Hiroya FUJISAKI and Sumio OHNO
Department of Applied Electronics, Science University of Tokyo
2641 Noda, 278 Japan
e-mail: fujisaki@te.noda.sut.ac.jp
Spoken language is clearly the primary means of human
communication, and will no doubt be the major means of man-machine
communication in the future. Compared with the written language,
however, messages of the spoken language are less well-formed, with
frequent occurrence of omissions, repetitions, word-order changes,
errors, etc. Yet these so-called ill-formed features do not seem to
present any obstacles to a human-to-human spoken dialogue. This leads
us to believe that these features possess their own reason of
existence and are actually utilized by the participants of a spoken
dialogue to facilitate smooth and efficient information exchange.
From this point of view, the present study analyzes human spoken
dialogues and then tries to construct a model for the process of a
spoken dialogue, with an aim of its utilization in human-machine
dialogues.
The results of our preliminary analysis of human dialogues have been
summarized in terms of
(1) Prerequisites to a spoken dialogue -- purposeful exchange of information
(2) Basic constrains of a spoken dialogue -- parsimony and real-time production
(3) Principles and methods for a smooth spoken dialogue -- cooperation
and reliability.
Most of these findings, however, have been obtained by studying the
recorded spoken dialogues, which directly reflect the processes of
selection and emission of information on the part of the speaker. The
processes of information processing on the part of the listener,
however, are only implicitly reflected in the recorded dialogues. For
a clear understanding of the human processes involved in a spoken
dialogue, further investigation is necessary into the perceptual and
cognitive capabilities of the listener.
The above-mentioned results have been further examined from the
point of view of their possible utilization in designing a system of
man-machine spoken dialogue. Our investigations clearly indicate that
the machine has to have appropriate models for the knowledge and
intention of the human counterpart in order to conduct an intelligent
and smooth dialogue, both in understanding the spoken messages uttered
by a human speaker and in generating the spoken messages to be given
to a human listener.
A study has also been started on the criterion and the
methodology for assessing the process of spoken dialogue. As a
criterion for the assessment, we introduce, at any point during the
course of a spoken dialogue, an index of achievement of its intended
goal(s). As for the cost function, we may adopt the weighted sum of
the time and the amount of information exchanged to attain a given
level of achievement. Further work is under way to develop a more
detailed theory and method of evaluating spoken dialogues.
Keywords: spoken dialogue, cognitive model, parsimony, real-time production, principle of cooperation, principle of reliability, models for the knowledge and intention of speaker/listener, assessment of spoken dialogue