nl-kr-request@CS.ROCHESTER.EDU (NL-KR Moderator Brad Miller) (12/01/88)
NL-KR Digest (11/30/88 21:06:43) Volume 5 Number 33
Today's Topics:
Subjectivity, Mud-slinging, Behaviorism
was predictive knowledge...
best grammatical theories
Generation and Recognition of Affixational Morphology (Unisys Seminar)
Submissions: NL-KR@CS.ROCHESTER.EDU
Requests, policy: NL-KR-REQUEST@CS.ROCHESTER.EDU
----------------------------------------------------------------------
Date: Wed, 9 Nov 88 20:06 EST
From: Clay M Bond <bondc@iuvax.cs.indiana.edu>
Subject: Subjectivity, Mud-slinging, Behaviorism
Walter Rolandi:
>Sure Rick. I think that the body of knowledge that is linguistics is
>somewhat obscure and scarcely known to anyone outside of the field.
>When linguists speak, mostly just linguists listen.
This is not without some element of truth, but only if you leave out
cognitive/computational/psycholinguists. Linguistics, at least that
done by computational linguists is more than scarcely known in AI, for
example, as is psycholinguistics in psychology, and cognitive linguistics
in psychology/neuroscience/compsci. As far as who listens when "linguists
speak" as much responsibility falls on the (potential) listener as does
the speaker. That's not blame that can be laid solely at the door of
linguistics.
>I am suggesting that
>this would not be the case if the discipline were to address questions
>which, if answered, would be of interest to the general scientific community.
>More specifically, I am suggesting that linguistics will take a giant
>step towards science when the field embraces causal analysis and
>controlled experimentation. This is the stuff of all real science.
And just who is to determine which questions are of interest? It is true
that analysis and experimentation has played next to no role in linguistics
*until recently*, but again, who determines what is "real science"? Besides
you, that is. And you can produce bullshit just as easily with "real
science" as you can with "irreal (?) science".
>Silly me, I supposed that to mean that pragmaticists were busy producing
>a body of data that defines the causal relationships between the
>people, places, and objects around us that effect [sic] the contents of what
>we say. What I found instead was a lot of pompous intellectualizing on
>the nature of "communication". _hat I have seen is little more than
>mentalistic, philosophical, even literary analysis of things like
>"information", "messages", and "intentions".
Pompous intellectualizing! Not a wee bit resentful here, are we? I can
sympathize with your objection to mentalism/philosophy/literature masquerading
as science, but I don't think you can remove information/messages/intentions
from language without throwing the baby out with the bath water. If language
is none of these things, then what, pray tell, do you think it is? When
you use language does it contain no information/messages/intentions?
>What's wrong with pragmatics and linguistics in general is that neither
>field addresses any issues of scientific import.
According to whom? What exactly has given you the authority to make this
bogus and unsubstantiated claim?
>Who cares about
>anti-homophones, u-umlaut and y (or rather, u-double-dot and y), and
>embedded, recursive prepositional phrases?
A lot of people, obviously. Your statement had no purpose but to be offensive.
>When you know all about
>these things, what do you know? I want to know why people say the
>things that they say and why their utterances take the forms that they do.
When you know all about particle physics, what do you know? The answer is
the same. I also want to know why people say the things that they say.
And the things that they say are intrinsically bound up with information,
messages, and intentions, those things you thumbed your nose at earlier.
>I want a scientific answer. Is it unfair or unkind of me to ask this
>of linguistics?
It is if you don't ask or require it of yourself. Double standards are
never fair or kind.
>>(BTW, use the term 'linguist' to refer
>>to linguists in general. If you are just complaining about generative
>>linguists, then say so.)
>
>I stand accused.
You do indeed, twice. And at least once by one who is not a generative
linguist.
>I want it [linguistics] to aspire to
>substance: give me a causal explanation of verbal behavior.
You cannot explain behavior without explaining cognition. Without cognition,
behaviorism *is* push-pull mechanism, which gives no more causal explanation
than does mentalism. Behaviorism without cognitive theory has no explanatory
power. I suggest that you find out a bit more about what is going on in
linguistics before you sling mud again. Chomsky is not the only linguist
out here.
