Kuniaki UEHARA, Minoru TANIGUCHI, and Hiroshi OGINO
Department of Computer and Systems Engineering, Kobe University
Nada, Kobe 657, Japan
e-mail: uehara@jedi.seg.kobe-u.ac.jp
Almost all current dialogue systems are generalization-based, in the
sense that they use explicit abstract generalizations. Within
computational linguistics, two main approaches have dominated, the
first using schemata to constrain the overall organization of the
text, and the second using plan operators to generate a sequence of
utterances given particular communicative goals. Schema-based
approaches cannot participate in an interactive dialogue with users.
In particular, they cannot elaborate on previous utterances or respond
to follow-up questions in the context of the on-going dialogue.
Plan-based approaches are limited because they know nothing of
stereotypes, and they treat each dialogue as completely novel. This
limitation dooms them to perform expensive text planning every time
they recognize a new instance of dialogue.
In this research, we will propose a case-based text planning approach
to dialogue systems in which feedback from the user is an integral
part of the text planning process. Case-based planning is the
paradigm that devises new plans by retrieving and adapting old ones
from memory. The basic model of case-based planning is obviously
incomplete if we adopt the model as the text planner of a dialogue
system. It is incomplete because this mechanism ignores the following
problems: (1) Users frequently ask follow-up questions requesting
clarification, elaboration, or re-explanation. (2) Users may
misunderstand the explanation if they do not have enough or accurate
knowledge about the domain.
In order to deal with the former problem, we make use of a reactive
approach which employs feedback from the user to guide subsequent
dialogues. To accomplish this task, the system is designed not only
to generate and execute text plans, but also to interrupt and modify
them, when the user asks a follow-up question or indicates that he
does not understand the explanation. In order to deal with the latter
problem, we adopt an approach where the system learns to anticipate
and avoid user's misunderstanding that it has previously encountered.
When the misunderstanding re-occur in later situations, the text
planner is reminding of the past misunderstanding and this reminding
serves as a warning to the text planner that it has to plan for the
fact that this misunderstanding is going to occur again. The text
planner also stores the repaired plans that were built in response to
past misunderstanding.
The system described here were implemented as a natural language
consultation system for the UNIX operating system, called ASSIST-R.
Users can ask ASSIST-R how to do things in UNIX, get definitions of
UNIX terminology, and get help debugging problems in using commands.
Currently, we are using an adaptation of Dyer's McDYPAR for the
Japanese query analyzer and Japanese Tree Adjoining Grammar for text
generation mechanism.
Keywords: Dialogue system, case-based planning, text planning