Computational linguistics

[marke@bilby.cs.uwa.oz.au (Mark Ellison)]

Bob Chevalier (lojbab%snark.thyrsus.com) writes (personal communication):

> re your posting on sci.lang (I can't post to it, but read it; feel
> free to relay this if you think it germane and useful)
> I agree with all of your arguments in favor of the programmer point of view
> but think something has been left out of the definitions involved.  Isn't the
> purpose of theoretical linguistics at least partially to determine how
> HUMANs process language.  An optimal computational algorithm may be able to 
> perfectly model the >effects< of the human processing, yet may have nothing
> to do with the actual human algorithm (if it can be called an algorithm).
> I've been told that the same thing goes in logic theorem-proving.  There are
> powerful theorem provers that can match anything humans can do with symbolic 
> logic, but using algorithms that humans are not able to use.  Yet it is more
> difficult to do a theorem prover that can prove theorems using only, say, the
> rules taught in a first year logic class.
>
>  --  lojbab = Bob LeChevalier, President, The Logical Language Group, Inc.
>               2904 Beau Lane, Fairfax VA 22031-1303 USA 703-385-0273
>      lojbab@snark.thyrsus.com
>
> e'osai ko sarji la lojban.

An interesting point. Certainly post-structuralist linguistics seems to
have seen itself as concerned with discovering mentally real structures of
language. So we are left with the question

	Is the determination of the structure of language (ignoring the
question of psychological reality) a valid part of theoretical linguistics?

My opinion is yes. It has no where else to go. Determining psychological
realities, and/or the _mechanism_ of language processing seems to me to
have no greater claim to the attention of linguists than language
structure itself. For instance, few would claim (I hope) that the
psychological processing of movement should be a filter on acceptable physical
theories. Do you believe in the psychological reality of the General
Theory of Relativity? Is it part of a `Universal Physics' innate in each
of us?

If my contention is correct, and concise natural language processing systems
must reflect the structure of language, then their independence of human
cognitive methods is neither here nor there.

Re the theorem provers, I think it is more the case that mechanical
theorem provers use techniques that are too tedious for humans to use,
than ones that are too hard. Anyway, that's beside the point. Good theorem
provers will reflect the nature of the formal system about which theorems
are being proved. The aspects of the formal system which are used in the
programs need not be the same that are used by human theorem-provers,
and so the program need not parallel human provers, but they will probably
be ones important to the structure of the formal system itself. For this
reason it will be harder for a first year logic student to understand what
goes on in an automatic theorem prover than for a trained logician with
many years in the field. Carrying the metaphor through to linguistics,
I contend that a computational mechanism for processing language system
need not reflect the human processing `algorithm', but will reflect
scientific discoveries concerning the object itself. Perhaps the human
mechanism will also reflect these discoveries, but perhaps not. It may not
be a very efficient implementation.

marke

-- 
--------------------------------------------------------------------------------
T Mark Ellison      marke@wacsvax.cs.uwa.oz.au    Department of Computer Science
                                                 University of Western Australia
Never necklace an albatross.                                W.A. 6009, Australia