--
<< Clay Bond -- IU Department of Leath-er, er, uh, Linguistics >>
<< bondc@iuvax.cs.indiana.edu AKA: Le Nouveau Marquis de Sade >>
<< {pur-ee,rutgers,pyramid,ames}!iuvax!bondc
------------------------------
Date: Fri, 11 Nov 88 14:14 EST
From: w.rolandi <rolandi@gollum.UUCP>
Subject: was predictive knowledge...
In response to Clay's:
>Please object. But if you have no alternative to offer which
>addresses your complaints by correcting what you perceive to
>be deficiencies, you are contributing nothing to anything. Or
>to be more colloquial, put your money where your mouth is.
What on earth do you think I've been talking about? From the onset,
I have been suggesting that the science of human behavior, behavior
analysis, offers specific advantages, both in its causal model and
in its experimental methodology for addressing the deficiencies of
linguistic knowledge.
Ironically, your emotional responses have unwittingly served to reinforce
my impression that linguists possess an impoverished understanding
of the purpose, spirit, and methods of science. I have stated
methodological and epistemological criticisms of linguistics. Curiously,
you do not address my specific complaints. Forfeiting an opportunity
to cite data that would prove me wrong--on my own terms, no less--you
act as if you are stepping forward to avenge the honor of an unjustly
maligned loved one.
Your first response questioned my authority to make such statements,
as if facts are untrue unless cited by authorities. Your second response
essentially entreats me to find something nice to say about linguistics
or to refrain from saying anything at all.
What can I say? I am talking about methodological issues associated
with obtaining a predictive knowledge of natural phenomena. The problem
is that I am talking specifics in an audience that is oblivious to the
general plan.
Who else but,
Walter Rolandi
rolandi@ncrcae.Columbia.NCR.COM
NCR Advanced Systems Development, Columbia, SC
------------------------------
Date: Fri, 11 Nov 88 16:11 EST
From: Avery Andrews <munnari!fac.anu.oz.au!andaling@uunet.UU.NET>
Subject: best grammatical theories
> How can you tell if one theory of syntax is better than another?
>Walter Rolandi
>rolandi@ncrcae.Columbia.NCR.COM
>NCR Advanced Systems Developmbou, Columbia, SC
Well, to begin with, you have to define what the NL syntax is actually doing.
(which I will henceforth call `grammar', to distinguish it from various other
things that people sometimes call `syntax'). I'd suggest that a grammar is a
scheme whereby a classification of the the vocabulary of a language into a
finite collection of (in general overlapping) categories, plus some additional
principles, is used to indicate how the meanings of the words (or formatives,
or morphemes) are to be put together. In a PSG-based account of English, for
example, we would say that there were lexical categories, or `parts of speech',
such as Det, Adj, N and V, and phrase-structure rules such as
S -> NP V (NP) (NP)
NP -> (Det) Adj* N
(grossly oversimplified, of course)
We'd also say something to the effect that NP's refer to entities, that their
components indicate the properties of the entity referred to, and that the V in
an S indicates the relation between the entities referred to. How this is said
depends on what approach to semantics we're taking.
So the PS rules and the semantic principles tell us how to sort out the
attributes between the entities in a sentence like
The naive herpetologist handed the beautiful princess a worried-looking
crocodile
Other languages employ different sorts of principles, often involving
grammatical feature-based methods such as case-marking and agreement, and other
theories describe the facts in rather different ways--categorial grammar, for
example, builds virtually all of the language-particular aspects of grammar
directly into the system of categories, with the rules being universal
principles (function application, composition, and type-raising, for example).
But these variations are always based on classifying the vocabulary into a
finite set of classes (including those defined by features).
Generative grammars frame these principles in such a way that they interact so
as to assign each string of words a possibly empty set of grammatically
legitimate ways of composing the meanings of its words. Ungrammatical strongs
have none, ambiguous ones more than one. (By speaking of ways of composing
word-meanings, rather than of meanings, we allow for a degree of fluidity in
the lexicon, and for the use of contextual knowledge to discover the meaning of
unknown words by reverse engineering.) In a successful utterance, knowledge of
context + common sense is sufficient to allow the hearer to choose the intended
meaning from the grammatically + lexically legitimate ones.
