"Real-time" Language Use by "Real" Agents Penelope Sibun

From: AAAI Technical Report FS-95-05. Compilation copyright © 1995, AAAI (www.aaai.org). All rights reserved.
Use by "Real"
Penelope Sibun
P. O. Box 620527
Woodsidc CA 94062
[email protected],corn
People do not speak in the well-formed sentences
that we find in written text. In fact, a careful transcription of spoken language will reveal that it is generally
composed of short fragments of text that are strung
together. While hierarchical sentence structure can often be imposed on such texts, an alternative analysis
that takes the fragment as the basic unit of language
accounts for most linguistic phenomena, and has the
advantage of constituting a simpler explanation. In
particular, a fragment-based analysis is well-suited to
a model of incremental text production.
In mythesis (Sibun 1991), I presented Salix, a program for incremental text generation. Salix produces
texts about things that are obviously structured, and
uses the structure of the subject matter to structure
the text it produces. The types of fragments that Salix
produces, and the strategies it uses for choosing what
text to generate next, were created by examining how
people generate texts in similar situations. Salix’s major limitations lie in the constraints imposed by the
initial selection of situations in whichit could produce
text. Salix has so far been limited largely to monologues about richly-structured objects such as families
and house layouts, about which it has readily available
all the information it will use.
I am working to broaden Salix’s horizons in two
directions simultaneously. First, it will be a greater
test of Salix’s ability to always find something to say
next if it is learning about its subject matter at the
same time at which it is describing it. For example,
suppose Salix takes its knowledgeof how to talk about
a building’s layout and explores a real building that it
has not seen before. In this case, Salix needs a way to
extract information from the environment that it can
map onto its strategies for producing text.
Second, I would like to expand Salix’s environment
to include other agents who would also be generating
text of their own. Salix’s context includes information about what has already been said. Heretofore,
the only speaker of interest has been Salix itself; in order to make an intelligent choice of what to say next
in a multiagent environment, Salix needs to be able
to incorporate into its context what other agents have
said as well. An obvious way of combining these two
goals is to join Salix’s language capabilities with those
of a mobile robot with vision.
A reM-time environment that is changing both
physically and by virtue of interaction with other
agents is a demandingone. I believe that for an artificial agent to have any hope of using language in such a
context it must be able to make simple choices quickly
and continuously
about what to say next. These
choices will always be highly constrained by the changing context, and will each result in a short segment of
appropriate text. The emphasis in such text generation will not be on producing well-formed grammatical
sentences but will instead be on the continuous ability
to produce text that is appropriate to the moment.
Penelope Sibun. "The Local Organization of Text."
In Proceedings of the Fifth International Workshop
on Natural Language Generation, pp 120-127, Linden
Hall, Dawson, PA, 1990.
Penelope Sibun. Locally Organized Tezt Generation.
COINSTechnical Report 91-73, Department of Computer and Information Science, University of Massachusetts, 1991. Also Report SSL-91-21/P91-00159,
Xerox Palo Alto Research Center, 1991.
Penelope Sibun. "Generating Text without Trees." In
Computational Intelligence: Special Issue on Natural
Language Generation, Volume8(1), pp 102-122, 1992.
Penelope Sibun. "Domain Structure, Rhetorical Structure, and Text Structure." In Proceedings of Intentionality and Structure in Discourse Relations, a Workshop
sponsored by the Special Interest Group on Generation of the Association for Computational Linguistics,
pp 118-121, Ohio State University,
Columbus, OH,