🚀 go-pugleaf

                                 S E M I N A R

WHERE:         University of Rochester
               Computer Science Department
               632 Computer Studies Building

WHEN:          Friday, October 30, 1992
               3:00 P.M.

SPEAKER:       REBECCA PASSONNEAU
               Columbia University


             Towards Intelligent Discourse Segmentation*

Lexico-grammatical choices and their semantic and pragmatic interpretation,
or what I will call the micro-structure of discourse, consisting of relations
among discourse segments defined in terms of their relative cohesiveness,
topicality, or relation to speaker intentions.  The nature of the constraints
between these two levels needs to be understood and verified in order to 
exploit them in building natural language systems.  Open questions include, how
many cognitively justifiable ways are there to segment an individual discourse?
Is discourse segment recognition continuous or discrete?  What are the most
reliable surface cues for signalling segmental structure, and under what
conditions?  I will report on the initial results of a long-term study of a
corpus of oral narrative monologues.  The two initial goals of our study are to
evaluate the ability of naive subjects to use speaker intention as a basis for
recognizing discourse segments, and to quantify how various linguistic cues
independently correlate with perceptually salient segment boundaries.
Statistically significant levels of agreement across subjects in the segmenta-
tion task suggest that segments are perceptually real, whatever their role in
on-line processing might be.  As a tool for analyzing linguistic cues of
segmental structure, I developed a simple algorithm for assigning segments
based on the form and distribution of referential noun phrases.  The percentage of perceptually salient boundaries identified by the algorithm varies with the 
level of agreement across subjects, suggesting that discourse segment recogni-
tion is a gradable phenomenon.  A preliminary comparison of the algorithm with
cue phrase distribution shows each cue to be roughly equally reliable, but
neither is reliable enough to be used alone.

*This work is being carried out jointly with Diane Litman (AT&T).