It turns out that there are lots of alternative ways of presenting such
principles, which are *not* created equal. We could replace the PSG above with
the following finite-state grammar (FSG):
S = (Det) Adj* N V ((Det) Adj* N) ((Det) Adj* N)
(Technically, this isn't really an FSG, since it describes the arrangement of
word-classes rather than individual words. Maybe it should be called a
Word-Class Grammar (WCG). Linguistically, FSG's have all the faults of WCG,
and more besides).
Although this generates the same strings as the PSG, and could be equipped with
an equivalent semantics, it can be read as making very different claims about
the causation of linguistic behavior.
The PSG says that there are two distinct causal factors, S-structure, and
NP-structure, which interact in a certain way, essentially, the NP-structure
principles govern the internal form of some sequental blocks of words, whose
external arrangement is determined by the S-structure principles. The FSG says
no such thing--there is just a big formulae saying what order words appear in.
The PSG says that there is one causal factor behind the relative order of Det,
Adj and N in the three positions, while the FSG says that there are three
different ones, which is definitely wrong. (If they were different, why would
they have the same effect, in position after position? And why would similar
phenomena appear, with different details of order, in language after language?)
PSGs are better than FSGs, but they get into trouble quickly, for example, with
subject-verb agreement, which, for `classical' (e.g. feature-free) PSGs,
requires that many rules be split into `singular' and `plural' verions. eg.:
S -> {NPsg Vsg (NP) (NP),
NPpl Vpl (NP) (NP)}
NP -> {NPsg, NPpl}
NPsg -> (Detsg) Adj* Nsg
NPpl -> (Detpl) Adj* Npl
What's wrong with this is that it says that there are different causal factors
behind the order of words in singular S and NP than in plural ones, and this is
clearly wrong: phrases differing properties such as number, gender or case
show essentially the same gross internal structure.
There are many known cures for this ailment: agreement transformations were
the first formal proposal widely accepted by linguists. Various kinds of
feature-sharing principles are today most popular, especially among those with
computational inclinations. All of these solutions have in common that they
treat the arrangement of the agreeing elements as one factor and the details of
what is shared as another.
Wh-movmement is a third example: here we have superposed on the basic
machinery for assigning semantic roles (on the basis of agreement, word order,
case-marking, etc., depending on the language) some additional machinery that
allows constituents to appear elsewhere than their basic expected position,
subject to interesting constraints. Many ways of accomodating the most obvious
facts are now known: transformational movements or deletions, coindexing
schemes, `slash-feature' manipulations, `functional uncertainty', functional
composition+type raising are some current contenders.
Thus one major desideratum a grammatical theory is to allow one to formulate
principles that correspond accurately to the underlying causal factors behind
grammatical phenomena, and correctly predict their interactions. The
corresponds pretty accurately to what Chomsky has sometimes called `descriptive
adequacy' (he has unfortunately also applied this term to less interesting
properties, such as simply generating the right set of sound-meaning
correspondences, without regard to whether it has been done right.) It also
corresponds pretty closely to being a pleasant language to specify grammars in.
Current grammatical theories (GTs) use a wide range of mechanisms to say to a
considerably extent the same kinds of things about how grammatical phenomena
are to be carved up, They thus have a wide range of overlap in the phenomena
they handle well. Nonetheless, they nonetheless differ significantly in how
well they approximate descriptive adequacy in various areas. LFG seems to me
to be unequalled for basic predicate-argument structure and the complement
system: proponents HPSG, GPSG, and the presently available varieties of
Categorial Grammar have not yet told in public any convincing story about
things like Quirky Case in Icelandic, or preverbal pronouns in French and
Spanish. The other theories seem to do better with Wh-movement and related
phenomena: proponents of LFG have not yet told in public a sensible story
about parasitic gaps or accross-the-board extraction. People tend to prefer
the theory that does best with the stuff they actually are working on at the
moment.
Beyond descriptive adequacy are somewhat murkier and more controversial
concerns, which tend to go under the label of `Explanatory Adequacy'. The most
generally accepted of these is the desideratum that a GT help explain why
language-learning is possible. The basis of this is the idea that the less one
has to specify to determine the language, the easier it is to imagine how
children can do it. The ultimate position here is that taken by Chomsky and
his followers in GB, who believe that an interesting portion of the grammar of
all languages (`core grammar') can be specified by setting the values of a
relatively small number of binary-valued (or maybe finite list-valued)
parameters. I find this idea pretty bizarre, but many linguists seem to accept
it.
