[Prev][Next][Index]
CSLI Calendar, 3 November 1994, vol.10:6
-
To: friends
-
Subject: CSLI Calendar, 3 November 1994, vol.10:6
-
From: Tom Burke <burke>
-
Date: Wed, 2 Nov 1994 13:42:11 -0800
C S L I C A L E N D A R O F P U B L I C E V E N T S
______________________________________________________________________________
3 November 1994 Stanford Vol. 10, No. 6
______________________________________________________________________________
A weekly publication of the Center for the Study of Language and
Information (CSLI), Stanford University, Ventura Hall, Stanford, CA 94305-4115
____________
CSLI ACTIVITIES DURING 3 -- 11 NOVEMBER 1994
THURSDAY, 3 NOVEMBER
10:00 - STASS Seminar
Cordura Hall, Room 100
An Introduction to Situation Theory
Keith Devlin, Saint Mary's College Mathematics
Abstract below
2:15 - CSLI Seminar
Cordura Hall, Room 100
Against the Canon: An Alternative to Canonical Representation
Ken Haase, MIT Media Laboratory
Abstract below
4:15 - SSP Forum
Building 60, Room 61-F
Spatial Thinking
Barbara Tversky, Stanford Psychology
Abstract below
7:30 - Berkeley/Stanford Phonology Workshop
2111 California Street, Berkeley
Nasal Consonant Harmony at a Distance: The Case of Kiyaka
Larry M. Hyman, UC Berkeley Linguistics
Abstract below
FRIDAY, 4 NOVEMBER
12:30 - HCI Seminar
McCullough Building, Room 134
Building Virtual Libraries on the Internet
Nick Arnett, Verity
Abstract below
3:30 - Linguistics Colloquium
Cordura Hall, Room 100
The Context-Dependency of Implicit Arguments
Cleo Condoravdi and Mark Gawron, CSLI and SRI
Abstract below
WEDNESDAY, 9 NOVEMBER
4:00 - Seminar on Computational Learning and Adaptation
Cordura Hall, Room 100
Density Estimation for Model-Based Computer Vision
Subutai Ahmad, Interval Research
Abstract below
THURSDAY, 10 NOVEMBER
10:00 - STASS Seminar
Cordura Hall, Room 100
Title and speaker to be announced
2:15 - CSLI Seminar
Cordura Hall, Room 100
The Digital Libraries Project
Terry Winograd, Stanford Computer Science
4:15 - Symbolic Systems Forum
Building 60, Room 61-F
Robots that Plan and Robots that React
Illah Nourbakhsh, Stanford Computer Science
Abstract below
FRIDAY, 11 NOVEMBER
12:30 - HCI Seminar
McCullough Building, Room 134
How to Support Design with Research When Potential Users
Can't Imagine Future Products
Bonnie Johnson, Interval Research
Abstract below
3:15 - Philosophy Colloquium
Building 90, Room 91-A
Reasoning in Hume
David Owen, University of Arizona Philosophy
Abstract below
3:30 - Linguistics Colloquium
Cordura Hall, Room 100
Underspecification by Optimization
Sharon Inkelas, UC Berkeley Linguistics
Abstract below
____________
The CSLI Calendar appears on Thursday of each week throughout the academic
year. Announcements, abstracts, and other information to appear in the
Calendar on a given Thursday should be submitted by e-mail to
incalendar@csli.stanford.edu by 5:00 p.m. on the previous Tuesday.
Past issues of the CSLI Calendar, a quarterly schedule of upcoming CSLI
events, and other information about CSLI are available on the World Wide Web:
<http://csli-www.stanford.edu/>. The Calendar, with available abstracts, is
also posted each week to the csli.bboard newsgroup.
____________
STASS SEMINAR
on Thursday, 3 November
10:00 a.m., Cordura Hall, Room 100
An Introduction to Situation Theory
Keith Devlin
Saint Mary's College Mathematics
<devlin@stmarys-ca.edu>
Keith Devlin continues his series of introductory lectures on situation
theory. In today's lecture, he will start to present a new framework for
analyzing discourse. This is work in progress. The aim is to develop a
situation-theoretic framework that can capture the insights of three well
known traditions: (i) situation semantics, (ii) speech acts theory, and (iii)
conversation analysis. Background reading remains the same as before: "Logic
and Information", by Keith Devlin (Cambridge University Press, 1991).
