Dialogue Management and Commonsense Server

(with Hideaki Takeda, Kenji Iino, Michiaki Iwadume, Katsuji Onishi,
Tadahisa Okimoto, Yoichi Kado, Motoyuki Takaai, Kazuhiro Takaoka,
Masatoshi Nishiki, Kenji Hanakawa, Shoji Yasumura)

Graduate School of Information Science, Nara Institute of Science and Technology
8916-5 Takayama, Ikoma, Nara 630-01, Japan
e-mail: nishida@is.aist-nara.ac.jp

In this research, we focus on commonsense knowledge referred to in spoken dialogue. We attempt to build a computational model of commonsense knowledge and develop a knowledge server that can provide commonsense knowledge for various components of spoken dialogue system.
In this fiscal year, we have made a study on the methodologies of constructing commonsense knowledge base with particular emphasis on ontology.
First, in order to obtain insights on the feasibility and cost of building an information base from existing documents, we conducted an experiment of producing a structured document consisting of an ontology and associated pieces of text fragments. After having worked out the whole process for a small fraction of a technical document, we took a chapter of a technical document (258 pages in Japanese) as sample and asked students to decompose them into a structured document. The resulting ontologies contain about 1,100 conceptual elements and 1,200 relations. We have partly implemented a prototype information system which can present information upon a given perspective by reconfiguring text fragments according to the ontology.
The result of the experiments was quite promising. About 116 hours are required to manually produce a structured document from the sample document. If the experiment had been made with the contextual media, the cost would be significantly decreased for the sake of various computational supports. Our working hypothesis is that building an information base from existing documents is quite feasible and cost effective even though it is to be handcrafted.
Second, we have developed the theoretical framework of the { contextual media}. The contextual media allows information of varying degree of structural sophistication to be integrated and evolve as information accumulated into the {information base} increases in quantity and quality.
Primary elements of the contextual media are {units}. A unit plays two mutually complementary roles. On the one hand, a unit represents a {concept} and is used as a lexical item in contextual media statements. On the other hand, a unit specifies a {context} in which statement in the contextual media is made. The most primitive relation in the contextual media is membership of a unit (as a concept) in another unit (as a context). The statements in a context may be arbitrarily sophisticated using known knowledge representation techniques and are organized into an abstraction hierarchy in which a more sophisticated representation layer is encapsulated into a less sophisticated representation layer. The technique is called {encapsulated representation}.
We have developed several techniques for making the contextual media approach useful. {Connection-based associative retrieval} allows the user to access information with little prior knowledge about the contents and structure of the information base with the contextual media. {Knowledge elaboration assistance package} helps the user elaborate an information base with respect to the terminology and structure, by mining tacit information structure in the information base.
We have made several preliminary studies, which suggest the feasibility and effectiveness of the contextual media approach.

Third, we have studied theoretical issues in the development of multiple ontologies in building a distributed and heterogeneous knowledge-base system.
We have analyzed relationship between ontology and agent in the { Knowledgeable Community} which is a framework of knowledge sharing and reuse based on a multi-agent architecture. Ontology is a minimum requirement for each agent to join the {Knowledgeable Community}.
We have explored a technique for mediation by ontology. A special agent {mediator} analyzes undirected messages and infer candidates of recipient agents by consulting ontology and relationship between ontology and agents.
We have developed a model ontology which is a combination of aspects each of which can represent a way of conceptualization. Aspects are combined either as {combination aspect} which means integration of aspects or {category aspect} which means choice of aspects. Since ontology by aspect allows heterogeneous and multiple descriptions for phenomenon in the world, it is appropriate as ontology of a heterogeneous knowledge-base system.
We have shown translation of messages as a way of interpreting multiple aspects. A translation agent can translate a message with some aspect to one with another aspect by analyzing dependency of aspects. Mediation and translation of messages make it easy to build each agent because it is required to have less knowledge on other agents.

Keywords: commonsense knowledge, knowledge sharing and reuse, knowledge media, ontology, mediation