A more modest claim is the difference between a transformational passive rule
and a `classic' LFG one:
TG Passive:
NP - X - V - NP - X =>
1 2 3 4 5
4 2 be+en+3 0 5+by#1
LFG Passive:
SUBJ -> OBLby/0
OBJ -> SUBJ
0 -> Pass.Part
The TG rule is a structure changing operation which has a complicated statement
because it has to recapitulate most of the basic facts of English clause
structure (the subject comes before the verb, the object after, auxiliaries
come between the subject and the verb, and prepositional phrases come after,
and before their NPs). And it doesn't *really* work anyway. The LFG rule is a
lexical rule operating on lexical items whose argument positions are associated
with grammatical relations: it creates a new lexical entry, altering the
grammatical relation/semantic role assignments and the feature composition of
the item as indicated. It is much simpler because it doesn't recapitulate
basic facts of clause structure, but depends on other components of the grammar
that account for these facts (the rules associating grammatical relations with
NPs, and connecting the feature-composition of verb forms with their
distribution). Furthermore, just about everything in it corresponds to an
actual parameter of variation in passive-like constructions: sometimes the old
subject is always suppresssed (the 0 option for SUBJ), sometimes never;
sometimes OBJ must be present and become SUBJ, sometimes not; sometimes the
grammatical category of the verb form become participial, sometimes not, It is
most unlikely than language learners would learn the TG rule, but it isn't
crazy to suppose that they could pick up the LFG one.
This concept of explanatory adequacy is actually closely connected with
descriptive adequacy, because the greater the extent to which the notation is
free from redundancy and formal fuss, the easier it is to imagine how a
learning procedure might successfully find its way through the space of
possibilities it provides. But there is still nothing in descriptive adequacy
that says that the GT shouldn't provide more facilities than one actually needs
to describe languages right, which this kind of explanatory adequacy does.
One can also ask that a GT illuminate actual facts about acquisition, rather
than just make it easier to imagine how it might be possible. Recent work by
Hyams and Pinker attempts to do this. But this requires an actual theory of
learning as well as a GT.
Another brand of explanatory adequacy pertains to processing. The way in which
a GT presents grammatical information can make a big difference to the
possibilities for using it in comprehension and production. TG's are pretty
terrible in this regard, many more recent theories are much better. (You can
get an LFG parser to run on your PC from the Institute for Machine Translation
at the University of Stuttgart; nothing like this is even envisionsable for
classical TG). A complicating factor is the possibility of compilation: maybe
language learning, and the halting performances of beginners, are based on
direct processing of some kind of augmented PSG, while the highly overlearned
performance of fluent speakers is based at least in part on the use of FSGs +
strategies generated by various kinds of compilation processes. Only
psycholinguistics will be able to sort that out. But the generalizations (e.g.
basic causal factors) detected by linguists still have to be taken as part of
the causal order, for otherwise there is no account of why they are there).
The import of Explanatory Adequacy is that the generalizations discovered by
linguists represent real causal factors. Thus, they ought to interact sensibly
with other kinds of causal factors, such as learning procedures and and
processing mechanisms.
There is at least one complicating factor. The generalizations (e.g. causal
factors) identified by grammatical analysis are supposed to reside in the minds
of language users. Because sometimes linguists get fooled. For example,
languages tend to undergo sequences of regular sound changes, which can turn
formerly regular and uninteresting systems of forms into things that look like
the results of applying long sequences of phonological rules to rather abstract
underlying representations. There is good reason to believe that there exist
things that look and act pretty much like standard generative phonological
rules (the original sound changes, at a minimum), but looking at say, the
residue of the great vowel change in English, it is hard to say what is
historical residue and what is genuine Modern English grammatical structure.
Chomsky + Halle's rules in SPE do roughly represent genuine causal factors in
linguistic behavior, but it's not so clear what there locus is: history, or
the brains of modern speakers.