____________
CSLI SEMINAR
on Thursday, 3 November
2:15 p.m., Cordura Hall, Room 100
Against the Canon: An Alternative to Canonical Representation
Ken Haase
MIT Media Laboratory
<haase@media.mit.edu>
An AI program without canonical descriptions seems a little like a car without
wheels: we have to wonder how it goes anywhere. Canonical descriptions enable
the specification of general procedures for inference, induction, and planning
without reference to the specifics of the domain of operation. However,
canonical descriptions are brittle when assumptions of reference or
correspondence change; consequently, much of the failure of AI systems to
scale can be blamed on the need to devise new representations and the parsers
or other programs for generating them. Over the past three years, we have
been exploring a model of representation where analogies between non-canonical
descriptions provide the structure usually provided by syntatic conventions
over canonical descriptions. Most of this work has focussed on story
understanding in the news domain: our current testbed consists of a diverse
database of a million words of short news summaries supporting various sorts
of semantic queries without any intervening canonical representation. I will
discuss this database, our plans for its future development, and the
consequences of a model of representation which eschews canonical description.
KENNETH HAASE is an assistant professor of Media Arts and Sciences at the
Massachusetts Institute of Technology. His main research interests are
knowledge representation and automated representational invention. His other
interests include machine discovery, story understanding, intelligent
interfaces, information retrieval, and philosophies of science and mind.
____________
SYMBOLIC SYSTEMS FORUM
on Thursday, 3 November
4:15 p.m., Building 60, Room 61-F
Spatial Thinking
Barbara Tversky
Stanford Psychology
<bt@psych.stanford.edu>
I will briefly review three projects in spatial thinking. The first
demonstrates systematic errors in cognitive maps, and uses those findings to
develop a theory of memory for space. The second explores the efficacy of
language in evoking spatial representations and reveals the way people think
about the space immediately around themselves. The third investigates
cognitive principles underlying external graphic representations, such as
diagrams, graphs, and charts.
____________
BERKELEY/STANFORD PHONOLOGY WORKSHOP
on Thursday, 3 November
7:30, p.m., 2111 California Street, Berkeley
Nasal Consonant Harmony at a Distance: The Case of Kiyaka
Larry M. Hyman
UC Berkeley Linguistics
<hyman@garnet.berkeley.edu>
In a number of Bantu languages, certain voiced consonants, e.g., the [d~l] of
the applicative suffix /-id-/, alternate with [n] after a nasal: -kub-il-
versus -tum-in- (van den Eynden 1968). Because these Bantu languages do not
allow nasalized vowels, it is necessary to view such assimilations as
operating "at a distance" (Poser 1983), with the intervening vowel(s) being
transparent. While this phenomenon is known to be widespread within Bantu and
has been known to phonologists for some time (cf. for example the Lamba
problem in Kenstowicz and Kisseberth 1979:71-2), there has, to my knowledge,
been no systematic treatment of it. In fact, it seems regularly to be
overlooked. In the most recent and comprehensive statement on consonant
harmony to date, Shaw (1991:137) states: "I am not aware of any nasal harmony
processes that target only consonants". Rather than focusing on the more
restricted Bantu case found in Lamba, Bemba, Luba, Mongo, Suku, etc., in this
phonology workshop I shall carefully document a much more extensive system of
nasal consonant harmony in Yaka (or Kiyaka), a Bantu language spoken in Zaire.
In this language, suffixes such as applicative /-id-/ are realized nasal even
when the nasal is not the immediately preceding consonant, e.g., -mat-in-,
-miituk-in-, etc. As seen in these examples, voiceless consonants are
transparent to nasal harmony. While prenasalized consonants do not condition
nasal consonant harmony (hence, -biimb-il-), they too are somehow transparent
to the process, hence: /-niimb-id-/ --> -niimb-in- because of the initial /n/.
I will begin by showing that the nasal harmony effects are pervasive in the
language (within the stem domain) and then argue that they should be captured
by a phonological rule of [+nasal] spreading (i.e., not by allomorphy or
feature copying). Drawing on the work of Pulleyblank (1989), Piggott (1992)
and others, I will present an analysis which accounts both for the long-
distance properties of the consonant assimilations as well as the transparency
of both voiceless and prenasalized consonants. The paper concludes by
inviting the audience to join in speculations as to why nasal consonant
harmony appears to be so rare.
The materials presented in this talk are based on Ruttenberg's (1969) Lexique
Yaka-Francais/Francais-Yaka, the first part of which has been scanned and
entered into the Comparative Bantu On-Line Dictionary (CBOLD) project at
Berkeley. Part of my goal at the workshop is also to show how wonderful it is
to do this kind of work with an electronic dictionary organized by fields.
The talk will take place at the home of Sharon Inkelas and Orhan Orgun, in
Berkeley. The address: 2111 California Street. Phone: (510) 649-1470.
Directions: from University Avenue (exit from I-80), turn south on California
(which is between Sacramento and MLK Blvd). Cross Addison and park anywhere.
2111 is in the block of condos, upstairs, near the main entrance to the
complex.
