6cf@senator-bedfellow.MIT.EDU> <DD7F77.CKw@cogsci.ed.ac.uk> <40ioik$579@senator-bedfellow.MIT.EDU> <DDAtxL.1z4@cogsci.ed.ac.uk> <40niv6$rk4@senator-bedfellow.MIT.EDU>
Date: Thu, 17 Aug 1995 14:58:19 GMT
Lines: 90

David Pesetsky <pesetsk@mit.edu> writes:

>OK, so let's agree you're not taking a radical position.  But 
>then please desist from characterizing your work and ideas in 
>contrast to a parody of alternative approaches.  It is no 
>doubt misleading to some readers of sci.lang.

Differences in emphasis need to be expressed and a variety of
important quantitative, currently largely unaddressed questions need
to be considered.  Debate is debate and I'm surely not the only one
who uses rhetoric around here :-)

>But here's something else I'm curious about.  In the final 
>analysis, doesn't the excitement of corpus evidence for you 
>actually boil down to the degree to which statistical 
>properties of corpora coincide with categories provided by -- 
>guess what? -- acceptability judgments?  For example, when you 
>viewed it as interesting that statistics on collocation can 
>categorize and group words in a particular way, wasn't the 
>excitement for you coming not from the fact that you can do 
>statistics on corpora (pretty trivial, you can do statistics 
>on anything), but from the fact that the fruits of your 
>numbers mirror categorizations established by traditional 
>introspective techniques.  Likewise, how do we decide if a 
>proposal about PP attachment is getting the attachments right 
>except by matching how the proposal works with people's 
>judgments about acceptability under a particular 
>interpretation in a particular context?

The excitement has to do with how one can build systems which process
texts and speech, and what implications this might have for how humans
process texts and speech.  The increasingly worrying position from the
point of view of theory is that artifacts which look as though they
can process language well appear to have few of the entities which
have come out of theoretical research on language over the last 30--40
years.  In particular they seem to lack grammars (and this is not for
want of trying to build NLP artifacts which contain grammars).  It is
most interesting that the artifacts which are built can be analysed in
such a way that various traditional constructs become evident.  This
does not strike me as evidence one way or the other as whether, in the
child, knowledge of these constructs is derived from building
sophisticated language processing devices and analysing them or
whether innate knowledge of these constructs is brought to bear to
build language processing devices.  I note only that people need
training to identify linguistic constructs with any degree of
precision, that experimental evidience which supports one of these two
views supports the other, and that there seem to be a mightly large
number of non-equivalent formalisms which are used to explain the same
data.

Thus it is impossible to say whether our knowledge of grammar and PP
attachment in a particular case is derived from analysis of a
sophisticated an accurate language processor (and may indeed be a
consequence of some principles of construction necessary for humans to
build language processors), or whether it is derived from innate
knowledge, for example from a correctly parameterised universal
grammar.

To put it another way, it is impossible to know whether "grammar"
plays a central and causative role in our processing of language, or
whether it is just one way of analysing the sophisticated language
processing mechanism we all possess.  There are reasons to believe
(which I can go in to if pressed) why the latter may be the case.
This position, if true, certainly does not invalidate investigation of
grammars, since these are by hypothesis excellent ways of
investigating the langauge processing mechanism; however it certainly
does not imply that the way to make a language processing mechanism is
to start with a grammar whereas the former position would seem to
imply that language processing without a grammar is a foolhardy (and
certainly inhuman) enterprise.

There now seems to be something of a mismatch between theory and
engineering, to such an extent that NLP engineers now have very little
applicable theory to fall back on since they have nothing worthy of
the name "linguistically justified grammar" in their systems.  This
does not mean they have no theory themselves; they view language
statistically and build artifacts which "learn" how to process
language, and statistics has its own well worked out theory.  To apply
this theory we need to get a statistical handle on what is going on in
language, and the questions I posed earlier are an attempt to provoke
discussion on this topic.

Corpus linguistics without the intellectual/political divide is
primarily theories of natural language processing systems in general
(since language exists in and of itself) and human's natural langauge
processing mechanisms in particular. The relationship between
processing systems and linguistic constructs such as grammars needs
explanation and examination.

Steve.