Happily, most of syntactic and contemporary phonological theory seems
reasonably secure against this line of attack, but one has to rely on common
sense to ascertain what is the actual locus of causal factors revealed by
grammatical analysis. (If a group of people started speaking a language taught
to them by aliens, its grammatical generalizations would not be of direct
relevance to human GT, since their utlimate locus would be the minds of the
aliens, not the human speakers. On the other hand, the changes that human
speakers introduced into the language might be highly revealing for human GT.)
Avery Andrews
The Australian National University
andaling%fac.anu.oz@seismo.css.gov
------------------------------
Date: Sun, 13 Nov 88 07:28 EST
From: Greg Lee <lee@uhccux.uhcc.hawaii.edu>
Subject: best grammatical theories
In the NL-KR Digest,
Avery Andrews <munnari!fac.anu.oz.au!andaling@uunet.UU.NET> replies to:
>
> > How can you tell if one theory of syntax is better than another?
>
> >Walter Rolandi
> >rolandi@ncrcae.Columbia.NCR.COM
by giving a very useful and readable summary of some syntactic theories
and their problems. Among other things, he compares the "causal
factors" they ascribe to some example expressions. But there is
a little problem here. If we were to suppose that any theory
causes the facts that it describes, then it would be fair to refer
to the parts of the theory as "causal factors", I suppose, as he
does. However I doubt that Walter, or any reasonable person, would agree
to this supposition.
Andrews says:
>... generalizations (e.g. basic causal factors) detected by linguists ...
This is even worse. Having discovered a general fact, a linguist may
assume he has uncovered its cause, apparently, because it causes
itself. This completely trivializes the idea that human behavior
has causes and that we should search for them.
So I suggest replacing 'causal' with 'descriptive' in Andrews' article
everywhere it occurs. It reads better.
Greg, lee@uhccux.uhcc.hawaii.edu
------------------------------
Date: Thu, 17 Nov 88 16:15 EST
From: Chrystopher Nehaniv <nehaniv@oreo.berkeley.edu>
Subject: Re: best grammatical theories
There is a very pratical way of comparing syntactic theories:
Compare how they explain a huge set of sentences. Do an
analysis of each sentence in each of theories being compared.
No known theory will be always right. Perhaps no possible non-trivial
theory will ever be always right.
If one theory explains more kinds of sentences, it's good. If
they explain about the same number, look for simplicity and
elegance as a comparision. As a related criterion, consider how
easy is it to do an analysis using each theory. In some cases,
one may choose a theory that gets less sentences right if its
much simpler than the competing theory. All the factors should
be considered to decide which theory is better.
----------------------------------------------------------
No theory should be dismissed for making wrong predications - all do.
A theory should only be superseded by a better theory. It's like this
for chemistry and physics. To reiterate, if Chomsky's GB or GPSG or
what have you fails on some sentence that's no reason to stop doing
that theory.
In fact, many grammarians do not aspire to explain all sentences
anymore, but only a "core" of grammatical utterances. They expect fuzzy
applications of any theory on the periphery. (We all make ungrammatical
utterances sometimes even when trying to be formal. Also idiosyncratic
properties of certain lexical items may make them fail to fit in a reasonable
grammar).
C.L. Nehaniv | " Things fall apart.
Dept. of Mathematics | It's scientific."
UC Berkeley, CA 94720 | -D. Byrne
C.L. Nehaniv | " Things fall apart.
Dept. of Mathematics | It's scientific."
UC Berkeley, CA 94720 | -D. Byrne
------------------------------
Date: Fri, 4 Nov 88 22:44 EST
From: finin@PRC.Unisys.COM
Subject: Generation and Recognition of Affixational Morphology (Unisys Seminar)
AI SEMINAR
UNISYS PAOLI RESEARCH CENTER
John Bear
SRI International
Generation and Recognition of Affixational Morphology
Koskenniemi's two-level morphological analysis system can be improved
upon by using a PATR-like unification grammar for handling the
morphosyntax instead of continuation classes, and by incorporating the
notion of negative rule feature into the phonological rule
interpreter. The resulting system can be made to do generation and
recognition using the same grammars.
1:00 am - November 7, 1988
R&D Conference Room
Unisys Paoli Research Center
Route 252 and Central Ave.
Paoli PA 19311
-- non-Unisys visitors who are interested in attending should --
-- send email to finin@prc.unisys.com or call 215-648-7446 --
------------------------------
End of NL-KR Digest
*******************