Copy Link
Add to Bookmark
Report
NL-KR Digest Volume 02 No. 18
NL-KR Digest (3/20/87 19:31:33) Volume 2 Number 18
Today's Topics:
proper domains of grammatical research
Tinlap3 Conference Report
dictionaries with part-of-speech noted
Seminar - Circumscriptive Theories (SU)
----------------------------------------------------------------------
Date: Thu, 19 Mar 87 11:28:51 EST
From: Bruce Nevin <bnevin@cch.bbn.com>
Subject: proper domains of grammatical research
Re: Proper domain(s) for linguistics
First, we are talking about the proper domain of grammatical research,
not of linguistics, which opens quite a broad umbrella. The original
question even limited the question to particular grammatical formalisms.
Perhaps an unstated assumption made my message unclear.
When I said the `proper domain' of grammatical research is not sentences
but discourses, I thought it obvious that sentence structure is subsumed
in a study of discourse structure. In answer to the question whether a
sentence is `always the proper datum for doing GB, GPSG, or other
roughly related research,' my claim was that this is too narrow a
domain, and that `sometimes' you have to deal with structures across
sentences of a discourse.
Unfortunately, discourse analysis is for most researchers limited to
conversation analysis and problems of coreference, and the deeper
regularities of discourse (`information structures') of which these are
but a part are almost universally ignored. (I gave references about
this in my response of March 4.)
To use Michael Maxwell's example, it would be improper to limit
grammatical research to the study of inflectional morphology.
The further point was that some of the regularities of discourse
`invade' the sentence level. Why is it that the examples I gave are so
utterly wrong as sentences? Each conjunct is perfectly fine
morphologically, syntactically, and semantically. How would you state
the restrictions on the conjunction `and' to capture facts such as
these? I conclude that one cannot satisfactorily account for sentences
(including in your account such restrictions as these) without a
logically prior account of discourse.
One reason an account of discourse is logically prior is as follows: it
is an empirical fact that sentences are not directly observable in a
language. A language is observable as a set of discourses. Discourses
can be segmented into sentences because there are recurrent regularities
in them. More exactly, discourses can be segmented into sections in
such a way that each section is an instance of one of a set of
well-formed sequences of word classes (or word-class sequences or, in
exceptional cases, an initial segment of such a well-formed sequence),
and it is these sections that are called sentences. For the domain of
research to be limited to sentences, some prior science would have to
determine sentencehood. Native speaker intuitions have been taken as
this `prior science', but this is merely a convenient (and necessary!)
heuristic. Results obtained in this way must always be verified on
formal grounds to have empirical standing. Rationalist and
biologistical claims to the contrary, there is so far no science of
intuition, and even as data linguistic intuitions are notoriously
unstable and difficult to replicate.
To forstall possible further misunderstanding, let me emphasize that I
believe intuitions are of vital importance in all science, and that I do
not expect to get a discovery procedure to derive a grammar from a
corpus of discourses.
By the way, why is it so hard to define the proper domain of physics?
All such definitions are subject to change as sciences evolve, of
course, but they need be neither arbitrary nor absurd. Philosophers of
science, including practicing scientists, do it all the time.
Bruce Nevin
bn@cch.bbn.com
(This is my own personal communication, and in no way expresses or
implies anything about the opinions of my employer, its clients, etc.)
------------------------------
Date: Thu, 19 Mar 87 11:40:11 MST
From: yorick%nmsu.csnet@RELAY.CS.NET
Subject: Tinlap3 Conference Report
Tinlap3
Las Cruces, New Mexico
January 7-9, 1987
Brief Conference Report
Yorick Wilks, Program Chairman.
The Tinlap3 workshop was held as scheduled on January 7-9 1986
with the financial support of the Association for Computational
Linguistics, the National Science Foundation, American Associa-
tion for Artificial Intelligence, the Association for Computing
Machinery, its special interest group SIGART, and the Computing
Research Laboratory at New Mexico State University. The schedule
was adhered to and the preprints for 43 of the 48 papers were
circulated to all invitees and registrants ahead of the workshop.
There were 170 registrants (in addition to the 41 invited panel-
ists) of whom about half were graduate students from the US and
abroad (countries represented by non-invited registrants included
Japan 5, UK 4, Ireland 4, France 3, Brazil 1, Germany 2, Mexico
2). A remarkable feature (as in previous Tinlaps) was the out-
standing quality of the audience, some of whom were more dis-
tinguished than the panels and who contributed frequently and at
length to the discussions (e.g. Dana Scott of CMU). Such audience
members made a magnificent contribution to the success of the
whole. The location proved no inhibition to coming, and most of
the visitors commented on the excellence of the weather and sur-
roundings.