____________
SEMINAR ON HUMAN-COMPUTER INTERACTION
on Friday, 4 November
12:30 p.m., McCullough Building, Room 134
Building Virtual Libraries on the Internet
Nick Arnett
Verity
<narnett@verity.com>
As anyone who's explored the Internet knows, millisecond access to terabytes
of information can be more of a problem than an opportunity. Intellectual
property owners of many kinds, eager to tap into the Internet as an
inexpensive distribution channel, are seeking ways to make libraries of
information available. These libraries can be "virtual," in the sense that
they exist only as a collection of pointers to documents, rather than a
physical collection. Thus, multiple "card catalogs" and other information
navigation tools, potentially reflecting many purposes and points of view, can
exist in parallel.
Copyright, performance and maintenance problems call for a distributed library
model. Formatting and conversion obstacles remain for both documents and
"meta-information," although emerging commercial products and standards claim
to solve the major problems.
Librarians, publishers and electronic forums are undergoing paradigm shifts as
their domains increasingly overlap. Publishers find themselves acting like
librarians as they build electronic collections of documents; librarians are
beginning to behave more like publishers by distributing documents and
"packaging" information; electronic forums such as Usenet and CompuServe are
becoming social information search and retrieval tools in their own right.
The World-Wide Web and related emerging standards for "Uniform Resource
Information," combined with databases and search engine-based information
"agents," hold great promise for construction of virtual libraries.
Important issues of maintainability and scalability continue to beg the
question of information structures. Opinion-based navigation and organic
models may address these issues.
NICK ARNETT is the World-Wide Web product manager at Verity Inc., Mountain
View, California. From 1988 through August of 1994, Arnett was president of
Multimedia Computing Corporation, the leading market research and consulting
firm tracking multimedia technologies and markets. He previously was a
journalist with publications including InfoWorld and American City Business
Journals.
Arnett's recent essay, "The Internet and the Anti-net: Two Public
Internetworks are Better than One", argues that current debate over
information infrastructure should embrace the idea of separate internetworks
for advertising revenue-based and customer revenue-based information products.
The paper draws parallels between today's mass media and the Catholic Church's
domination of popular information in early modern Europe, suggesting that
technology today will play a similar role in restoring a balance of influence.
____________
LINGUISTICS COLLOQUIUM
on Friday, 4 November
3:30 p.m., Cordura Hall, Room 100
The Context-Dependency of Implicit Arguments
Cleo Condoravdi and Mark Gawron
CSLI and SRI
<condorav@csli.stanford.edu,gawron@ai.sri.com>
Relational predicates such as _win_, _local_, _within a day_, _imminent_,
exhibit a type of context-dependence that calls for an integration of
indexicality with anaphora (Mitchell 1986, Partee 1989, Nunberg 1992). Their
implicit arguments have indexical, discourse anaphoric and bound variable
readings:
(1) A local bar is selling cheap beer. (indexical)
(2) A reporter for the Times got seriously drunk.
A local bar was selling cheap beer. (discourse anaphoric)
(3) Every sports fan watched the Superbowl in a local bar. (bound variable)
In this talk we provide a unified account of these distinct readings within a
dynamic system of semantic interpretation extended so as to incorporate the
context of utterance. We analyze context-dependent predicates as having
presuppositions similar to those of definite descriptions (Heim 1982) with
respect to their implicit argument. The descriptive conditions associated
with an implicit argument as part of its meaning specification are minimal,
basically the sortal properties required by the predicate it is an argument
of; in a given context, however, they end up being richer as a result of
accommodation necessary for presupposition satisfaction. Our proposal
accounts for the varieties of context-dependent elements, preserves Kaplan's
(1989) insights about the semantics and logic of indexicals, and resolves the
problem of the scopal interaction between operators and context-dependent
predicates on readings determined by the context of utterance (Nunberg 1992).
____________
SEMINAR ON COMPUTATIONAL LEARNING AND ADAPTATION
on Wednesday, 9 November
4:00 p.m., Cordura Hall, Room 100
Density Estimation for Model-Based Computer Vision
Subutai Ahmad
Interval Research
<ahmad@interval.com>
A fundamental problem in computer vision involves the automatic learning of
useful, structured object models. I will argue that density estimation is a
critical component in learning these models, particularly when dealing with
complex non-rigid objects. If the image models incorporate both the underlying
joint probability distributions and the posterior distributions then it is
possible to optimally solve several important tasks. These include tasks such
as classification in the presence of noisy features, dealing with occlusions
and missing data, and the infamous problem of image-model correspondence. I
consider the specific case of Gaussian basis function (GBF) networks and show
that, under appropriate training regimes (such as the EM algorithm), GBF
networks allow for a closed form approximation to the above solutions.
Experiments with a complex problem (3D hand gesture recognition) validate the
theory for each of the above tasks. Just estimating the posterior does not
guarantee good performance on these tasks: estimating the underlying joint
distribution is vital.
This talk describes work done jointly with Volker Tresp.