Content-The content and intellectual results of such meetings are
notoriously hard to assess. Even though the program and partici-
pants had been chosen in the most distributed and objective way
possible (with the extensive input from the ACL executive and its
book series editorial advisory board), nevertheless there were
individuals unhappy that certain issues were assigned time for
panel discussion at all. That is inevitable; yet it is hard to
imagine how any process could have produced a program of topics,
or even panel members, more representative of the state of the
field at the present time. There were, again, grumbles heard that
the panels came so obviously from non-randomly-selected parts of
the country, with very high input from the San Francisco Bay
Area. That again is inevitable if such panels are to gather the
best people, since there is, at the moment, a high concentration
of CL talent there, and the fact cannot be ignored. The format
closely followed that of previous Tinlaps except that at Tinlap2
there were some invited papers in addition to panels. Some few
attendees in 1987 hankered after that but most accepted that that
could introduce an unfortunate two-tier status into the proceed-
ings. In the event, two participants in the connectionism panel
were given about 20-25 minutes each to explain their systems in
more detail than their position papers allowed. This was entirely
justified, given the novelty and complexity of detail of that new
approach, and well within the discretion of that panel chair. But
on the whole I think that the Tinlap1 and Tinlap3 formats found
general acceptance: they discouraged the selling and promotion of
researchers systems (for which there are many opportunities
throughout the conference season) and encouraged concentration on
fundamental and general issues. Chairs were given total discre-
tion over the conduct of their sessions, although I gave them
some guidelines, pointing out that, for example, it would be
better to sacrifice restatement of position papers time in favor
of panel interaction and audience participation. Some chairs
showed more control over their panelists than others, but this
was adhered to in general, and many of the floor contributions
were outstanding.
Some very personal reactions to the panels Panel I (Words and
Worlds) This brought more to the center of CL what had been a
peripheral topic. It emphasized that real dictionaries are a real
world knowledge resource as well as a linguistic one, if extrac-
tion algorithms can be devised to get at what they contain. It
was argued strongly that AI/CL will not be able to hand code the
whole world and language from scratch in the next few years, and
to achieve real scale it will have to use the kinds of resources
that existing dictionaries and encyclopedias constitute. Panel
II (Unification grammar) This was a meaty session, largely in-
formational for the audience. Those who object to the whole
school made few protests and it was a very positive session that
brought out how much the current practitioners of context free
grammar are a family, sharing insights, whatever their differ-
ences of notation. The key dispute was over whether unification
is a programming language for grammar or makes theoretical claims
as well. Panel III (Connectionism) This was the hottest topic
and in some ways the liveliest. Dave Waltz as Chair gave an ex-
cellent overview of the issues and theories in a tree diagram of
intellectual inheritance. The key issues were refining the types
of connectionism there are about, what, if anything, they share,
and whether there are any firm criteria for defining the ``local-
ist'' version that would force inclusion of what, say, Charniak
or Lehnert do, into that category. Panel IV (Discourse theory)
An excellent positive session, this emphasized theoretical ad-
vance, and ways in which AI/CL innovations have passed quite
beyond the linguistic and philosophical foundations of the sub-
ject. Panel V (Why has NLP made so little progress) This was an
airing of grievances for those who had been seeking some kind of
theoretical showdown, but was unsatisfactory. However, it did
have the virtue of pointing up how tired and outdated the argu-
ment about the ``power of syntax'' versus ``semantics'' is!
Panel VI (Formal versus commonsense semantics) This was another
session where strong views were aired, but some degree of con-
sensus emerged here that the dispute about the role of logic in
CL is more a commitment to a certain kind of theoretical language
of description than to any strong claims about explanatory
theories, and that was progress of a sort. Panel VII (Reference)
Theories and their powers were compared and explored in this po-
sitive, consensus-seeking, session. Furthermore, a philosophical
challenge, that reference is really the province of philosophers,
was successfully fought off. Panel VIII (Metaphor). There was
considerable consensus here, in a way there would not have been
eight or ten years ago, that metaphor and its computational ex-
plication are central and not peripheral to NLP. Panel IX (Gen-
eration) Another excellent session, this sought consensus among
notations and claims and in particular about the key issue of how
much commitment a machine or human generator must make to content
and form and at what point in the process.
Domestic detail The preprints contained a number of misprints
(most of these present in the hard copy as originally submitted
by panelists). A corrected copy of the preprints (to include all
the missing ones that were distributed only at the workshop it-
self) is being assembled now, one to which authors will be invit-
ed to submit a corrected version of their position paper. The
Computing Research Laboratory at NMSU itself assembled and print-
ed the preprints used at the workshop, largely because of the
shortage of time to distribute the finished product, and the fact
that that method enabled the distributing process to be carried
out more easily. However, the revised and updated set will be
reprinted by the Association for Computational Linguistics and
will be made widely available by them. The oral presentation, the
panel discussions, and some of the audience discussion will be
transcribed and the panelists will be consulted later to see if
they are interested in some later book form of publication to a
yet wider international audience. A straw poll at the workshop
suggested that nearly all of them were. In addition, a private
company videotaped the workshop and will make commercially avail-
able to institutions and individuals a record of those panels for
which the participants are prepared to sign releases. To make
life simple, the Computing Research Laboratory signed a release
of its interest to the company for no fee. All in all, the
results of the Tinlap3 workshop should be very widely available
in a variety of forms. The physical arrangements of the workshop
seemed smooth and relatively free of hitches. Given the high de-
gree of external financial support for the workshop from spon-
sors, the registration fee was kept very low (and even lower for
students), certainly way below the actual marginal cost of pro-
viding for each registrant. By far the largest expense was the
travel and accommodation costs for all invitees, though again the
costs there were kept down by the special rate the local hotels
gave to participants, and the fact that many invitees got APEX
rates.
Conclusion Everyone seemed to have an excellent time, intellectu-
ally and socially. The truth of the matter may be that CL (and
that part of AI) is not now in an exciting period, but a profit-
able consensus-building one. From a scientific point of view,
however, that may be a great asset.
------------------------------
Date: 19 Mar 87 19:18:15 GMT
From: dave@mimsy.umd.edu (Dave Stoffel)
Organization: U of Maryland, Dept. of Computer Science, Coll. Pk., MD 20742
Subject: dictionaries with part-of-speech noted
I recently queried the net community about computerized
dictionaries which contained part-of-speech information. Here's
a digest of the responses.
----
>From the Oxford Text Archives:
Oxford Advanced Learner's Dictionary of Contemporary English
Collins English Dictionary.
>From ?
Webster's Pocket Dictionary (Amsler's thesis used this one)
Longmans Dictionary of Contemporary English.
>From Gage Publishers:
Gage Canadian Dictionary
----
Automated Language Processing Systems
190 West 800 North
Provo, UT 84601
Tel. (801) 375-0090
They have a wide variety of machine readable dictionaries (in several
languages). They are not on USENET but you could get in touch with
them by telephone or mail. Talk to either Robert Goode or Logan Wright.
----
You may wish to consult a report by Robert Amsler on computerized
dictionaries that appeared in the Annual review for Inf Sc and Tech
Vol 19, 1984, pp 161-209.
----
A book you may be interested in:
Erik Akkerman
Pieter Masereeuw
Willem Meijs
1985
Designing a Computerized Lexicon for Linguistic Purposes
ASCOT Report No. 1
Rodopi
Amsterdam
A comparison of the Longman Dictionary of Contemporary English and
the Oxford Advanced Learner's Dictionary for the purposes of NLP
research.
Both dictionaries are apparently available on tape, and both have part of
speech info included. (The report favors Longman's dictionary.)
--
Dave Stoffel (703) 790-5357
seismo!mimsy!dave
dave@Mimsy.umd.edu
Amber Research Group, Inc.
------------------------------
Date: 18 Mar 87 1142 PST
From: Vladimir Lifschitz <VAL@SAIL.STANFORD.EDU>
Subject: Seminar - Circumscriptive Theories (SU)
[Excerpted from AIList]
CIRCUMSCRIPTIVE THEORIES
Vladimir Lifschitz
Thursday, March 19, 4pm
Bldg. 160, Room 161K
The use of circumscription for formalizing commonsense knowledge and
reasoning requires that a circumscription policy be selected for each
particular application: we should specify which predicates are
circumscribed, which predicates and functions are allowed to vary,
what priorities between the circumscribed predicates are established,
etc. The circumscription policy is usually described either informally
or using suitable metamathematical notation. In this talk a simple and
general formalism will be proposed which permits describing circumscription
policies by axioms, included in the knowledge base along with the axioms
describing the objects of reasoning.
------------------------------
End of NL-KR Digest
*******************