The goal of this seminar is to increase communication among local researchers
with interests in computational approaches to learning and adaptation. If you
would like to be added to (or removed from) the mailing list, or if you are
interested in giving a talk in the seminar, please send email to
<langley@cs.stanford.edu>.
____________
STASS SEMINAR
on Thursday, 10 November
10:00 a.m., Cordura Hall, Room 100
Speaker/Title to be announced
____________
CSLI SEMINAR
on Thursday, 10 November
2:15 p.m., Cordura Hall, Room 100
The Digital Libraries Project
Terry Winograd
Stanford Computer Science
<winograd@cs.stanford.edu>
____________
SYMBOLIC SYSTEMS FORUM
on Thursday, 10 November
4:15 p.m., Building 60, Room 61-F
Robots that Plan and Robots that React
Illah Nourbakhsh
Stanford Computer Science
<illah@cs.stanford.edu>
Robotics has been an area punctuated by extreme approaches for several years.
Many roboticists, most notably nearer the Atlantic Ocean, have constructed
systems that react to changes in their sensors' input from the environment.
Other roboticists have created planning systems, demonstrating flawless
simulations of robotic navigators. We will take a short guided tour of these
two extremes and spend some time examining the middle ground, where
roboticists are very gradually converging. The issues that, in the end, must
be combined to create robotic systems include reactivity, interleaving
planning and execution, and assumptive systems that defer planning for the
possibility of worst-case scenarios.
____________
SEMINAR ON HUMAN-COMPUTER INTERACTION
on Friday, 11 November
12:30 p.m., McCullough Building, Room 134
How to Support Design with Research When Potential Users
Can't Imagine Future Products
Bonnie Johnson
Interval Research
<johnson@interval.com>
At one time, new product development was something done in a lab and presented
to stereotypical consumers for comment and refinement. "New and Improved" was
sufficient to generate interest. Now many new products are beyond the
understanding of their potential users. And many users are likewise beyond the
understanding of the marketers and developers who need to reach them. At a
time when thorough, accurate and perceptive market insights are most needed,
they have become increasingly difficult to get. This presentation will discuss
methods of "Applied Exploration" from observing "life contexts" of potential
users to synthesizing conclusions in the form of guidelines.
____________
PHILOSOPHY COLLOQUIUM
on Friday, 11 November
3:15 p.m., Building 90, Room 91-A
Reasoning in Hume
David Owen
University of Arizona Philosophy
Most discussion of Hume's views on reasoning, especially demonstrative
reasoning, have assumed that Hume is working with a concept of deduction,
formally construed either in the traditional syllogistic way or in the modern
sense. Hume's own work, and the background of Descartes and Locke, makes it
clear that he rejected any formal conception of deduction. This paper
attempts to reconstruct Hume's own views on demonstration, and puts forward
new solutions to several problems traditionally found in Hume: the relation
between conceivability of the contrary and demonstrability, the limitation of
demonstration to algebra and arithmetic, empirical measurement, what it is to
imply a contradiction, and the nature of demonstrative and probable reasoning.
____________
LINGUISTICS COLLOQUIUM
on Friday, 11 November
3:30 p.m., Cordura Hall, Room 100
Underspecification by Optimization
Sharon Inkelas
UC Berkeley Linguistics
<inkelas@cogsci.berkeley.edu>
Past theories of underspecification (e.g., Ringen 1975, Kiparsky 1982,
Archangeli 1984, Steriade 1987) have based underspecification on such
principles as markedness, redundancy in inventories, or predictability.
Critics of underspecification (e.g., Mohanan 1991, McCarthy and Taub 1992,
Steriade 1994) have, however, pointed out specific problem cases for each such
principle, concluding that underspecification is not a viable tool in the
analysis of various phenomena previously claimed to require it.
I argue in this paper that while it is true that past principles are
inadequate to fully characterize the distribution of underspecification, it is
not true that underspecification is dispensable or uncontrollable. I propose
a new theory, called Archiphonemic Underspecification, in which
underspecification is regulated by simple, deterministic principles of
optimization. Lexicon Optimization, adapted from work in Optimality Theory
(Prince and Smolensky 1993), selects underlying form so as to optimize the
phonological mapping from input to output. Grammar Optimization, adopted from
Kiparsky 1993, ensures that the input-output mapping is maximally
structure-preserving (in the sense of not deleting structure). Together,
these principles predict underspecification in exactly the case of predictable
alternations, and nowhere else. Specifically, predictable but nonalternating
structure is fully specified underlyingly.
>From the perspective of this new account, the failure of past principles of
underspecification is understandable: they all attempt to regulate
underspecification without regard for the alternations in any given language.
Archiphonemic Underspecification underspecifies only in case of alternations,
enabling the description of ternary contrasts and numerous other phenomena
which past proposals could not account for.
____________